Content recommendation method, device, equipment, storage medium and computer program product

Abstract

The present application discloses a content recommendation method, device, equipment, storage medium and computer program product, which relate to the field of machine learning. The method includes: obtaining account basic data and historical interaction data corresponding to a target account, the historical interaction data indicating a historical interaction sequence; obtaining target interaction data in the historical interaction sequence; fusing the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data and the target feature representation corresponding to the target interaction data to obtain a fusion feature; predicting the interest distribution of the content in the content library based on the fusion feature to obtain an interest distribution prediction result. That is, in the process of predicting the recommended content based on the target account and its corresponding interaction data, the feature representation corresponding to the target interaction data is added to the feature representation corresponding to the target account and the historical interaction data to perform feature fusion, and prediction is performed based on the obtained fusion feature, thereby improving the accuracy of content recommendation.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of machine learning, and in particular to a content recommendation method, apparatus, device, storage medium, and computer program product. Background Art

[0002] With the continuous development of science and technology, people's social methods have gradually changed from offline to online. "Conversation" refers to the interaction data between users and projects in chronological order, and the conversation-based recommendation system determines the user's interest preferences based on the corresponding historical conversation data between users and projects, and then recommends content.

[0003] In the related art, a session-based recommendation model is usually constructed, and sample data corresponding to anonymous users is input into the recommendation model to train its model parameters, so as to obtain a target recommendation model for recommending content to users.

[0004] However, in the above method, only one recommendation model is trained, so that when the model is actually used to recommend content to target users, the recommended content often ignores the interest preferences of the target users themselves, resulting in low accuracy of content recommendation and poor human-computer interaction efficiency. Summary of the invention

[0005] The embodiments of the present application provide a content recommendation method, apparatus, device, storage medium and computer program product, which can improve the accuracy of content recommendation. The technical solution is as follows.

[0006] On the one hand, a method for identifying entity words is provided, the method comprising:

[0007] Acquire account basic data and historical interaction data corresponding to the target account, wherein the account basic data is used to indicate basic account information of the target account, and the historical interaction data is used to indicate a historical interaction sequence constructed by interaction data of the target account within a historical time period;

[0008] Acquire target interaction data within a specified interaction interval in the historical interaction sequence, where the specified interaction interval is related to the interest time distribution of the target account;

[0009] fusing the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data, and the target feature representation corresponding to the target interaction data to obtain a fused feature;

[0010] Based on the fusion feature, interest distribution prediction is performed on the content in the content library to obtain an interest distribution prediction result, and the interest distribution prediction result is used to determine the content to be recommended to the target account.

[0011] In another aspect, a content recommendation device is provided, the device comprising:

[0012] An acquisition module, used to acquire account basic data and historical interaction data corresponding to a target account, wherein the account basic data is used to indicate basic account information of the target account, and the historical interaction data is used to indicate a historical interaction sequence constructed by interaction data of the target account within a historical time period;

[0013] The acquisition module is further used to acquire target interaction data within a specified interaction interval in the historical interaction sequence, where the specified interaction interval is related to the interest time distribution of the target account;

[0014] A fusion module, used to fuse the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data, and the target feature representation corresponding to the target interaction data to obtain a fusion feature;

[0015] The prediction module is used to predict the interest distribution of the content in the content library based on the fusion feature to obtain an interest distribution prediction result, and the interest distribution prediction result is used to determine the content to be recommended to the target account.

[0016] On the other hand, a computer device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement a content recommendation method as described in any of the above-mentioned embodiments of the present application.

[0017] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction, at least one program, a code set or an instruction set is stored in the storage medium, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by a processor to implement a content recommendation method as described in any of the above-mentioned embodiments of the present application.

[0018] On the other hand, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs any of the content recommendation methods described in the above embodiments.

[0019] The beneficial effects brought by the technical solution provided by the embodiment of the present application include at least:

[0020] In the process of fusing the account feature representation of the target account's corresponding account basic data and the interaction feature representation of the target account's corresponding historical interaction data, the target feature representation corresponding to the historical interaction sequence constructed by the target account's interaction data within the historical time period is added, the three are fused to determine the fusion feature, and the interest distribution of the content is predicted based on the fusion feature. Finally, the content recommended to the target account is determined. By adding the interaction data containing the interaction sequence for feature fusion, it is easier to capture the interest preference of the target account in the content recommendation process, thereby improving the accuracy of content recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 is a schematic diagram of a content recommendation method provided by an exemplary embodiment of the present application;

[0023] Figure 2 It is a schematic diagram of an implementation environment provided by an exemplary embodiment of the present application;

[0024] Figure 3 is a flow chart of a content recommendation method provided by an exemplary embodiment of the present application;

[0025] Figure 4 is a flow chart of a content recommendation method provided by another exemplary embodiment of the present application;

[0026] Figure 5 is a schematic diagram of splicing features provided by an exemplary embodiment of the present application;

[0027] Figure 6 is a global fusion schematic diagram provided by another exemplary embodiment of the present application;

[0028] Figure 7 is a schematic diagram of global fusion and local fusion provided by another exemplary embodiment of the present application;

[0029] Figure 8 is a flow chart of a content recommendation method provided by another exemplary embodiment of the present application;

[0030] Fig. 9 is a schematic diagram of a target knowledge graph provided by an exemplary embodiment of the present application;

[0031] Fig.10 is a schematic diagram of a content recommendation method provided by another exemplary embodiment of the present application;

[0032] Fig.11 is a structural block diagram of a content recommendation device provided by an exemplary embodiment of the present application;

[0033] Fig.12 is a structural block diagram of a content recommendation device provided by another exemplary embodiment of the present application;

[0034] Fig.13 It is a schematic diagram of the structure of a server provided by an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0035] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0036] First, a brief introduction is given to the terms involved in the embodiments of the present application.

[0037] Session: is the interaction behavior of the target user within a specified time period. Session-based recommendation is to recommend the next item clicked by the target user based on the current session. Session recommendation divides the historical interaction sequence of the target user into several sessions according to relevant logic, models the current session, and obtains the current interest preference of the target user, which is used to predict the next item of interest to the user.

[0038] Natural Language Processing (NLP): It is an important direction in the fields of computer science and artificial intelligence. It studies various theories and methods that can achieve effective communication between people and computers using natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field will involve natural language, that is, the language people use in daily life, so it is closely related to the study of linguistics. Natural language processing technology usually includes text processing, semantic understanding, machine translation, robot question answering, knowledge graph and other technologies.

