Hierarchical Attention Based Recurrent Neural Network Framework for Mobile MOBA Game Recommender Systems
Qiongjie Yao, Xiaofei Liao, Hai Jin · 2018
The mobile multiplayer online battle arena (MOBA) game is a genre of real-time strategy video games on mobile devices, such as King of Glory. The main business model is to drive players to purchase items like heroes or skins. Recommending items based on player interest is the core task of recommender systems. In the MOBA game, player interest changes over the game experience, which is implied in player behavior based on historical game matches. Match sequences, that consist of every match in the timeline, indicate how players interact with the game and the change process of player interest. Recurrent neural networks (RNNs) are employed by many recommendation scenes to model sequence data to profile user preference for better recommendation accuracy. However, their RNNs based frameworks ignore the interpretability of recommendation results, which is an important requirement for mobile MOBA games. To solve this challenge, we propose an interpretable RNN framework based on hierarchical attention in this work, which is inspired by the attention mechanism applied in machine translation. The main component long short-term memory (LSTM), that is the RNN variant, models player interest from historical match sequences, and the hierarchical attention is used to measure the effect factors of matches and behavior events happened in a match. To verify effectiveness, we train several models on real mobile MOBA game King of Glory datasets. Compared to non-sequence models, our model achieves 2% higher accuracy; with hierarchical attention, the proposed model can interpret the recommendation results effectively compared to naive RNN based models.