A method and system for accurately recommending university library subject documents based on a multi-dimensional knowledge graph

Abstract

The application provides a university library subject literature accurate recommendation method and system based on a multidimensional knowledge graph, and relates to the technical field of information retrieval. The method comprises the following steps: the application performs text logical analysis on a target literature set, extracts basic metadata and deep academic logical feature entities, constructs a multidimensional subject knowledge hypergraph and a cross-dimensional hyperedge; obtains a user real-time interaction behavior sequence and a subject research cycle state feature vector, and obtains a dynamic space-time intention representation matrix through feature alignment and fusion; positions a candidate literature node based on the dynamic space-time intention representation matrix, constructs a heterogeneous meta-path along the cross-dimensional hyperedge, and performs attention aggregation and probability fusion calculation, and pushes a recommendation list. The application overcomes the cross-disciplinary semantic fault, realizes the accurate mapping of the space-time dynamic scientific research intention, and significantly improves the literature recommendation hit rate.

Description

Technical Field

[0001] This invention relates to the field of information retrieval technology, and in particular to a method and system for precise recommendation of subject literature in university libraries based on multidimensional knowledge graphs. Background Technology

[0002] University digital libraries house massive amounts of academic digital resources. The continuous growth of these resources increases the difficulty of implementing precise cross-disciplinary document retrieval. Knowledge graph architecture has been introduced into the document distribution module of library and information systems. Existing document recommendation technologies typically extract author, abstract, and keywords to construct entity nodes in the knowledge graph. For example, Chinese patent document CN113761323A (publication date: December 7, 2021) discloses a document recommendation system and method. This existing patent technology calculates semantic similarity based on document abstracts. This prior art scheme combines keyword similarity to rank documents and construct a knowledge graph to output recommendation results.

[0003] During the actual algorithm iteration process, researchers discovered limitations in the aforementioned conventional technical approaches. These existing technologies primarily rely on external explicit metadata and shallow abstract features for similarity calculation. This extraction logic fails to deeply analyze the research questions and underlying reasoning logic within the main text of the document. In interdisciplinary recommendation scenarios, single explicit feature associations are insufficient to establish cross-links based on underlying research paradigms. Traditional recommendation mechanisms often depend on discrete user interaction records for preference matching. Researchers' demands for document depth exhibit significant dynamic shifts at different research stages. Existing mechanisms lack dynamic tracking and modeling of the project's lifecycle and spatiotemporal state. Static matching strategies cannot adapt to the high-granularity document requirements at various stages of the research cycle. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for accurate recommendation of subject literature in university libraries based on multidimensional knowledge graphs. This invention solves the problems of shallow graph representation dimensions and low cross-domain recommendation accuracy caused by the lack of spatiotemporal modeling of the research cycle in existing technologies.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for precise recommendation of subject-specific literature in university libraries based on multidimensional knowledge graphs includes: Obtain the target document set and perform textual logic parsing to extract basic metadata feature entities and deep academic logic feature entities; A multidimensional subject knowledge hypergraph is constructed based on the basic metadata feature entities and the deep academic logic feature entities, and cross-dimensional hyperedges are constructed in the multidimensional subject knowledge hypergraph according to the semantic dependency relationships between the deep academic logic feature entities. Obtain the real-time interaction behavior sequence of the target user, and extract the feature vector of the current research cycle state of the target user; The real-time interactive behavior sequence is encoded with temporal features to obtain a basic interest vector, and the basic interest vector is fused with the research cycle state feature vector of the topic to obtain a dynamic spatiotemporal intent representation matrix. Using the dynamic spatiotemporal intent representation matrix as the inference benchmark, candidate document nodes are located in the multidimensional subject knowledge hypergraph, and heterogeneous meta-paths between the dynamic spatiotemporal intent representation matrix and the candidate document nodes are constructed along the cross-dimensional hyper-edges. Attention aggregation is performed on the node features of the heterogeneous meta-paths to calculate the local semantic matching degree. A global probability fusion calculation is performed on the local semantic matching degree corresponding to each heterogeneous meta-path to obtain a comprehensive inference probability value, and a subject literature recommendation list is pushed to the target user based on the comprehensive inference probability value.

[0006] A precise recommendation system for subject-specific literature in university libraries based on multidimensional knowledge graphs, comprising: The entity feature extraction module is used to obtain the target document set and perform text logic parsing to extract basic metadata feature entities and deep academic logic feature entities; The subject knowledge hypergraph construction module is used to construct a multidimensional subject knowledge hypergraph based on the basic metadata feature entities and the deep academic logic feature entities, and to construct cross-dimensional hyperedges in the multidimensional subject knowledge hypergraph according to the semantic dependency relationship between the deep academic logic feature entities; The behavior sequence and state feature extraction module is used to obtain the real-time interaction behavior sequence of the target user and extract the state feature vector of the current research cycle of the target user. The dynamic intent feature fusion module is used to encode the real-time interactive behavior sequence with temporal features to obtain a basic interest vector, and to perform feature alignment and fusion of the basic interest vector with the research cycle state feature vector of the topic to obtain a dynamic spatiotemporal intent representation matrix. The path construction and local matching inference module is used to locate candidate document nodes in the multidimensional subject knowledge hypergraph using the dynamic spatiotemporal intent representation matrix as the inference benchmark, and construct heterogeneous meta-paths between the dynamic spatiotemporal intent representation matrix and the candidate document nodes along the cross-dimensional hyperedges, so as to perform attention aggregation on the node features of the heterogeneous meta-paths and calculate the local semantic matching degree. The global fusion and recommendation push module is used to perform global probability fusion calculation on the local semantic matching degree corresponding to each heterogeneous meta-path to obtain a comprehensive inference probability value, and push a subject literature recommendation list to the target user based on the comprehensive inference probability value.

