Artificial Intelligence Based Employee Attrition Analysis and Prediction

The use of classification models to predict the probability of an employee leaving the job can greatly improve the ability of human resources to intervene in a timely manner and rectify the situation to prevent attrition. After the so-called "The Great Resignation", which took place after the pandemic period, studies on attrition have increased in order to regain the lost workforce and increase the loyalty of existing employees. In Turkey, however, studies in this field remain insufficient. The aim of this study is to predict the possible loss of employees, to take relevant measures and to reduce the financial losses of companies. The study was made available as SaaS (Software as a Service) based in order to facilitate accessibility to the target audience. For the analysis phase, the columns in the data set were analyzed with Pandas and Scikit-Learn library. Exploratory Data Analysis was performed by visualizing the analyzed data with Plotly and Seaborn libraries. Various insights were obtained with these inferences. The data was subjected to pre-processing stages such as cleaning, missing data completion, scaling, feature selection, data set balancing, dimension reduction, etc. to make it useful. Logistic Regression, KNN, SVM, Desicion Tree, Random Forest, ADABoost and Naive Bayes were used to train the model. In order to enrich the data set and increase the efficiency of the model, artificial intelligence-based synthetic data generation (Data Augmentation) was performed. In case of missing columns in the trained model, KNN-Data Imputation, one of the missing data completion methods, was used. In the optimization process of the model, hyperparameter optimizations were performed to achieve maximum efficiency with improvements. 5-fold crossvalidation prevented the model from over-learning. Performance was analyzed on the basis of Accuracy, Recall, Precision and F1-score metrics and the success criterion was determined as F1-Score. The model was presented as a web service with an accuracy rate of 0.99.