CASE: Predict User Behaviors via Collaborative Assistant Sequence Embedding Model

The order of behaviors implies that sequential patterns play an important role of the user-behavior prediction problem. Traditional behavior-prediction models use large-scale static matrices which ignore sequential information. Moreover, although Markov chains and deep learning methods consider sequential information, they still suffer the problems of the behavior uncertainty and data sparsity in real life scenarios. In this paper, we propose a collaborative-assistant sequence embedding prediction (named CASE) model as a solution to address these shortcomings. The idea is to mine sequential behavior patterns with strong intention-expressing ability based on a collaborative selector, and construct original behavior graph and intent determination graph (IDG), following which we predict user behavior based on graph embedding and recurrent neural networks. The experiments on three public datasets demonstrate that CASE outperforms many advanced methods based on a variety of common evaluation metrics.