Knowledge graph-based campus football in-school and out-school cooperation resource recommendation and docking method

Abstract

The invention provides a knowledge graph-based campus football in-school and out-school cooperation resource recommendation docking method, which comprises the following steps of: coding a user-resource time sequence interaction behavior through a time sequence diagram neural network model, and generating a dynamic interest vector; mining candidate resources in the knowledge graph by using a limited multi-hop semantic propagation method, fusing semantic span and functional heterogeneity to carry out resource diversity measurement, respectively evaluating a resource matching degree and diversity gain by using a two-channel scoring function, and carrying out semantic distance constraint based on a minimum spanning tree algorithm to prevent the resources from gathering a single interest field; and the diversity threshold is dynamically adjusted in combination with real-time feedback of the user, online closed-loop optimization of the recommendation system is realized, interest matching accuracy and result diversity of campus football resource recommendation are improved, and adaptive evolution of recommendation strategies is realized.

Description

Technical Field

[0001] This invention relates to the field of personalized resource intelligent recommendation technology in the field of campus football, and in particular to a method for recommending and connecting campus football resources for cooperation with external partners based on knowledge graphs. Background Technology

[0002] In recent years, intelligent resource recommendation and ranking systems have been widely used in the field of smart campuses, especially in campus football cooperation, resource integration, and project matching, where the demand for recommendation algorithms has increased significantly. In the public domain, most mainstream resource matching and recommendation systems adopt collaborative filtering, content recommendation (such as similarity measurement based on user profiles or resource tags), or knowledge graph-driven relevance screening methods. With the development of knowledge graph technology, more and more scenarios such as campus resource management, teacher collaboration, and venue and event matching are introducing entity relationship modeling and graph path reasoning mechanisms from knowledge graphs, aiming to improve the matching degree between recommendation results and user interests. However, in current technologies, most traditional recommendation methods primarily aim to improve relevance to user interests, which easily leads to highly homogenized recommended content. This results in a lack of sufficient diversity in the recommendation list, failing to effectively cover users' potential diverse interests or meet the needs of resource integration across multiple scenarios both inside and outside the school. While some methods introduce diversity or novelty objectives, most employ simple weighted fusion strategies (such as manually adjusting relevance and diversity weights in the final scoring model). In practical applications, these methods heavily rely on weight parameter selection and lack a systematic balance for semantic structure, making it difficult to achieve both accuracy and coverage in recommendation performance. Furthermore, existing multi-objective recommendation systems rarely conduct in-depth mining of the semantic path structure of knowledge graphs, lacking explicit modeling of the semantic span and functional coverage breadth of the recommendation list. Most recommendation algorithms also fail to fully integrate the dynamic diffusion mechanism of user interests on the knowledge graph, thus failing to achieve structurally novel recommendations and struggling to avoid the interest solidification and "information silo" phenomenon caused by collaborative behavioral signals. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method for recommending and connecting on-campus and off-campus cooperative resources for campus football based on knowledge graphs.

[0004] The technical solution of this invention is implemented as follows: a knowledge graph-based method for recommending and connecting on-campus and off-campus cooperative resources for campus football, comprising: S1: Extract the user's historical interest entity set from the campus football knowledge graph. The historical interest entities include structured knowledge nodes such as schools the user has cooperated with, associated coach qualifications, training venue types, and competitions participated in. S2: Encode the interaction sequence associated with the user's historical interest entities and timestamps, use a time-series graph neural network model to capture the evolution trend of user interests, and generate a dynamic interest vector with time-series awareness. S3: Based on the dynamic interest vector, perform restricted multi-hop semantic propagation in the campus football knowledge graph, expand the candidate resource set along predefined semantic relationships such as 'cooperation relationship', 'project matching', and 'complementary ability', and record the semantic span value of each propagation path; S4: Construct a dual-channel scoring function. The first channel calculates the matching score between candidate resources and dynamic interest vectors based on cosine similarity. The second channel calculates the diversity gain value based on path semantic span and functional overlap between resources. S5: Set up a semantic distance constraint module based on minimum spanning tree to calculate the weighted semantic distance between candidate resources and set a dynamic semantic distance threshold to avoid semantic homogeneous clustering of resources in the recommendation list; S6: A lightweight Pareto front search algorithm is used to perform non-dominated ranking of the candidate resource set, and to find the local Pareto optimal solution set between matching degree and diversity objectives, and generate a preliminary recommended ranking list. S7: Introduces an adaptive controller based on real-time user feedback signals, which dynamically adjusts the semantic distance threshold according to the browsing depth and jump frequency in user click behavior, thereby adjusting the exploration intensity of the recommendation list; S8: Implement a closed-loop optimization and update mechanism to re-inject user feedback behavior in the recommendation list into the dynamic interest vector generation model, thereby realizing the online evolution and continuous optimization of the recommendation strategy.

[0005] The knowledge graph-based method for recommending and connecting on-campus and off-campus cooperative resources for campus football, provided by this invention, has the following beneficial effects: (1) This invention constructs a time-series graph neural network that integrates dynamic interest evolution and a constraint mechanism based on semantic distance to achieve collaborative optimization ranking of resource relevance and diversity. This method can not only accurately identify the user's interest migration trend on the knowledge graph path, but also effectively explore and introduce resources with large differences in semantic structure by utilizing restricted multi-hop semantic propagation, which greatly improves the coverage and novelty of the recommended content and overcomes the problem of traditional collaborative filtering methods easily falling into recommendation homogenization. (2) By introducing a dual-channel scoring function and a semantic distance constraint based on the minimum spanning tree, this invention can effectively distinguish between the functional overlap of resources and potential semantic clustering. This mechanism can dynamically adjust the structural differences between resources in the recommendation list, significantly reducing the phenomenon of semantic homogeneous clustering. This makes the improvement of the diversity index of the recommendation method dependent on the real semantic structure between entities, rather than simply relying on randomization algorithms or weight parameter tuning. It has the technical advantages of strong interpretability and high adjustment sensitivity. (3) This invention adopts a lightweight Pareto front search and adaptive optimization closed loop, which dynamically introduces real-time user feedback into the regulation of recommendation strategy. It can optimize semantic distance threshold and diversity strategy based on user behavior, and achieve a dynamic balance between relevance and diversity objectives, making the recommendation scheme more in line with individual preferences. This closed loop mechanism improves the response speed and self-evolution ability of the recommendation method to actual needs, and effectively adapts to the complex scenario of connecting diverse resources inside and outside the school for campus football. Attached Figure Description

[0006] Figure 1 This is a flowchart of the knowledge graph-based method for recommending and connecting on-campus and off-campus cooperative resources for campus football, as described in this invention. Figure 2 This is a sub-flowchart of the knowledge graph-based method for recommending and connecting on-campus and off-campus cooperative resources for campus football, as presented in this invention. Figure 3 This is another sub-flowchart of the knowledge graph-based method for recommending and connecting campus football resources both inside and outside the school. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides a method for recommending and connecting on-campus and off-campus cooperative resources for campus football based on knowledge graphs, specifically including: S1: Extract the user's historical interest entity set from the campus football knowledge graph. The historical interest entities include structured knowledge nodes such as schools the user has cooperated with, associated coach qualifications, training venue types, and competitions participated in. S2: Encode the interaction sequence associated with the user's historical interest entities and timestamps, use a time-series graph neural network model to capture the evolution trend of user interests, and generate a dynamic interest vector with time-series awareness. S3: Based on the dynamic interest vector, perform restricted multi-hop semantic propagation in the campus football knowledge graph, expand the candidate resource set along predefined semantic relationships such as 'cooperation relationship', 'project matching', and 'complementary ability', and record the semantic span value of each propagation path; S4: Construct a dual-channel scoring function. The first channel calculates the matching score between candidate resources and dynamic interest vectors based on cosine similarity. The second channel calculates the diversity gain value based on path semantic span and functional overlap between resources. S5: Set up a semantic distance constraint module based on minimum spanning tree to calculate the weighted semantic distance between candidate resources and set a dynamic semantic distance threshold to avoid semantic homogeneous clustering of resources in the recommendation list; S6: A lightweight Pareto front search algorithm is used to perform non-dominated ranking of the candidate resource set, and to find the local Pareto optimal solution set between matching degree and diversity objectives, and generate a preliminary recommended ranking list. S7: Introduces an adaptive controller based on real-time user feedback signals, which dynamically adjusts the semantic distance threshold according to the browsing depth and jump frequency in user click behavior, thereby adjusting the exploration intensity of the recommendation list; S8: Implement a closed-loop optimization and update mechanism to re-inject user feedback behavior in the recommendation list into the dynamic interest vector generation model, thereby realizing the online evolution and continuous optimization of the recommendation strategy.

