Research On Personalized Paper Recommendation Method Based On Heterogeneous Network Embedding

With the rapid increase in the number of electronic papers,as well as the intersection and continuous development of different disciplines and fields,researchers have to spend a lot of time and effort to find the papers they are truly interested in.In recent years,the emergence and development of heterogeneous networks have made paper recommendation methods based on heterogeneous network embedding very popular,among which using meta-paths to mine heterogeneous relationships has become one of the most important ways.However,through analysis and summarization,this research finds that these paper recommendation methods based on heterogeneous network embedding still have some shortcomings.This research proposes corresponding solutions to these shortcomings in turn:Firstly,the existing sampling methods that use meta-path for data collection(such as random walking and backtracking)have insufficient diversity in the collected information,and may even collect invalid or noisy information.And a large number of recommendationrelated works only mine user reading preferences and paper audience preferences on static data,ignoring the fact that these features will dynamically change over time.To address these shortcomings,this research considers the directionality of links in heterogeneous information networks and designs a novel meta-path sampling method using the node’s in-degree and out-degree to reduce noise and increase diversity in the sampled paths.Furthermore,this research utilizes an attention-aware Bi-LSTM to mine short-term features and designs a compensation recommendation mechanism to integrate the short-term feature-based recommendation results into the long-term feature-based recommendation results,in order to capture the dynamics of features.Secondly,many recommendation methods that rely on meta-paths for network embedding inevitably require manual design and selection of meta-paths,which is very timeconsuming and labor-intensive.Therefore,this research proposes a universal HIN construction method and combines it with multi-attention mechanisms to achieve automatic design and selection of meta-paths.Finally,the existing recommendation methods do not consider whether the recommended papers are consistent with the user’s current research direction.Senior researchers often have rich historical reading data and involve multiple research directions.If the recommendation does not consider the user’s current research direction,the recommendation results will lack timeliness.To address this issue,this research innovatively utilizes BERT to pre-train the keyword features of papers,and combines them with self-attention mechanism to obtain the content features of papers and the research direction features of users.These features are matched by dot product to assist the main task of recommendation,ensuring the timeliness of the recommendation results.The proposed methods are verified by extensive experiments with various representative baseline methods on multiple data sets.The experimental results demonstrate the effectiveness of the proposed methods.Based on the proposed personalized paper recommendation methods,this research designs and implements a new personalized paper recommendation system,which comprehensively considers various practical needs of users in the paper recommendation scenario.