Research And Implementation Of Session-Based Recommendation Algorithms Based On Graph And Deep Learning

Session-based recommendation system is a system that analyzes sessions,i.e.,the interaction history of users over a period of time,in order to provide recommendation services that meet users’ personalized needs.Session-based recommendation systems,which have received a lot of attention from researchers recently,are more focused than other recommendation systems on capturing short-term and dynamic changes in users’ interests and improving the accuracy and real-time of recommendations by analyzing the context of the sessions.This thesis focuses on session-based recommendation algorithms to address some shortcomings of the current mainstream models: First,there are complex dependencies between the user’s next choice and the interactions in the session,and most of the current mainstream approaches treat the session sequence as a simple sequence of items,ignoring the rich contextual information in the session,resulting in a less comprehensive and accurate model of the user’s interests.Second,some recent works have tried to use the behavior type information in a session to improve the performance of recommendation systems,mainly focusing on capturing the sequential dependency patterns between behavior types,failing to make good use of the behavior semantic information.This thesis suggests solutions from the perspectives of attention mechanisms absorbing side information and building behavioral semantic awareness networks,respectively,with the research goal of increasing the performance of session-based recommendation algorithms.First,a multi-information enhanced attention network incorporating multiple pieces of information is proposed to address the problem of inaccurate user interest modeling due to insufficient utilization of session information.A more comprehensive and reasonable representation of user interests is constructed by fusing information about items,categories,and behaviors in interactions in a session to model users’ current interests and using an augmented attention fusion mechanism to reduce the influence of irrelevant items in a session on modeled user interests.Experimental results on two real datasets demonstrate the effectiveness of the proposed model while revealing the importance of incorporating category and behavior information to improve the performance of the session-based recommender system.Secondly,based on revealing the effectiveness of multiple behavior information,the way of modeling and incorporating information of multiple behavior types is further deepened,and a multi-behavior type dependent complex challenge is addressed by proposing a multi-behavior semantic-aware graph neural network-based session-based recommendation algorithm.The algorithm learns item dependency information under different behavior patterns by constructing three behavior semantic dependency graphs from the perspective of behavior type semantics.Then,it achieves the full mining and utilization of behavior types in a session by fusing the intention strength information contained in the behavior types through a behavior-aware self-attentive mechanism.Experiments on three datasets demonstrate the rationality of the proposed model for the utilization of multiple types of behavioral information and improve the performance of the session-based recommendation algorithm.Finally,based on the research of the session-based recommendation algorithm proposed in this thesis,a remote sensing image recommendation system is developed based on the demand for remote sensing image recommendation in actual business,which can push remote sensing images to users in real time based on the user’s behavioral data,proving the application value of the algorithm proposed in this thesis.