A personalized recommendation system and recommendation method based on real-time library context

Abstract

The present invention belongs to the technical field related to information recommendation and discloses a personalized recommendation system and method based on real-time library context. In the system, a user attribute fusion module performs feature weighted fusion on attribute features; a context information fusion module performs weighted fusion on library context information features; a user preference identification module calculates similarity feature vectors between candidate book representations and historical book representations and calculates the weights of the corresponding historical books. All historical book representations are weighted and input into a pooling layer to obtain reading preference features; and a recommendation probability calculation module concatenates the features and then passes them through a linear layer, an activation layer, and a fully connected layer in sequence to obtain the recommendation probability of the candidate book. This system not only captures users' long-term preferences and individual characteristics but also provides users with more tailored recommendations in real-time context, thereby generating highly personalized recommendation results.

Description

Technical Field

[0001] The present invention belongs to the technical field related to information recommendation, and more specifically, relates to a personalized recommendation system and recommendation method based on real-time library context. Background Art

[0002] With the rapid development of information technology and the widespread adoption of digital libraries, the number and variety of library resources have grown exponentially. Libraries not only include a wide range of professional books, journals, and academic papers, but also continuously update and introduce a wide variety of digital resources, such as e-books, audiobooks, and video tutorials. This massive accumulation of information places immense pressure on users to sift through the information when searching for specific resources. Helping readers find the content they need within this rich library has become a pressing issue.

[0003] Traditional book recommendation systems primarily rely on collaborative filtering or content matching techniques. Collaborative filtering recommends resources by identifying the reading history of other readers with similar interests, while content matching makes recommendations based on the similarity between a user's past borrowed content and the content of the resource. These methods can meet basic user needs to a certain extent, but with the increasing diversification of reader needs and the increasing richness of library resources, traditional recommendation methods tend to lead to generalized and repetitive recommendations, failing to dynamically respond to user needs. Summary of the Invention

[0004] In response to the above-mentioned defects or improvement needs of the prior art, the present invention provides a personalized recommendation system and recommendation method based on the real-time context of the library, which aims to dynamically respond to the personalized needs of users and recommend books that meet their personalized needs.

[0005] To achieve the above objectives, the present invention provides a personalized recommendation system based on real-time library context, which includes:

[0006] The user attribute fusion module is used to construct the attribute features of the current user and input them into the feature weighted fusion network unit to obtain attribute fusion features; the attributes include the user's gender and age;

[0007] A context information fusion module is used to construct context information features of the current library and input them into a feature weighted fusion network unit to obtain context fusion features; the context information includes the current library's geographical location, current visit time, and current environment;

[0008] The user preference identification module is used to construct candidate book representations and representations of historical books that the current user has read, calculate the similarity feature vectors of the candidate book representations and historical book representations, and then concatenate the candidate book representations, historical book representations, and the similarity feature vectors of the two. After passing them through the activation layer and the linear layer, the weights of the corresponding historical books are obtained. After all the historical book representations are weighted, they are input into the pooling layer to obtain the reading preference features corresponding to the candidate books;

[0009] A recommendation probability calculation module is used to concatenate the attribute fusion features, context fusion features, reading preference features, and candidate book representations, and then pass them through a linear layer, an activation layer, and a fully connected layer in sequence to obtain the recommendation probability of the candidate book;

[0010] Among them, the feature weighted fusion network unit is used to perform pairwise interaction on the input features to obtain interactive features, each of the interactive features is input into the linear layer and the activation layer in turn to obtain the weight of the corresponding interactive feature, and all the interactive features are weighted summed to achieve feature fusion.

[0011] Optionally, the weight of the historical book and / or the weight of the interactive feature is a normalized weight.

[0012] Optionally, the attributes further include education level, profession and interests and hobbies.

[0013] Optionally, the context information also includes the subject area where the user is located in the library.

[0014] Optionally, the representation of each book includes a vector formed by concatenating the book ID, the title text, and the encoding of the book category.

[0015] Optionally, similar feature vectors of the candidate book representation and the historical book representation are obtained by calculating the outer product of the candidate book representation and the historical book representation.

[0016] Optionally, for feature v i x i and feature v j x j Perform two-to-two interactions to obtain the interaction feature f ij The calculation formula is:

[0017] f ij ={(v i ⊙v j )x i x j}

[0018] Where, v i is the embedding vector of the i-th attribute, x i is the value of the i-th attribute, v j is the embedding vector of the jth attribute, xj is the value of the j-th attribute.

