Cross-domain book recommendation method and system based on neural network

Abstract

The invention provides a cross-domain book recommendation method and system based on a neural network, and relates to the technical field of intelligent book recommendation, and the method comprises the steps: obtaining first information containing user reading history, retrieval behaviors, research directions and paper information, and converting the first information into an initial interest topological structure and an evolution trend of a user; then, a cross-domain interest migration graph of the user is obtained through Markov clustering; and inputting the interest migration graph and the structured book text information into an extended probability latent variable semantic model to obtain a semantic embedding expression fusing user interests. Potential association scores between the user and the books are calculated through association analysis of the expression and the interest migration graph, a cross-domain recommendation list is generated according to sorting, and personalized and structured book recommendation is achieved. According to the method, deep alignment of the interest migration structure and the semantic space is realized, and the accuracy and practicability of cross-domain recommendation are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent book recommendation, and in particular to a cross-domain book recommendation method and system based on neural network. Background Art

[0002] In the current era of information explosion and rapid knowledge update, traditional book recommendation systems are mainly built based on classic methods such as collaborative filtering and content matching. Although they have achieved certain results in the same field, they have shown obvious limitations when facing users' cross-disciplinary and cross-field information needs. Especially among high-level user groups such as scientific researchers and university teachers, their interest migration paths have highly heterogeneous and evolutionary characteristics, and traditional methods are difficult to capture complex interest migration trajectories and multi-dimensional semantic associations.

[0003] Therefore, there is an urgent need for a cross-domain book recommendation method and system based on neural networks to meet users' comprehensive recommendation needs for diverse and highly relevant books. Summary of the invention

[0004] The purpose of the present invention is to provide a cross-domain book recommendation method and system based on neural network to improve the above problems. In order to achieve the above purpose, the technical solution adopted by the present invention is as follows: In a first aspect, the present application provides a cross-domain book recommendation method based on a neural network, comprising: Acquire first information, wherein the first information includes the user's reading history data, historical search data, research direction label, and historically published paper information; The first information is converted into an initial interest topology structure and the user's interest evolution trend, and Markov clustering processing is performed based on the converted result to obtain the user's cross-domain interest migration map; The user's cross-domain interest migration map and the preset book text structured information are input into the extended probabilistic latent variable semantic model for processing, and semantic embedding is performed based on the processed preliminary book topic distribution to obtain the semantic embedding expression of cross-domain books that integrates user interests; Performing correlation analysis on the semantic embedding expression and the user's cross-domain interest migration graph to obtain a cross-domain potential correlation score result between the user and the book; Sorting is performed based on the cross-domain potential association scoring results between users and books, and a cross-domain book recommendation list is obtained based on the sorting results.

[0005] In a second aspect, the present application also provides a cross-domain book recommendation system based on a neural network, comprising: An acquisition unit is used to acquire first information, wherein the first information includes the user's reading history data, historical search data, research direction label and historically published paper information A processing unit, configured to convert the first information into an initial interest topology structure and the user's interest evolution trend, and perform Markov clustering processing based on the converted result to obtain a cross-domain interest migration graph of the user; An embedding unit is used to input the user's cross-domain interest migration map and preset book text structured information into the extended probabilistic latent variable semantic model for processing, and perform semantic embedding based on the processed preliminary book topic distribution to obtain a semantic embedding expression of cross-domain books that integrates user interests; An analysis unit, configured to perform association analysis on the semantic embedding expression and the user's cross-domain interest migration graph to obtain a cross-domain potential association scoring result between the user and the book; The sorting unit is used to perform sorting based on the cross-domain potential association scoring results between users and books, and obtain a cross-domain book recommendation list based on the sorting results.

[0006] The beneficial effects of the present invention are: By constructing the initial topological structure of user reading behavior and research interests, and combining Markov clustering to generate a cross-domain interest migration graph, the structured book semantic information is further integrated, and multi-granular semantic expressions are obtained through the extended probabilistic latent variable model and Poisson embedding; the deep association modeling of users and books is then realized through gated recursive networks, heterogeneous attention autoencoders and exponential quantum graph networks to obtain potential scoring results; finally, through Bayesian sorting, restricted polynomial regression and Lévy flight mechanism joint sorting optimization, personalized and diverse recommendation results are generated. The present invention realizes the deep alignment of interest migration structure and semantic space, significantly improving the accuracy and practicality of cross-domain recommendations.

[0007] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or be understood by implementing the embodiments of the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0009] Figure 1A schematic diagram of a cross-domain book recommendation method based on a neural network according to an embodiment of the present invention; Figure 2 Schematic diagram of the structure of a cross-domain book recommendation system based on neural network described in an embodiment of the present invention.

[0010] In the figure: 701, acquisition unit; 702, processing unit; 703, embedding unit; 704, analysis unit; 705, sorting unit. DETAILED DESCRIPTION

[0011] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0012] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0013] Example 1

[0014] This embodiment provides a cross-domain book recommendation method based on neural network.

[0015] See also Figure 1 , the figure shows that the method includes step S1, step S2, step S3, step S4 and step S5.

[0016] Step S1, obtaining first information, wherein the first information includes the user's reading history data, historical search data, research direction label and historically published paper information; It can be understood that this step comprehensively collects the key feature information of users in academic and reading behaviors through the fusion of multi-source heterogeneous data, and constructs the basic data set required for user portraits. Specifically, reading history data reflects the user's past literature reading preferences in various academic platforms, e-book platforms or knowledge communities; historical search data reflects the user's interest orientation in actively obtaining information, such as keyword search records, query frequency and time distribution; research direction labels are usually composed of research field labels that users actively declare in scientific research communities or academic platforms, and can also be indirectly extracted through paper keywords, conference classifications, etc.; and historically published paper information covers the user's past scientific research results, including structured information such as titles, abstracts, keywords and subject areas. This step can break the recommendation limitations brought by a single interest tag, map the user's academic interest behavior into a dynamic representation basis with multi-dimensional features, and provide a highly accurate starting basis for subsequent interest modeling and migration. Especially in recommendation tasks involving user research direction migration or cross-border fusion of knowledge, this dynamic interest acquisition method driven by static labels is more in line with the interest evolution path of real users, and significantly improves the recommendation system's ability to grasp user intentions.