[0007] The present invention discloses the following technical effects: This invention provides a method and system for precise recommendation of subject literature in university libraries based on a multidimensional knowledge graph. By performing textual logical analysis on target documents to extract deep academic logical feature entities, and constructing cross-dimensional hyperedges in the multidimensional subject knowledge hypergraph, this invention effectively overcomes the cross-disciplinary semantic fragmentation problem caused by relying solely on shallow explicit metadata, achieving deep knowledge association based on underlying methodological levels. Simultaneously, it aligns and fuses the real-time interactive behavior sequence of target users with the research cycle state feature vector to generate a dynamic spatiotemporal intent representation matrix, accurately capturing the dynamic intent drift characteristics of researchers at different research stages and avoiding homogeneous data loops in recommendation results. Furthermore, based on the dynamic spatiotemporal intent representation matrix, heterogeneous meta-paths are constructed along the cross-dimensional hyperedges, and attention aggregation and global probability fusion calculations are performed, achieving precise mapping and inference between the target user's spatiotemporal dynamic intent and deep academic logical entities. Ultimately, this significantly improves the hit rate and recommendation accuracy of cross-disciplinary literature distribution in university libraries. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating a method for precise recommendation of subject-specific literature in university libraries based on a multidimensional knowledge graph, as provided in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0012] like Figure 1 As shown, this invention provides a method for precise recommendation of subject-specific literature in university libraries based on multidimensional knowledge graphs, including: Step 100: Obtain the target document set and perform textual logic parsing to extract basic metadata feature entities and deep academic logic feature entities; Step 200: Construct a multidimensional subject knowledge hypergraph based on the basic metadata feature entities and the deep academic logic feature entities, and construct cross-dimensional hyperedges in the multidimensional subject knowledge hypergraph according to the semantic dependency relationships between the deep academic logic feature entities; Step 300: Obtain the real-time interaction behavior sequence of the target user and extract the feature vector of the current research cycle state of the target user; Step 400: Encode the real-time interactive behavior sequence with temporal features to obtain a basic interest vector, and then perform feature alignment and fusion of the basic interest vector with the research cycle state feature vector to obtain a dynamic spatiotemporal intent representation matrix. Step 500: Using the dynamic spatiotemporal intent representation matrix as the inference benchmark, locate candidate document nodes in the multidimensional subject knowledge hypergraph, and construct heterogeneous meta-paths between the dynamic spatiotemporal intent representation matrix and the candidate document nodes along the cross-dimensional hyper-edges, so as to perform attention aggregation on the node features of the heterogeneous meta-paths and calculate the local semantic matching degree. Step 600: Perform global probability fusion calculation on the local semantic matching degree corresponding to each heterogeneous meta-path to obtain a comprehensive inference probability value, and push a subject literature recommendation list to the target user based on the comprehensive inference probability value.

[0013] Furthermore, the specific implementation process of steps 100-200 is as follows: This embodiment acquires a target document set and uses a pre-defined natural language processing (NLP) model to perform textual logic parsing on the content of the target document set. The NLP model employs a multi-head attention encoder architecture. The textual logic parsing process first transforms the target document set into a standard text sequence format and strictly truncates it into a sequence of standard syntactic units of 512 characters each. Subsequently, this embodiment performs named entity recognition through textual logic parsing, extracting author and journal attributes from the target document set to obtain basic metadata feature entities. During the data processing for named entity recognition, this embodiment performs part-of-speech tagging and boundary prediction on the standard syntactic unit sequences, constructing an input tensor composed of part-of-speech tagging features and entity boundary features. This input tensor is then used to perform sequence labeling tasks at a conditional random field layer, forcibly extracting entity nodes with confidence scores greater than a pre-defined score threshold of 0.85. This pre-defined score threshold is derived from the cross-validation baseline of a historical document annotation database, ensuring the accuracy and absolute reliability of the generated basic metadata feature entities.

[0014] After extracting basic metadata, this embodiment performs semantic slot filling through textual logic parsing to extract implicit technical elements representing research problems and theoretical models from the target document set, thereby obtaining deep academic logic feature entities. Deep academic logic feature entities refer to structured feature expressions that are detached from the external surface labels of documents and are specifically used to represent the underlying scientific research methodology and mathematical deduction mechanism of the documents. This embodiment constructs a semantic slot template containing three core elements, corresponding to the core research objective, the underlying derivation hypothesis, and the constraint boundary conditions. To achieve accurate text-to-slot mapping, this embodiment inputs the extracted text unit semantic feature vector and the pre-initialized preset slot label vector into a multilayer perceptron for dimensional alignment, projecting them into a shared semantic space of the same dimension. Then, the spatial Euclidean distance between the aligned semantic feature vector and the preset slot label vector is calculated. This embodiment forces the text fragment with the shortest spatial Euclidean distance to fill the corresponding empty slot, thereby directly extracting unstructured long text into structured deep academic logic feature entities, achieving quantitative capture and solidification of implicit technical elements.

[0015] After entity extraction, this embodiment directly uses the extracted basic metadata feature entities and deep academic logic feature entities as graph network nodes to initialize and construct a multi-dimensional subject knowledge hypergraph. Subsequently, this embodiment calls a graph autoencoder to extract features from deep academic logic feature entities belonging to different disciplines, obtaining initial semantic embedding vectors for each deep academic logic feature entity. The graph autoencoder consists of an encoding processing layer and a decoding and restoration layer. In this embodiment, the adjacency matrix of the graph network node and the node's own attribute feature matrix are multiplied by an inner product and then synchronously input into the encoding processing layer. A nonlinear activation function mapping is performed by modifying the linear unit function to filter out negative noise, and finally, a dense continuous non-negative vector with a fixed dimension of 256 is output as the initial semantic embedding vector. This initial semantic embedding vector completely preserves the high-order network topology structure of the entity nodes and the feature distribution state of the original semantic dimensions in the low-dimensional space.

