An intelligent recommendation system for library document procurement based on reader borrowing behavior

Abstract

The application relates to the technical field of library information management, in particular to a kind of intelligent recommendation system for library literature procurement based on reader borrowing behavior, and discloses a system named "intelligent recommendation system for library literature procurement based on reader borrowing behavior", which aims to solve the problems of lagging procurement mode and lack of foresight in new book evaluation. The system comprises the following modules: data acquisition and preprocessing, reader behavior time series modeling, discipline evolution dynamics analysis, future demand feature prediction, new book semantic correlation matching and intelligent generation of procurement decision. The system constructs behavior sequence through a sliding window, predicts future discipline demand using a Markov model, and calculates the semantic matching degree of new books based on deep learning. The application realizes accurate prediction of demand drift, significantly improves the foresight of procurement, resource allocation efficiency and the turnover rate of library collection.

Description

Technical Field

[0001] This invention belongs to the field of library information management technology, specifically relating to an intelligent recommendation system for library document procurement based on reader borrowing behavior. Background Technology

[0002] With the dramatic changes in the global information environment and the profound evolution of higher education and scientific research models, libraries, as hubs for knowledge storage, dissemination, and service, have undergone a strategic transformation from simple "document preservation" to "precise resource support" and "decision-making assistance." In the current scientific research and education system, the quality of document resource development directly relates to the breadth and depth of discipline construction and is a key indicator of an academic institution's core competitiveness. However, with the significantly accelerated pace of knowledge iteration in related disciplines and the increasingly complex evolutionary paths of the knowledge structures of professional reader groups (such as university faculty and students and cutting-edge researchers), some inherent characteristics of the aforementioned traditional procurement technologies are gradually revealing deep-seated technical limitations in addressing the new challenges of the modern scientific research environment. First, from the perspective of the timeliness of supply and demand matching, the existing "reader-recommended purchase" model is essentially a severely lagging feedback mechanism. Because the acquisition, sorting, cataloging, and shelving of literature have inherent business cycles, when readers submit a purchase request for research purposes, they are often at the peak of their thirst for that particular knowledge point. However, by the time the literature is actually acquired and available for borrowing, the reader's research focus or learning stage has often shifted, leading to the embarrassing situation of low utilization rates upon arrival at the library. This temporal misalignment essentially reflects the current system's lack of awareness of the evolving patterns of reader needs. Existing technological paradigms have failed to construct a deep semantic association model between reader behavior trajectories and the evolution of knowledge topology. Traditional classification methods or keyword matching only remain at the surface-level feature comparison level, unable to analyze the dynamic mechanisms of readers' migration from the "known" to the "unknown." When readers' interests drift from basic mathematics to artificial intelligence, if the system cannot recognize the inherent logic of this disciplinary evolution, it cannot proactively allocate relevant advanced literature before readers actually engage in search behavior. Therefore, how to deeply mine the temporal characteristics of readers' borrowing behavior in multi-dimensional spatiotemporal dimensions, identify the drift trajectory of disciplinary interests, and based on this, predict the semantic associations of new books in the upstream of publishing, thereby constructing an intelligent procurement decision-making method with forward-looking perception capabilities, has become a key technological challenge and an urgent bottleneck problem facing the field of intelligent resource construction in libraries. Summary of the Invention

[0003] To address the technical problems in existing library document procurement models, such as delayed feedback, lack of awareness of academic evolution, and lack of foresight in new book evaluation, this invention provides an intelligent recommendation system for library document procurement based on reader borrowing behavior. The system achieves accurate capture and forward-looking prediction of the drift trajectory of readers' knowledge needs by constructing a temporal evolution model of reader behavior.

[0004] A library document procurement intelligent recommendation system based on reader borrowing behavior is characterized in that the system includes a data acquisition and preprocessing module, a reader behavior time series modeling module, a subject evolution dynamics analysis module, a future demand feature prediction module, a new book semantic association matching module, and a procurement decision intelligent generation module.

[0005] The data acquisition and preprocessing module establishes a real-time communication connection with the underlying database of the Library Integrated Management System (ILS) through a standard data exchange interface to extract the original borrowing records of the target reader group within a preset time window. The original borrowing records include a unique reader identifier, a unique document identifier, a borrowing timestamp, a return timestamp, and the corresponding Chinese Library Classification number of the document. The data acquisition and preprocessing module has a built-in classification number normalization processing unit that transforms the acquired multidimensional raw data into a subject path vector with time as the axis.

[0006] In a preferred embodiment of the present invention, the reader behavior time-series modeling module receives subject path vectors from the preprocessing module and constructs a reader behavior time-series sequence using a sliding window algorithm; the step size of the sliding window is set to one month, and the window length is set to six months; the module constructs a feature matrix reflecting the knowledge acquisition focus of the reader group within a specific time period by weighting the frequency of occurrence of subject classification numbers within the window; furthermore, the reader behavior time-series modeling module also includes a noise filtering unit to remove atypical borrowing records with a borrowing period of less than 24 hours, so as to ensure that the feature expression of the time-series sequence can truly reflect the systematic learning behavior of readers.

