Warranty operation enhancement through social media knowledge: a deep-learning methods

Abstract

People use social media as free channels to share their sentiments and experiences during all stages of consuming a product or service. Likewise, corporations depend on social media as feedback sources that influence the positioning of their products/services in the market. This paper aims to recognise the frequent product flaws and warranty issues through social network mining. We have performed ontology-based methods, text mining, and sentiment analysis using deep learning methods on social media data to investigate product failures, symptoms, and the correlation between warranty programs and customer behaviour. Correspondingly, a multi-sources mining approach has been incorporated into social media mining to cover all the occasions. Furthermore, we promoted a decision support system to learn practically through customer feedback. Finally, to validate the accuracy and reliability of the results, we used the claimed data of the laptop industry to compare our derivatives and machine learning validation metrics to ensure accuracy.

Keywords:

• Social media analysis
• data mining
• decision support system
• warranty service
• deep learning

1. Introduction

The most important way to elevate products and develop services is by using customer comments and preferences (Cui et al. 2017). Classic techniques for gathering data about product¹ failures and defects were questionnaires, warranty claims, customer complaints, and after-sales service providers’ data. However, these methods are restricted and not comprehensive (Alkahtani et al. 2018; Zheng et al. 2020). Since time and population constraints affect the data gathering comprehensiveness in products’ failures during the warranty period, it is recommended to use a more considerable population of users in addition to the classical techniques. Considering the position of social networks in today’s business and their vital role in shaping the reputation of organizations, products and services, some of the most important channels to examine the reflection of a product/service are social networks (Alkahtani et al. 2018; Guo et al. 2020; Zheng et al. 2020, 2021).

As mentioned above, discovering all kinds of product defects is not reachable in warranty channels. Besides, a warranty as a contractual agreement obligates the warranty service providers to repair all the product failures using the repair or replacement solutions. Therefore, it is essential to assume a suitable approach to discover the frequent failures and then adopt the right warranty strategy to increase customer satisfaction and decrease expenses (Fang and Hsu 2019; Shokouhyar et al. 2021). One of the most significant issues that warranty service providers face is recognizing the frequency of product defects. Understanding the relationship between the product’s failures and the designing, manufacturing, repairing, and maintaining processes are essential in order to find opportunities for enhancing quality and reliability due to the costs associated with failures and the consequence of product failure on customer satisfaction (Alkahtani et al. 2018; Fang and Hsu 2019; Grambau et al. 2019; Shokouhyar et al. 2021).

Previous literature such as (Shokouhyar et al. 2021; Zheng et al. 2021; Sarmast, et al. 2023) confess social media data as a vital information reference for discovering product weaknesses. In fact, the importance of the combination of social media mining with machine learning tools is stated in Pourranjbar and Shokouhyar (2023) and is used for electronic waste management prediction which is a significant post-sale concern along with other post-sale costs. However, they mainly focus on the organization and classification of defect-related texts and ignore the details of these defects, which can provide worthy oversight understanding (Zheng et al. 2020; Zheng et al. 2021). In the preceding literature, researchers used product defect data from social networks to sustain decision-support models in manufacturing to produce more qualified products. In fact, what has been mainly discussed is product troubleshooting through design modification and manufacturing mechanisms. Moreover, papers focusing on warranty solutions mainly used warranty claims data.

This paper focuses on social media mining using data mining methods based on ontology to investigate and reveal the knowledge beyond the users’ feedback and use the informative threads extracted in a sentiment analysis process through deep learning techniques, which are used to scan the sentences and find the related texts (Trappey et al. 2018; Zheng et al. 2020). We tried to keep our approach domain-neutral² and reveal the potential aspects of social networks in integrating the complete information from multi-source channels³. A framework is determined to use the information gained from social media to lead the warranty service providers in the decision-making process (Jeong et al. 2017; Alkahtani et al. 2018; Wang et al. 2020). This framework discovers the most frequent failures, vulnerable components and main causes for defects of a product and provides the service provider with insight into suitable strategy. We will confirm that the useful threads found in social media data can cover customers’ most frequent failures compared to the claimed data. The specific help of the current model is that the company would get aware of the problems before a warranty claim occurs (Kulesza et al. 2019; Shokouhyar et al. 2021).

The current research widens the sights of how social media data can assist warranty service providers in discovering the most vulnerable components of their products and improving their service by maintaining the right choice between repair or replacement strategy using online data, which is barely discussed in other researches. The article’s innovations are considered in more detail as follows.

This paper contributes to three main aspects.

• First, using the multi-channels of social media data leads to a wider source of data related to warranty services and is not limited to warranty claims.

• Second, it reveals the frequent failures, including the related component and symptoms of the defects and the main causes for a product’s features’ flaws, using a deep learning method based on the product’s ontology.

• Third, it presents a model to improve the warranty services strategies by considering the details of product defects to discover the vulnerable components, symptoms, and main causes to organize the warranty plan and optimized product-service⁴ approach.

2. Research questions and process

Research questions are raised in the following, and try to discuss the main concerns of the research, the study procedure, and calculations:

1. In relation to stated data, how much valid information can be obtained via social network data mining to improve products and services?

2. Can a warranty service provider enhance and adequately support its decisions, operations, and strategy using the information they have obtained?

3. Which optimization factors can be used to assist the use of a decision support system in warranty service??

The above questions are answered through the following approaches and procedures:

1. Examining the literature on warranty service models and product defect identification with an emphasis on social media mining, employing multi-channel social networks, the calculation method, and the numerical example with a statistical comparison between social media findings and claimed data⁵ (Tuarob and Tucker 2015; Liu et al. 2019).

2. Applying deep learning techniques based on ontology and data mining to social media data related to the relationship between product defects/warranty issues and customer usage, providing new insights into social media data in warranty service provider strategies. Establish a role.

3. A comprehensive decision support system is presented to model decision factors and gain broader clarity through customer feedback, leading to better warranty operations.

The whole model seeks to answer the questions through a systematic and methodological approach, and its summary is shown in Figure 1 (Shokouhyar et al. 2021).

Figure 1. Research methodology.

We defined three steps to structure the model. Since our research focuses on three main areas: (1) Product defect detection, (2) Warranty services, (3) Social media mining, the first step begins with a literature review based on the following specifications. We reviewed related publications and studies to identify examples and similar models that have worked for product defects from user feedback (including social media). In addition, the warranty rules and standards of warranty service providers have been revised. Considering text mining and deep learning methodology as the main research method, we explored various models and tools. Then we focused on the connections of defects identified by social media with warranty programs (repair/exchange solutions). In the second step, we collected data from LENOVO's warranty service provider to explore ontology and terminology, device architecture, and common customer failures. Next, we explored social media sources using the Product Failure Ontology and mapped them to warranty models, plans, and devices. We then proposed a decision support system to detect equipment failures and repair warranty services. In the third step, we presented the conclusions as part of a comprehensive calculation and discussion. Finally, we presented strategic proposals, conclusions and recommendations for future research (Shokouhyar et al. 2021).

3. Literature review

In this section, according to the research approach, which is the warranty service improvement based on the discovery of the product’s frequent defects through social media, there is a review of warranty service solutions and their combination with the Product-Service system. Then, the literature on using social network mining and its relative methods in order to analyse the results with warranty services will be presented.

3.1. Warranty service

The existing literature on warranty service is extensive. First, consider the need for warranty management and its main functions. Murthy and Blischke (2010) showed that from the customer’s perspective, warranty service should be considered as the next role. (1) ensure the necessary reliability, and (2) ensure sustainability. In addition, from the manufacturer’s point of view, three important features in warranty service have been introduced. (1) limiting bad services and accountability, (2) distinguishing products and related services from competing products, and (3) using effective promotional tools. In fact, good and robust warranty management is only successful if the design system and development process are synchronized with customer feedback, can be changed and improved based on data obtained during the warranty period, and proper guidelines are applied. These are confessed by Tapiero et al. (2019) but only the post-sales product failure costs are admitted in their dynamic approach to quality control when manufacturing product using the economic design of quality control samples. However, this paper proves the importance of social media mining in bringing the result of failures and their frequencies to business management decisions. According to Blischke and Murthy (1992), warranty policies have classifications that serve as primary levels for further investigation and group policies based on post-sales product development.