[0017] Step S2: converting the first information into an initial interest topology structure and the user's interest evolution trend, and performing Markov clustering processing based on the converted result to obtain a user's cross-domain interest migration graph; It can be understood that this step not only performs multi-scale clustering of static interests, but also accurately captures the behavioral trend of users diffusing from the main research field to the marginal field through dynamic modeling of the graph structure. Compared with traditional vector space models or classification methods, this graph model combined with Markov clustering provides stronger structural expression capabilities and migration pattern mining capabilities, improves the structural integrity and evolutionary interpretability of interest modeling, and effectively enhances the system's perception and prediction capabilities of users' long-term, cross-domain interest migration, laying the foundation for improving the time sensitivity and cross-border diversity of the recommendation algorithm. In this step, step S2 includes step S21, step S22, step S23 and step S24.

[0018] Step S21, performing self-organizing mapping network processing on the first information, wherein a two-dimensional topology preserving structure is constructed to perform nonlinear dimensionality reduction and interest cluster mapping on a high-dimensional interest vector containing an embedded representation of the first information, so as to obtain an initial interest topology structure of the user; It can be understood that this step first converts the user's reading history, search behavior, research direction, and published papers into a unified high-dimensional interest embedding vector through semantic encoding (BERT model). This vector contains both the content similarity at the semantic level and the contextual relationship between the temporal characteristics of user behavior and subject labels, and has rich information expression capabilities.

[0019] Secondly, this step introduces a self-organizing map network for unsupervised learning. The self-organizing map network constructs a neural network structure with two-dimensional grid nodes, in which each neuron represents the prototype of an interest cluster. Through competitive learning and neighborhood function update mechanism, the system finds the mapping position of each interest vector in the high-dimensional interest space, and maps vectors that are similar in semantics or behavior to adjacent or neighboring neurons in the two-dimensional topology, thereby achieving nonlinear dimensionality reduction while retaining the structural characteristics of the original interest distribution. This process allows different interest points to present clear clustering characteristics in the two-dimensional space, while maintaining the correlation and evolution direction between the original interests. The two-dimensional interest topological structure constructed in this step is a visual mapping result of the user's interest state. Each topological node represents a cluster interest center, and the adjacency relationship between nodes reflects the potential semantic transfer or behavior migration path. This structure provides a solid foundation for subsequent migration modeling and path identification based on graph methods.

[0020] Among them, the competition rules of the competitive learning mechanism are as follows:

[0021] in, is the weight of the node of interest closest to the input vector, is the learning rate, is the current user interest vector.

[0022] Among them, the update formula of the neighborhood function is as follows:

[0023] in, Representation Node In time step The weight of Representation Node In time step The weight of represents the learning rate, represents the weight calculated based on the neighborhood function, Represents the current user interest vector.

[0024] Step S22, performing discrete cosine transform processing on the initial interest topology structure of the user, wherein the long-term and short-term interest change trends are decoupled by extracting the frequency domain pattern of the time series weight signal in the same interest node trajectory, and the evolution trend of the user's interest among multiple preset research fields is obtained; It can be understood that this step takes each interest node in the user's initial interest topology as a processing unit, and extracts its corresponding time series weight signal, which is usually composed of the user's behavior frequency or interest intensity within a time window, such as the number of views, search clicks, and literature citation frequency of a certain research direction. These behaviors are quantified into time series vectors. Then, these sequences are transformed into the frequency domain using discrete cosine transform, and its mathematical expression is that the time series performs the following transformation:

[0025] in, Indicates Discrete cosine transform coefficients, represents the total length of the input signal, The original input signal The time domain signal of sampling points, represents the time index of the input signal, represents the frequency index after transformation, It is pi.

[0026] The core function of this process is to map the user's behavior trajectory in a certain interest direction from the time domain to the frequency domain, and then identify the interest change patterns represented by different frequency bands: low-frequency coefficients reflect long-term stable interest tendencies, while high-frequency coefficients reveal the user's recent interest mutations or short-term behavioral interference. By setting the frequency domain energy threshold (for example, selecting the first few coefficients with cumulative energy reaching 90%), the system can effectively compress the input dimension and remove high-frequency noise, retaining only representative interest change features.

[0027] Furthermore, this step embeds these frequency domain features into the multidimensional research field space, and constructs an evolution tensor based on the frequency feature similarity between nodes to capture the user's interest shift path between multiple preset research fields. Unlike traditional time sliding window averaging or first-order difference methods, DCT has stronger expressiveness and compression capabilities, and can more clearly distinguish between trend interests (such as gradual growth) and cyclical preferences (such as staged regression), and has significant advantages when facing cyclical alternation or implicit migration signals.

[0028] This step realizes the frequency domain modeling and decomposition of user interest temporal behavior, enabling the system to structurally decouple the interest evolution process and accurately divide the composition of long-term stable interests and short-term fluctuating interests. This not only improves the foresight of cross-domain recommendations, but also provides physically interpretable frequency labels for the subsequent construction of time weights in interest migration graphs, thereby achieving more refined and interpretable dynamic interest analysis.

[0029] Step S23: performing field label matching and nested community mining on the evolution trend of the user's interests among multiple preset research fields, wherein each interest node is mapped to a preset research field through an interest label attribution mechanism based on bidirectional label propagation to obtain a labeled interest evolution graph; It can be understood that this step first introduces a domain label matching mechanism, which is based on a two-way label propagation method to align user interest nodes with preset research field labels. The specific operation is: construct a two-layer graph structure, one layer is the nodes in the user interest evolution graph, and the other layer is a preset research field label set (specifically from the Web of Science subject classification), and then iterate through the label propagation algorithm. On the one hand, the known field labels are propagated to the interest nodes, and on the other hand, the labels of high-weight nodes in the interest trajectory are back-propagated to the research field label layer, thereby realizing two-way attribution judgment. This mechanism not only utilizes the topological similarity of the graph structure, but also combines the dynamics of node weights during time evolution.

