A book intelligent purchasing method and system based on borrowing heat

Abstract

The application discloses a kind of based on borrowing heat's book intelligent procurement method and system, its method includes steps: the metadata of collection book and historical borrowing data are collected, and according to the data, construct heterogeneous characteristics, the heterogeneous characteristics include sparse category feature, dense numerical characteristic and dense semantic feature;According to the historical borrowing data, calculate borrowing heat, as training label;Improved DeepFM model is constructed and model is trained;According to the metadata of candidate book, using the improved DeepFM model of trained, obtain the predicted borrowing heat of each candidate book;Based on the predicted borrowing heat, the total budget of pre-set procurement, each subject minimum procurement volume constraint and copy upper limit constraint, execute two-stage greedy optimization algorithm, generate final book procurement book and recommended procurement volume.The application can realize the accurate prediction of new book borrowing heat, and improve the scientificity, efficiency and the quality of collection resource construction of book procurement.

Description

Technical Field

[0001] This invention belongs to the field of book procurement and artificial intelligence technology, specifically, it relates to a book intelligent procurement method and system based on borrowing popularity. Background Technology

[0002] The acquisition of print books is a core component of library collection development. Currently, library print book acquisition mainly relies on librarian experience, annual subject plans, or reader recommendations for manual selection. This traditional model has significant drawbacks: First, it is difficult to systematically integrate complex factors such as reader borrowing preferences, subject development trends, and the intrinsic value of books, resulting in a highly subjective decision-making process. Second, it lacks the ability to scientifically predict future book borrowing needs, often leading to the absence of popular books or delayed acquisitions, while some books are over-purchased due to misjudgments, resulting in resource waste and low fund utilization. Finally, faced with massive amounts of bibliographic information, manual screening is highly subjective and inefficient, failing to meet the needs of modern libraries for refined and intelligent acquisition.

[0003] In recent years, some studies have attempted to introduce machine learning technology to assist in procurement decisions, but the following limitations still exist:

[0004] (1) Feature construction failed to fully fit the book procurement scenario. For example, the feature selection lacked verification of its relevance to the procurement decision; or the procurement decision problem was simply equated with the personalized recommendation problem, focusing on user behavior profiles and failing to effectively integrate the key metadata of the book itself.

[0005] (2) The models mostly use general machine learning methods to process all features indiscriminately, failing to fully consider the significant differences in the intrinsic importance of different features in the book purchasing scenario, resulting in deviations between the prediction results and the practice in the field.

[0006] (3) Existing solutions mostly predict whether a single book will be purchased and the number of copies, but fail to integrate multiple constraints such as the total purchase budget and the minimum guaranteed number of copies for each subject into the global optimization framework, making it difficult to achieve optimal allocation of global resources.

[0007] Therefore, there is an urgent need for an intelligent decision-making method that can deeply integrate multi-dimensional information about books, accurately predict future demand, and automatically generate the overall optimal procurement plan under multiple practical constraints. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings and deficiencies of existing technologies and provide a smart book procurement method based on borrowing popularity. This method can accurately predict the borrowing popularity of new books and generate an optimal procurement list using a two-stage greedy optimization algorithm under constraints such as budget, subject coverage, and copy limit, thereby improving the efficiency of book procurement and enhancing the scientific nature of library resource construction.

[0009] The second objective of this invention is to provide a smart book purchasing system based on borrowing popularity.

[0010] The objective of this invention is achieved through the following technical solution: a smart book procurement method based on borrowing popularity, comprising the following steps:

[0011] S1. Collect metadata and historical borrowing data of the library's collection of books, and construct heterogeneous features for each book in the collection based on the metadata and historical borrowing data. The heterogeneous features include sparse category features, dense numerical features and dense semantic features.

[0012] S2. Calculate the borrowing popularity of each book in the library collection within a preset time window based on the historical borrowing data, and use it as a training label;

[0013] S3. Construct an improved DeepFM model, and train the model using the heterogeneous features as input and the borrowing popularity as output; wherein, the improved DeepFM model includes a dense input layer, a DNN sub-network and an improved FM sub-network, the dense numerical features and dense semantic features are processed by the dense input layer and then input into the DNN sub-network but not into the improved FM sub-network; the improved FM sub-network introduces a subject cross-enhancement module to enhance the weights of feature cross-terms containing the Chinese Library Classification number;

[0014] S4. Based on the metadata of the candidate books, use the trained improved DeepFM model to obtain the predicted borrowing popularity of each candidate book.

[0015] S5. Based on the predicted borrowing popularity, the preset total procurement budget, the minimum number of books to be purchased for each subject, and the upper limit of duplicate copies, execute a two-stage greedy optimization algorithm to generate the final book procurement list and the suggested number of books to be purchased.

[0016] Preferably, in step S1, the sparse category features include the Chinese Library Classification number, publisher, author, and series name; the dense numerical features include the publication year, price, average number of books borrowed by the author, and the recent borrowing popularity ranking of the category; and the dense semantic features include the book title and keywords.

[0017] Preferably, constructing the recent borrowing popularity ranking feature of the category includes: calculating the quantile of the total number of borrowings of books belonging to the Chinese Library Classification number c within a preset time window among all Chinese Library Classification numbers in the library's collection:

[0018] R norm = ,

[0019] Among them, the quantile R normUsed to represent the recent borrowing popularity ranking of books in a category, N is the total number of Chinese Library Classification numbers for all books in the library's collection, R c It is the ranking of the total number of times a book belonging to the Chinese Library Classification number C was borrowed within a preset time window among all the books in the library's collection under all Chinese Library Classification numbers.

