Research On Multi-Behavior Commodity Recommendation Model Based On Deep Learning

With the continuous expansion of the e-commerce market,the number of online commodities has grown exponentially.Faced with a large number of products,how to efficiently obtain products that meet the needs has become the most concerned issue for e-commerce merchants and users.The product recommendation system can spontaneously find the products that best meet the user’s preferences,which is an important method to solve the problem of "information overload".As the user needs are more closely integrated with the recommendation model,the recommendation performance is also facing higher requirements.How to further improve the accuracy of system recommendation is a continuous concern of academia and industry.Most of the existing recommendation algorithms based on deep learning only consider a single type of user behavior,ignoring the multiple behaviors of users when browsing products,such as clicking,adding to shopping cart,collecting,purchasing,etc.These different behaviors contain a large amount of user preference information that is beneficial for recommendation.This thesis designs a recommendation model for a variety of interactive behavior data in product recommendation,mainly including three tasks:(1)In order to fine-grained consider the dependencies between various types of user behaviors,this thesis proposes a multi-behavior product recommendation model(IBDM for short)that integrates behavior dependencies into a multi-task learning framework.The IBDM model learns a separate interaction function for each behavior type,introduces a gating mechanism to adaptively learn the relationship between behaviors according to the actual situation,and introduces a tower structure for each behavior to output the user’s predicted value under this behavior.The recommendation effect of the model is verified on two real-world datasets.Compared with the classic recommendation model,the IBDM model has improved HR and NDCG indicators.(2)In order to better learn the high-level interaction information between user items and provide higher-quality recommendation results,this thesis proposes a multi-behavior item recommendation model based on relational graph convolutional networks,referred to as RGCMB.The RGCMB model introduces a relation-aware graph convolutional propagation layer,which fuses nodes(user nodes and product nodes)and relationship representations through combined operations and combines graph convolutional networks to obtain high-order connectivity between heterogeneous feedback data.By comparing with several other multi-behavior recommendation methods,the RGCMB model has obvious improvement in HR and NDCG indicators.(3)In order to consider the user’s dynamic preference,this thesis proposes a multi-behavior product recommendation model based on deep Q network fusion of long-term and short-term preferences,referred to as ILSPDQN.The ILSPDQN model uses the self-attention block to extract the user’s short-term preference representation from the user’s recent interaction sequence,extracts the user’s long-term preference representation from the user’s demographic characteristics and interacted products,and then fuses the long-term and short-term preference representations to generate dynamic preferences.The method of maximizing the cumulative expected reward value is used to learn the model and give the recommendation result.Experiments on the JData dataset show that the ILSPDQN model outperforms other baseline models.