Scientific and technological information recommendation method based on multi-dimensional weight fusion and grouping collaborative filtering
Patent Information
- Application Number
- CN202610607238.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-06
- Publication Date
- 2026-08-11
AI Technical Summary
[0003]然而，当前主流的资讯推送平台普遍难以满足专家用户高度专业化、定制化的信息需求
本发明针对专业领域用户高度定制化的信息需求与现有推荐技术适应性不足的矛盾，本方案提出了一种多维度权重融合与分组协同过滤的科创资讯推荐方法及系统。首先，通过多维度权重融合用户历史交互资讯，构建精准用户内容表征：基于科创资讯内容与用户画像的相似度加权突出主题相关性，依据资讯聚类分组（见下）的组别加权强化主导兴趣模式（同组交互资讯越多权重越高），结合时间衰减加权确保近期行为主导表征更新，利用多类型行为信号（如点击、收藏）加权捕捉用户内容偏好。其次，引入用户组-资讯组协同表征学习：根据资讯内容表征对资讯聚类分组，根据用户与聚类后资讯组的交互行为对用户聚类分组，构建“用户组-资讯组”交互矩阵，分别学习用户和资讯的协同表征，缓解专业场景下用户兴趣高度分化、交互数据稀疏的难题。最终，分别将用户与资讯的内容表征与协同表征融合，获取各自的联合表征，通过匹配双方的联合表征实现推荐。本方案同步设计了对应的推荐系统架构以支持快速应用。该方法提升了专业科创资讯推荐的精准性，突破信息茧房限制，同时增强系统对稀疏数据和时效性需求的鲁棒性，为高端用户提供高效知识获取通道。
Smart Images
Figure CN122548032A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of technology, and in particular to a method for recommending science and technology innovation information based on multi-dimensional weight fusion and grouped collaborative filtering. Background Technology
[0002] Experts and researchers in their respective fields are the core driving force behind national scientific and technological innovation and industrial upgrading. The efficiency and quality of their acquisition of cutting-edge knowledge directly affect the progress of technological breakthroughs. Science and technology information services for these high-end users are a key infrastructure supporting their efficient scientific research, understanding of industry trends, and enhancement of professional capabilities.
[0003] However, current mainstream news delivery platforms generally fail to meet the highly specialized and customized information needs of expert users. General recommendation algorithms often employ a broad content distribution strategy, pushing content with wide coverage but deviating from the user's core research area, resulting in a large amount of irrelevant or low-value information mixed in. Expert users are forced to expend considerable valuable time on secondary filtering, leading to low information acquisition efficiency and generally low satisfaction. This not only severely weakens the platform's appeal and stickiness to professional users but also restricts the platform's ability to realize its core value in vertical professional fields.
[0004] Furthermore, the limitations of existing technological frameworks are particularly evident in specialized fields. While single-content recommendation models can analyze the semantic features of information and ensure the relevance of recommended topics, they struggle to effectively capture users' potential, cross-domain, or dynamically evolving deep interests. Collaborative filtering, on the other hand, is difficult to implement in scenarios with high-frequency information updates due to data cold start and real-time limitations. More importantly, the niche nature of expert user groups results in an extremely sparse user-project interaction matrix, making it difficult for traditional collaborative algorithms to uncover effective cross-entity relationships.
[0005] To this end, we propose a hybrid recommendation method and system that integrates multi-dimensional weighted representation learning and group collaborative optimization to achieve accurate delivery and value release of scientific research information. Summary of the Invention
[0006] This invention overcomes the shortcomings of existing technologies and provides a method for recommending science and technology innovation information based on multi-dimensional weight fusion and grouped collaborative filtering.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention provides a method for recommending science and technology innovation information based on multi-dimensional weight fusion and grouped collaborative filtering, including the following steps: Acquire basic information and professional attribute data, and construct static user profile representations and information content representations based on the basic information and professional attribute data; The static user profile representation and information content representation are obtained by extracting the title, summary and keywords of the information using a pre-trained large language model; Information content representations are clustered to form information groups, and users are clustered to form user groups based on user interaction behavior with information groups; Based on historical interaction data between users and information, a user group-information group interaction matrix is constructed, and collaborative representations of users and information are learned respectively. By integrating content representations and collaborative representations from the user side and the information side respectively, a joint representation of users and information containing comprehensive information is generated; Based on user profile representation, information joint representation, and user joint representation, the matching degree between target information and users is calculated, candidate information is sorted, and the information with the highest matching degree is accurately pushed to the target user.
