Stolen prediction in contactless shopping of SCO registry systems with machine learning

Nowadays, machine learning algorithms are used to make examinations such as prediction, classification and clustering in many different areas. Machine learning algorithms provide benefits in many areas and are preferred because of their high performance. Within the scope of this study, it is used for the classification and estimation of stolen detection in contactless purchases in Self Checkout (SCO) cash registers in the retail sector. In some companies in the retail sector, there are SCO cash registers that can be used with or without a cashier. In SCO cash register systems, the customer can complete the payment and shopping steps on his own. In this study, sales and cancellation data from contactless shopping experiences in SCO cash registers are used. Since the cancellation data caused theft, it was classified by detecting whether the data was canceled or sold using machine learning algorithms. In classification, Logistic Regression, Decision tree (C4.5), K - Nearest Neighbor (KNN), Gradient Augmented Trees (GBT) and Random Forest (RF) algorithms, which are very common in literature research, were applied and the results were compared. Comparisons were made on models' accuracy, f1-score (f1-score), precision, recall and Area Under The Curve (AUC) values. In addition, basket analysis was performed on the canceled receipts with the Apriori algorithm, and the products purchased together were evaluated. As a result of the study, when the algorithms are compared, the Decision Tree algorithm gives the best performance with an accuracy rate of 96.49% and an AUC value of 0.98%. Here, the best hyper parameter values determined as Gini index, maximum depth 40 and maximum feature 8 were taken as the division criteria. The lowest performance belongs to Logistic Regression model with 65.62% accuracy rate and 0.72 AUC value. As a result of the study, it is seen that it is appropriate to classify with the Decision Tree model, which exhibits the best performance. At the same time, taking into account the product associations in the canceled receipts, it will be possible to take measures to prevent product loss and theft.