[0016] To achieve precise alignment of deep interdisciplinary logic, this embodiment constructs an orthogonal projection matrix comprising a theoretical constraint subspace and an experimental parameter subspace. The orthogonal projection matrix is ​​then used to map each initial semantic embedding vector to both the theoretical constraint subspace and the experimental parameter subspace, resulting in a multi-view projection feature set. The specific data processing procedure for this orthogonal projection matrix is ​​as follows: a co-occurrence matrix of high-frequency interdisciplinary words is extracted from historical literature data; singular value decomposition is performed on this co-occurrence matrix; and the two largest sets of orthogonal singular vectors are extracted as the initial weights of the orthogonal projection matrix. The theoretical constraint subspace is used to rigorously define the boundary condition feature set of the mathematical model, while the experimental parameter subspace is used to characterize the physical variable feature set of the experimental environment. This embodiment calculates the distribution divergence values ​​between the multi-view projection feature sets of different disciplines within both the theoretical constraint subspace and the experimental parameter subspace to quantify the semantic dependencies between entities representing deep academic logic features. The data processing of the distribution divergence value is achieved by calculating the relative entropy between two probability distribution vectors (i.e., the expected value of the information difference between the actual distribution and the reference distribution). In this embodiment, the divergence threshold is strictly fixed at 0.15. This value is taken from the statistical lower boundary of successful cross-disciplinary cases. When the calculated distribution divergence value is closer to the starting zero value, the quantitative characterization of the semantic dependency between the two in the underlying academic methodology is stronger.

[0017] The expression for the numerical value of the distribution divergence is: ; in, This represents the numerical value of the distribution divergence. The total number of feature dimensions of the theoretical constraint subspace or experimental parameter subspace; For specific dimension indexes; The first deep academic logic feature entity in the subspace 3D probability distribution components; For the second deep academic logic feature entity in the subspace 3D probability distribution components; The probability distribution mean vector of the first deep academic logic feature entity; This is the probability distribution mean vector of the second deep academic logic feature entity; It is an exponential function with the natural constant as its base; The Euclidean distance norm is the two mean vectors that characterize the spatial distance penalty factor. When the divergence values ​​of the distributions of a pair of deep academic logical feature entities are all below a preset divergence threshold in all subspaces, this embodiment determines that there is a strong transfer possibility between the deep academic logical feature entities, and establishes non-isomorphic edges between them. This strong transfer possibility limits the direct borrowing and cross-disciplinary inspiration value of these two documents belonging to different macro-disciplines in terms of underlying inference algorithms or experimental verification mechanisms. Finally, this embodiment uses hypergraph convolution operators to adaptively reorganize the high-order topology of all established non-isomorphic edges, and unifies the reorganized non-isomorphic edge network clusters as cross-dimensional hyperedges in the multi-dimensional subject knowledge hypergraph. The execution process of the hypergraph convolution operator adopts a graph information transmission network architecture with a receptive field depth of 2 layers. During the transmission process, this embodiment calculates the correlation weight matrix based on the number of hyperedges to which the entity node belongs and the total number of nodes contained in the hyperedge. By aggregating the allocation values ​​of entity nodes in the correlation weight matrix, the discrete non-isomorphic edges are fused and denoised. The resulting cross-dimensional hyperedge breaks through the limitation of traditional graphs connecting two nodes at a time, and realizes high-dimensional association and aggregation of multiple cross-disciplinary feature entities in the same hidden semantic space.

[0018] Furthermore, the specific implementation process of step 300 is as follows: Furthermore, this embodiment acquires the real-time interaction behavior sequence of the target user, sets the historical time period to the past 30 days, and analyzes the real-time interaction behavior sequence to statistically analyze the target user's search term generalization index and single-discipline literature access density index within this historical time period. The search term generalization index is used to characterize the semantic span of the user's input search term in the subject classification tree system, wherein the subject classification tree system is derived from the basic book classification standard structure library mounted at the bottom layer of this system. This structure library is fixed as a hierarchical node architecture with a fixed depth of 5 levels, including top-level macro-disciplines and bottom-level sub-research fields. The generalization index value is determined by the hierarchical depth of the least common ancestor node of the last subject node to which each search term belongs in the subject classification tree. This embodiment divides the interaction behavior sequence according to 24-hour time slices, and generates the search term generalization index by calculating the variance of the node hierarchical depth corresponding to all search terms triggered in each time slice. The single-discipline literature access density index is used to characterize the depth of a user's research into literature in a specific sub-field. In this embodiment, it is calculated by counting the total number of interactions of literature browsing and full-text download belonging to the same lowest-level discipline node within a single time slice, and then dividing it by the overall literature interaction base within that time slice for normalization. The overall literature interaction base is strictly defined as the absolute sum of the number of independent triggers of browsing, bookmarking, and downloading actions actually performed by the target user for all discipline nodes within that time slice, thereby ensuring that the physical boundary of the normalization denominator is clear and the processing logic is strictly closed-loop.

[0019] Subsequently, this embodiment extracts the numerical change gradient of the search term generalization index and the peak frequency of the single-discipline literature access density index to construct an observation sequence for state inference. The numerical change gradient refers to the slope data obtained by dividing the difference in the search term generalization index between two adjacent time slices by the corresponding time interval. This gradient feature can quantitatively reflect the convergence or divergence rate of the user's scientific research search perspective. This embodiment sets a continuous 7-day sliding observation window to extract this gradient data. The peak frequency refers to the cumulative number of times the single-discipline literature access density index exceeds a preset high-frequency threshold within the sliding observation window. This embodiment strictly fixes this high-frequency threshold to 15 deep interactions within a unit time slice. By accumulating the number of days that the search reaches or exceeds this high-frequency threshold, a peak frequency feature reflecting the user's continuous deep-focus behavior is generated.