[0020] Preferably, in step S3, the sparse category features are input to the improved FM sub-network and processed by the dense embedding layer of the improved DeepFM model to obtain a dense vector, which is then input to both the improved FM sub-network and the DNN sub-network.

[0021] Preferably, in step S3, the working process of the interdisciplinary enhancement module includes:

[0022] S31. Constructing a mask matrix for subject-specific perception:

[0023] ,

[0024] Where n is the total number of sparse class features, M st Let M be the element in the s-th row and t-th column of the mask matrix M, where M is related to a subject when the intersection of the s-th feature and the t-th feature is related to that subject. st =1, otherwise M st =0;

[0025] S32. Introduce trainable parameters w β Calculate the interdisciplinary enhancement factor β:

[0026] β=σ(w β ), where σ is the Sigmoid function;

[0027] S33, Improved FM Subnetwork Output y FM for:

[0028] ,

[0029] Among them, w s For the first-order feature weights, x s x t For eigenvalues, , These are the dense vectors corresponding to the s-th and t-th features, respectively. It is the inner product of dense vectors.

[0030] Preferably, the two-stage greedy optimization algorithm includes:

[0031] In the first phase, within the constraint of the total procurement budget B, books are selected in descending order of predicted borrowing popularity, prioritizing the fulfillment of the minimum pre-set procurement quantity L for each subject. k Require;

[0032] In the second stage, if there is a surplus in the total procurement budget after the first stage, virtual books with representative copies are generated for the candidate books and selected greedily according to the predicted borrowing popularity after decay, so as to maximize the total borrowing popularity and determine the final list of books to be procured and the recommended number of books to be procured.

[0033] Preferably, the specific process of the first stage includes:

[0034] S51. Sort the candidate books in descending order of predicted borrowing popularity, and initialize the cumulative total amount P=0 and the number of selected books S for each subject k. k =0, the first phase of the procurement list is empty;

[0035] S52. Process each book in the sorted list in descending order, and perform the following steps for the currently processed book:

[0036] S521. Determine if the termination condition is met: If the cumulative total amount P equals the total procurement budget B, or the number of selected books for all subjects S. k All are equal to the corresponding minimum purchase quantity L k If the condition is met, the process terminates and proceeds to step S53; otherwise, step S522 is executed.

[0037] S522. Select one book i in sequence, with a price of p. i ;

[0038] S523, Obtain the subject k to which the current book belongs;

[0039] S524. Determine the number of books S selected for subject k. k Is it equal to the minimum number of copies L to be purchased? k If yes, skip the current book and return to step S522 to process the next book; otherwise, proceed to step S525.

[0040] S525. Calculate the expected total expenditure after purchasing the current book: P' = P + p i ;

[0041] S526. Determine whether P' is greater than the total procurement budget B. If yes, proceed to step S53; otherwise, proceed to step S527.

[0042] S527. Add the current book to the first phase of the procurement list, update the cumulative total amount P=P', and let S k Add 1, then return to step S521;

[0043] S53. Output the first phase procurement list as the first phase procurement result.

[0044] Preferably, the specific process of the second stage includes:

[0045] S54, Generate (U) for each candidate book i i -1) This virtual book, in which U i Let y be the upper limit of the number of copies of the candidate books, and let y be the predicted borrowing popularity of the m-th virtual book. i (m) =α m y i y i The predicted borrowing popularity of candidate books is given by α, which is a preset decay coefficient and 0 < α < 1.

[0046] S55. Sort the candidate books and all virtual books that were not purchased in the first phase in descending order of predicted borrowing popularity; initialize the cumulative total amount P to the final value of the first phase, and initialize the procurement list for the second phase to the procurement list for the first phase.

[0047] S56. Process each book in the order arranged in step S55 in descending order, and perform the following steps for the currently processed book:

[0048] S561. If the total cumulative amount P equals the total procurement budget B, then terminate the process and proceed to step S57; otherwise, proceed to step S562.

[0049] S562. Select one book i in sequence, with a price of p. i ;

[0050] S563. Calculate the expected total expenditure after purchasing the current book: P' = P + p i ;

[0051] S564. Determine whether P' is greater than the total procurement budget B. If yes, proceed to step S57; otherwise, proceed to step S565.

[0052] S565. Add the current book to the second phase of the purchase list, update the total cumulative amount P=P', and then return to step S561;

[0053] S57. Based on the second-stage procurement list, merge the candidate books and virtual books to obtain the final book procurement list and the recommended number of books to be procured.

[0054] A book procurement system based on borrowing popularity, applied to the above-described method, includes:

[0055] The feature construction module is used to collect metadata and historical borrowing data of the library's collection of books, and to construct heterogeneous features for each book in the collection based on the metadata and historical borrowing data. The heterogeneous features include sparse categorical features, dense numerical features and dense semantic features.

[0056] The tag generation module is used to calculate the borrowing popularity of each book in the library collection within a preset time window based on the historical borrowing data, and use it as training tags.

