Book recommendation method and system based on knowledge graph and multi-objective optimization

Abstract

The invention belongs to the technical field of book recommendation, and discloses a book recommendation method based on a knowledge graph and multi-objective optimization, which utilizes a multi-task learning framework and a ripple propagation network to learn user and project representation at the same time. In order to more accurately obtain the degree of interest of the user on the article and the potential preference of the user, a ripple propagation network is introduced to extract key information from the browsing history of the user to mine the potential interest preference of the user, and the representation of the user entity is obtained. The high-order semantic propagation attenuation is then solved by adopting a gated loop unit (GRU) combined with an attention mechanism, and then a final representation of the user is generated. The method aims at solving the problems that most book recommendation methods at present only pay attention to the accuracy of recommendation results, but only consider that the accuracy cannot meet the personalized requirements of readers, and only one index is used for measuring the advantages and disadvantages of a recommendation algorithm, so that a multi-target optimization algorithm is introduced to further optimize a recommendation list; and obtaining a final recommendation list.

Description

Technical Field

[0001] The present invention belongs to the technical field of book recommendation, and in particular to a book recommendation method and system based on knowledge graph and multi-objective optimization. Background Art

[0002] Books are a valuable asset of human wisdom and culture, carrying the accumulation of all kinds of knowledge, ideas and history. By reading books, we can acquire professional knowledge, broaden our horizons, enhance the depth of thinking, and cultivate critical thinking and innovation. Books not only record the process of human civilization and the spiritual world, but also are an important bridge for dialogue across time and space and different cultures. They inspire our creativity, enrich our emotional world, and provide a continuous source of power for personal growth and social progress. The power of books is infinite, and they play an irreplaceable role in education, science and technology, art and other fields. With the continuous development of Internet technology and the popularization of electronic products, people's attention has gradually changed from traditional paper book resources to electronic book resources. The "2022 China Digital Reading Report" shows that the scale of digital reading users in my country will reach 530 million in 2022, a year-on-year increase of 4.75%. Among them, the scale of digital reading users in China continued to expand during the period of 2017-2022, and the digital reading market has great development potential.

[0003] In the Internet era, the rapid growth and accumulation of data and information in the field of digital reading have brought about serious information overload problems. In this case, how to help users find the most suitable choice among many books has become an important topic. Recommendation algorithm is one of the most effective ways to solve information overload. The recommendation system analyzes the user's historical behavior, mines the user's interest preferences, and recommends the products of interest to the user. It is now widely used in online shopping, movies, music and other scenarios. Traditional recommendation systems are affected by problems such as cold start, sparsity, lack of recommendation accuracy and interpretability, and it is difficult to meet the diverse needs of users. Introducing knowledge graphs as auxiliary information into recommendation algorithms has become an increasingly popular research direction in the recommendation field, thereby alleviating the problems of cold start and sparse matrix in traditional recommendation algorithms, and thus improving the overall effect of the recommendation system.

[0004] Recommendations based on knowledge graphs are mainly divided into three categories: recommendation methods based on knowledge graph embedding, recommendation methods based on knowledge graph paths, and joint recommendation methods.

[0005] Through the above analysis, the problems and defects of the prior art are as follows:

[0006] (1) Recommendation method based on knowledge graph embedding: By embedding entities and relationships in the knowledge graph into a low-dimensional vector space, its latent semantic information is captured for recommendation. It has the advantages of high computational efficiency and ability to capture complex relationships, but lacks solvability and is highly dependent on high-quality embedding.

[0007] (2) Recommendation method based on knowledge graph path: recommending items by analyzing the connection path between users and items. The advantage is that the recommendation results are highly interpretable, but the disadvantage is that the path calculation is complex and has high requirements on the integrity of the knowledge graph.

[0008] (3) Joint-based recommendation method: Recommendations are made by combining path information and embedded representation. The principle is to use the interpretability of the path and the low-dimensional representation of the embedding to complement each other. The advantage is that it takes into account both recommendation effect and interpretability. The disadvantage is that in the process of information dissemination, it focuses on optimizing and updating the representation of the user side or the project side, and the degree of modeling of the user side and the project side is uneven. Summary of the invention

[0009] In view of the problems existing in the prior art, the present invention provides a book recommendation method based on knowledge graph and multi-objective optimization.

[0010] The present invention is implemented as follows: a book recommendation method based on knowledge graph and multi-objective optimization includes:

[0011] Step 1, initialization parameters;

[0012] Step 2: knowledge graph embedding task;

[0013] Step 3, recommend tasks;

[0014] Step 4: Joint optimization;

[0015] Step 5: Multi-objective optimization.

[0016] Further, the initialization parameters:

[0017] Initialize model parameters, including user embedding vectors, item embedding vectors, entity embeddings in the knowledge graph, and relationship embeddings; these parameters can be randomly initialized or obtained through pre-training.

[0018] Furthermore, the knowledge graph embedding task:

[0019] For a given knowledge triple (h, r, t), the head entity and the corresponding relation are taken as input, and the original feature vectors of the head entity h and the relation r are processed respectively using multiple cross-compression units and non-linear layers; then their latent features are concatenated together, and then an L-layer MLP is used to predict the tail t:

[0020] h L =Ev～S(h) [C L (v,h)[e]]

[0021] Among them C L is a cross compression unit, used to output the latent features of the L-layer header entity h, and S(h) is the relevant item of entity h;

[0022] r L =M L (r)

[0023]

[0024] The original vectors of h and r can be feature vectors, including descriptions of ID, type, or text; then their potential features are connected to predict the tail entity; finally, the score of the triple (h, r, t) is calculated by the similarity function, as follows:

[0025]

[0026] In the above formula, f KG The function is defined in the form of an inner product and the output is standardized by the nonlinear activation function sigmoid;

[0027]

[0028] The purpose of the cross compression unit is to achieve information sharing between the recommendation module and the knowledge graph embedding module;

[0029] For item v and its corresponding entity e in the knowledge graph, construct the cross feature C between them l ∈R d×d , where d is the dimension of the hidden layer;

[0030]

[0031] The output of item v and entity e after the cross compression unit is:

[0032]

[0033] in, is the cross compression unit weight, is the bias parameter. Through the cross compression unit, the model can adaptively adjust the weight of knowledge transfer and learn the correlation between the two tasks. For simplicity, the cross compression unit is expressed as:

[0034] [v l+1 , e l+1 ]=C(v l ,e l ).

[0035] Furthermore, the recommended tasks:

[0036] In the recommendation module, the ripple propagation network is used to replace the original user vector representation method, and the final user representation U and item representation V are used as the input of the module;

[0037] Given the interaction matrix Y and the knowledge graph G, the h-layer neighbor nodes of user u are defined as As follows:

[0038]

[0039] Where: h = 1, 2..., H is the number of propagation layers; A collection of historical items for user u;

[0040] Then As the seed set of the knowledge graph, the hierarchical interest set of user u is constructed along the connection path in the knowledge graph through a ripple-like propagation method. The interest set is defined as:

[0041]

[0042] The item vector h and the candidate entity vector v are concatenated and normalized by the Softmax function to obtain the weight p. The user's historical items are dynamically fused with the weight p to obtain the user's interest representation, which is defined as:

[0043]

[0044] After the above calculations, we get a multi-layer user interest representation set:

[0045]

[0046] Use reset gate r t Control the speed of forgetting information; use update gate z t The hidden state information of the previous moment is forgotten and selectively retained; the candidate hidden state is calculated based on the activation result of the update gate On this basis, the hidden state h at the current time t is calculated t , the calculation process is as follows:

[0047] r t =σ(W r x t +U r h t-1 )

[0048] z t =σ(W z x t +U z h t-1 )

[0049]

[0050] Where σ represents the Sigmoid activation function; W r and U r Represents the weight parameter of the reset gate; W z and U z Represents the weight parameter of the update gate; the symbol * represents the Hadamard product;

[0051] The entity responses of user u at different link lengths in the knowledge graph are input into GRU in the order from inner layer to outer layer. In order to extract the key features of the preference diffusion set, the attention mechanism is used to assign weights to the feature vectors extracted by GRU, and then the final vector representation u of the user is generated, as shown in the following formula:

[0052]

[0053] Where n represents the maximum link length of the preference diffusion set diffusion; is the hidden state vector representation of user u’s entity response o to the preference diffusion set in the GRU unit; θ jn is the weight corresponding to the current j-th hidden state vector, indicating the relative importance of the hidden state vector;

[0054] Finally, the user's final vector representation and the candidate item vector representation are combined and the predicted user click probability is calculated. The formula is as follows:

[0055]

[0056] Furthermore, the joint optimization:

[0057] In order to optimize the model performance, the present invention adopts the following loss function:

[0058]

[0059] Where: the first term F is the cross entropy loss; the second term is the knowledge graph embedding unit loss, λ1 is the loss value parameter; the third term is the L2 regularization loss, and λ2 is the regularization parameter.

