Machine learning with real-world HR data: mitigating the trade-off between predictive performance and transparency

Figures & data

Figure 1. Inductive research process using machine learning with an out-of-sample test.

Figure 2. Acquired longitudinal data for historical voluntary turnover. The direct impact (black) is investigated in this study.

Table 1. Descriptive statistical overview of available variables in the acquired dataset.

Nominal variables	N (679463)	Continuous variables	Mean	Std. Dev.
A1_Gender		A2_Age	49.09	9.13
woman	71.0%	A3_Number_of_children	1.13	0.94
men	25.7%	A5_Age_youngest_children	20.46	10.09
…unknown	3.3%	B1_Commute_distance_in_km	15.28	17.62
		B3_Early_retirement_rate	6.08	17.12
A4_Children_under_18_years		B4_Sickness_days	2.07	4.73
0	72.6%	B5_Degree_of_employment	90.75	14.00
1	27.4%	B6_Gross_working_days	20.06	2.40
		B7_Vacation_days	2.33	3.55
A6_Education_level		B8_Net_working_days	15.34	5.78
lower degree  university	49.5%	B9_Salary_increase_last_year	0.16	0.22
vocational training	46.7%	B10_Tenure_in_month	272.17	127.18
higher graduation  university	3.8%	B11_Salary_today_EUR	4068.64	889.97
		B12_Salary_increase_last_5_years	0.27	0.26
B2_Public_service_status_GER		B13_Parental_leave	0.06	1.11
No	78.0%	C1_Overall_employee_satisfaction	3.36	0.63
Yes	22.0%	C2_Employee_satisfaction_moving_average	3.50	0.32
		D1_Monthly_unemployment_rate_Germany	5.36	0.48
Voluntary turnover		D2_Monthly_number_of_vacancies_Germany	721.10	94.60
No_Turnover	99.2%	D3_Monthly_short_time_workers_Germany	1134.91	1659.11
Turnover	0.8%	E1_Monthly_Covid_strigency_index_Germany	19.67	29.15
		E2_Monthly_New_Covid_Cases_Germany	69835.00	181549.00
		E3_Monthly_Covid_Google_trends_Germany	74.70	116.21

Table 2. Selected ML algorithms with their primary advantage according to the transparency vs. performance trade-off.

Algorithm	R base function/package	Advantage
Generalised Linear Model	glm	Transparency
Elastic Net Regression	glmnet	Transparency
Classification Tree	rpart2	Transparency
Naïve Bayes	naivebayes	Transparency
Random Forest	ranger	Performance
Extreme Gradient Boosting	xgbDart	Performance
Generalised Boosted Machine	gbm	Performance
Feed Forward Neural Network	nnet	Performance

Table 3. Predictive performance measures of all used algorithms on test data.

Algorithm	Advantage	κ	ROC	Precision	Recall
Random Forest*	Performance	0.26*	0.87	0.18	0.54
Extreme Gradient Boosting	Performance	0.24	0.85	0.18	0.36
Generalised Boosted Machine	Performance	0.22	0.87	0.18	0.34
Classification Tree	Transparency	0.18	0.68	0.23	0.15
Feed Forwards Neural Network	Performance	0.16	0.68	0.17	0.16
Generalised Linear Model	Transparency	0.12	0.77	0.23	0.09
Elastic Net Regression	Transparency	0.12	0.76	0.20	0.09
Naïve Bayes	Transparency	0.07	0.59	0.05	0.23

Table 4. Confusion matrices on test data of random forest (highest predictive performance overall) and classification tree (most successful alternative algorithm with advantage transparency).

Random forest				Classification tree
		Reference				Reference
Prediction		No Turnover	Turnover	Prediction		No Turnover	Turnover
	No Turnover	111,089 (96.60%)	513 (0.45%)		No Turnover	113,324 (98.53%)	940 (0.82%)
	Turnover	2,800 (2.43%)	598 (0.52%)		Turnover	565 (0.49%)	171 (0.15%)

Figure 3. Predictive performance vs. transparency trade-off on test data (confidence interval 0.95).

Figure 4. Permutated feature importance for the top 20 predictors (relative change in performance, confidence interval 0.95).

Figure 5. Accumulated local effect plots for the top 12 predictors according to permutated feature importance. Grey = all employees, blue = men, red = women.

Figure 6. Top-10 SHAP values for two employees successfully predicted turnover candidates (true-positive).

Data availability statement

Data available on request due to privacy/ethical restrictions.