[0039] Machine Learning (ML): It is a multi-disciplinary interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

[0040] Heterogeneous Graph Neural Networks (HGNN): The purpose of heterogeneous graph representation learning is to seek a meaningful vector representation for each node to facilitate subsequent applications, such as link prediction, personalized recommendation, node classification, etc. Therefore, a neural network with the same structure as the main node and a different structure from the main node is constructed to obtain different types of feature information. Among them, each node included in the heterogeneous graph neural network is called the "neighbor" of the main node, so the "neighbors" include homogeneous neighbors and heterogeneous neighbors, and the weight value assigned to each "neighbor" is different.

[0041] For illustration, please refer to Figure 1 , which shows a schematic diagram of a content recommendation method provided by an exemplary embodiment of the present application, such as Figure 1 As shown, the account basic data 101 and the historical interaction data 102 corresponding to the target account 100 are obtained, wherein the historical interaction data 102 includes a historical interaction sequence constructed by interaction data within a historical time period (such as: interaction data 1, interaction data 2, ..., interaction data n, wherein interaction data 1 to interaction data n are interaction data arranged in sequence), the target interaction data 103 within a specified interaction interval in the historical interaction data 102 is obtained, the account feature representation 104 corresponding to the account basic data 101, the interaction feature representation 105 corresponding to the historical interaction data 102, and the target feature representation 106 corresponding to the target interaction data 103 are determined, the account feature representation 104, the interaction feature representation 105, and the target feature representation 106 are fused to obtain a fused feature, and interest distribution prediction is performed on the content in the content library based on the fused feature, wherein the content library includes at least one content to be predicted (such as: content 1, content 2, ..., content n), the interest distribution prediction result corresponding to each content in the content library is obtained, and the target content 107 finally recommended to the target account is determined according to the interest distribution prediction result.

[0042] An embodiment of the present application provides a content recommendation method. In the process of fusing the account feature representation of the target account's corresponding account basic data and the interaction feature representation of the target account's corresponding historical interaction data, the target feature representation corresponding to the historical interaction sequence constructed by the interaction data of the target account in the historical time period is added. The three are fused to determine the fusion feature, and the interest distribution of the content is predicted based on the fusion feature. Finally, the content recommended to the target account is determined. By adding the interaction data containing the interaction sequence for feature fusion, it is easier to capture the interest preference of the target account in the content recommendation process, thereby improving the accuracy of content recommendation.

[0043] In combination with the above-mentioned noun introduction and method introduction, the application scenarios of the embodiments of the present application are illustrated by examples.

[0044] 1. Applied to shopping recommendation scenarios. Obtain the account basic data corresponding to the user's shopping account and the historical shopping data corresponding to the shopping account, wherein the account basic data includes the user's corresponding shopping preferences, and the historical shopping data includes historical shopping records arranged in order of purchase time, obtain the historical shopping record closest to the current time point in the historical shopping records as the target interaction data, extract the feature representations corresponding to the account basic data, the target interaction data and the historical shopping data respectively, perform feature fusion on the three feature representations, and infer the product with the highest purchase expectation of the current user based on the fusion result and recommend it to the user;

[0045] 2. Applied to video recommendation scenarios. The basic account data corresponding to the user's audience account and the historical viewing data corresponding to the audience account are obtained, wherein the basic account data includes the user's corresponding video viewing preferences, and the historical viewing data includes historical viewing records corresponding to the historical viewing videos arranged in order of viewing time. A historical viewing video closest to the current time point in the historical viewing videos is obtained as the target interaction data, and feature representations corresponding to the basic account data, the target interaction data, and the historical viewing records are extracted respectively. The three feature representations are feature fused, and based on the fusion result, the video with the highest probability of being watched by the user in the current time period is inferred and recommended to the user.

[0046] It is worth noting that the content recommendation method provided in the embodiment of the present application can be implemented by a terminal, can be implemented by a server, or can be implemented by a terminal and a server in collaboration.

[0047] Among them, when the terminal and the server collaborate to implement the solution provided by the embodiment of the present application, the terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited by the embodiment of the present application.

[0048] For illustration, please refer to Figure 2 , which shows a schematic diagram of an implementation environment provided by an exemplary embodiment of the present application, such as Figure 2 As shown, the implementation environment includes a terminal 210 and a server 220 , and the terminal 210 and the server 220 are connected via a communication network 230 .

[0049] In the embodiment of the present application, an application with a content recommendation function is installed in the terminal 210. The application providing the content recommendation function can be implemented as a browser, a video player, an e-book reading program, an online shopping application, an instant messaging application, etc., which is not limited in the embodiment of the present application. After the user performs at least one trigger operation on the historical browsing content, the terminal 210 sends at least one historical browsing content triggered by the trigger operation as historical interaction data to the server 220.

[0050] After receiving at least one historical browsing content corresponding to the user sent by the terminal, the server 220 obtains the account basic data including the basic information of the user account, and constructs a historical interaction sequence according to the historical browsing content, determines the target interaction data within the specified interaction interval in the historical interaction sequence, that is, the target browsing content, obtains the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data, and the target feature representation corresponding to the target interaction data, and fuses the three feature representations to obtain the fused feature 221.

[0051] In addition, the server 220 also includes a content library 222, which includes at least one candidate content to be recommended. The fusion feature 221 is input into the content library 222, and the interest distribution of the candidate content included in the content library 222 is predicted, and the user's interest distribution prediction result corresponding to the candidate content is determined. Based on the interest distribution prediction result corresponding to at least one candidate content, the target recommended content 223 to be recommended to the user is determined.

[0052] The server 220 feeds back the target recommended content to the terminal 210, and the terminal 210 displays the target recommended content, wherein the display method includes at least one of voice display, text display or interface display.

[0053] Among them, the terminal 210 includes at least one of a smart phone, a tablet computer, a portable laptop computer, a desktop computer, a smart speaker, a smart wearable device, a smart voice interaction device, a smart home appliance, a car terminal and the like.

[0054] It is worth noting that the above-mentioned communication network 230 can be implemented as a wired network or a wireless network, and the communication network 230 can be implemented as any one of a local area network, a metropolitan area network or a wide area network, which is not limited in the embodiments of the present application.

[0055] It is worth noting that the above-mentioned server 220 can be implemented as a cloud server in the cloud, wherein cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and network in a wide area network or a local area network to realize data calculation, storage, processing, and sharing. Cloud technology is a general term for network technology, information technology, integration technology, management platform technology, application technology, etc. based on the cloud computing business model application, which can form a resource pool, which is used on demand and is flexible and convenient. Cloud computing technology will become an important support. The background services of the technical network system require a large amount of computing and storage resources, such as video websites, picture websites, and more portal websites. With the high development and application of the Internet industry, in the future, each item may have its own identification mark, which needs to be transmitted to the background system for logical processing. Data of different levels will be processed separately, and all kinds of industry data require strong system backing support, which can only be achieved through cloud computing.

[0056] In some embodiments, the server 220 may also be implemented as a node in a blockchain system.