[0057] The model training module is used to construct an improved DeepFM model, and train the model with the heterogeneous features as input and the borrowing popularity as output. The improved DeepFM model includes a dense input layer, a DNN sub-network, and an improved FM sub-network. The dense numerical features and dense semantic features are processed by the dense input layer and then input into the DNN sub-network but not into the improved FM sub-network. The improved FM sub-network introduces a subject cross-enhancement module to enhance the weights of feature cross-terms containing the Chinese Library Classification number.

[0058] The borrowing popularity prediction module is used to obtain the predicted borrowing popularity of each candidate book based on the metadata of the candidate books and using the trained improved DeepFM model.

[0059] The intelligent procurement decision module is used to generate the final book procurement list and suggested procurement number based on the predicted borrowing popularity, the preset total procurement budget, the minimum number of books to be purchased for each subject and the upper limit of duplicate copies, by executing a two-stage greedy optimization algorithm.

[0060] A storage medium storing a program that, when executed by a processor, implements the aforementioned intelligent book procurement method based on borrowing popularity.

[0061] The present invention has the following advantages and effects compared with the prior art:

[0062] (1) Based on the characteristics of book metadata and borrowing data, this invention constructs a heterogeneous feature system composed of sparse categories, dense numerical values ​​and dense semantics, and innovatively designs a feature input path isolation mechanism: dense numerical features and dense semantic features are only input into the DNN sub-network. This design enables the improved FM sub-network to focus on processing the first-order features and second-order interaction features of sparse features, while the DNN sub-network processes the high-order interactions of dense features. This avoids the computational burden caused by the invalid crossover of dense features in the improved FM sub-network and the defect of the crossover results being uninterpretable, thereby improving the model processing efficiency and prediction accuracy.

[0063] (2) In this invention, a subject cross-enhancement module is introduced into the improved FM sub-network. Through the mask matrix and learnable subject cross-enhancement factors, the weight of the feature cross-items involving the Chinese Library Classification number field is dynamically increased. This design enables the model to explicitly focus on high-value combinations related to subjects, thereby increasing the contribution rate of subject-related feature combinations to the prediction, improving the model's adaptability to book procurement business, and improving the model's accuracy in predicting book borrowing popularity.

[0064] (3) This invention employs a two-stage greedy optimization algorithm to transform predicted borrowing popularity into purchasing decisions. The first stage ranks books based on predicted borrowing popularity, ensuring the minimum number of books purchased for each subject under the constraint of the total purchasing budget. The second stage generates virtual books with decreasing marginal borrowing popularity for each candidate book, and greedily purchases books within the remaining budget to maximize the total borrowing popularity. This algorithm unifies predicted borrowing popularity, total purchasing budget, minimum purchasing number constraints for each subject, and upper limit constraints on duplicate copies into an efficient and interpretable decision-making framework, overcoming the limitations of independent prediction models and achieving a globally approximate optimal purchasing list. Furthermore, by providing a basis for book purchasing decisions through accurate predicted borrowing popularity, it significantly improves the scientific nature of book purchasing decisions.

[0065] (4) The modular design of the system in this invention can be connected to the existing library information management system through a standard interface without large-scale modification of the existing infrastructure, and has good application prospects. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating an intelligent book procurement method based on borrowing popularity according to the present invention.

[0067] Figure 2 This is a schematic diagram of the structure of the improved DeepFM model of the present invention.

[0068] Figure 3 This is a flowchart illustrating the two-stage greedy optimization algorithm of this invention. Detailed Implementation

[0069] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0070] Example 1

[0071] like Figure 1 As shown, this invention provides a smart book purchasing method based on borrowing popularity, including the following steps:

[0072] Step 1: Data Acquisition and Heterogeneous Feature Construction

[0073] Collect book metadata (including Chinese Library Classification number, publisher, author, series name, publication year, price, book title, and subject terms) and book borrowing data (including historical book borrowing records) from the library's collection database and borrowing log system.

[0074] The collected book metadata and book borrowing data are preprocessed (including noise reduction, deduplication, completion, and format standardization), and three types of heterogeneous features are constructed for each book in the collection: sparse category features, dense numerical features, and dense semantic features.

[0075] (1) Sparse Category Features: Sparse category features are constructed using integer IDs mapped from the Chinese Library Classification number, publisher, author, and series name. Specifically, the Chinese Library Classification number, publisher, author, and series name are treated as strings, where the series name without a series name is used as...<NOT_SERIES> As a string, for unknown data, it is used <unk>As strings, the strings of Chinese Library Classification Number, Publisher, Author, and Series Name are mapped to integer IDs respectively (e.g., Higher Education Press is mapped to 100).

[0076] (2) Dense Numerical Features: Publication year, price, average number of books borrowed by the author, and recent borrowing popularity ranking within the category were normalized to construct dense numerical features for the books. Specifically:

[0077] Publication year: Converted to the current year and then normalized;

[0078] Pricing: Directly normalized;

[0079] Average number of book borrowings per author: This refers to the average number of times all books under each author's name were borrowed within a preset time window T. avg And through the formula Calculate the average number of books borrowed by the author after smoothing and normalization, b. norm , where b max The maximum number of times an author's books are borrowed in the library's collection is set by the preset time window T, which is based on the book acquisition cycle. For example, T can be six months or one year. In this embodiment, books whose acquisition time is less than the preset time window T are not included in the statistics.