[0010] Step S1: Extract the user's historical interest entity set from the campus football knowledge graph. These historical interest entities include structured knowledge nodes such as schools the user has previously collaborated with, associated coach qualifications, training venue types, and competitions participated in. Specifically, this includes: S1.1: Based on user identification information in the campus football knowledge graph, obtain the set of school entities associated with the user in the historical period in order to identify its inter-school cooperation network structure; Based on user identification information in the campus football knowledge graph, an entity retrieval algorithm (parameters: unique user ID key value, entity category filter) is used to achieve preliminary extraction of all school nodes directly associated with user identifiers in the knowledge graph database; Furthermore, by using the relationship traversal method (parameters: entity relationship type = "inter-school cooperation", time range filter = historical period), the cooperation relationship path between the user's directly associated school nodes is parsed, and the timestamp attribute set of the school entity is obtained; Furthermore, a graph structure analysis algorithm (parameters: node degree threshold, connected component decomposition rules) is used to perform local network structure analysis on the extracted school entity set in the knowledge graph and generate adjacency matrix data of the inter-school cooperation network. Furthermore, by using an attribute aggregation algorithm (parameters: regional code, school level, football project implementation level), the attribute completion and feature combination of school nodes in the adjacency matrix are realized, and an attribute vector set is generated for subsequent network structure recognition. Furthermore, a network centrality calculation method (parameters: degree centrality, betweenness centrality, and compactness centrality) is adopted to achieve a quantitative evaluation of the inter-institutional cooperation network structure. The formula for degree centrality calculation is as follows:

[0011] in, Let be the number of direct connections to a given school node. The total number of nodes in the inter-school network is represented by the formula used to quantify the connection density of nodes in the network. Through the above graph structure analysis and centrality calculation processing, the results of the previous step are transformed into the set of school entities associated with the user in the historical period and their structural indicators in the inter-school cooperation network, so as to achieve accurate identification of the user's inter-school cooperation network structure. For example, in a campus football knowledge graph implementation environment, the user identifier is "UID_20231108", the entity category filter is set to "school", and the historical period is set to 5 years. The system calls the graph database interface to retrieve 12 school nodes directly associated with the user and parses the "inter-school cooperation" relationship paths between them, forming a subgraph containing 28 cooperation relationship edges. Through connected component decomposition, a single connected component with 12 nodes is obtained, with an adjacency matrix dimension of 12×12. In the attribute aggregation stage, the geographical code (East / West), educational level (primary school, middle school), and football program development level (basic, advanced, professional) of each school node are added. Centrality analysis is used to calculate the degree centrality of each school node, where the maximum value is... The value ≈0.545 indicates that this school has the densest cooperative connections in the network. The final output is the set of school entities associated with the user, along with the structural labels and quantitative indicators of this set within the inter-school cooperation network, providing a precise data source for the coach qualification association analysis in the next sub-step S1.2. S1.2: For each school node in the school entity set, query the 'employment relationship' edge between it and the coach entity, and extract the set of coach qualification entities associated with the user to represent the distribution of the professional ability of the user's cooperating coaches; Based on the school entity set data obtained in the previous step, a knowledge graph relationship query algorithm (parameters: relationship type = "employment relationship", node type filter = coach entity) is used to realize the semantic association path retrieval from school entities to coach entities and return a list of coach entities that have a direct employment relationship with each school node; Furthermore, through the coach qualification extraction model (parameters: attribute fields = qualification level, teaching experience, professional certification), the structured extraction of the coach entity attribute set is realized, and the qualification feature vector corresponding to each coach is obtained; Furthermore, an entity association filtering algorithm (parameter: user identifier filtering condition) is adopted to filter coaches in the coach entity list that have intersection with the target user in the historical cooperation record, and generate a subset of user-associated coaches; Furthermore, by using a qualification feature vector normalization processing algorithm (parameters: range standardization method = Min-Max, vector length alignment dimension = K), the qualification features are compared under a unified scale, providing consistent input for subsequent professional ability distribution analysis; Furthermore, a qualification distribution modeling algorithm (parameters: statistical method = kernel density estimation, dimension = qualification level, professional field) is adopted to estimate the distribution density of the professional ability space of the user-associated coach and generate an ability distribution matrix as output; Through the above algorithm chain, the set of school entities obtained in the previous step is transformed into a set of coach qualification entities associated with the user and their professional ability distribution data, so as to accurately depict the composition of coach resources in the user's campus football cooperation relationship. For example, in an inter-school cooperation network containing 10 school entities, when executing the knowledge graph relationship query algorithm, the relationship type is set to "employment relationship," and the node type is restricted to coach entities. This returns an average of 3 coaches associated with each school, totaling 30 coach entities. A qualification extraction model is then applied to these 30 coach entities to extract three attributes: qualification level (range 1-5), teaching experience (years range 1-20), and professional certification (Boolean feature), forming a qualification feature vector. For example, a coach's qualification feature might be [4, 15, 1], where 4 represents the qualification level, 15 represents the teaching years, and 1 represents professional certification. A user identifier filter is applied to compare historical cooperation records, filtering out 12 coaches who have cooperated with the target user, forming a subset of user-associated coaches. Min-Max normalization is then applied to this subset, normalizing the qualification level to the [0, 1] interval, the teaching years to the [0, 1] interval, and the professional certification as a binary feature, remaining unchanged, for example, [0.75, 0.78, 1]. A kernel density estimation algorithm is used to construct capability distribution matrices for both qualification level and professional field dimensions. Two peak regions are observed in the qualification level dimension, corresponding to a higher proportion of coaches with medium to high qualifications. The output capability distribution matrix is ​​directly used as a professional capability constraint in the subsequent training venue resource matching step, significantly improving the relevance of resource recommendations to users' actual needs. S1.3: Based on the set of coach qualification entities, traverse the training venue entities connected to them to obtain the set of training venue types that the user has used, so as to reflect their resource usage preferences; Based on the set of coach qualification entities, a knowledge graph adjacency traversal algorithm (parameters: maximum number of hops = 1, relation type filtering = training site associated edge) is used to realize the mapping and extraction from coach nodes to directly connected training site entity nodes. Furthermore, through an entity attribute filtering algorithm (parameters: site category identifier = campus football field, function and purpose label = training), the extracted training site entities are finely screened by attributes, and a candidate set of training sites that conforms to the user's historical actual usage records is obtained. Furthermore, a relationship strength calculation method (parameters: interaction frequency weight = 0.6, time span weight = 0.4) is adopted to evaluate the usage correlation between each entity in the candidate set of training venues and the coach qualification entity, and generate a training venue usage intensity index. Furthermore, by using a type clustering algorithm (parameters: similarity threshold = 0.75, clustering method = hierarchical clustering), the candidate set of training sites is aggregated and classified according to site type, and a set of training site types is generated. By performing set normalization, the training site type set from the previous step is transformed into a standardized site type representation vector, thereby achieving a unified output of user resource usage preference features. For example, in a recommendation system for collaborative resources between schools and universities, for a user's historical set of coach qualifications, an adjacency traversal algorithm is used, traversing only once along the training field association edges to extract an initial set containing entities of artificial turf fields, standard soccer fields, and multi-purpose gymnasium fields. Attribute filtering is applied to this set to remove non-school soccer fields, retaining only artificial turf fields and standard soccer fields. The relationship strength is then determined using an interaction frequency weight of 0.6 and a time span weight of 0.4. Calculation formula:

