Abstract: A factor-level feature and attribute preference joint learning session-based recommendation model is proposed to address the problem of low recommendation accuracy caused by short sequences, sparse data, and difficulty in generalization. The model first learns user global-level session item embeddings by constructing a global level session item dependency perception graph. Then, using the method of disentanglement representation learning, the items in the conversation are decomposed into multiple relatively independent factor-level features to learn user factor-level interest preferences. Then, using contextualized self-attention graph neural networks, user preferences for session item attributes are captured. Finally, factor-level interest preferences and the project attribute preferences are jointly learned to obtain the user's final interest preferences, which in turn completes the session recommendation. Multiple experiments on two publicly available datasets, Diginetica and Cosmetics, have shown that our model outperforms the baseline model in comparison, verifying the recommendation performance and design rationality of our model.

Key words: session-based recommendation, factor-level feature, disentangled representation learning, global-level item dependency, graph neural network