[0080] Recent Borrowing Popularity Ranking by Category: This ranking calculates the total number of borrowings of books belonging to the Chinese Library Classification number c within a preset time window T, ranking them among all Chinese Library Classification numbers in the library's collection. The ranking is then calculated using the quantile formula R. norm = Calculate the recent borrowing popularity ranking of the category after normalization (R). norm Where N is the total number of all classification numbers of the library's collection, and R c This is the ranking of the total number of times a book belonging to the Chinese Library Classification number C was borrowed within a preset time window among all Chinese Library Classification numbers in the library's collection.

[0081] (3) Dense semantic features: Using a pre-trained lightweight Sentence-BERT model, the book title and the concatenated topic words are converted into 384-dimensional sentence vectors respectively, and then the dimensions are reduced to 64-dimensional by a learnable linear layer projection. Finally, the 64-dimensional sentence vectors of the book title and topic words together constitute 128-dimensional dense semantic features.

[0082] Step 2: Calculate the borrowing popularity of books as training labels for the model:

[0083] For each book in the library's collection, the total number of times it was borrowed, y, within a preset time window T (e.g., the past year) is calculated. To avoid the influence of extreme values, the formula is used... Log-smoothing and normalization are performed on y to obtain y norm This refers to the book's borrowing popularity, which serves as the training label for the model. Here, y... max This is the maximum number of times all books in the library can be borrowed.

[0084] Step 3: Build an improved DeepFM model and train it.

[0085] Build as Figure 2 The improved DeepFM model is shown below. Based on the standard DeepFM architecture, this model incorporates two key structural improvements tailored to the data characteristics and decision-making logic of book procurement operations. This results in a prediction model consisting of a dense embedding layer, a dense input layer, an improved FM subnetwork (i.e., an FM subnetwork embedding interdisciplinary enhancement modules), a DNN subnetwork, and an output layer. Its structure is as follows:

[0086] 1. Input and feature processing layer (including dense embedding layer and dense input layer):

[0087] To efficiently handle the heterogeneous features of books and avoid invalid feature interactions, this embodiment has specifically designed the model's input path:

[0088] Sparse Categorical Features: Sparse categorical features consisting of the Chinese Library Classification number, publisher, author, and series name (already mapped to integer IDs) are input into the improved FM sub-network and simultaneously into a dense embedding layer. Each feature field is mapped into a low-dimensional dense vector (e.g., 16-dimensional) through an independent dense embedding layer. Multiple dense vectors obtained after processing by the dense embedding layer are then fed into both the improved FM sub-network and the DNN sub-network for subsequent feature crossing and deep feature extraction.

[0089] Dense numerical and dense semantic features: Normalized dense numerical features such as publication year, price, author popularity, and category popularity ranking, as well as dense semantic features constructed by dimensionality reduction through book title and subject word mapping, are received and concatenated through a dense input layer. The concatenated feature vector is only fed into the DNN sub-network and not into the improved FM sub-network.

[0090] Design Principles: The above design achieves a clear division of labor in the model's feature processing stages. The improved FM subnetwork focuses on processing sparse categorical features and explicit, interpretable low-order feature combinations between features (such as the combined effect of a specific publisher and a specific Chinese Library Classification number). The DNN subnetwork is responsible for automatically learning complex deep nonlinear patterns from the mixed representations of all features (including embedded dense features and original dense features). This path isolation avoids meaningless cross-operations of dense features in the improved FM subnetwork (such as the combination of a specific publisher and a certain price), which helps improve the overall computational efficiency and interpretability of the model.

[0091] 2. Improved FM subnetwork:

[0092] The improved FM subnetwork is used to learn sparse features and explicit cross features (such as classification number-publisher, classification number-author) that are crucial to the value of a book. Figure 2 As shown, the improved FM subnetwork extracts linear features of books from the sparse class features of the input, extracts second-order cross features of books from the dense vector output by the dense embedding layer, and then outputs the sum of the two to the output layer.

[0093] In book procurement, the subject classification (Chinese Library Classification number) of books is one of the key attributes influencing borrowing demand. To enable the model to effectively capture the cross-effects between subject attributes and other features (such as publisher and author), the core improvement in this embodiment is the introduction of a subject cross-enhancement module into the improved FM subnetwork. This module dynamically increases the contribution weight of feature cross-items involving the Chinese Library Classification number in the prediction. The specific working process of the subject cross-enhancement module is as follows:

[0094] S31. Constructing a mask matrix for subject-specific perception:

[0095] Define the mask matrix The rules for the values ​​of its elements are as follows:

[0096] ,

[0097] Where n is the total number of sparse class features, M st This represents the element in the s-th row and t-th column of the mask matrix M. This mask matrix is ​​used to identify all feature cross pairs involving subject-specific information.

[0098] S32, Computational Interdisciplinary Enhancement Factor:

[0099] Introduce a globally trainable parameter w β (w) β The initial value is 0), and the interdisciplinary enhancement factor β = σ(w) is calculated using the activation function σ (σ is the Sigmoid function). β The interdisciplinary enhancement factor is adaptively adjusted based on loss feedback during model training.