[0020] After extracting the underlying feature data, this embodiment integrates the numerical change gradient and the peak frequency into a continuous two-dimensional observation data stream, and inputs it into a pre-trained state classification model. In this embodiment, the state classification model is specifically instantiated as a pre-trained Hidden Markov Model (HMM). An HMM is a statistical data processing model involving a dual stochastic process, containing an invisible set of hidden states and a visible set of observation data. This embodiment utilizes a labeled database containing 5000 real-world samples of university researchers' behavior to perform offline training of the model parameters using the Expectation-Maximization (EM) algorithm. This solidifies and generates a state transition probability matrix describing the inherent transition patterns between hidden research states, and an emission probability distribution function describing the generation of observation data streams under specific hidden states, thereby establishing the global inference parameters of the HMM.

[0021] This embodiment uses the Hidden Markov Model to perform state sequence deduction, outputting one of the following states: the initial research stage, the experimental intensive stage, or the thesis writing stage, to obtain the current research stage. In the specific deduction data processing stage, this embodiment calls a dynamic programming path optimization algorithm. This algorithm extracts the local optimal cumulative probability of the previous time step within each deduction time step and iteratively multiplies it with the corresponding transition probability value in the state transition probability matrix and the emission probability distribution function value of the current observed data stream in the corresponding state. Thus, under the condition of inputting a two-dimensional observed data stream, it inversely solves for the hidden state stream with the maximum joint posterior probability, and the final node output of this state stream is taken as the user's current research stage. In terms of physical logic mapping, when the observed numerical change gradient is positive and the peak frequency is less than 3, the model converges and outputs the initial research stage with high probability; when the numerical change gradient shows a sharply decreasing negative value and the peak frequency is as high as 5 or more, the model locks in the experimental intensive stage; other stable fluctuating states are mapped and determined as the thesis writing stage. This embodiment enforces a lower limit for inference confidence. Only when the calculated maximum joint posterior probability exceeds the confidence threshold of 0.75 will the current research stage be confirmed for output.

[0022] Finally, this embodiment performs vectorized encoding on the current research stage to obtain the research cycle state feature vector representing the user's research status. To preserve the probability distribution evolution information during the hidden state sequence deduction process and avoid semantic truncation loss caused by hard discrete encoding, this embodiment extracts three continuous posterior probability distribution values ​​calculated by the Hidden Markov Model at the final moment for the project initiation research stage, the experimental tackling stage, and the thesis writing stage. This embodiment concatenates these three normalized probability values ​​(between zero and one) according to a preset fixed stage dimension order, generating a dense continuous vector with a fixed 3-dimensional dimension as the final research cycle state feature vector. This 3-dimensional feature vector not only accurately locates the research stage in which the user currently holds absolute dominance but also precisely preserves the user's intention to shift to or remain in other adjacent research stages through the magnitude of the floating-point values.

[0023] Furthermore, the specific implementation process of step 400 is as follows: This embodiment analyzes the real-time interactive behavior sequence and extracts the interaction timestamps corresponding to each interactive behavior from the underlying data records. The interaction timestamps are used to accurately pinpoint the absolute physical time coordinates of the target user's triggering of academic actions such as browsing, saving, or downloading documents. This embodiment extracts the 100 most recent consecutive historical interactive behaviors as time-series processing objects, calculates the absolute value of the time difference between the current system time and the interaction timestamps, and fixes the unit of measurement for this absolute time difference in hours. Subsequently, this embodiment inputs the absolute value of the time difference into a preset exponential decay function for normalization calculation to obtain the time decay weight coefficient corresponding to each interactive behavior. The specific data processing process of this exponential decay function is as follows: using the natural constant as the base, and using the product of the negative absolute value of the time difference and the preset decay factor as the exponent, a power operation is performed.

[0024] The expression for the time decay weighting coefficient is: ; in, The time decay weight coefficient corresponding to each interactive behavior; It is an exponential function with the natural constant as its base; The preset attenuation factor is constant at 0.005. The absolute timestamp of the current system time; For the first The timestamp of the interaction corresponding to each interactive behavior; To calculate the absolute value of the obtained time difference; In this embodiment, the preset decay factor is strictly set to a constant of 0.005. The value of this constant is derived from the statistical fitting mean of the Ebbinghaus forgetting curve in the historical borrowing database of the digital library. Its physical meaning corresponds to a time half-life of about 140 hours. This processing mechanism makes the interaction behavior closer to the current system time obtain a weight coefficient that is closer to the value of 1, thereby forcibly amplifying the intention expression weight of recent behavior in the underlying mathematical mapping.

[0025] Next, in this embodiment, the real-time interactive behavior sequence is input into a Long Short-Term Memory (LSTM) network with a self-attention mechanism for temporal feature encoding. The LSM network contains a data filtering structure with forget gates, input gates, and output gates, specifically designed to capture long-range implicit dependencies in the interactive sequence. In this embodiment, the single-layer hidden state feature dimension of the LSM network is set to 128 dimensions, and a self-attention mechanism layer containing four independent computation heads is connected to the top layer of its network structure. During the encoding process, the self-attention mechanism layer calculates the global feature interdependence matrix between each interactive behavior node in the sequence, outputting an initial temporal feature vector carrying global sequence context information. Subsequently, this embodiment uses the time decay weight coefficients calculated in the aforementioned steps for feature weighting. The specific weighting calculation operation is as follows: the value of each time step dimension of the initial temporal feature vector is multiplied element-wise with the corresponding time decay weight coefficient. After completing the weight penalty for each time step, this embodiment performs a global average pooling operation along the step length dimension of the time sequence, compressing and accumulating the feature tensors of multiple time steps and calculating the mean, finally outputting a basic interest vector corrected by the time decay penalty. This basic interest vector quantitatively characterizes the overall dynamic-static academic preferences of users after incorporating the time-forgetting effect.