[0060] Furthermore, the multi-objective optimization:

[0061] The Top-K candidate recommendation list is used as the initial population of the multi-objective optimization algorithm, and the Sobol sequence is used to initialize the population; secondly, the fuzzy dominance relationship is used: assuming x i and x j Represent any two individuals in the population, then use Bt (x i , x j ) represents x i X j The number of targets with good performance, W s (x i , x j ) represents x i X j The number of targets with poor performance, so the fuzzy dominance membership c is as follows, that is, the proportion of the number of better targets is used to express the degree of dominance between individuals, which is called fuzzy dominance membership c:

[0062]

[0063] When the fuzzy dominance membership c is greater than the set threshold, that is, the proportion of the number of better targets is within an acceptable range, the dominance relationship is established; otherwise, the dominance relationship is not established;

[0064] In addition, to avoid cyclic domination of individuals in the population, an energy parameter Q(x i ), as the overall standard for measuring the value of the individual objective function, is calculated as follows:

[0065]

[0066] f j (x i ) is individual x i The jth (j∈[1,m]) objective function value of z j is the minimum value of the jth (j∈[1,m]) objective function; when the true minimum value of the objective function is difficult to obtain, z j The value of takes the minimum value of the jth objective function corresponding to the current population;

[0067] Then the discoverer in the sparrow population is introduced into the t-distribution disturbance, and the improved discoverer position update formula is:

[0068]

[0069] in, represents the information position of the jth dimension when the i-th sparrow iterates to the t-th time; random number α∈(0,1]; iter maxDenote the maximum number of iterations; $R2 \in [0, 1]$, representing the warning value; $ST \in [0.5, 1]$, representing the safety value; $Q$ is a random number following a normal distribution; $L$ is a $1 \times d$ matrix where each element is 1. When $R2 \lt ST$, that is, there is no danger around, the population continues to forage. When $R2 \geq ST$, it indicates that there is danger in this area, and the discoverer needs to issue an alarm and lead other individuals in the population to move to a safe area; $t(iter)$ is a t-distribution whose degrees of freedom increase as the number of iterations increases. In the initial stage of algorithm iteration, the t-distribution is similar to the Cauchy distribution, which enables the algorithm to have strong global search ability. In the later stage of algorithm iteration, the t-distribution is similar to the Gaussian distribution, which helps the algorithm to have good local search ability, so as to enable the algorithm to maintain the diversity of the population in both local and global aspects;

[0070] Finally, calculate the crowding degree, update the sparrow population and the external archive set, and output the final recommendation list.

[0071] Another object of the present invention is to provide a book recommendation system based on knowledge graph and multi-objective optimization, including:

[0072] An initialization module for initializing parameters;

[0073] An embedding module for knowledge graph embedding tasks;

[0074] A recommendation module for recommendation tasks;

[0075] A joint optimization module for joint optimization;

[0076] A multi-objective optimization module for multi-objective optimization.

[0077] Another object of the present invention is to provide a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the book recommendation method based on knowledge graph and multi-objective optimization.

[0078] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the book recommendation method based on knowledge graph and multi-objective optimization.

[0079] Another object of the present invention is to provide an information data processing terminal for implementing the book recommendation system based on knowledge graph and multi-objective optimization.

[0080] Combined with the above technical solutions and the solved technical problems, the advantages and positive effects of the technical solution to be protected by the present invention are:

[0081] First, the present invention is a book recommendation method based on knowledge graph and multi-objective optimization. This method makes full use of the advantages of multi-task learning framework and ripple propagation network, and extracts features from both users and items at the same time. First, for the extraction of user features, the ripple propagation network is used to obtain the user's representation, and the user features at different levels are combined with the gated recurrent unit (GRU) combined with the attention mechanism to mine the user's potential dynamic interest preferences and generate the user's final representation. Secondly, for the extraction of item features, the cross compression unit is used to realize the information sharing between item information and head entities, and learn a richer item entity representation. The obtained user vector and entity vector are input into the prediction layer to predict the probability of interaction between the user and the entity item. Finally, the Top-K candidate recommendation list results are used as the initialization population of the multi-objective optimization algorithm, and the candidate recommendation list is optimized using the multi-objective optimization algorithm. The recommendation effect is comprehensively judged from the three aspects of recommendation accuracy, diversity and novelty, and then a better recommendation list is generated to improve user satisfaction with the recommendation results.

[0082] The purpose of the present invention is to further optimize the results of the improved knowledge graph-based recommendation algorithm using a multi-objective optimization algorithm to obtain the final recommendation result, so as to effectively solve the problems existing in book recommendations. The ripple propagation network is a recommendation system model based on the knowledge graph, which improves the accuracy and diversity of recommendations by simulating the "ripple effect" of user interests in the knowledge graph. It uses the structured information of the knowledge graph to explore and predict user potential interests through an iterative propagation mechanism to achieve personalized recommendations. The present invention addresses the problems of uneven modeling of traditional knowledge graph-based recommendation algorithms on the user side and the project side, as well as the problem of attenuation of entities in high-order semantic propagation. At the same time, the practice of using only the accuracy of the recommendation results as the only criterion for recommendation can no longer meet the needs of users. The invention uses a multi-task learning framework and a ripple propagation network to extract features on both the user side and the project side. First, for user feature extraction, the ripple propagation network is used to obtain the user's representation. When the ripple network propagation is attenuated in high-order semantic propagation, the user features at different levels are solved by using a gated recurrent unit combined with an attention mechanism to mine the user's potential dynamic interest preferences, thereby generating the user's final representation. For the extraction of item features, a cross-compression unit is used to achieve information sharing between item information and head entities, and learn richer item entity representations; finally, the Top-K candidate recommendation list results are used as the initialization population of the multi-objective optimization algorithm to further optimize the recommendation results and output the final recommendation list. Experiments have shown that the method proposed in the present invention solves the above-mentioned problems well, and at the same time comprehensively evaluates the recommendation effect from three aspects: accuracy, diversity, and novelty of the recommendation, thereby generating a better recommendation list and improving user satisfaction with the recommendation results.

[0083] Inspired by the ripple network and multi-objective optimization algorithm, this paper proposes a book recommendation algorithm based on knowledge graph and multi-objective optimization to solve the book recommendation problem. Through multi-task learning and ripple propagation network, user and project representations can be learned simultaneously, and then the gated recurrent unit (GRU) combined with the attention mechanism is used to solve the problem of high-order semantic propagation attenuation in the ripple propagation network. Finally, the Top-K candidate recommendation list is further optimized using a multi-objective optimization algorithm to obtain the final recommendation list. The present invention has the following advantages:

[0084] (1) The present invention utilizes a multi-task learning framework and a ripple propagation network to simultaneously learn user and item representations.

[0085] (2) In order to more accurately obtain the user's interest in items and the user's potential preferences, the present invention introduces a ripple propagation network to extract key information from the user's browsing history to explore the user's potential interests and preferences, and obtain the representation of the user entity. Then, a gated recurrent unit (GRU) combined with an attention mechanism is used to solve the attenuation of high-order semantic propagation, and then generate the final representation of the user.