[0100] S33, Improved FM Subnetwork Output:

[0101] Improved FM subnetwork output y FM for:

[0102] ,

[0103] Among them, w s For the first-order feature weights, x s x t For eigenvalues, , These are the dense vectors corresponding to the s-th and t-th features, respectively. The term represents the inner product of dense vectors. The first summation term is a first-order linear term used to capture the independent influence of each feature. The second summation term is a standard second-order cross term used to capture the interactive influence of all feature pairs. The third summation term is a subject-specific cross-enhancement term, which, based on the standard second-order cross-enhancement term, uses a mask matrix M to filter out all subject-related feature cross-pairs, and then multiplies their inner product by the product of the eigenvalues ​​by a learnable subject-specific cross-enhancement factor β before summing them up.

[0104] Through the interdisciplinary enhancement module, the model can explicitly distinguish and strengthen the role of feature combinations containing subject attributes (such as CLC number-publisher, CLC number-author) in prediction. This design makes the model more aligned with the decision-making logic of prioritizing subject attributes in book procurement, thereby improving the accuracy and business adaptability of predictions.

[0105] The two targeted improvements mentioned above enable the model to more accurately quantify the potential borrowing value of books and more effectively capture the key impact of subject attributes on borrowing demand. Overall, they improve the accuracy and efficiency of borrowing popularity prediction and provide more reliable input for subsequent optimized procurement decisions under constraints.

[0106] 3. DNN sub-network:

[0107] The DNN subnetwork receives all dense vectors from the dense embedding layer and the dense input layer, mines deep nonlinear patterns from the concatenated mixed dense features, and extracts high-order features from the dense vectors. The DNN subnetwork contains three hidden layers, which are fully connected. In this embodiment, the number of neurons in each layer is 256, 128, and 64, respectively, and the ReLU activation function is used.

[0108] 4. Output layer:

[0109] The improved FM subnetwork outputs y FM With the output y of the DNN subnetwork DNN The sums are then mapped using the Sigmoid activation function to obtain the final normalized predicted borrowing popularity value: .

[0110] Model training:

[0111] Using the heterogeneous characteristics of the library's collection as input and the corresponding historical borrowing frequency as training labels, the improved DeepFM model is trained end-to-end in a supervised manner using the mean squared error (MSE) as the loss function and the gradient descent algorithm. During training, all parameters are simultaneously optimized, including the weights of the linear projection, the weights of the dense embedding layer, the weights of the DNN subnetwork, the weights of the improved FM subnetwork, and the globally trainable parameter w of the interdisciplinary enhancement module. β All parameters are saved after the model training reaches convergence or a predetermined accuracy. These parameters will be used for subsequent prediction of the borrowing popularity of candidate books.

[0112] Step 4: Predicting the borrowing popularity of candidate books:

[0113] Based on the metadata of the candidate book list (including Chinese Library Classification number, book title, author, publisher, etc.), and combined with the book borrowing data, sparse category features, dense numerical features, and dense semantic features of the candidate books are constructed according to step one. The sparse category features, dense numerical features, and dense semantic features of the candidate books are then input into the trained improved DeepFM model to obtain the predicted borrowing popularity of each book. The candidate books are then sorted in descending order according to the predicted borrowing popularity.

[0114] Step 5: Use a two-stage greedy optimization algorithm to determine the book list and recommended number of copies to be purchased.

[0115] After obtaining the predicted borrowing popularity of each candidate book in this step, based on the preset total procurement budget B and the minimum number of books to be purchased for each subject k L... k The maximum number of copies U for each candidate book i i Given constraints, a two-phase greedy optimization algorithm is executed to generate the final book procurement list and recommended number of copies to be purchased. The algorithm flow is as follows: Figure 3 As shown.

[0116] Phase 1: Minimum guaranteed procurement for each subject based on borrowing popularity.

[0117] The goal of the first phase is to prioritize meeting the minimum number of books L that are pre-set for each subject k, within the constraint of the total procurement budget B. k The requirements and specific process are as follows:

[0118] Input: A list of candidate books, where each candidate book includes its predicted borrowing popularity, price, and subject (determined by the Chinese Library Classification number);

[0119] S51. Sorting and Initialization: Sort candidate books in descending order of predicted borrowing popularity. Initialize the total cumulative amount P=0, and the number of selected books S for each subject k. k =0, the first phase of the procurement list List1 is empty;

[0120] S52. Sequential Processing and Selection: Process the candidate books in descending order as described above. For each book processed, perform the following steps in sequence:

[0121] S521. Determine if the loop termination condition is met:

[0122] If the current cumulative total amount P equals the total procurement budget B, or the number of selected books S across all disciplines. k All are equal to the corresponding minimum purchase quantity L k If the condition is met, the process will terminate immediately and proceed to step S53; otherwise, step S522 will be executed.

[0123] S522. Select one book i in sequence, with a corresponding price p. i ;

[0124] S523, Get the subject k to which the current book i belongs;

[0125] S524. Determine the subject constraint: Determine the number S of selected volumes for subject k. k Is it equal to its preset minimum purchase quantity L? k If yes, skip the current book i and return to step S522 to process the next book; otherwise, continue to step S525.

[0126] S525. Calculate budget feasibility: Calculate the expected total expenditure P' = P + p after purchasing the current book i. i ;

[0127] S526. Determine the budget constraint: Determine whether P' > B is satisfied. If yes, proceed to step S53; otherwise, proceed to step S527.

