Personalized article recommendation method and device, electronic equipment and storage medium

Abstract

The invention provides a personalized article recommendation method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence and data processing. According to the implementation scheme, context attention learning is carried out on a historical interaction behavior sequence of a target user for an article, an interest portrait of the target user is obtained, and the historical interaction behavior sequence comprises a plurality of article semantic vectors which are arranged according to a time sequence and interact with the target user; determining an information cocoon house strength coefficient of the target user on the basis of a semantic relationship and an article category distribution relationship among the article semantic vectors in the historical interaction behavior sequence; based on the interest portrait of the target user and the information cocoon house strength coefficient, generating an exploration strategy parameter vector of the target user; and inputting the exploration strategy parameter vector of the target user into the article recommendation model to obtain an article recommendation list of the target user output by the article recommendation model. Therefore, the recommendation accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of artificial intelligence and data processing. The present disclosure specifically relates to a personalized item recommendation method and device, an electronic device and a storage medium. BACKGROUND

[0002] In the information age of explosive data growth, recommendation systems, as a key technology for information filtering, effectively alleviate the modern problem of information overload. At the individual level, recommendation systems achieve precise matching of personalized information services by constructing high-dimensional vector representations of users and items, significantly reducing user search costs and improving information acquisition efficiency and subjective experience. At the platform and business level, recommendation systems create direct business value by improving user engagement, retention rate, and conversion rate; at the same time, through an efficient content distribution mechanism, they promote the circulation of long-tail goods and ensure the diversity and health of the content ecosystem. At the social macro level, recommendation systems, as the information infrastructure of the digital age, their algorithm design deeply affects the flow pattern of public information and the breadth and depth of cultural dissemination. Therefore, the study of recommendation systems not only involves technical efficiency, but also relates to the healthy development of individual cognition, market structure, and social information environment.

[0003] Early recommendation systems mainly relied on collaborative filtering and its variants, which achieved item recommendation by mining the similarity in the user-item interaction data matrix. Subsequently, deep neural networks (such as NeuMF), sequence models (such as GRU4Rec, SASRec), and attention mechanism-based models (such as BERT4Rec) were widely applied to recommendation systems, greatly improving the recommendation accuracy of recommendation systems. The core optimization goal of these models is to maximize the accuracy of predicting the next interaction (such as clicks or purchases) by learning from user historical behavior. However, the training of such models highly depends on the statistical bias in historical data, and their optimization process is essentially a fitting and amplification of existing user preferences, leading the system to recommend content highly similar to the user's historical behavior, thus forming an information cocoon. This phenomenon not only reduces the diversity of recommended content, but also limits the user's possibility of exploring new interests, causing user cognitive bias and narrow vision. In addition, this also affects the market ecosystem, limiting the exposure of long-tail content and inhibiting the development of new products and creators. SUMMARY

[0004] The present disclosure provides a personalized item recommendation method, device, electronic device and storage medium.

[0005] According to an aspect of the present disclosure, a personalized item recommendation method is provided, comprising: learn context attention on a historical interaction behavior sequence of a target user for an item, to obtain an interest portrait of the target user, wherein the historical interaction behavior sequence includes a plurality of item semantic vectors that are arranged in time sequence and interacted with the target user; determine an information cocoon strength coefficient of the target user based on semantic relationships and item category distribution relationships between the item semantic vectors in the historical interaction behavior sequence; generate an exploration strategy parameter vector of the target user based on the interest portrait and the information cocoon strength coefficient of the target user; input the exploration strategy parameter vector of the target user into an item recommendation model, to obtain an item recommendation list of the target user output by the item recommendation model.

[0006] According to another aspect of the present disclosure, there is provided an individualized item recommendation device, comprising: an interest portrait determination module configured to learn context attention on a historical interaction behavior sequence of a target user for an item, to obtain an interest portrait of the target user, wherein the historical interaction behavior sequence includes a plurality of item semantic vectors that are arranged in time sequence and interacted with the target user; an information cocoon determination module configured to determine an information cocoon strength coefficient of the target user based on semantic relationships and item category distribution relationships between the item semantic vectors in the historical interaction behavior sequence; a strategy vector determination module configured to generate an exploration strategy parameter vector of the target user based on the interest portrait and the information cocoon strength coefficient of the target user; a recommendation list determination module configured to input the exploration strategy parameter vector of the target user into an item recommendation model, to obtain an item recommendation list of the target user output by the item recommendation model.

