A paper recommendation method, system and storage medium based on a heterogeneous information network

Abstract

This invention discloses a paper recommendation method, system, and storage medium based on heterogeneous information networks, belonging to the field of computer technology. The method includes the following steps: First, feature extraction and classification of paper texts are performed using a domain dictionary and a composite feature extraction algorithm. Then, a heterogeneous information network containing multiple entities and semantic relationships is constructed based on the extracted features. Next, a temporally aware bimodal research interest vector model is constructed by fusing user historical behavior and a temporal decay function. Then, multi-source similarity of text, structure, and relationships is calculated in the heterogeneous network and dynamically weighted and fused with the user's research interest similarity. Finally, reinforcement learning is used to optimize the fusion weights to generate a personalized Top-N recommendation list. This invention effectively solves the problems of traditional methods, such as difficulty in mining deep semantic associations, neglecting the temporal evolution of interests, and lack of adaptive optimization, significantly improving the accuracy, timeliness, and personalization of academic paper recommendations.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a paper recommendation method, system, and storage medium based on heterogeneous information networks. Background Technology

[0002] In recent years, with the rapid development of science and technology, the number of academic papers has grown exponentially. Researchers face the significant challenge of quickly and accurately locating papers relevant to their research interests from a massive amount of academic literature. Against this backdrop, personalized paper recommendation systems have emerged, becoming an important tool for alleviating information overload and helping users discover relevant academic achievements. Academic social networking platforms contain rich, diverse, and interconnected academic information resources, providing researchers with an important platform for sharing academic findings and exploring new research areas.

[0003] Currently, mainstream recommendation system algorithms can be mainly divided into three categories: (1) content-based recommendation methods, which mainly recommend based on the matching degree between the content of the paper and the user's interests; (2) collaborative filtering-based recommendation methods, which utilize the behavioral patterns of user groups to mine potential interest associations; and (3) hybrid recommendation methods, which attempt to combine the advantages of multiple recommendation strategies. However, these traditional methods have obvious technical limitations: First, they are mostly limited to modeling the binary relationship between users and papers, and fail to fully explore the rich semantic associations in academic networks; second, in terms of feature processing, existing methods usually adopt fixed word segmentation dictionaries and static feature weighting methods, lacking the ability to adapt to academic terminology, and cannot effectively handle linguistic phenomena such as polysemy.

[0004] Traditional methods have significant shortcomings in modeling research interests. Users' research interests exhibit distinct temporal evolution characteristics, encompassing both long-term, stable, and continuous research interests, as well as short-term, changing, and immediate research interests. However, existing recommendation algorithms generally lack sophisticated modeling of these temporal characteristics, failing to accurately capture the dynamic changes in users' research interests. This leads to systems often recommending papers that are related to a user's historical interests but are no longer of interest, severely impacting the practical value of the recommendation results.

[0005] Furthermore, in the area of ​​academic network representation learning, traditional methods have failed to fully utilize the rich structural and semantic information within heterogeneous information networks. Academic networks contain various types of entities, such as papers, authors, institutions, journals, conferences, and keywords, as well as complex semantic relationships like citation, writing, affiliation, and publication, constituting a typical heterogeneous information network. Most existing recommendation methods neglect the rich interrelationships between these entities, thus limiting the accuracy and practicality of recommendations.

[0006] At the system architecture level, traditional recommendation systems generally suffer from single module functions and lack of collaborative optimization. Specifically, this manifests in the following ways: (1) the feature extraction module often adopts a fixed processing flow and cannot be adaptively adjusted according to the characteristics of the domain; (2) the text understanding module is mostly based on shallow machine learning models and is difficult to capture deep semantic features; (3) the network modeling module does not fully explore the relationships between heterogeneous networks; (4) the interest modeling module lacks consideration of temporal characteristics; (5) the recommendation generation module usually adopts static weight configuration and lacks a dynamic optimization mechanism based on user feedback; (6) the system lacks continuous learning ability and cannot perform online updates and performance improvements based on user behavior feedback.

[0007] These technical shortcomings severely limit the practical effectiveness of paper recommendation systems in real academic scenarios. There is an urgent need for an intelligent recommendation solution that can deeply integrate heterogeneous network information, accurately model the temporal evolution characteristics of user interests, and possess continuous learning capabilities. Summary of the Invention

[0008] In view of this, in order to solve the technical problems existing in the prior art, such as coarse feature extraction granularity, lack of temporal awareness in interest modeling, insufficient utilization of heterogeneous network relationships, and lack of continuous optimization capability of the system, the purpose of this invention is to provide a paper recommendation method, system and storage medium based on heterogeneous information networks. By constructing a multi-layered technical architecture, intelligent paper recommendation is achieved, which significantly improves the accuracy, practicality and effectiveness of the recommendation function.