[0012] in, This is a standardized value for the frequency of venue use. To obtain the strength value of the artificial turf field using the time span standardized values, the strength value is... The strength value of a standard football field is The candidate set is input into a hierarchical clustering algorithm, with a similarity threshold of 0.75, generating two types of venue type clusters: artificial turf and natural turf. Through normalization, the venue type clusters are mapped to preference vectors of length 2 [0.48, 0.52]. This vector serves as the standardized output of user resource usage preferences and is used in subsequent step S1.4 for association queries of the event set. This process achieves accurate mapping and preference quantification from coach qualifications to training venue types, significantly improving the completeness and accuracy of historical interest entity construction. S1.4: For each type of training venue in the set of training venue types, query the set of events it has participated in, extract the set of campus football event entities that the user has participated in in the past, and construct a user activity trajectory map. Node relationship retrieval processing is performed on the set of training venue types. A reverse indexing algorithm based on knowledge graph relationship matching is used (parameters: relationship type = 'participating in events', query depth = 1 hop) to realize the mapping relationship between training venue types and the set of events they participate in. Furthermore, by using a semantic relationship filtering method (parameters: event type filtering condition = 'campus football event', semantic tag similarity threshold = 0.85), non-campus football event nodes in the mapping relationship are removed, and the original set of event projects in the limited domain is obtained; Furthermore, a node attribute merging algorithm (parameters: merge key = 'event identifier ID', conflict resolution strategy = 'timestamp priority') is adopted to deduplicatize and integrate duplicate event nodes in the original set of event items, and generate a unique set of event items; Furthermore, by utilizing a trajectory construction method based on node centrality (parameters: centrality calculation index = 'weighted degree centrality', weight source = venue usage frequency), the unique event set is sorted according to the association strength with the training venue type, and an ordered list of events in which users have participated in the past is generated; By using graph structure serialization processing, the list of user-participated event projects is transformed into time-series activity trajectory data, thereby standardizing the input for subsequent construction of user activity trajectory maps and dynamic interest modeling. For example, in the implementation environment of a campus football cooperative recommendation system, the user's historical training field type set includes two types of nodes: grass fields and indoor artificial turf fields. A reverse indexing algorithm based on knowledge graph relationship matching is executed on this set, with the relationship type set to 'participate in events' and the query depth set to 1 hop. The knowledge graph retrieves nodes associated with grass fields as Campus Football League A and Provincial Friendly Match B, and nodes associated with indoor artificial turf fields as Winter Indoor Invitational Tournament C and Campus Football League A. Using a semantic relationship filtering method, only nodes with a tag similarity greater than 0.85 and the type being campus football events are retained, resulting in League A and Winter Indoor Invitational Tournament C. A node attribute merging algorithm is used, with the event identifier ID as the merging key. When resolving duplicate conflicts, the latest timestamp record is prioritized, ultimately generating a unique set of event nodes: {League A, Winter Indoor Invitational Tournament C}. Next, based on the weighted degree centrality calculation of venue usage frequency, the frequency of use of the grass field (20 times) and the frequency of use of the indoor artificial turf field (12 times) correspond to event association strengths of 0.62 and 0.38, respectively. After sorting, a list of user's historical participation in events [League A, Winter Indoor Invitational Tournament C] is generated. Finally, the sorted list is combined with event date information for time-series processing to form activity trajectory data containing event identifiers, types, and times of occurrence, providing time-dependent standardized input for subsequent dynamic interest vector modeling. S1.5: The four types of entity sets, namely schools, coaches, venues, and events, are structurally integrated to form a unified set of user historical interest entities, which serves as the input data source for subsequent user interest evolution modeling; Based on the school entity set, coach qualification entity set, training venue type set, and competition project entity set output by steps S1.1 to S1.4, a multi-source entity fusion algorithm (parameters: entity primary key field, type label, semantic embedding vector) is used to achieve structured alignment of the four types of entity sets in a unified semantic space. Furthermore, through the entity feature mapping function (parameters: entity type, attribute set, relational adjacency information), the standardized encoding of attribute fields of different types of entities is realized, and a common feature dimension index data structure for cross-type entities is generated; Furthermore, through a relation merging algorithm (parameters: knowledge graph relation set, type priority constraint), cross-type edge merging is achieved in the four types of entity sets, and a unified relation adjacency matrix is ​​obtained to describe the global connectivity of the user's historical cooperation network. Furthermore, an entity association matrix concatenation and sparsification method (parameter: sparsity threshold 0.3) is adopted to realize the joint representation of the relationship matrices of the four types of entity sets and generate a global feature matrix of user historical interests; The embedding vectors of four types of entities are fused into a high-dimensional unified interest representation vector using a vector fusion algorithm (parameters: weighting ratio α=0.5, β=0.3, γ=0.2). As the input data source for subsequent user interest evolution modeling, it achieves the structured integration and semantic alignment effect from a collection of scattered entities to a unified collection of interest entities; For example, in a campus football cooperation dataset, the school entity set contains 50 school nodes (with attribute fields: region, level, and number of historical cooperations), the coach qualification entity set contains 120 coach nodes (with attribute fields: qualification level and years of coaching experience), the training field type set contains 15 types of field nodes (with attribute fields: paving type and whether nighttime lighting is available), and the event entity set contains 80 event nodes (with attribute fields: event scale and frequency of holding). In the entity fusion stage, a semantic embedding method is used to encode each entity node into a 128-dimensional vector, which is then input into a feature mapping function to standardize the "region" field of schools, the "qualification level" field of coaches, the "paving type" field of fields, and the "event scale" field of events according to a unified numerical encoding rule. In the relationship merging stage, cross-type merging is performed on the edges of "cooperation relationship," "employment relationship," "use relationship," and "participation relationship" in the knowledge graph to generate a unified adjacency matrix with a size of 265×265 and a non-zero element ratio of 0.28. After concatenating the association matrices, sparsification is performed by removing edges with weights less than 0.05, resulting in a global feature matrix with a sparsity of 0.31. In the vector fusion stage, the embedded vectors of the four entity classes are weighted and summed according to the following formula: α=0.5 (schools), β=0.3 (coaches), and γ=0.2 (venues and events combined).

[0013] in, The average value of the embedding vectors of the school set. Embed the average vector of the coach set. The average value of the embedding vector for the site set. The average value of the embedded vectors is used to represent the event set. The fusion result is a 128-dimensional unified interest representation vector, which can significantly improve the semantic consistency and global dependency capture capabilities in subsequent temporal interest evolution modeling.

[0014] Step S2: Encode the interaction sequence associated with the user's historical interest entities and timestamps, and use a time-series graph neural network model to capture the evolution trend of user interests, generating a dynamic interest vector with time-series awareness. Specifically, this includes: S2.1: Preprocess the user's historical interest entities and their associated timestamp sequences to extract time interval features and event sequence sequence features in order to construct a time-enhanced user-resource interaction graph structure; The set of user historical interest entities and their corresponding timestamp sequence data obtained through the structured integration in step S1 are processed using a time series preprocessing algorithm (parameters: timestamp precision in milliseconds, missing value filling strategy is linear interpolation) to achieve complete and accurate processing of time information. Furthermore, by using a time interval calculation method (parameters: time format is uniformly UTC, minimum time unit is seconds), the time difference between adjacent interactive events is extracted, and a time interval feature matrix is ​​obtained; Furthermore, an event sequence encoding algorithm (parameters: sequence length truncation threshold is 100, event type dictionary encoding dimension is 64) is adopted to realize the discretized representation of the sequence features of interactive events and generate event sequence sequence encoding vectors; Furthermore, using a feature fusion method (parameters: fusion mode is weighted concatenation, weight ratio is 0.6 for time interval features and 0.4 for sequence features), the time interval feature matrix and the event sequence encoding vector are fused together to generate a time-series feature composite representation vector set; Furthermore, a graph structure construction algorithm is adopted (parameters: node type is distinguished as user entity and resource entity, edge type is distinguished as time-related edge and resource access edge) to map the temporal feature composite representation vector to the feature attributes of user-resource interaction graph nodes and edges, forming a temporal enhanced user-resource interaction graph structure; Through the above processing method, the original user historical interest entities and timestamp sequences are transformed into graph structure data with time feature weights and event sequence dependencies, so as to provide high-fidelity and semantically complete time-series input for the subsequent spatiotemporal aggregation process in S2.2; For example, in a campus football resource recommendation system, the input set of user historical interest entities contains four types of nodes: A. School, B. Coach qualifications, C. Training field type, and D. Participated events, with 50 instances in each type. The corresponding timestamp sequence precision is uniformly set to milliseconds and calibrated using the UTC standard. The matrix size obtained by calculating the time interval feature is 200×1, with a value range of 0 to 86400 seconds; the event type dictionary encoding is set to 64-dimensional one-hot encoding, and the sequence truncation threshold is 100 entries, resulting in a final sequential feature matrix size of 200×64. In the fusion processing stage, the time interval feature matrix and the sequential feature matrix are weighted and concatenated at a weight ratio of 0.6:0.4, resulting in a fused vector set with a dimension of 200×65, where the first column is the weighted time interval feature and the remaining columns are the weighted sequential features. In the graph structure construction stage, a dual-type node graph is created, containing 1 user node, 200 resource nodes, and a total of 200 edges. Each edge is appended with a fused vector as a feature attribute. The above processing has been verified to significantly improve the temporal sensitivity of node embedding when calling the spatiotemporal aggregation module of the TGNN model, and can capture the dynamic trend of user interests within a 30-day period in the simulated session. S2.2: Based on the aforementioned temporal enhanced user-resource interaction graph structure, a temporal graph neural network (TGNN) framework is used to perform spatiotemporal aggregation of node features in order to integrate the dynamic correlation information between users and resource entities in the time dimension; S2.3: Perform parameter initialization and temporal sensitivity calibration on the time gating mechanism in TGNN to enhance the model's responsiveness to key time nodes in the user interest evolution path; S2.4: Based on the node embedding vector output by the time gating mechanism, the user's historical interaction path is time-weighted and fused to generate a user interest evolution trajectory vector with time-series awareness. S2.5: Perform normalization and dimension alignment operations on the user interest evolution trajectory vector to generate a standardized dynamic interest vector, which serves as the input representation for the subsequent restricted multi-hop semantic propagation module.