The warranty service offerings in our product service system can be categorized into the following basic warranty types: (1) Free Replacement Warranty (FRW): This means that consumers can have their products repaired and defective components replaced free of charge. (2) Prorated Guarantee (PRW): This is a type in which the customer bears part of the predetermined repair costs within the warranty period. (3) Combined Warranty (CW): The CW type is a mixed type of FRW and PRW, and the customer can select the type of product warranty (Fang and Hsu 2019). In addition, there are also complex types of warranties. In such cases, different warranty terms apply to different product components (He et al. 2018). The warranty period is not fixed to a specific period and is determined by product usage and similar factors (Murthy and Blischke 2000). Iskandar and Murthy (2003) developed a complex warranty-type repair/replacement strategy model that describes a method to repair when a defect occurs at a certain age of the product or when it reaches a certain level of use. Regardless of the different types of warranties, this paper focuses on the failures that could happen at any time during the warranty period or even after the warranty period. Therefore, the whole data mining system could provide results for any manager seeking improvement in any steps of post-sale development.

Further aspects of warranty management need to be considered by companies; costs and customer satisfaction are a direct variable, which needs to be managed so that it remains at an optimal level. In order to ensure that both factors are balanced, companies need to carefully process customer feedback in the duration of warranty periods with a view to determining an optimum period and policy. Special effort has been made by researchers to improve the operation as a whole. Detailed and accurate data on the performance of a system can be derived from modern production systems by means of information systems. For example, Jeong et al. (2007) proposed a decision support system that brings process-state knowledge from several crucial channels and uses these elements to identify the most potential cause for a product issue. Also, Fang and Hsu (2019) describe a preventive maintenance policy to decrease customer loss, ameliorate dissatisfaction, and reduce warranty expenses. Moreover, (Dai et al. 2017) identify causes of failure in designing a suitable warranty policy based on two-dimensional warranty data. In addition, Arabi et al. (2017) provide a model based on game theory to promote a novel approach to illustrate the optimal warranty period and the out-of-warranty replacement period. Zuo et al. (2000) promotes a policy for warranty service regarding multi-state repairable products to determine the repair/replacement policies by using models for forecasting warranty cost and the effecting factors such as a combination of replacement and minimal repair. However, there is little effort in providing a streaming system using an LSTM network to deliver defects of a product from social media mining to a warranty cost management system.

3.2. Social media data for warranty service

providing valuable information and data for customers and interested parties is a key purpose of websites and content on Social Media (D’Haen et al. 2016). It has been established that, given the characteristics of various external stakeholders, customer satisfaction can be adversely affected by social media analytics (Wang et al. 2020). The data from the website has been shown in previous reviews to give more detailed analyses than the information available on the market (Meire et al. 2017). A useful tool for corporate awareness of the community and its products and services is social media. Because the social media pages are updated daily, they offer more benefits than websites. Platforms are also more readily available and available at hand, such as Facebook and Twitter posts, in view of the regular use of social networks. As shown by Meire et al. (2017), social media data is much easier to analyse than unstructured text on websites (Chan et al. 2015). As a result of this study, with data analysis using the Samsung Mobile Facebook page, the role of social media in improving operating and production decision making has been proven (Chan et al. 2015). That is why this research concentrates on different channels of social media and ensures their results by comparing them with operational data that are derived from warranty claims (Sharif & Shokouhyar 2021).

For corporations, the tacit knowledge beyond social network data provides new opportunities and challenges to identify and exploit this source of information with a view to improving operational processes (Ramanathan et al. 2017; Choi et al. 2020). For example, Matthias et al. (2017) has focused on applying and exploiting Big Data to create competitive advantages by introducing a model to show how they help sustainable improvement and customer satisfaction. Beyond the sentiment analysis in social media channels, Ullah et al. (2023) contributes to the creation of a quality-related lexical dictionary based on an evaluation framework and the automatic labelling of the data in accordance with those labels prior to their use as training data for the BiLSTM model. A dataset of Amazon product reviews is used to assess the suggested model which is approximately near to the deep learning approach of the current paper but lacks the effort of converting the output result to useful data for the cost management system. Facebook now has 300 million, and Twitter is 396.5 million users, as stated by Statista. And among users aged 25–34, Twitter is the most popular. Therefore, in order to gain valuable customer insights and use them to advance business performance, it is necessary to be agile enough to take advantage of the knowledge results (Wieneke and Lehrer 2016; Trappey et al. 2018). So, Kühl et al. (2019) present a feasibility study by prioritizing and quantifying the customer essentials and requirements from social media data using a supervised machine-learning method on the Twitter data from the field of e-mobility. Extracting useful knowledge from user-generated media is also explained and confirmed by Wang et al. (2018). In this regard, to cover the risks of biased ideas in one social media channel, such as Twitter, we utilized different social networks that attract a wide diversity of people with different levels of education and professionality. To facilitate the organizations using the unstructured data from social media, Yadav and Vishwakarma (2020) provide a survey of famous deep learning models utilized in sentiment analysis using taxonomy implications. Moreover, to complete the approach of using data analysis in intelligent manufacturing and service systems, Shukla et al. (2019) review the papers providing big data analytics to improve the decision-making process, and Ali Hasan et al. (2018) adopted a hybrid approach that involves a machine Learning-based sentiment analysis for Twitter Accounts. Besides, Zhan et al. (2020) has researched how social media can deliver virtual platforms to enable organizational learning and innovation in the new product development process. However, we chose the LSTM networks to provide deep learning since Long Short-Term Memory networks are a type of recurrent neural network capable of learning order dependence in sequence prediction problems. This is the behaviour in complex problem areas such as machine translation and clustering of words in social media channels.

Focusing on social media data for defects and warranty service, there has been some research about identifying the product’s discontinuation and disruption from social media with illustration (Zavala et al. 2019). Moreover, Jin et al. (2015) demonstrate discovering customer needs based on data mining to improve product design. It also has worked on sentences belonging to specific features and components of products and then discovered the engineering factors that can be enhanced. Since we have provided an ontology-based system and taught our system to investigate the defects related to the products’ features and components, the mentioned usage is also applicable. To continue the investigation through social media data for product corrigible features, Zheng et al. (2020) uses three filters for sentiment analysis, component-symptom examination, and similarity investigation to present a product defect detection model (PDDM) based on deep learning methods while Agüero-Torales et al. (2021) work on deep learning methods based on multilingual sentiment analysis on 11 social media data. To collect and analyse the customers’ feedback based on ontology, Alkahtani et al. (2018) provided a Decision Support System (DSS) to decide on warranty failure information correctly which is only based on operational data that are derived from warranty claims and do not cover the out-of-warranty period or defects that are not resulted in a claim. and (Gavrilova et al. 2017) used a managerial practice for effective knowledge management based on ontology engineering and knowledge structuring techniques.

3.3. Decision Support Systems

Decision Support Systems primary focus is to encourage good decision making by organisations. By providing sufficient solutions and improvements in planning, the DSS has managed to reduce complexity and uncertainty over the past 50 years. To exfoliate the DSS, Kumar and Thakurta (2021) have revealed the three underlying themes of DSS as ‘Plan’, ‘Design’, and ‘Use’ by using Lexical analysis, topic modelling, and other data mining techniques to develop content analysis procedure. Ko and Gillani (2020) has developed a taxonomy for DSS using semantic text mining to improve business analytics. However, these systems lack a live decision support system that could facilitate decision-making and help managers with a predictive approach. Also, complexity and uncertainty have led organizations to collaborative planning, and multiple partners participate in deciding on a subject successfully. Rockwell et al. (2009) has provided a Decision Support Ontology (DSO) to enable the process of decision-making within a joint approach, and it has been built upon the Web Ontology Language (OWL) which is the base of our research in analysing the social media data.

Using customers’ feedback and perspectives improves the total approach to managing a company. Ducange et al. (2019) Uses a DSS that can efficiently help companies in managing promotional and marketing campaigns on different social media channels. Furthermore, Nave et al. (2018) have provided a DSS framework to follow customer sentiments in social media, with a focus on the tourism industry and the SMEs⁶ that can’t afford heavy expenses for monitoring their reputation through social channels and reviews. Warranty issues and the prediction of defects frequencies and using them for different managerial targets, such as warranty cost management issues, are the potentials that are hardly found in researches. In close relation to this paper’s objective, Abrahams et al. (2014) has provided a framework to investigate product defects from user-generated texts and social media data with a focus on the automotive industry which relies only on one source of data. Considering the growth in today’s competitive market, there is always a petition to diminish repair and warranty costs. So, Pérez-Fernández et al. (2017) have introduced a DSS to decrease costs belonging to warranty services and promote a method in which the manufacturers can prioritize the decisions about the product’s parts, improving to accelerate cost optimization. In addition, the efficacious root causes diagnosis of faults and providing a first-time repair would decrease the operation cost. The mentioned research limited the data input on the real customer claims and ignores the potential dissatisfactions that are not found in operational data sources.