[0009] The technical solution adopted in this invention is: a paper recommendation method based on heterogeneous information networks, comprising the following steps: Feature extraction and text classification: Key features are extracted from the paper text data using a domain dictionary and a composite feature extraction algorithm, and a model is trained to classify the papers. Heterogeneous information network construction: Based on the extracted features, a heterogeneous information network is constructed from diverse academic entities and relationships to store complex academic information in a structured form; Temporal-aware interest modeling: Integrating user historical behavior with temporal decay functions to construct an instantaneous and continuous dual-modal research interest vector model; Multi-source similarity calculation: In the heterogeneous information network, multi-granularity similarity is calculated by comprehensively measuring text attribute similarity, network structure similarity, and relationship similarity to fully measure node correlation; Research interest similarity calculation: Based on the bimodal research interest vector model, the similarity between the user's immediate research interests and continuous research interests is calculated separately, and then dynamically weighted and fused. Personalized recommendation generation: The multi-source similarity and the research interest similarity are fused, and reinforcement learning is used to optimize the fusion weights to generate a Top-N paper recommendation list.

[0010] As a further improvement to the above scheme, the feature extraction and text classification include the following steps: We construct a domain-adaptive word segmentation dictionary, integrate a keyword database from academic papers, and use the jieba word segmentation algorithm for context-aware word segmentation. The TF-IDF algorithm combined with information gain is used for feature weighting and selection. By calculating the word frequency and inverse document frequency of each keyword, the term weight is calculated to determine the importance of each keyword to a text or a corpus, thereby achieving highly discriminative feature extraction. The CBOW model based on Word2Vec is used to train word vectors, and a Naive Bayes classifier is used for text classification. Cross-validation is then used to optimize the classifier's performance.

[0011] As a further improvement to the above solution, the construction of the heterogeneous information network includes the following steps: Define multiple entity types, including papers, authors, institutions, journals / conferences, and keywords; Define semantic relationship edges, including reference relationships, writing relationships, membership relationships, and co-occurrence relationships; Create nodes and edges based on different types of entities, map structured data into a heterogeneous information network graph structure, and import node attributes and edge weights.

[0012] As a further improvement to the above scheme, the temporal-aware interest modeling includes the following steps: Research interest time series are constructed based on user historical behavior data, which includes paper reading records, download records, citation records, and collection records; A sliding time window mechanism was used to divide the research interest periods. The continuous research interests covered the long-term research history, and the time window was set to T1 years. The immediate research interests focused on recent research dynamics, and the time window was set to T2 years. T1 is longer than T2. Interest evolution is dynamically quantified using an improved Logistic time decay function; Based on the quantitative results, continuous research interest vectors and immediate research interest vectors are extracted to construct a bimodal interest fusion model, which is then dynamically adjusted based on user behavior feedback.

[0013] As a further improvement to the above scheme, the multi-source similarity calculation includes the following steps: Node representation vectors are learned based on network embedding technology, and the metapath2vec algorithm is used to perform random walk sampling on heterogeneous information networks to generate semantically preserved node embeddings. Calculate multi-granularity semantic similarity between different types of nodes; A multi-source similarity fusion function is constructed to achieve adaptive weight allocation based on the attention mechanism, dynamically adjusting the weights of each similarity component according to the node type and query context.

[0014] As a further improvement to the above scheme, the calculation of multi-granularity semantic similarity between different types of nodes includes the following steps: (1) Text attribute similarity: The semantic matching degree of the title, abstract and keywords is calculated based on the cosine similarity of word vectors; (2) Network structure similarity: The Jaccard similarity of the network topology is used to calculate the overlap of the neighborhood of nodes; (3) Relationship similarity: The PathSim algorithm based on meta-path is used to calculate the strength of multi-hop relationships between nodes.

[0015] As a further improvement to the above scheme, the calculation of research interest similarity includes the following steps: Obtain the user's real-time research interest vector and ongoing research interest vector respectively; Obtain keywords and time information related to candidate papers; Obtain word vectors of interests and paper keywords using the Word2Vec model; The similarity between candidate papers and immediate research interests and ongoing research interests is calculated separately and assigned weight parameters. Finally, a weighted sum is performed to obtain the research interest similarity.

[0016] As a further improvement to the above solution, the personalized recommendation generation includes the following steps: Construct a multi-level recommendation scoring function that integrates multi-source similarity and research interest similarity; A Q-learning-based reinforcement learning algorithm is used to dynamically optimize the weight parameters; A Top-N recommendation list is generated based on rating ranking, and a variety of re-ranking strategies are used to balance accuracy and diversity. Implement a real-time feedback mechanism to dynamically update the recommendation model based on user click behavior and dwell time.