[0030] After completing the mapping of interest tags, the nested community mining algorithm is used to further identify the intrinsic structural relationships in the interest graph. This process first divides the user interest graph into hierarchically nested sub-communities, each of which corresponds to a research field or a cross-field combination, and presents the user's interest migration pattern of deepening or jumping across layers in these fields. Community division not only considers the evolutionary similarity and temporal consistency between nodes, but also integrates the frequency domain features extracted in the previous step, making the nested structure more consistent in time and semantics.

[0031] The unique effect of this step is to transform the original continuous interest evolution data into a semantically clear and clearly structured graph representation, so that the subsequent recommendation model can perform multi-scale learning based on the "interest-domain" structure; at the same time, the two-way label propagation avoids the inaccurate attribution problem caused by sparse labels or domain intersections in traditional one-way matching, and the introduction of the nested community structure effectively improves the hierarchy and interpretability of interest structure modeling. The technical effect of this step is reflected in the precise alignment of the user's interest evolution path with the domain knowledge structure, providing a semantic basis and structural support for precise matching and reasoning in subsequent cross-domain recommendations.

[0032] Step S24, the annotated interest evolution graph is subjected to Markov clustering processing, wherein, by constructing a state transition matrix between user interest nodes and performing random walk simulation, potential inter-domain interest migration paths are identified and summarized to obtain the user's cross-domain interest migration graph.

[0033] It can be understood that this step first constructs a state transfer matrix based on the state transfer relationship between the user's interest nodes. In the graph, each node represents a specific interest point (such as a research field or a specific topic), and the edge between the nodes represents the potential transition probability from one interest node to another. At this time, the elements of the state transfer matrix reflect the probability of the user transferring from a certain interest state (research field, topic, etc.) to another interest state. This transition probability not only considers the direct connection relationship between nodes, but also adjusts the weight according to the user's historical behavior (such as clicks, reading, retrieval, etc.), so that the probability distribution is more in line with the actual interest migration trajectory.

[0034] Next, perform a random walk simulation. The random walk algorithm simulates the process of users starting from a certain interest node and randomly walking to other nodes, and repeats multiple iterations. During the random walk, based on the state transition matrix, each step is made according to the transition probability between nodes. The purpose of the random walk simulation is to discover and identify potential cross-domain interest migration paths through multiple simulations. These paths reflect the migration rules of user interests from one field to another. By simulating different interest paths, it is possible to identify which fields have more frequent migrations and which fields have slower or less obvious user interest transfers.

[0035] After completing the random walk simulation, the interest nodes in the interest graph are divided into several groups based on the clustering algorithm in the Markov process (K-means clustering is used in this step). These groups can effectively reveal the user's cross-domain interest migration map, that is, the potential rules and preferences of users migrating between different research fields. For example, some users may focus on a certain subject area for a certain period of time and then turn to related or adjacent fields, while other users may show a more dispersed interest migration path.

[0036] This step uses the Markov clustering method to reveal the user's potential interest migration pattern from a large-scale interest graph and provide a more refined interest trajectory analysis. Unlike traditional keyword-based recommendation systems, Markov clustering introduces a dynamic transfer process in the model, making the identification of cross-domain interest migration more consistent with actual user behavior patterns. At the same time, the random walk process helps the system overcome the limitations of a single path through a large number of random simulations, and can better explore the possible interest transfer direction of users, providing strong data support for subsequent personalized recommendations.

[0037] Step S3: input the user's cross-domain interest migration map and the preset book text structured information into the extended probabilistic latent variable semantic model for processing, and perform semantic embedding based on the processed preliminary book topic distribution to obtain a semantic embedding expression of cross-domain books that integrates the user's interests; It can be understood that this step combines two seemingly independent information sources, the cross-domain interest migration graph and the book theme, and uses the extended probabilistic latent variable semantic model to model the relationship between user interests and book texts. This fusion method not only improves the accuracy of interests, but also can handle cross-domain interest migration, making up for the limitation that traditional recommendation systems can only make recommendations within a single field. In this step, step S3 includes step S31, step S32, and step S33.

[0038] Step S31, the user's cross-domain interest migration map and the preset structured book information are processed by an extended probabilistic latent variable semantic model, wherein the book content, chapter structure and domain label are jointly modeled through a preset hierarchical Bayesian inference mechanism to obtain a multi-granular topic distribution of each book; It can be understood that this step first builds a joint input set that integrates semantic features and interest evolution features based on the user's cross-domain interest migration graph, combined with the transfer path and evolution trajectory between interest nodes. In this set, each interest node not only represents a research field or topic, but also carries the weight given by the user's historical behavior, such as reading frequency, duration or interaction density under a specific topic. These features are quantitatively expressed through the interest state vector to drive subsequent topic modeling.

[0039] Next, the above set and the preset structured book information are input into the extended probabilistic latent variable semantic model. In this model, multi-layer semantic modeling is performed by introducing a hierarchical Bayesian inference mechanism. This mechanism constructs a joint probability distribution between book content, chapter structure and domain labels, and introduces latent variables in each layer to represent hidden topics. Among them, this step uses Gibbs sampling to perform parameter inference, so as to model the topics that may be contained in each book at different levels (such as the whole book, chapter, paragraph). In the process of topic extraction, the semantic connection between chapters maintains coherence through the context window mechanism, and the label information is used as a priori to guide the model to optimize the attribution of topics.

[0040] Then, based on the model inference results, a multi-granular topic distribution is generated for each book, including the dominant semantics at the book level, local topics at the chapter level, and fine-grained semantic clusters of nested subtopics. These topic distributions not only reflect the semantic diversity of the book itself, but can also be semantically compared with the research field preferences shown by users in the interest migration path, laying a solid foundation for subsequent semantic embedding and matching recommendations.