[0007] Furthermore, the subject evolution dynamics analysis module constructs a subject association topology model based on knowledge graph technology. Nodes in the subject association topology model represent sub-disciplinary fields, and the edges between nodes represent logical progression relationships or interdisciplinary connections between subjects. The module maps the temporal sequence of reader behavior to the subject association topology model and uses path mining algorithms to identify the current knowledge mastery stage of the reader group. The subject evolution dynamics analysis module identifies the interest drift trend of the reader group by calculating the migration rate and deflection angle of the subject path vector in the topology model. The migration rate is defined as the change in logical distance between subject nodes per unit time, and the deflection angle is defined as the angle between the current subject path vector and the historical average path vector.

[0008] In another preferred embodiment of the present invention, the future demand feature prediction module calculates the subject-specific keywords and their corresponding academic depth levels that the reader group is most likely to retrieve in the next stage based on a Markov prediction model; the state transition probability matrix of the Markov prediction model... It is constructed using the following mathematical formula: in, This indicates that the reader is from the subject area Shift to academic disciplines The probability, This indicates that the reader completes the process from the historical sample. arrive Number of migrations This indicates that the reader is from the subject area The total number of all transfer behaviors that occurred; the future demand feature prediction module determines the demand keywords that are likely to appear in the next three months by calculating the steady-state distribution vector.

[0009] The academic depth level is calculated by analyzing the difficulty coefficient of the literature borrowed by the reader group in the past; the difficulty coefficient is based on a comprehensive score of the literature’s citation frequency, inclusion in core journals, and the complexity of the chapter structure; the demand feature output by the prediction module is a multi-dimensional vector, which includes a set of keywords, a predicted demand intensity value, and a target depth range.

[0010] Furthermore, the new book semantic association matching module obtains new book metadata information from an external publisher's database in real time via an application programming interface (API); the metadata information includes book title, abstract, table of contents, author biography, keywords, and ISBN number; the module uses a deep learning-based semantic representation model to transform the new book metadata into a high-dimensional semantic feature vector; the new book semantic association matching module calculates the cosine similarity between the predicted demand feature vector and the new book semantic feature vector. : in, For the predicted demand feature vector, The semantic feature vector of the new book; the system sets a similarity threshold. ,when At that time, it was determined that the new book was strongly related to the future needs of readers.

[0011] As a further improvement of the present invention, the intelligent procurement decision generation module receives the evaluation results of the new book semantic association matching module and generates an intelligent procurement recommendation list in conjunction with the current collection data; the current collection data includes the existing number of copies of similar documents in the library, the average turnover rate in the past year, and the publication timeliness of the document; the intelligent procurement decision generation module uses a multi-criteria decision analysis algorithm (MCDA) to calculate the procurement priority score of each candidate new book. in, The preset weighting coefficients, This refers to the number of copies of similar documents already in the library. This represents the upper limit for the number of copies in this discipline. The time span is from the publication date to the present; the system sorts the data from high to low based on the $Score$ values ​​and automatically generates tiered purchasing recommendations.

[0012] This invention also provides a hardware architecture for an intelligent recommendation system for library document procurement based on reader borrowing behavior. The hardware architecture includes a high-performance computing cluster, a distributed storage array, and a load balancer. The high-performance computing cluster consists of multiple computing nodes, each configured with at least two 16-core processors with a clock speed of 3.0 GHz and at least 256 GB of synchronous dynamic random access memory (DDR4 SDRAM). The distributed storage array is used for persistent storage of reader behavior logs, subject knowledge graphs, and new book metadata, and employs redundant disk array (RAID 6) technology to ensure high data reliability. The load balancer is located at the network access layer and uses a round-robin scheduling algorithm to distribute concurrent requests generated by the data acquisition module, ensuring low-latency response when processing large-scale concurrent borrowing data.

[0013] As a specific embodiment of the present invention, during the operation of the data acquisition and preprocessing module, for documents with incomplete classification numbers, a text classification-based completion mechanism is initiated; this mechanism uses a convolutional neural network (CNN) to scan the bibliographic information of the documents, extract feature maps, and output the secondary subject classification number with the highest probability as the completion result; the preprocessed subject path vector is stored in key-value pair format to facilitate rapid retrieval by subsequent modules.

[0014] Furthermore, the subject evolution dynamics analysis module introduces the concept of "path weight" from the knowledge graph when identifying the current knowledge mastery stage of the reader group. The path weight is dynamically adjusted according to the co-occurrence frequency between two subject areas and the logical sequence of subject evolution. For example, when the system detects that the reader group is shifting from "discrete mathematics" to "data structures" on a large scale, the corresponding path weight in the topology model will increase, thereby causing the prediction model to tilt towards this evolution path when calculating the transition probability.

[0015] As a further refinement of the present invention, the time series analysis in the future demand feature prediction module adopts a Long Short-Term Memory (LSTM) network; the LSTM network includes an input gate, a forget gate, and an output gate, which can effectively capture the long-term dependencies in readers' borrowing behavior; the network training process adopts the backpropagation algorithm and the Adam optimizer, and the loss function adopts the mean squared error function; by inputting the subject demand intensity sequence of the past twelve months, the demand evolution curve for the next three months is predicted.