[0019] The present invention also provides a personalized recommendation method based on real-time library context, which includes:

[0020] Acquiring input data, wherein the input data includes attributes of the current user, context information of the current library, candidate books, and historical books that the current user has consulted;

[0021] Inputting the input data into the trained personalized recommendation system based on real-time library context as described in any one of the above items to obtain a recommendation probability for each candidate book;

[0022] The candidate books are ranked according to the size of the recommendation probability, and the greater the probability, the higher the ranking.

[0023] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.

[0024] The present invention also provides a computer program product, comprising a computer program or instructions, which implement the steps of the above method when executed by a processor.

[0025] In general, the above technical solutions conceived by the present invention have the following beneficial effects compared with the prior art:

[0026] 1. This paper proposes a personalized recommendation system based on real-time library context. The system includes a user attribute fusion module, a context information fusion module, a user preference identification module, and a recommendation probability calculation module. By concatenating attribute fusion features, context fusion features, reading preference features, and candidate book representations, a final concatenation vector is generated. This vector not only incorporates the user's long-term reading preferences but also incorporates specific interests in the current context, thereby more accurately expressing the user's overall preference for candidate books. This ensures that the recommendation system not only captures the user's long-term preferences and individual characteristics, but also provides recommendations that better meet the user's needs in real-time context, thereby generating highly personalized recommendation results.

[0027] 2. In this invention, both the user attribute fusion module and the context information fusion module effectively capture the potential relationships between high-dimensional features through feature crosstalk, thereby fully utilizing multidimensional information. Weights are then automatically learned based on interaction features, assigning different weights to different interaction features. This allows the relative importance of multidimensional attributes to be distinguished, leading to more accurate identification and differentiation of core features.

[0028] 3. This invention introduces learnable linear and activation layers when acquiring the weights of user and contextual features. Compared with the direct multiplication of the QKV matrix in the traditional Transformer encoder, the standard model requires only one-tenth of the number of parameters and greatly improves the online inference efficiency while maintaining similar inference results. This is suitable for scenarios based on real-time library recommendations and can be deployed in offline hosts with higher real-time requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a structural diagram of a personalized recommendation system in one embodiment of the present invention;

[0030] Figure 2 is a data processing flow chart of a feature weighted fusion network unit in one embodiment of the present invention;

[0031] Figure 3 FIG. 4 is a diagram of the data processing process of a user preference identification module in one embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0033] Example 1

[0034] The present invention provides a personalized recommendation system based on real-time library context, such as Figure 1 The figure shows a structural diagram of a personalized recommendation system in one embodiment of the present invention, which includes a user attribute fusion module, a context information fusion module, a user preference identification module and a recommendation probability calculation module. Each module is introduced in detail below.

[0035] The user attribute fusion module is used to construct the attribute features of the current user and input them into the feature weighted fusion network unit to obtain attribute fusion features. The attributes include the user's gender and age.

[0036] Specifically, the user attribute fusion module includes a user attribute feature construction unit and a feature weighted fusion network unit connected in sequence. The user attribute feature construction unit is used to construct the current user attribute feature (embedded representation) based on the current user's attribute information and input it into the subsequent feature weighted fusion network unit for feature fusion. The feature weighted fusion network unit is used to perform pairwise interaction on the input features to obtain interaction features, and each interaction feature is input into the linear layer and activation layer in sequence to obtain the weight of the corresponding interaction feature. The weighted sum of all interaction features is then performed to achieve feature fusion.

[0037] like Figure 2 The figure shows a data processing flow chart of a feature weighted fusion network unit in one embodiment of the present invention. The feature weighted fusion network unit includes an interaction unit, a weight identification network, and a weighted summation unit. The interaction unit is used to perform pairwise interaction on the features output by the feature construction unit to obtain interaction features. The weighted network unit includes a linear layer and an activation layer. Each interaction feature output by the interaction unit passes through the linear layer and the activation layer in sequence to obtain the weight of the corresponding interaction feature. The summation unit is used to perform weighted summation on each interaction feature to achieve feature fusion.

[0038] Specifically, for feature v i x i and feature vjx j Perform two-to-two interactions to obtain the interaction feature f ij The calculation formula is: ij ={(v i ⊙v j )x i x j}, where v i is the embedding vector of the i-th attribute, x i is the value of the i-th attribute, v j is the embedding vector of the jth attribute, x j is the value of the j-th attribute.

[0039] In a specific embodiment, in addition to gender and age, attributes may also include education level, major, interests and hobbies, etc., and the above information can be read in the library cloud data warehouse.