[0008] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering, basic information and professional attribute data are obtained, and static user profile representation and information content representation are constructed based on the basic information and professional attribute data, specifically as follows: Upload basic information and professional attribute data, and extract professional attributes from user registration information and historical behavior data based on the basic information and professional attribute data; Numerical features are extracted from the professional attributes in the user registration information and historical behavior data, and the numerical features are converted into discrete features. One-hot encoding is used for features with fewer than a preset threshold. For discrete features with no fewer than a preset threshold of categories, the word embedding model is used to convert them into dense vectors and average pooling is performed. One-hot encoding and dense vectors are then concatenated, and multilayer perceptron encoding is used to obtain user profile representations. By using prompt templates, the title, keywords, and summary of the information are extracted using a pre-trained large language model, and then the text representation of the information is obtained. After the large model returns the extracted information, a pre-trained embedding model is used to embed this information to obtain the content representation of the information.
[0009] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering, the information content representation is clustered to form information groups, and users are clustered to form user groups based on the interaction behavior between users and information groups. Specifically: Input all user sets All information collection Information content representation set User interaction behavior sequence ,in The length of the action sequence. Information data representing behavior lengths from 1 to u is clustered using K-means based on its content representation. Behavioral vectors are constructed based on the frequency of user interactions with information groups, and spectral clustering is performed on users. For new ungrouped information, it is assigned to the group with the highest cosine similarity between its content representation and the class center vector. For new, ungrouped users, they are temporarily assigned to groups based on their user profiles. Once their interaction sequence reaches a preset length, they are then assigned to groups based on their behavior vectors.
[0010] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering, a user group-information group interaction matrix is constructed based on the historical interaction data between users and information, and the collaborative representations of users and information are learned respectively, specifically: Construct a group interaction matrix, where the group interaction matrix The following relationship must be satisfied: , in, Represents the group interaction matrix The Middle line, number The element value of the column represents the first element. The user group and the first The intensity of interaction between information groups; Indicates the first The first group Each user group contains several users with similar characteristics; No. The first group Each information group contains several pieces of information with similar characteristics; Indicates user Information The interaction weights are a weighted sum of multiple behaviors, among which... ; Learning collaborative representations through singular value decomposition satisfies the following relationship: , in, For user collaborative representation, For information collaboration representation; Use a circular queue to store a set of matrices. These matrices record the most recent consecutive period of time. User interaction behavior information, where each matrix records non-overlapping consecutive... The interaction information of the day, when the interval between the time of a matrix and the current time exceeds Then clear the matrix data and use it to record the next... Interactive information between the heavens; When performing matrix factorization to learn collaborative representations, the following relation is satisfied: , Represents the interaction matrix at time j. One of the data points.
[0011] Furthermore, the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering also includes: Based on user interaction behavior sequence For each piece of information in the interaction, four sets of weights are calculated and then fused, including similarity weighting, group weighting, time series weighting, and behavioral signal weighting. In similarity-weighted calculations, the representation of information content is calculated. User profile representation Cosine similarity; In group weighting, the frequency of occurrence of each group of information in the user interaction information sequence is counted, and the group weight is scaled logarithmically by the frequency. In time series weighting, an exponential decay function is used to weight the time series, and finally the weighted results are fused.
[0012] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering, the content representations and collaborative representations from the user side and the information side are fused respectively to generate a joint representation of users and information containing comprehensive information, specifically including: User content representation is obtained based on the content representation of weighted fused interactive information, where the following relationship is satisfied: , in, This represents the weighted result in the weighted fusion process. This represents the content features of the i-th user.
[0013] Furthermore, the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering also includes: pass Functions generate user joint representations and information joint representation where the following relationship is satisfied: , , in, is the joint representation vector of the user, which is the final output of the model; w is a learnable weight matrix used to perform linear transformation on the concatenated features; To represent the user's content characteristics; To represent the collaborative characteristics of users. To characterize the content features of information; This represents the collaborative characteristics of information.