[0016] Furthermore, the semantic association matching module of the new book introduces an attention mechanism when processing text metadata. By assigning higher weights to core technical terms in the summary and table of contents, the accuracy of semantic vector generation is improved. The semantic representation model is pre-trained on a general encyclopedia and a professional terminology corpus, and has the ability to understand synonym transformations and contextual logic of professional terms.

[0017] In a preferred embodiment of the present invention, the intelligent procurement decision generation module further includes a closed-loop feedback submodule. This submodule is used to monitor the actual utilization of newly procured documents after they arrive at the library, including the number of borrowings, the number of people queuing for reservations, and the reader satisfaction ratings in the first month after they are put on the shelves. The closed-loop feedback submodule uses the above-mentioned actual utilization data as labels to feed back to the future demand feature prediction module, and dynamically corrects the state transition probability matrix of the Markov model through a reinforcement learning algorithm to achieve continuous iterative optimization of the system's prediction accuracy.

[0018] Furthermore, the system provides a graphical management interface for displaying the generated intelligent procurement recommendation list to interviewers. The interface supports multi-dimensional filtering and sorting, and interviewers can view the predicted score details of each recommended book, the prediction logic of the subject's evolution path, and the distribution map of similar resources in the library. The interface also has a manual intervention interface, allowing interviewers to temporarily adjust the weight coefficients of specific subject categories based on sudden research tasks or specific subject development plans.

[0019] In this embodiment of the invention, the system's workflow is as follows: First, the data acquisition and preprocessing module extracts daily borrowing increment data from the library management system via a database connector and associates it with the reader's identity information (such as grade, major, and research direction) in the reader database to form annotated behavior log; Second, the reader behavior time series modeling module aggregates the above logs by reader group to generate a subject interest time series for each professional group, and uses a Gaussian filter to smooth out random fluctuations caused by holidays or emergencies; Third, the subject evolution dynamics analysis module, combined with the library's subject map, calculates the current knowledge life cycle stage of each reader group and identifies different states such as "foundational consolidation," "professional advancement," or "cutting-edge exploration"; Fourth, the future demand characteristic prediction module calls a trained Markov model. The model, in conjunction with the LSTM network, outputs a demand forecast report for the next procurement cycle. This report, encapsulated in XML format, includes the expected subject distribution density, keyword cloud map, and suggested document depth indicators. The fifth step involves a new book semantic association matching module that synchronizes the latest book catalogs from the external publishing market via multi-threaded crawlers or a data bus, and uses deep semantic analysis algorithms to scan each book for "demand fit." The sixth step involves an intelligent procurement decision generation module that applies a multi-criteria scoring model, comprehensively considering predicted demand, duplicate redundancy, publication timeliness, and budget constraints to generate a final procurement list recommendation. The seventh step involves the system pushing the recommended list to the acquisition management terminal, and after the documents arrive at the library, achieving precise "book-to-person" service through personalized recommendation interfaces for readers (such as mobile app push notifications), while simultaneously collecting feedback data to complete the closed-loop adaptive adjustment of the algorithm.

[0020] The present invention provides an intelligent recommendation system for library document procurement based on reader borrowing behavior, which has the following beneficial effects: 1. Significantly enhances the forward-looking nature of literature procurement: By constructing a dynamic model of disciplinary evolution, the system can identify the dynamic drift pattern of readers' knowledge needs and accurately predict the knowledge stage that readers will enter before they actually conduct a search, thereby achieving advance planning and effectively solving the problem of lag in traditional procurement models.

[0021] 2. Optimized resource allocation efficiency: The system utilizes a multi-criteria decision analysis algorithm to scientifically balance demand intensity, duplicate quantity, and publication timeliness, avoiding blind investment in overheated disciplines or redundant resources. This allows the limited procurement budget to be precisely tilted towards the next stage of discipline evolution, significantly improving the return on investment of collection funds.

[0022] 3. Enhanced the scientific rigor of new book evaluation: Based on deep learning-based semantic association matching technology, it breaks through the cold start dilemma of new books having no historical borrowing data for reference. By calculating the matching degree between the features of new books and the features of predicted demand in a high-dimensional semantic space, it achieves an objective evaluation of the utilization potential of newly entered documents.

[0023] 4. Improved turnover rate of library resources: Because the procurement decisions are precisely aligned with readers' learning curves and research progress, newly acquired documents can be quickly matched with readers at the corresponding knowledge stage. Through a closed-loop feedback mechanism, it has been verified that the initial borrowing conversion rate and average turnover speed of the documents recommended by this invention are better than the traditional manual book selection model.

[0024] 5. The system has achieved automation and intelligence in the decision-making process: It has built a closed loop from raw data collection to final decision generation, reducing tedious manual screening work. This not only lowers the professional threshold for interviewers, but also ensures the ability of procurement strategies to automatically optimize with changes in the subject environment through continuous algorithm iteration.