[0040] Traditional recommendation systems usually only distinguish user characteristics by item ID, and fail to fully utilize the user's multidimensional attribute information, such as gender, age, education level, profession, and interests and hobbies. This method is difficult to accurately portray the user's multifaceted preferences and personalized needs. In fact, the user's attribute characteristics contain rich information (such as gender, age, education level, profession, and interests and hobbies, etc.). Simply relying on ID distinction ignores these details and cannot provide a comprehensive user portrait for the recommendation system. To this end, the present invention introduces a user attribute fusion module, which effectively captures the potential relationship between high-dimensional features through feature cross-linking, thereby making full use of multidimensional information, and then automatically learns weights based on interactive features, assigning different weights to different interactive features, thereby distinguishing the relative importance of multidimensional attributes. For example, the interaction between the user's age and interests and hobbies may reflect their preferences better than other feature combinations, while the interactive effect of gender and education level is relatively weak. Through this method, the model can more accurately identify and distinguish core features. In addition, the weight recognition network introduces learnable linear layers and activation layers, which only requires one-tenth of the parameters under the standard model and greatly improves the online reasoning efficiency while maintaining similar reasoning effects. It is suitable for scenarios based on real-time library recommendations and can be deployed in offline hosts with higher real-time requirements.

[0041] by Figure 2 Take this as an example to illustrate. The input data of user attributes is X = {x1, x2, x3, x4, x5, x6}, and each value represents the numerical value of a different attribute. For example, if the attribute is a categorical attribute, such as education level, the values ​​of elementary school, junior high school, high school, and university can be taken as 0.2, 0.4, 0.6, and 0.8 respectively; if the attribute is missing, enter 0 directly; for example, if the gender is male and the gender is female, the value is 1 if it belongs to the category, and 0 if it does not belong; if the attribute is a numerical attribute, such as age, the value can be the age value or the normalized age value. The user attribute feature construction unit obtains the embedded representation of the attribute based on the input attribute data ε = {v i x i} i∈x , v i is the trainable embedding representation of the corresponding attribute, x i is the value of the corresponding attribute. After the interaction unit crosses the embedded vectors, the interaction feature f is obtained. ij ={(v i ⊙v j )x i x j}. The interaction feature f ij Input the trainable linear layer and activation layer in sequence to obtain the interaction feature f ij The summation unit performs weighted summation on each interactive feature to achieve feature fusion.

[0042] In one embodiment, the activation layer selects the ReLU activation function, the weight is the normalized weight, and the weight a ij The acquisition process can be expressed as:

[0043] a' ij =hT T ReLU(W(v i ⊙v j )x i x j +b)

[0044]

[0045] Where h T is a learnable weight vector. For example, the output of the ReLU activation layer is a 4*1 vector, so h T It is a 1*4 learnable parameter vector, which is used to transform the dimension into 1*1. The calculated result a' ij Is a scalar value. W and b are the trainable vectors and biases in the linear layer. is an unordered set of any two combinations of i and j.

[0046] The context information fusion module is used to construct the context information features of the current library and input them into the feature weighted fusion network unit to obtain the context fusion features. The context information includes the geographical location of the current library, the current visit time and the current environment.

[0047] Specifically, the context information fusion module includes a context information feature construction unit and a feature weighted fusion network unit, which are connected in sequence. The context information feature construction unit is used to construct the current library's context information features (embedded representation) based on the current library's context information and input them into the subsequent feature weighted fusion network unit for feature fusion. The feature weighted fusion network unit is used to perform pairwise interaction on the input features to obtain interaction features, and each interaction feature is input into the linear layer and activation layer in sequence to obtain the weight of the corresponding interaction feature. The weighted sum of all interaction features is then performed to achieve feature fusion.

[0048] In a specific embodiment, the context information may include, in addition to the current library's geographical location, current visit time, and current environment, information such as the subject area the user is in the library.

[0049] Specifically, the feature weighted fusion network unit in the context information fusion module can refer to the introduction of the feature weighted fusion network unit in the user attribute fusion module, which will not be repeated here.

[0050] In the invention, in addition to focusing on the long-term attributes of users, real-time contextual information is also introduced. This contextual information can help the recommendation system generate more personalized recommendation content in the specific context of the user's current situation.

[0051] In this embodiment, context information can be divided into the following three categories.

[0052] The first category is the current library's geographic location and current visit time, including the user's city, library name, season, month, and specific visit time. This type of macro information not only provides a geographical and temporal context for recommendations, but also can influence the user's borrowing preferences from a broader perspective. For example, in winter, users may be more inclined to borrow literature and philosophy books, while during specific holidays, travel or relaxing reading books may be more popular. This type of time and location information is encoded as a user's macro-contextual feature, which helps the model better understand the changes in users' reading needs in different seasons and geographical environments.