3.4. Text mining

The increasing use of social media for making a framework to analyse social content has been confirmed and demonstrated by Ampofo et al. (2015). It has explained the challenges of conducting text mining in online social media environments, such as the availability of social media data, research ethics, and the integrity of the data collection. Some research has also presented a detailed survey of the text mining tools and methods as well as the methodology adopted by Kim and Ha (2022) to propose a new deep learning residual control chart based on the asymmetrical count response variable with highly correlated explanatory variables. Moreover, Maynard et al. (2017) has revealed a framework that incorporates collecting data, semantic research, aggregation, semantic inquiry, and visualization tools. This approach allows reviewers to deeply explore the data and execute complicated semantic search queries over social media content (He et al. 2016). In this regard, Kabakus et al. (2017) proposes a TwitterSentiDetector that classifies tweet sentiments into positive, negative, and neutral using polarity scores derived from sentiment lexicons together with the suggested linguistic methodologies. Furthermore, Shen et al. (2019) employ bilingual text mining to analyse the trends of online to offline commerce from the social media point of view while the current research relies on machine learning and deep learning techniques to enhance the dynamic approach to mining live data.

Digging through social media content relies on specific reasons and purposes (Trappey, et al. 2018). This paper proposes a DSS framework based on social media data with a focus on customer opinions on after-sale services, especially warranty services (Aggarwal and Wang 2011). In this regard, Khare and Chougule (2012) have worked on a ‘Domain Aware Text & Association Mining (DATAM)’ system that enhances after-sales service and repairing for the automobile industry. Using the most frequent symptoms of defects demonstrated that DATAM can recognize all anomalous signs through the following steps: (1) analysing the service manuals and instructions to identify the repairs, (2) association mining to find the social media content, (3) semantic comparison and mapping the previous steps, and, (4) checking for taking suitable action that are the close basis for designing our ontology-based system. Rajpathak et al. (2012) has presented a novel model to monitor the devices continuously and take corrective actions in the case of defection. So, it proposes a ‘domain-specific Association and Text mining system for Knowledge discovery’ (ASTEK) that integrates both history and knowledge-based approaches. A summary of the most related literature and gaps covered in this research are shown in Table 1.

Table 1. Summery of literature in different fields of study-sorted by year.

Year	Authors	Research Domain	Social media Data mining	Multi- source data	Defect- related discovery	Case study	Technique/Method	Main findings
2000	Zuo et al.	Warranty Service	–	–	✓	–	Mathematical Programming Numerical Application	Warranty service policy for repairable products with the approach of repairing or replacing a failed item based on the damage size and the length of the residual warranty period to minimize the expected warranty servicing cost
2010	Murthy Blischke	Warranty Service	–	–	–	–	Review methods and tools	Difficulties encountered in the research of warranty data practical tools for examination of one- and two-dimensional warranties failure detection and reporting algorithm
2014	Abrahams et al.	Product defect analysis	–	✓	✓	Automotive Industry- Honda, Toyota, and Chevrolet websites	Defect tagging, product feature tagging, text mining	Propose an integrated framework based on text analytics to discover the product defect from social media content which is called SMART (social media Analytic framework using Text)
2015	Chan et al.	Product R&D	✓	✓	–	Samsung mobile Channels: website & Facebook	Content Analysis Pearson correlation coefficient	Finding the interrelationships amongst critical factors through the social media data investigation with the statistical cluster analysis technique
2016	D’Haen et al.	Product Marketing	✓	–	–	–	Web Crawling Text Preprocessing Spherical Clustering	A decision support system to assist sales representatives in selecting leads based on sounder reasoning using web crawling and data mining
2017	Matthijs Meire et al.	Product R&D	✓	–	–	Coca Cola Channels: website & Facebook	Data mining	A decision support system showing the added value of Facebook data in the customer acquisition process
2017	Perez-Fernandez et al.	Warranty Service	–	–	–	Automotive industry	Delphi Method Analytical Hierarchy Process (ALENOVO)	Developing a DSS to prioritize the improving part of a product and decrease the warranty costs
2018	Hasan et al.	Text sentiment analysis	✓	–	–	Twitter	machine learning sentiment analyzer	Adoption of various sentiment analyzers with machine-learning
2018	Alkahtani et al.	Product defect analysis	–	–	✓	automobile manufacturing	MASON Ontology, Self- Organizing Maps, text mining	Proposing a Decision Support System based on ontology, self-organizing maps, reliability and cost optimization to determine manufacturing faults and map them to design information
2018	Marta Nave et al.	Data mining method	✓	✓	–	4615 different businesses in 51 cities are classified into 22 business categories	CRISP-DM methodology for data mining	Using a text mining and sentimental analysis technique to assist the process of decision-making by analyzing online reviews and using them on a decision support system.
2020	Zheng et al.	Product defect analysis	✓	–	✓	automobile industry	novel probabilistic graphic model text mining using machine learning	Using data mining techniques to propose a novel probabilistic graphic model for defect-discovery from social media data called PDDM
2021	Shokouhyar et al.	Warranty Service	✓	–	✓	Samsung products Twitter	Data preparation Text-mining Machine Learning	Using social media data mining to find out the consequences of social networks on warranty claim forecast
2021	Zheng et al.	Product defect analysis	✓	–	✓	automobile industry	Probabilistic graphic model Text mining FMEA	Using an integrated probabilistic graphic model to discover product flaws among the social media data with the FMEA approach to identify the severity of the failure

4. Methodology

In this section, the methodology used is presented. This framework includes four steps, (1) data sources and data blending, (2) data analysis with deep learning technique, (3) data model and Decision Support System and (4) result evaluation. Each step is detailed and comprised of the following paragraphs:

4.1. Data sources and data blending

The collection of customers’ reviews and comments on devices and products can be done from a variety of sources. Some of them include organizational platforms, such as surveys, email marketing, contact centre content and complaints channels. The review of a product’s reputation and the overall atmosphere shall also be conducted on social media platforms, including Twitter, Facebook, Instagram, Skype, WhatsApp etc. Some channels are also for professionals to exchange highly technical feedback and bugs, such as device/brand forums, Reddit threads, XDA developer sites, GitHub, etc.

In this case, in order to cover all possible probabilities, we’re combining data collection sources. In order to gather data, the advantages of using more than one source are as follows:

• The sources are supporting one another. Different channels of feedback allow for individual freedom. It can increase accuracy, reliability and efficiency to collect data from a variety of sources.

• Results attribute the clustering. Tracking different sources for consumer reviews can help attribute the feedback and cluster the results by different levels of people, items, and technologies.

• The process of determining the effectiveness of different strategies enhances. Given that each channel has its own applications and users, the result would have a better match and integration.

• Merging multiple datasets often demonstrates valuable knowledge that might not be covered unless the data is combined. Useful information provides a new and unique perspective that might lead to better decisions.

• Data blending includes getting data from different sources and creating a single, unique dataset for data visualization and analysis. Data blending qualifies a data analyst to use any type of data from any source in their examination for accurate perspicuity.

We define the following steps for data collection and data blending:

1. The first step in gathering data is to demonstrate the information and data sources that are practical enough to cover the research questions. In this paper, we determined the information related to defects in multiple sources.

2. The next step would be investigating the data sources such as Twitter and Facebook for the general assessment of the product and its defects and exploring through LENOVO Forums, Reddit application, GitHub, and XDA-developers site to analyse the technical information and cluster the findings (Shokouhyar et al. 2021).

3. We then identify the appropriate data set from various sources. There could be a vast range of formats, or several file types can be found. Each data source must have the same dimension to be blended.

4. Then, we merge the data from different sources and customize each connection based on the common dimension to confirm that the data blend is integrated.

That is the procedure we take to examine social media data collection. Since we want to examine the validation of the social media data and the methodology adopted, we analyse the results and match them with the secondary data. In fact, we receive operational data⁷ from the warranty service companies related to the LENOVO brand, discover the product’s defect ontology, and compare the social media data with operational data results. We cooperated with LENOVO company representatives, which delivers aftersales and warranty services for its products, such as laptops. The primary data from the LENOVO representative includes the warranty period’s length, the ratio of the warranty usage and costs, warranty usage categories and strategies, and different defects and failures related to the product.