[0007] According to another aspect of the present disclosure, there is provided an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform any one of the individualized item recommendation methods in the embodiments of the present disclosure.

[0008] According to another aspect of the present disclosure, there is provided a non-transitory computer readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to perform any one of the individualized item recommendation methods in the embodiments of the present disclosure.

[0009] According to the technology of the present disclosure, the historical interaction behavior sequence of a target user for an item is learned for context attention, obtaining an interest portrait of the target user, wherein the historical interaction behavior sequence includes a plurality of item semantic vectors that interact with the target user in chronological order; based on the semantic relationship and item category distribution relationship between each item semantic vector in the historical interaction behavior sequence, the information cocoon intensity coefficient of the target user is determined; based on the interest portrait of the target user and the information cocoon intensity coefficient, the exploration strategy parameter vector of the target user is generated; the exploration strategy parameter vector of the target user is input into the item recommendation model, and the item recommendation list of the target user output by the item recommendation model is obtained. In this way, the appropriate interest portrait and information cocoon intensity coefficient are extracted from the historical interaction behavior sequence of the user for the item, and the exploration strategy parameter vector of the target user is determined according to the two, so that the item recommendation model can consider the user interest and information cocoon intensity fed back by the exploration strategy parameter vector of the target user, and obtain an accurate item recommendation list. The item recommendation list can break the information cocoon and provide more accurate recommended items.

[0010] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the present disclosure, nor to limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings are used to better understand the present scheme and do not limit the present disclosure. Among them: Figure 1 is a flowchart of a personalized item recommendation method according to an embodiment of the present disclosure; Figure 2 is a flowchart of a data preprocessing process according to an embodiment of the present disclosure; Figure 3 is a schematic diagram of an information cocoon intensity calculation process according to an embodiment of the present disclosure; Figure 4 is a schematic diagram of a processing process of a strategy network according to an embodiment of the present disclosure; Figure 5 is a schematic diagram of an updating process and an application process of an item recommendation model according to an embodiment of the present disclosure; Figure 6 is a structural block diagram of a building structure modeling device according to an embodiment of the present disclosure; Figure 7 is a block diagram of an electronic device according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0012] Exemplary embodiments of the present disclosure are described herein with reference to the accompanying drawings, which are cited as illustrative examples. Various details of the embodiments of the present disclosure are described herein in order to provide a thorough understanding of the present disclosure. It will be understood by those of ordinary skill in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope of the present disclosure. Also, for the sake of brevity and clarity, descriptions of well-known functions and constructions are omitted herein.

[0013] Figure 1 is a flowchart of a personalized item recommendation method according to an embodiment disclosed herein.

[0014] As shown in Figure 1 , the personalized item recommendation method can include: S110, performing context attention learning on a historical interaction behavior sequence of a target user for an item to obtain an interest profile of the target user, wherein the historical interaction behavior sequence includes a plurality of item semantic vectors that are arranged in a time sequence and interacted with by the target user; S120, determining an information cocoon strength coefficient of the target user based on semantic relationships and item category distribution relationships between the item semantic vectors in the historical interaction behavior sequence; S130, generating an exploration strategy parameter vector of the target user based on the interest profile of the target user and the information cocoon strength coefficient; S140, inputting the exploration strategy parameter vector of the target user into an item recommendation model to obtain an item recommendation list of the target user output by the item recommendation model.

[0015] Exemplarily, as shown in Figure 2 , the generation process of the historical interaction behavior sequence of the target user for the item can be as follows: First, the interaction behavior log of the target user for the item is collected, for example, including the time length, frequency and time of operations such as clicking, purchasing, collecting and browsing the item. Metadata of the item, such as category, label and theme, and static information of the target user, such as user biological related attributes and device information, can also be used.

[0016] Then, the collected data undergoes preprocessing operations such as missing value handling, outlier detection and repair, and data deduplication. Missing value handling includes: for missing numerical features, the median is used to fill the missing value; for missing categorical features, the mode is used to fill the missing categorical feature; for missing important identifiers (such as user ID or item ID), the entire record is discarded. Outlier detection and repair involves: based on business rules, identifying obviously illogical abnormal data, treating it as a missing value, and repairing it according to the missing value handling strategy described above. Data deduplication includes: deduplicating data based on a unique key, such as a combination of user ID-item ID-timestamp, retaining only the last valid record.