[0057] In combination with the above-mentioned noun introduction and application scenarios, the content recommendation method provided in the embodiment of the present application is described, and the method is performed by the server as an example for description. For schematic illustration, please refer to Figure 3 , which shows a flowchart of a content recommendation method provided by an exemplary embodiment of the present application, the method includes the following steps.

[0058] Step 301: Obtain basic account data and historical interaction data corresponding to the target account.

[0059] The account basic data is used to indicate the basic account information of the target account, and the historical interaction data is used to indicate the historical interaction sequence constructed by the interaction data of the target account in the historical time period.

[0060] Illustratively, the basic account information includes the account profile information corresponding to the target account, such as: the user gender, age, hobbies, etc. corresponding to the target account; or, the basic account information also includes the account social information corresponding to the target account, such as: if there is an associated account that has a social relationship (such as a friend relationship) with the target account, then the account social information is used to indicate the social relationship between the target account and the associated account; or, the basic account information includes the account preference information corresponding to the target account, and the account preference information indicates that within a specific period of time, based on the historical operation records corresponding to the target account, the interest preference corresponding to the target account is determined, such as: when the target account is the corresponding account in a video application, the video type corresponding to the video content browsed by the target account within a specified historical time period is used as the account preference information corresponding to the target account.

[0061] Among them, the account basic data will be updated after the target account updates the account basic information.

[0062] Illustratively, historical interaction data represents interaction data generated between a target account and historical content or historical items within a historical time period. For example, taking the target account as a video application, when the target account generates a trigger operation on a certain video content, or browses the video content, the triggering behavior / browsing behavior corresponds to the interaction data generated between the target account and the video content.

[0063] In some embodiments, the historical interaction sequence represents a series of interaction data corresponding to the interaction behaviors (interaction behaviors such as click operations and browse operations in video scenarios and purchase operations in shopping scenarios, which are not limited here) generated by the target account within the historical time period, wherein the series of interaction data is arranged in order, and an example of a shopping scenario is used for illustration:

[0064] In a shopping scenario, the target account is a shopping account corresponding to a shopping application. In a historical time period, the target account purchases at least one item. The historical purchase record generated by the purchase behavior is the historical interaction data. The sequence corresponding to the historical purchase records arranged in order within the historical time period corresponding to the target account is the historical interaction series.

[0065] Optionally, arranging in order includes: arranging in chronological order of associated behaviors, or arranging according to the number of repetitions corresponding to interactive behaviors (such as arranging from high to low according to the number of historical repeated browsing times of users in a video scenario), which is not limited here.

[0066] It is understandable that in the specific embodiments of the present application, related data such as account basic data and historical interaction data are involved. When the above embodiments of the present application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0067] Step 302: Acquire target interaction data within a specified interaction interval in a historical interaction sequence.

[0068] The specified interaction interval is related to the interest time distribution of the target account.

[0069] Optionally, when the historical interaction sequence is constructed in chronological order, the specified interaction interval includes a specified time period or a specified time point, wherein the specified time period or the specified time point is the interest time distribution of the historical interaction data corresponding to the target account, and the target interaction data is one or more historical interaction data within the specified time period, or one or more historical interaction data corresponding to the specified time point, such as: from the historical interaction data arranged in chronological order, one or more historical interaction data closest to the current time point are obtained as the target interaction data; when the historical interaction sequence is constructed in the order of interactive behavior, the specified interaction interval includes setting an interaction threshold, that is, the interaction threshold is the interest time distribution corresponding to the target account, and when the interactive behavior between the target account and the historical interaction data reaches the interaction threshold, the historical interaction data is used as the target interaction data.

[0070] Step 303: fuse the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data, and the target feature representation corresponding to the target interaction data to obtain a fused feature.

[0071] Illustratively, the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data, and the target feature representation corresponding to the target interaction data are extracted, wherein, when the target interaction data is one interaction data, the target feature representation corresponds to the target interaction data, and when the target interaction data contains multiple interaction data, the target feature representation is a common feature representation corresponding to multiple target interaction data.

[0072] Optionally, the feature representation methods corresponding to the account feature representation, the interaction feature representation, and the target feature representation include representation in vector form, or representation in set form, or representation in array form, which is not limited here.

[0073] Optionally, the fusion method includes at least one of the following methods:

[0074] 1. The account feature representation, interaction feature representation, and target feature representation are sequentially concatenated to obtain a concatenated result, which is used as a fusion feature;

[0075] 2. Establish a feature fusion network, input the account feature representation, interaction feature representation and target feature representation into the feature fusion network at the same time, and obtain an output result as the corresponding fusion feature of the three. For example, establish a convolutional neural network, perform convolution, pooling, connection and other operations on the input feature representation, and obtain a new feature representation as the fusion feature.

[0076] It is worth noting that the above description of the fusion method is only an illustrative example and the embodiments of the present application do not limit this.

[0077] Step 304: predict the interest distribution of the content in the content library based on the fusion features to obtain an interest distribution prediction result.

[0078] Among them, the interest distribution prediction result is used to determine the content to be recommended to the target account.

[0079] Illustratively, interest distribution prediction includes calculating the interest probability of the target account corresponding to each content in the content library, that is, predicting the interest probability value of the target account corresponding to each content in the content library. The larger the interest probability value, the greater the probability that the target account will interact with it.

[0080] In some embodiments, the content library is a pre-established library, and the content contained in the content library is the recommended content to be predicted, including content that has interacted with other accounts, or content that has not interacted with any account; the content of the content library includes content of the same type or different types, which is not limited here.

[0081] Optionally, the method for predicting interest distribution includes at least one of the following prediction methods:

[0082] 1. Obtain the content features corresponding to each content in the content library, compare the similarity of each content feature with the fusion feature, obtain a similarity score, and use the similarity score as the prediction probability corresponding to the content. Finally, the similarity scores corresponding to the contents of the content library are used as the interest distribution prediction results;

[0083] 2. Obtain the fusion vector representation corresponding to the fusion feature, and obtain the content features corresponding to the content in the content library, where the content features are represented in the form of feature vectors. Calculate the distance between the fusion vector and the feature vector, and obtain the distance result as the prediction probability corresponding to the content. Finally, each content in the content library corresponds to a distance result as the interest distribution prediction result.

[0084] It is worth noting that the above description of the interest distribution prediction method is only an illustrative example and the embodiments of the present application are not limited to this.

[0085] After determining the predicted probability corresponding to each content in the content library, determine the content with the highest predicted probability or a specified number of contents in the content library as recommended content, or set a probability threshold, and use the content in the content library with the predicted probability reaching the probability threshold as recommended content, and recommend the recommended content to the target account.