[0014] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering, the matching degree between target information and users is calculated according to user profile representation, information joint representation, and user joint representation, specifically as follows: Calculate the matching degree between users and candidate information. where the following relationship is satisfied: , , , Calculate the cosine similarity of content representations Dot product similarity with co-representations By absorbing the advantages of late-stage fusion, a first-order interaction is explicitly constructed to provide a correlation signal between the two. User preference characteristics; By retaining user profile representations and joint representations from both parties, and leveraging the advantages of early fusion, the model can directly acquire the complete semantic information learned by the user tower and the information tower respectively, thus helping it capture fine-grained relationships that cannot be expressed by the inner product. Output Joint representation with each candidate information Matching score .
[0015] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and group collaborative filtering, candidate information is ranked, and the information with the highest matching degree is accurately pushed to the target user, specifically as follows: For all candidate information, a joint representation is formed, based on the joint representation of candidate information and... Matching score Sort in descending order; The matching score of the information is adjusted based on its popularity and business needs. The top-K candidate information with the highest matching scores are selected and returned to be displayed to the user.
[0016] This invention addresses the shortcomings of the prior art and has the following beneficial effects: This invention addresses the contradiction between the highly customized information needs of users in professional fields and the insufficient adaptability of existing recommendation technologies. This solution proposes a multi-dimensional weighted fusion and grouped collaborative filtering method and system for recommending science and technology innovation information. First, it constructs accurate user content representations by fusing historical user interaction information with multi-dimensional weights: weighted similarity between science and technology innovation information content and user profiles highlights topic relevance; group-based weighted reinforcement of dominant interest patterns (higher weight for more interactive information within the same group) is used according to information clustering (see below); time decay weighting ensures recent behavior dominates the representation update; and multiple types of behavioral signals (such as clicks and favorites) are used to capture user content preferences. Second, it introduces user-information group collaborative representation learning: information is clustered based on content representation, and users are clustered based on their interaction behavior with the clustered information groups, constructing a "user-information group" interaction matrix. The collaborative representations of users and information are learned separately, alleviating the challenges of highly differentiated user interests and sparse interaction data in professional scenarios. Finally, the content representations and collaborative representations of users and information are fused separately to obtain their respective joint representations. Recommendations are achieved by matching the joint representations of both parties. This solution includes a corresponding recommendation system architecture designed to support rapid application. This method improves the accuracy of professional science and technology information recommendations, overcomes information silo limitations, and enhances the system's robustness to sparse data and timeliness requirements, providing high-end users with an efficient knowledge acquisition channel. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.
[0018] Figure 1 The overall flowchart of the science and technology innovation information recommendation method based on multi-dimensional weight fusion and group collaborative filtering is shown. Detailed Implementation
[0019] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.
[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0021] like Figure 1 As shown, this invention provides a method for recommending science and technology innovation information based on multi-dimensional weight fusion and grouped collaborative filtering, including the following steps: Acquire basic information and professional attribute data, and construct static user profile representations and information content representations based on the basic information and professional attribute data; The static user profile representation and information content representation are obtained by extracting the title, summary and keywords of the information using a pre-trained large language model; Information content representations are clustered to form information groups, and users are clustered to form user groups based on user interaction behavior with information groups; Based on historical interaction data between users and information, a user group-information group interaction matrix is constructed, and collaborative representations of users and information are learned respectively. By integrating content representations and collaborative representations from the user side and the information side respectively, a joint representation of users and information containing comprehensive information is generated; Based on user profile representation, information joint representation, and user joint representation, the matching degree between target information and users is calculated, candidate information is sorted, and the information with the highest matching degree is accurately pushed to the target user.
[0022] It should be noted that this solution also includes a corresponding recommendation system architecture to support rapid application. This method improves the accuracy of professional science and technology innovation information recommendations, breaks through the limitations of information cocoons, and enhances the system's robustness to sparse data and timeliness requirements, providing high-end users with an efficient knowledge acquisition channel.