[0086] (3) The present invention aims at the fact that most of the current book recommendation methods only care about the accuracy of the recommendation results. However, considering only the accuracy can no longer meet the personalized needs of learners, and it is not advisable to use only one indicator to measure the pros and cons of the recommendation algorithm. Therefore, a multi-objective optimization algorithm is introduced to further optimize the recommendation list to obtain the final recommendation list.

[0087] Second, as auxiliary evidence of the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0088] (1) The expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0089] Improve user satisfaction: By introducing a multi-task learning framework, ripple propagation network, and multi-objective optimization algorithm, the accuracy, diversity, and novelty of recommendation results have been significantly improved, which can better meet user needs.

[0090] Optimize recommendation performance: Use multi-objective optimization algorithms to further optimize recommendation results, comprehensively evaluate recommendation effects from multiple dimensions (accuracy, diversity, novelty), and provide users with better recommendation services.

[0091] Solve the problem of information overload: effectively alleviate the problem of information overload caused by the growth of digital reading resources and help users quickly find books of interest.

[0092] The commercial value of the technical solution of the present invention after transformation is reflected in the following aspects:

[0093] Adapting to market demand: As the scale of digital reading users continues to grow (for example, the scale of digital reading users in China will reach 530 million in 2022), this technology can provide a competitive advantage for e-book platforms and digital reading companies.

[0094] Improve platform stickiness: By optimizing recommendation effects, more users can be attracted to use the platform, and the user retention rate and platform usage time can be increased.

[0095] Expand application scenarios: The technology can not only be applied to book recommendations, but can also be extended to other recommendation scenarios such as movies, music, and e-commerce, and has broad commercial prospects.

[0096] (2) The technical solution of the present invention fills the technical gap in the industry at home and abroad:

[0097] The present invention introduces a ripple network and a multi-objective optimization algorithm to simultaneously model the user side and the project side, thereby solving the problem of insufficient user side modeling in traditional recommendation algorithms.

[0098] To address the attenuation problem of traditional knowledge graph-based recommendation algorithms in high-order semantic propagation, the present invention proposes a gated recurrent unit (GRU) combined with an attention mechanism, which greatly improves the accuracy of user interest feature extraction.

[0099] The present invention takes the accuracy, diversity and novelty of recommendations as comprehensive optimization goals, and achieves a comprehensive improvement in the recommendation effect through a multi-objective optimization algorithm.

[0100] (3) The technical solution of the present invention solves the technical problems that people have been eager to solve but have never been able to solve successfully:

[0101] High-order semantic propagation attenuation problem: During the embedding and propagation of knowledge graphs, the high-order semantic information of user interests gradually decays. This invention uses GRU combined with attention mechanism to alleviate this problem.

[0102] The problem of uniformity of recommendation results: Traditional algorithms only focus on recommendation accuracy, ignoring the optimization of diversity and novelty. The present invention alleviates this problem through a multi-objective optimization algorithm.

[0103] Uneven modeling of users and items: Traditional recommendation algorithms mainly model items on the item side, ignoring information sharing and feature extraction on the user side. This invention optimizes both user-side and item-side modeling through a ripple propagation network.

[0104] (4) The technical solution of the present invention overcomes technical prejudice:

[0105] Emphasis on a single goal: Traditional recommendation systems only optimize accuracy. The present invention breaks the single-goal bias through a multi-objective optimization algorithm, taking into account accuracy, diversity, and novelty.

[0106] Over-reliance on item-side features: Traditional algorithms tend to optimize item-side modeling and ignore user-side dynamic features. This paper overcomes this bias through the ripple propagation network and attention mechanism.

[0107] Ignoring the interpretability of recommendation results: Traditional algorithms focus on performance improvement but lack interpretability. This invention combines knowledge graph paths and embedded information to make recommendation results more interpretable. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] Figure 1 It is a flow chart of a book recommendation method based on knowledge graph and multi-objective optimization provided in an embodiment of the present invention.

[0109] Figure 2 It is a structural block diagram of a book recommendation system based on knowledge graph and multi-objective optimization provided by an embodiment of the present invention.

[0110] Figure 3 is a diagram of a cross compression unit provided by an embodiment of the present invention.

[0111] Figure 4 It is a gated cycle unit diagram provided by an embodiment of the present invention.

[0112] Figure 5 It is a flow chart of a book recommendation algorithm based on knowledge graph and multi-objective optimization provided in an embodiment of the present invention.

[0113] Figure 6 It is a diagram of the Top-k experimental results provided by an embodiment of the present invention.

[0114] Figure 7 It is a bar chart of the accuracy index of the multi-objective optimization provided by the embodiment of the present invention.

[0115] Figure 8 It is a bar chart of diversity indicators of multi-objective optimization provided by an embodiment of the present invention.

[0116] Fig. 9 It is a histogram of novelty index of multi-objective optimization provided by an embodiment of the present invention.

[0117] Fig.10 This is a main page display diagram of the book recommendation system provided by an embodiment of the present invention.

[0118] Fig.11 This is a diagram of a book details page of a book recommendation system provided by an embodiment of the present invention.

[0119] Fig.12 This is a recommendation page diagram of the book recommendation system provided by an embodiment of the present invention.

[0120] Fig.13 This is a search page diagram of a book recommendation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0121] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0122] The following are two specific industrial application examples of the book recommendation method based on knowledge graph and multi-objective optimization:

[0123] Example 1: Personalized teaching material recommendation on online education platform

[0124] Online education platforms need to recommend personalized learning materials to users (students or teachers), including e-books, course resources or workbooks.

[0125] 1. Knowledge graph construction: Build an educational knowledge graph, including entities and relationships such as subject knowledge points, textbook content, course relationships, and user learning behaviors. For example, the prerequisite relationship between knowledge points, the age group that textbooks are suitable for, and the user's learning history.

[0126] 2. Recommended implementation:

[0127] The system embeds various semantic information of learning behavior, course content and teaching materials based on students' learning progress and historical learning records, combined with knowledge graphs.

[0128] A multi-objective optimization method is used to balance the accuracy and diversity of recommendations, ensuring that the recommended teaching materials not only meet the user's learning needs but also have a certain breadth of knowledge coverage.

[0129] 3. Result output: Recommend teaching materials that match their current learning progress to students, or recommend teaching resources that suit their teaching needs to teachers, so as to improve teaching efficiency and user satisfaction.

[0130] Improved personalization of recommendations;

[0131] The knowledge graph enhances the adaptability of textbook recommendations;

[0132] Multi-objective optimization ensures the diversity and fairness of user experience.

[0133] Example 2: Interest-oriented recommendation on an electronic book sales platform

[0134] E-book sales platforms (such as the Kindle Store or Douban Reading) need to recommend books that match users' reading interests while taking into account the breadth of recommendations and sales growth goals.

[0135] 1. Knowledge graph construction: Build a knowledge graph covering book content, user interests and sales data. For example, the author, classification, content keywords, user evaluation relationship of the book, and user reading preference and behavior data.

[0136] 2. Recommended implementation:

[0137] Match user interests with book content through knowledge graph embedding.

[0138] In multi-objective optimization, the user ratings, user interest matching and sales contribution of books are comprehensively considered to ensure that the recommended books are both in line with user interests and beneficial to the platform's sales goals.

[0139] 3. Result output: Generate a set of recommended book lists based on the user's interests and historical purchasing behavior, covering both the books that the user commonly prefers and the books with high sales potential, thereby enhancing the commercial value of the recommendation.

[0140] Improve user purchase rate and reading experience;

[0141] Balance business goals and user interests in recommendations;

[0142] The knowledge graph enhances the relevance recommendations between books and promotes the discovery of long-tail books.

[0143] These two embodiments demonstrate the specific application scenarios of the method in the two major industries of online education and electronic book sales.

