Research On Session-aware Recommendation Algorithm Based On Deep Learning

With the vigorous development of internet technology and the popularization of mobile intelligent terminals,internet applications have covered various aspects of people’s clothing,food,housing,and transportation.The resulting massive data contains rich information and enormous value,but the colorful and complex massive data seriously interferes with users’ processing and selection of valuable information.Recommendation systems can recommend information that users are interested in without clear requirements,effectively solving the problem of "information overload" on the internet.As an important direction in the research field of recommendation systems,session-aware recommendation systems are further developed from session-based recommendation systems.It can simultaneously use the user’s current session information and historical interaction records to segment them according to the session,learn the user’s long-term preference information,and determine the user’s current session interest through the session level,so as to model the user’s preference and provide users with more accurate recommendation results.This thesis conducts research on session-aware recommendation systems.At present,current session-aware recommendation models fail to take into account the item transfer relationship between users’ implicit behavior sequences,and fail to consider the impact of users’ historical interests on the current session.In response to this issue,this thesis proposes a Session-Aware Recommendation with Self-Attentive Mechanism and Graph Neural Networks(SR-SGN)model based on graph neural networks and self-attention mechanism.SR-SGN can effectively capture the information propagation in the session sequence and the rich item conversion relationships in the session through building the user behavior diagram and the gating graph neural network,and then capture the long-term preference information in the user’s historical session and short-term interest in the current session through the self-attention mechanism.In addition,SR-SGN further utilizes the self-attention mechanism to model the impact of users’ long-term interests on current conversation behavior.In addition,this thesis also designs a gated fusion mechanism that adaptively assigns weights to long-term and short-term interests to accurately calculate user interest representations.On the basis of the SR-SGN model,this paper proposes an adaptive weight denoising mechanism to address the issue of losing a large amount of item information when generating histo rical session representations in current research and failing to effectively handle item-level noise within historical sessions.This mechanism aims to adaptively mask item-level noise in historical session sequences.In response to the current session-aware recommendation model’s problem of obtaining worse recommendation results due to session level noise when facing longer historical sessions,this paper introduces the prob-sparse selfattention mechanism to alleviate the impact of session level noise in long historical sessions by discarding tail query operations with lower session level contributions.Finally,this thesis proposes a model named Session-aware Recommendation Based on Adaptive Weight Denoising and Prob-Sparse Self-attention Mechanism(SRPAM).This thesis conducted a large number of experiments,such as comparative analysis and ablation validation on real datasets between the above two models and mainstream session recommendation and session-aware recommendation methods,verifying the effecti veness and rationality of the proposed method.Among them,the SRPAM model proposed in this thesis outperforms the most advanced session-aware recommendation methods at home and abroad.