[0015] like Figure 2 As shown, step S3 involves performing restricted multi-hop semantic propagation in the campus football knowledge graph based on the dynamic interest vector, expanding the candidate resource set along predefined semantic relationships such as 'cooperation relationship', 'project matching', and 'complementary ability', and recording the semantic span value of each propagation path. Specifically, this includes: S3.1: Semantic type annotation is performed on the entity relationships in the campus football knowledge graph. Predefined semantic relationships such as 'cooperative relationship', 'project matching', and 'complementary ability' are used as guiding rules for the propagation path, providing a semantic constraint basis for subsequent multi-hop propagation; Based on the full entity relationship data in the campus football knowledge graph, a semantic type annotation method (parameters: full set of relationship edge attributes, entity category dictionary, predefined semantic relationship set) is used to realize the semantic category mapping of relationship edges; Furthermore, by using a relation pattern analysis algorithm (parameters: statistical distribution of relation triples, entity category combination rules), the pattern structure of relation edges is identified, and the structure type label of the relation edges in the knowledge graph is obtained; Furthermore, by utilizing the semantic mapping table construction method (parameters: semantic category set, relation edge ID mapping table), the corresponding connection between predefined semantic relations and relation edge encoding is realized, and a relation edge semantic annotation index matrix is ​​generated; Furthermore, a semantic consistency verification algorithm (parameters: annotation index matrix, entity category constraint rules) is adopted to achieve global consistency detection of semantic annotations for relation edges and generate a qualified annotation screening set; By combining the annotation index matrix with the qualified annotation filter set, the original relation edge data from the previous step is transformed into a relation edge data structure with three types of semantic tags: "cooperative relationship", "project matching" and "complementary capabilities", thus realizing the basic configuration of semantic propagation path guidance rules. For example, in a scenario containing 5000 relationship edges in a campus football knowledge graph, the input is a complete set of relationship edge attributes (including fields such as relationship type, direction, timestamp, and entity category). The entity category dictionary includes four types of nodes: school, coach, field, and event. A semantic type annotation method is used to label relationship edges that conform to the school-school relationship pattern and whose attribute field includes the number of inter-school collaborations as "cooperative relationship," those that conform to the field-event matching pattern and whose attribute field includes event type compatibility as "project matching," and those that conform to the coach-school complementary pattern and whose attribute field includes skill difference as "ability complementary." During the semantic mapping table construction process, a semantic annotation index matrix of relationship edges is generated, with a dimension of 5000×3, corresponding to the Boolean mapping of the three types of semantic relationships. Through semantic consistency verification, 152 semantically conflicting relationship edges are removed, retaining 4848 qualified annotated relationship edges as the propagation path guidance rule base. In subsequent multi-hop semantic propagation execution, this rule base ensures that the expanded candidate resource set propagates only along relationship edges guided by the three types of semantic tags, significantly improving the accuracy and controllability of semantic expansion. S3.2: Locate the initial interest node in the knowledge graph based on the user's dynamic interest vector. The dynamic interest vector is generated by encoding the user's historical interaction behavior by a temporal graph neural network and is used to drive the selection of the starting point and the priority ranking of the propagation direction of semantic propagation. S3.3: A restricted depth-first search algorithm is adopted. Starting from the initial interest node, multi-hop propagation is carried out along the labeled semantic relationship edge. During the propagation process, a maximum hop number threshold and a path repetition detection mechanism are set to prevent the propagation path from expanding indefinitely and semantic drift. S3.4: During the generation of each propagation path, the semantic span value of the path is calculated. The semantic span value is calculated based on the path hop count and the heterogeneity of semantic relationships in the path, and is used to quantify the breadth of the semantic space covered by the path. In the process of multi-hop semantic propagation, the number of hops and the sequence of relation types recorded during the path construction process are used as the calculation input to form a path span index for quantifying the breadth of semantic coverage. A path hop count method (parameters: maximum hop count threshold, unit hop weight set to 1) is used to measure the length of the propagation path on the knowledge graph. Furthermore, by using a relation heterogeneity measurement method (parameter: relation type set weight mapping table), the weighted statistics of the number of different semantic relation transformations occurring in the path are realized, and the relation heterogeneity factor value is obtained; Furthermore, a weighted fusion formula is used to weight and fuse the hop count statistics and the relation heterogeneity factor values ​​to generate a semantic span value, as shown in the following formula:

[0016] in, This is the semantic span value. This represents the number of hops in the path. As a factor of relational heterogeneity, and These are weighting coefficients obtained through expert experience or data-driven optimization. Furthermore, by using a normalization method (parameter: maximum and minimum normalization range [0,1]), the comparability of semantic span values ​​across different paths is achieved, and a standardized semantic span index is obtained. Using the above calculation method, the hop count and relational heterogeneity data output from the previous propagation step are transformed into structured semantic span values, thereby achieving a quantitative representation of the breadth of the semantic space covered by the path. For example, when calculating the semantic span of the propagation path of campus football cooperative resources, assume that a certain path contains a number of hops. Given a relationship sequence of {cooperation, project matching, complementary capabilities, complementary capabilities}, and assigning a weight of 1.0 to cooperation, 0.8 to project matching, and 0.9 to complementary capabilities based on a relationship weight mapping table, the relationship heterogeneity factor is... The calculation is the average weight of all changes in adjacent relationship types, with a total number of changes. The average weight is .set up = , = Then the semantic span value = + = Normalize this value within the range [0,1]. If the maximum span of all paths in the current session is 4.8 and the minimum span is 1.5, then the normalization result is: = Ultimately, the standardized semantic span index corresponding to this path is 0.26, which is used for diversity gain calculation and ranking optimization, achieving a significant improvement in the functional coverage breadth of recommendation resources; S3.5: Summarize the set of resource entity nodes reached by multi-hop propagation to form a candidate resource set, and use the semantic span value corresponding to each path as an additional semantic attribute of the resource node to provide structured input for subsequent diversity assessment and ranking decision.