[0144] like Figure 1 As shown, a book recommendation method based on knowledge graph and multi-objective optimization provided by an embodiment of the present invention includes the following steps:

[0145] S101, initialization parameters;

[0146] S102, knowledge graph embedding task;

[0147] S103, recommended tasks;

[0148] S104, joint optimization;

[0149] S105: Multi-objective optimization.

[0150] The following is the detailed working principle of the book recommendation method based on knowledge graph and multi-objective optimization:

[0151] 1. Parameter initialization

[0152] In the initialization stage of the recommendation system, the model needs to set a series of parameters, including user embedding vectors, item embedding vectors, entity embeddings in the knowledge graph, and relationship embeddings.

[0153] The parameters can be initialized by random distribution or by using the results of a pre-trained model to ensure that the initial embedding vector has certain semantic information.

[0154] The initialization parameters lay the foundation for subsequent embedding training and recommendation optimization, enabling the model to effectively process multi-source information.

[0155] 2. Knowledge graph embedding task

[0156] The core of the knowledge graph embedding task is to map the entities and relations in the knowledge graph into a high-dimensional vector space.

[0157] Use embedding algorithms (such as TransE, DistMult, etc.) to represent entities and relations in the knowledge graph as vectors, thereby capturing the semantic and structural information in the knowledge graph.

[0158] The embedding results not only enhance the semantic understanding ability of book recommendations, but also provide rich contextual information for modeling the association between users and items.

[0159] 3. Recommended tasks

[0160] In the recommendation task, the recommendation score is calculated by the user embedding vector, item embedding vector and knowledge graph embedding vector.

[0161] Calculate the matching degree between users and books, and use the scoring function to predict the books that users are interested in.

[0162] In addition, the semantic relevance in the knowledge graph is combined to improve the diversity and accuracy of the recommendation results.

[0163] The goal of the recommendation task is to maximize the matching score between users and correct items while minimizing the scores of incorrect items.

[0164] 4. Joint optimization and multi-objective optimization

[0165] Joint optimization simultaneously trains the model parameters of the knowledge graph embedding task and the recommendation task, enabling the two to share information and improve the expressiveness of the embedding.

[0166] Joint optimization: By sharing the objective function (such as the loss function), the model can simultaneously improve the accuracy of knowledge graph embedding and recommendation performance.

[0167] Multi-objective optimization: During the optimization process, the system balances multiple objectives (such as recommendation accuracy, diversity, fairness, etc.) and uses weighted loss functions or Pareto optimal solutions to ensure that the system meets the requirements of multiple performance indicators.

[0168] This optimization method can improve the quality of recommendations while avoiding overfitting to a certain target and enhancing the adaptability and versatility of the system.

[0169] Through the above steps, the present invention makes full use of the semantic information of the knowledge graph and the multi-objective optimization technology to achieve accurate and efficient book recommendations.

[0170] Initialization parameters provided by the embodiment of the present invention:

[0171] Initialize model parameters, including user embedding vectors, item embedding vectors, entity embeddings in the knowledge graph, and relationship embeddings; these parameters can be randomly initialized or obtained through pre-training.

[0172] The knowledge graph embedding task provided by the embodiment of the present invention is:

[0173] For a given knowledge triple (h, r, t), the head entity and the corresponding relation are taken as input, and the original feature vectors of the head entity h and the relation r are processed respectively using multiple cross-compression units and non-linear layers; then their latent features are concatenated together, and then an L-layer MLP is used to predict the tail t:

[0174] h L =E v～S(h) [C L (v,h)[e]]

[0175] Among them C L is a cross compression unit, used to output the latent features of the L-layer header entity h, and S(h) is the relevant item of entity h;

[0176] r L =M L (r)

[0177]

[0178] The original vectors of h and r can be feature vectors, including descriptions of ID, type, or text; then their potential features are connected to predict the tail entity; finally, the score of the triple (h, r, t) is calculated by the similarity function, as follows:

[0179]

[0180] In the above formula, f KG The function is defined in the form of an inner product and the output is standardized by the nonlinear activation function sigmoid;

[0181]

[0182] The purpose of the cross unit is to achieve information sharing between the recommendation module and the knowledge graph embedding module;

[0183] For item v and its corresponding entity e in the knowledge graph, construct the cross feature C between them l ∈R d×d , where d is the dimension of the hidden layer;

[0184]

[0185] The output of item v and entity e after the cross compression unit is:

[0186]

[0187] in, is the cross compression unit weight, is the bias parameter. Through the cross compression unit, the model can adaptively adjust the weight of knowledge transfer and learn the correlation between the two tasks. For simplicity, the cross compression unit is expressed as:

[0188] [v l+1 , e l+1 ]=C(v l ,e l ).

[0189] The recommended tasks provided by the embodiment of the present invention are:

[0190] In the recommendation module, the ripple propagation network is used to replace the original user vector representation method, and the final user representation U and item representation V are used as the input of the module;

[0191] Given the interaction matrix Y and the knowledge graph G, the h-layer neighbor nodes of user u are defined as As follows:

[0192]

[0193] Where: h = 1, 2..., H is the number of propagation layers; A collection of historical items for user u;

[0194] Then As the seed set of the knowledge graph, the hierarchical interest set of user u is constructed along the connection path in the knowledge graph through a ripple-like propagation method. The interest set is defined as:

[0195]

[0196] The item vector h and the candidate entity vector v are concatenated and normalized by the Softmax function to obtain the weight p. The user's historical items are dynamically fused with the weight p to obtain the user's interest representation, which is defined as:

[0197]

[0198] After the above calculations, we get a multi-layer user interest representation set:

[0199]

[0200] Use reset gate r t Control the speed of forgetting information; use update gate z t The hidden state information of the previous moment is forgotten and selectively retained; the candidate hidden state is calculated based on the activation result of the update gate On this basis, the hidden state h at the current time t is calculated t , the calculation process is as follows:

[0201] r t =σ(W r x t +U r h t-1 )

[0202] z t =σ(W z x t +U z h t-1 )

[0203]

[0204] Where σ represents the Sigmoid activation function; W r and U r Represents the weight parameter of the reset gate; W z and U z Represents the weight parameter of the update gate; the symbol * represents the Hadamard product;

[0205] The entity responses of user u at different link lengths in the knowledge graph are input into GRU in the order from inner layer to outer layer. In order to extract the key features of the preference diffusion set, the attention mechanism is used to assign weights to the feature vectors extracted by GRU, and then the final vector representation u of the user is generated, as shown in the following formula:

[0206]

[0207] Where n represents the maximum link length of the preference diffusion set diffusion; is the hidden state vector representation of user u’s entity response o to the preference diffusion set in the GRU unit; θ jn is the weight corresponding to the current j-th hidden state vector, indicating the relative importance of the hidden state vector;

[0208] Finally, the user's final vector representation and the candidate item vector representation are combined and the predicted user click probability is calculated. The formula is as follows:

[0209]

[0210] The joint optimization provided by the embodiment of the present invention:

[0211] In order to optimize the model performance, the present invention adopts the following loss function:

[0212]

[0213] Where: the first term F is the cross entropy loss; the second term is the knowledge graph embedding unit loss, λ1 is the loss value parameter; the third term is the L2 regularization loss, and λ2 is the regularization parameter.

