Research On Sequential Recommendation Algorithm Based On Fusion Of Heterogeneous Data

The development of cloud computing and big data has promoted the explosive growth of Internet data.In this context,people increasingly rely on recommendation systems to filter information.Whether it is information producers or information consumers,the emergence of recommendation systems has brought them a lot of benefits.For information producers,the recommendation system allows them to display the massive amount of information they hold to different users in a targeted manner,which improves the utilization of information while improving the quality of service and helps to improve user satisfaction.For information consumers,the recommendation system quickly and effectively provides them with the information they need,reducing time wastage,improving search efficiency,and facilitating people's lives.Since users' historical behavior records are usually in chronological order,the sequential recommendation problem has become the focus of research in recent years.In the traditional sequential recommendation system,it is generally considered that both the user's identity information and the user's historical behavior sequence are available.In fact,in many scenarios,the user's identity information is not available due to privacy protection and other reasons.At this time,the traditional sequential recommendation algorithm cannot work.The session-based recommendation algorithm is an algorithm that recommends the next item through an anonymous user behavior sequence in this scenario.As a branch of sequence recommendation algorithm,this kind of algorithm has become a research hotspot of various universities.The current session-based recommendation mainly use recurrent neural networks and graph neural networks as basic models to model session sequences.However,the algorithm based on recurrent neural network has strong sequence dependence and cannot capture the nonsequence dependence in the sequence.Although the graph neural network can model the complex transfer relationship between the actions in the sequence,it achieves a better recommendation result.However,this type of method only conducts modeling from a single level,and there is a problem of insufficient utilization of session information.Therefore,this paper studies the session-based recommendation algorithm fusing heterogeneous data.The main work and innovations are as follows:(1)Aiming at the problem that the session-based recommendation model based on graph neural network is only modeled from a single level,this paper proposes a session-based recommendation algorithm enhanced by category graph(Ca Se4SR).Firstly,a directed graph is constructed separately for the item sequence and the category sequence,and the gated graph neural network is used to model the two sequences respectively,and then the learned item representation and category representation are fused,and then the attention mechanism is used to obtain the user's global preferences,combined with local preferences to obtain the final representation of the session sequence to predict the next behavior.(2)Aiming at the problem of neighbor information sparsity in neighbor session selection,this paper proposes a category-aware collaborative session-based recommendation algorithm(Ca Co SR),which proposes a category-aware collaborative session-based recommendation algorithm(Ca Co SR).The neighbor selection mechanism,which combines the item level and category level information to find similar neighbor sessions,effectively alleviates the problem of sparse neighbor information.(3)Aiming at the problem that the user's personalized preferences cannot be captured due to the unavailability of user identity information,this paper attempts to model the user's behavior type sequence and proposes a category-aware collaborative personalized conversation recommendation algorithm(P-Ca Co SR).The recurrent neural network is used to model the user's behavior type sequence,and the user's general preference information is obtained,which is used as personalized information to assist the modeling of the session.(4)In terms of practical application,this paper uses FPMC,GRU4Rec-top K,NARM,STAMP,SR-GNN,CSRM and other models as the benchmark model through experimental verification on two real datasets,and uses Recall@20 and MRR@20 is used as the evaluation index.The experimental results show that the recommendation performance of the proposed algorithm on all datasets is consistently better than these benchmark models,and can predict the user's next behavior more accurately.