Multi-interest Aware Recommendation in CrowdIntell Network

In CrowdIntell Network, there are various transactions between different Digital-selfs, and these transactions are affected by the mind of the intelligence subjects. Mining and analyzing the mind of the digital-selfs is conducive to the better decision-making of the Intelligent subjects, and promotes the conclusion of transactions. We focused on how to model the interests of Intelligent subject in a complex environment, and how to use interests to make the recommendation. Considering the relevance and sequential characteristics of the behavior of the Intelligent subject in CrowdIntell Network is of great value for predicting future behavior. Most of the previous work was directly modeling the sequence patterns on interactive history, and then training a DNN model to learn the interest representation. Recent research focuses on interest mining by embedding both sequential and original characteristics of behaviors. However, in a real environment, the interests of users (for the convenience of description, we simplify intelligent subjects to users) are complex and diverse (e.g., a certain user likes documentaries and also likes science fiction movies), and a user may has multiple interests. In order to solve this problem, a Multi-interest Aware Recommendation in CrowdIntell Network is proposed, which involves the embedding layer, two-stage feature extraction layer and full connection layer. Specifically, the model is firstly constructed and trained to learn the user's multi-interest representation by considering the interaction between items in the behavior sequences, and then the convolutional attention mechanism is used to extract local features on the interests "image" as a higher-order user interest representation. Experiments on public datasets show that our method outperforms state-of-the-art sequential recommendation methods.