[0053] The second category is the current environment, such as weather, temperature, and humidity. These external contexts significantly influence users' immediate reading preferences. On sunny and warm days, users may prefer more relaxing reading, while in cooler or quieter environments, they may prefer in-depth reading. These objective factors, as contextual characteristics, influence users' real-time reading preferences.

[0054] The third category is the user's real-time browsing and search behavior, including the content of the book they are currently viewing, search keywords, and the physical area of ​​the library they are in (e.g., literature section, science section, etc.). This type of information is crucial in the context of book recommendations because it instantly reflects the user's current interests and needs. Specifically, real-time browsing and search behavior directly reveals the user's current reading preferences and focus. For example, when a user is browsing books in a specific field or entering specific search keywords, the system can instantly capture these dynamic signals and understand the user's current interests. This real-time information not only compensates for the shortcomings of traditional recommendation systems in modeling users' long-term and short-term interests, but also enables rapid adjustment of recommendation strategies as user behavior changes. By monitoring and analyzing users' reading and search behavior in real time, the recommendation system can capture these changes in real time and provide recommendations that are more tailored to the user's immediate needs. In the context of library recommendations, real-time behavior information, after being processed with expanded dimensions, significantly enhances its representational power in the model. Using these real-time features as the primary input, the system can accurately model users' short-term interest dynamics. For example, if a user is searching for a science book or frequently browses a specific subject, the recommendation system can respond immediately and prioritize highly relevant resources, thereby improving user satisfaction and the relevance of recommendations. This is especially true given the user's physical location within the library. This is the biggest difference between traditional e-commerce recommendations and real-time library recommendations. Traditional e-commerce recommendations cannot utilize this kind of contextual information and can only make recommendations based on user and interaction history.

[0055] The user preference identification module is used to construct candidate book representations and representations of historical books that the current user has read, calculate the similarity feature vectors of the candidate book representations and historical book representations, and concatenate the candidate book representations, historical book representations and the similarity feature vectors of the two. After passing through the activation layer and the linear layer in sequence, the weights of the corresponding historical books are obtained. After all the historical book representations are weighted, they are input into the pooling layer to obtain the reading preference features corresponding to the candidate books.

[0056] like Figure 3 The following diagram illustrates the data processing process of the user preference identification module in one embodiment of the present invention. First, the book information (such as book ID, title, and category) in the user's interaction history is mapped and concatenated to form a historical sample, thereby forming a sequence of the user's reading behavior. Correspondingly, the features of candidate books are generated through the same process to facilitate consistency comparison and correlation with historical data.

[0057] In order to determine the probability of the user's interest in the current candidate book, this embodiment performs a dot product operation on each book in the history sequence and the candidate book, and then performs splicing, activation, and linear mapping. In this way, the matching degree (weight) of each historical book and the candidate book can be calculated item by item to highlight the correlation between the user and the candidate book. In one embodiment, the similarity feature vectors of the candidate book representation and the historical book representation are obtained by calculating the outer product of the candidate book representation and the historical book representation. For example, the encoding of the book ID, title text, and book category are respectively represented as e ID ,e text ,e cate , then there are historical books representing: e k =concat(e ID ,e text ,e cate ). Let the representation of the candidate book be V U (A), then calculate the candidate books V U (A) Interact with any history book k The formula for similarity weight is: Then, all historical book representations are weighted and input into the pooling layer to obtain the reading preference features corresponding to the candidate books.

[0058] The recommendation probability calculation module is used to splice the attribute fusion features, context fusion features, reading preference features and candidate book representations and then pass them through the linear layer, activation layer and fully connected layer in sequence to obtain the recommendation probability of the candidate book.

[0059] In a specific embodiment, the activation layer in the recommendation probability calculation module selects the PReLU activation function or the dice activation function. Due to the inconsistent quality of book data, when the feature values ​​of the data set input are relatively sparse, dice can be considered. If the data set quality is high and there are fewer missing features, PRelu is used.

[0060] In this paper, attribute fusion features, context fusion features, reading preference features, and candidate book representations are concatenated to produce a final concatenated vector. This vector not only captures the user's long-term reading preferences but also incorporates specific interests in the current context, more accurately expressing the user's overall preference for candidate books. This ensures that the recommendation system not only captures the user's long-term preferences and individual characteristics but also provides recommendations that better meet the user's needs in real-time contexts. Further calculations are performed on the final output vector to determine the user's preference probability for the candidate books. By comparing the recommendation probabilities of all candidate books, the candidate books can be ranked, with higher recommendation probabilities resulting in higher rankings.