[0023] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering, basic information and professional attribute data are obtained, and static user profile representation and information content representation are constructed based on the basic information and professional attribute data, specifically as follows: Upload basic information and professional attribute data, and extract professional attributes from user registration information and historical behavior data based on the basic information and professional attribute data; Numerical features are extracted from the professional attributes in the user registration information and historical behavior data, and the numerical features are converted into discrete features. One-hot encoding is used for features with fewer than a preset threshold. For discrete features with no fewer than a preset threshold of categories, the word embedding model is used to convert them into dense vectors and average pooling is performed. One-hot encoding and dense vectors are then concatenated, and multilayer perceptron encoding is used to obtain user profile representations. By using prompt templates, the title, keywords, and summary of the information are extracted using a pre-trained large language model, and then the text representation of the information is obtained. After the large model returns the extracted information, a pre-trained embedding model is used to embed this information to obtain the content representation of the information.
[0024] During pre-training, the user set is divided into two opposing subsets in an 8:2 ratio, with their interaction behavior data serving as the training and validation datasets, respectively. Training is then performed using the training data, with the user interaction behavior data input into the training module.
[0025] It should be noted that professional attributes are extracted from user registration information and historical behavior data, including age, gender, education, job title, research field, technical direction, and attention tags (extracted from user behavior). Numerical features such as age are binned and converted into discrete features. For discrete features with fewer categories, one-hot encoding is used; otherwise, for such features, they are converted into dense vectors using a word embedding model (such as word2vec) and then average pooled. Finally, the one-hot encoding and dense vectors are concatenated, and a multilayer perceptron (MLP) encoding is used to obtain the user profile representation. Using prompt templates, a pre-trained large language model is used to extract the title, keywords, and summary of the information, and then further obtain the textual representation of the information. For example, the large model is asked, "There is a science and technology innovation news item. Please extract the title, keywords, and a summary of the news content. The number of keywords is limited to ten, and the summary word count is limited to 200 words. Please reply in the format of 'Title: ... \nKeywords: ... \nSummary: ...'". After the large model returns the extracted information, a pre-trained embedding model (such as BERT) is used to embed this information to obtain the information content representation. .
[0026] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering, the information content representation is clustered to form information groups, and users are clustered to form user groups based on the interaction behavior between users and information groups. Specifically: Input all user sets All information collection Information content representation set User interaction behavior sequence ,in The length of the action sequence. This represents information data with a behavior length of 1 to u. K-means clustering is performed on the information based on its content representation.
[0027] Behavioral vectors are constructed based on the frequency of user interactions with information groups, and spectral clustering is performed on users. For new ungrouped information, it is assigned to the group with the highest cosine similarity between its content representation and the class center vector. For new, ungrouped users, they are temporarily assigned to groups based on their user profiles. Once their interaction sequence reaches a preset length, they are then assigned to groups based on their behavior vectors.
[0028] It should be noted that information clustering and grouping: information is clustered using K-means based on its content representation, satisfying the following relationship: , Among them, For the number of information groups, For the first The center vectors of the clusters, This is a collection of knowledge content clustering results, containing several knowledge content clusters. This represents the nth knowledge content cluster, composed of knowledge content that meets specific similarity conditions, i.e., it is formed by calculating the information content representation. with each cluster center Cosine similarity, information Assigned to the group with the highest similarity In the context of knowledge content, v represents the set of knowledge content, and V represents a single piece of knowledge content (such as a document, video, etc.). Information The content feature vector.
[0029] Behavioral vectors are constructed based on the frequency of user interactions with newsgroups. Based on this, further spectral clustering is performed on users to satisfy the following relationship: , in Number of user groups For the first The centroid of each cluster, Let be the user behavior vector, where Indicates user With the Information Group The number of interactions; The set of user clustering results, where Indicates the first User clusters. The optimization objective of user clustering is to minimize the vector of all user behaviors. With its cluster centroid The sum of distances. M is the number of user groups, i.e., the number of clusters for user clustering; : No. The centroid vector of a user cluster is the mean of all user behavior vectors in that group.
[0030] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering, a user group-information group interaction matrix is constructed based on the historical interaction data between users and information, and the collaborative representations of users and information are learned respectively, specifically: Construct a group interaction matrix, where the group interaction matrix The following relationship must be satisfied: , in, Represents the group interaction matrix The Middle line, number The element value of the column represents the first element. The user group and the first The intensity of interaction between information groups; Indicates the first The first group Each user group contains several users with similar characteristics; No. The first group Each information group contains several pieces of information with similar characteristics; Indicates user Information The interaction weights are a weighted sum of multiple behaviors, among which... ; Learning collaborative representations through singular value decomposition satisfies the following relationship: , in, For user collaborative representation, For information collaboration representation; Use a circular queue to store a set of matrices. These matrices record the most recent consecutive period of time. User interaction behavior information, where each matrix records non-overlapping consecutive... The interaction information of the day, when the interval between the time of a matrix and the current time exceeds Then clear the matrix data and use it to record the next... Interactive information between the heavens; When performing matrix factorization to learn collaborative representations, the following relation is satisfied: , Represents the interaction matrix at time j. One of the data points.