4.1.1. Data collection from Twitter (Schwitter and Liebe 2020)

Twitter as a social network was launched in 2006, in which users communicate with each other on their personalized pages by sharing limited messages. Although its main focus was the exchange of data and facts, additional features for social interactions have been developed recently. Such as sharing other tweets (retweeting) or replying, giving like mentioning users, sharing pictures, links, and other multimedia data, and using Hashtags (#) before words is a popular tool to tag specific tweets.

Twitter proposes APIs⁸ to allow the public to access their data: the Streaming API and the REST API (Representational State Transfer). Specific conditions have to be observed to access any one of the APIs.

Since Twitter has established the developer account by providing a name, description, and domain, the developer’s Twitter application is registered. This application certifies the user for the application of the Twitter API and authorizes the user the access key and token through the application management tool.

4.1.1.1. Streaming API

utilizing the streaming API, Twitter gives access to (a sample of) all published tweets on Twitter. The specific data will be sent to the user constantly After sending a request to the API. The Streaming API only sends out real-time tweets.

4.1.1.2. REST API

REST (Representational State Transfer) is a general software that represents a set of limitations for developing web services. All requesters will access the data and manage the web service using a predefined set of stateless operations.

4.1.2. Data collection from Facebook

Facebook is an American online social media founded in 2004. Users on Facebook create a profile displaying personal information about themselves. They can communicate with any other users who have agreed to be their ‘friend’ or, with different privacy settings, to the public, by sharing posts containing text, photos, and multimedia.

Facebook has established a new feature and analytics API for researchers called ‘Facebook Open Research & Transparency’ (FORT). To find the trends and discover the derivatives on Facebook pages, academics and researchers can use these collections of APIs.

The sample API code in FORT:

4.1.3. Data collection from other sources

Reddit is an American social news website for ranking web content and creating discussions between different people. Members can post content such as links, text posts, and multimedia to the site that can be rated by other users. Prioritizing the posts is by topic and are in ‘communities’ or ‘subreddits’. Reddit has an API that authorizes you to access a lot of information on Reddit with JSON formatting. Most pages on Reddit can also be accessed through json by simply adding ‘.json’ to the end of a URL.

The LENOVO forum community platform helps developers and users get interested in developing or using apps and services for LENOVO platforms, services, and devices. Anyone can join the discussion, regardless of experience level or where they exist in the ecosystem of LENOVO company. LENOVO forum is one of the data collection sources used in this research and can be analyzed for the target keywords by using a Crawler. Crawler is a program developed to crawl through web pages and get indexed data. The indexed data (usually in .json format) is the raw data prepared to get clean and structured.

XDA-Developers stands for a social network consisting of millions of developers from any kind of Android and Windows operating system that use this platform and forums to share their issues, abate about the OS⁹ versions, and communicate about their experiences and different devices. To determine the software failures more accurately, the XDA-Developers is also one of the data gathering sources that can be examined using the Crawlers developed to get the indexed data and target keywords. The raw data of the XDA-Developers is ready to use after the data preparation process.

GitHub is a platform with more than 73 million users that provides Internet hosting for software development and proposes distributed version control and source code management. This platform is used by millions of developers, engineers, and companies to share their issues and debate about the OS versions, device software, and hardware and share their experiences about different devices such as laptops. Crawlers can explore this platform to get the indexed data and target keywords. The raw data from GitHub will be blended with other sources after the data preparation process. After collecting data from multiple data sources, we should step forward to data cleaning and preparation.

4.2. Data analysis with deep learning technique

Data preparation is a fundamental stage of data analysis. Data preparation uses the techniques of qualifying data to be ready for analysis. The whole process is presented in Figure 2. There are steps to prepare data for examination:

Figure 2. Data analysis with deep learning technique.

4.2.1. Data cleaning

Data cleaning means identification and modification of errors and removing unwanted words. Since the data sources are vast and complicated, some operations need to be done to cleanse and purify the data, such as removing the spam words, symbols, duplicates, and irrelevant values and converting the word types into a typical kind.

4.2.2. Data structuring and integration

This means modelling and organizing data from different data channels into a unified format that will match the requirements. It is necessary to sort them into defined data structures and patterns to continue more operations on them. Sometimes a change in the data scale or distribution function is needed (Shokouhyar, Ahmadi and Ashrafzadeh 2021). After that, the conversion of words to lowercase and applying the lemmatization were done in the Gensim library to get the root words.

4.2.3. Ontology development

Social media sentiment analysis based on domain ontology has been developed to determine the defects among user-generated comments effectively. The Web Ontology Language (OWL) is manipulated to describe ontology due to its complete framework for illustrating ontologies. OWL advocates the following features Classes, Taxonomic relations, Datatype properties, Objects properties, Individuals and Restrictions. According to Bechhofer-2009, classes show the concepts of the domain and relationships are determined in Protégé (Alkahtani et al. 2018). The ontology has been designed using the warranty service instructions and guidelines. The output is used as one of the inputs of the next step for the creation of feature vectors. Sample data is shown in the following Figures 3 and 4.

Figure 3. Sample ontology of laptop’s components and features in Protégé.

Figure 4. Sample ontology of processor’s components and features in Protégé.

4.2.4. Creation of feature vectors and word-embedding

In the current step, the outputs of steps two and three will be used to empower the learning method to analyze the social media data. In this regard, we manipulated the TF-IDF¹⁰ features with word embeddings obtained from the Word2Vec¹¹ model. These features are combined because the TF-IDF operates on the bag-of-words model and extracts the most defective components by capturing the semantic relationship based on the product ontology and failures terminology (Afaq and Manocha 2023). Besides, the Word2Vec model sustains the syntactic and semantic relations among words. Consequently, the word vectors are created by the Word2Vec model in the Gensim library. Then, the Scikit library sets each document’s TF-IDF information. The online latent semantic indexing model uses 2 LSTM¹² networks for each incoming data stream, compares the document and incrementally updates the term-defect document matrix and defect-document matrix (Pathak et al. 2021). Reminding the paper’s purpose that seeks the defects among the social media threads, the sentimental filter is used to determine and highlight the negative and neutral sentences. Considering the regular components and symptoms of the device, the filter also classifies the data into subjectively related threads (Zheng et al. 2020; Li et al. 2022). That means the ones containing a defect in the device components or showing defect symptoms are determined and can be derived from a post and its replies and comments (Zheng et al. 2020).

Assuming $N\left(\alpha ,x\right)$ is the number of terms $\alpha$ occur in document x, which is the sentence obtaining streaming data from multi-channel sources of social networks, |X| is the length of document x, |X| is the total number of documents in the entire collection of the social media data streaming and $\left|x\in X:\alpha \in x\right|$ Denotes the total number of documents in which term $\alpha$ occurs. As a result, the TF-IDF is calculated according to Equation (1)(1) $$\mathit{\text{TF}}-\mathit{\text{IDF}}=\frac{N\left(\alpha ,x\right)}{\left|x\right|}\times \mathrm{\text{log}}\hspace{0.17em}\frac{\left|X\right|}{\left|x\in X:\alpha \in x\right|}$$(1) : (1) $$\mathit{\text{TF}}-\mathit{\text{IDF}}=\frac{N\left(\alpha ,x\right)}{\left|x\right|}\times \mathrm{\text{log}}\hspace{0.17em}\frac{\left|X\right|}{\left|x\in X:\alpha \in x\right|}$$(1)

The word vectors are created by the Word2Vec model in the Gensim library, and TF-IDF is applied to word vectors for the documents $X=\left\{x_1,x_2,\dots ,x_n\right\},$ which denotes a vector x with M dimensions. in which $m_{\mathit{\text{th}}}$ the item represents the weight of the $m_{\mathit{\text{th}}}$ term calculated by the TF-IDF score using Equation (1)(1) $$\mathit{\text{TF}}-\mathit{\text{IDF}}=\frac{N\left(\alpha ,x\right)}{\left|x\right|}\times \mathrm{\text{log}}\hspace{0.17em}\frac{\left|X\right|}{\left|x\in X:\alpha \in x\right|}$$(1) , and the actual term is shown by its word vector obtained from Word2Vec. The LSTM Networks 1 and 2 have been used in the research of Pathak et al. (2021), with the learning mechanism shown in Figure 5. The previous iteration extracts the term-defect matrix $X_{t-1}$ and finds latent space representation $y_t.$ LSTM network 2 obtains $y_t$ and inputs document representation $x_t$ and discovers the term-defect matrix $X_t.$ Therefore, the proposed model works online by altering $X_i$ and $y_i$ while maintaining only one LSTM model active at time t. The LSTM networks are shown in Figure 6.