[0017] like Figure 3 As shown, step S4 involves constructing a dual-channel scoring function. The first channel calculates the matching score between candidate resources and dynamic interest vectors based on cosine similarity, while the second channel calculates the diversity gain value based on path semantic span and functional overlap between resources. Specifically, this includes: S4.1: Perform vector space mapping on the semantic relationship between the candidate resource entity set and the dynamic interest vector to obtain the representation vector of the candidate resource in the user interest space; S4.2: Based on the cosine similarity calculation model between the representation vector of the candidate resource and the dynamic interest vector, perform the matching score calculation of the first channel to quantify the semantic relevance strength between the candidate resource and the user's current interest. Based on the correspondence between the representation vectors of candidate resources and dynamic interest vectors, a 128-300 dimensional semantic vector cosine similarity calculation method is adopted (parameter: candidate resource representation vector). Dynamic interest vectors This enables the semantic relevance measurement function between candidate resources and user interests; Furthermore, by performing vector inner product and modulus normalization, the two sets of vectors are mapped to the same unit space to reduce the interference of vector size differences on similarity calculation and obtain the original similarity metric, as shown in the following formula:

[0018] in, For dynamic interest vectors, This is the semantic representation vector of the candidate resource. This represents the vector dot product operation. Indicates the magnitude of the vector; Furthermore, a numerical stabilization algorithm (parameter: ε is a very small positive number to prevent division by zero error) is used to fine-tune the cosine similarity to address the numerical fluctuation problem of high-dimensional sparse vectors during the calculation process and achieve stable similarity score output. Furthermore, based on the time-weighted characteristics of dynamic interest vectors, a time-decay weighting function is used to adjust the similarity results, as shown in the following formula:

[0019] in, This is a weighted similarity metric. To fine-tune the corrected similarity metric, The time decay coefficient, The time interval between the most recent user interaction with the candidate resource; Furthermore, the weighted similarity score is assigned to the [0,1] interval by a vector normalization mapping function, so that it can be weighted and fused with the diversity gain value in the future to achieve the distribution consistency of different scoring channels; By using cosine similarity calculation and time-weighted correction, the representation vector from the previous step is transformed into a quantified matching score, achieving a high-precision semantic relevance evaluation effect between candidate resources and user interests. For example, in a campus football inter-school cooperation recommendation scenario, let the candidate resource representation vector be a feature vector of length 128, and the dynamic interest vector be an interest weight vector of length 128. Using a cosine similarity calculation model, the vector dot product result is 75.26, and the vector magnitudes are 9.14 and 8.21 respectively, yielding the original similarity value. The initial value was approximately 1.005, which was corrected to 0.985 after numerical stabilization (ε=1e-6). Assuming a time decay coefficient λ=0.05 and a time interval Δt=20 days between the user's most recent interaction with this type of resource, the weighted similarity was calculated. ≈0.658. After normalizing this value to the [0,1] interval, it is saved as the final matching score, which is used to merge and rank it with the diversity scores of the same batch of candidate resources. The results show that this resource has significantly improved in relevance indicators, reduced the repeated recommendation rate in the test set, and greatly improved the coverage of recommendation results in different interest dimensions. S4.3: Perform weighted statistical processing on the semantic span of the knowledge graph path between candidate resources, and generate a semantic distance matrix based on path length and relation heterogeneity factor to reflect the potential semantic differences between resources; Before performing weighted statistical processing on the set of knowledge graph path semantic span values ​​between candidate resources, the input data includes path hop count information and a set of semantic relationship heterogeneity weight factors derived from S3.4 and appended to resource nodes; A weighted summation algorithm is used (parameter: path hops). Relationship Heterogeneity Weight This function calculates the semantic span weighting value of a single path and obtains the path weighting span value matrix entries between each candidate resource pair. Furthermore, a multi-path aggregation algorithm (parameter: a set of weighted span values ​​of different paths for the same resource pair) is used to achieve cross-path weighted mean fusion processing and generate an average semantic span index for resource pairs to eliminate the influence of single-path randomness. Furthermore, a normalization method (parameters: average semantic span index set, maximum span value, minimum span value) is adopted to achieve interval standardization mapping of average semantic span and obtain a normalized semantic span matrix, thereby improving the comparability of different resource pairs. Furthermore, by using a relation type sensitive weighting algorithm (parameters: normalized semantic span value, corresponding relation type difference factor matrix), the product operation of semantic span and relation type heterogeneity is realized to quantify the potential semantic differences between resources and obtain the final semantic distance matrix. Through the above-mentioned weighted statistics and normalized matrix construction process, the path semantic span value of the previous step is transformed into a semantic distance matrix that fully reflects the potential semantic differences between resources, thereby realizing a structured and quantitative expression of resource differences. For example, the candidate set of the campus football knowledge graph includes resource nodes R1, R2, and R3. R1 and R2 have three multi-hop propagation paths with 2, 3, and 4 hops respectively, and their corresponding relation heterogeneity weights are 0.8, 0.9, and 0.7. A weighted summation formula is used.

[0020] in Let be the heterogeneity weight of the i-th path. Let n be the number of hops for the i-th path and n be the total number of paths. Calculate the average weighted span value. The result is (0.8×2+0.9×3+0.7×4) / 3=9.1 / 3≈3.033. The average weighted span value of all resource pairs is normalized, with the maximum value set to 5 and the minimum value set to 1. The mapping formula is...

[0021] Substituting S=3.033, Smax=5, Smin=1, we get ≈0.508. Combining this with the relation type difference factor F set to 1.2, the final semantic distance D = S' × F ≈ 0.61 is calculated. The above process is repeated for resource pairs R1 and R3, and R2 and R3, to obtain a complete 3×3 semantic distance matrix. This matrix serves as the core input in the subsequent calculation of the diversity gain value in S4.4, effectively improving the controllability of resource differences in multi-objective recommendation. S4.4: Based on the semantic distance matrix and the semantic overlap calculation model of the resource function description text, perform the diversity gain value calculation of the second channel to measure the novelty and differentiation contribution of the candidate resource in the recommendation list. For the semantic distance matrix generated by S4.3 and the set of functional description texts of candidate resources, a semantic overlap calculation model based on word vector embedding (parameters: embedding dimension is 100-300, preferably 300, and the pre-training corpus is the campus sports domain) is used to construct the semantic space of resource functions. Furthermore, text similarity calculation is performed on any two candidate resource entities using the vectorized functional description text (method: cosine similarity, parameters: stop word removal and stemming preprocessing), and a resource function semantic similarity matrix is ​​obtained; Furthermore, through The formula calculates the functional difference value between resources, where This represents the functional text similarity between resource i and resource j, ensuring that the difference value is normalized between 0 and 1. Furthermore, the aforementioned functional dissimilarity matrix and the semantic distance matrix generated in S4.3 are weighted and fused (parameters: functional weight coefficient α is 0.4-0.9, preferably 0.6; semantic path weight coefficient β is 0.1-0.6, preferably 0.4), and a comprehensive dissimilarity matrix is ​​generated using the fusion calculation formula. The formula is:

[0022] in This is a functional difference matrix. This is the semantic distance matrix; Furthermore, the comprehensive difference matrix is ​​statistically analyzed row by row, and the average difference of each candidate resource in the current set is calculated as the diversity gain value of that resource. Normalization is then performed to ensure that the index is within a uniform dimension. Through the above diversity gain value calculation model, functional differences and path semantic span are linked into a unified metric, realizing a quantitative evaluation of the novelty and differentiated contribution of each candidate resource in the composition of the recommendation list. For example, in a set of 50 candidate campus football resources, each resource is accompanied by functional description text (approximately 200 words). A 300-dimensional embedding is performed using a word vector model specific to the campus sports domain. The functional text matrix is ​​input into a cosine similarity calculation module, which outputs a 50×50 functional semantic similarity matrix. The similarity matrix is ​​then used for... The calculated functional dissimilarity matrix has an average value of approximately 0.72, reflecting a high degree of functional dissimilarity among resources. This matrix is ​​then fused with a semantic distance matrix generated from path semantic span, with α=0.6 and β=0.4, and the following steps are performed. The resulting comprehensive dissimilarity matrix improved the mean value to 0.81, significantly enhancing the comprehensiveness of the dissimilarity measurement. After calculating the average dissimilarity value for each resource row by row in this matrix, a group of resource nodes with high diversity gain values ​​were identified. These nodes effectively improved the structural diversity of the recommendation list in subsequent recommendation ranking. S4.5: Perform weighted normalization fusion processing on the matching score and diversity gain value to generate a comprehensive score vector of candidate resources to support the subsequent Pareto front search and non-dominated ranking.

