Research On Sequence Recommendation Method Based On Attribute And Behavior Type Awareness

Recommendation systems have become an indispensable application in the fields of e-commerce and online services.Sequential recommendation systems are an important form of recommendation system,which use users’ historical behavior sequences to model their interests and can capture changes in user interests.However,if only users’ historical behavior is considered and information such as item attribute information and user behavior type is ignored,it will face problems such as poor generalization,data sparsity,and cold start.To solve these problems,integrating side information is a very effective method.By integrating side information,sequential recommendation systems can more accurately understand users’ interests and behaviors,thereby improving the accuracy and coverage of recommendations.In this context,this article proposes two auxiliary information-aware sequential recommendation methods.First,using item attributes in sequence recommendation can more accurately capture changes in user interests.There are two problems with current algorithms that integrate item attributes.Firstly,there is a lack of constraints on attribute information.Secondly,directly concatenating vectors or using multi-layer perceptron fusion methods can introduce noisy information and disrupt vector space consistency.To address these issues,this paper proposes the CAFICA-SR(using cross-attention fusion item category attributes Sequence recommendation)model,which uses cross-attention to integrate item category attributes into sequence recommendation.Specifically:(1)Extract item category interests to capture changes in user interests and use an auxiliary loss function to optimize the embedding representation of item categories.(2)Use sparse attention to reduce the impact of noisy items and use cross-attention to integrate item category attributes,without invading item embedding vectors.Second,using multi-behavior modeling in sequence recommendation can more accurately predict user interests under target behavior.There are two problems with single-behavior sequence recommendation models.Firstly,when the single-behavior sequence is short and there is less data on target behavior,there is a data sparsity problem.Secondly,when mixing behavior sequences and processing them only according to one type of behavior,the influence relationship between different behavior types is ignored.To fully utilize user behavior types,this paper proposes the MAMB-SR(Multi-attribute multi-behavior sequential recommendation)model.Specifically:(1)Use multiple types of behavior data and consider both item attributes and behavior type as auxiliary information to alleviate data sparsity.(2)In response to the characteristics of multi-behavior sequences,improve the Transformer structure by adding a behavior relationship matrix to the multi-head attention layer and innovatively generating the behavior relationship matrix,while extending the multi-layer perceptron to a multi-behavior multi-layer perceptron.Experimental results show that the two methods proposed in this paper have better performance than baseline models and similar models in terms of HR and NDCG indicators,mitigating data sparsity and cold start problems,and improving the accuracy and coverage of recommendations.