Figure 5. LSTM 1 learning mechanism.

Figure 6. LSTM networks.

4.2.5. Deep learning analysis validation metrics

According to the literature review (Zheng et al. 2020), the performance evaluation metrics for deep learning and sentiment analysis methods are Precision, Recall and F-Measure. We have measured the mentioned metrics to evaluate the methodology we have used. The following metrics are calculated according to Equations (7)–(9). (7) $$\mathrm{\text{Precision}}=\frac{\mathit{\text{TP}}}{\mathit{\text{TP}}+\mathit{\text{FP}}}$$(7) (8) $$\mathrm{\text{Recall}}=\frac{\mathit{\text{TP}}}{\mathit{\text{TP}}+\mathit{\text{FN}}}$$(8) (9) $$\mathrm{F}‐\mathrm{\text{Measure}}=\frac{2\times \mathrm{\text{Precision}}\times \mathrm{\text{Recall}}}{\mathrm{\text{Precision}}+\mathrm{\text{Recall}}}$$(9)

*TP = True Positive; FP = False Positive; FN = False Negative.

4.3. The DSS model with the ontology approach

4.3.1. Data model

After data analysis is completed, the threads, which are the inputs of the decision support system (DSS), have the structure described in Table 2. The structure used is augmented in Rockwell et al. (2009). Each property object belongs to a data table with the following format where the $i,j,p,q\in N.$ Samples of tables are shown in Table 3.

Table 2. Information captured for threads.

Property	Description	Data table format
Issue	Text description of the defect/ failure
Symptom	Whether the defect has a symptom and its description	Spi = {Sp1, Sp2, …, Spn}
Has a cause	Whether the thread specifies a cause for the failure	Caj = {Ca1, Ca2, …, Can}
Has a component	Whether the defect affects a component(s)	Cop = {Co1, Co2, …, Con}
Has a solution^a	Whether the thread contains a solution or action of repair/replace	Slq = {Sl1, Sl2, …, Sln}

^aEach object of the solutions table carries a tag of repair/ replace.

Table 3. Table samples.

The interactions between the objects are shown in Figure 7.

Figure 7. Objects interactions.

4.3.2. Data association

For data association, we use the joint table of different data obtained from the deep learning method, which relates the objects with a specific logic as follows:

• Symptoms and related components, causes, and solutions are in one thread containing replies or comments.

• The above symptoms and issues are repeated in other threads, and the related components, causes, and solutions are mentioned.

Table

Joint table of Symptoms and related Causes:	JT(Spi, Caj,…,n)
Joint table of Causes and related Components:	JT(Caj, Cop,…,n)
Joint table of Symptoms and related Components:	JT(Spi, Cop,…,n)
Joint table of Components and related Solutions:	JT(Cop, Slq,…,n)

The generated DSS model incorporates many rules such as the below (Rajpathak et al. 2012), to call the symptoms, causes, components, and solutions:

If the issue happened, use the related ‘symptom_i’, find the related ‘cause_j’ from the {Joint table of Symptoms and related Causes}, then find the corresponding ‘component_p’ from the {Joint table of Causes and related Components}, then consider the ‘solution_q’ from the {Joint table of Components and related Solutions}.

Using the data obtained from social networks is not enough to obtain the critical results and outputs from this research. Therefore, we provide the claimed data in after-sales service and warranty agencies and the related manuals to create a database of symptoms, causes, components, and solutions based on the claimed data. Consequently, we received relevant information to compare the results gained from social media and agencies to have the highest degree of compliance with the actual marks and deliver more efficient consequences. The first output of the whole process is to determine the components among the threads which had the most numbers of defects from the social media point of view:

Failures found in social media are shown with F_s, and F_sCo_j represents the components related to the failures. Due to the 80–20 rule, known as the ‘Pareto Principle’, that asserts 80% of outputs result from 20% of all causes, to find the most defective components from the social media point of view, we search through the data until the number of failures related to the most defective components reaches the 80% of total failures.

The same operation has been done for the operational data from the claimed data in after-sales service and warranty agencies. Failures found in claimed data are shown with F_o, and F_oCo_m represents the components related to the failures.

The whole model is shown in Figure 8.

Figure 8. Distinguishing the components.

We divided the results into two groups:

1. Failures common with the agencies’ results.

2. Failures not common with the agencies’ results.

4.3.2.1. Failures common with the agencies’ results

For the first group, the most defective and known components are determined. These items are common between the representatives and the social media reports for making failures. Co_t represents the components related to the failures and is common with the agencies’ results.

To represent the common failures (F_c (Co_t)), we use the following: (1) $$F_{\mathrm{c}}\left({\mathit{\text{Co}}}_{\mathrm{t}}\right)=F_{\mathrm{s}}\cap F_{\mathrm{o}}\left({\mathit{\text{Co}}}_{\mathrm{t}}\right)=\left\{{\mathit{\text{Co}}}_{\mathrm{t}}\right|{\mathit{\text{Co}}}_{\mathrm{t}}\in F_{\mathrm{s}},F_{\mathrm{o}}\}$$(1)

The (1) shows the critical components that should be focused on.

To represent the symptoms of the common failure (F_c (Sp_t)), we use the following: (2) $$F_{\mathrm{c}}\left({\mathit{Sp}}_{\mathrm{t}}\right)=\left\{{\mathit{Sp}}_{\mathrm{t}}\right|{\mathit{\text{Co}}}_{\mathrm{t}},{\mathit{Sp}}_{\mathrm{t}}\in \mathit{\text{JT}}({\mathit{Sp}}_{\mathrm{i}},{\mathrm{Co}}_{\mathrm{p}}\text{,}\mathrm{\dots }\text{,n})\}$$(2)

4.3.2.2. Failures not common with the agencies’ results

The second group needs a more technical examination and deeper analysis. These items are the ones that agencies have encountered less than the people using the products. If these kinds of failures won’t be reviewed, they can cause severe dissatisfaction and warranty costs rise. So, all the Symptoms, Causes, and solutions related to these defective components should be studied and reviewed. Co_r represents the components related to the failures and are not common with the agencies’ results.

To represent these failures (F_nc (Co_r)), we use the following: (3) $$F_{\text{nc}}({\mathit{\text{Co}}}_{\mathrm{r}})=F_{\mathrm{s}}-F_{\mathrm{o}}\mathrm{ }({\mathit{\text{Co}}}_{\mathrm{r}})=\left\{{\mathit{\text{Co}}}_{\mathrm{r}}\right|{\mathrm{Co}}_{\mathrm{t}}\in F_{\mathrm{s}},{\mathrm{Co}}_t\notin F_{\mathrm{o}}\}$$(3)

To represent the symptoms of the failure (F_nc (Sp_r)), we use the following: (4) $$F_{\text{nc}}\left({\mathit{\text{Sp}}}_{\mathrm{r}}\right)=\left\{{\mathit{\text{Sp}}}_{\mathrm{r}}\right|{\mathit{\text{Co}}}_{\mathrm{r}},\mathit{\text{Sp}}\mathrm{r}\in \mathit{\text{JT}}({\mathit{\text{Sp}}}_{\mathrm{i}},{\mathit{\text{Co}}}_p\text{,}\mathrm{\dots }\text{,n})\}$$(4)

4.3.3. Optimizing the warranty service solution

Another way of optimizing the whole system is to make the best decision between replacing a failed tool and repairing it. Many factors affect the choice, such as age and the condition of the defect detected, The mean time between failures (MTBF), the item’s mean time to first failure (MTTF), the remaining time of the warranty period, and costs of the repair/replacement choices. For example, (Rao 2011) has adopted an algorithmic approach and a DSS for making the best choice between repairing or replacing a failed item. The next step after investigating the most defective items and the most probable failures is to figure out the best conditions for the warranty period to optimize the cost. Jack and Murthy (2001) has a simple function for optimizing the warranty cost by making the best choice between repairing and replacing an item. To integrate the whole model using the theorem of Jack and Murthy (2001), let’s consider the cost of repairing a failed item as $C_{\mathit{\text{RP}}}={\int }_0^w{r\left(x\right)}^{*}\mathit{\text{dx}}$ and the cost of replacing it as C_RC. (*r(x) indicates the hazard rate). The duration of the warranty period is in $(0,W)$ intervals and K and L in $(0,W)$ are the optimal parameters that are the best time to satisfy the $(K,L)$ strategy in Jack and Murthy (2001) by minimizing the total cost as $J(K,L).$ In fact, in the Jack and Murthy (2001) the MTBF and MTTF of the product are the most effective factors for determining the L and K, which are the intervals at which the product should be repaired or replaced based on the time of failure. According to the theorem, during $(0,K)$ and $\left(L,W\right)$ the product should be repaired. While, during the $(K,L)$ the product should be replaced. In this research, we involve the components of the product. So, the warranty service provider could obtain two different approaches. First, the warranty service provider can define different warranty periods (W) for each component of the product and consequently, the optimum L and K would be different for each component due to different MTBF and MTTF. Second, it can define the warranty period (W) and the optimum L and K for the product based on the most defective components due to their MTBF and MTTF.