[0214] The multi-objective optimization provided by the embodiment of the present invention:

[0215] The Top-K candidate recommendation list is used as the initial population of the multi-objective optimization algorithm, and the Sobol sequence is used to initialize the population; secondly, the fuzzy dominance relationship is used: assuming x i and x j Represent any two individuals in the population, then use B t (x i , x j ) represents x i X j The number of targets with good performance, W s (x i , x j ) represents x i X j The number of targets with poor performance, so the fuzzy dominance membership c is as follows, that is, the proportion of the number of better targets is used to express the degree of dominance between individuals, which is called fuzzy dominance membership c:

[0216]

[0217] When the fuzzy dominance membership c is greater than the set threshold, that is, the proportion of the number of better targets is within an acceptable range, the dominance relationship is established; otherwise, the dominance relationship is not established;

[0218] In addition, to avoid cyclic domination of individuals in the population, an energy parameter Q(x i ), as the overall standard for measuring the value of the individual objective function, is calculated as follows:

[0219]

[0220] f j (x i ) is the j-th (j ∈ [1, m]) objective function value of individual x; z i is the minimum value of the j-th (j ∈ [1, m]) objective function; when it is difficult to obtain the true minimum value of the objective function, the value of z j takes the minimum value of the j-th objective function corresponding to the current population; j

[0221] Then, the discoverers in the sparrow population are introduced to the t-distribution perturbation, and the updated formula for the position of the improved discoverer is:

[0222]

[0223] where represents the information position of the j-th dimension when the i-th sparrow iterates to the t-th time; the random number α ∈ (0, 1]; iter max represents the maximum number of iterations; R2 ∈ [0, 1], representing the warning value; ST ∈ [0.5, 1], representing the safety value; Q is a random number obeying the normal distribution; L is a 1×d matrix with each element being 1. When R2 < ST, that is, there is no danger around, the population continues to forage. When R2 ≥ ST, it means that there is danger in this area, and the discoverer needs to issue an alarm and lead other individuals in the population to move to the safe area; t(iter) is the t-distribution that increases continuously with the increase of the degree of freedom as the number of iterations increases. In the initial stage of algorithm iteration, the t-distribution is similar to the Cauchy distribution, which enables the algorithm to have strong global search ability. In the later stage of algorithm iteration, the t-distribution is similar to the Gaussian distribution, which helps the algorithm to have better local search ability, so that the algorithm can maintain the diversity of the population in both local and global aspects;

[0224] Finally, calculate the crowding degree, update the sparrow population and the external archive set, and output the final recommendation list.

[0225] As Figure 2 shown, a book recommendation system based on knowledge graph and multi-objective optimization provided by an embodiment of the present invention includes:

[0226] An initialization module for initializing parameters;

[0227] An embedding module for knowledge graph embedding tasks;

[0228] A recommendation module for recommendation tasks;

[0229] A joint optimization module for joint optimization;

[0230] A multi-objective optimization module for multi-objective optimization.

[0231] The book recommendation system based on knowledge graph and multi-objective optimization provided by the embodiment of the present invention realizes accurate and personalized book recommendation by integrating knowledge graph technology and multi-objective optimization algorithm. The working principle of the system will be described in detail below.

[0232] First, the initialization module in the system is responsible for initializing various parameters and data structures to prepare for the subsequent recommendation process. Specifically, this module loads the basic information of users and books, and builds preliminary user portraits and book feature vectors. At the same time, the basic structure of the knowledge graph is initialized, including entities (such as authors, publishers, categories, etc.) and relationships (such as the relationship between authors and books, the relationship between books and categories, etc.). The initialization module ensures that the system has complete basic data and reasonable parameter configuration at startup, laying the foundation for the efficient operation of subsequent modules.

[0233] Next, the embedding module undertakes the task of embedding the knowledge graph and converts the entities and relations in the knowledge graph into low-dimensional vector representations. By using graph embedding techniques such as TransE, GraphSAGE or other advanced graph neural network models, the embedding module can capture the rich semantic information and structural features in the knowledge graph. These embedding vectors can not only reflect the multidimensional characteristics of books, but also reveal the potential relationship between users and books, thereby providing high-quality input data for the recommendation module.

[0234] In the recommendation module, the system makes personalized book recommendations based on the user's historical behavior, preferences, and the book vectors generated by the embedding module. The recommendation module uses collaborative filtering, content recommendation, or hybrid recommendation algorithms, combined with knowledge graph embedding vectors, to generate a preliminary recommendation list. By utilizing the semantic relationships in the knowledge graph, the recommendation module can better understand user needs, provide book recommendations that are highly relevant to user interests, and improve the accuracy of recommendations and user satisfaction.

[0235] The joint optimization module is responsible for integrating the outputs of the recommendation module and the embedding module, and improving the overall recommendation effect through joint optimization. Specifically, this module jointly trains the recommendation task with the knowledge graph embedding task, so that the embedding vector can not only accurately represent books and users, but also optimize the recommendation results. The joint optimization module shares parameters and a common loss function to ensure that the embedding vector effectively captures the semantic information of the knowledge graph while improving the recommendation performance, thus achieving synergy between the two.

[0236] The multi-objective optimization module further improves the performance of the recommendation system by optimizing multiple objectives at the same time to meet different recommendation needs. Common optimization objectives include recommendation accuracy, recommendation diversity, novelty, and user satisfaction. The multi-objective optimization module uses advanced optimization algorithms, such as genetic algorithms, particle swarm optimization, or weight adjustment methods, to balance the conflicting relationships between various objectives and generate the best recommendation solution that comprehensively considers multiple factors. This module ensures that the system can provide accurate recommendations while also meeting users' needs for diversity and novelty, improving the overall user experience.

[0237] Finally, the fusion module integrates the output results of each module to generate the final recommendation list. The fusion module adopts weighted average, sorting algorithm or other fusion strategies to comprehensively consider the recommendation results of each module to ensure that the final recommended books are in line with the user's interests and have diversity and novelty. At the same time, the system also includes computer equipment, computer-readable storage media and information data processing terminals, which together support the efficient operation and wide application of the system. The computer equipment is responsible for processing and storing data, the storage medium saves the programs that implement the functions of each module, and the information data processing terminal serves as an interface for user interaction, realizing the efficient operation of the book recommendation system based on knowledge graph and multi-objective optimization in actual application scenarios.

[0238] In summary, the present invention fully utilizes the multi-dimensional information of users and books through modular design and deep integration of knowledge graph and multi-objective optimization, and realizes efficient, accurate and personalized book recommendations. This system has significant advantages in practical applications, can significantly improve the reading experience and satisfaction of users, and has broad market application prospects and significant practical value. The purpose of the present invention is to further optimize the results of the improved recommendation algorithm based on knowledge graph using a multi-objective optimization algorithm to obtain the final recommendation results, so as to effectively solve the problems existing in book recommendations. The ripple propagation network is a recommendation system model based on knowledge graph, which improves the accuracy and diversity of recommendations by simulating the "ripple effect" of user interests in the knowledge graph. It uses the structured information of the knowledge graph to explore and predict the potential interests of users through an iterative propagation mechanism to achieve personalized recommendations. The present invention aims at the problem of uneven modeling of the traditional recommendation algorithm based on knowledge graph at the user end and the project end, and the problem of attenuation of entities in high-order semantic propagation. At the same time, the practice of using only the accuracy of the recommendation results as the only criterion for recommendation in the recommendation can no longer meet the needs of users. The invention uses a multi-task learning framework and a ripple propagation network to extract features on both the user side and the project side. First, for user feature extraction, the ripple propagation network is used to obtain the user's representation. When the ripple network propagation is attenuated at high-order semantic propagation, the user features at different levels are solved by using a gated recurrent unit combined with an attention mechanism to mine the user's potential dynamic interest preferences, thereby generating the user's final representation. For the extraction of item features, a cross-compression unit is used to achieve information sharing between user information and head entities, and learn richer item entity representations; finally, the Top-K candidate recommendation list results are used as the initialization population of the multi-objective optimization algorithm to further optimize the recommendation results and output the final recommendation list. Experiments have shown that the method proposed in the present invention solves the above-mentioned problems well, and at the same time comprehensively evaluates the recommendation effect from three aspects: accuracy, diversity, and novelty of the recommendation, thereby generating a better recommendation list and improving user satisfaction with the recommendation results.