[0017] Next, for the preprocessed interaction behavior logs, a fixed time window is set, such as the last 20 interactions of each user. All behavior records within this window are extracted from the sorted data and generated into an ordered interaction sequence according to their chronological order of occurrence. This sequence is stored in the form of an ordered list, where each element corresponds to a behavior time and contains key information such as item text content, item ID, behavior type, and timestamp.

[0018] Check if each item ID in the interaction sequence exists in the global item pool. IDs that do not exist in the global pool (such as delisted products and test data) are considered invalid IDs, and the corresponding meta-data is deleted from the sequence to ensure data consistency.

[0019] Then, the item content in each element of the interaction sequence is processed as follows: This study employs a keyword extraction algorithm combining TF-IDF and TextRank to automatically extract keywords representing the core theme of items. The SBERT model is then used to encode these keywords, mapping them to a low-dimensional dense vector space. This generates a fixed-dimensional (e.g., 384-dimensional) dense vector that accurately represents the semantic information of the item. This semantic information is then associated with key information including item ID, behavior type, and timestamp to obtain the item's semantic vector. This allows for the retrieval of the target user's historical interaction sequence with the item.

[0020] For example, the sequence of historical interaction behaviors is input into an encoder based on the Transformer architecture, and then passed through a multi-head attention network, a feedforward neural network and an attention pooling layer in sequence to obtain a global vector output by the attention pooling layer. This global vector is then used as the interest profile of the target user.

[0021] For example, such as Figure 3As shown, based on the semantic relationship between the semantic vectors of each item in the historical interaction behavior sequence and the item category distribution relationship, the homogeneity degree of item consumption, the decay degree of exploration behavior and the homogeneity degree of item feedback of the target user are determined, the homogeneity degree of item consumption, the decay degree of exploration behavior and the homogeneity degree of item feedback are weighted and summed to obtain the information cocoon intensity coefficient of the target user.

[0022] The homogeneity degree of item consumption is determined according to the average cosine similarity of the target user's recent consumption items in the semantic vector space. The higher the average cosine similarity, the higher the homogeneity degree of item consumption, indicating that the user's consumption content is more single.

[0023] The decay degree of exploration behavior is determined by counting the proportion of items outside the user's historical interest "comfort zone" (for example, the top three most frequently interacted item categories) in the user's recent interactions. The lower the value of the item proportion, the more serious the decay of the user's exploration behavior.

[0024] The homogeneity degree of item feedback is determined by analyzing the category distribution of the user's positive feedback items such as likes and collections, and calculating the Gini coefficient of the category distribution. The higher the Gini coefficient, the more concentrated the positive emotional feedback.

[0025] It should be noted that the homogeneity degree of item consumption, the decay degree of exploration behavior and the homogeneity degree of item feedback are normalized, for example, normalized to the interval [0, 1] to eliminate dimensional differences. Then, the normalized homogeneity degree of item consumption, the decay degree of exploration behavior and the homogeneity degree of item feedback are weighted and summed to obtain the information cocoon intensity coefficient of the target user.

[0026] Exemplarily, based on the interest profile of the target user and the information cocoon intensity coefficient, the exploration strategy parameter vector of the target user is generated, including: splicing the interest profile of the target user and the information cocoon intensity coefficient to obtain a comprehensive state representation vector; inputting the comprehensive state representation vector into the action and state based policy network to obtain the exploration strategy parameter vector output by the policy network.

[0027] Understandably, the policy network is a trained network, which takes the comprehensive state representation vector as the state, and the exploration strategy parameter vector as the action. The training sample composed of the reward value of the action and the next comprehensive state representation vector is used to train the policy network, and in the training process, the expectation value of the long-term cumulative reward is maximized as the training target, and the PPO algorithm is used to alternately optimize the network parameters of the action network and the state network in the policy network.

[0028] Exemplarily, the reward function of the long-term cumulative reward can be as follows: ; wherein, represents a long-term cumulative reward, respectively represent a user retention rate weight, a user activity weight, an interest exploration breadth weight and an information cocoon intensity coefficient weight, represents a user retention rate, represents a user activity, represents an interest exploration breadth, represents an information cocoon intensity coefficient.