So, the objective is: (5) $$\text{Minimum}J\left(K,L\right)=C_{\mathit{\text{RP}}}(K,L)+C_{\mathit{\text{RC}}}(K,L)$$(5) (6) $$\text{Subject to}0\le K\le L\le W$$(6)

5. Calculations and deployment

The data preparation and calculation have been implemented for the LENOVO Laptops to ensure the DSS model validation through the social media channels mentioned in the previous sections. Operational data about this device was also captured from the LENOVO representatives to prove the model’s efficiency. The calculations are sorted into six parts. We have presented the operational and social media data statistics in the following. Then, we analysed the association between both sources of data. Afterwards, the research method was validated through a comparison of different methodologies and the calculation of related metrics. Finally, the warranty cost for the operational data is presented to validate the optimisation method.

5.1. Statistics of operational data

In this section, we provide the data collected from the agencies about the device’s custom failures, the related symptoms and components, and the corresponding solutions they maintained to resolve the device issues (Shokouhyar et al. 2021). We have studied the data between 2021 Feb to 2022 Jan. We have examined 431 issues that had a warranty claim and had been submitted into the agency’s database according to Table 4. The total components being affected were 28 that have been repaired or replaced by the warranty service provider. According to the operational results, the number of replacements is about 28% of the total issues, and the number of replacements of the hard disks was the most.

Table 4. Total statistics of operational data for LENOVO laptops.

	Issues	symptom	component	Cause	Replace	Repair
Total Numbers	431	58	28	45	120	311
Repetitive issues	141	128	8	23	52	89
Serious Problems^a	172	37	12	26	63	109

^aSerious problems are the ones the laptop doesn’t work properly.

The operation described in the previous section has been done for the operational data from the claimed data in the after-sales service. Failures and all the defective components found in the claimed data are shown in Table 5. Due to the 80–20 rule, known as the ‘Pareto Principle’, to find the most defective components, we search through the data until the number of failures related to the most defective components reaches 80% of total failures. Examinations stated that among 28 components that seemed to be the vulnerable components, only 3 of the components were the most defective ones. According to Table 5, the Hard disk was the most vulnerable component. It has created most of the expenses. More than 70% of the replacements belonged to these most defective components. Figure 9 shows the most defective items of the laptop according to the warranty claimed data.

Figure 9. Graph of claimed data based on the Pareto Principle.

Table 5. Details of claimed data.

	Component	Number of defects	Cumulative percentage	Related symptom	Related cause	Replace	Repair	Replace Cost	Repair Cost
FoCo1	Hard disk	142	33%	6	3	31	111	6200	6660
FoCo2	Battery	123	61%	5	2	30	93	4500	4185
FoCo3	Keyboard	80	80%	3	4	27	53	2160	1855
FoCo4	Charger	15	84%	2	1	8	7	640	70
FoCo5	Display Screen	12	86%	2	2	7	5	700	450
FoCo6	Top Panel	8	88%	2	2	0	8	0	160
FoCo7	Base Panel	6	90%	1	1	0	6	0	180
FoCo8	Back Camera	5	91%	2	1	0	5	0	125
FoCo9	Motherboard	5	92%	1	1	2	3	600	900
FoCo10	Cooling Fan	4	93%	1	1	4	0	800	0
FoCo11	Touchpad	3	94%	2	2	0	3	0	330
FoCo12	USB Port	3	94%	2	1	0	3	0	30
FoCo13	SSD	3	95%	3	1	3	0	180	0
FoCo14	Video Card	3	96%	2	1	0	3	0	300
FoCo15	Speaker	3	96%	1	3	2	1	40	10
FoCo16	Mic Port	3	97%	1	1	0	3	0	30
FoCo17	CPU	2	97%	2	1	2	0	1600	0
FoCo18	RAM	1	98%	2	2	1	0	35	0
FoCo19	ROM	1	98%	1	1	1	0	40	0
FoCo20	Wireless Card	1	98%	2	1	1	0	18	0
FoCo21	Optical Drive	1	98%	3	2	1	0	19	0
FoCo22	Hinges	1	99%	2	2	0	1	0	140
FoCo23	Palm Rest Assembly	1	99%	1	1	0	1	0	15
FoCo24	VGA port	1	99%	2	2	0	1	0	80
FoCo25	Ethernet Cable port	1	99%	2	2	0	1	0	50
FoCo26	Headphone Jack Port	1	100%	1	2	0	1	0	35
FoCo27	Front Camera	1	100%	2	1	0	1	0	60
FoCo28	Antenna	1	100%	2	1	0	1	0	110
$\sum F_o{\mathit{\text{Co}}}_{\text{m}}$	n = 28	431	-	58	45	120	311	17,532	15,775

Besides, the agencies’ manuals and procedures also have been studied to examine the recommendations for the failures. These manuals have mentioned the probable defects and failures for different product components and described the failure symptoms, the influenced components, and the corresponding causes. They have been used to choose the #data, ontology and related text categories correctly. There are some related procedures to clarify the agencies’ and customers’ responsibilities and the scope of the warranty services that can assist the companies in clearing the rules and optimising the additive expenses.

5.2. Statistics for social media report

In this section, we provide the data collected from different social media channels to figure out the device’s custom failures, the related symptoms and components (Ko et al. 2018), and the corresponding causes mentioned in the posts. We have studied the data between 2021 Feb to 2022 Jan. In order to classify the total data captured from the different sources, the number of issues and related details are presented in Table 6. We have examined 5483 threads from different social media channels that had an issue and had made the customer show negative sentiments or contain any of the #texts shown in Table 3. The detailed statistics of the mentioned experiment have been presented in Table 6. Most of the threads belonged to Twitter and Facebook. The scattering data used from different social channels is shown in Figure 10.

Figure 10. Threads scattering in social media.

Table 6. Statistics for social media issues.

	Twitter	Facebook	Reddit	LENOVO forums	GitHub	XDA.developers
Numbers of threads	2145	1872	473	534	321	138
Numbers of defect-related threads	1532	1231	432	376	256	98
Numbers of words in threads	33,704	36,930	14,688	13,536	13,312	1666
Numbers of issues	884	745	174	234	341	84
Numbers of symptoms	1203	897	562	420	287	127
Numbers of components	453	280	122	111	154	64

The operation described in Section 4.3.2 has been done for the social media data and Failures, and the most defective components found in different sources are shown in Table 7. Due to the 80–20 rule, known as the ‘Pareto Principle’, we search through the data until the number of failures related to the most defective components reaches 80% of total failures. Examinations stated that only 5 of the components are the most defective among the different channel results. According to Table 7, the Hard disk was the most vulnerable component in the four channels of social networks. Figure 11 shows the most defective items of the laptop according to social media data.

Figure 11. Sample detail data for social media issues.

Table 7. The most defective items according to social media.

	Component	Twitter	Facebook	Reddit	LENOVO forums	GitHub	XDA.developers
FsCo1	Hard disk	640	468	143	120	96	41
FsCo2	Battery	421	459	183	94	62	–
FsCo3	Keyboard	–	54	19	86	47	18
FsCo4	Ports	165	–	–	–	–	–
FsCo5	OS	–	–	–	–	–	19
$\sum F\text{sCoj}$	–	1226	981	345	300	205	78
Numbers of total defect-related threads		1532	1231	432	376	256	98
Percentage		80%	80%	80%	80%	80%	80%

Beside the most defective components, we have also discovered the related symptoms of each component. The most frequent symptoms for the most defective items are shown in Tables 8–12. Moreover, the graphs for all frequent symptoms are presented in Figures 12–16. Remarkably, the Pareto Principle has also detected the frequent symptoms for each defective component. In fact, symptoms common in 80% of the failures of the component have been discovered and shown in the following tables and graphs.

