Research On Deep Collaborative Filtering Recommendation Algorithm Integrating Attention Mechanism

The rapid development of Internet technology result in quantity of information increased on the network.Faced with such a massive amount of data,users cannot satisfy the needs of information in a timely and effective manner,which makes for the problem of "information overload".As an effective strategy to settle this matter,recommender system has received extensive attention.The final recommendation performance is largely determined by the recommendation algorithm,which is the kernel of the recommendation system.Collaborative filtering recommendation algorithm as an effective method to build recommendation system has been applied in many fields.However,this kind of algorithm considers that a group of historical items that users interact with have the same influence on users.Therefore,it is difficult to capture users' preferences in a finer granularity.In addition,when modeling user historical behavior data,only the user-item relationship or the item-item relationship is considered,so that the data information in user historical behavior cannot be fully mined,which restricts the performance of recommendation algorithm.Faced with the above problems,in order to improve the performance of collaborative filtering algorithm,we take the collaborative filtering algorithm as the primary research content,and focus on the combination of relevant technologies in deep learning and collaborative filtering recommendation algorithm to build the corresponding deep collaborative filtering recommendation model.Specifically,the attention mechanism is introduced and a dual most relevant attention network is designed to model the user's historical behavior sequence,in which to settle this matter that the previous algorithms lack the importance distinction of user's historical interaction items.In terms of the issue that the collaborative filtering algorithm only considers the user-item relationship or the item-item relationship when modeling user historical behavior data,a relational collaborative filtering model combining attention mechanism is proposed to model the two kinds of relationships in user historical behavior data simultaneously.The main work of this dissertation includes:1.This dissertation firstly describes the theoretical basis which includes recommender system,collaborative filtering algorithm,deep learning technology,and investigates the current research situation and existing weakness of recommendation algorithms at domestic and overseas.Secondly,the collaborative filtering recommendation algorithm and technologies of deep learning are introduced at length.Finally,aiming at some problems in current collaborative filtering algorithms,a collaborative filtering recommendation algorithm on account of the dual most relevant attention network and a relational collaborative filtering recommendation algorithm integrating attention mechanism are proposed on account of the existing collaborative filtering algorithm combined with deep learning technology.2.In view of the current collaborative filtering algorithms that ignore the impact of a set of historical items interacted by the user,a new collaborative filtering recommendation algorithm on account of the dual most relevant attention network is proposed(DMRACF),which contains two layers of attention network.First,the item-level attention network is used to assign different weights to different historical items to capture the most relevant items in the user's historical interactive items;Then,the item-level attention network is used to perceive the degree of interaction between different historical items and the target item;Finally,the twolayer attention network is used to simultaneously capture the user's fine-grained preferences on historical interactive items and target items,so as to make better recommendations for the next step.The experimental results on two real data sets show that the algorithm is is better than the benchmark model.3.Some current collaborative filtering algorithms,including DMRACF,only use the user-item relationship or the item-item relationship in the user's historical behavior data alone,and rarely consider these two relationships at the same time,which may potentially limit the recommendation performance.For the sake of this problem,a relational collaborative filtering recommendation algorithm(SARCF)that integrates attention mechanism is proposed to jointly learn two kinds of relations in user historical interaction data to learn effective user item feature representation and improve recommendation quality.Specifically,the SARCF algorithm proposed in this dissertation divides the modeling of user historical interaction data into two parts: user-item preference modeling to capture user preference relationships with items;itemitem relationship data modeling to capture the relationships between items.At the same time,the self-attention mechanism is introduced to capture the influence of the user-item relationship on the item-item relationship modeling,and distinguish important historical items for the user.Finally,multi-layer perceptron is used for joint learning of the two parts to mine the deep feature representation of user and item data.Experimental results on two real data sets show that the algorithm has superior performance.