[0029] Exemplarily, as shown in Figure 4 , the policy network includes an Actor network and a Critic network, both of which include a hidden layer, a fully connected layer and an activation layer. The comprehensive state representation vector is input into the deployed Actor network, and through forward propagation, an individualized, continuous value policy parameter vector is output. The policy parameter vector and the comprehensive state representation vector are input into the deployed Critic network, and through forward propagation, a scalar estimate of the value of the current state is output. Through the scalar estimate , a next comprehensive state representation vector is generated. In this way, multiple optimizations are performed, and finally a suitable policy parameter vector is output.

[0030] Exemplarily, the finally generated policy parameter vector is input into the item recommendation model, and the item recommendation model generates a suitable item recommendation list according to the policy parameter vector . The item recommendation list can include related information of one or more items. The item recommendation model can be a SASRec, NeuMF or the like.

[0031] According to the above embodiment, the historical interaction behavior sequence of the target user for the item is subjected to context attention learning to obtain an interest portrait of the target user, wherein the historical interaction behavior sequence includes a plurality of item semantic vectors that are arranged in time sequence and interacted with the target user; based on the semantic relationship and the item category distribution relationship between the item semantic vectors in the historical interaction behavior sequence, the information cocoon strength coefficient of the target user is determined; based on the interest portrait of the target user and the information cocoon strength coefficient, the exploration strategy parameter vector of the target user is generated; and the exploration strategy parameter vector of the target user is input into the item recommendation model to obtain the item recommendation list of the target user output by the item recommendation model. In this way, the appropriate interest portrait and information cocoon strength coefficient are extracted from the historical interaction behavior sequence of the user for the item, and the exploration strategy parameter vector of the target user is determined according to the two, so that the item recommendation model can consider the user interest and information cocoon strength fed back by the exploration strategy parameter vector of the target user to obtain an accurate item recommendation list. The item recommendation list can break the information cocoon and provide more accurate recommended items.

[0032] In an embodiment, based on the semantic relationship and the item category distribution relationship between the item semantic vectors in the historical interaction behavior sequence, the information cocoon strength coefficient of the target user is determined, including: based on the similarity between any two adjacent item semantic vectors in the historical interaction behavior sequence, the item consumption homogenization degree of the target user is determined; based on the appearance frequency of the item semantic vectors of each item category in the historical interaction behavior sequence, the exploration behavior attenuation degree of the target user is determined; the set of item semantic vectors with positive feedback semantics is determined from the historical interaction behavior sequence, and based on the proportion of each item category in the set of item semantic vectors, the item feedback homogenization degree of the target user is determined; and the item consumption homogenization degree, the exploration behavior attenuation degree and the item feedback homogenization degree are weighted and summed to obtain the information cocoon strength coefficient of the target user.

[0033] It can be understood that the set of item semantic vectors includes one or more item semantic vectors with positive feedback semantics. The positive feedback semantics can be identified by the target user's like, collection and other positive feedback information for the item semantic vector.

[0034] Exemplarily, the average, median or range of the similarity between any two adjacent item semantic vectors in the historical interaction behavior sequence can be taken as the item consumption homogenization degree.

[0035] Exemplarily, based on the appearance frequency of the item semantic vectors of each item category in the historical interaction behavior sequence, each item category is classified to obtain a familiar area and an exploration area, and the exploration behavior attenuation degree of the target user is determined according to the appearance frequency of the item semantic vectors in the exploration area in the historical interaction behavior sequence.

[0036] Exemplarily, based on the distribution of the proportion of each item category in the item semantic vector set, a Gini coefficient corresponding to the distribution is determined, and the Gini coefficient is taken as the homogenization degree of item feedback of the target user.

[0037] According to the above-mentioned embodiments, the homogenization degree of item consumption, the decay degree of exploration behavior and the homogenization degree of item feedback of the target user can be determined through the semantic relationship between each item semantic vector in the historical interaction behavior sequence and the distribution relationship of the item category, and then the information cocoon intensity coefficient of the target user can be obtained by weighted summation of the three.

[0038] In an embodiment, based on the similarity between any two adjacent item semantic vectors in the historical interaction behavior sequence, the homogenization degree of item consumption of the target user is determined, including: calculating the arithmetic mean of the similarity between any two adjacent item semantic vectors in the historical interaction behavior sequence; and determining the homogenization degree of item consumption based on the quantile of the arithmetic mean of the similarity in the historical arithmetic mean of the similarity distribution data.