[0025] Building upon this foundation, the system's data acquisition and preprocessing module further integrates a real-time data quality monitoring unit. This unit identifies anomalous data streams by calculating the entropy value of the data distribution. If borrowing data exhibits unnatural clustering within a certain time period (e.g., a short-term surge in borrowing due to specific course assignments), the unit will automatically identify it as "course impulse noise" and reduce its interference weight on long-term subject interest shift predictions using exponential smoothing when building the time-series model.

[0026] As another important improvement of this invention, the discipline evolution dynamics analysis module adopts a multi-layered composite graph architecture when handling interdisciplinary connections. The first layer is a hierarchical logical graph based on the Chinese Library Classification; the second layer is a semantic connection graph based on the spatial distribution of word vectors. When a reader's borrowing behavior simultaneously shows migration characteristics towards a certain emerging interdisciplinary field (such as bioinformatics, computational finance, etc.) in both graph layers, the system will automatically trigger an "interdisciplinary warning" and significantly increase the weight score of that field in the demand prediction vector.

[0027] The future demand feature prediction module constructs a "knowledge gradient evaluation model" based on document metadata features when calculating the academic depth level. This model extracts parameters such as formula density, average citation years of references, and terminology complexity from the documents, and calculates the academic depth value (ranging from 0 to 1) using a support vector regression (SVR) algorithm. When generating predictive features, the system not only provides the subject direction but also explicitly indicates the required document depth level. For example, for readers learning basic calculus, the system predicts their next need will focus on "medium-depth" real analysis literature, rather than "basic" popular science mathematics or "extremely advanced" cutting-edge professional papers, thus ensuring a precise match between the purchased literature and the reader's cognitive level.

[0028] In the feature extraction stage, the semantic association matching module for new books employs a "chapter directory weight enhancement algorithm" tailored to the specific characteristics of academic works. This algorithm extracts the book's table of contents structure using optical character recognition (OCR) or PDF parsing technology, identifying the logical hierarchy between chapter titles. Compared to conventional methods that rely solely on the title and abstract, this algorithm reveals a deeper level of the distribution of knowledge points contained in the document. For new books containing numerous chapters with a large number of predicted high-frequency keywords, the system assigns extremely high semantic matching scores, even if the relevant keywords do not directly appear in the book title.

[0029] The intelligent procurement decision generation module calculates the impact factor of duplicate quantity. At that time, the concept of a dynamic inventory warning line was introduced. This warning line is automatically adjusted based on the average annual borrowing growth rate of that subject category. For subjects in a period of rapid growth, the system will automatically lower the warning line. A negative contribution to the rating encourages the purchase of more copies to cope with the expected borrowing peak; for technology fields that are in decline or where knowledge is changing very rapidly, the system will strictly limit the number of copies to ensure the simplicity and efficiency of the collection structure.

[0030] Furthermore, the distributed storage array in the hardware architecture employs non-volatile memory caching (NVMe Cache) technology to accelerate the reading and writing of knowledge graph topology data. During large-scale Markov chain state transition calculations, the system can utilize graphics processing units (GPUs) in the high-performance computing cluster for parallel acceleration, reducing the update cycle of the predictive model from the traditional weekly cycle to an hourly cycle, significantly improving the real-time performance of decision-making.

[0031] As a supplementary explanation of the present invention, the closed-loop feedback submodule of the system, when collecting reader reviews, not only collects explicit star ratings but also analyzes readers' comment texts using natural language processing technology. Through sentiment analysis and topic extraction, the system can identify readers' specific evaluations of document depth, print quality, and content timeliness. This refined feedback data is transformed into feature vectors, which are used to fine-tune the parameters of the new book semantic association matching module, enabling the system's criteria for judging "high-quality documents" to adaptively adjust as the library's readers' tastes evolve.

[0032] In a real-world deployment environment, the intelligent recommendation system for library document acquisition based on reader borrowing behavior can be decoupled from and linked with other library subsystems (such as the OPAC query system and electronic resource management system) through a standardized microservice architecture. The acquisition suggestion reports regularly generated by the system can be directly connected to the pending approval pool of the acquisition workflow software, supporting one-click transfer to purchase orders, thereby building a fully digitalized chain process from demand perception to resource acquisition.

[0033] In summary, this invention introduces a prediction-based dynamic decision-making mechanism into the field of library resource development through multi-dimensional data mining and the construction of complex evolutionary models. This mechanism not only addresses the inherent shortcomings of traditional solutions at the technical level but also represents a leap in management philosophy, moving from an "experience-driven" approach to a deep integration of "data-driven" and "logic-driven" approaches. The entire system architecture is robust, with clear logic and efficient collaboration between modules, providing solid technical support for enhancing libraries' knowledge assurance capabilities in the information age. The algorithms and hardware selections involved in this invention are all feasible within the scope of existing engineering technologies, possessing extremely high practical application value and promising prospects for widespread adoption. Attached Figure Description

