Research And Application Of Multi-interest Sequential Recommendation

The sequential recommender system attempts to predict the next interaction based on user's historical behaviors,which is a challenging problem due to intricate sequential dependencies and user's various interests underneath the interactions.Existing works regard each item that the user interacts with as an interest unit and apply advanced deep learning techniques to learn a unified interest representation.Although all of the methods mentioned above have achieved promising results,there are still limitations.1)A single representation is insufficient to capture multiple user interests.2)User's interests vary in multiple granularities.An item mirrors preferences for a specific item,while a set of items reflect general user interests,which are barely captured by a unified representation at the same granularity level.3)Due to the positional embedding and the attentive mechanism,the model will be dominated by the most recent interactions and irrelevant items still hold a few attention weights,which introduce noises.In order to solve these shortcomings,in this paper,we propose MiGRU4 Rec and CAN model.Then,we design a multi-interest sequential recommendation system and visualize the model's internal details,so that the researchers can have a comprehensive insight into how model works under the hood.We also propose a feasible deployment method for CAN model which reduces the computational cost of the model.The main contributions of this paper are as follows:1.Multi-Interest GRU for Recommendation MiGRU4 Rec We propose the Mi-GRU4Rec model to capture users' multiple interests.The multiple slots in the memory matrix are used to store the hidden state of different units,and each mem-ory slot represents a particular interest of the user,which captures users' multiple interests explicitly.Besides,we try to learn user's preferences towards each inter-est and introduce user's personalized information.Through our experiments,the MiGRU4 Rec model achieves a 2%-5% improvement over existing approaches.2.Capsule Attentive Network CAN In order to capture diverse users' interests at different granularity levels,we propose a capsule attentive network for the se-quential recommendation.We split the sequences into two parts according to sequential order and model the sequence at two different granularity levels to better learn and utilize the user's multiple interests.Specifically,the historical in-teractions converge to multiple user's interests at coarse-grained levels,and will be further analyzed with the recent interactions in the following modules at fine-grained levels,which retains favorable information in the historical interactions and eliminates the influence of the unrelated items to some extent.We conduct extensive experiments on three real-world datasets to verify the effectiveness of our proposed model and provide insight into how CAN works under the hood.3.Multi-Interest Sequential Recommendation System To address the problem that researchers are unsure about what the model learns during the training pro-cess and hardly make concrete improvements towards the model,we implement the multi-interest sequential recommendation system and deploy the CAN model online.In addition to the main functionality,we visualize the internal details of the model in the recommendation process,which provides researchers insight into how the model works under the hood and helps improve the model design.Be-sides,we improve our implementation of the CAN model and reduce the compu-tational overhead.