[0039] Exemplarily, the quantile of the arithmetic mean of the similarity in the historical arithmetic mean of the similarity distribution data is taken as the homogenization degree of item consumption.

[0040] According to the above-mentioned embodiments, the homogenization degree of item consumption of the user can be accurately determined In an embodiment, based on the frequency of occurrence of the item semantic vector of each item category in the historical interaction behavior sequence, the decay degree of exploration behavior of the target user is determined, including: taking the item category with the highest frequency of occurrence in the historical interaction behavior sequence as the familiar area, and taking the item category in the historical interaction behavior sequence as the exploration area; and determining the decay degree of exploration behavior based on the ratio between the number of item semantic vectors in the exploration area and the number of item semantic vectors in the historical interaction behavior sequence.

[0041] Exemplarily, the ratio between the number of item semantic vectors in the exploration area and the number of item semantic vectors in the historical interaction behavior sequence is taken as the decay degree of exploration behavior.

[0042] According to the above-mentioned embodiments, the decay degree of exploration behavior of the user can be accurately determined.

[0043] In an embodiment, based on the proportion of each item category in the item semantic vector set, the homogenization degree of item feedback of the target user is determined, including: summing the absolute value of the difference between the proportions of any two item categories in the item semantic vector set, and determining the homogenization degree of item feedback of the target user based on the summation result.

[0044] For example, the ratio of the summation result to the total number of item categories is used as the degree of homogeneity of item feedback.

[0045] Based on the above implementation method, the degree of homogeneity of user item feedback can be accurately determined.

[0046] In one implementation, the method further includes: determining a base loss, a diversity regularization loss, a novelty regularization loss, and a fairness regularization loss based on the item recommendation list corresponding to the exploration strategy parameter vector; parsing the exploration strategy parameter vector to obtain diversity weights, novelty weights, and fairness weights; performing a weighted summation of the diversity regularization loss, novelty regularization loss, and fairness regularization loss based on the diversity weights, novelty weights, and fairness weights, and adding the weighted summation result to the base loss to obtain a first loss function; and updating the model parameters of the item recommendation model based on the first loss function.

[0047] For example, the policy parameter vector This is parsed into corresponding dynamic weight coefficients, such as diversity weights. Novelty weight And fairness weight .

[0048] For example, the first loss function The calculation formula is as follows: ; in, Indicates the basic loss. These represent the regularization loss for diversity, the regularization loss for novelty, and the regularization loss for fairness, respectively.

[0049] For example, basic loss This is the basic relevance used to ensure that the recommendation results are related to the user's interests, and its calculation formula can be as follows: ; in, Indicates user For positive sample items The predicted score Indicates user For negative sample items The predicted score for The function records the predicted score, or recommendation score, for each item in the item recommendation list.

[0050] For example, diversity regularization loss It is used to improve the distribution width of recommendation results. Its calculation formula is as follows: ; wherein, represents the probability distribution of the items in the item recommendation list belonging to the item category. The higher the value, the better the diversity of the item recommendation list.

[0051] Exemplarily, the novelty regularization loss is used to improve the novelty of the recommendation result. The calculation formula is as follows: ; wherein, is the global historical interaction times of the item in the recommendation list, represents the sum of the global historical interaction times of all items in the recommendation list. The higher the value, the better the novelty of the recommendation list.

[0052] Exemplarily, the calculation formula of the fairness regularization loss is as follows: ; wherein, represents the exposure of the item arranged at the tail of the recommendation list, .

[0053] Alternatively, the calculation formula of the fairness regularization loss may be as follows: ; wherein, wherein, represents the exposure of the item in the recommendation list, represents the exposure of the item in the recommendation list, represents the total number of items in the recommendation list.

[0054] In the present example, the Gini coefficient is used to evaluate the fairness regularization loss , which can measure the degree of imbalance of the exposure distribution between different items in the recommendation list. The closer the value is to 0, the more fair the distribution is.

[0055] Exemplarily, as shown in Figure 5 , the value of may be 3. On the basis of the above example, the self-game loss is introduced into the first loss function , and the first loss function is updated as follows: .

[0056] ​​Then, the item recommendation model, e.g., the SASRec model, is updated and trained by using the first loss function. Thus, in the application stage, the trained item recommendation model is used to generate the corresponding item recommendation list.