[0017] The present invention also provides a paper recommendation system based on heterogeneous information networks, which applies a paper recommendation method based on heterogeneous information networks as described above, including an intelligent feature extraction module, a text understanding module, a heterogeneous network engine, a temporal awareness module, a recommendation generation module, and a feedback learning module; The intelligent feature extraction module is configured to perform multi-dimensional feature extraction and dimensionality reduction, including: a domain dictionary management unit, responsible for building and maintaining a domain-adaptive word segmentation dictionary; a feature weighting unit, which implements composite feature selection based on TF-IDF and information gain; and a vectorization unit, which generates high-quality word vectors based on the Word2Vec model. The text understanding module is configured for deep learning-based text classification, including: a classifier training unit for training and optimizing a Naive Bayes classifier; a classification verification unit for evaluating classification performance through cross-validation; and a semantic understanding unit for extracting deep semantic features of the text. The heterogeneous network engine is configured to construct and query heterogeneous information networks, including: a network construction unit that realizes graph structure modeling of multiple entities and relationships; a network embedding unit that learns node representations using the metapath2vec algorithm; and a similarity calculation unit that realizes multi-granularity semantic similarity calculation. The temporal awareness module is configured for bimodal research interest modeling, including: a time series analysis unit for processing user historical behavior data; an interest modeling unit for realizing vector representations of immediate and ongoing research interests; and a decay function calculation unit for dynamically quantifying the interest evolution process. The recommendation generation module is configured for multi-source information fusion and recommendation list generation, including: a scoring calculation unit that implements a multi-level recommendation scoring function; a weight optimization unit that dynamically adjusts parameters based on reinforcement learning; and a list generation unit that generates personalized Top-N recommendation results. The feedback learning module is configured to continuously optimize the recommendation effect based on user interaction behavior, and includes: a behavior collection unit to collect feedback data, including user clicks, dwell time, and favorites; and a model update unit to realize online learning and incremental updates of the recommendation model.

[0018] The present invention also provides a storage medium storing instructions for a processor to execute, wherein the processor executes the instructions to perform a paper recommendation method based on a heterogeneous information network as described above.

[0019] Compared with the prior art, the technical solution provided by the present invention has at least the following beneficial effects: (1) At the feature extraction level, by introducing a domain dictionary and a composite feature extraction algorithm for feature extraction and text classification, the traditional method relies on fixed word segmentation rules and static feature weighting, which leads to inaccurate identification of domain terms and poor handling of polysemy. It can adaptively capture the deep semantic features of academic texts, providing a high-quality and highly discriminative data foundation for subsequent accurate similarity calculation and interest modeling.

[0020] (2) At the network modeling level, by constructing a heterogeneous information network that integrates multiple entities and complex relationships, the traditional recommendation system has changed the shallow mode of only analyzing the binary relationship between "user-paper". It can systematically represent and store the rich semantic associations between entities such as authors, institutions, journals, and keywords, so that the recommendation decision is not only based on content matching, but also on academic context and community relationship, which significantly improves the rationality of the recommendation.

[0021] (3) At the level of interest modeling, the innovative integration of user historical behavior data and temporal decay function has constructed a dual-modal research interest vector model, which effectively solves the defect of the traditional model that treats user interests as a static single vector and cannot capture its evolution over time. It can accurately distinguish and quantify the user's long-term stable research direction and short-term emerging hot topics, thereby avoiding outdated or deviating from the current interest in the recommended content, and greatly improving the timeliness and personalization of the recommendation.

[0022] (4) At the recommendation generation level, by combining the multi-source relationship information of heterogeneous information networks with the user's temporal perception interest model and adopting a multi-source similarity fusion strategy, an intelligent recommendation mechanism that comprehensively evaluates the relevance of papers from multiple angles and levels is realized.

[0023] (5) At the system architecture level, a complete technical loop is built from data input, complex network construction, dynamic interest modeling to intelligent recommendation output. Continuous optimization is achieved through feedback learning mechanism to ensure effective connection and collaborative optimization between various technical modules. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall steps of the paper recommendation method based on heterogeneous information networks of the present invention; Figure 2 This is a schematic diagram illustrating the specific process of the paper recommendation method based on heterogeneous information networks of the present invention. Figure 3 This is a structural block diagram of the paper recommendation system based on heterogeneous information networks according to the present invention. Detailed Implementation

[0025] The present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. The step numbers in the embodiments of the present invention are only set for ease of explanation and description, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0026] like Figure 1 , Figure 2 As shown, this embodiment provides a paper recommendation method based on heterogeneous information networks, including the following steps: Feature extraction and text classification: Key features are extracted from the paper text data through a domain dictionary and a composite feature extraction algorithm, and the model is trained to accurately classify the paper. Heterogeneous information network construction: Based on the extracted features, a heterogeneous information network is constructed from diverse academic entities and relationships to store complex academic information in a structured form; Temporal-aware interest modeling: Integrating user historical behavior with temporal decay functions to construct an instantaneous and continuous dual-modal research interest vector model; Multi-source similarity calculation: In the heterogeneous information network, multi-granularity similarity is calculated by comprehensively measuring text attribute similarity, network structure similarity, and relationship similarity to fully measure node correlation; Research interest similarity calculation: Based on the bimodal research interest vector model, the similarity between the user's immediate research interests and continuous research interests is calculated separately, and then dynamically weighted and fused. Personalized recommendation generation: The multi-source similarity and the research interest similarity are fused, and reinforcement learning is used to optimize the fusion weights to generate an accurate and diverse list of Top-N paper recommendations.

[0027] Specifically, this invention addresses the problem of paper recommendation in academic social networks. In this embodiment, the paper text data includes basic user information and paper content information. The user's basic information includes name, institution, research team, and research interests, while the paper content information includes title, abstract, keywords, contextual information, citations, publication date, and the conference or journal in which it was published. These features not only effectively represent the important content information of the paper but also reflect the user's research interests and the team's research methods, thereby improving the paper recommendation task in academic social networks.