[0041] Step S32: input the multi-granularity topic distribution into a preset Poisson embedding model for processing, wherein the preset Poisson embedding model generates a low-rank sparse embedding vector by simulating the high-dimensional Poisson counting characteristics of each book in the semantic space, thereby obtaining a sparse semantic expression that matches the interest evolution structure; It can be understood that this step first takes the multi-granular topic distribution of each book as input to construct a set of high-dimensional semantic count vectors. These count vectors are based on the word frequency of each topic at each granularity level, reflecting the "semantic strength" characteristics of the book in the entire semantic space. Since multi-granular topics have a hierarchical structure, the high-dimensional vector not only contains the distribution frequency of the main topic, but also includes the count information of local sub-topics and marginal topics, forming a sparse and structured basis for information expression.

[0042] Next, the high-dimensional semantic count information is input into the preset Poisson embedding model for processing. The Poisson embedding model is a type of probabilistic graphical model with Poisson distribution as the generation mechanism, which is suitable for processing high-dimensional discrete count data. In actual operation, it is assumed that the semantic count of each book is a Poisson distribution sample generated by a potential low-dimensional embedding vector. Specifically, the model introduces a low-rank semantic space matrix to factorize the original high-dimensional semantic count matrix, and decouples the latent variable expression of the basic topic dimension and the semantic features of the book in the semantic space. The embedding process uses maximum likelihood estimation to solve the model parameters to ensure that the generated embedding vector has lower dimension and stronger sparsity while maintaining the original semantic information. Among them, Through this process, we can obtain low-dimensional sparse embedding vectors corresponding to each book. These vectors not only maintain consistency with the original topic distribution in the semantic space, but also have good structural matching characteristics for the semantic migration relationship formed by the interest evolution path. The sparse semantic expression finally output not only compresses redundant semantic dimensions, but also highlights the semantic core areas with potential migration value, so that the subsequent recommendation model can more efficiently capture the deep matching relationship between user interests and book content.

[0043] Step S33: Adaptively perform relevant entropy embedding processing based on the sparse semantic expression and the user's jump path in the interest transfer graph, wherein a personalized cross-domain semantic embedding expression aligned with the interest transfer pattern is obtained by constructing a semantic covariance matrix weighted by the user path and minimizing its information entropy offset.

[0044] It can be understood that this step first constructs a user path weighted semantic covariance matrix based on the sparse semantic expression vector obtained in the previous step and the user's jump path information in the interest migration graph. This covariance matrix not only captures the statistical correlation between different books in various dimensions in the semantic embedding space, but also assigns personalized weights to the semantic relationship between each pair of books by integrating the jump trajectories in the interest migration graph. Specifically, for the book pairs involved in the user's jump from one research field to another, a higher path importance weight will be assigned based on indicators such as jump frequency, path length, and node similarity, so that the covariance matrix is ​​closer to the real interest evolution dynamics.

[0045] Subsequently, based on the weighted semantic covariance structure, the correlation entropy embedding mechanism is introduced for optimization. The core idea of ​​this method is to minimize the information entropy offset between the semantic covariance generated by the embedding representation guided by the user path and the expected distribution, ensuring that the embedding space is not only sparse but also truly reflects the semantic connection driven by interest migration. The optimization process uses the KL divergence entropy in information theory as a metric. By iteratively minimizing the uncertainty in the embedding space, the semantic vectors involved in the user's interest transfer are made more concentrated and discriminative, thereby constructing a semantic representation that is highly consistent with the evolution trajectory of individual interests.

[0046] The final output is a set of personalized cross-domain semantic embedding expressions that align with the user's interest migration pattern. These embeddings not only retain the advantages of semantic sparsity and low-dimensional expression, but also structurally accurately adapt to the user's transfer path in the interest graph, forming a deep expression pattern of user-semantic-domain synergy.

[0047] Step S4, performing association analysis on the semantic embedding expression and the user's cross-domain interest migration graph to obtain a cross-domain potential association scoring result between the user and the book; It can be understood that this step not only accurately describes the potential interest association between users and books in different fields, but also effectively explores the potential needs that have not been explicitly expressed in the interest evolution path by integrating semantic embedding with the interest migration graph structure. Compared with traditional content-based or collaborative filtering recommendation methods, this method can capture the deep dynamic evolution of user interests at both semantic and structural levels, greatly improving the accuracy and personalization of cross-domain recommendations. In this step, step S4 includes step S41, step S42, step S43 and step S44.

[0048] Step S41, performing gated recursive hybrid network processing according to the semantic embedding expression and the user's cross-domain interest migration graph, wherein a gating mechanism is introduced to control the information flow of the semantic embedding sequence and the interest graph jump sequence, and the interest diffusion characteristics are captured at the long-term and short-term levels to obtain a hybrid interest evolution trajectory between the user and the book; It can be understood that this step first uses the user's personalized sparse semantic embedding sequence as the input main channel, and time-aligns it with the node jump sequence contained in the user interest migration graph (i.e., the evolution path of user interests between various fields), and ensures that the two types of information have a consistent evolution order through timestamps or event sequence numbers. Then, a dual-channel recursive network structure with a gating mechanism is introduced: the main channel uses a GRU (Gated Recurrent Unit) unit with an update gate and a reset gate to model the semantic embedding information, and the secondary channel uses another GRU structure to model the interest jump sequence, and the two are fused through a cross-gating layer.

[0049] In the cross-gating layer, a trainable gating weight vector is used to dynamically weight the hidden states of the main channel and the secondary channel. The weight distribution is adjusted according to the joint index of the semantic context change rate and the interest jump frequency in the current time step, so as to ensure that the model can dynamically switch between semantic dominance and structure dominance. This design can effectively avoid the problem of information loss or excessive memory in long sequence modeling, and at the same time achieve rapid response of information at the mutation point of interest evolution (such as jumping from one research field to another).

[0050] In order to further enhance the model's ability to perceive interest diffusion characteristics at different time scales, the gated recursive hybrid network introduces a hierarchical recursive structure in its upper layer, that is, to capture the user's long-term interest evolution trend in a long time window and to capture local hot spot interest fluctuations in a short time window. This hierarchical capture mechanism calculates the recursive hidden states at two different time scales in parallel and generates the final hybrid interest evolution trajectory representation through attention weighted aggregation.

