Research On Cross-Domain Recommendation Based On Attention Mechanism

With the advent of the information age and big data era,the ever-expanding information and data have flooded every aspect of people's life.In order to obtain information efficiently and accurately,recommendation system has become an indispensable tool in people's daily life.At present,most researchers focus on the improvement of recommendation performance in a single domain.Most of these recommender models have the problem of data sparsity,so it is difficult to model the user's interest accurately.In fact,user information is usually cross platforms or cross domains.Information in different domains can be shared and complemented with each other.However,for a specific recommendation task,the importance degree of information in different fields is different.Recommendation algorithms based on attention mechanism can give different importance to different information,so that the attention of the model can focus on the key feature or important information,and reduce the role of other irrelevant features.In view of this,this paper proposes two cross-domain recommendation models based on attention mechanism between two domains.One is a cross-domain collaborative recommendation model based on adversarial adaptation and attention network(DAAN).The model considers both domain-shared and domain-specific knowledge between source domain and target domain.In this framework,we design a private encoder and an adversarial encoder to extract the private features and shared features of the two domains respectively,and use the attention network to couple them.In particular,in the adversarial encoder,we use deep adversarial adaptation to capture the shared characteristics of the common users between the two domains.In the experiment,we construct three cross-domain recommendation scenarios based on Amazon review data to verify the effectiveness of the DAAN model for top-N recommendation task.The experimental results show that the performance of the DAAN model is significantly better than that of the single domain,cross domain and the recommendation algorithm based on the confrontation learning.The other is a cross-domain recommendation model with auxiliary reviews and two-layer attention network(RACDR).In order to model the user's interest more accurately,we combine user reviews.In this framework,we design a private encoder and an auxiliary review encoder.The former is used to capture the features of users and items in the target domain,and the latter is used to capture the features of users in the source domain.In particular,in the auxiliary review encoder,we consider the importance of each comment in the user reviews document for modeling user preference,and use the attention network to automatically allocate the weight.At the top of the model,we stack a new attention layer to obtain the importance of user characteristics in source domain and target domain.In the experiment,we construct a crossdomain recommendation scenario based on Amazon review data and verify the effectiveness of RACDR model in the rating prediction task.Experimental results show that the recommendation performance of RACDR model is significantly better than the comparison algorithm used in the experiment.