[0034] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the structure of an intelligent recommendation system for library document procurement based on reader borrowing behavior according to the present invention; Figure 2 This is a schematic diagram illustrating the workflow of an intelligent recommendation system for library document procurement based on reader borrowing behavior, as described in this invention. Detailed Implementation

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0036] This invention provides an intelligent recommendation system for library document procurement based on reader borrowing behavior. Its overall architecture employs a highly integrated modular design, achieving a closed-loop process from automated perception of raw reader behavior data to precise generation of procurement decision instructions. The system mainly consists of a data acquisition and preprocessing module, a reader behavior time-series modeling module, a subject evolution dynamics analysis module, a future demand feature prediction module, a new book semantic association matching module, and a procurement decision intelligent generation module. To ensure the real-time performance of massive data processing and the stability of algorithm operation, the system is deployed on a hardware environment supported by a high-performance computing cluster, distributed storage array, and load balancer.

[0037] Specifically, the data acquisition and preprocessing module serves as the system's data entry point, establishing a real-time communication connection with the library's existing Integrated Management System (ILS) underlying database through a standardized data exchange interface. In engineering practice, this connection can be implemented using JDBC (Java Database Connectivity) or a RESTful API, extracting raw borrowing records from the target reader group at a preset acquisition frequency (e.g., incremental synchronization once per hour or once per day). This raw data exists in unstructured or semi-structured log format, containing unique reader identifiers, unique document identifiers, borrowing and return timestamps accurate to the second, and the corresponding Chinese Library Classification number for each document. To address the issue of inconsistent data formats from different sources, the data acquisition and preprocessing module integrates a classification number normalization processing unit. This unit can identify and convert different versions of the Chinese Library Classification code, mapping them uniformly to a standardized subject classification tree structure. Furthermore, this module integrates a real-time data quality monitoring unit. This unit assesses the stability of data distribution by calculating the entropy of the data stream. When it detects unnatural clustered borrowing behavior within a specific time period, the system automatically identifies it as "course impulse noise." For example, in a university environment, midterm assignments for a particular course often lead to a concentrated borrowing of related reference books within a short period. This short-term fluctuation does not represent a shift in readers' long-term knowledge needs. Therefore, the real-time data quality monitoring unit reduces the weight of such data in subsequent time-series modeling using exponential smoothing, thereby ensuring the purity of the subject path vector.

[0038] After obtaining the cleaned subject path vectors, the reader behavior temporal modeling module starts its operation. This module uses a sliding window algorithm to represent readers' borrowing behavior temporally, specifically setting the step size of the sliding window to one calendar month and the window length to six calendar months. Through this rolling data slicing method, the system can capture readers' knowledge acquisition trajectory over a period of half an academic year or longer. Within each window, the system performs a weighted calculation based on the frequency of occurrence of subject classification numbers, assigning higher weight coefficients to borrowing behaviors closer to the current time point, thereby constructing a feature matrix reflecting the knowledge acquisition focus of the reader group in the current period. Addressing the common problem of missing or incomplete classification numbers in library data, the module incorporates a text classification-based completion mechanism. It uses a convolutional neural network (CNN) to extract features from the title, keywords, and abstract of documents, outputting the secondary subject classification number with the highest probability as the completion result, and storing the processed data in a high-performance key-value pair format.

[0039] As one of the core technical logics of this invention, the subject evolution dynamics analysis module constructs a multi-layered, composite subject association topology model using knowledge graph technology. The first layer of the graph is constructed based on the hierarchical logic of the *Chinese Library Classification*, reflecting the inclusion and affiliation relationships between subjects; the second layer is based on the distribution patterns of word vector space, reflecting the horizontal semantic relationships between emerging and interdisciplinary subjects. During operation, this module maps the reader's subject path vectors into this topology model and identifies the reader's current knowledge acquisition stage through path mining algorithms. The system quantifies the drift trend of interest by calculating the migration rate and deflection angle of the path vectors in the model. Specifically, the migration rate characterizes the speed at which the reader moves from basic theory to applied practice or cutting-edge exploration along the subject chain, while the deflection angle reveals whether the reader is crossing subject boundaries into emerging fields. When the reader's borrowing behavior simultaneously shows a migration towards a specific field (such as the intersection of artificial intelligence and biology) in both layers of the graph, the system triggers an interdisciplinary warning and automatically increases the weight of that field.

[0040] The future demand characteristic prediction module, based on the results of dynamic analysis, uses a combined algorithm of a Markov prediction model and a Long Short-Term Memory (LSTM) network to predict demand. The Markov model's state transition probability matrix... By analyzing the number of times readers migrated between different subject areas in historical samples Dynamic construction is performed. Formula This ensures the statistical objectivity of the transition probability. Meanwhile, the LSTM network, utilizing its unique gating mechanism (input gate, forget gate, output gate), processes the demand intensity sequence of the past twelve months, effectively capturing periodic patterns in academic research. The demand characteristics output by this module not only include the predicted set of keywords but also an academic depth level determined based on a "knowledge gradient evaluation model." This model, through the Support Vector Regression (SVR) algorithm, comprehensively analyzes the formula density, reference timeliness, and terminology complexity of the literature, quantifying the literature depth as a value between 0 and 1. This means the system can not only predict whether a reader needs books in the field of "quantum computing" but also accurately indicate whether they need "introductory popular science" or "doctoral research" literature.