[0057] Exemplarily, the self-game loss The DPO algorithm can be used for evaluation, as follows:

[0058] wherein, represents an experience replay buffer, which is a dataset storing historical data, each piece of data in the dataset containing an context and its corresponding output positive sample (high reward recommendation list) and negative sample (low reward recommendation list), represents the data sampled from the experience replay buffer , the expectation is calculated, represents a temperature coefficient, represents the current strategy model to be optimized , when given the context , the probability of generating the positive sample , represents the probability of generating the positive sample by the reference strategy model under the same context , represents the probability of generating the negative sample by the current strategy model to be optimized when given the context , represents the probability of generating the negative sample by the reference strategy model when given the context , the reference strategy model is the strategy model parameter used by the item recommendation system (i.e., the item recommendation model described above) before the last large-scale iteration update, and its parameters are frozen and do not participate in gradient update.

[0059] After obtaining the above loss function, the model parameter of the item recommendation model is updated by using the gradient information in the loss function. The updated item recommendation model can be used for item recommendation to generate a suitable item recommendation list. In some examples, each item in the item recommendation list can also be rescored and sorted, and the reordered item recommendation list can be output.

[0060] According to the above-mentioned embodiments, the item recommendation model can be updated by using the loss function, and the accuracy of the item recommendation model can be improved.

[0061] Figure 6 FIG. 1 is a structural block diagram of an individualized item recommendation device according to an embodiment of the present disclosure.

[0062] As shown in Figure 6 the individualized item recommendation device comprises: An interest portrait determination module 610 is configured to perform context attention learning on a historical interaction behavior sequence of a target user with respect to an item, to obtain an interest portrait of the target user, wherein the historical interaction behavior sequence comprises a plurality of item semantic vectors that are arranged in time sequence and interacted with the target user. An information cocoon determination module 620 is configured to determine an information cocoon intensity coefficient of the target user based on semantic relationships and item category distribution relationships between the item semantic vectors in the historical interaction behavior sequence. A strategy vector determination module 630 is configured to generate an exploration strategy parameter vector of the target user based on the interest portrait and the information cocoon intensity coefficient of the target user. A recommendation list determination module 640 is configured to input the exploration strategy parameter vector of the target user into an item recommendation model, to obtain an item recommendation list of the target user output by the item recommendation model.

[0063] In an embodiment, the information cocoon determination module 620 comprises: A first degree determination unit is configured to determine an item consumption homogenization degree of the target user based on a similarity between any two adjacent item semantic vectors in the historical interaction behavior sequence. A second degree determination unit is configured to determine an exploration behavior attenuation degree of the target user based on a frequency of occurrence of item semantic vectors of each item category in the historical interaction behavior sequence. A third degree determination unit is configured to determine a set of item semantic vectors with positive feedback semantics from the historical interaction behavior sequence, and determine an item feedback homogenization degree of the target user based on a proportion of each item category in the set of item semantic vectors. A fourth degree determination unit is configured to perform weighted summation on the item consumption homogenization degree, the exploration behavior attenuation degree, and the item feedback homogenization degree, to obtain the information cocoon intensity coefficient of the target user.

[0064] In an embodiment, the first degree determination unit is specifically configured to: calculate an arithmetic mean of the similarity between any two adjacent item semantic vectors in the historical interaction behavior sequence. determine the item consumption homogeneity degree based on a quantile of the arithmetic mean of the similarity in the historical arithmetic mean of the similarity distribution data.

[0065] In an implementation, the second degree determining unit is specifically configured to: take the item category with the highest frequency of occurrence in the historical interaction behavior sequence as a familiar area, and take the item category in the historical interaction behavior sequence that is not in the familiar area as an exploration area; determine the exploration behavior decay degree based on a ratio between a number of item semantic vectors in the exploration area and a number of item semantic vectors in the historical interaction behavior sequence.

[0066] In an implementation, the third degree determining unit is specifically configured to: sum absolute values of differences between proportions of any two item categories in the set of item semantic vectors, and determine the item feedback homogeneity degree of the target user based on a sum result.

[0067] In an implementation, the strategy vector determining module 630 includes: a state vector determining unit configured to concatenate a target user interest portrait and an information cocoon intensity coefficient to obtain a comprehensive state representation vector; a strategy vector determining unit configured to input the comprehensive state representation vector into an action and state based policy network to obtain an exploration strategy parameter vector output by the policy network.