[0086] In summary, an embodiment of the present application provides a content recommendation method. In the process of fusing the account feature representation of the target account's corresponding account basic data and the interaction feature representation of the target account's corresponding historical interaction data, the target feature representation corresponding to the historical interaction sequence constructed by the interaction data of the target account within the historical time period is added. The three are fused to determine the fusion feature, and the interest distribution of the content is predicted based on the fusion feature. Finally, the content recommended to the target account is determined. By adding the interaction data containing the interaction sequence for feature fusion, it is easier to capture the interest preference of the target account in the content recommendation process, thereby improving the accuracy of content recommendation.

[0087] In an optional embodiment, the fusion method includes concatenating the account feature representation, the interaction feature representation, and the target feature representation. Figure 4 , which shows a flow chart of a content recommendation method provided by an exemplary embodiment of the present application, such as Figure 4 As shown, the method comprises the following steps:

[0088] Step 401, obtaining basic account data and historical interaction data corresponding to the target account.

[0089] The account basic data is used to indicate the basic account information of the target account, and the historical interaction data is used to indicate the historical interaction sequence constructed by the interaction data of the target account in the historical time period.

[0090] The contents of the account basic data and historical interaction data in step 401 have been described in detail in the above step 301 and will not be repeated here.

[0091] Step 402: Obtain the latest n interaction data in the historical interaction sequence as target interaction data, where n is a positive integer.

[0092] In some embodiments, the most recent interaction data in the historical interaction sequence is obtained as the target interaction data.

[0093] Schematically, the historical interaction sequence includes multiple historical interaction data arranged in chronological order, wherein, taking the current time point as the anchor point, the historical interaction data at the end of the sequence in the historical interaction sequence is the interaction data closest to the current time point, and is used as the target interaction data.

[0094] Among them, the order of the historical interaction sequence indicates the content interest direction of the target account in different time periods. For example, taking video content as an example, when the first three historical browsing videos in the historical interaction sequence are all animation videos, it indicates that the target account has a higher interest in animation videos in the browsing time period corresponding to the first three historical browsing videos. If the last historical browsing video in the historical interaction sequence is a variety show video, it indicates that the video content with the highest interest of the target account at the current time point is variety show videos. Therefore, selecting the most recent interaction data in the historical interaction sequence as the target interaction data can better express the target account's recent short-term interests.

[0095] Step 403 , concatenate the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data, and the target feature representation corresponding to the target interaction data to obtain a concatenated feature.

[0096] Illustratively, the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data, and the target feature representation corresponding to the target interaction data are sequentially spliced ​​together, or the account feature representation, the interaction feature representation, and the target feature representation are randomly arranged each time and spliced ​​together to obtain the spliced ​​features, which is not limited here.

[0097] Indicatively, Figure 5 As shown, it shows a schematic diagram of splicing features provided by an exemplary embodiment of the present application, such as Figure 5 As shown, Figure 5 The spliced ​​feature 500 includes the account feature representation 501 corresponding to the account basic data, the interaction feature representation 502 corresponding to the historical interaction data, and the target feature representation 503 corresponding to the target interaction data, wherein the splicing method is shown in Formula 1:

[0098] Formula 1: X = [f u ||f s ||f short ]

[0099] Among them, X represents the splicing feature, || represents splicing, and f u represents the account feature representation, f s represents the interactive feature representation, f short Represents the target feature representation.

[0100] It is worth noting that, in this embodiment, the account feature representation, the interaction feature representation, and the target feature representation are features expressed using the same type of feature representation.

[0101] Step 404, performing feature fusion extraction on the spliced ​​features to obtain fusion features.

[0102] In some embodiments, the splicing features are input into a feedforward neural network to obtain a first feedforward feature vector; based on the splicing features and the first feedforward feature vector, a first fusion result is determined; the attention vector corresponding to the first fusion result is determined through a multi-head self-attention mechanism; based on the first fusion result and the attention vector, a second fusion result is determined; based on the second fusion result and the convolution vector corresponding to the second fusion result, a third fusion result is obtained; based on the third fusion result and the second feedforward feature vector corresponding to the third fusion result, a fusion feature is obtained.

[0103] Illustratively, in this embodiment, two methods are used to fuse and extract the splicing features to obtain fused features.

[0104] The first one uses the Multi-Head Self Attention (MHSA) mechanism to globally fuse the splicing features.

[0105] For illustration, please refer to Figure 6 , which shows a schematic diagram of global fusion provided by an exemplary embodiment of the present application, such as Figure 6 As shown, the spliced ​​feature 601 is input into the first neural network 602 constructed by the multi-head self-attention mechanism, and the fused feature 603 is output.

[0106] In the neural network 602, the input concatenated feature 601 is firstly obtained through a feedforward neural network (FNN) to obtain a first feedforward feature vector, and the first feedforward feature vector is linearly converted from Q, K, V to corresponding features with different parameters through a self-attention network. and Among them, W corresponds to different superscripts to indicate different parameters, that is, refer to the following formula 2:

[0107] Formula 2:

[0108] Among them, Q stands for Query, K stands for Key, and V stands for Value. These three parameters represent the three sequence vectors in the attention mechanism.

[0109] After linear transformation, the three sequence vectors determine their corresponding attention values ​​through the attention function, that is, the weight values ​​corresponding to Value. For example, refer to Formula 3:

[0110] Formula 3:

[0111] in, d Q represents the dimension, To prevent the gradient from vanishing.

[0112] After determining the weight value corresponding to Value, use it as a single self-attention value, connect multiple self-attention values, and obtain a multi-head attention vector. The total consumption is reduced by reducing the dimension. For illustration, please refer to Formula 4 and Formula 5:

[0113] Formula 4:

[0114] Formula 5: MultiHead(Q, K, V) = [head1, head2,..., head h ]W O +b O ;

[0115] After determining multiple self-attention values, they are passed through a feedforward neural network and layer normalization, and the final output is the fused feature.

[0116] The second method uses a multi-head self-attention mechanism for global fusion, and then performs local fusion based on the results of global fusion.

[0117] For illustration, please refer to Figure 7 , which shows a schematic diagram of global fusion and local fusion provided by an exemplary embodiment of the present application, such as Figure 7 As shown, the splicing feature 701 is input into the second neural network 702, and the second neural network 702 includes the first neural network corresponding to the multi-head self-attention mechanism, and the fusion feature 703 is output.

[0118] Among them, firstly, the splicing feature is input into the feedforward neural network to obtain the first feedforward feature vector, and the first fusion result is determined according to the splicing feature and the first feedforward feature vector. Schematically, the method for determining the first fusion result is shown in Formula 6:

[0119] Formula 6:

[0120] in, represents the first fusion result, and FFM(X) represents the first feedforward feature vector.