[0028] In this embodiment, the feature extraction and text classification include the following steps: We construct a domain-adaptive word segmentation dictionary, integrate a keyword database from academic papers, and use the jieba word segmentation algorithm for context-aware word segmentation. Specifically, a four-level word segmentation dictionary system is constructed, comprising a basic dictionary, an academic keyword database, an abbreviation dictionary, and a domain-specific dictionary.

[0029] The TF-IDF algorithm combined with information gain is used for feature weighting and selection. By calculating the word frequency and inverse document frequency of each keyword, the term weight is calculated to determine the importance of each keyword to a text or a corpus, thereby achieving highly discriminative feature extraction. Specifically, the weights of feature words are determined through word frequency statistics, inverse document frequency calculation, and information entropy analysis.

[0030] The CBOW model based on Word2Vec is used to train word vectors, and high-quality distributed text representations are generated by using dynamic window size and negative sampling optimization techniques.

[0031] Text classification is performed using an improved Naive Bayes classifier, incorporating TF-IDF weighting and Bayesian smoothing techniques, and combining k-fold cross-validation to ensure the model's accuracy and generalization ability. Cross-validation is then used to optimize classifier performance. This embodiment significantly enhances the representational ability of text features through a domain-adaptive word segmentation dictionary and a composite feature selection algorithm. Based on the processed paper text data, a heterogeneous information network containing multiple entities and complex relationships is constructed using a heterogeneous network engine.

[0032] In this embodiment, the construction of the heterogeneous information network includes the following steps: Define multiple entity types, including papers, authors, institutions, journals / conferences, and keywords; Define semantic relationship edges, including reference relationships, writing relationships, membership relationships, and co-occurrence relationships; Create nodes and edges based on different types of entities, map structured data into a heterogeneous information network graph structure, and import node attributes and edge weights.

[0033] In this embodiment, the temporal-aware interest modeling includes the following steps: Research interest time series are constructed based on user historical behavior data, which includes paper reading records, download records, citation records, and collection records; A sliding time window mechanism is used to divide research interest periods. Continuous research interests cover long-term research history, with a time window set to T1 years; immediate research interests focus on recent research developments, with a time window set to T2 years; where T1 is greater than T2. ​​In this embodiment, the time window T1 for continuous research interests can be 3-5 years; the time window T2 for immediate research interests can be 0.5-1 years.

[0034] An improved Logistic time decay function is used to dynamically quantify the evolution of interests. Based on the quantification results, continuous research interest vectors and immediate research interest vectors are extracted to construct a bimodal interest fusion model, which is dynamically adjusted based on user behavior feedback. This embodiment innovatively proposes a temporally aware bimodal interest model that accurately captures the dynamic evolution of user research interests.

[0035] In this embodiment, the multi-source similarity calculation includes the following steps: Node representation vectors are learned based on network embedding technology, and the metapath2vec algorithm is used to perform random walk sampling on heterogeneous information networks to generate semantically preserved node embeddings. Calculate multi-granularity semantic similarity between different types of nodes; specifically, it includes the following steps: (1) Text attribute similarity: The semantic matching degree of the title, abstract and keywords is calculated based on the cosine similarity of word vectors; (2) Network structure similarity: The Jaccard similarity of the network topology is used to calculate the overlap of the neighborhood of nodes; (3) Relationship similarity: The PathSim algorithm based on meta-path is used to calculate the strength of multi-hop relationships between nodes.

[0036] A multi-source similarity fusion function is constructed to achieve adaptive weight allocation based on an attention mechanism, dynamically adjusting the weights of each similarity component according to node type and query context. Heterogeneous information networks are utilized to fully mine the complex semantic relationships between academic entities, and high-quality node representations are learned through the metapath2vec algorithm.

[0037] In this embodiment, the calculation of research interest similarity includes the following steps: Obtain the user's real-time research interest vector and ongoing research interest vector respectively; Obtain keywords and time information related to candidate papers; Obtain word vectors of interests and paper keywords using the Word2Vec model; The similarity between candidate papers and immediate research interests and ongoing research interests is calculated separately and assigned weight parameters. Finally, a weighted sum is performed to obtain the research interest similarity.

[0038] In this embodiment, the personalized recommendation generation includes the following steps: Construct a multi-level recommendation scoring function that integrates multi-source similarity and research interest similarity; A Q-learning-based reinforcement learning algorithm is used to dynamically optimize the weight parameters; a 28-dimensional state space and a 6-dimensional action space are designed for adaptive parameter adjustment. A Top-N recommendation list is generated based on rating ranking, and a variety of re-ranking strategies are used to balance accuracy and diversity. Implement a real-time feedback mechanism to dynamically update the recommendation model based on user click behavior and dwell time.

[0039] Reinforcement learning is employed to adaptively optimize parameters, ensuring the recommendation system can adapt to changes in user interests. Continuous optimization is achieved through a feedback learning mechanism, constructing a complete intelligent recommendation closed loop. Real-time collection of user interaction data such as clicks, dwell time, and favorites is used to design an incremental learning algorithm for online updates of the recommendation model. A multi-objective reward function is constructed to comprehensively consider both short-term interaction feedback and long-term user satisfaction.