[0051] Step S42, performing heterogeneous attention autoencoder processing according to the hybrid interest evolution trajectory, wherein by constructing a multi-channel attention map based on preset user behavior types, the distinguishability of interest expression is enhanced in the dimension of behavior heterogeneity, and a structurally enhanced user-book interaction representation is obtained; It can be understood that this step first classifies and maps the behavior nodes corresponding to the user at each time step based on the hybrid interest evolution trajectory obtained in the previous stage, and constructs a multi-channel heterogeneous attention map. In this map, each channel represents a user behavior type, and the edge weights between users and books under different channels reflect the importance of the behavior in the semantic context. The nodes are represented by the semantic embedding vectors and interest jump sequences in the hybrid trajectory.

[0052] Next, an independent attention scoring network is introduced for each behavior type through a heterogeneous attention mechanism to calculate the importance coefficient of the user node to the adjacent book node in the current behavior channel. The calculation of the attention score not only considers the intensity of the user's historical behavior (such as frequency and duration), but also introduces multiple factors such as domain label consistency, semantic embedding similarity, and contextual weight of domain jumps, and constructs an attention weight tensor with behavioral semantic perception capabilities. In the autoencoder structure, the attention weight of each channel is used to guide the aggregation process of adjacent node information in the encoding stage, so that the final embedding of each user node not only retains its original semantic features, but also integrates its interactive preference features under different behavior channels. Among them, the process of calculating the importance coefficient of the user node to the adjacent book node in the current behavior channel is to first input two parallel RNN structures into the user interest sequence at the same time:

[0053]

[0054] in, For the The short-term interest hidden state of the moment, Indicates The user interest input features at the moment, Indicates The short-term interest hidden state of the moment, Indicates The long-term interest hidden state of the moment, Indicates The long-term interest hidden state of the moment, represents a recursive neural network structure for modeling short-term interest dynamics. Represents a recurrent neural network structure for modeling long-term interest evolution.

[0055] Then, the importance coefficient is calculated by weighted fusion, where the calculation formula of the importance coefficient is:

[0056] in, represents the importance coefficient, Indicates The attention score associated with each item, represents the number of all items considered, It is a sum index variable, indicating the The score of the candidate element.

[0057] After encoding, the user embedding generated by each channel will be sent to the decoder for reconstruction. By minimizing the difference between the original graph structure (such as the connection strength of the user-book edge) and the reconstructed graph, the model is driven to learn an embedding representation that is more sensitive to the user's real behavior preferences. Finally, the reconstructed vectors of each channel are weighted fused to form a structurally enhanced user-book interaction representation.

[0058] Step S43: performing exponential quantum graph network processing according to the structurally enhanced user-book interaction representation, constructing a quantum state extension path graph with non-commutative adjacency characteristics, introducing an exponential decay coefficient and a quantum state coupling mechanism in the node propagation process, and obtaining a cross-domain semantic coupling score vector between the user and the book; It can be understood that in this step, a "quantum state extension path diagram" is first constructed based on the structurally enhanced user-book interaction representation. In this diagram, the user and book nodes are regarded as quantum systems in a certain semantic superposition state, and the edges between the nodes no longer represent only adjacency relationships. Instead, an adjacency weight matrix with non-commutative characteristics is introduced, allowing semantic propagation to proceed under an asymmetric structure, reflecting the user's non-uniform perception of different books.

[0059] In the node propagation mechanism, quantum graph networks introduce two core mechanisms: one is the exponential decay coefficient, which is used to simulate the information decay phenomenon caused by the increase of path depth during the propagation of semantics in the graph. That is, the longer the propagation path, the smaller the coupling effect between nodes. The coefficient is in the form of ,in, represents the path length, The first is the attenuation intensity factor, which is used to control the suppression intensity of long-term dependence; the second is the quantum state coupling mechanism, which is used to capture the degree of coupling between the user's interest state and the book's semantic state in the superposition space. The second is the quantum state coupling mechanism, which is used to capture the degree of coupling between the user's interest state and the book's semantic state in the superposition space. This mechanism constructs a coupling tensor based on the Heisenberg quantum coupling model, quantifies the intensity and directionality of the mutual influence between nodes, and allows interference terms and superposition interference between multiple interest topics, which is especially suitable for expressing users' fuzzy interest preferences between interdisciplinary books.

[0060] In the iterative propagation process of the graph network, each round of information update is not only affected by the characteristics of adjacent nodes, but also by the attenuation correction of path weights and the interaction of quantum states. Finally, by aggregating the quantum state interaction information between all nodes, a cross-domain semantic coupling score vector is generated. This score vector can reflect the implicit interest coupling strength of users in different fields and different themes of books in a fine-grained manner, thereby providing a more discriminating feature basis for downstream recommendation sorting.

[0061] Step S44, performing potential scoring result estimation processing according to the cross-domain semantic coupling scoring vector, wherein the scoring matrix is ​​soft-filled and boundary convergence optimized by constructing a collaborative reasoning mechanism with a migration normalization item to obtain the cross-domain potential association scoring result between the user and the book.

[0062] It can be understood that this step first constructs a user-book rating matrix, in which the blank items indicate that the user has not yet interacted with the book. The system uses the coupled rating vector obtained in the previous step as the initial soft rating input to guide the collaborative filtering framework to fill it in. In order to improve the robustness of migration, the collaborative reasoning mechanism introduces a migration normalization term, which normalizes the rating value according to the jump strength of the user in the interest migration graph to prevent a certain interest source from over-dominantly dominating the rating generation process. The normalization function form is as follows:

[0064] in, is the normalized potential score, Score semantic coupling, For users in the graph from the node To other nodes Jump strength.

[0065] Subsequently, the system uses collaborative matrix soft filling methods (such as matrix completion based on kernel regression, low-rank filling with graph regularization, etc.) to optimize and solve the normalized rating matrix. In the solution process, a boundary convergence optimization mechanism is introduced to ensure that the rating prediction does not fall into overfitting or propagation distortion by setting dynamically adjusted convergence thresholds and local loss constraints. Especially when dealing with cold start users or books with few interactions, the boundary mechanism can stabilize the rating output by integrating the semantic coupling features of neighboring users or books.