[0041] After extracting the demand features, the new book semantic association matching module accesses an external publisher's database via an API interface to obtain full-dimensional metadata, including book title, abstract, table of contents, and ISBN number. This module employs a deep semantic representation model based on an attention mechanism to map the new book's metadata into a high-dimensional space. To improve matching accuracy, this module specifically designs a "chapter directory weight enhancement algorithm," which uses parsing techniques to extract the logical hierarchy within the directory and calculates the cosine similarity between the distribution of knowledge points contained in the directory and the predicted demand feature vector. The calculated similarity... Exceeding the preset threshold At that time, the new book was marked as a potential purchase target.

[0042] The intelligent procurement decision generation module combines current collection data with a multi-criteria decision analysis (MCDA) algorithm to generate the final procurement priority score. The calculation formula is: in, , , These are weights based on demand relevance, collection saturation, and publication timeliness. The system introduces a dynamic inventory warning line concept; for subjects with high borrowing growth rates during their upward trend, the warning line will be automatically lowered. The negative contribution of (existing duplicate quantity) to the scoring is used to encourage procurement. The final intelligent procurement recommendation list is displayed to the interviewee through a graphical management interface 11, supporting one-click transition to the procurement process.

[0043] To verify the technical effects of the present invention, the following detailed description will be provided through specific embodiments and comparative examples.

[0044] Example 1: In this example, the system described in this invention is deployed in the library of a large comprehensive university of science and technology. The system hardware configuration includes four computing nodes, each equipped with two 3.0GHz Intel Xeon series sixteen-core processors, 512GB DDR4 memory, and a 100TB distributed storage array composed of NVMe SSDs, using RAID 6 parity mode. The experimental period was set from September 2022 to August 2023, a total of 12 months. The data acquisition module was connected to the university's ILS system, acquiring 2.4 million original borrowing records covering 35,000 on-campus readers. In the preprocessing stage, the system identified that approximately 12% of the records needed to be completed using CNN due to incomplete classification numbers, with a completion accuracy rate of 94.5%. In the analysis of the dynamics of disciplinary evolution, the system monitored that in the spring semester of 2023, the disciplinary path of first-year graduate students in the School of Computer Science rapidly shifted from "Discrete Mathematics" and "Fundamentals of Programming" to "Deep Learning Theory," with a migration rate increase of 45%, and the shift angle pointing towards "Computer Vision." Predictions based on Markov models and LSTM show that this group's demand for in-depth literature related to "Transformer architecture" and "generative AI" will experience explosive growth from May to July 2023, with a prediction strength of 0.88 (out of 1.0). Based on this demand, the new book matching module selected 150 relevant new books from an external bibliography, including 25 original foreign language works. Mean semantic matching score... The weight is set to 0.82. , , The system ultimately generated a recommended list containing 120 books.

[0045] In Comparative Example 1, during the same period, another subject category of the library (humanities and social sciences) adopted a traditional acquisition model. Acquisition staff primarily selected books manually based on bestseller lists provided by publishers, recommended purchase lists submitted by subject matter experts, and borrowing statistics of similar books from previous years. Purchasing decisions relied mainly on static statistics of the previous year's borrowing data, without incorporating dynamic time-series forecasting models or semantic association matching techniques. The determination of duplicate copies mainly referenced empirical values, lacking adjustments based on dynamic inventory warning lines.

[0046] Experimental data comparison and result analysis: During the actual utilization tracking period from September to December 2023, a multi-dimensional quantitative comparison was conducted on the literature purchased in Example 1 (recommended by the system of this invention) and Comparative Example 1 (traditional manual book selection). The test results are shown in Table 1.

[0047] Table 1: Comparison of the effectiveness of the recommendation system of this invention with the traditional procurement model The data in Table 1 clearly demonstrates the significant superiority of the system provided by this invention in several core indicators. Firstly, regarding the "first-month borrowing rate of new books," Example 1 achieved 87.5%, meaning that the vast majority of the documents recommended by the system were borrowed within the month of acquisition, compared to only 42.3% in Comparative Example 1. This strongly proves the high degree of alignment between demand characteristics predicted based on the dynamics of disciplinary evolution and actual reader needs, achieving precise supply through "books finding people." Secondly, the average borrowing turnover period was shortened from 45.8 days to 12.4 days, greatly improving the circulation efficiency of document resources. The most crucial indicator is the "reader demand satisfaction lag time." Traditional models, being responsive procurement (i.e., processing only after readers generate demand and provide feedback), exhibit significant lag; while the system of this invention achieves -15 days, meaning that relevant documents are prepared for acquisition approximately two weeks before readers engage in large-scale borrowing behavior, achieving forward-looking planning. In terms of budget execution, Example 1 achieved an effective copy rate of 94.2%, while the proportion of documents with zero borrowing rate was only 2.1%. In contrast, nearly one-fifth of the documents in the traditional model are never borrowed within six months of being acquired, resulting in a serious waste of resources. This demonstrates that by using the Multi-Criterion Decision Analysis (MCDA) algorithm and introducing a dynamic inventory warning line, this system can scientifically suppress the generation of redundant copies, ensuring that every penny of procurement funds is invested in areas with truly high demand.