[0026] To bridge the heterogeneous physical representation gap between time-series features and research state features, this embodiment constructs a cross-modal alignment network. This network nonlinearly maps the basic interest vector and the research cycle state feature vector to a shared hidden space of unified dimension, yielding aligned candidate interest vectors and aligned candidate state vectors. The cross-modal alignment network consists of two sets of parallel deep feedforward neural networks with independent weight parameters. In this embodiment, the dimension of the shared hidden space for unified mapping is strictly defined as 256 dimensions. The specific feature mapping process is as follows: the aforementioned basic interest vector (fixed to 128 dimensions) and the aforementioned research cycle state feature vector (fixed to 3 dimensions) are input into the corresponding deep feedforward neural networks. A multilayer fully connected perceptron performs matrix multiplication for feature dimensionality upscaling, and a hyperbolic tangent function is used for nonlinear activation mapping, projecting the heterogeneous features into a 256-dimensional feature space with a unified physical scale metric. The resulting aligned candidate interest vectors and aligned candidate state vectors are not only completely aligned in dimensionality but also possess a fundamental basis for direct Euclidean distance measurement and inner product operations in mathematical space.

[0027] This embodiment further extracts the stage evolution gradient features from the research cycle state feature vector. The stage evolution gradient features refer to the difference in evolution rate between the probability distributions of different research states in the research cycle state feature vector, used to quantitatively determine the maturity and transition trend of the target user's research cycle. The specific data processing procedure for these stage evolution gradient features is as follows: This embodiment retrieves the target user's 3D historical state vector from the cache in the previous sliding time window, performs a vector subtraction operation on the corresponding dimension between the current research cycle state feature vector and the 3D historical state vector, and outputs a 3D gradient vector representing the probability transition difference. This embodiment generates a cycle gating threshold for dynamically filtering redundant information based on the stage evolution gradient features. The generation process of the gating threshold is as follows: the evolution gradient features of the current stage are input into a single-layer logistic classifier. The weight matrix inside the single-layer logistic classifier is pre-trained and converged based on the frequency distribution of user cross-stage search behavior in the historical literature database. In this embodiment, the weight matrix is ​​used to perform linear weighted summation on the 3-dimensional gradient vector, and then a dynamically scaled scalar between 0 and 1 is output through sigmoid nonlinear activation function mapping. In the derivation of this specific calculation cycle, the cycle gating threshold is calculated to be 0.6. In the shared hidden layer space, this embodiment performs a mask filtering operation on the aligned candidate interest vector according to the cycle gating threshold to filter out feature parameters that do not match the current research cycle. The data execution logic of the mask filtering is as follows: the component values ​​of all 256 feature dimensions of the aligned candidate interest vector are traversed in sequence. When the absolute value of a certain dimension is lower than the cycle gating threshold of 0.6, the feature of that dimension is determined to be an irrelevant redundant item in the current research stage, and the feature value of that dimension is forcibly reset to zero and erased, thereby achieving strict feature alignment of different modal features at the semantic purity level.

[0028] After feature mask cleaning, this embodiment calculates the inner product correlation degree between the filtered aligned candidate interest vector and the aligned candidate state vector. The data processing procedure for the inner product correlation degree is as follows: the two dense feature vectors in the shared hidden layer space are multiplied element-wise according to their corresponding dimensions, and then a global summation is performed to output a global scalar value that characterizes the intensity of the overlap and resonance between the two features. Finally, this embodiment uses the inner product correlation degree as the weight distribution coefficient to perform an outer product expansion operation on the filtered aligned candidate interest vector and the aligned candidate state vector to complete feature alignment fusion and obtain a dynamic spatiotemporal intent representation matrix. The specific calculation mechanism of this outer product expansion operation is as follows: the aligned candidate interest vector in column vector form and the aligned candidate state vector in row vector form are subjected to tensor multiplication derivation to directly generate a two-dimensional feature matrix with a size of 256 rows by 256 columns. Then, all elements inside the two-dimensional feature matrix are uniformly multiplied by the inner product correlation degree global scalar to achieve joint feature activation. The resulting dynamic spatiotemporal intent representation matrix not only retains the independent spatial distribution information of the two different dimensional features, but also depicts the spatiotemporal coupling and deduction intent of the user's underlying interactive actions and the macroscopic scientific research status in a panoramic closed loop on a two-dimensional plane.

[0029] The expression for the dynamic spatiotemporal intent representation matrix is: ; in, The generated dynamic spatiotemporal intent representation matrix; This is the filtered, aligned candidate interest vector in column vector form; It is the transpose of the aligned candidate state vector in row vector form; The outer product extension operator for a two-dimensional tensor; The symbol for element-wise multiplication of a global scalar with a matrix; The inner product is the sum of the corresponding dimensions of two dense feature vectors. and These are the L2 norms of the two vectors, and the ratio of the inner product to the L2 norm product represents the global scalar of the inner product correlation degree. This embodiment also provides a precise recommendation system for subject-specific literature in university libraries based on multidimensional knowledge graphs, including: The entity feature extraction module is used to obtain the target document set and perform text logic parsing to extract basic metadata feature entities and deep academic logic feature entities; The subject knowledge hypergraph construction module is used to construct a multidimensional subject knowledge hypergraph based on the basic metadata feature entities and the deep academic logic feature entities, and to construct cross-dimensional hyperedges in the multidimensional subject knowledge hypergraph according to the semantic dependency relationship between the deep academic logic feature entities; The behavior sequence and state feature extraction module is used to obtain the real-time interaction behavior sequence of the target user and extract the state feature vector of the current research cycle of the target user. The dynamic intent feature fusion module is used to encode the real-time interactive behavior sequence with temporal features to obtain a basic interest vector, and to perform feature alignment and fusion of the basic interest vector with the research cycle state feature vector of the topic to obtain a dynamic spatiotemporal intent representation matrix. The path construction and local matching inference module is used to locate candidate document nodes in the multidimensional subject knowledge hypergraph using the dynamic spatiotemporal intent representation matrix as the inference benchmark, and construct heterogeneous meta-paths between the dynamic spatiotemporal intent representation matrix and the candidate document nodes along the cross-dimensional hyperedges, so as to perform attention aggregation on the node features of the heterogeneous meta-paths and calculate the local semantic matching degree. The global fusion and recommendation push module is used to perform global probability fusion calculation on the local semantic matching degree corresponding to each heterogeneous meta-path to obtain a comprehensive inference probability value, and push a subject literature recommendation list to the target user based on the comprehensive inference probability value.