[0121] Secondly, the first fusion result is input into the first neural network corresponding to the multi-head self-attention mechanism to determine the multi-head attention vector, and the second fusion result is determined according to the multi-head attention vector and the first fusion result. Schematically, the method for determining the second fusion result is shown in Formula 7:

[0122] Formula 7:

[0123] in, represents the multi-head attention vector, and X′ represents the second fusion result.

[0124] Next, the second fusion result is input into the convolution layer for convolution to obtain the convolution vector corresponding to the second fusion result. The third fusion result is determined according to the second fusion result and the convolution vector corresponding to the second fusion result. Schematically, the method for determining the third fusion result is shown in Formula 8:

[0125] Formula 8: X″=X′+ConvM(X′)

[0126] Among them, X″ represents the third fusion result, and ConvM(X′) represents the convolution vector corresponding to the second fusion result.

[0127] Finally, the third fusion result is input into the feedforward neural network and layer normalization to obtain the second feedforward feature vector corresponding to the third fusion result. According to the third fusion result and the second feedforward feature vector corresponding to the third fusion result, the fusion feature is determined. Schematically, the method for determining the fusion feature is shown in Formula 9:

[0128] Formula 9:

[0129] Among them, X fusion represents the fusion feature, and FFM(X″) represents the second feedforward feature vector corresponding to the third fusion result.

[0130] It is worth noting that the two fusion methods included in this embodiment are only illustrative examples and are not limited to this embodiment of the present application.

[0131] Step 405 , predicting the interest distribution of the content in the content library by concatenating features and fusing features to obtain an interest distribution prediction result.

[0132] After obtaining the fusion features, the interest distribution of the content in the content library is predicted by combining the splicing features and the fusion features. For example, please refer to Formula 10 and Formula 11 for the interest distribution prediction method:

[0133] Formula 10: h final =W f [X||X fusion ]

[0134] Formula 11:

[0135] In formula 10, h final The predicted feature representation represents the combination of the concatenated feature and the fused feature. The predicted feature representation is used to compare with the content in the content library to determine the similarity of the corresponding content.

[0136] In Formula 11, z iRepresents the predicted probability corresponding to the i-th content, and finally obtains a set of predicted probability distributions corresponding to each content in the content library as the interest distribution prediction result.

[0137] Indicatively, z i It is also necessary to perform normalization processing through the softmax function to make the predicted probability distribution more accurate. The predicted probability corresponding to each content indicates the probability that the target account will interact with it next time. The higher the predicted probability, the more likely the target account will interact with it next time. For example, taking video content as an example, the higher the predicted probability corresponding to the video, the more likely the target account will browse the video next time.

[0138] In summary, an embodiment of the present application provides a content recommendation method. In the process of fusing the account feature representation of the target account's corresponding account basic data and the interaction feature representation of the target account's corresponding historical interaction data, the target feature representation corresponding to the historical interaction sequence constructed by the interaction data of the target account within the historical time period is added. The three are fused to determine the fusion feature, and the interest distribution of the content is predicted based on the fusion feature. Finally, the content recommended to the target account is determined. By adding the interaction data containing the interaction sequence for feature fusion, it is easier to capture the interest preference of the target account in the content recommendation process, thereby improving the accuracy of content recommendation.

[0139] In this embodiment, two fusion methods are listed. The first is to globally fuse the splicing features through a multi-head self-attention mechanism to obtain a fused feature. The second is to first globally fuse the splicing through a multi-head self-attention mechanism, and then locally fuse the fusion result through a convolution layer to finally obtain a fused feature. Through the global fusion / global + local fusion method, the obtained fused feature can more comprehensively integrate the corresponding information of the account feature representation, the interaction feature representation and the target feature representation, thereby improving the accuracy of the fused feature.

[0140] In this embodiment, the content in the content library is predicted by combining the splicing feature and the fusion feature, thereby improving the prediction accuracy corresponding to the content recommendation degree.

[0141] In an optional embodiment, the concatenation process further includes concatenating the account feature representation with the interaction vector representation corresponding to the historical interaction data. Figure 8 , which shows a flow chart of a content recommendation method provided by an exemplary embodiment of the present application, such as Figure 8 As shown, the method comprises the following steps:

[0142] Step 801, obtaining basic account data and historical interaction data corresponding to the target account.

[0143] The account basic data is used to indicate the basic account information of the target account, and the historical interaction data is used to indicate the historical interaction sequence constructed by the interaction data of the target account in the historical time period.

[0144] The contents of the account basic data and historical interaction data in step 801 have been described in detail in the above step 301 and will not be repeated here.

[0145] Step 802: Acquire associated account data and associated interaction data corresponding to the associated account.

[0146] The associated account data is used to indicate basic account information of the associated account, and the associated interaction data includes historical interaction data corresponding to the associated account.

[0147] Illustratively, the associated accounts include accounts that have an associated relationship with the target account, wherein the associated relationship includes at least one of a friend relationship, a lover relationship, a family relationship, a work partnership relationship, etc., which is not limited here.

[0148] Among them, the associated account data includes the account profile information corresponding to the associated account, such as: the user gender, age, hobbies, etc. corresponding to the associated account; or, the account basic information includes the account preference information corresponding to the associated account, and the account preference information indicates that within a specific period of time, based on the historical operation records corresponding to the associated account, the interest preferences corresponding to the associated account are determined, such as: when the associated account is the corresponding account in a video application, the video type corresponding to the video content browsed by the associated account within a specified historical time period is used as the account preference information corresponding to the associated account.

[0149] Among them, the associated account data will be updated after the associated account updates the basic account information.

[0150] Illustratively, the associated interaction data represents the interaction data generated between the associated account and historical content or historical items within a historical time period. For example, taking the associated account as a video application, when the associated account generates a trigger operation on a certain video content, or browses the video content, the triggering behavior or browsing behavior generates the interaction data between the associated account and the video content.

[0151] It is understandable that in the specific embodiments of the present application, related data such as associated account data and associated interaction data are involved. When the above embodiments of the present application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0152] It is worth noting that when the target account and the associated account have the same historical interaction data, the historical interaction data serves as both the historical interaction data corresponding to the target account and the associated interaction data corresponding to the associated account, which is not limited here.

[0153] Step 803, based on the account basic data, historical interaction data, associated accounts and associated interaction data, determine the target knowledge graph corresponding to the target account.

[0154] Among them, the target knowledge graph is used to indicate the relationship between account basic data, historical interaction data, associated accounts and associated interaction data.

[0155] For illustration, please refer to Fig. 9 , which shows a schematic diagram of a target knowledge graph provided by an exemplary embodiment, such as Fig. 9 As shown, Fig. 9 It includes account basic data 901 (located in the center of the graph), historical interaction data 902 ( Fig. 9 ) and associated account data 903 ( Fig. 9 The target knowledge graph 910 is represented by black circles except the center of the graph. Fig. 9 The historical interaction data 902 and the associated interaction data are the same interaction data, and the associated interaction data is not displayed here.