[0128] S527, Execute Procurement: Add the current book i to List1; update the cumulative total amount P=P'; update the number of books selected for procurement by subject, and let S k Add 1, then return to step S521;

[0129] S53. Output the first-stage procurement results, including the first-stage procurement list List1 and the updated cumulative total amount P, and calculate the remaining budget B'=BP.

[0130] Specifically, in the first phase, the algorithm first checks whether the termination condition has been met (budget exhausted or the minimum purchase quantity requirement for all subjects has been met). If not, it processes the book with the highest predicted borrowing popularity: if its subject has not yet met the minimum purchase quantity requirement, the book is purchased and its status is updated; if it has, the book is skipped. After processing, the algorithm returns to the beginning of the loop and checks the termination condition again, repeating this process until the termination condition is triggered or the budget is exhausted during the loop. In the actual application scenario of this method, the library will set a reasonable total purchase budget B (B>0) and a minimum purchase quantity L for each subject. k (at least one L) k >0). If parameters that do not conform to business logic are entered due to special circumstances (such as B=0), this method will still give a definite result (empty purchase list) according to the algorithm flow.

[0131] Phase Two: Optimized procurement of duplicate copies driven by marginal borrowing popularity.

[0132] If there is still a remaining budget after the first phase (i.e., B' > 0), then the second phase is executed. The goal of the second phase is to maximize the remaining budget and the maximum number of copies U for each candidate book i. i Under constraints, by introducing virtual books (i.e., duplicate copies) to simulate diminishing marginal returns, the total borrowing activity is maximized. The specific process of the second stage is as follows:

[0133] S54. Virtual Book Generation: Generate virtual books representing copies of all candidate books. In this embodiment, the maximum number of copies of each candidate book i is the corresponding U. i U i This indicates that the maximum number of copies that can be purchased for each candidate book may vary. For example, a maximum of 2 copies can be purchased for computer science books, and a maximum of 5 copies for novels. Each candidate book is generated according to its corresponding maximum number of copies (U). i -1) This virtual book. The predicted borrowing popularity y of the m-th virtual book. i (m) According to formula y i (m) = α m ·y i The attenuation calculation is that the predicted borrowing popularity of the m-th virtual book is the m-th power of α, which is the original predicted borrowing popularity of the corresponding candidate book, where y i The predicted borrowing popularity of candidate book i is given by α, which is a preset attenuation coefficient where 0 < α < 1. In this embodiment, α is set to 0.6, but it can also be set according to actual needs.

[0134] S55. List Merging, Sorting and Initialization: The books in the second phase include candidate books not purchased in the first phase and virtual books of all candidate books. Therefore, the candidate books not purchased in the first phase are merged with all virtual book lists generated in step S54 and sorted in descending order according to the predicted borrowing popularity (for virtual books, the attenuated predicted borrowing popularity). The cumulative total amount P is initialized to the final value of the first phase, and the second phase purchase list List2 is initialized to List1.

[0135] S56. Process in order: Process each book i (candidate book or virtual book) in descending order of the arrangement in step S55, and perform the following steps on the currently processed book:

[0136] S561. Determine the termination condition:

[0137] If the current cumulative total amount P is equal to the total procurement budget B, the process is terminated immediately and the process is skipped to step S57; otherwise, step S562 is executed.

[0138] S562. Select Books: Select one book i in sequence, with a corresponding price p. i ;

[0139] S563. Calculate the cumulative amount: Calculate the expected total expenditure P' after purchasing the current book i = P + p i ;

[0140] S564. Determine budget constraints: Determine whether P' is greater than the total procurement budget B. If yes, proceed to step S57; otherwise, proceed to step S565.

[0141] S565. Execute the purchase: Add the current book i to List2 and update the total accumulated amount P=P'. Then return to step S561.

[0142] S57. Generate Final Results: Based on List2, count the total number of times each book was selected (including the number of times candidate books and their corresponding virtual books from the first and second stages were selected). This total number is the final recommended number of copies to be purchased for the candidate books. Output the final book list and the recommended number of copies to be purchased.

[0143] Specifically, the second-stage algorithm first checks whether the termination condition has been met (i.e., whether the cumulative total amount P has reached the total procurement budget B). If not, it processes the candidate book or virtual book with the highest predicted borrowing popularity (including the predicted borrowing popularity of decaying virtual books): it is directly added to the second-stage procurement list, and its price is added to the cumulative total amount P. After processing, the algorithm returns to the beginning of the loop and checks the termination condition again, repeating this process until the budget is exhausted. During this process, all candidate books and virtual books are sorted in descending order of predicted borrowing popularity, and subject constraints are not considered. Its core objective is to maximize the total borrowing popularity within the remaining budget.

[0144] Example 2

[0145] A book procurement system based on borrowing popularity, applying the method of Embodiment 1 above, specifically includes:

[0146] The feature construction module is used to collect metadata and historical borrowing data of the library's collection of books, and to construct heterogeneous features for each book in the collection based on the metadata and historical borrowing data. The heterogeneous features include sparse categorical features, dense numerical features and dense semantic features.

[0147] The tag generation module is used to calculate the borrowing popularity of each book in the library collection within a preset time window based on the historical borrowing data, and use it as training tags.