[0031] Furthermore, the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering also includes: Based on user interaction behavior sequence For each piece of information in the interaction, four sets of weights are calculated and then fused, including similarity weighting, group weighting, time series weighting, and behavioral signal weighting. In similarity-weighted calculations, the representation of information content is calculated. User profile representation Cosine similarity; In group weighting, the frequency of occurrence of each group of information in the user interaction information sequence is counted, and the group weight is scaled logarithmically by the frequency. In time series weighting, an exponential decay function is used to weight the time series, and finally the weighted results are weighted and fused.
[0032] It should be noted that this is based on user interaction behavior sequences. Each piece of information in the interaction is weighted separately into four groups and then fused.
[0033] 1. Similarity-weighted calculation: Calculating the representation of information content User profile representation cosine similarity : , 2. Group Weighting: This involves calculating the frequency of each group of information in a user interaction information sequence, scaling the group weights logarithmically by the frequency. The group weighting result... The following relationship must be satisfied: , Time series weighted results The calculation uses an exponential decay function and satisfies the following relationship: , in, The length of the action sequence. For exponential parameters, Time-weighted results Indicates the length of the action sequence At what moment, The time at which the length of the behavior sequence is indicated.
[0034] 4. Weighted results of behavioral signals The following relationship must be satisfied: , Weight fusion: , in These are the parameters that the model needs to learn.
[0035] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering, the content representations and collaborative representations from the user side and the information side are fused respectively to generate a joint representation of users and information containing comprehensive information, specifically including: Obtain user content representation based on the content representation of weighted fused interactive information. where the following relationship is satisfied: , in, This represents the weighted result in the weighted fusion process. This represents the content features of the i-th user.
[0036] Furthermore, the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering also includes: pass Functions generate user joint representations and information joint representation where the following relationship is satisfied: , , in, is the joint representation vector of the user, which is the final output of the model; w is a learnable weight matrix used to perform linear transformation on the concatenated features; To represent the user's content characteristics; For the collaborative characteristics of users, To characterize the content features of information; This represents the collaborative characteristics of information.
[0037] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and grouped collaborative filtering, the matching degree between target information and users is calculated according to user profile representation, information joint representation, and user joint representation, specifically as follows: Calculate the matching degree between users and candidate information. where the following relationship is satisfied: , , , Calculate the cosine similarity of content representations Dot product similarity with co-representations By absorbing the advantages of late-stage fusion, a first-order interaction is explicitly constructed to provide a correlation signal between the two. User preference characteristics; By retaining user profile representations and joint representations from both parties, and leveraging the advantages of early fusion, the model can directly acquire the complete semantic information learned by the user tower and the information tower respectively, thus helping it capture fine-grained relationships that cannot be expressed by the inner product. Output Joint representation with each candidate information Matching score .
[0038] Furthermore, in the science and technology innovation information recommendation method based on multi-dimensional weight fusion and group collaborative filtering, candidate information is ranked, and the information with the highest matching degree is accurately pushed to the target user, specifically as follows: For all candidate information, a joint representation is formed, based on the joint representation of candidate information and... Matching score Sort in descending order; The matching score of the information is adjusted based on its popularity and business needs. The top-K candidate information with the highest matching scores are selected and returned to be displayed to the user.
[0039] In addition, this method also includes: Acquire multi-source scientific information within a preset time period, and extract patent application volume, paper publication volume, and investment and financing event number based on the multi-source scientific information within the preset time period; A technology popularity index is constructed based on the number of patent applications, the number of papers published, and the number of investment and financing events. A change point detection algorithm is used to divide the technology popularity index into technology stages. For hot data in the nascent stage of technology, we prioritize recommending original patent texts and top conference papers; for hot data in the mature stage of technology, we focus on industry analysis and competitive intelligence. A weighting function is constructed, and weighting coefficients are calculated based on the hot data of the mature technology stage or the hot data of the nascent technology stage. The user matching degree is then updated according to the weighting coefficients.