[0023] Step S5: Set up a semantic distance constraint module based on minimum spanning tree to calculate the weighted semantic distance between candidate resources and set a dynamic semantic distance threshold to avoid semantic homogeneous clustering of resources in the recommendation list. Specifically, this includes: S5.1: Perform semantic path extraction processing on each pair of resource entities in the candidate resource set. Based on the predefined relational path structure in the campus football knowledge graph, obtain the multi-hop semantic association path between resource entities to construct a semantic adjacency relationship network. S5.2: Based on the semantic adjacency relationship network, a graph embedding algorithm is used to map the resource entities into a low-dimensional vector space to generate a set of semantic embedding vectors, so as to quantify the semantic representation distance of the resource entities in the knowledge graph. S5.3: Perform Euclidean distance calculation on each pair of resource entities in the semantic embedding vector set to obtain an initial semantic distance matrix, so as to reflect the degree of semantic similarity of resource entities in the embedding space; S5.4: Based on the category attributes, functional tags, and domain characteristics of resource entities, the initial semantic distance matrix is ​​weighted to generate a weighted semantic distance matrix, thereby enhancing the sensitivity to functional differences and domain coverage. S5.5: Perform the minimum spanning tree (MST) construction algorithm on the weighted semantic distance matrix to generate a semantic distance constraint tree structure, and set a dynamic semantic distance threshold based on the preset diversity control parameters to filter out resource combinations with semantic differences higher than the threshold, so as to avoid semantic homogeneous clustering of resources in the recommendation list. Based on the weighted semantic distance matrix as input data, the minimum spanning tree (MST) construction algorithm (parameters: weighted semantic distance matrix, resource node set) is used to generate a semantic distance constraint tree, so as to ensure that the minimum necessary path length and the maximum difference are preserved in the semantic connection between resources. Furthermore, using the Kruskal or Prim algorithm (parameters: row and column indices of the distance matrix, edge weights sorted in ascending order), the weighted semantic distance between all candidate resource pairs is selected edge by edge, and a tree structure covering all resource nodes is constructed under the conditions of ensuring full connectivity and no loops. Furthermore, based on the constructed semantic distance constraint tree, a traversal calculation method (parameter: the weighted distance set of all edges in the tree) is used to extract the statistical features of the global distance distribution, including the mean, standard deviation, and maximum and minimum distance values, and obtain the resource semantic difference matrix. Furthermore, through the diversity control parameter generation module (parameters: preset diversity target, relevance constraint weights), statistical features are mapped to dynamic semantic distance thresholds, calculated using a formula:

[0024] in, For dynamic semantic distance threshold, The mean of semantic distance. The standard deviation of semantic distance. This is the diversity adjustment coefficient; Furthermore, by using a threshold filtering algorithm (parameters: dynamic semantic distance threshold, resource difference matrix), resource combinations with semantic distance below the threshold are eliminated, combinations with semantic difference above the threshold are retained, and an optimized resource set is generated as the input for subsequent sorting. By using the minimum spanning tree construction algorithm and dynamic threshold filtering, the weighted semantic distance matrix is ​​transformed into structured resource difference constraint data, thereby achieving the expected technical effect of avoiding semantic homogeneous clustering and improving diversity and coverage in the recommendation list. For example, in a campus football knowledge graph scenario containing 10 candidate resource nodes, the weighted semantic distance matrix is ​​generated by calculating the weights of resource category attributes, with the category difference weight set to 1.2, the functional label difference weight to 1.5, and the domain coverage weight to 1.0. When constructing the minimum spanning tree using the Prim algorithm, the edge with the smallest weighted semantic distance is selected first, gradually expanding to cover all nodes, resulting in a set of edge weights in the tree as {0.82, 1.05, 1.08, 1.12, 1.20, 1.25, 1.33, 1.38, 1.45}. The mean semantic distance is then calculated. = Standard deviation = Set a diversity adjustment coefficient = The dynamic semantic distance threshold is calculated using the formula. = During the screening process, resource combinations with a semantic distance below 1.305 were removed, while combinations with a distance above this threshold were retained, ultimately resulting in a resource set with significantly improved diversity. Validation results show that the recommendation results significantly improved in terms of functional domain coverage, while maintaining a high level of relevance to user interests.

[0025] Step S6: A lightweight Pareto front search algorithm is used to perform non-dominated ranking of the candidate resource set, searching for a locally Pareto optimal solution set between matching degree and diversity objectives, and generating a preliminary recommended ranking list. Specifically, this includes: S6.1: Normalize the matching score and diversity gain value for each resource in the candidate resource set to eliminate the interference of different dimensions on the multi-objective optimization process and obtain the standardized matching vector and the standardized diversity vector. S6.2: Based on the standardized matching degree vector and the standardized diversity vector, construct a set of resource distribution points in a two-dimensional target space, where each resource corresponds to a coordinate point in the target space, the horizontal axis represents the matching degree, and the vertical axis represents the diversity. S6.3: A lightweight Pareto front search algorithm is used to perform non-dominated sorting on the resource distribution point set, identify all non-dominated solutions, and form a Pareto front solution set. The non-dominated solutions are resource nodes that cannot further improve diversity without reducing the matching degree, or vice versa. Based on the input condition of a resource distribution point set in a two-dimensional target space, where each point consists of a standardized matching degree vector and a standardized diversity vector, a lightweight Pareto front search algorithm is adopted (parameters: resource point set size N, non-dominance criterion is bi-objective optimization, matching degree threshold). diversity threshold This allows for the determination of non-dominance relationships for all resources; Furthermore, a dominance detection method based on target coordinate comparison (parameters: coordinate point Pi=(ri,di), coordinate point Pj=(rj,dj)) is used to determine the dominance between each pair of resource nodes and obtain the dominance Boolean matrix as the input result for subsequent filtering. Furthermore, an iterative elimination strategy (parameters: dominance Boolean matrix, resource point set index) is adopted to eliminate the set of resources that are completely dominated by other nodes, and retain the nodes that are not simultaneously superior to any node in all objectives, forming an initial non-dominated set data structure. Furthermore, the remaining non-dominated nodes are classified through sorting and hierarchical processing (parameter: non-dominated set data structure), Pareto front set is constructed according to the distribution of matching degree and diversity coordinates, and an index list of nodes in the set is generated for subsequent weighted filtering and local optimum determination. By using the above non-dominated sorting method, the resource distribution point set results of the previous step are transformed into Pareto front solution set technical indicators, achieving a comprehensive and balanced screening effect between matching degree and diversity objectives. For example, in the campus football knowledge graph recommendation scenario, the size of the input resource distribution point set is... There are 1 resource node, and the matching degree r and value range of each node are [ ]. , The range of diversity d is [ , The lightweight Pareto search algorithm is used with the following parameter configurations: matching degree threshold. = diversity threshold = The dominance test formula is: ,in For the i-th resource point, This is the matching degree value. This represents the diversity value. During algorithm execution, the number of non-dominated decision entries in the dominance Boolean matrix is ​​reduced to [value missing]. Each resource node is selected. Through iterative elimination and hierarchical sorting, a list of resource nodes is obtained. The Pareto front set of nodes. These nodes are not completely superior to other nodes in both matching degree and diversity objectives. The subsequent weighted score calculation significantly improves the balance of the recommendation list in matching resources inside and outside the school. S6.4: Perform a local optimal solution screening operation on the Pareto front solution set. Based on the preset matching degree priority weight and diversity priority weight, calculate the weighted comprehensive score of each non-dominated solution, and select the resource nodes with the highest comprehensive scores to form a local Pareto optimal solution set. S6.5: Sort the local Pareto optimal solution set according to the comprehensive score from high to low to generate a preliminary recommendation ranking list, which serves as the input data for the subsequent adaptive controller module and is used to dynamically adjust the semantic distance threshold in the recommendation strategy. After receiving the comprehensive score data of the local Pareto optimal solution set, a sorting algorithm (parameters: sorted in descending order of comprehensive score, with a stability requirement of O(n log n) complexity) is used to implement the priority sorting function of resource nodes; Furthermore, by using a key-value mapping-based index construction method (parameters: resource unique identifier, comprehensive score), efficient element location is achieved during the sorting process, and a sorted list of resource indexes is obtained. Furthermore, a serialization processing algorithm (parameters: JSON structured format, fields including resource ID, comprehensive score, matching degree value, and diversity gain value) is adopted to transform the sorted resource index list into a structured recommendation list data object and generate a business data packet that can be directly transmitted. Furthermore, by using a list truncation and step size control method (parameters: number of truncation items K, step size S), the size of the initial recommendation list is controlled, and a dataset that meets the input requirements of the subsequent adaptive controller is obtained. By using a list version identifier and timestamp embedding algorithm (parameters: timestamp format ISO-8601, version number auto-increment strategy), the recommendation list results from the previous step are transformed into recommendation data with version control and effective time, thus providing a stable and traceable data source for subsequent dynamic adjustment of semantic distance threshold; For example, the resource comprehensive score range of the local Pareto optimal solution set is set to 0 to 1, with a matching weight of 0.6 and a diversity weight of 0.4. When executing the sorting algorithm, a quicksort method is used, and the original order of nodes with the same score is stably preserved. During the index construction phase, a hash table-based key-value storage is used to associate resource IDs (e.g., R001, R002, etc.) with comprehensive scores (e.g., 0.85, 0.82, etc.) to generate a lookup table. During serialization, the following JSON structure is constructed: {"ResourceID":"R001","Score":0.85,"MatchScore":0.89,"DiversityGain":0.81}. Multiple objects are grouped into an array and encapsulated into a recommendation list data package. During the list truncation phase, K=50 and step size S=5 are set, retaining the top 50 resource nodes as the initial recommendation list. During the version identifier generation phase, the current timestamp (e.g., 2024-06-15T08:00:00Z) and version number (e.g., V3) are embedded into the list metadata, resulting in the format {"Version":"V3","Timestamp":"2024-06-15T08:00:00Z","Recommendations":[...]}. This format is directly invoked in the next step when the adaptive controller dynamically adjusts the semantic distance threshold. This process ensures the accuracy and traceability of the ranking results, significantly improving the usability and adaptability of the recommendation list while balancing matching degree and diversity.