Figure 12. Graph for the most frequent symptoms of Hard disk failure.

Figure 13. Graph for the most frequent symptoms of Battery failure.

Figure 14. Graph for the most frequent symptoms of Keyboard failure.

Figure 15. Graph for the most frequent symptoms of Ports failure.

Figure 16. Graph for the most frequent symptoms of OS failure.

Table 8. The most frequent symptoms of Hard disk.

Symptom	Twitter	Facebook	Reddit	LENOVO forums	GitHub	XDA.developers
File Corruption	153	121	87	65	53	21
Drive Not Detected	122	84	54	33	14	15
Strange Loud Noises	75	92	44	36	27	9
Sum	350	297	185	134	94	45
Total	439	371	232	167	118	56

Table 9. The most frequent symptoms of Battery.

Symptom	Twitter	Facebook	Reddit	LENOVO forums	GitHub	XDA.developers
Too Hot Charging	133	56	54	51	17	–
High Battery Usage	116	132	55	43	41	–
Too Long Charging	93	100	67	28	30	–
Sum	342	288	176	122	88	–
Total	423	359	221	153	110	–

Table 10. The most frequent symptoms of Keyboard.

Symptom	Twitter	Facebook	Reddit	LENOVO forums	GitHub	XDA.developers
Screen Error	–	29	23	16	16	11
Keys Not Working	–	70	25	25	17	9
Some Keys Not Working	–	33	40	39	15	12
Sum	–	132	88	80	48	32
Total	–	167	109	100	59	40

Table 11. The most frequent symptoms of Ports.

Symptom	Twitter	Facebook	Reddit	LENOVO forums	GitHub	XDA.developers
Compatible Error Code	143	–	–	–	–	–
Device Not Detected	82	–	–	–	–	–
Device Disconnecting During Usage	45	–	–	–	–	–
Sum	270	–	–	–	–	–
Total	341	–	–	–	–	–

Table 12. The most frequent symptoms of OS.

Symptom	Twitter	Facebook	Reddit	LENOVO forums	GitHub	XDA.developers
Blue Screen	–	–	–	–	–	17
System Error	–	–	–	–	–	5
System Freezing	–	–	–	–	–	3
Sum	–	–	–	–	–	25
Total	–	–	–	–	–	31

5.3. Social media and operational data association

The most defective components of LENOVO laptops are presented in Table 13. These items are the most defective components common between the agent’s results and the social media reports. Other defective components found in social media mining are the same as those found in operational data. The only difference between the two groups of items is OS. Obviously, problems with OS are not related to the laptop warranty service provider’s responsibilities. According to the tables above and below, all of the defects and components found in the operational data sources were the exact defects and components found in social media. However, the number of defects and the most vulnerable components found in social media was more than the operational data, as we see according to Tables 5 and 7. That means social media can provide more information than those that exist in agencies. As presented in Figures 8 and10, the most defective components of social media and the operational data examination are the same, which can prove the accuracy of the data mining through social networks. Figure 17 shows the frequency of the defects in social media compared to operational data.

Figure 17. Sample of common failures in operation vs social media.

Table 13. The most defective components common in operations and social media.

	Cot	%Defects in operation	%Defects in social media
FsCo1	Hard disk	33%	38%
FsCo2	Battery	29%	31%
FsCo3	Keyboard	19%	6%
∑FcCot	3	80%	76%

According to operational data and social media mining, some failure symptoms are more common and frequent among users. The graph of the symptoms’ frequency is presented in Figure 18. Consequently, this data is used to discover the failures’ main causes. This part has been investigated through the operational data, which are more accurate. In fact, the frequent symptoms have been matched with the main reasons for the related failure according to the agencies’ claimed data. The main reasons for the failures are shown in Table 14. Besides, the frequency of the main reasons has also been described and illustrated in Table 14 and Figure 19.

Figure 18. The frequency of symptoms.

Figure 19. Graph for most frequent symptoms’ reasons.

Table 14. Main causes for the most frequent symptoms according to operational data.

File Corruption		Drive Not Detected		High Battery Usage		Too Long Charging
Main causes	percentage	Main causes	percentage	Main causes	percentage	Main causes	percentage
Hard disk malfunction	48%	Hard driver issue	38%	Decreasing battery capacity	43%	Old Battery	31%
Virus	32%	Connection issue	37%	Old battery	33%	Damaged Battery	48%
Software Error	20%	Faulty BIOS settings	25%	Power malfunction	24%	Damaged Cable	21%

As shown, one of the most frequent failures belongs to the Battery which is a concern in today’s energy management systems. As mentioned in Shi et al. (2023), the outputs of the failures especially the battery drains could be used as an input for optimizing the battery management systems development. By fusing the power of physical process models with the adaptability of machine learning approaches, this will significantly increase the predictive and modelling capability for long-range connections across multiple timelines.

5.4. Comparison of methods in sentiment analysis

After data preparation and extracting informative threads, we tagged 2000 threads with their real sentiments. Then, we used 10-fold cross-validation to conduct on different classifiers, which can analyze the sentiment of the threads in order to evaluate their performance and compare them with each other same as the comparison with the model proposed in this research. The classifiers used are presented in Table 15, and the F-measure metric is calculated for all of them (Zheng et al. 2020).

Table 15. F-measures for different classifiers.

Classifier	F-measure
Random Forest	0.8965
Decision Tree	0.8733
Gradient Boosting Decision Tree	0.9322
Logistic Regression	0.8954
K-Nearest Neighbor	0.9121

According to Table 15, all the classifiers are below 0.95. However, the Gradient Boosting Decision Tree and K-Nearest Neighbor have the highest F-measures. Since F-measure allows a model to be evaluated by taking both the precision and recall into account using a single score, we have presented the mentioned metric to compare the different classifiers and also make a basement for our methodology to learn more accurately (Zheng et al. 2020).

5.5. Deep learning analysis validation

According to Section 4.2.5 and the previous section, we have collected the data for the research’s method’s validation. In this case, we tagged 2000 threads with their real sentiments. Then, we used 10-fold cross-validation to conduct our research model in order to evaluate its performance and accuracy. The result of the validation is presented in Table 16 (Zheng et al. 2020).

Table 16. Methodology validation metrics.

	Precision	Recall	F-Measure
Evaluation	0.9779	0.9773	0.9763

The comparison of the F-measure metric for the current research and other sentiment analyzers and classifiers is presented in Figure 20. As shown in the figure below, the research method has the top accuracy according to the F-measure metric (Zheng et al. 2020).

Figure 20. F-measure comparison between different methods of sentiment analysis.

5.6. The warranty cost function based on optimum L and K

As explained in Section 4.3.3, the warranty cost function is optimized according to Jack and Murthy (2001). We have collected the data for the cost calculations, and all of the factors are shown in Table 17 and Figure 21. According to the following table, we have decreased the cost of the warranty service due to an optimized strategy for warranty service based on the repair/replacement solution. In this table, we defined the optimum K and L according to each component’s replacement and repair cost based on a three-year warranty period. Then, we reviewed the operational data and considered those of the replacements which could have been repaired instead and vice versa. Finally, the cost was recalculated to measure the decrease in warranty service costs.

Figure 21. Graph for decreasing the warranty service cost.

Table 17. Warranty service cost for operational data.

Component	Crp	Crc	p	K	L	Total Cost 1	Total Cost 2	Decrease %
Hard disk	60	200	3.3	1.21	2.62	12,860	10,060	−22%
Battery	45	150	3.3	1.21	2.62	8685	8055	−7%
Keyboard	35	80	2.2	1.21	2.82	4015	3340	−17%

Total Cost 1: The real cost that the warranty service provider has spent for the warranty service.

Total Cost 2: The cost that the warranty service provider may spend for the warranty service according to the (Jack and Murthy 2001) strategy.