[0040] Through the above complete technical implementation scheme, the present invention can effectively integrate the structural characteristics of heterogeneous information networks and the temporal awareness of user research interests, and realize intelligent paper recommendation throughout the entire process from feature processing, semantic understanding, network construction, interest modeling to recommendation generation and continuous optimization, which significantly improves the accuracy, practicality and effectiveness of the recommendation.

[0041] The specific implementation steps of this embodiment will be described in detail below: Step 1: Feature extraction and text classification; First, TF-IDF is used to extract features from the text. The TF-IDF algorithm is used to evaluate the importance of a word to a document or a document in a corpus. It is a weighting method based on statistical information. TF represents the frequency of a word in the document set containing that word; it is a normalized representation of the word's occurrences and can effectively measure the weight of the word in the document set. The formula is: ; in, Indicator i Appear in the document j The number of times, ∑ k This indicates that all words in the file j The total number of times it appears in the text.

[0042] IDF stands for Inverse Document Frequency, used to evaluate the general importance of the obtained word segmentation in a document set. Its formula is: ; in, This indicates the total number of files in the document set. Indicates the occurrence of words The number of documents; Indicates the first j One document.

[0043] Therefore, the formula for TF-IDF is as follows: ; Based on the TF-IDF algorithm formula, text preprocessing, word weight calculation, and keyword extraction were performed. It was found that the weight value of a word increases with the number of times it appears in the document, and decreases with the number of times it appears in the corpus file.

[0044] Furthermore, after using TF-IDF for text feature extraction, this embodiment of the invention employs the Information Gain method to extract keywords from the paper text that can better distinguish the paper type. Specifically, by calculating the entropy value of a word in a document set, assuming the range of values ​​for a random variable in an entity is: X = ｛x 1 ,x 2 ,…,x n ｝ And the probability set corresponding to the value of each variable is: P={p 1 ,p 2 ,…,p n } At this point, the formula for calculating the information entropy of the variable is: ; Where C represents the paper category; Indicates category The probability of appearing in a document set.

[0045] Furthermore, assuming that feature word T appears in or does not appear in the text of paper of type C, the formula for calculating the conditional entropy of type C is as follows: ; in, This indicates that the document set contains characteristic words. The probability, Indicates the presence of characteristic words The document belongs to the category The conditional probability. This indicates that the document set does not contain the characteristic words. The probability of.

[0046] For example, when the information entropy of feature word A is smaller than that of feature word B, the uncertainty of feature word A in determining whether a paper belongs to type C is lower than that of feature word B. Therefore, feature word A has a better distinguishing effect than feature word B when differentiating whether a paper belongs to type C. Information gain represents the amount of information obtained after eliminating uncertainty and can assess the importance of a word within the entire corpus.

[0047] In summary, the formula for calculating information gain is as follows: .

[0048] Step 2: Construction of heterogeneous information network; In some embodiments, based on the text data after feature extraction, the entities and relationships between entities are respectively used as nodes and edges of a graph to construct a heterogeneous information network for storing and retrieving text data.

[0049] Specifically, this embodiment of the invention assumes a set containing n papers: P={ P 1 , P 2 , … , P n } Each paper in the set is represented by m different types of features. F={F 1 , F 2 , ... , F m } Each feature F i It consists of a set of features, i.e. F i ={ f 1i , f 2i , ..., f ni } .

[0050] For example, a collection of multiple entity types: ; Semantic relation edge set: ; Different types of entity data are represented as interconnected nodes, and the relationships between entities are represented as edges, thus constructing a heterogeneous information network. ; in, V It is a set of nodes. E It is a set of edges. It is a node type mapping function. It is an edge-type mapping function.

[0051] A heterogeneous information network is a vast network graph with rich semantic relationships, consisting of entities and the relationships between them. After extracting features using TF-IDF and information gain methods, basic information such as author, title, abstract, keywords, citations, publication date, and conference or journal is obtained. Different entities contain different relationships, including: citation and reference relationships between papers, relationships between users and papers, relationships between papers and paper types, relationships between users and research interests, relationships between users, and relationships between paper types, such as a multi-entity type set. E ={paper P, author A, mechanismI , Journals / Conferences J, Keywords K By representing different types of entity data as interconnected nodes and the relationships between entities as edges, a heterogeneous information network is constructed, enabling more efficient storage and retrieval of thesis text data.

[0052] In this embodiment of the invention, using the data of nodes and edges in a heterogeneous information network, Word2Vec technology is employed to train word vectors on the data in the nodes (including title, abstract, keywords, citations, publication time, and journal or conference), thereby obtaining the similarity between any two data nodes in the heterogeneous information network. The specific formula for the similarity between any two keyword data nodes is as follows: ; ; in , , Each represents any pair of keywords. and The similarity score between the title, abstract, and keywords can be obtained by calculating the cosine similarity of their corresponding word vectors. ， , Let represent the weights of the title, abstract, and keywords in the similarity calculation within a node of a heterogeneous information network, respectively, and satisfy the following conditions: This invention assigns different weight ratios to different types of papers, where when... , , It has the best effect.

