ELECTRONIC GAMING MACHINE PLAYSTYLE DETECTION AND RAPID PLAYSTYLE CLASSIFICATION USING MULTIVARIATE CONVOLUTIONAL LSTM NEURAL NETWORK ARCHITECTURE

Abstract

Electronic Gaming Machines (EGM) are common, anonymous, stateless gambling machines operated by a region’s lottery and situated in licensed venues. Previous work have shown that problem gambling detection is possible using EGM data, however, real-time customer personae identification might be even more important for stopping problem gamblers or suspicious playing behaviors. The following clustering algorithms were used and analyzed for the task of identifying different behaviours and personae types based on play style: CURE, DBSCAN, K-means with random initialization, TK-means++ (TKMPP), BIRCH, EMA, OPTICS and BANG. The results show that K-means with random initialization is the most suitable method for this task since it can scale well for the immense player data, it is efficient enough for multiple tests and analysis, and, most importantly, the results by K-means (clusters) are interpretable and meaningful. The experiments indicate that DBSCAN can be used before K-means to refine the results as it can identify results that cannot be identified by K-means. Inferred personae are used as labels for the playing sessions, and this data is used to train classifiers for playstyle detection. We identified methods suitable for real-time customer analysis, and the minimal number of initial transactions needed to successfully conduct this task. The classic classification methods such as perceptron, decision trees, and random forests are compared to deep learning based methods, showing that the best performance is obtained with a Multivariate Convolutional LSTM neural network. An important results is that increasing the number of analyzed transactions to more than 40 does not result in a large increase in performance.