[0148] The model training module is used to construct an improved DeepFM model, and train the model with the heterogeneous features as input and the borrowing popularity as output. The improved DeepFM model includes a dense input layer, a DNN sub-network, and an improved FM sub-network. The dense numerical features and dense semantic features are processed by the dense input layer and then input into the DNN sub-network but not into the improved FM sub-network. The improved FM sub-network introduces a subject cross-enhancement module to enhance the weights of feature cross-terms containing the Chinese Library Classification number.

[0149] The borrowing popularity prediction module is used to obtain the predicted borrowing popularity of each candidate book based on the metadata of the candidate books and using the trained improved DeepFM model.

[0150] The intelligent procurement decision module is used to generate the final book procurement list and suggested procurement number based on the predicted borrowing popularity, the preset total procurement budget, the minimum number of books to be purchased for each subject and the upper limit of duplicate copies, by executing a two-stage greedy optimization algorithm.

[0151] Example 3

[0152] A storage medium storing a program, which, when executed by a processor, implements the intelligent book purchasing method based on borrowing popularity described in Embodiment 1, as follows:

[0153] S1. Collect metadata and historical borrowing data of the library's collection of books, and construct heterogeneous features for each book in the collection based on the metadata and historical borrowing data. The heterogeneous features include sparse category features, dense numerical features and dense semantic features.

[0154] S2. Calculate the borrowing popularity of each book in the library collection within a preset time window based on the historical borrowing data, and use it as a training label;

[0155] S3. Construct an improved DeepFM model, and train the model using the heterogeneous features as input and the borrowing popularity as output; wherein, the improved DeepFM model includes a dense input layer, a DNN sub-network and an improved FM sub-network, the dense numerical features and dense semantic features are processed by the dense input layer and then input into the DNN sub-network but not into the improved FM sub-network; the improved FM sub-network introduces a subject cross-enhancement module to enhance the weights of feature cross-terms containing the Chinese Library Classification number;

[0156] S4. Based on the metadata of the candidate books, use the trained improved DeepFM model to obtain the predicted borrowing popularity of each candidate book.

[0157] S5. Based on the predicted borrowing popularity, the preset total procurement budget, the minimum number of books to be purchased for each subject, and the upper limit of duplicate copies, execute a two-stage greedy optimization algorithm to generate the final book procurement list and the suggested number of books to be purchased.

[0158] The specific processing steps described above are as in Example 1 and will not be repeated here.

[0159] In this embodiment, the storage medium may be a disk, optical disk, computer memory, read-only memory, random access memory, USB flash drive, portable hard drive, or other media.

[0160] The above embodiments are preferred embodiments of the present invention and are not intended to limit the present invention. Any changes or other equivalent substitutions made without departing from the technical solution of the present invention are included within the protection scope of the present invention.< / unk>

Claims

1. A book procurement method based on borrowing popularity, characterized in that, Includes the following steps: S1. Collect metadata and historical borrowing data of the library's collection of books, and construct heterogeneous features for each book in the collection based on the metadata and historical borrowing data. The heterogeneous features include sparse category features, dense numerical features and dense semantic features. S2. Calculate the borrowing popularity of each book in the library collection within a preset time window based on the historical borrowing data, and use it as a training label; S3. Construct an improved DeepFM model, and train the model using the heterogeneous features as input and the borrowing popularity as output; wherein, the improved DeepFM model includes a dense input layer, a DNN sub-network, and an improved FM sub-network, the dense numerical features and dense semantic features are processed by the dense input layer and then input into the DNN sub-network but not into the improved FM sub-network; the improved FM sub-network introduces a subject cross-enhancement module to enhance the weights of feature cross-terms containing the Chinese Library Classification number; the sparse category features are input into the improved FM sub-network, and simultaneously processed by the dense embedding layer of the improved DeepFM model to obtain a dense vector, the dense vector being simultaneously input into the improved FM sub-network and the DNN sub-network; S4. Based on the metadata of the candidate books, use the trained improved DeepFM model to obtain the predicted borrowing popularity of each candidate book. S5. Based on the predicted borrowing popularity, the preset total procurement budget, the minimum number of books to be purchased for each subject and the upper limit of duplicate copies, execute a two-stage greedy optimization algorithm to generate the final book procurement list and the suggested number of books to be purchased. The working process of the interdisciplinary enhancement module includes: S31. Constructing a mask matrix for subject-specific perception: ， Where n is the total number of sparse class features, M st Let M be the element in the s-th row and t-th column of the mask matrix M, where M is related to a subject when the intersection of the s-th feature and the t-th feature is related to that subject. st =1, otherwise M st =0; S32. Introduce trainable parameters w β Calculate the interdisciplinary enhancement factor β: β=σ(w β ), where σ is the Sigmoid function; S33, Improved FM Subnetwork Output y FM for: ， Among them, w s For the first-order feature weights, x s x t For eigenvalues, , These are the dense vectors corresponding to the s-th and t-th features, respectively. It is the inner product of dense vectors.

2. The intelligent book procurement method based on borrowing popularity according to claim 1, characterized in that, In step S1, the sparse category features include the Chinese Library Classification number, publisher, author, and series name; the dense numerical features include the publication year, price, average number of books borrowed by the author, and the recent borrowing popularity ranking of the category; the dense semantic features include the book title and keywords.