[0156] Among them, the account basic data 901 is connected to the historical interaction data 902, the account basic data 901 is connected to the associated account data 903, and the historical interaction data 902 with a data association relationship are connected, wherein the data association relationship includes two historical interaction data being of the same type or two interaction data corresponding to the same project, etc., which are not limited here.

[0157] Illustratively, since the data types corresponding to the associated account data 903 and the account basic data 901 are both account data and have the same data structure type, the associated account data 903 is used as the “homogeneous neighbor” corresponding to the account basic data 901. Since the data type corresponding to the historical interaction data 902 is interaction data, which belongs to a different data structure type from the account basic data 901, the historical interaction data 902 is used as the “heterogeneous neighbor” corresponding to the account basic data 901.

[0158] Step 804: Based on the target knowledge graph, determine the account node corresponding to the account basic data and the interaction node corresponding to the historical interaction data.

[0159] In some embodiments, based on the target knowledge graph, the associated node information corresponding to the associated account data and the interaction node information corresponding to the historical interaction data are determined; based on the associated node information and the interaction node information, the account node corresponding to the account basic data and the interaction node corresponding to the historical interaction data are determined.

[0160] Illustratively, the account node includes node information corresponding to the account data base, and the interaction node includes node information corresponding to the historical interaction data.

[0161] Indicatively, Fig. 9 As shown, in the target knowledge graph 910, each data (including account basic data 901, historical interaction data 902 and associated account data 903) is represented as a node in the target knowledge graph 910, so each node corresponds to a node information. When the data is updated, the node information corresponding to the data is also updated. It is worth noting that the target knowledge graph 910 is a layered structure diagram, and each layer contains historical interaction data 902 or associated account data 903. Therefore, in the update process, it is necessary to finally determine the updated node information through layer-by-layer updating.

[0162] In the target knowledge graph, since there is a connection between each node (that is, the "line" between nodes represents the connection between the two nodes), when updating the account basic data corresponding to the target account, it is necessary to determine the importance of the "homomorphic neighbors" and "heterogeneous neighbors" corresponding to the account basic data and aggregate the two to determine the update information corresponding to the nodes in the current layer.

[0163] First, we need to calculate the importance of each node in the target knowledge graph. For example, refer to Formula 12 and Formula 13 for determining the importance:

[0164] Formula 12:

[0165] Formula 13:

[0166] Among them, α uv represents the importance score, e uv is α uv The attention weight obtained after normalization, and are all learnable parameters, u, v∈U, represents the node u on the l-1th layer, N u Represents the “neighbors” corresponding to node u (including homogeneous neighbors and heterogeneous neighbors).

[0167] As can be seen from Formula 12, by connecting the associated node information and the interactive node information in the target knowledge graph, the importance score corresponding to each node is determined, and then the importance score is normalized by Formula 13 to obtain the attention weight.

[0168] Indicatively, Fig. 9 As shown, Fig. 9 Also shown is a current layer aggregation method, including after determining the associated node information corresponding to the associated account data 903, constructing an associated relationship graph 911 with the associated node information and the account basic data 901, and after determining the interactive node information corresponding to the historical interactive data 902, constructing an interactive relationship graph 912 with the interactive node information and the account basic data 901. When the associated account data 903 or the historical interactive data 902 is updated, the associated node information or the interactive node information is updated accordingly, that is, the associated relationship graph 911 or the interactive relationship graph 912 is updated.

[0169] Among them, please refer to Formula 14 and Formula 15 for the aggregation method:

[0170] Formula 14:

[0171] Formula 15:

[0172] Among them, in Formula 14, W3 and b are learnable parameters, and MLP in Formula 15 represents a multilayer perceptron. Formulas 14 and 15 can be used to determine the nodes corresponding to each node u in a single layer in the target knowledge graph, including associated nodes corresponding to associated account data and interactive nodes corresponding to historical interaction data.

[0173] Indicatively, Fig. 9 As shown, the association nodes are used to construct the association relationship graph 911, and the interaction nodes are used to construct the interaction relationship graph 912. The association relationship graph 911 and the interaction relationship graph 912 are stacked through the stacking layer 913 to finally obtain the updated knowledge graph 914. The updated knowledge graph 914 includes the account node representation corresponding to the account basic data and the interaction node representation corresponding to the historical interaction data.

[0174] Step 805: Acquire target interaction data within a specified interaction interval in the historical interaction sequence.

[0175] The specified interaction interval is related to the interest time distribution of the target account.

[0176] The content of the target interaction data in step 805 has been described in detail in the above step 302 and will not be repeated here.

[0177] It is worth noting that steps 8061 to 8063 are the process of determining the interactive feature representation, and steps 8071 to 8072 are the process of determining the target feature representation. In this embodiment, the two processes are performed simultaneously.

[0178] Step 8061, determining an account feature representation based on the account node, and determining an interaction vector representation based on the interaction node.

[0179] In this embodiment, the account node and the interaction node are input into the embedding layer to obtain the account feature representation corresponding to the account node and the interaction vector representation corresponding to the interaction node.

[0180] Step 8062, concatenate the account feature representation and the interaction vector representation to obtain an interaction concatenation vector.

[0181] In this embodiment, the account feature representation and the interaction vector representation are concatenated left and right, and the concatenated result is used as the interaction concatenated vector. For the concatenating method, please refer to Formula 16, Formula 17 and Formula 18:

[0182] Formula 16:

[0183] Formula 17:

[0184] Formula 18:

[0185] Among them, W u , W i 、b u and b i are learnable parameters, Represents an account node, represents the interaction vector, is the initial representation of node i, that is, the interaction concatenation vector.

[0186] Step 8063, extract features from the interactive concatenation vector to obtain interactive feature representation.

[0187] In this embodiment, the interactive splicing vector is feature extracted through a graph attention network (GAT) to obtain an interactive feature representation, wherein the feature extraction method is shown in Formula 19, Formula 20 and Formula 21:

[0188] Formula 19:

[0189] Formula 20:

[0190] Formula 21:

[0191] Among them, W1, W2 and W3 are learnable parameters. It represents the feature representation in the t-th layer structure. When the graph attention mechanism contains an L-layer network structure, the sum of the feature representations corresponding to each layer of the L layers is the interactive feature representation (not shown in the formula).

[0192] Step 8071, obtaining a target stitching vector in the interactive stitching vector that is within a specified interactive interval.

[0193] Schematically, the last vector in the interactive splicing vector is obtained as the target splicing vector.

[0194] Step 8072, perform feature extraction on the target concatenated vector to obtain target feature representation.

[0195] Schematically, by extracting the feature representation in the target concatenated vector, the extraction result is obtained as the target feature representation.