[0048] Further analysis of the closed-loop feedback process of the system in actual operation reveals the following: In Example 1, the system continuously monitors the utilization of newly purchased documents through a closed-loop feedback submodule. When it detects that the number of people queuing to borrow certain documents exceeds five, the feedback module automatically adjusts the weight parameters in the future demand prediction module and triggers a "secondary supplementary purchase" suggestion. Simultaneously, natural language processing technology is used to perform sentiment analysis on readers' comments on the mobile library app, extracting readers' feedback on the depth of the documents. For example, if readers generally report that a textbook predicted to be of "medium depth" is actually too shallow, the system automatically corrects the knowledge gradient evaluation model parameters in the new book semantic association matching module. This adaptive learning capability allows the system's recommendation accuracy to show a significant logarithmic growth trend after a complete procurement cycle iteration.

[0049] The hardware architecture design of this invention also provides a solid foundation for the implementation of the aforementioned algorithms. The distributed storage array employs NVMe Cache technology to cache and accelerate frequently accessed subject knowledge graph topology data, keeping the latency of a single path mining operation within 50 milliseconds. The computing nodes in the high-performance computing cluster achieve smooth distribution of concurrent requests through a load balancer, ensuring extremely high response speeds even during peak data access periods such as university-wide course selection weeks. When performing large-scale matrix operations on Markov state transition probabilities, the system can utilize GPU resources within the cluster for parallel acceleration, reducing training tasks that would normally take hours to minutes, thus ensuring the real-time nature of decision generation.

[0050] In summary, this invention, through the deep mining of reader borrowing behavior characteristics and combined with a nonlinear dynamic model of disciplinary evolution, constructs a complete intelligent construction solution for library document resources. This solution not only technically addresses long-standing pain points in the library community such as procurement delays and resource mismatch, but also achieves the intelligent transformation of library services through data-driven logic. The embodiments disclosed in this specification are merely typical applications of the technical concept of this invention. Any equivalent substitutions or partial improvements made within the scope of the claims of this invention should be included within the protection scope of this invention.

Claims

1. A library document procurement intelligent recommendation system based on reader borrowing behavior, characterized in that, The system includes: a data acquisition and preprocessing module, used to establish a real-time communication connection with the underlying database of the library integrated management system through a standard data exchange interface, extract the original borrowing records of the target reader group within a preset time window, and use a built-in classification number normalization processing unit to transform the multidimensional raw data into subject path vectors with time as the axis; a reader behavior temporal modeling module, used to receive the subject path vectors, use a sliding window algorithm to construct a feature matrix reflecting the knowledge acquisition focus of the reader group within a specific time period, and use a noise filtering unit to remove atypical borrowing records with borrowing cycles shorter than a preset threshold; and a subject evolution dynamics analysis module, used to construct a multi-layered composite subject association topology model containing subject hierarchical logic and semantic associations based on knowledge graph technology, and to calculate the migration rate of the subject path vectors in the topology model by mapping the temporal sequence of reader behavior to the topology model. The system includes several modules: a bias angle module to identify the interest drift trend and current knowledge acquisition stage of the reader group; a future demand feature prediction module to calculate the subject keywords and corresponding academic depth levels of the reader group in the next stage based on Markov prediction models and long short-term memory networks, and output a multi-dimensional demand feature vector containing a keyword set, predicted demand intensity value, and target depth range; a new book semantic association matching module to obtain new book metadata information from external publisher databases in real time, use a deep learning-based semantic representation model to transform the new book metadata into a high-dimensional semantic feature vector, and calculate the cosine similarity between the demand feature vector and the new book semantic feature vector; and a procurement decision intelligent generation module to calculate the procurement priority score of candidate new books based on the cosine similarity and combined with the current collection data, using a multi-criteria decision analysis algorithm, and automatically generate a tiered intelligent procurement recommendation list.

2. The intelligent recommendation system for library document procurement based on reader borrowing behavior as described in claim 1, characterized in that, The data acquisition and preprocessing module also integrates a real-time data quality monitoring unit. This unit identifies abnormal data distribution by calculating the entropy value of the data stream. When it detects that the borrowing data within a certain period exhibits unnatural clustering, it determines it to be course impulse noise and reduces the weight of the data in that period using exponential smoothing when establishing the time series model. The classification number normalization processing unit is used to identify and convert different versions of the Chinese Library Classification code, mapping them uniformly to a standardized subject classification tree structure.