[0066] Ultimately, the cross-domain potential association scoring results output by this step not only fully retain the user's preference pattern on the interest evolution path, but also take into account the global consistency of the scoring matrix structure, providing an accurate and explanatory numerical basis for subsequent book ranking recommendations.

[0067] Step S5: performing sorting based on the cross-domain potential association scoring results between users and books, and obtaining a cross-domain book recommendation list based on the sorting results.

[0068] It can be understood that this step improves the personalized expression ability and domain expansion ability of the recommended content while ensuring the accuracy of semantic coupling through multi-factor sorting and dynamic weighting mechanism, so that the recommendation system has stronger interpretability, adaptability and user knowledge gain potential. In this step, step S5 includes step S51, step S52 and step S53.

[0069] Step S51: performing Bayesian ranking processing according to the potential scoring results, estimating the preference level of each book on the basis of maximizing the posterior probability of the scoring distribution by constructing a hierarchical ranking structure model of the scoring confidence, and obtaining a preliminary ranking list of candidate books; It can be understood that this step first regards the rating result of each user-book pair as an observation sample generated from a hidden preference level, and the preference level itself is a hidden variable to be estimated, belonging to a discrete level space (such as "very interested", "average", "not interested", etc.). The system further introduces a confidence factor for each rating value, which is comprehensively estimated by indicators such as the semantic coupling degree of the rating, the number of interest jump paths, and semantic sparsity, as the basis for modeling the prior uncertainty of the rating accuracy.

[0070] The confidence factor is:

[0072] in, represents the confidence factor of the score, Represents the similarity between user interests and book content, Indicates the frequency of user interest migration, Indicates the sparseness of the match between interests and book topics.

[0073] In the model building phase, the system uses a hierarchical Bayesian modeling strategy to model the rating generation process into a three-layer structure: the first layer models the user's preference distribution prior (Beta distribution is used in this step); the second layer models the variance structure of the rating confidence, and the third layer models the observed value of the actual rating, which is regarded as a sampling result generated by noise perturbation based on the preference level. On this basis, the system uses a variational inference method to maximize the posterior probability of the rating of each candidate book, thereby inferring its ranking level under the current user's interest preference.

[0074] Different from the traditional point value-based sorting, Bayesian ranking fully considers the confidence differences and potential uncertainties of the ratings during the estimation process, and has stronger fault tolerance and ranking stability. It is especially suitable for dealing with sparse ratings or semantic drift in cross-domain recommendation systems, ensuring that the initial ranking list is more in line with the actual preferences of users.

[0075] Step S52: performing restricted minimum polynomial regression processing on the preliminary ranking list of the candidate books, wherein a low-order nonlinear mapping model between the ranking sequence and the semantic similarity function is established, and a structural smoothing penalty term is added to correct the extreme ranking deviation, so as to obtain a ranking trend result after semantic calibration; It can be understood that this step uses the sorting sequence (i.e., the preliminary sorting list of candidate books) as the independent variable and the semantic similarity function (the similarity between the book content and the user's interests calculated by multidimensional semantic embedding, etc.) as the dependent variable to construct a low-order polynomial regression model. The purpose of this regression model is to capture the nonlinear mapping law between the preliminary sorting list and the semantic similarity by fitting the relationship between them. For example, some books may be ranked at the top of the preliminary sorting list due to their high scores, but their semantic similarity is low. The regression model will adjust their sorting position according to this deviation to ensure that the final sorting is more consistent with the user's interests and semantic matching.

[0076] Among them, the low-order polynomial regression model is as follows:

[0077] in, represents the dependent variable, represents the independent variable, represents the regression coefficient, represents the error term.

[0078] Secondly, in the regression process, this step introduces a structural smoothing penalty term to correct extreme deviations in the ranking. The role of this penalty term is to limit the complexity of the regression model, avoid overfitting the noise or outliers in the data, and ensure that the ranking adjustment process is smoothly transitioned within the effective range. For example, when the initial ranking of some books has a very low match with the semantic similarity, the structural smoothing penalty term will moderately control its ranking changes to prevent its extreme position deviation in the final ranking, thereby avoiding the imbalance of the overall ranking due to individual outliers.

[0079] Step S53, performing Lévy flying sorter processing according to the sorting trend results after the semantic calibration, wherein the Lévy flying sorter is used to simulate the non-Gaussian jump distribution characteristics of the user's cognitive path, randomly insert book nodes with a long distance in the feature space, and obtain a cross-domain book recommendation list with enhanced diversity.

[0080] It can be understood that the Lévy flight in this step is a random process, which is characterized by the jump length obeying the Lévy distribution, which has a heavy-tailed characteristic, meaning that it can simulate long-distance jumps or jumps of non-Gaussian nature. Compared with traditional Gaussian random processes, Lévy flights can produce longer jumps, which enables it to better represent the long-distance jumps in the user's interest exploration behavior when simulating the user's cognitive path, that is, the user may suddenly become interested in some books that are far away in the feature space. This behavior often does not conform to the conventional neighbor recommendation model, but reflects the exploratory characteristics of human cognition. For example, after being exposed to knowledge in multiple related fields, the user's interests may span multiple different fields or topics, thereby generating a broader preference.

[0081] In specific operations, the Lévy flight sorter first processes the sorting trend results after semantic calibration. Based on the characteristics of Lévy flight, it simulates the process of users jumping from a known interest field to a farther interest node, and adjusts the order of books in the recommendation list. In this way, the system will randomly insert book nodes that are far away in the feature space. These books may not be the most relevant to the current user's interest field, but by simulating long-distance jumps in the cognitive process, they can bring new interest inspiration and knowledge expansion. Such insertion can increase the diversity of the recommendation list, allowing users not only to be exposed to closely related books, but also to discover cross-domain knowledge and information, thereby enriching their reading experience.

[0082] Embodiment 2: like Figure 2 As shown, this embodiment provides a cross-domain book recommendation system based on neural network, see Figure 2 The system includes an acquisition unit 701 , a processing unit 702 , an embedding unit 703 , an analysis unit 704 and a sorting unit 705 .

