Temporal-Based Action Graph with Sequential Pattern Mining for ChuRN Detection: a Playtika Case Study

Detecting churn is a critical task for any company, especially for online gaming companies, where early user churn rates are high. To improve customer retention rates, recent research has focused on using action graphs and deep neural networks to encode users' in-app transition patterns. However, existing action graphs approaches lack the ability to capture important characterizations of the given events data, such as sequential patterns and sparsity. In this paper, we propose a novel action-graph approach that addresses these limitations by incorporating sequence mining and temporal analysis techniques with graph neural networks. Specifically, we use a graph-based representation that captures the sequence of actions and the distribution of time-intervals between them. We then apply a convolutional neural network to learn the downstream task. Using real-world action data, we demonstrate the superiority of our proposed method compared to state-of-the-art approaches with respect to classification precision and recall. Overall, our approach presents a promising new direction for improving churn prediction in online gaming, as well as in other industries. We believe that our findings will be of interest to researchers and practitioners working on customer retention and churn prediction. The code and data for our approach are readily available on our GitHub repository.