[0068] In an implementation, the apparatus described above further includes: a loss determining module configured to determine a basic loss, a diversity regularization loss, a novelty regularization loss, and a fairness regularization loss based on the item recommendation list corresponding to the exploration strategy parameter vector; a weight determining module configured to parse the exploration strategy parameter vector to obtain a diversity weight, a novelty weight, and a fairness weight; a loss function determining module configured to weight-sum the diversity regularization loss, the novelty regularization loss, and the fairness regularization loss based on the diversity weight, the novelty weight, and the fairness weight, and add a weighted sum result to the basic loss to obtain a first loss function; a parameter updating module configured to perform model parameter updating on the item recommendation model based on the first loss function.

[0069] The specific functions and examples of the modules and sub-modules of the apparatus of the embodiments of the present disclosure are described in the related description of the corresponding steps in the above method embodiments, which will not be described here.

[0070] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0071] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0072] Figure 7 This is a structural block diagram of an electronic device according to an embodiment of the present disclosure. Figure 7 As shown, the electronic device includes a memory 410 and a processor 420. The memory 410 stores a computer program that can run on the processor 420. There can be one or more memories 410 and processors 420. The memory 410 can store one or more computer programs, which, when executed by the electronic device, cause the electronic device to perform the methods provided in the above-described method embodiments. The electronic device may also include a communication interface 430 for communicating with external devices and performing data exchange and transmission.

[0073] If the memory 410, processor 420, and communication interface 430 are implemented independently, they can be interconnected via a bus to communicate with each other. This bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0074] Optionally, in a specific implementation, if the memory 410, processor 420 and communication interface 430 are integrated on a single chip, the memory 410, processor 420 and communication interface 430 can communicate with each other through an internal interface.

[0075] It is to be understood that the above-described processor can be a Central Processing Unit (CPU), but can also be other general purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic components, discrete hardware components, or the like. The general purpose processor can be a microprocessor or any conventional processor, or the like. It is to be appreciated that the processor can be an Advanced RISC Machines (ARM) architecture processor.

[0076] Further, the memory can include a read-only memory and a random access memory, and can further include a non-volatile random access memory. The memory can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memory. The non-volatile memory can include a Read-Only Memory (ROM), a Programmable ROM (PROM), an Erasable PROM (EPROM), an Electrically EPROM (EEPROM), or a flash memory. The volatile memory can include a Random Access Memory (RAM) used as an external cache. By way of example, but not limitation, many forms of RAM can be used. For example, a Static Random Access Memory (SRAM), a Dynamic Random Access Memory (DRAM), a Synchronous DRAM (SDRAM), a Double Data Rate SDRAM (DDR SDRAM), an Enhanced SDRAM (ESDRAM), a Synchlink DRAM (SLDRAM), and a Direct RAMBUS RAM (DR RAM) can be used.

[0077] In the above embodiments, all or part of the steps can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the steps can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present disclosure are generated. The computer can be a general purpose computer, a special purpose computer, a computer network or other programmable apparatus. The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired (for example: coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (for example: infrared, Bluetooth, microwave, etc.) mode. The computer readable storage medium can be any available medium accessible by a computer, or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be a magnetic medium (for example: floppy disk, hard disk, magnetic tape), an optical medium (for example: digital versatile disc (DVD)) or a semiconductor medium (for example: solid state disk (SSD)) etc. It is worth noting that the computer readable storage medium mentioned in the present disclosure can be a non-volatile storage medium, in other words, it can be a non-transitory storage medium.

[0078] A person of ordinary skill in the art can understand that all or part of the steps of the above embodiments can be completed by hardware, or the program can instruct the related hardware to complete, and the program can be stored in a computer readable storage medium. The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0079] In the description of the embodiments of the present disclosure, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples without contradiction.

[0080] In the description of the embodiments of the present disclosure, unless otherwise specified, " / " means or, for example, A / B can mean A or B. "And / or" in this document only describes the relationship between associated objects, which means that there can be three relationships, for example, A and / or B, which can mean: A exists alone, A and B exist together, and B exists alone.

[0081] In the description of the embodiments of the present disclosure, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present disclosure, unless otherwise specified, the meaning of "multiple" is two or more.