3. The intelligent recommendation system for library document procurement based on reader borrowing behavior according to claim 1, characterized in that, When constructing the feature matrix, the reader behavior time-series modeling module sets the sliding window step size to one month and the window length to six months. It calculates the frequency of occurrence of subject classification numbers within the window by weighting the borrowing behavior closer to the current time node, and assigns a higher weight coefficient to the borrowing behavior. The reader behavior time-series modeling module also has a built-in text classification-based completion mechanism. For documents with missing classification numbers, it uses a convolutional neural network to extract features from the bibliographic information of the document and outputs the second-level subject classification number with the highest probability as the completion result. The preprocessed subject path vector is stored in key-value pair format.

4. The intelligent recommendation system for library document procurement based on reader borrowing behavior according to claim 1, characterized in that, The subject evolution dynamics analysis module constructs a subject association topology model with a multi-layered composite architecture. The first layer is a hierarchical logical graph based on the Chinese Library Classification, and the second layer is a semantic association graph based on the spatial distribution of word vectors. When calculating the interest drift trend, the module defines the migration rate as the change in logical distance between subject nodes per unit time and the deflection angle as the angle between the current subject path vector and the historical average path vector. When a reader's borrowing behavior shows a migration characteristic towards interdisciplinary fields in both layers of the graph, an interdisciplinary warning is triggered, and the weight of that interdisciplinary field is automatically increased.

5. The intelligent recommendation system for library document procurement based on reader borrowing behavior according to claim 1, characterized in that, The future demand feature prediction module constructs the state transition probability matrix of a Markov prediction model. When using the formula in Indicates from the subject area Shift to academic disciplines The probability, Indicates the historical sample from arrive Number of migrations Indicates from the subject area The total number of all transfer behaviors that occurred; at the same time, the module processes the subject demand intensity sequence of the past twelve months through a long short-term memory network, predicts the demand evolution curve for the next three months, and trains the network with the Adam optimizer through the backpropagation algorithm.

6. The intelligent recommendation system for library document procurement based on reader borrowing behavior according to claim 1, characterized in that, When determining the academic depth level, the future demand feature prediction module constructs a knowledge gradient evaluation model based on the support vector regression algorithm. The model extracts formula density, average citation years of references, and terminology complexity parameters from the document metadata to calculate an academic depth value ranging from 0 to 1. The demand prediction report output by the future demand feature prediction module is packaged in XML format and includes, in addition to the subject term set, the expected subject distribution density, keyword cloud map, and specific document depth indicators to achieve matching between the purchased documents and the readers' cognitive level.

7. The intelligent recommendation system for library document procurement based on reader borrowing behavior according to claim 1, characterized in that, The semantic association matching module for new books introduces an attention mechanism when processing new book metadata. It improves the accuracy of semantic feature vectors by assigning weights to core technical terms in the book summary and table of contents. The module also includes a chapter / table of contents weight enhancement algorithm. This algorithm extracts the book's table of contents structure using optical character recognition or PDF parsing technology, identifies the logical hierarchy between chapter titles, and calculates the cosine similarity between the distribution of knowledge points in the table of contents and the predicted demand feature vector. The calculation formula is: in For the demand feature vector, This is the semantic feature vector of the new book.

8. The intelligent recommendation system for library document procurement based on reader borrowing behavior according to claim 1, characterized in that, The intelligent procurement decision generation module employs a multi-criteria decision analysis algorithm when calculating the procurement priority score $Score$: in For preset weighting coefficients, For semantic similarity, This refers to the number of copies of similar documents already in the library. This represents the upper limit for the number of copies in this discipline. This module introduces the concept of a dynamic inventory warning line, which automatically adjusts based on the average annual borrowing growth rate of each subject category, to determine the publication duration. For disciplines that are on the rise, the value can be reduced by lowering... Negative contributions in the rating are used to encourage additional purchases.

9. The intelligent recommendation system for library document procurement based on reader borrowing behavior according to claim 1, characterized in that, The system also includes a closed-loop feedback submodule and a graphical management interface. The closed-loop feedback submodule monitors the actual utilization rate of newly acquired documents after they arrive at the library, and uses natural language processing technology to perform sentiment analysis and topic extraction on readers' comments. The extracted feedback data is then used as tags to feed back to the future demand feature prediction module, which dynamically corrects the state transition probability matrix of the Markov model through reinforcement learning algorithms. The graphical management interface displays the recommendation list and score details for each dimension, and includes a manual intervention interface that allows interviewers to temporarily adjust the weight coefficients according to specific discipline development plans.

10. A library document procurement intelligent recommendation system based on reader borrowing behavior according to any one of claims 1 to 9, characterized in that, The system is deployed on a hardware architecture that includes a high-performance computing cluster, a distributed storage array, and a load balancer. The high-performance computing cluster consists of multiple computing nodes, each configured with a multi-core processor with a main frequency of no less than 3.0 GHz and no less than 256 GB of synchronous dynamic random access memory, and supports parallel acceleration computing using a graphics processing unit. The distributed storage array uses non-volatile memory caching technology to accelerate the reading and writing of knowledge graph topology data, and uses redundant disk array technology to store data. The load balancer is set at the network access layer and distributes concurrent requests generated by the data acquisition module through a round-robin scheduling algorithm.

Images