[0026] Step S7: Introducing an adaptive controller based on real-time user feedback signals, dynamically adjusting the semantic distance threshold according to the browsing depth and jump frequency in user click behavior, and regulating the exploration intensity of the recommendation list. Specifically, this includes: S7.1: Collect user click behavior data on the recommendation list page. The click behavior data includes page dwell time, number of clicked resources, browsing depth sequence and page jump frequency to obtain a real-time user interaction behavior feature sequence. S7.2: Based on the user's real-time interaction behavior feature sequence, the sliding window statistical method is used to calculate the user's behavioral activity index in the current recommendation session, so as to generate a behavioral activity quantification value, which is used as one of the input parameters of the adaptive controller; S7.3: Perform fuzzy logic control rule matching on the measured value of the behavior activity, and generate a preliminary semantic distance threshold adjustment signal based on the preset fuzzy control strategy to achieve preliminary dynamic adjustment of the exploration intensity of the recommendation list; Based on the input conditions of user real-time interactive behavior feature sequence and behavior activity measurement value, a fuzzy logic control rule matching algorithm (parameters: user activity value, preset semantic distance adjustment rule set) is used to map continuous activity index into fuzzy membership degree. Furthermore, through the membership function calculation module (parameters: the endpoints of the triangular membership function and the trapezoidal membership function), the fuzzification processing of the low, medium and high levels of user activity is realized, and a multi-level activity membership vector is obtained; Furthermore, through a fuzzy rule reasoning engine (parameter: the rule base contains rules such as "low activity → increase exploration intensity" and "high activity → decrease exploration intensity"), the membership vector is matched with the control strategy rules, and a corresponding set of fuzzy output variables is generated to represent the increasing or decreasing trend of the semantic distance threshold. Furthermore, through the synthesis operation of fuzzy inference (parameter: Max-Min synthesis method), the output results of multiple rules are aggregated and processed to obtain a comprehensive fuzzy control output surface, which is used to capture the global influence of different activity levels on the exploration intensity. Furthermore, through the membership degree defuzzification module (parameter: centroid method), the comprehensive fuzzy surface is converted into a single semantic distance threshold adjustment value, and a preliminary semantic distance threshold adjustment signal is generated, which serves as the input for the subsequent exponentially weighted moving average algorithm. By using a fuzzy logic-based control rule matching approach, the behavioral activity metric is transformed into an executable semantic distance threshold adjustment signal, thereby enabling the initial dynamic adjustment of the recommendation list exploration intensity. For example, in a campus football resource recommendation system, the quantifiable value of behavioral activity is set to range from 0 to 100. A triangular membership function is used to define the low activity range as [0, 20, 40], the medium activity range as [30, 50, 70], and the high activity range as [60, 80, 100]. The rule base is set as follows: low activity is mapped to an increase of 0.3 in exploration intensity, high activity is mapped to a decrease of 0.3 in exploration intensity, and medium activity remains unchanged. When the input activity is 45, the fuzzification process calculates the low, medium, and high membership values ​​as 0.25, 0.75, and 0, respectively. The inference engine generates a fuzzy output set based on the rule base, weighting and aggregating the 0.3 increment corresponding to low activity and the 0 increment corresponding to medium activity, resulting in an exploration intensity increment of approximately 0.075. Defuzzification uses the centroid method to calculate the adjustment signal, as shown in the following formula:

[0027] in, To output the increment of exploration intensity within the variable range, The comprehensive membership function value of this increment is calculated to output an adjustment signal of 0.075. After being applied to the semantic distance constraint module, it enables dynamic fine-tuning of the exploration intensity of the recommendation list, effectively improving the diversity coverage of the recommendation results. S7.4: Based on the semantic distance threshold adjustment signal and historical recommendation feedback data, the updated dynamic semantic distance threshold is calculated using an exponentially weighted moving average algorithm to generate diversity control parameters suitable for the current user state; S7.5: Inject the dynamic semantic distance threshold into the semantic distance constraint module based on the minimum spanning tree to dynamically filter the weighted semantic distance between candidate resources in order to generate an optimized recommendation list that conforms to the current user interest evolution state; For the updated dynamic semantic distance threshold, a parameter injection mechanism (parameters include threshold value, timestamp label, and user session identifier) ​​is adopted to input it into the semantic distance constraint module based on minimum spanning tree, so as to realize the diversity filtering control in the recommendation list generation stage. Furthermore, by using a weighted semantic distance matching algorithm (parameters include resource category weight, function tag weight, and domain coverage factor), a dynamic threshold comparison operation is performed on the semantic distance matrix between candidate resources in the semantic distance constraint module, and a set of resource pairs that match the current user's interest status is obtained. Furthermore, using the minimum spanning tree path pruning algorithm (parameters include a dynamic threshold and a set of semantic embedding vectors for resource nodes), edges with semantic distances lower than the dynamic threshold are removed from the constrained tree structure, and a new set of highly dissimilar resource connectivity branches is generated. Furthermore, a resource branch sorting algorithm (parameters including the average matching degree within the branch and the cross-domain diversity index of the branch) is adopted to calculate the priority of each connected branch and generate a sorted index list that conforms to the dynamic interest evolution state. By combining resource filtering and branch sorting, the dynamic threshold adjustment results from the previous step are transformed into optimized recommendation list data, achieving the expected technical effect of balancing diversity and relevance. For example, in a certain application of recommending resources for campus football both on and off campus, the dynamic semantic distance threshold was updated to... The timestamp corresponds to Seconds, session identifier is an integer The candidate resource set contains There are 1 resource node, and the semantic embedding vector dimension is 1. In the semantic distance constraint tree constructed based on the minimum spanning tree, a weighted semantic distance matching algorithm is used to set the weights of resource categories. Functional tag weight Domain Coverage Factor Calculate the weighted semantic distance of each edge. And compare it with a dynamic threshold. The formula is as follows:

[0028] in, For the semantic distance of resource categories, For the semantic distance of function tags, To cover semantic distance within the domain. During pruning, all Below After deleting the edges, the remaining edges form several connected components. For each component, calculate the mean matching degree within the component. With cross-domain diversity index , and according to Generate a comprehensive priority score. The final output recommendation list includes... The resources, distributed across more than [number] resources in the original candidate set, are [number]. The category domain is % and the matching degree is maintained at a high relevance level, effectively achieving both functional diversification and interest relevance of the recommendation results.