[0053] Furthermore, the similarity between two papers can be measured based on the similarity between the keywords to be recommended. Additionally, other similar papers exist within the paper and user entities, collectively constructing the paper similarity model. In this paper, the similarity of recommended words is primarily obtained by training with word2vec to obtain the similarity between the keywords to be recommended, using the following formula: ; in, This represents the similarity between the recommended keyword pairs in the titles of any two papers. This represents the similarity between the keyword pairs to be recommended in any two papers. This represents the similarity between pairs of keywords to be recommended in the abstracts of any two papers. ， , Let represent the weights of each attribute in the similarity calculation, and satisfy . This invention assigns different weight ratios to different types of papers, where when... , , It has the best effect.

[0054] Step 3: Temporally Aware Bimodal Research Interest Modeling; Temporal attributes are introduced into users' research interests, which are then divided into immediate research interests and ongoing research interests. The similarity of users' research interests is calculated using the Logistic time decay function.

[0055] Specifically, in this embodiment of the invention, given the large intervals between publication dates in the papers and the limited decay range, a decay factor of 0-0.5 is primarily selected as the main coefficient for time decay. Therefore, this embodiment employs the Logistic time decay function to calculate the similarity of research interests, incorporating the time decay factor to measure research interest and thus assessing the changing process of user research interests. The formula for the traditional Logistic time decay function is as follows: ; in, This indicates the historical time points when a user reads, saves, or otherwise interacts with the content. This indicates the current time point when the recommended task was executed; This represents the attenuation coefficient.

[0056] Specifically, the following steps are included: 3.1. Users' research interests are divided into immediate research interests and ongoing research interests. First, a time series of research interests is constructed based on users' multidimensional historical behavioral data. The result is: ; Each time slice contains: ; Behavioral weight learning: ;in, These represent user features, contextual features, and temporal features, respectively. The weights of each behavior are obtained by fusing them using a multilayer perceptron (MLP) and then normalizing them using Softmax. .

[0057] 3.2. Using an adaptive time window mechanism to learn continuous research interests and immediate research interests: (1) Continuing research interests: (Year); in, For user active days, This represents the total number of days.

[0058] (2) Immediate research interests: (Year); in, It is the sigmoid function; For learnable parameter vectors, This is the feature vector of recent activities.

[0059] 3.3 Using an improved temporal decay function: Logistic time decay function: ; in, As an adaptive attenuation factor, For the initial time of interest, This is the time offset. This is the interest retention bias term.

[0060] 3.4 Continuous Research Interest Modeling: Basic persistent interest vector: ;in, This represents the i-th paper on user interaction. It measures the persistence of users' attention to a topic by comprehensively calculating a weighted score obtained by combining the duration of users' interaction behavior on a specific topic, the frequency of historical interactions, and the standard deviation (regularity) of the time intervals. This score is used to measure the persistence and reliability of the research interest.

[0061] Evolutionary interest vector: ; in, Indicates the first j The feature gradient of each interest topic on the time axis This represents the partial derivative of the decay function with respect to time, capturing the rate at which the weights change over time. This represents a trend detection operator.

[0062] Overall ongoing research interests: ;in, 。

[0063] 3.5 Real-time Research Interest Modeling: Sudden Interest Detection: Detecting Sudden Changes in Interest Using the CUSUM Algorithm ; in, express t The intensity of user interaction observed at all times. This indicates the average level of historical behavior. This represents the allowable deviation margin, used to filter random fluctuation noise. Let... h As the detection threshold, when > h At that time, it was determined that the user had developed a "sudden interest".

[0064] 3.6. Dual-modal interest fusion: Modal importance assessment: ; ; in, and These are learnable parameters.

[0065] Final expression of interest: ; Step 4: Calculate multi-source similarity; 4.1 Text attribute similarity; Multi-granular text similarity: ;in, .

[0066] in, , , These are the semantic vector representations of the title, abstract, and keywords of text u, respectively. , , The semantic vector representations of the title, abstract, and keywords of text v are respectively obtained through models such as Word2Vec or BERT.

[0067] 4.2 Structural similarity; Jaccard similarity: ;in, Represents the set of a node's direct neighbors.

[0068] 4.3 Relationship Similarity; PathSim similarity: ; in, For a predefined set of symmetric metapaths, This represents a metapath instance from node u to v.

[0069] 4.4 Multi-source similarity fusion; Attention weight calculation: ; in, This corresponds to three well-calculated multi-source similarities: text attribute similarity, network structure similarity, and relation similarity. The feature representation of the s-th similarity type; and These are learnable parameters.

[0070] Final network similarity: .

[0071] Step 5: Calculate the similarity of research interests; 5.1. Continuous Interest Similarity: ; in, This represents the time alignment factor, used to measure the degree of consistency in the evolution of two users' interests over time.

[0072] 5.2 Instant Interest Similarity: ; This represents the context similarity factor, used to measure the consistency of the user's current research environment.

[0073] 5.3 Overall Interest Similarity: ; in, It is dynamically adjusted based on user behavior feedback.

[0074] Step Six: Personalized Recommendation Generation; 6.1 Multi-level scoring function; Construct a comprehensive scoring function: ; in, Novelty factor based on paper publication time and user engagement history; : A measure of the diversity of topic distribution in the recommendation list. Let be the weight parameters, and satisfy ? .