[0030] Furthermore, the specific implementation process of steps 500-600 is as follows: This embodiment uses the dynamic spatiotemporal intent representation matrix as the inference benchmark to calculate the distance metric between the dynamic spatiotemporal intent representation matrix and the deep academic logic feature entities in the multidimensional subject knowledge hypergraph. Since the dynamic spatiotemporal intent representation matrix generated in the preceding steps is a two-dimensional tensor, this embodiment first performs a global max pooling operation along the column feature dimension direction of this two-dimensional tensor (i.e., the spatial direction corresponding to the preceding aligned candidate state vector), strictly reducing the dimensionality of the 256-row multiplied-256-column two-dimensional feature matrix into a single 256-dimensional query intent vector. Subsequently, this embodiment calculates the Euclidean distance between this 256-dimensional query intent vector and the 256-dimensional initial semantic embedding vectors of each deep academic logic feature entity in the hypergraph within the shared hidden space. This embodiment strictly limits the preset distance range of the Euclidean distance to a value less than 0.45, marking deep academic logic feature entities whose Euclidean distance satisfies this preset distance range as recommendation anchors. These recommendation anchors represent the underlying methodological elements most directly relevant to the target user's current research intent. Next, in this embodiment, a breadth-first search algorithm is invoked in the knowledge hypergraph network topology to locate first-order physical neighbor document entities that have a direct connection relationship with the recommendation anchor point as candidate document nodes, thereby completing the initial delineation of cross-disciplinary recommendation targets.

[0031] After locating all candidate document nodes, this embodiment uses the dynamic spatiotemporal intent representation matrix as the starting node, connects the nodes via the associated cross-dimensional hyperedges, and terminates with the candidate document nodes to construct the heterogeneous meta-path. A heterogeneous meta-path refers to a directed topologically connected sequence formed by alternating combinations of semantic nodes of different types and implicit edges of different properties in a multi-level complex graph. To prevent semantic information decay and cross-domain noise diffusion caused by infinite graph traversal, this embodiment, based on the underlying academic deductive rules, forcibly limits the maximum jump depth of the constructed heterogeneous meta-path to 3 hops. Specifically, the constructed heterogeneous meta-path must and can only follow a fixed spatial topological sequence of "query intent starting node—recommendation anchor entity—cross-dimensional hyperedge—candidate document endpoint node." Through directional traversal in the graph network, this embodiment generates multiple sets of associated heterogeneous meta-paths representing different interdisciplinary migration logics for each located candidate document node.

[0032] Subsequently, this embodiment employs a graph neural network model to calculate the feature transfer weights between adjacent nodes on the heterogeneous path, and uses these feature transfer weights to perform a weighted summation of the node features on the heterogeneous path to achieve attention aggregation, thereby calculating the local semantic matching degree. In this embodiment, the graph neural network model is specifically instantiated as a graph attention network. In the specific data processing stage for calculating the feature transfer weights, this embodiment concatenates the feature vectors of adjacent source nodes and target nodes on the heterogeneous path along their first and last dimensions, and inputs the concatenated hybrid vector into a single-layer feedforward neural network with a leakage linear correction unit activation function having a fixed negative half-axis slope of 0.2, to calculate a non-normalized raw attention score. Next, this embodiment uses a normalized exponential function to perform probability space mapping on all raw attention scores pointing to the same target node, outputting feature transfer weights whose sum is strictly constant to the value 1. Finally, this embodiment uses the feature transfer weight to perform scalar multiplication and summation operations on the feature vector of the source node, completes the attention aggregation of node features on the entire heterogeneous path, and outputs a floating-point scalar with a value between 0 and 1 as the local semantic matching degree corresponding to a single heterogeneous path. This matching metric quantifies the logical alignment quality on a single source tracing path.

[0033] The expression for the feature transfer weights is: ; in, Weights are transferred for the features from the source node to the target node along the heterogeneous path; It is a normalized exponential function; The activation function for the leakage linear correction unit with a fixed negative half-axis slope of 0.2; This is the learnable feature mapping matrix of a single-layer feedforward neural network; The source node feature vector; The feature vector of the target node; This is the operator for concatenating the first and last dimensions. It is the set of all first-order physical neighbor nodes pointing to the target node; The feature vector of physical neighbor node k; To integrate the local inference results of multiple heterogeneous paths, this embodiment extracts the topological feature vectors of each heterogeneous path and assigns global confidence weights to the corresponding local semantic matching degrees based on these topological feature vectors. The topological feature vector refers to a set of structured attribute values ​​representing the degree distribution of nodes included in a single path, the length of path hops, and the type of cross-disciplinary hyperedges. This embodiment sequentially concatenates these three types of discrete attribute values ​​to construct a continuous feature vector with a fixed total dimension of 16. This continuous feature vector is then input into a multilayer perceptron network consisting of three hidden layers for nonlinear decoding. Simultaneously, an sigmoid nonlinear activation function is applied to the top-level output of this multilayer perceptron network, forcing an independent floating-point value between 0 and 1 as the global confidence weight for each path. Next, this embodiment performs scalar multiplication and weighted summation operations on the local semantic matching degrees output by each heterogeneous path using the global confidence weights to obtain the output result of global probability fusion calculation. Furthermore, in this embodiment, the weighted summation output is mapped to a normalized value through a nonlinear activation function to obtain the comprehensive inference probability value. This nonlinear activation function uses a mapping method combining the hyperbolic tangent function and linear transformation to forcibly compress and map the fusion calculation result of any real number field to the absolutely convergent numerical interval (0, 1). This comprehensive inference probability value ultimately constitutes the ultimate quantitative criterion for determining the interdisciplinary matching degree of the corresponding candidate documents.