[0239] The original combination of multi-task learning and knowledge graph recommendation algorithm (MKR) steps:

[0240] Step 1. Initialize parameters: First, randomly or pre-train the embedding vectors of users, items, and knowledge graphs;

[0241] Step 2. Recommendation task: Positive and negative samples are sampled through user-item interaction data, so that the item representation in the recommendation task and the head entity representation embedded in the knowledge graph can transfer information to each other, and then predict the matching score of the user and the item, and finally optimize the loss of the recommendation task.

[0242] Step 3. Knowledge graph embedding task: Sampling is performed through triple data, allowing the item representation in the recommendation task and the head entity representation of the knowledge graph embedding to communicate with each other, calculating the triple distance, and finally optimizing the knowledge graph embedding loss.

[0243] Step 4. Joint optimization: Combine the losses of the recommendation task and the knowledge graph task to achieve multi-task training.

[0244] When the algorithm combines knowledge graph and multi-objective optimization algorithm to solve the book recommendation problem, the algorithm does not model the user side and the project side equally. When the algorithm performs the recommendation task in step 2, it only considers modeling on the project side, allowing the project side to share information with the head entity of the indicator graph to complete the modeling of the project side, while the user side knowledge only uses MLP to extract user representations, so there is a problem of uneven modeling on both ends. When the algorithm proceeds to step 4, using the accuracy of the recommendation results as the only criterion for recommendation can no longer meet the needs of users, and the recommendation results need to be measured from multiple aspects.

[0245] To this end, the present invention is influenced by the ripple propagation network and designs a book recommendation method based on knowledge graph and multi-objective optimization algorithm. In this model, the advantages of the multi-task learning framework and the ripple propagation network are utilized to simultaneously extract features from both users and items. First, for the extraction of user features, the ripple propagation network is used to obtain the user's representation, and the user features at different levels are combined with the gated recurrent unit (GRU) combined with the attention mechanism to mine the user's potential dynamic interest preferences and generate the user's final representation. Secondly, for the extraction of item features, the cross compression unit is used to realize information sharing between user information and head entities, and learn richer item entity representations. Then, the obtained user vector and entity vector are input into the prediction layer to predict the probability of user interaction with the entity item. Finally, the recommended results are used as the initial population of the multi-objective optimization algorithm, and the recommendation results are further optimized to obtain the final results.

[0246] The improvement proposed in this invention combines multi-task learning and knowledge graph recommendation algorithm (MKUIP) with the following steps:

[0247] Input: interaction matrix Y, knowledge graph G

[0248] Output: Top-K final recommendation list

[0249] 1) Initialization parameters: First, randomly or pre-train the embedding vectors of users, items, and knowledge graphs;

[0250] 2) Recommendation task: Positive and negative samples are sampled through user-item interaction data, so that the item representation in the recommendation task and the head entity representation embedded in the knowledge graph can transfer information to each other, and then predict the matching score of the user and the item, and finally optimize the loss of the recommendation task.

[0251] 3) Knowledge graph embedding task: Sampling is performed through triple data, allowing the item representation in the recommendation task and the head entity representation of the knowledge graph embedding to communicate with each other, calculating the triple distance, and finally optimizing the knowledge graph embedding loss.

[0252] 4) Joint optimization: Combine the losses of the recommendation task and the knowledge graph task to achieve multi-task training.

[0253] 5) Multi-objective optimization: The multi-objective optimization algorithm is used to optimize the candidate recommendation list, and the recommendation effect is comprehensively evaluated from three aspects: accuracy, diversity, and novelty, to obtain the final recommendation list.

[0254] The following is a detailed description of the technical solution. First, the algorithm input is provided: user-item interaction matrix Y, knowledge graph G. The following is a detailed description of the steps of applying the improved combination of multi-task learning and knowledge graph recommendation algorithm to the book recommendation problem:

[0255] Step 1: Initialize parameters

[0256] Initialize model parameters, including user embedding vectors, item embedding vectors, entity embeddings in the knowledge graph, and relationship embeddings. These parameters can be initialized randomly or obtained through pre-training.

[0257] Step 2: Knowledge graph embedding task

[0258] Knowledge graph embedding is to embed entities and relations into a continuous vector space while preserving their structure. In this paper, we adopt a deep semantic matching architecture. For a given knowledge triple (h, r, t), the head entity and the corresponding relation are taken as input, and the original feature vectors of the head h and relation r are processed respectively using multiple cross compression units and non-linear layers. Then their latent features are concatenated together, and then an L-layer MLP is used to predict the tail t:

[0259] h L =E v～S(h) [C L (v,h)[e]]

[0260] Among them C L It is a cross compression unit used to output the latent features of the L-layer header entity h, and S(h) is the relevant item of entity h.

[0261] r L =M L(r)

[0262]

[0263] The original vectors of h and r can be feature vectors, including descriptions of ID, type, or text. Then their potential features are connected to predict the tail entity. Finally, the score of the triple (h, r, t) is calculated by the similarity function, as follows:

[0264]

[0265] In the above formula, f KG The function is defined in the form of an inner product and the output is normalized by a nonlinear activation function sigmoid.

[0266]

[0267] The purpose of the cross compression unit is to achieve information sharing between the recommendation module and the knowledge graph embedding module. Its structure is as follows Figure 3 shown.

[0268] For item v and its corresponding entity e in the knowledge graph, construct the cross feature C between them l ∈R d×d , where d is the dimension of the hidden layer.

[0269]

[0270] The output of item v and entity e after the cross compression unit is:

[0271]

[0272] in, is the cross compression unit weight, is the bias parameter. Through the cross compression unit, the model can adaptively adjust the weight of knowledge transfer and learn the correlation between the two tasks. For simplicity, the cross compression unit is expressed as:

[0273] [v l+1 , e l+1 ]=C(v l ,e l )

[0274] Step 3: Recommend tasks

[0275] In the recommendation module, the ripple propagation network is used to replace the original user vector representation method, and the final user representation U and item representation V are used as the input of this module.

[0276] Given the interaction matrix Y and the knowledge graph G, the h-layer neighbor nodes of user u are defined as As follows:

[0277]

[0278] Where: h = 1, 2..., H is the number of propagation layers; A collection of historical items for user u.

[0279] Then As the seed set of the knowledge graph, the hierarchical interest set of user u is constructed along the connection path in the knowledge graph through a ripple-like propagation method. The interest set is defined as:

[0280]

[0281] User historical behavior is important information for modeling user interests. Mining user historical behavior can improve the model's interest expression ability. Therefore, this paper first uses the attention mechanism to fuse user historical items, and uses the fused user feature vector as the user's shallow interest representation. The input of the attention mechanism is the user's historical items and candidate items. The item vector h and the candidate entity vector v are concatenated and normalized by the Softmax function to obtain the weight c. The user's historical items are dynamically fused with the weight c to obtain the user's interest representation, which is defined as:

[0282]

[0283] After the above calculations, we get a multi-layer user interest representation set:

[0284]

[0285] After obtaining the entity response of user u in the preference diffusion set, a preference feature sequence from the inner layer to the outer layer is formed. The gated recurrent unit structure is introduced for deep feature extraction and the attention mechanism is integrated for weight allocation, so as to capture the feature representation that is more important to the user in the user preference diffusion set and refine the granularity of the user interest description. The model framework of GRU is as follows Figure 4 shown.

[0286] GRU uses reset gate r t Control the degree of information forgetting; use update gate z t The hidden state information of the previous moment is forgotten and selectively retained; the candidate hidden state is calculated based on the activation result of the update gate On this basis, the hidden state h at the current time t is calculated t , the calculation process is as follows:

[0287] r t =σ(W r x t +Ur h t-1 )

[0288] z t =σ(W z x t +U z h t-1 )

[0289]

[0290] Among them, σ represents the Sigmoid activation function; W r and U r Represents the weight parameter of the reset gate; W z and U z Represents the weight parameter of the update gate; the symbol * represents the Hadamard product.