[0083] The acquisition unit 701 is used to acquire first information, wherein the first information includes the user's reading history data, historical search data, research direction label and historically published paper information The processing unit 702 is used to convert the first information into an initial interest topology structure and the user's interest evolution trend, and perform Markov clustering processing based on the converted result to obtain a user's cross-domain interest migration graph; The embedding unit 703 is used to input the user's cross-domain interest migration map and the preset book text structured information into the extended probability latent variable semantic model for processing, and perform semantic embedding based on the processed preliminary book topic distribution to obtain a semantic embedding expression of the cross-domain book that integrates the user's interests; An analyzing unit 704 is used to perform association analysis on the semantic embedding expression and the user's cross-domain interest migration graph to obtain a cross-domain potential association scoring result between the user and the book; The sorting unit 705 is used to perform sorting based on the cross-domain potential association scoring results between users and books, and obtain a cross-domain book recommendation list based on the sorting results.

[0084] It should be noted that, regarding the system in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0086] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A cross-domain book recommendation method based on neural network, characterized in that: include: Acquire first information, wherein the first information includes the user's reading history data, historical search data, research direction label, and historically published paper information; The first information is converted into an initial interest topology structure and the user's interest evolution trend, and Markov clustering processing is performed based on the converted result to obtain the user's cross-domain interest migration map; The user's cross-domain interest migration map and the preset book text structured information are input into the extended probabilistic latent variable semantic model for processing, and semantic embedding is performed based on the processed preliminary book topic distribution to obtain the semantic embedding expression of cross-domain books that integrates user interests; Performing correlation analysis on the semantic embedding expression and the user's cross-domain interest migration graph to obtain a cross-domain potential correlation score result between the user and the book; Sorting is performed based on the cross-domain potential association scoring results between users and books, and a cross-domain book recommendation list is obtained based on the sorting results.

2. The cross-domain book recommendation method based on neural network according to claim 1 is characterized in that: The first information is converted into an initial interest topology structure and a user's interest evolution trend, and Markov clustering processing is performed based on the converted result, including: Processing the first information through a self-organizing mapping network, wherein a high-dimensional interest vector including an embedded representation of the first information is subjected to nonlinear dimensionality reduction and interest cluster mapping by constructing a two-dimensional topology preserving structure to obtain an initial interest topology structure of the user; The initial interest topology of the user is processed by discrete cosine transformation, wherein the long-term and short-term interest change trends are decoupled by extracting the frequency domain pattern of the time series weight signal in the same interest node trajectory, and the evolution trend of the user's interest among multiple preset research fields is obtained; The evolution trend of the user's interests among multiple preset research fields is processed by field label matching and nested community mining, wherein each interest node is mapped to a preset research field through an interest label attribution mechanism based on two-way label propagation to obtain a labeled interest evolution map; The annotated interest evolution graph is subjected to Markov clustering processing, in which the state transition matrix between user interest nodes is constructed and random walk simulation is performed to identify and summarize potential inter-domain interest migration paths, thereby obtaining the user's cross-domain interest migration graph.

3. The cross-domain book recommendation method based on neural network according to claim 1 is characterized in that: The user's cross-domain interest migration map and the preset book text structured information are input into the extended probabilistic latent variable semantic model for processing, and semantic embedding is performed based on the processed preliminary book topic distribution, including: The user's cross-domain interest migration graph and the preset structured book information are processed by an extended probabilistic latent variable semantic model. The book content, chapter structure and domain label are jointly modeled through a preset hierarchical Bayesian inference mechanism to obtain a multi-granular topic distribution for each book. Inputting the multi-granularity topic distribution into a preset Poisson embedding model for processing, wherein the preset Poisson embedding model generates a low-rank sparse embedding vector by simulating the high-dimensional Poisson counting characteristics of each book in the semantic space, thereby obtaining a sparse semantic expression that matches the interest evolution structure; Adaptive correlation entropy embedding processing is performed based on the sparse semantic expression and the user's jump path in the interest transfer graph, wherein a personalized cross-domain semantic embedding expression aligned with the interest transfer pattern is obtained by constructing a semantic covariance matrix weighted by the user path and minimizing its information entropy offset.

4. The cross-domain book recommendation method based on neural network according to claim 1 is characterized in that: Performing correlation analysis on the semantic embedding expression and the user's cross-domain interest migration graph, including: A gated recursive hybrid network is processed based on the semantic embedding expression and the user's cross-domain interest migration graph, wherein a gating mechanism is introduced to control the information flow of the semantic embedding sequence and the interest graph jump sequence, and the interest diffusion characteristics are captured at the long-term and short-term levels to obtain a hybrid interest evolution trajectory between the user and the book; Performing heterogeneous attention autoencoder processing according to the hybrid interest evolution trajectory, wherein by constructing a multi-channel attention map based on preset user behavior types, the distinguishability of interest expression is enhanced in the dimension of behavior heterogeneity, and a structurally enhanced user-book interaction representation is obtained; An exponential quantum graph network is processed based on the structurally enhanced user-book interaction representation. By constructing a quantum state extension path graph with non-commutative adjacency characteristics, an exponential decay coefficient and a quantum state coupling mechanism are introduced in the node propagation process to obtain a cross-domain semantic coupling score vector between the user and the book. The potential scoring result estimation process is performed according to the cross-domain semantic coupling scoring vector, wherein the scoring matrix is ​​soft-filled and boundary convergence optimized by constructing a collaborative reasoning mechanism with a migration normalization item, so as to obtain the cross-domain potential association scoring result between the user and the book.

5. The cross-domain book recommendation method based on neural network according to claim 1 is characterized in that: Sorting is performed based on the cross-domain potential correlation score results between users and books, and a cross-domain book recommendation list is obtained based on the sorting results, including: Performing Bayesian ranking processing according to the potential rating results, estimating the preference level of each book on the basis of maximizing the posterior probability of the rating distribution by constructing a hierarchical ranking structure model of rating confidence, and obtaining a preliminary ranking list of candidate books; Performing restricted minimum polynomial regression processing on the preliminary ranked list of candidate books, wherein a low-order nonlinear mapping model between the ranked sequence and the semantic similarity function is established, and a structural smoothing penalty term is added to correct the extreme ranking deviation, thereby obtaining a ranking trend result after semantic calibration; The Lévy flying sorter is processed according to the sorting trend result after the semantic calibration, wherein the Lévy flying sorter is used to simulate the non-Gaussian jump distribution characteristics of the user's cognitive path, randomly insert book nodes with a long distance in the feature space, and obtain a cross-domain book recommendation list with enhanced diversity.