[0034] After obtaining the final quantitative criteria for all candidate document nodes, this embodiment uses a fast sorting algorithm to globally sort all candidate document nodes in descending order based on the comprehensive inference probability values. After the sorting process is completed, this embodiment selects the candidate document nodes that rank within the top preset number to generate the subject literature recommendation list. This embodiment strictly fixes the top preset number to 20, that is, it forcibly selects the top 20 document resources with the highest comprehensive inference probability values ​​as the final exposure targets. These extracted recommended documents not only match the target user's historical search preferences in terms of explicit metadata, but also achieve deep cross-temporal alignment in terms of underlying experimental data model constraints and the user's current research lifecycle. Finally, this embodiment pushes the encapsulated subject literature recommendation list to the target user's terminal device through a data interface, thus completing the complete business processing loop of cross-domain multimodal feature fusion and high-granularity subject literature accurate recommendation based on multidimensional knowledge graph.

[0035] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0036] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for precise recommendation of subject-specific literature in university libraries based on multidimensional knowledge graphs, characterized in that, include: Obtain the target document set and perform textual logic parsing to extract basic metadata feature entities and deep academic logic feature entities; A multidimensional subject knowledge hypergraph is constructed based on the basic metadata feature entities and the deep academic logic feature entities, and cross-dimensional hyperedges are constructed in the multidimensional subject knowledge hypergraph according to the semantic dependency relationships between the deep academic logic feature entities. Obtain the real-time interaction behavior sequence of the target user, and extract the feature vector of the current research cycle state of the target user; The real-time interactive behavior sequence is encoded with temporal features to obtain a basic interest vector, and the basic interest vector is fused with the research cycle state feature vector of the topic to obtain a dynamic spatiotemporal intent representation matrix. Using the dynamic spatiotemporal intent representation matrix as the inference benchmark, candidate document nodes are located in the multidimensional subject knowledge hypergraph, and heterogeneous meta-paths between the dynamic spatiotemporal intent representation matrix and the candidate document nodes are constructed along the cross-dimensional hyper-edges. Attention aggregation is performed on the node features of the heterogeneous meta-paths to calculate the local semantic matching degree. A global probability fusion calculation is performed on the local semantic matching degree corresponding to each heterogeneous meta-path to obtain a comprehensive inference probability value, and a subject literature recommendation list is pushed to the target user based on the comprehensive inference probability value.

2. The method for precise recommendation of subject literature in university libraries based on multidimensional knowledge graphs according to claim 1, characterized in that, The process of acquiring the target document set and performing textual logical parsing to extract basic metadata feature entities and deep academic logical feature entities includes: The content of the target document set is subjected to textual logical parsing using a pre-defined natural language processing model; Named entity recognition is performed through the logical parsing of the main text, and author attributes and journal attributes are extracted from the target document set to obtain basic metadata feature entities. Semantic slot filling is performed through the logical parsing of the main text to extract the implicit technical elements representing research problems and theoretical models in the target document set, thereby obtaining deep academic logic feature entities.

3. The method for precise recommendation of subject literature in university libraries based on multidimensional knowledge graphs according to claim 1, characterized in that, The construction of a multidimensional subject knowledge hypergraph based on the basic metadata feature entities and the deep academic logic feature entities, and the construction of cross-dimensional hyperedges in the multidimensional subject knowledge hypergraph according to the semantic dependency relationships between the deep academic logic feature entities, includes: The basic metadata feature entities and the deep academic logic feature entities are used as graph network nodes to initialize and construct a multidimensional subject knowledge hypergraph. The graph autoencoder is invoked to extract features from the deep academic logic feature entities belonging to different disciplines, thereby obtaining the initial semantic embedding vectors of each deep academic logic feature entity. Construct an orthogonal projection matrix that includes a theoretical constraint subspace and an experimental parameter subspace, and use the orthogonal projection matrix to map each of the initial semantic embedding vectors to the theoretical constraint subspace and the experimental parameter subspace respectively, to obtain a multi-view projection feature set; Within the theoretical constraint subspace and the experimental parameter subspace, the distribution divergence values ​​between the multi-view projection feature sets of different disciplines are calculated respectively to quantify the semantic dependency relationships between the deep academic logic feature entities. When the distribution divergence values ​​of the same pair of deep academic logic feature entities are all lower than the preset divergence threshold in all subspaces, it is determined that there is a strong possibility of migration between the deep academic logic feature entities, and non-isomorphic edges are established between them. The hypergraph convolution operator is used to adaptively reorganize the high-order topology of all the established non-isomorphic edges, and the reorganized non-isomorphic edge network cluster is used as a cross-dimensional hyperedge in the multi-dimensional subject knowledge hypergraph.

4. The method for precise recommendation of subject literature in university libraries based on multidimensional knowledge graphs according to claim 1, characterized in that, The process of acquiring the real-time interaction behavior sequence of the target user and extracting the feature vector of the target user's current research cycle state includes: The real-time interactive behavior sequence is analyzed to statistically analyze the target user's search term generalization index and single-discipline literature access density index within a historical time period. By inputting the generalization index of the search terms and the access density index of the single-discipline literature into the status classification model, the current research stage of the target user can be obtained. The current research stage is vectorized and encoded to obtain the research cycle state feature vector representing the user's research status.

5. The method for precise recommendation of subject literature in university libraries based on multidimensional knowledge graphs according to claim 4, characterized in that, The step of inputting the generalization index of the search terms and the access density index of single-discipline documents into the state classification model to obtain the current research stage of the target user includes: Extract the numerical change gradient of the generalization index of the search terms and the peak frequency of the access density index of the single-discipline literature. The numerical change gradient and the peak frequency are input into a pre-trained Hidden Markov Model; By performing state sequence deduction using the Hidden Markov Model, one of the following states can be output: the initial research stage, the experimental tackling stage, or the thesis writing stage, thus obtaining the current research stage.