[0291] Considering that the user preference intensity decreases with the increase of link length in the preference diffusion set, the entities with farther distance are less important for the representation of user interest preferences. Therefore, the entity responses of user u at different link lengths in the knowledge graph are input into GRU in the order from inner layer to outer layer. In order to extract key features of the preference diffusion set, the attention mechanism is used to assign weights to the feature vectors extracted by GRU, and then generate the final vector representation u of the user, as shown in the following formula:

[0292]

[0293] Where n represents the maximum link length of the preference diffusion set diffusion; is the hidden state vector representation of user u’s entity response o to the preference diffusion set in the GRU unit; θ jn is the weight corresponding to the current j-th hidden state vector, indicating the relative importance of the hidden state vector.

[0294] Finally, the user's final vector representation and the candidate item vector representation are combined and the predicted user click probability is calculated. The formula is as follows:

[0295]

[0296] Step 4: Joint Optimization

[0297] In order to optimize the model performance, the present invention adopts the following loss function:

[0298]

[0299] Where: the first term F is the cross entropy loss; the second term is the knowledge graph embedding unit loss, λ1 is the loss value parameter; the third term is the L2 regularization loss, and λ2 is the regularization parameter.

[0300] Step 5: Multi-objective optimization

[0301] The Top-K candidate recommendation list is used as the initial population of the multi-objective optimization algorithm, and the Sobol sequence is used to initialize the population; secondly, the fuzzy dominance relationship is used: assuming x i and x j Represent any two individuals in the population, then use B t (x i , x j ) represents x i X j The number of targets with good performance, W s (x i , x j ) represents x i X j The number of targets with poor performance, so the fuzzy dominance membership c is as follows, that is, the proportion of the number of better targets is used to express the degree of dominance between individuals, which is called fuzzy dominance membership c:

[0302]

[0303] When the fuzzy dominance membership c is greater than the set threshold, that is, the proportion of the better target number is within an acceptable range, the dominance relationship is established; otherwise, the dominance relationship is not established.

[0304] In addition, to avoid cyclic domination of individuals in the population, an energy parameter Q(x i ), as the overall standard for measuring the value of the individual objective function, is calculated as follows:

[0305]

[0306] f j (x i ) is individual x i The jth (j∈[1,m]) objective function value of z j is the minimum value of the jth (j∈[1,m]) objective function; when the true minimum value of the objective function is difficult to obtain, z j The value of is the minimum value of the jth objective function corresponding to the current population.

[0307] Then the discoverer in the sparrow population is introduced into the t-distribution disturbance, and the improved discoverer position update formula is:

[0308]

[0309] in, It represents the information position of the j-th dimension when the i-th sparrow iterates to the t-th time; the random number α ∈ (0, 1]; iter max It represents the maximum number of iterations; R2 ∈ [0, 1], representing the warning value; ST ∈ [0.5, 1], representing the safety value; Q is a random number following a normal distribution; L is a 1×d matrix where each element is 1. When R2 < ST, that is, there is no danger around, the population continues to forage. When R2 ≥ ST, it indicates that there is danger in this area, and the discoverer needs to issue an alarm and lead other individuals in the population to move to the safe area; t(iter) is a t-distribution that increases continuously as the degree of freedom increases with the number of iterations. In the initial stage of algorithm iteration, the t-distribution is similar to the Cauchy distribution, which enables the algorithm to have strong global search ability; in the later stage of algorithm iteration, the t-distribution is similar to the Gaussian distribution, which helps the algorithm to have good local search ability, so that the algorithm can maintain the diversity of the population in both local and global aspects.

[0310] Finally, calculate the crowding degree, update the sparrow population and the external archive set, and output the final recommendation list.

[0311] The flow chart of the book recommendation algorithm based on knowledge graph and multi-objective optimization for recommendation is as follows Figure 5 shown:

[0312] The parameter settings of the present invention are shown in the following table, where d is the dimension of the knowledge graph vector representation, n is the sampling number of the interest set, h is the number of propagation layers, t is the training frequency of the model recommendation unit, λ1 and λ2 are hyperparameters, P is the population size, G max is the maximum number of evolutionary generations, M is the number of objective functions, K is the length of the final recommendation list, P c is the crossover probability, P m is the mutation probability.

[0313] Table 1 Experimental parameter settings

[0314] Table1 Model parameters

[0315]

[0316] To verify the practicality of MKUIP, in the Top-K experiment, accuracy (Precision@K) and recall (Recall@K) are used to measure the model recommendation effect. The experimental results of Top-K are as Figure 6 :

[0317] To verify the effectiveness of the multi-objective optimization algorithm, different metrics are used to evaluate the recommendation performance of the recommendation algorithm. In this section, the performance of the recommendation algorithm is measured from three dimensions, and the accuracy, diversity, and novelty of the recommendation are used as the evaluation metrics of the algorithm.

[0318] UserCF MF MORS MKUIP Accuracy 0.29 0.26 0.33 0.32 Diversity 0.49 0.45 0.47 0.54 Novelty 0.78 0.92 0.93 0.98

[0319] The present invention studies the problem of rating prediction in the recommendation system and proposes a book recommendation method MKUIP based on knowledge graph and multi-objective optimization. MKUIP first introduces GRU combined with attention mechanism in the interest propagation method to alleviate the user interest decay characteristics to generate user representation, thereby improving the user interest expression ability of the model; then, the project representation is generated through multi-layer cross compression operation, and multi-task learning is used to optimize the recommendation unit and knowledge graph embedding unit, thereby enhancing the generalization ability of the model; then, the project and user representation are used to recommend to the user; finally, the Top-K recommendation result is further optimized by the multi-objective optimization algorithm to obtain the final recommendation result. The book data set in the recommendation system field is selected for experiment. The results show that the MKUIP proposed in this paper is improved in diversity and novelty indicators compared with the other three reference methods. It can also be seen from the experimental results of Top-K that the improved recommendation algorithm of the present invention has obvious improvements in precision and recall compared with other methods, thereby verifying the effectiveness and feasibility of the present invention.

[0320] Book recommendation system main page: Fig.10 This is the main page display of the book recommendation system, which is mainly divided into popular books module, you may also like module, and recommended books module.

[0321] Book details page: In the three modules of popular books, recommended books and recommended books on the homepage, click on any book to jump to the book details page, where you can see the relevant information of the book and rate it. The operation results are shown in Fig.11 :

[0322] Book recommendation page: The book recommendation interface finds users with similar preferences based on the user's book rating data, and recommends books that these similar users like to the user to meet the user's preferences. The user-based collaborative filtering algorithm is mainly used here. The operation results are as follows: Fig.12 :

[0323] Book search page: Provide a search box on the homepage for retrieval. For example, search for Red Tent as shown below and the corresponding results will be given. Fig.13 :

[0324] In order to verify the effectiveness of the MKUIP algorithm recommendation, the experimental data set is the book knowledge graph data, the keyword data of 3682 books, 943 users and 100000 rating data. The test is conducted on this book data set, and the average accuracy, diversity and novelty of the recommendation results of all target users are calculated.

[0325] as follows Figure 7-Figure 9As shown in the figure, the accuracy, diversity, and novelty of different algorithms on the book dataset are compared. As can be seen from the figure, for this book dataset, in terms of the accuracy of the recommendation results, although the accuracy of MKUIP is slightly lower than that of MORS, it is better than UserCF and MF, among which MF's recommendation results perform the worst. In terms of the diversity and novelty of the recommendation results, MKUIP performs well and is better than other comparison algorithms. The diversity of the recommendation results generated by UesrCF is lower than that of MF, but MF is better than UesrCF in terms of novelty.