6. A cross-domain book recommendation system based on neural network, characterized in that: include: An acquisition unit, configured to acquire first information, wherein the first information includes a user's reading history data, historical search data, research direction labels, and historically published paper information; A processing unit, configured to convert the first information into an initial interest topology structure and the user's interest evolution trend, and perform Markov clustering processing based on the converted result to obtain a cross-domain interest migration graph of the user; An embedding unit is used to input the user's cross-domain interest migration map and preset book text structured information into the extended probabilistic latent variable semantic model for processing, and perform semantic embedding based on the processed preliminary book topic distribution to obtain a semantic embedding expression of cross-domain books that integrates user interests; An analysis unit, configured to perform association analysis on the semantic embedding expression and the user's cross-domain interest migration graph to obtain a cross-domain potential association scoring result between the user and the book; The sorting unit is used to perform sorting based on the cross-domain potential association scoring results between users and books, and obtain a cross-domain book recommendation list based on the sorting results.

7. The cross-domain book recommendation system based on neural network according to claim 6 is characterized in that: The processing unit comprises: A first processing subunit is used to perform self-organizing mapping network processing on the first information, wherein a high-dimensional interest vector containing an embedded representation of the first information is subjected to nonlinear dimensionality reduction and interest cluster mapping by constructing a two-dimensional topology preserving structure to obtain an initial interest topology structure of the user; A second processing subunit is used to perform discrete cosine transform processing on the initial interest topology structure of the user, wherein the long-term and short-term interest change trends are decoupled by extracting the frequency domain mode of the time series weight signal in the same interest node trajectory, and the evolution trend of the user's interest among multiple preset research fields is obtained; The third processing sub-unit is used to perform field label matching and nested community mining on the evolution trend of the user's interests among multiple preset research fields, wherein each interest node is mapped to a preset research field through an interest label attribution mechanism based on bidirectional label propagation to obtain a labeled interest evolution map; The fourth processing sub-unit is used to perform Markov clustering processing on the annotated interest evolution graph, wherein, by constructing a state transition matrix between user interest nodes and performing random walk simulation, potential inter-domain interest migration paths are identified and summarized to obtain the user's cross-domain interest migration graph.

8. The cross-domain book recommendation system based on neural network according to claim 6, characterized in that: The embedding unit comprises: The first embedding subunit is used to process the user's cross-domain interest migration map and the preset structured book information into an extended probabilistic latent variable semantic model, wherein the book content, chapter structure and domain label are jointly modeled through a preset hierarchical Bayesian inference mechanism to obtain a multi-granular topic distribution of each book; A second embedding subunit is used to input the multi-granularity topic distribution into a preset Poisson embedding model for processing, wherein the preset Poisson embedding model generates a low-rank sparse embedding vector by simulating the high-dimensional Poisson counting characteristics of each book in the semantic space, thereby obtaining a sparse semantic expression matching the interest evolution structure; The third embedding subunit is used to perform adaptive correlation entropy embedding processing according to the sparse semantic expression and the user's jump path in the interest transfer graph, wherein a personalized cross-domain semantic embedding expression aligned with the interest transfer pattern is obtained by constructing a semantic covariance matrix weighted by the user path and minimizing its information entropy offset.

9. The cross-domain book recommendation system based on neural network according to claim 6, characterized in that: The analysis unit comprises: The first analysis subunit is used to perform gated recursive hybrid network processing according to the semantic embedding expression and the user's cross-domain interest migration graph, wherein the information flow of the semantic embedding sequence and the interest graph jump sequence is controlled by introducing a gating mechanism, and the interest diffusion characteristics are captured at the long-term and short-term levels to obtain a hybrid interest evolution trajectory between the user and the book; A second analysis subunit is used to perform heterogeneous attention autoencoder processing according to the hybrid interest evolution trajectory, wherein by constructing a multi-channel attention map based on preset user behavior types, the distinguishability of interest expression is enhanced in the dimension of behavior heterogeneity, and a structure-enhanced user-book interaction representation is obtained; The third analysis subunit is used to perform exponential quantum graph network processing according to the structurally enhanced user-book interaction representation, by constructing a quantum state extension path graph with non-commutative adjacency characteristics, introducing an exponential decay coefficient and a quantum state coupling mechanism in the node propagation process, and obtaining a cross-domain semantic coupling score vector between the user and the book; The fourth analysis subunit is used to estimate and process potential rating results based on the cross-domain semantic coupling rating vector, wherein the rating matrix is ​​soft-filled and boundary convergence optimized by constructing a collaborative reasoning mechanism with a migration normalization item to obtain the cross-domain potential association rating result between the user and the book.

10. The cross-domain book recommendation system based on neural network according to claim 6, characterized in that: The sorting unit comprises: A first sorting subunit is used to perform Bayesian ranking processing according to the potential scoring results, estimate the preference level of each book on the basis of maximizing the posterior probability of the scoring distribution by constructing a hierarchical ranking structure model of the scoring confidence, and obtain a preliminary ranking list of candidate books; The second sorting subunit is used to perform restricted minimum polynomial regression processing based on the preliminary sorting list of candidate books, wherein a low-order nonlinear mapping model between the sorting sequence and the semantic similarity function is established, and a structural smoothing penalty term is added to correct the extreme sorting deviation, so as to obtain the sorting trend result after semantic calibration; the third sorting subunit is used to perform Lévy flying sorter processing based on the sorting trend result after semantic calibration, wherein the non-Gaussian jump distribution characteristics of the user cognitive path are simulated by the Lévy flying sorter, and book nodes with a long distance in the feature space are randomly inserted to obtain a cross-domain book recommendation list with enhanced diversity.

Images