6. The method for precise recommendation of subject literature in university libraries based on multidimensional knowledge graphs according to claim 1, characterized in that, The process of encoding the real-time interactive behavior sequence with temporal features to obtain a basic interest vector, and then aligning and fusing the basic interest vector with the research cycle state feature vector to obtain a dynamic spatiotemporal intent representation matrix, includes: Extract the timestamps of each interaction in the real-time interaction sequence, and calculate the time decay weight coefficient of each interaction using the exponential decay function with the current system time as the reference. The real-time interactive behavior sequence is input into a long short-term memory network with a self-attention mechanism for temporal feature encoding, and the features are weighted using the time decay weight coefficient to obtain the basic interest vector. A cross-modal alignment network is constructed, which nonlinearly maps the basic interest vector and the research cycle state feature vector of the topic to a shared hidden layer space of the same dimension, to obtain the alignment candidate interest vector and the alignment candidate state vector. Extract the stage evolution gradient features from the periodic state feature vector of the research topic, and generate a periodic gating threshold for dynamically filtering redundant information based on the stage evolution gradient features; In the shared hidden space, a masking operation is performed on the aligned candidate interest vectors according to the periodic gating threshold to filter out feature parameters that do not match the current research period, thereby achieving strict feature alignment. The inner product correlation degree between the filtered aligned candidate interest vector and the aligned candidate state vector is calculated, and the outer product expansion operation is performed on the aligned candidate interest vector and the aligned candidate state vector using the inner product correlation degree as the weight distribution coefficient to complete the feature alignment fusion and obtain the dynamic spatiotemporal intent representation matrix.

7. The method for precise recommendation of subject literature in university libraries based on multidimensional knowledge graphs according to claim 6, characterized in that, The step of extracting the timestamps of each interaction in the real-time interaction sequence and calculating the time decay weight coefficient of each interaction using the current system time as a reference and an exponential decay function includes: Analyze the real-time interactive behavior sequence and extract the interaction timestamp corresponding to each interactive behavior; Calculate the absolute value of the time difference between the current system time and the timestamp of the interaction; The absolute value of the time difference is input into a preset exponential decay function for normalization calculation to obtain the time decay weight coefficient corresponding to each of the interactive behaviors.

8. The method for precise recommendation of subject literature in university libraries based on multidimensional knowledge graphs according to claim 1, characterized in that, The process involves locating candidate document nodes in the multidimensional subject knowledge hypergraph using the dynamic spatiotemporal intent representation matrix as an inference benchmark, constructing heterogeneous meta-paths between the dynamic spatiotemporal intent representation matrix and the candidate document nodes along the cross-dimensional hyperedges, and performing attention aggregation on the node features of the heterogeneous meta-paths to calculate the local semantic matching degree, including: Calculate the distance metric between the dynamic spatiotemporal intent representation matrix and the deep academic logic feature entities in the multidimensional subject knowledge hypergraph, mark entities that meet the preset distance range as recommended anchor points, and locate document entities that have a direct connection relationship with the recommended anchor points as candidate document nodes; The heterogeneous meta-path is constructed by connecting the dynamic spatiotemporal intent representation matrix as the starting node, via the associated cross-dimensional hyperedge, and with the candidate document node as the ending node. A graph neural network model is used to calculate the feature transfer weights between adjacent nodes on the heterogeneous path, and the node features of the heterogeneous path are weighted and summed using the feature transfer weights to achieve attention aggregation, thereby calculating the local semantic matching degree.

9. The method for precise recommendation of subject literature in university libraries based on multidimensional knowledge graphs according to claim 1, characterized in that, The step of performing a global probability fusion calculation on the local semantic matching degree corresponding to each heterogeneous meta-path to obtain a comprehensive inference probability value, and then pushing a subject literature recommendation list to the target user based on the comprehensive inference probability value, includes: Extract the topological feature vector of each heterogeneous metapath, and assign global confidence weights to the corresponding local semantic matching degree based on the topological feature vector; The local semantic matching degree of each heterogeneous metapath is weighted and summed using the global confidence weight to obtain the result of performing global probability fusion calculation; The weighted summation result is mapped to a normalized value through a nonlinear activation function to obtain the comprehensive inference probability value; Based on the comprehensive inference probability values ​​in descending order, all candidate document nodes are sorted, and the candidate document nodes that rank first by a preset number are selected to generate the subject literature recommendation list, which is then pushed to the target user.

10. A precise recommendation system for subject-specific literature in university libraries based on multidimensional knowledge graphs, characterized in that: include: The entity feature extraction module is used to obtain the target document set and perform text logic parsing to extract basic metadata feature entities and deep academic logic feature entities; The subject knowledge hypergraph construction module is used to construct a multidimensional subject knowledge hypergraph based on the basic metadata feature entities and the deep academic logic feature entities, and to construct cross-dimensional hyperedges in the multidimensional subject knowledge hypergraph according to the semantic dependency relationship between the deep academic logic feature entities; The behavior sequence and state feature extraction module is used to obtain the real-time interaction behavior sequence of the target user and extract the state feature vector of the current research cycle of the target user. The dynamic intent feature fusion module is used to encode the real-time interactive behavior sequence with temporal features to obtain a basic interest vector, and to perform feature alignment and fusion of the basic interest vector with the research cycle state feature vector of the topic to obtain a dynamic spatiotemporal intent representation matrix. The path construction and local matching inference module is used to locate candidate document nodes in the multidimensional subject knowledge hypergraph using the dynamic spatiotemporal intent representation matrix as the inference benchmark, and construct heterogeneous meta-paths between the dynamic spatiotemporal intent representation matrix and the candidate document nodes along the cross-dimensional hyperedges, so as to perform attention aggregation on the node features of the heterogeneous meta-paths and calculate the local semantic matching degree. The global fusion and recommendation push module is used to perform global probability fusion calculation on the local semantic matching degree corresponding to each heterogeneous meta-path to obtain a comprehensive inference probability value, and push a subject literature recommendation list to the target user based on the comprehensive inference probability value.

Images