[0029] Step S8: Execute a closed-loop optimization and update mechanism, re-injecting user feedback behavior in the recommendation list into the dynamic interest vector generation model to achieve online evolution and continuous optimization of the recommendation strategy. Specifically, this includes: S8.1: Collect real-time feedback data on user behavior in the recommendation list, such as clicks, browsing, jumping, and dwell time, to obtain timestamps and interaction intensity features of user behavior sequences, which serve as input signals for closed-loop optimization. S8.2: Based on the collected user behavior sequence, the sliding time window statistical method is used to calculate the user behavior pattern features in the recommendation list, including but not limited to click depth, jump frequency, and page dwell time distribution, to generate user behavior feedback feature vectors as update inputs for the dynamic interest evolution model; S8.3: The user behavior feedback feature vector is fused with the original dynamic interest vector. The fused vector is modeled in time series using a gated recurrent unit (GRU) model to generate an updated dynamic interest state representation to reflect the real-time evolution trend of user interests. S8.4: Based on the updated dynamic interest state representation, incremental semantic propagation is performed in the campus football knowledge graph. New candidate resource sets are expanded along the relationship edges such as 'cooperation relationship', 'project matching', and 'complementary ability', and the path semantic span value is updated to support resource matching and diversity control in the next cycle. S8.5: Input the updated candidate resource set into the dual-channel scoring function, recalculate the matching degree score and diversity gain value, and combine the semantic distance constraint module based on minimum spanning tree and Pareto front search algorithm to generate a new round of recommendation ranking list, realizing closed-loop feedback and online evolution of recommendation strategy.

[0030] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A knowledge graph-based method for recommending and connecting on-campus and off-campus cooperative resources for campus football, characterized in that, Includes the following steps: S1: Extract the set of user historical interest entities from the campus football knowledge graph. The user historical interest entities include the schools the user has cooperated with, the qualifications of associated coaches, the types of training venues, and the structured knowledge nodes of the competitions participated in. S2: Encode the interaction sequence associated with the user's historical interest entities and timestamps, and use a time-series graph neural network model to capture the evolution trend of user interests and generate dynamic interest vectors; S3: Based on the dynamic interest vector, perform restricted multi-hop semantic propagation in the campus football knowledge graph, expand the candidate resource set along predefined semantic relations, and record the semantic span value of each propagation path; S4: Construct a dual-channel scoring function. The first channel calculates the matching score between candidate resources and dynamic interest vectors based on cosine similarity. The second channel calculates the diversity gain value based on path semantic span and functional overlap between resources. S5: Set up a semantic distance constraint module based on minimum spanning tree to calculate the weighted semantic distance between candidate resources and set a dynamic semantic distance threshold; S6: The lightweight Pareto front search algorithm is used to perform non-dominated sorting on the candidate resource set, and to find the local Pareto optimal solution set between matching degree and diversity objectives, and generate a recommended sorting list. S7: Introduce an adaptive controller based on real-time user feedback signals, dynamically adjust the semantic distance threshold according to the browsing depth and jump frequency in the user's click behavior, and adjust the exploration intensity of the recommended ranking list.

2. The method for recommending and connecting on-campus and off-campus cooperative resources for campus football based on knowledge graphs according to claim 1, characterized in that, Following step S7, the following is also included: S8: Execute a closed-loop optimization and update mechanism to re-inject user feedback behavior from the recommendation list into the dynamic interest vector generation model.

3. The method for recommending and connecting on-campus and off-campus cooperative resources for campus football based on knowledge graphs according to claim 1, characterized in that, Step S1 specifically includes: Based on user identification information in the campus football knowledge graph, obtain the set of school entities associated with the user in the historical period and identify its inter-school cooperation network structure. For each school node in the school entity set, query the "employment relationship" edge between it and the coach entity, and extract the set of coach qualification entities associated with the user; Based on the set of coach qualification entities, traverse the training venue entities connected to them to obtain the set of training venue types that the user has used. For each type of training venue in the set of training venue types, query the set of events it has participated in, extract the set of campus football event entities that the user has participated in in the past, and construct a user activity trajectory map; The four types of entities—schools, coaches, venues, and events—are structurally integrated to form a unified set of user historical interest entities.

4. The method for recommending and connecting on-campus and off-campus cooperative resources for campus football based on knowledge graphs according to claim 3, characterized in that, Step S1 further includes employing entity retrieval algorithms, relation traversal methods, and graph structure analysis algorithms to extract school node sets from the knowledge graph based on user IDs, and to generate inter-school cooperation network structures and quantitative indicators through cooperation relationship paths and attribute aggregation.

5. The method for recommending and connecting on-campus and off-campus cooperative resources for campus football based on knowledge graphs according to claim 1, characterized in that, Step S2 specifically includes: Preprocess the user's historical interest entities and their associated timestamp sequences, extract time interval features and event sequence sequence features, and construct a time-enhanced user-resource interaction graph structure; Based on the aforementioned temporally enhanced user-resource interaction graph structure, a temporal graph neural network framework is used to perform spatiotemporal aggregation of node features; For the time gating mechanism in the time sequence graph neural network, perform parameter initialization and time sensitivity calibration; Based on the node embedding vector output by the time gating mechanism, the user's historical interaction path is time-weighted and fused to generate a user interest evolution trajectory vector. Normalization and dimension alignment operations are performed on the user interest evolution trajectory vector to generate a standardized dynamic interest vector.

6. The method for recommending and connecting on-campus and off-campus cooperative resources for campus football based on knowledge graphs according to claim 1, characterized in that, Step S3 specifically includes: Semantic type annotation is performed on the entity relationships in the campus football knowledge graph, and predefined semantic relationships are used as guiding rules for the propagation path; Initial interest nodes are located in the knowledge graph based on user dynamic interest vectors, wherein the dynamic interest vectors are generated by encoding user historical interaction behavior by a temporal graph neural network. A restricted depth-first search algorithm is adopted. Starting from the initial interest node, multi-hop propagation is carried out along the labeled semantic relationship edge. During the propagation process, a maximum hop number threshold and a path duplication detection mechanism are set. During the generation of each propagation path, the semantic span value of the path is calculated; The set of resource entity nodes reached by multi-hop propagation is aggregated to form a candidate resource set, and the semantic span value corresponding to each propagation path is used as an additional semantic attribute of the resource node.

7. The method for recommending and connecting on-campus and off-campus cooperative resources for campus football based on knowledge graphs according to claim 6, characterized in that, The semantic span value is calculated based on a weighted average of the path hop count and the heterogeneity of semantic relationships within the path.

8. The method for recommending and connecting on-campus and off-campus cooperative resources for campus football based on knowledge graphs according to claim 1, characterized in that, Step S4 specifically includes: The semantic relationship between the candidate resource entity set and the dynamic interest vector is processed by vector space mapping to obtain the representation vector of the candidate resource in the user interest space. Based on the cosine similarity calculation model between the representation vector and the dynamic interest vector, the matching score of the first channel is calculated. We perform weighted statistical processing on the semantic span of knowledge graph paths between candidate resources, and generate a semantic distance matrix based on path length and relational heterogeneity factor; Based on the semantic distance matrix and the semantic overlap calculation model of the resource function description text, the diversity gain value of the second channel is calculated. The matching score and the diversity gain value are weighted and normalized to generate a comprehensive score vector for the candidate resources.

9. The method for recommending and connecting on-campus and off-campus cooperative resources for campus football based on knowledge graphs according to claim 8, characterized in that, The dual-channel scoring function uses 128-300 dimensional semantic vector cosine similarity combined with time weighting in the first channel, and embeds functional description text between resources (dimension 100-300) based on cosine similarity and semantic distance weighting. The functional weight coefficient is 0.4-0.9, and the semantic path weight coefficient is 0.1-0.

6.

10. The method for recommending and connecting on-campus and off-campus cooperative resources for campus football based on knowledge graphs according to claim 1, characterized in that, Step S5 specifically includes: Semantic path extraction is performed on each pair of resource entities in the candidate resource set. Based on the predefined relational path structure in the campus football knowledge graph, multi-hop semantic association paths between resource entities are obtained, and a semantic adjacency relationship network is constructed. Based on the semantic adjacency network, a graph embedding algorithm is used to map resource entities into a low-dimensional vector space to generate a set of semantic embedding vectors. Perform Euclidean distance calculation on each pair of resource entities in the semantic embedding vector set to obtain an initial semantic distance matrix; Based on the category attributes, functional tags, and domain characteristics of resource entities, the initial semantic distance matrix is ​​weighted to generate a weighted semantic distance matrix; The minimum spanning tree construction algorithm is executed on the weighted semantic distance matrix to generate a semantic distance constraint tree structure. Based on the preset diversity control parameters, a dynamic semantic distance threshold is set to filter out resource combinations with semantic differences higher than the threshold.

Images