[0326] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0327] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A book recommendation method based on knowledge graph and multi-objective optimization, characterized in that: The following steps are involved: Step 1, parameter initialization: Initialize the user embedding vector, item embedding vector, entity embedding and relationship embedding in the knowledge graph, and the parameters are obtained by random initialization or pre-training; Step 2: Knowledge graph embedding task: Based on the entities and relations in the knowledge graph, the knowledge graph is embedded into a high-dimensional vector space using an embedding model to capture the semantic and structural information of the knowledge graph. Step 3, recommendation task: Based on the user embedding vector, item embedding vector and knowledge graph embedding vector, calculate the matching score between the user and the book, and predict the books that the user is interested in; Step 4: Joint optimization: Combine the objective functions of the knowledge graph embedding task and the recommendation task to optimize model parameters and improve embedding expression ability and recommendation accuracy. Step 5, multi-objective optimization: Based on multiple objectives such as recommendation accuracy, diversity, and fairness, use the weighted loss function or Pareto optimal solution method to achieve multi-objective optimization of the recommendation task.

2. The book recommendation method based on knowledge graph and multi-objective optimization as claimed in claim 1, characterized in that: The initialization parameters: Initialize model parameters, including user embedding vectors, item embedding vectors, entity embeddings in the knowledge graph, and relationship embeddings; these parameters can be randomly initialized or obtained through pre-training.

3. The book recommendation method based on knowledge graph and multi-objective optimization as claimed in claim 1, characterized in that: The knowledge graph embedding task: For a given knowledge triple (h, r, t), the head entity and the corresponding relation are taken as input, and the original feature vectors of the head entity h and the relation r are processed respectively using multiple cross-compression units and non-linear layers; then their latent features are concatenated together, and then an L-layer MLP is used to predict the tail t: h L =E v～S(h) [C L (v,h)[e]] Among them C L is a cross compression unit, used to output the latent features of the L-layer header entity h, and S(h) is the relevant item of entity h; r L ＝M L (r) The original vectors of h and r can be feature vectors, including descriptions of ID, type, or text; then their potential features are connected to predict the tail entity; finally, the score of the triple (h, r, t) is calculated by the similarity function, as follows: In the above formula, f KG The function is defined in the form of an inner product and the output is standardized by the nonlinear activation function sigmoid; The purpose of the cross unit is to achieve information sharing between the recommendation module and the knowledge graph embedding module; For item v and its corresponding entity e in the knowledge graph, construct the cross feature C between them l ∈R d×d , where d is the dimension of the hidden layer; The output of item v and entity e after the cross compression unit is: in, is the cross compression unit weight, b . l ∈R d is the bias parameter. Through the cross compression unit, the model can adaptively adjust the weight of knowledge transfer and learn the correlation between the two tasks. For simplicity, the cross compression unit is expressed as: [v l+1 ,e l+1 ]=C(v l ,e l )。 4. The book recommendation method based on knowledge graph and multi-objective optimization as claimed in claim 1, characterized in that: The recommended tasks: In the recommendation module, the ripple propagation network is used to replace the original user vector representation method, and the final user representation U and item representation V are used as the input of the module; Given the interaction matrix Y and the knowledge graph G, the h-layer neighbor nodes of user u are defined as As follows: Where: h = 1, 2..., H is the number of propagation layers; A collection of historical items for user u; Then As the seed set of the knowledge graph, the hierarchical interest set of user u is constructed along the connection path in the knowledge graph through a ripple-like propagation method. The interest set is defined as: The item vector h and the candidate entity vector v are concatenated and normalized by the Softmax function to obtain the weight p. The user's historical items are dynamically fused with the weight p to obtain the user's interest representation, which is defined as: After the above calculations, we get a multi-layer user interest representation set: Use reset gate r t Control the degree of information forgetting; use update gate z t Forget and selectively retain the hidden state information of the previous moment; The candidate hidden state is calculated based on the activation result of the update gate On this basis, the hidden state h at the current time t is calculated t , the calculation process is as follows: r t =σ(W r x t +U r h t-1 z t =σ(W z x t +U z h t-1 Where σ represents the Sigmoid activation function; W r and U r Represents the weight parameter of the reset gate; W z and U z Represents the weight parameter of the update gate; the symbol * represents the Hadamard product; The entity responses of user u at different link lengths in the knowledge graph are input into GRU in the order from inner layer to outer layer. In order to extract the key features of the preference diffusion set, the attention mechanism is used to assign weights to the feature vectors extracted by GRU, and then the final vector representation u of the user is generated, as shown in the following formula: Where n represents the maximum link length of the preference diffusion set diffusion; is the hidden state vector representation of user u’s entity response o to the preference diffusion set in the GRU unit; θ jn is the weight corresponding to the current j-th hidden state vector, indicating the relative importance of the hidden state vector; Finally, the user's final vector representation and the candidate item vector representation are combined and the predicted user click probability is calculated. The formula is as follows:

5. The book recommendation method based on knowledge graph and multi-objective optimization as claimed in claim 1, characterized in that: The joint optimization: In order to optimize the model performance, the present invention adopts the following loss function: Where: the first term F is the cross entropy loss; the second term is the knowledge graph embedding unit loss, λ1 is the loss value parameter; the third term is the L2 regularization loss, and λ2 is the regularization parameter.

6. The book recommendation method based on knowledge graph and multi-objective optimization as claimed in claim 1, characterized in that: The multi-objective optimization: The Top-K candidate recommendation list is used as the initial population of the multi-objective optimization algorithm, and the Sobol sequence is used to initialize the population; secondly, the fuzzy dominance relationship is used: assuming x i and x j Represent any two individuals in the population, then use B t (x i , x j ) represents x i X j The number of targets with good performance, W s (x i , x j ) represents x i X j The number of targets with poor performance, so the fuzzy dominance membership c is as follows, that is, the proportion of the number of better targets is used to express the degree of dominance between individuals, which is called fuzzy dominance membership c: When the fuzzy dominance membership c is greater than the set threshold, that is, the proportion of the number of better targets is within an acceptable range, the dominance relationship is established; otherwise, the dominance relationship is not established; In addition, to avoid cyclic domination of individuals in the population, an energy parameter Q(x i ), as the overall standard for measuring the value of the individual objective function, is calculated as follows: f j (x i ) is individual x i The jth (j∈[1,m]) objective function value of z j is the minimum value of the jth (j∈[1,m]) objective function; When the true minimum value of the objective function is difficult to obtain, z j The value of takes the minimum value of the jth objective function corresponding to the current population; Then the discoverer in the sparrow population is introduced into the t-distribution disturbance, and the improved discoverer position update formula is: Among them, represents the information position of the j-th dimension when the i-th sparrow iterates to the t-th time; the random number α ∈ (0, 1]; iter max represents the maximum number of iterations; R2 ∈ [0, 1], representing the warning value; ST ∈ [0.5, 1], representing the safety value; Q is a random number obeying the normal distribution; L is a 1×d matrix with each element being 1; when R2 < ST, that is, there is no danger around, the population continues to forage, and when R2 ≥ ST, it indicates that there is danger in this area, and the discoverer needs to issue an alarm and lead other individuals in the population to move to the safe area; t(iter) is a t-distribution that increases continuously with the increase of the degree of freedom as the number of iterations increases; in the initial stage of algorithm iteration, the t-distribution is similar to the Cauchy distribution, which enables the algorithm to have strong global search ability; in the later stage of algorithm iteration, the t-distribution is similar to the Gaussian distribution, which helps the algorithm to have good local search ability, so that the algorithm can maintain the diversity of the population in both local and global aspects; Finally, the crowding degree is calculated, the sparrow population and the external archive set are updated, and the final recommendation list is output.

7. A book recommendation system based on knowledge graph and multi-objective optimization that implements the book recommendation method based on knowledge graph and multi-objective optimization as described in any one of claims 1 to 6, characterized in that: The book recommendation system based on knowledge graph and multi-objective optimization includes: Initialization module, used to initialize parameters; Embedding module, used for knowledge graph embedding tasks; Recommendation module, used for recommendation tasks; Joint optimization module, used for joint optimization; Multi-objective optimization module, used for multi-objective optimization.

8. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the book recommendation method based on knowledge graph and multi-objective optimization as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the book recommendation method based on knowledge graph and multi-objective optimization as described in any one of claims 1 to 6.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the book recommendation system based on knowledge graph and multi-objective optimization as described in claim 7.

Images