[0075] 6.2 Enhanced learning optimization; The Q-learning algorithm is used to dynamically optimize the weight parameters: ; Among them, state Includes user history and current context, actions Adjusting the weighting parameters, rewards Calculated based on user click-through rate and dwell time.

[0076] 6.3 Generation of Top-N Recommendation List; Personalized recommendations are generated based on rating ranking: ; in, It is a collection of candidate papers. It is the list diversity score.

[0077] like Figure 3 As shown, this embodiment also provides a paper recommendation system based on heterogeneous information networks, which applies a paper recommendation method based on heterogeneous information networks as described above, including an intelligent feature extraction module, a text understanding module, a heterogeneous network engine, a temporal awareness module, a recommendation generation module, and a feedback learning module; The intelligent feature extraction module is configured to perform multi-dimensional feature extraction and dimensionality reduction, including: a domain dictionary management unit, responsible for building and maintaining a domain-adaptive word segmentation dictionary; a feature weighting unit, which implements composite feature selection based on TF-IDF and information gain; and a vectorization unit, which generates high-quality word vectors based on the Word2Vec model. The text understanding module is configured for deep learning-based text classification, including: a classifier training unit for training and optimizing a Naive Bayes classifier; a classification verification unit for evaluating classification performance through cross-validation; and a semantic understanding unit for extracting deep semantic features of the text. The heterogeneous network engine is configured to construct and query heterogeneous information networks, including: a network construction unit that realizes graph structure modeling of multiple entities and relationships; a network embedding unit that learns node representations using the metapath2vec algorithm; and a similarity calculation unit that realizes multi-granularity semantic similarity calculation. The temporal awareness module is configured for bimodal research interest modeling, including: a time series analysis unit for processing user historical behavior data; an interest modeling unit for realizing vector representations of immediate and ongoing research interests; and a decay function calculation unit for dynamically quantifying the interest evolution process. The recommendation generation module is configured for multi-source information fusion and recommendation list generation, including: a scoring calculation unit that implements a multi-level recommendation scoring function; a weight optimization unit that dynamically adjusts parameters based on reinforcement learning; and a list generation unit that generates personalized Top-N recommendation results. The feedback learning module is configured to continuously optimize the recommendation effect based on user interaction behavior, and includes: a behavior collection unit to collect feedback data, including user clicks, dwell time, and favorites; and a model update unit to realize online learning and incremental updates of the recommendation model.

[0078] This embodiment also provides a storage medium storing instructions for a processor to execute, wherein the processor executes the paper recommendation method based on a heterogeneous information network as described above.

[0079] Compared to existing technologies, this invention constructs a heterogeneous information network for storing and retrieving text data, introduces temporal attributes, and finally obtains paper recommendation results through comprehensive measurement.

[0080] Compared to existing technologies, this invention uses temporal awareness to perceive the academic characteristics of users and papers; inputs the data into a model, dynamically weights the data, and then outputs the results to recommend papers. Four temporal awareness functions are used to compare the recommendation results, and multiple evaluation metrics are used to evaluate our model. This method effectively considers the characteristics of papers and changes in scholars' interests, solves problems such as cold start and sparsity of data, filters out outdated papers, and effectively controls the timeliness of data, thereby greatly improving the accuracy of paper recommendations.

[0081] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A paper recommendation method based on heterogeneous information networks, characterized in that, Includes the following steps: Feature extraction and text classification: Key features are extracted from the paper text data using a domain dictionary and a composite feature extraction algorithm, and a model is trained to classify the papers. Heterogeneous information network construction: Based on the extracted features, a heterogeneous information network is constructed from diverse academic entities and relationships to store complex academic information in a structured form; Temporal-aware interest modeling: Integrating user historical behavior with temporal decay functions to construct an instantaneous and continuous dual-modal research interest vector model; Multi-source similarity calculation: In the heterogeneous information network, multi-granularity similarity is calculated by comprehensively measuring text attribute similarity, network structure similarity, and relationship similarity to fully measure node correlation; Research interest similarity calculation: Based on the bimodal research interest vector model, the similarity between the user's immediate research interests and continuous research interests is calculated separately, and then dynamically weighted and fused. Personalized recommendation generation: The multi-source similarity and the research interest similarity are fused, and reinforcement learning is used to optimize the fusion weights to generate a Top-N paper recommendation list.

2. The paper recommendation method based on heterogeneous information networks according to claim 1, characterized in that, The feature extraction and text classification include the following steps: We construct a domain-adaptive word segmentation dictionary, integrate a keyword database from academic papers, and use the jieba word segmentation algorithm for context-aware word segmentation. The TF-IDF algorithm combined with information gain is used for feature weighting and selection. By calculating the word frequency and inverse document frequency of each keyword, the term weight is calculated to determine the importance of each keyword to a text or a corpus, thereby achieving highly discriminative feature extraction. The CBOW model based on Word2Vec is used to train word vectors, and a Naive Bayes classifier is used for text classification. Cross-validation is then used to optimize the classifier's performance.