3. The intelligent book procurement method based on borrowing popularity according to claim 2, characterized in that, Constructing a recent borrowing popularity ranking for the aforementioned category includes: calculating the quantile of the total number of borrowings of books belonging to the Chinese Library Classification number c within a preset time window, ranking them among all Chinese Library Classification numbers in the library's collection. R norm = ， Among them, the quantile R norm Used to represent the recent borrowing popularity ranking of books in a category, N is the total number of Chinese Library Classification numbers for all books in the library's collection, R c It is the ranking of the total number of times a book belonging to the Chinese Library Classification number C was borrowed within a preset time window among all the books in the library's collection under all Chinese Library Classification numbers.

4. The intelligent book procurement method based on borrowing popularity according to claim 1, characterized in that, The two-stage greedy optimization algorithm includes: In the first phase, within the constraint of the total procurement budget B, books are selected in descending order of predicted borrowing popularity, prioritizing the fulfillment of the minimum pre-set procurement quantity L for each subject. k Require; In the second stage, if there is a surplus in the total procurement budget after the first stage, virtual books with representative copies are generated for the candidate books and selected greedily according to the predicted borrowing popularity after decay, so as to maximize the total borrowing popularity and determine the final list of books to be procured and the recommended number of books to be procured.

5. The intelligent book procurement method based on borrowing popularity according to claim 4, characterized in that, The specific process of the first stage includes: S51. Sort the candidate books in descending order of predicted borrowing popularity, and initialize the cumulative total amount P=0 and the number of selected books S for each subject k. k =0, the first phase of the procurement list is empty; S52. Process each book in the sorted list in descending order, and perform the following steps for the currently processed book: S521. Determine if the termination condition is met: If the cumulative total amount P equals the total procurement budget B, or the number of selected books for all subjects S. k All are equal to the corresponding minimum purchase quantity L k If the condition is met, the process terminates and proceeds to step S53; otherwise, step S522 is executed. S522. Select one book i in sequence, with a price of p. i ; S523, Obtain the subject k to which the current book belongs; S524. Determine the number of books S selected for subject k. k Is it equal to the minimum number of copies L to be purchased? k If yes, skip the current book and return to step S522 to process the next book; otherwise, proceed to step S525. S525. Calculate the expected total expenditure after purchasing the current book: P' = P + p i ; S526. Determine whether P' is greater than the total procurement budget B. If yes, proceed to step S53; otherwise, proceed to step S527. S527. Add the current book to the first phase of the procurement list, update the cumulative total amount P=P', and let S k Add 1, then return to step S521; S53. Output the first phase procurement list as the first phase procurement result.

6. The intelligent book procurement method based on borrowing popularity according to claim 4, characterized in that, The specific process of the second stage includes: S54. Generate a U for each candidate book i i -1 virtual book, of which U i Let y be the upper limit of the number of copies of the candidate books, and let y be the predicted borrowing popularity of the m-th virtual book. i (m) =α m y i y i The predicted borrowing popularity of candidate books is given by α, which is a preset decay coefficient and 0 < α < 1. S55. Sort the candidate books and all virtual books that were not purchased in the first phase in descending order of predicted borrowing popularity; initialize the cumulative total amount P to the final value of the first phase, and initialize the procurement list for the second phase to the procurement list for the first phase. S56. Process each book in the order arranged in step S55 in descending order, and perform the following steps for the currently processed book: S561. If the total cumulative amount P equals the total procurement budget B, then terminate the process and proceed to step S57; otherwise, proceed to step S562. S562. Select one book i in sequence, with a price of p. i ; S563. Calculate the expected total expenditure after purchasing the current book: P' = P + p i ; S564. Determine whether P' is greater than the total procurement budget B. If yes, proceed to step S57; otherwise, proceed to step S565. S565. Add the current book to the second phase of the purchase list, update the total cumulative amount P=P', and then return to step S561; S57. Based on the second-stage procurement list, merge the candidate books and virtual books to obtain the final book procurement list and the recommended number of books to be procured.

7. A book intelligent procurement system based on borrowing popularity, applied to the method described in any one of claims 1-6, characterized in that, include: The feature construction module is used to collect metadata and historical borrowing data of the library's collection of books, and to construct heterogeneous features for each book in the collection based on the metadata and historical borrowing data. The heterogeneous features include sparse categorical features, dense numerical features and dense semantic features. The tag generation module is used to calculate the borrowing popularity of each book in the library collection within a preset time window based on the historical borrowing data, and use it as training tags. The model training module is used to construct an improved DeepFM model, and train the model with the heterogeneous features as input and the borrowing popularity as output. The improved DeepFM model includes a dense input layer, a DNN sub-network, and an improved FM sub-network. The dense numerical features and dense semantic features are processed by the dense input layer and then input into the DNN sub-network but not into the improved FM sub-network. The improved FM sub-network introduces a subject cross-enhancement module to enhance the weights of feature cross-terms containing the Chinese Library Classification number. The borrowing popularity prediction module is used to obtain the predicted borrowing popularity of each candidate book based on the metadata of the candidate books and using the trained improved DeepFM model. The intelligent procurement decision module is used to generate the final book procurement list and suggested procurement number based on the predicted borrowing popularity, the preset total procurement budget, the minimum number of books to be purchased for each subject and the upper limit of duplicate copies, by executing a two-stage greedy optimization algorithm.

8. A storage medium storing a program, characterized in that, When the program is executed by the processor, it implements a book intelligent procurement method based on borrowing popularity as described in any one of claims 1-6.

Images