[0040] It should be noted that the formula for calculating the weight system is as follows: , Where A, B, and C are correction coefficients, x is the technology age, y is the heat intensity growth rate, and z is the user technology stage matching.
[0041] It should be noted that this method fully considers the significant differences in information value at different technological stages, such as patents in the nascent stage versus news in the declining stage, which can improve recommendation accuracy and achieve differentiated recommendations.
[0042] In addition, this method also includes: Construct a reinforcement learning recommendation agent, and use user profiles and browsing technology distribution within a preset time as the state space, and explore known interest-related information and emerging technology information as the action space; Click-through rate, dwell time, and subsequent behavior are used as the reward space. The reinforcement learning recommendation agent is trained using the state space, action space, and reward space. The information list is then recommended using the reinforcement learning recommendation agent. A Bayesian network is introduced, and the information list is input into the Bayesian network to estimate the uncertainty of users' interest in emerging technology information. Information with high uncertainty receives exploration bonus points, and the technology information with exploration bonus points is statistically analyzed to generate a list of uncertain information. Calculate the cosine distance between the list of uncertain information and the user's historical attention. The list of uncertain information with a cosine distance less than a preset cosine distance threshold is used as a scientific information recommendation list, and the information is pushed in batches based on the scientific information recommendation list.
[0043] It should be noted that over-reliance on users' historical interests in existing technologies can lead to "information cocoons," causing users to miss out on emerging disruptive technologies. It is advisable to appropriately push information with high novelty but strong technical relevance, such as recommending basic advances in quantum computing to integrated circuit engineers.
[0044] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.
[0045] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.
[0046] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.
[0047] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0048] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.
[0049] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A scientific and technological information recommendation method based on multi-dimensional weight fusion and grouping collaborative filtering, characterized in that, Includes the following steps: Acquire basic information and professional attribute data, and construct static user profile representations and information content representations based on the basic information and professional attribute data; The static user profile representation and information content representation are obtained by extracting the title, summary and keywords of the information using a pre-trained large language model; Information content representations are clustered to form information groups, and users are clustered to form user groups based on user interaction behavior with information groups; Based on historical interaction data between users and information, a user group-information group interaction matrix is constructed, and collaborative representations of users and information are learned respectively. By integrating content representations and collaborative representations from the user side and the information side respectively, a joint representation of users and information containing comprehensive information is generated; Based on user profile representation, information joint representation, and user joint representation, the matching degree between target information and users is calculated, candidate information is sorted, and the information with the highest matching degree is accurately pushed to the target user.
2. The scientific and technological information recommendation method based on multi-dimensional weight fusion and grouping collaborative filtering according to claim 1, characterized in that, Acquire basic information and professional attribute data, and construct static user profile representations and information content representations based on the basic information and professional attribute data, specifically as follows: Upload basic information and professional attribute data, and extract professional attributes from user registration information and historical behavior data based on the basic information and professional attribute data; Numerical features are extracted from the professional attributes in the user registration information and historical behavior data, and the numerical features are converted into discrete features. One-hot encoding is used for features with fewer than a preset threshold. For discrete features with no fewer than a preset threshold of categories, the word embedding model is used to convert them into dense vectors and average pooling is performed. One-hot encoding and dense vectors are then concatenated, and multilayer perceptron encoding is used to obtain user profile representations. By using prompt templates, the title, keywords, and summary of the information are extracted using a pre-trained large language model, and then the text representation of the information is obtained. After the large model returns the extracted information, a pre-trained embedding model is used to embed this information to obtain the content representation of the information. 3.The scientific and technological information recommendation method based on multi-dimension weight fusion and group collaborative filtering according to claim 1, characterized in that, The information content representation is clustered to form information groups, and users are clustered to form user groups based on their interaction behavior with the information groups. Specifically: Input all user sets All information collection Information content representation set User interaction behavior sequence ,in The length of the action sequence. Information data representing behavior lengths from 1 to u is clustered using K-means based on its content representation. Behavioral vectors are constructed based on the frequency of user interactions with information groups, and spectral clustering is performed on users. For new ungrouped information, it is assigned to the group with the highest cosine similarity between its content representation and the class center vector. For new, ungrouped users, they are temporarily assigned to groups based on their user profiles. Once their interaction sequence reaches a preset length, they are then assigned to groups based on their behavior vectors.