[0196] Step 808 , the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data, and the target feature representation corresponding to the target interaction data are fused to obtain a fused feature.

[0197] The relevant contents in step 808 have been described in detail in the above step 303 and will not be repeated here.

[0198] Step 809: predict the interest distribution of the content in the content library based on the fusion features to obtain an interest distribution prediction result.

[0199] Among them, the interest distribution prediction result is used to determine the content to be recommended to the target account.

[0200] The content of the interest distribution prediction result in step 809 has been described in detail in the above steps 304 and 405 and will not be repeated here.

[0201] In summary, an embodiment of the present application provides a content recommendation method. In the process of fusing the account feature representation of the target account's corresponding account basic data and the interaction feature representation of the target account's corresponding historical interaction data, the target feature representation corresponding to the historical interaction sequence constructed by the interaction data of the target account within the historical time period is added. The three are fused to determine the fusion feature, and the interest distribution of the content is predicted based on the fusion feature. Finally, the content recommended to the target account is determined. By adding the interaction data containing the interaction sequence for feature fusion, it is easier to capture the interest preference of the target account in the content recommendation process, thereby improving the accuracy of content recommendation.

[0202] In this embodiment, the accuracy of the account basic data and the historical interaction data can be improved by constructing a target knowledge graph including homogeneous neighbors and heterogeneous neighbors and updating the account basic data and the historical interaction data according to the information of each node.

[0203] In an alternative embodiment, please refer to Fig.10 , which shows a schematic diagram of a content recommendation method provided by an exemplary embodiment of the present application, such as Fig.10 As shown, the method includes four parts:

[0204] The first part is the graph embedding module 1010.

[0205] Through the association relationship between the target account and the associated account, and the interaction sequence formed by the historical interaction data 1011 corresponding to the target account and the associated interaction data 1012 corresponding to the associated account, a target knowledge graph 1013 in the graph embedding module 1010 is constructed, wherein the target knowledge graph 1013 includes isomorphic neighbors (i.e., associated account data) and heterogeneous neighbors (i.e., historical interaction data and associated interaction data) corresponding to the account basic data corresponding to the target account, which are used to obtain the node information corresponding to each data node, calculate the importance score corresponding to each neighbor according to the heterogeneous graph attention network (HGNN), and update and confirm the account node 1014 corresponding to the account basic data and the interaction node 1015 corresponding to the historical interaction data 1011.

[0206] The second part is a session embedding module 1020 based on the target account.

[0207] After the account node 1014 and the interaction node 1015 are obtained, they are input into the target account-based session embedding module 1020. First, the account node 1014 and the interaction node 1015 are input into the embedding layer to obtain the account feature representation 1021 corresponding to the account node 1014 and the interaction vector representation 1022 corresponding to the interaction node 1015.

[0208] Secondly, the account feature representation 1021 and the interaction vector representation 1022 are concatenated through the neural network 1023 corresponding to the graph attention mechanism to obtain an interaction concatenated vector 1024, and features are extracted from the interaction concatenated vector 1024 to obtain an interaction feature representation.

[0209] Finally, in the process of obtaining the interactive splicing vector, the interactive splicing vector at the end of the sorting is selected as the target splicing vector 1025, and features are extracted from the target splicing vector 1025 to obtain a target feature representation.

[0210] That is, the target account-based session embedding module 1020 is used to obtain the account feature representation 1021, the interaction feature representation, and the target feature representation.

[0211] The third part is the feature fusion module 1030.

[0212] The account feature representation 1021, the interaction feature representation and the target feature representation are spliced ​​to obtain a spliced ​​feature 1031, and the spliced ​​feature 1031 is fused through a high-level feature fusion network to obtain a fused feature 1032, wherein the spliced ​​feature 1031 is first globally fused through a multi-head self-attention mechanism in the high-level feature network, and then the fusion results corresponding to the global fusion are locally fused through a convolutional network to finally obtain a fused feature 1032.

[0213] The fourth part is the prediction module 1040.

[0214] After obtaining the fusion feature 1032, the prediction layer calculates the similarity between the content in the content library and the fusion feature 1032 to obtain the prediction probability corresponding to each content, and finally obtains the interest distribution prediction result corresponding to the content library. The recommended content 1041 that is finally recommended to the target account is determined based on the interest distribution prediction result.

[0215] In summary, an embodiment of the present application provides a content recommendation method. In the process of fusing the account feature representation of the target account's corresponding account basic data and the interaction feature representation of the target account's corresponding historical interaction data, the target feature representation corresponding to the historical interaction sequence constructed by the interaction data of the target account within the historical time period is added. The three are fused to determine the fusion feature, and the interest distribution of the content is predicted based on the fusion feature. Finally, the content recommended to the target account is determined. By adding the interaction data containing the interaction sequence for feature fusion, it is easier to capture the interest preference of the target account in the content recommendation process, thereby improving the accuracy of content recommendation.

[0216] The beneficial effects of this program also include:

[0217] (1) In this scheme, we first generate node representations corresponding to the target account and historical interaction data through heterogeneous knowledge graph embedding, which can model the association relationship between accounts and interaction data at the same time. Then, a graph attention network is used to model the session representation based on the corresponding target account transformation. A high-level feature fusion module is proposed to generate high-level fusion features between the target account and interaction data, including global and local fusion mechanisms. The former is the basis and the latter is the supplementary part. Finally, the concatenated features and fused features are used for item recommendation. This helps to improve the accuracy of content recommendation.

[0218] (2) The feature fusion module mentioned in this scheme helps to improve the performance of the model, and this scheme effectively utilizes social relationships to enhance the recommendation performance.

[0219] In addition, this application uses graph embedding for end-to-end training, and other supervised methods can be used to offline calculate the embedding layer feature representation of user and interaction data. The session embedding module based on the target account uses the graph attention mechanism, and other graph convolutional neural network algorithms can be used here.

[0220] Fig.11 is a structural block diagram of a content recommendation device provided by an exemplary embodiment of the present application. Fig.11 As shown, the device comprises:

[0221] An acquisition module 1110 is used to acquire account basic data and historical interaction data corresponding to a target account, wherein the account basic data is used to indicate basic account information of the target account, and the historical interaction data is used to indicate a historical interaction sequence constructed by interaction data of the target account within a historical time period;

[0222] The acquisition module 1110 is further configured to acquire target interaction data within a specified interaction interval in the historical interaction sequence, wherein the specified interaction interval is related to the interest time distribution of the target account;

[0223] A fusion module 1120, configured to fuse the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data, and the target feature representation corresponding to the target interaction data to obtain a fusion feature;

[0224] The prediction module 1130 is used to predict the interest distribution of the content in the content library based on the fusion feature to obtain an interest distribution prediction result, and the interest distribution prediction result is used to determine the content to be recommended to the target account.

