Research On Personalized Recommendation Based On Sequence Data And Session Dat

The advent of the big data era has brought a lot of convenience to our work and life,as well as the problem of “information overload”.In order to solve the increasingly serious problem of “information overload” in the internet environment,personalized recommendation technology has attracted extensive attention from academia and industry.The basis of personalized recommendation is user interaction data,and the core is the representation of user dynamic preference.Around this core,this paper propose three specific research problems,based on sequence data and session data most concerned in personalized recommendation.(1)When constructing recommendation model based on sequence data,how to more accurately represent users’ interaction intention and dynamic preference?(2)When building a recommendation model based on session data,how to accurately convert the user’s interaction sequence into the user’s real interaction mode?(3)How to accurately identify different interaction purposes of users in interactive sessions?Most of the existing sequence recommendation models take the recurrent neural network as the main structure,and adopt a unified modeling framework: take a single interaction behavior as the input unit to model the user’s interaction intention,and take the last hidden state of the recurrent neural network as the user’s dynamic preference representation to generate the recommendation list.In fact,the user’s interaction intention is often represented by multiple interaction items in the form of session.Therefore,how to construct a more reasonable input unit to better represent the user’s interaction intention is an important factor to promote recommendation.In addition,it is not enough to use only the last hidden state of the recurrent neural network to capture user preferences,because the memory cell of the recurrent neural network is difficult to retain the global sequence pattern.Therefore,how to effectively preserve the global sequence pattern is another key factor to enhance the effect of recommendation.To solve the above problems,this paper proposes a new hierarchical time-based directional attention(HTDA)network framework,which includes user intention representation layer and user dynamic preference representation layer.Specifically,in order to solve the matching deviation between interaction sequence and user’s real interaction intention,a fine-grained user intention representation method is constructed in this paper.Firstly,by introducing time threshold,the original interaction records of users are divided to obtain coarse-grained representation of user intention.Then,the multi-dimensional self-attention mechanism is applied to the coarse-grained user intention representation to capture the transformation mode between the interaction features of interaction items,and then obtain the fine-grained user intention representation.In order to capture the global sequence of interaction,this paper constructs a new time-based directional attention mechanism.By introducing the time factor,the attention mechanism can perceive the time attribute of the interaction item,and by introducing the mask matrix,the attention mechanism can capture the directivity between the input units.Experimental analysis demonstrates the effectiveness of the HTDA framework proposed in this paper.Compared with sequence recommendation,session-based recommendation can not directly obtain users’ historical interaction records,and the interaction data is more sparse.Therefore,it has higher requirements for the ability of the model to accurately capture users’ preferences.Most of the existing session-based recommendation models are built under the following basic assumptions,that is,the sequence relationship between the user’s interaction items in the interactive session is one-to-one corresponding to the user’s real interaction mode,that is,the sequence relationship between the adjacent interaction items in the session is strictly irreversible.In fact,the strict sequential relationship between interactive items in a session is not completely equivalent to the user’s real interaction mode.Because the real interaction mode of users in the session is not only the sequence mode,but also the co-occurrence mode,which is the alternating existence of multiple modes.The so-called co-occurrence mode means that two interaction items always appear in the interaction list in pairs,but there is no obvious sequential relationship between them.Therefore,how to accurately convert the sequence relationship between interactive items in the session into the user’s real interaction mode will help to expand the recommendation model and further enhance the recommendation effect.To solve the above problems,this paper proposes an Enhanced Graph Neural network(E-GNN)for session-based recommendation.In E-GNN,this paper focuses on how to identify users’ real interaction patterns based on the sequence relationship between interaction items in the session.In order to identify the real interaction mode of users in the session,E-GNN comprehensively considers the global interaction behavior of all users and the interaction behavior of the target user in the current session.Specifically,in order to explicitly model the coupling relationship between various interaction modes in the current session,this paper first constructs a Weighted Global Item Graph(WGIG)based on the historical interaction sessions of all users.The weight of directed edges between any two adjacent nodes in WGIG is determined by the ratio of the number of edges in the same direction to the sum of all edges between the two nodes.The weight value in WGIG represents the sequence and co-occurrence between its corresponding interaction items.The closer the weight value is to 1,the stronger the sequence is,and the closer it is to 0.5,the stronger the co-occurrence is.Then,a Local Session Graph(LSG)is constructed based on the sequence relationship between interactive items in the current session.Each node in the LSG represents a non recurring interaction item in the session,and the directed edges between nodes correspond to the sequential relationship between the interaction items in the session.Finally,using the fusion algorithm proposed in this paper,the weight values of the corresponding edges in LSG and WGIG are integrated to obtain the enhanced graph E-GNN representing the user’s real interaction mode.The E-GNN learns the representation vector of each node in the graph through the gated recurrent unit,and uses the attention mechanism to generate the representation of the whole session graph.Experimental analysis demonstrates the effectiveness of the E-GNN framework proposed in this paper.In session-based recommendation,how to more accurately capture users’ current preferences according to limited interaction records has always been a research hotspot.Previously,we have explored the relationship between user interaction sequence and actual interaction mode in session,and proposed a new framework based on graph neural network.In addition to the above-mentioned problem of user interaction pattern recognition,user interaction purpose recognition in interactive session is also a very important research problem.The identification of user interaction purpose in a session will directly affect the distance between the overall representation of subsequent sessions and the user’s real preferences.The existing session-based recommendation usually uses the attention mechanism to aggregate the representation vectors of each interaction item in the session to generate the overall representation of the session,which represents the user preference.A basic assumption of the above research paradigm is that all interaction items in a session represent the same or similar user interaction purposes.In fact,the interaction purposes of users in a session are often diverse,and there are usually great differences between them.Therefore,how to accurately identify the user’s interaction purpose in the session to enhance the overall representation of the session is the key to enhance session-based recommendation.To solve the above problems,this paper proposes a new Purpose Aware Sessionbased Recommendation(PASR)model.In PASR framework,firstly,an interaction session graph is constructed based on the user’s current interaction sequence to capture the rich association information between adjacent interaction items,so as to obtain the representation vector of interaction items with strong expressiveness.Subsequently,according to the similarity between interaction items,an Interaction Purpose Recognition Algorithm(IPRA)is proposed to divide the original user interaction sessions,and the sub sessions represent different interaction purposes of users.Then,representation vectors representing different interaction purposes of users are generated for different interaction sub sessions.Finally,the user purpose representation vector,short-term preference representation vector and long-term preference representation vector are mapped to a unified vector space by linear transformation,and matched with the item representation in the candidate set to obtain the final recommendation list.Experimental analysis demonstrates the effectiveness of the pasr framework proposed in this paper.