4. The scientific and technological information recommendation method based on multi-dimensional weight fusion and grouping collaborative filtering according to claim 3, characterized in that, Based on historical user-information interaction data, a user group-information group interaction matrix is constructed, and collaborative representations of users and information are learned separately, specifically: constructing a group interaction matrix, wherein the group interaction matrix satisfies the following relationship: ， in, Represents the group interaction matrix The Middle line, number The element value of the column represents the first element. The user group and the first The intensity of interaction between information groups; Indicates the first The first group Each user group contains several users with similar characteristics; No. The first group Each information group contains several pieces of information with similar characteristics; Indicates user Information The interaction weights are a weighted sum of multiple behaviors, among which... ; Learning collaborative representations through singular value decomposition satisfies the following relationship: ， in, For user collaborative representation, For information collaboration representation; Use a circular queue to store a set of matrices. These matrices record the most recent consecutive period of time. User interaction behavior information, where each matrix records non-overlapping consecutive... The interaction information of the day, when the interval between the time of a matrix and the current time exceeds Then clear the matrix data and use it to record the next... Interactive information between the heavens; When performing matrix factorization to learn collaborative representations, the following relation is satisfied: ， represents one data in the interaction matrix at the jth time point represents one data in the interaction matrix at the jth time point 5. The scientific and technological information recommendation method based on multi-dimensional weight fusion and group collaborative filtering according to claim 4, characterized in that, Also includes: User interaction behavior sequence-based After calculating four groups of weights for each piece of information of the interaction respectively, including similarity weighting, group weighting, time series weighting and behavior signal weighting, the four groups of weights are fused. In the similarity weighting, the cosine similarity of the information content representation and the user profile representation is calculated. In group weighting, the frequency of occurrence of each group of information in the user interaction information sequence is counted, and the group weight is scaled logarithmically by the frequency. In time series weighting, an exponential decay function is used to weight the time series, and finally the weighted results are fused.
6. The method for recommending science and technology innovation information based on multi-dimensional weight fusion and grouped collaborative filtering according to claim 5, characterized in that, By integrating content representations and collaborative representations from both the user and information sides, a joint representation of users and information containing comprehensive information is generated, specifically including: User content representation is obtained based on the content representation of weighted fused interactive information, where the following relationship is satisfied: ， wherein, is the weighting result in the weighted fusion, is the content feature representation of the i-th user.
7. The method for recommending science and technology innovation information based on multi-dimensional weight fusion and grouped collaborative filtering according to claim 6, characterized in that, Also includes: pass Functions generate user joint representations and information joint representation where the following relationship is satisfied: ， ， in, is the joint representation vector of the user, which is the final output of the model; w is a learnable weight matrix used to perform linear transformation on the concatenated features; To represent the user's content characteristics; For the collaborative characteristics of users, To characterize the content features of information; This represents the collaborative characteristics of information.
8. The scientific and technological information recommendation method based on multi-dimensional weight fusion and group filtering of claim 7, wherein, Based on user profile representation, information joint representation, and user joint representation, the matching degree between target information and users is calculated, specifically as follows: Calculating a match score for a user and a candidate information where the following relation is satisfied: ， ， ， Calculate the cosine similarity of content representations Dot product similarity with co-representations By absorbing the advantages of late-stage fusion, a first-order interaction is explicitly constructed to provide a correlation signal between the two. User preference characteristics; By retaining user profile representations and joint representations from both parties, and leveraging the advantages of early fusion, the model can directly acquire the complete semantic information learned by the user tower and the information tower respectively, thus helping it capture fine-grained relationships that cannot be expressed by the inner product. Output Joint representation with each candidate information Matching score .
9. The method of claim 1, wherein the method further comprises: The candidate information is sorted, and the information with the highest matching degree is accurately pushed to the target user, specifically: joint representation of all candidate information, descending order sorting according to the matching score of the joint representation of the candidate information and the target user . The matching score of the information is adjusted based on its popularity and business needs. The top-K candidate information with the highest matching scores are selected and returned to be displayed to the user.