[0082] The above only describes exemplary embodiments of the present disclosure, and is not intended to limit the present disclosure. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A personalized item recommendation method characterized by, The method comprises: learning context attention of a historical interaction behavior sequence of a target user for an item, to obtain an interest portrait of the target user, wherein the historical interaction behavior sequence comprises a plurality of item semantic vectors that are arranged in time sequence and interacted with the target user; determining an information cocoon strength coefficient of the target user based on semantic relationships and item category distribution relationships between the item semantic vectors in the historical interaction behavior sequence; generating an exploration strategy parameter vector of the target user based on the interest portrait and the information cocoon strength coefficient of the target user; inputting the exploration strategy parameter vector of the target user into an item recommendation model to obtain an item recommendation list of the target user output by the item recommendation model.

2. The method of claim 1, wherein, The method comprises: determining an item consumption homogenization degree of the target user based on similarity between any two adjacent item semantic vectors in the historical interaction behavior sequence; determining an exploration behavior attenuation degree of the target user based on appearance frequencies of item semantic vectors of each item category in the historical interaction behavior sequence; determining an item feedback homogenization degree of the target user based on proportions of each item category in the item semantic vector set; weighting and summing the item consumption homogenization degree, the exploration behavior attenuation degree and the item feedback homogenization degree to obtain the information cocoon strength coefficient of the target user.

3. The method of claim 2, wherein, The method comprises: calculating an arithmetic mean of similarity between any two adjacent item semantic vectors in the historical interaction behavior sequence; determining the item consumption homogenization degree based on a quantile of the arithmetic mean of similarity in historical arithmetic mean of similarity distribution data.

4. The method of claim 2, wherein, The method comprises: taking an item category with the highest appearance frequency in the historical interaction behavior sequence as a familiar area, and taking item categories in the historical interaction behavior sequence that are not in the familiar area as exploration areas; determining the exploration behavior attenuation degree based on a ratio between a number of item semantic vectors in the exploration areas and a number of item semantic vectors in the historical interaction behavior sequence.

5. The method of claim 2, wherein, The method comprises: summing absolute values of differences between proportions of any two item categories in the item semantic vector set, and determining the item feedback homogenization degree of the target user based on a summing result.

6. The method according to any one of claims 1-5, characterized in that, The generating of the exploration strategy parameter vector of the target user based on the interest portrait of the target user and the information cocoon intensity coefficient comprises: The interest portrait of the target user and the information cocoon intensity coefficient are feature spliced to obtain a comprehensive state representation vector; The comprehensive state representation vector is input into a policy network based on actions and states to obtain an exploration strategy parameter vector output by the policy network.

7. The method of claim 6, wherein, Further comprising: Based on the item recommendation list corresponding to the exploration strategy parameter vector, determining a basic loss, a diversity regularization loss, a novelty regularization loss and a fairness regularization loss; Analyzing the exploration strategy parameter vector to obtain a diversity weight, a novelty weight and a fairness weight; Based on the diversity weight, the novelty weight and the fairness weight, the diversity regularization loss, the novelty regularization loss and the fairness regularization loss are weighted and summed, and the weighted sum result is added to the basic loss to obtain a first loss function; Based on the first loss function, the model parameter of the item recommendation model is updated.

8. An individualized item recommendation apparatus, characterized by, Comprise: An interest portrait determination module is configured to perform context attention learning on a historical interaction behavior sequence of a target user with respect to an item to obtain an interest portrait of the target user, wherein the historical interaction behavior sequence comprises a plurality of item semantic vectors that are interacted with the target user in chronological order; An information cocoon determination module is configured to determine an information cocoon intensity coefficient of the target user based on semantic relationships and item category distribution relationships between the item semantic vectors in the historical interaction behavior sequence; A strategy vector determination module is configured to generate an exploration strategy parameter vector of the target user based on the interest portrait of the target user and the information cocoon intensity coefficient; A recommendation list determination module is configured to input the exploration strategy parameter vector of the target user into an item recommendation model to obtain an item recommendation list of the target user output by the item recommendation model.

9. An electronic device, comprising: Comprise: At least one processor; And The memory is in communication connection with the at least one processor; wherein The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method of any one of claims 1-7.

10. A non-transitory computer-readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to enable the computer to execute the method according to any one of claims 1-7.

Images