Research On Dynamic Knowledge Graph Recommender System Based On User Behavior Propagation

In the Internet era of information overload,recommender system plays an increasingly important role in helping users find the content they need in a large amount of information.However,there are often problems of cold start and sparse data in the recommender system.The user ’s implicit preference mining is not deep enough,the real-time performance of the recommendation results is poor,and the system ’s self-improvement correction function is lacking.At present,due to the difficulty of data acquisition,the available effective data is less,and the extractability of user behavior preference is poor.The recommender system has the following difficulties : 1)User preferences are flexible,selection bias and exposure bias,resulting in repeated recommendation of popular goods,user potential preference mining is not deep enough,and information utilization is not comprehensive enough.2)Most of the existing researches on shilling attack detection rely on the explicit false profiles presented by the user ’s historical data to find shilling attackers.Important features such as user propagation and social feedback in social networks have not been applied to the detection of shilling attackers.3)Cold start and data sparsity are usually unavoidable,and there will be situations where side information is not available.The application of deep learning in recommender systems relies on big data,which limits its application in many scenarios.In order to improve the diversity,adaptability,security and accuracy of recommendation results,this paper proposes a dynamic knowledge graph recommender system based on user behavior propagation.The specific work includes :(1)A Recommendation Approach Based on Heterogeneous Network and Dynamic Knowledge Graph(HN-DKG)is proposed.According to the composite behavior of users in cross-domain and multi-platform,multi-modal heterogeneous nodes are mined,heterogeneous network graphs are constructed,and basic knowledge graphs are established.The multi-head attention mechanism of the graph attention network(GAT)component is used to focus on the temporal information and the graph structure information of the knowledge graph,so as to enhance the relationship between multimodal heterogeneous nodes and construct a dynamic knowledge graph.The improved Ripple Net is used to discover the potential interest of users,and the predicted products are scored and sorted.The mechanism of user seed cluster,propagation blocking and random seed is set to make the recommendation results more accurate and diverse.Using the public test data set,compared with the baseline algorithm,experiments show that the proposed method has better performance in the effectiveness and diversity of recommendations.In addition,ablation experiments also prove the effectiveness of each link in the algorithm.(2)A security detection approach based on autonomy-oriented user sensor in social recommendation network(AOUSD)is proposed,which simulates users as autonomous social sensors.Based on the mechanism of information dissemination,information feedback and information disappearance of user autonomous sensors in social networks,the interaction behavior model of users in social recommendation is constructed,and the variable time function related to events is considered to generate the dynamic knowledge graph of user relationship based on social sensors.The hierarchical clustering method is used to generate the initial suspicious candidate group,and the graph group detection clustering method is applied to obtain the attacker on the dynamic knowledge graph.The performance of AOUSD is tested by simulating the shilling attack environment on the Netlogo platform.The Amazon data set is used for experiments,and the proposed AOUSD algorithm is compared with other algorithms.The experimental results show that the proposed AOUSD method has advantages in efficiency and accuracy of shilling attack detection.(3)A dual learning-based recommendation approach(DLRA)is proposed.DLRA regards the recommendation task as two independent sub-tasks-main task and dual task,which show strong duality in DLRA.The initial task is item-centered,aiming to find users who have high evaluation of items,while the dual task is user-centered,aiming to recommend favorite items to users.These two tasks have strong duality in recommendation space,selection probability and recommendation basis.Based on the data sets of Movielens and Book Crossing,the data sparse and cold start recommendation scenarios are simulated.Experimental results show that DLRA achieves significant improvement in the case of sparse labeled data,and is superior to other hybrid recommendation methods and deep learning strategies,with smaller prediction error and better recommendation accuracy.