3. The paper recommendation method based on heterogeneous information networks according to claim 1, characterized in that, The construction of the heterogeneous information network includes the following steps: Define multiple entity types, including papers, authors, institutions, journals / conferences, and keywords; Define semantic relationship edges, including reference relationships, writing relationships, membership relationships, and co-occurrence relationships; Create nodes and edges based on different types of entities, map structured data into a heterogeneous information network graph structure, and import node attributes and edge weights.

4. The paper recommendation method based on heterogeneous information networks according to claim 1, characterized in that, The temporal-aware interest modeling includes the following steps: Research interest time series are constructed based on user historical behavior data, which includes paper reading records, download records, citation records, and collection records; A sliding time window mechanism was used to divide the research interest periods. The continuous research interests covered the long-term research history, and the time window was set to T1 years. The immediate research interests focused on recent research dynamics, and the time window was set to T2 years. T1 is longer than T2. Interest evolution is dynamically quantified using an improved Logistic time decay function; Based on the quantitative results, continuous research interest vectors and immediate research interest vectors are extracted to construct a bimodal interest fusion model, which is then dynamically adjusted based on user behavior feedback.

5. The paper recommendation method based on heterogeneous information networks according to claim 1, characterized in that, The multi-source similarity calculation includes the following steps: Node representation vectors are learned based on network embedding technology, and the metapath2vec algorithm is used to perform random walk sampling on heterogeneous information networks to generate semantically preserved node embeddings. Calculate multi-granularity semantic similarity between different types of nodes; A multi-source similarity fusion function is constructed to achieve adaptive weight allocation based on the attention mechanism, dynamically adjusting the weights of each similarity component according to the node type and query context.

6. The paper recommendation method based on heterogeneous information networks according to claim 5, characterized in that, Calculating multi-granularity semantic similarity between different types of nodes includes the following steps: (1) Text attribute similarity: The semantic matching degree of the title, abstract and keywords is calculated based on the cosine similarity of word vectors; (2) Network structure similarity: The Jaccard similarity of the network topology is used to calculate the overlap of the neighborhood of nodes; (3) Relationship similarity: The PathSim algorithm based on meta-path is used to calculate the strength of multi-hop relationships between nodes.

7. A paper recommendation method based on heterogeneous information networks according to claim 1 or 4, characterized in that, The calculation of research interest similarity includes the following steps: Obtain the user's real-time research interest vector and ongoing research interest vector respectively; Obtain keywords and time information related to candidate papers; Obtain word vectors of interests and paper keywords using the Word2Vec model; The similarity between candidate papers and immediate research interests and ongoing research interests is calculated separately and assigned weight parameters. Finally, a weighted sum is performed to obtain the research interest similarity.

8. The paper recommendation method based on heterogeneous information networks according to claim 1, characterized in that, The personalized recommendation generation includes the following steps: Construct a multi-level recommendation scoring function that integrates multi-source similarity and research interest similarity; A Q-learning-based reinforcement learning algorithm is used to dynamically optimize the weight parameters; A Top-N recommendation list is generated based on rating ranking, and a variety of re-ranking strategies are used to balance accuracy and diversity. Implement a real-time feedback mechanism to dynamically update the recommendation model based on user click behavior and dwell time.

9. A paper recommendation system based on heterogeneous information networks, employing a paper recommendation method based on heterogeneous information networks as described in any one of claims 1-8, comprising an intelligent feature extraction module, a text understanding module, a heterogeneous network engine, a temporal awareness module, a recommendation generation module, and a feedback learning module; characterized in that, The intelligent feature extraction module is configured to perform multi-dimensional feature extraction and dimensionality reduction, including: a domain dictionary management unit, responsible for building and maintaining a domain-adaptive word segmentation dictionary; a feature weighting unit, which implements composite feature selection based on TF-IDF and information gain; and a vectorization unit, which generates high-quality word vectors based on the Word2Vec model. The text understanding module is configured for deep learning-based text classification, including: a classifier training unit for training and optimizing a Naive Bayes classifier; a classification verification unit for evaluating classification performance through cross-validation; and a semantic understanding unit for extracting deep semantic features of the text. The heterogeneous network engine is configured to construct and query heterogeneous information networks, including: a network construction unit that realizes graph structure modeling of multiple entities and relationships; a network embedding unit that learns node representations using the metapath2vec algorithm; and a similarity calculation unit that realizes multi-granularity semantic similarity calculation. The temporal awareness module is configured for bimodal research interest modeling, including: a time series analysis unit for processing user historical behavior data; an interest modeling unit for realizing vector representations of immediate and ongoing research interests; and a decay function calculation unit for dynamically quantifying the interest evolution process. The recommendation generation module is configured for multi-source information fusion and recommendation list generation, including: a scoring calculation unit that implements a multi-level recommendation scoring function; a weight optimization unit that dynamically adjusts parameters based on reinforcement learning; and a list generation unit that generates personalized Top-N recommendation results. The feedback learning module is configured to continuously optimize the recommendation effect based on user interaction behavior, and includes: a behavior collection unit to collect feedback data, including user clicks, dwell time, and favorites; and a model update unit to realize online learning and incremental updates of the recommendation model.

10. A storage medium storing instructions for execution by a processor, characterized in that, When the processor executes instructions, it performs a paper recommendation method based on heterogeneous information networks as described in any one of claims 1-8.

Images