6. Discussion and research limitations

This paper promotes the advantages of social media data contribution to warranty service enhancement. Many popular social networks have been mined, blended, and prepared to generate valuable and informative threads. The proposed approach enables the model to analyse data at the sentence level to extract the sentiment of the threads and frequent failures of the product using product components and defects’ ontology. The method uses online latent semantic indexing with regularisation constraints in long short-term memory networks. The informative threads deriving from social media networks are processed serially, and the deep learning-based model incrementally develops the model according to online streaming data. The model categorises the whole data into different properties like issues, symptoms, and components. The research methodology is tested by Precision, Recall and F-Measure metrics and is compared to other sentiment classifiers to evaluate the model’s performance and accuracy. The results show that the highest F-measure amount (0.9763) belongs to the proposed research model. The next high F-measure amount was recorded for the Gradient Boosting Decision Tree, while the lowest F-measure amount was documented for the Decision Tree.

Finally, the threads have been analysed, and the frequent failures have been identified. The results indicate that according to the 80–20 rule, known as the ‘Pareto Principle’, 80% of the failures occurred to 20% of the components of the product. This conclusion is proved by both operational and social media data with the same results. So, to answer the research questions, it is identified that the obtained data from social media is sufficient to determine the system’s current state and leads to valid information for enhancing the products and services. In fact, the study demonstrates a correlation between the social media data and the operational data that support each other. According to Tables 5 and 7 and Figures 8 and 10, the most frequent defects found in a warranty service provider have been covered 100% by social media data totally.

Firstly, using social media data widens the scope of data gathering and lets in all data from different sources. So, the warranty database is not restricted to the claimed data. Secondly, it reveals the most repeated failed items and their frequency with online streaming data that help find the most vulnerable components, the main causes of the defects, and the frequent symptoms. Third, it proposes a model that facilitates warranty strategies by considering the details of product defects to systematise the warranty plan and costs. Table 17 and Figure 21 show that finding the most vulnerable components leads to adopting the correct strategy and can decrease the total cost of warranty services. As shown in Table 17 and Figure 21, finding the vulnerable components, their failure cause, and the details of warranty solutions (warranty period, MTBF, and MTTF) using Equations (5)(5) $$\text{Minimum}J\left(K,L\right)=C_{\mathit{\text{RP}}}(K,L)+C_{\mathit{\text{RC}}}(K,L)$$(5) and (6)(6) $$\text{Subject to}0\le K\le L\le W$$(6) , we have defined the optimum L, k according to the (Jack and Murthy 2001). Then, we rechecked the operational data to find the improving points according to the component’s condition in claim time. We have found that some component replacements were technically unnecessary due to Jack and Murthy (2001) strategy. So, we have modified the replacements with repair costs for the most defective components to see the result. The costs have decreased for all the most defective components. The Hard disk, as the most defective component of the LENOVO laptops, could have a 22% lower cost if the above strategy is used.

The presented method in this research has some constraints and limitations. The gathered social media data is not clean, containing many issues such as spelling or errors in referring, mistyping and inaccurate data associated with a defect claim. Besides, it can contain duplicated data on the company’s different official pages and sources have been used. Thus, the companies should carefully cleanse the data using the cleaning and preparation process that contains structuring, integrating, and associating according to the 4.2 section. Also, this study was limited to English language posts, which causes issues with its generalizability and should be developed with multilingual mining methods to increase the model’s accuracy. Another challenge facing is the considerable volume of data accumulated from multiple sources causes the data preparation process, complicated and extended. This problem is solved using advanced methods and developed tools such as machine learning techniques and data mining software. Despite the constraints and the complicated data mining process, the results are aligned with the research hypothesis. The data shown in Table 13 supports the above statement, as the common failures in social media cover 100% of the repetitive defects recorded in internal databases. Therefore, monitoring the online data improves the company’s ability to detect product issues before the customer claims warranty services.

7. Conclusion

It is believed that not only focusing on the customer’s needs and expectations before manufacturing products is necessary, but also additional attention is needed after a product and service are delivered. While the warranty service centres are the frontline to encounter customer problems, the article’s motivation was the concern of them to reach the following objectives that are among the most important concerns of top managers:

• Analysing the frequent product failures to optimize the warranty services during the warranty period: Investigating the likely defects and the most frequent failures that may be a potential warranty claim will widen the horizons of a manager prioritising the company’s development opportunities and cost management issues.

• Keeping the customers satisfied and fulfilled by detecting the root causes of defects and rectifying them: Being informed about the most frequent dissatisfactions and inconveniences in their products and services will support them to be prepared enough to confront the failures quickly and will equip them to address the customers’ problems as soon as possible.

• Being the active source and channel to transfer customer feedback: Along with the customer service department and call centre feedback, the new source of data will support the managers to enhance their input data and check the accuracy and validity of internal data. This will help managers and companies to enrich their infrastructure and performance to bring their operational results closer to the results perceived by customers.

• Determining the frequent failures and related symptoms to improve the production process and after-sale services: Since the methodology introduced in this paper is not only restricted to warranty issues and any issue that is related to a product or service can be investigated through social media channels and is able to be learned by an artificial intelligence system, the system is ready to be used to enhance different departments and infrastructures and be the basis of improvement opportunities in a company.

Hence, this paper promotes a decision support system (DSS) to cover the above concern using the streaming data of online social channels in a data preparation process based on ontology and deep learning methods. Besides, the claimed data has been collected from the LENOVO Agencies. Then the most defective components and frequent failures were detected from both data sources to optimise the warranty solutions. The result stated that both data sources are matched with each other, and the conclusions support the main objectives of the research. So, companies can use social networks to enhance their warranty services as well as manufacturing operations. As mentioned in the previous researches, it is considered that companies use social media mainly to send and receive data and information for different purposes such as marketing, promotions, etc. In that respect, social media, a helpful tool for corporations, carries more benefits than websites since it is up to date most of the time due to daily usage. In this regard, the research has shown how companies and managers can benefit from social media data, incorporating data preparation and ontology-based deep learning from multi-source data channels like Twitter, Facebook, Reddit, etc. A case study in the laptop industry is used to validate the efficacy and reliability of the research approach. The method used is not only beneficial for any IT devices such as laptops and mobile phones but also can be used for any product such as automobiles or home furniture and electrical tools. In fact, any device which is negotiated by users on social media channels is able to be investigated for frequent defects and failures. Regardless of using social media data, other machine learning methods, such as Naive Bayes are also applicable to the current methodology, and since Most Artificial Intelligence work involves either Machine Learning or Deep Learning, and the ‘intelligent’ behavior of machines requires knowledge, the whole system can be used for a failure prediction system including failure patterns and frequencies used in warranty and guarantee periods. Moreover, the model is also usable as an input for Low-fidelity and High-fidelity models which are considered in Müller (2019) and provides customer preferences and priorities to the mentioned models and enriches the algorithms that approximate the simulations at both fidelity levels. Future research will incorporate the following components:

• A more complex cost optimization function to cover all the elements affecting the cost management issue.

• Maintaining the customer behaviour to the whole approach and solving the issues that can occur in the customer-agent relationship.

• Application of ‘game theory’ to the research for promoting the product-service system using the product, use, or result-oriented system.

Abbreviations
Pss	=	Product-Service System
DSS	=	Decision Support System
API	=	Application Programming Interface
ML	=	Machine Learning
DL	=	Deep Learning
LSTM	=	Long Short-Term Memory Networks
OWL	=	Web Ontology Language
SME	=	Small and Medium-sized Enterprises
FORT	=	Facebook Open Research and Transparency
OS	=	Operating System
TF-IDF	=	Term Frequency-Inverse Document Frequency
MBTF	=	Mean time Between Failures
MTTF	=	Mean time To First Failure

Disclosure statement

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability statement

Data sharing is not applicable to this article as no new data were created or analysed in this study.

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Notes

1 Any product containing consumer, industrial, service products can be included in this paper’s scope.

2 The approach ignores any area outside of its narrowly defined functions.

3 The paper gets data from several social channels instead of a single channel.

4 A product-service approach is built on the notion that businesses must deliver the product’s function rather than the actual product. The approach is common as a way to facilitate cooperative consumption of both goods and services.

5 Data provided from customers’ warranty claims.

6 Small and medium-sized enterprises.

7 Data provided from warranty claims in companies.

8 An API (Application Programming Interface) is a way for two or more programs to communicate with each other.

9 Operating Systems.

10 Term Frequency - Inverse Document Frequency.

11 Word2vec is a method used in natural language processing that utilizes a neural network model to understand the connections between words from a vast collection of text.

12 long short-term memory networks.