[0061] In general, the present invention extracts features from user information, fully exploits multidimensional attributes and contextual features, and then fuses user representations, historical interaction records, real-time contextual features, and candidate items to ultimately output the recommendation probability of the target book, thereby significantly improving the relevance and real-time nature of the recommendations and generating highly personalized recommendation results.

[0062] Example 2

[0063] The present invention also relates to a personalized recommendation method based on real-time library context, which comprises:

[0064] Acquiring input data, wherein the input data includes attributes of the current user, context information of the current library, candidate books, and historical books that the current user has consulted;

[0065] Inputting the input data into a trained personalized recommendation system based on real-time library context to obtain a recommendation probability for each candidate book;

[0066] The candidate books are ranked according to the size of the recommendation probability, and the greater the probability, the higher the ranking.

[0067] The personalized recommendation system is the one described in Example 1. After training, the system can quickly output the recommendation probability of each candidate book after receiving input data through processing by various modules, thereby achieving book recommendations. It is understood that the recommendation system can be trained using traditional network training methods, and the training samples can be historical user book selection records.

[0068] Example 3

[0069] The present invention also relates to a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when the computer program is executed by a processor.

[0070] Specifically, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0071] Example 4

[0072] An embodiment of the present invention provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the method of the above embodiment of the present invention.

[0073] The technical features of the above-described embodiments may be combined in any manner. To simplify the description, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. It should be noted that the phrases "in one embodiment," "for example," "and another example," etc., of the present invention are intended to illustrate the present invention and are not intended to limit the present invention.

[0074] The above-described embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A personalized recommendation system based on real-time library context, characterized by: include: User attribute fusion module, used to construct the attribute features of the current user and input them into the feature weighted fusion network unit to obtain attribute fusion features; The attributes include the user's gender and age; A context information fusion module is used to construct context information features of the current library and input them into a feature weighted fusion network unit to obtain context fusion features; the context information includes the current library's geographical location, current visit time, and current environment; The user preference identification module is used to construct candidate book representations and representations of historical books that the current user has read, calculate the similarity feature vectors of the candidate book representations and historical book representations, and then concatenate the candidate book representations, historical book representations, and the similarity feature vectors of the two. After passing them through the activation layer and the linear layer, the weights of the corresponding historical books are obtained. After all the historical book representations are weighted, they are input into the pooling layer to obtain the reading preference features corresponding to the candidate books; A recommendation probability calculation module is used to concatenate the attribute fusion features, context fusion features, reading preference features, and candidate book representations, and then pass them through a linear layer, an activation layer, and a fully connected layer in sequence to obtain the recommendation probability of the candidate book; Among them, the feature weighted fusion network unit is used to perform pairwise interaction on the input features to obtain interactive features, each of the interactive features is input into the linear layer and the activation layer in turn to obtain the weight of the corresponding interactive feature, and all the interactive features are weighted summed to achieve feature fusion.

2. The personalized recommendation system based on real-time library context as claimed in claim 1, characterized in that: The weight of the historical book and / or the weight of the interactive feature are normalized weights.

3. The personalized recommendation system based on real-time library context as claimed in claim 1, characterized in that: The attributes also include education, profession, and interests and hobbies.

4. The personalized recommendation system based on real-time library context as claimed in claim 1, characterized in that: The context information also includes the subject area where the user is located in the library.

5. The personalized recommendation system based on real-time library context according to any one of claims 1 to 4, characterized in that: The representation of each book includes a vector composed of the book ID, title text, and the encoding of the book category.

6. The personalized recommendation system based on real-time library context according to any one of claims 1 to 4, characterized in that: The similarity feature vectors of the candidate book representation and the historical book representation are obtained by calculating the outer product of the candidate book representation and the historical book representation.

7. The personalized recommendation system based on real-time library context according to any one of claims 1 to 4, characterized in that: For feature v i x i and feature v j x j Perform two-to-two interactions to obtain the interaction feature f ij The calculation formula is: f ij ={(v i ⊙v j )x i x j } Where, v i is the embedding vector of the i-th attribute, x i is the value of the i-th attribute, v j is the embedding vector of the jth attribute, x j is the value of the j-th attribute.

8. A personalized recommendation method based on real-time library context, characterized in that: include: Acquiring input data, wherein the input data includes attributes of the current user, context information of the current library, candidate books, and historical books that the current user has consulted; Inputting the input data into the trained personalized recommendation system based on library real-time context according to any one of claims 1 to 7 to obtain a recommendation probability for each candidate book; The candidate books are ranked according to the size of the recommendation probability, and the greater the probability, the higher the ranking.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to claim 8 are implemented.

10. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instructions are executed by a processor, the steps of the method according to claim 8 are implemented.

Images