[0225] In an optional embodiment, the acquisition module 1110 is further configured to acquire the latest n interaction data in the historical interaction sequence as the target interaction data, where n is a positive integer.

[0226] In an optional embodiment, the acquisition module 1110 is further configured to acquire the most recent interaction data in the historical interaction sequence as the target interaction data.

[0227] In an optional embodiment, the fusion module 1120 includes:

[0228] A splicing unit 1121, configured to splice the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data, and the target feature representation corresponding to the target interaction data to obtain a spliced ​​feature;

[0229] The extraction unit 1122 is used to perform feature fusion extraction on the splicing features to obtain the fusion features.

[0230] In an optional embodiment, the prediction module 1130 is further used to predict the interest distribution of the content in the content library by using the splicing feature and the fusion feature to obtain the interest distribution prediction result.

[0231] In an optional embodiment, the extraction unit 1122 is also used to input the splicing feature into a feedforward neural network to obtain a first feedforward feature vector; determine a first fusion result based on the splicing feature and the first feedforward feature vector; determine the attention vector corresponding to the first fusion result through a multi-head self-attention mechanism; determine a second fusion result based on the first fusion result and the attention vector; obtain a third fusion result based on the second fusion result and the convolution vector corresponding to the second fusion result; obtain the fusion feature based on the third fusion result and the second feedforward feature vector corresponding to the third fusion result.

[0232] In an optional embodiment, the fusion module 1120 further includes:

[0233] The extraction unit 1122 is further configured to extract an account feature representation corresponding to the account basic data and an interaction vector representation corresponding to the historical interaction data;

[0234] The concatenation unit 1121 is further used to concatenate the account feature representation and the interaction vector representation to obtain an interaction concatenation vector;

[0235] The extraction unit 1122 is further used to perform feature extraction on the interactive splicing vector to obtain the interactive feature representation.

[0236] In an optional embodiment, the fusion module 1120 further includes:

[0237] An acquiring unit 1123 is configured to acquire a target stitching vector in the interactive stitching vector that is within the specified interactive interval;

[0238] The extraction unit 1122 is further used to perform feature extraction on the target splicing vector to obtain the target feature representation.

[0239] In an optional embodiment, the fusion module 1120 further includes:

[0240] The acquisition unit 1123 is further configured to acquire associated account data and associated interaction data corresponding to the associated account, wherein the associated account data is used to indicate basic account information of the associated account, and the associated interaction data includes historical interaction data corresponding to the associated account;

[0241] A determining unit 1124 is configured to determine a target knowledge graph corresponding to the target account based on the account basic data, the historical interaction data, the associated account, and the associated interaction data, wherein the target knowledge graph is used to indicate an association relationship between the account basic data, the historical interaction data, the associated account, and the associated interaction data;

[0242] The determining unit 1124 is further configured to determine, based on the target knowledge graph, the account node corresponding to the account basic data and the interaction node corresponding to the historical interaction data.

[0243] In an optional embodiment, the determining unit 1124 determines the account feature representation based on the account node, and determines the interaction vector representation based on the interaction node.

[0244] In an optional embodiment, the determination unit 1124 is further used to determine, based on the target knowledge graph, the associated node information corresponding to the associated account data and the interaction node information corresponding to the historical interaction data.

[0245] In summary, an embodiment of the present application provides a content recommendation device. In the process of fusing the account feature representation of the target account corresponding to the account basic data and the interaction feature representation of the target account corresponding to the historical interaction data, the target feature representation corresponding to the historical interaction sequence constructed by the interaction data of the target account in the historical time period is added, the three are fused to determine the fusion feature, and the interest distribution of the content is predicted based on the fusion feature, and finally the content recommended to the target account is determined. By adding the interaction data containing the interaction sequence for feature fusion, it is easier to capture the interest preference of the target account in the content recommendation process, thereby improving the accuracy of content recommendation.

[0246] It should be noted that the content recommendation device provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the content recommendation device provided in the above embodiment and the content recommendation method embodiment belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0247] Fig.13The following is a schematic diagram showing the structure of a server provided by an exemplary embodiment of the present application. Specifically:

[0248] The server 1300 includes a central processing unit (CPU) 1301, a system memory 1304 including a random access memory (RAM) 1302 and a read only memory (ROM) 1303, and a system bus 1305 connecting the system memory 1304 and the central processing unit 1301. The server 1300 also includes a mass storage device 1306 for storing an operating system 1313, application programs 1314, and other program modules 1315.

[0249] The mass storage device 1306 is connected to the central processing unit 1301 through a mass storage controller (not shown) connected to the system bus 1305. The mass storage device 1306 and its associated computer readable media provide non-volatile storage for the server 1300. That is, the mass storage device 1306 may include a computer readable medium (not shown) such as a hard disk or a compact disc read only memory (CD-ROM) drive.

[0250] Without loss of generality, computer-readable media may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer storage media include RAM, ROM, Erasable Programmable Read Only Memory (EPROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other solid-state storage technology, CD-ROM, Digital Versatile Disc (DVD) or other optical storage, cassettes, tapes, disk storage or other magnetic storage devices. Of course, those skilled in the art will appreciate that computer storage media are not limited to the above. The above-mentioned system memory 1304 and mass storage device 1306 can be collectively referred to as memory.

[0251] According to various embodiments of the present application, the server 1300 can also be connected to a remote computer on the network through a network such as the Internet. That is, the server 1300 can be connected to the network 1312 through the network interface unit 1311 connected to the system bus 1305, or the network interface unit 1311 can be used to connect to other types of networks or remote computer systems (not shown).

[0252] The above-mentioned memory also includes one or more programs, and the one or more programs are stored in the memory and configured to be executed by the CPU.

[0253] An embodiment of the present application also provides a computer device, which includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, at least one program, a code set or an instruction set is loaded and executed by the processor to implement the content recommendation method provided by the above-mentioned method embodiments.

[0254] An embodiment of the present application also provides a computer-readable storage medium, on which is stored at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by a processor to implement the content recommendation method provided by the above-mentioned method embodiments.

[0255] The embodiment of the present application also provides a computer program product, the computer program product or computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs any of the content recommendation methods described in the above embodiments.

[0256] Optionally, the computer readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a solid state drive (SSD), or an optical disk. Among them, the random access memory may include a resistance random access memory (ReRAM) and a dynamic random access memory (DRAM). The serial numbers of the above embodiments of the present application are only for description and do not represent the advantages and disadvantages of the embodiments.

[0257] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0258] The above description is only an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A content recommendation method, It is characterized in that The method comprises: Acquire account basic data and historical interaction data corresponding to the target account, wherein the account basic data is used to indicate basic account information of the target account, and the historical interaction data is used to indicate a historical interaction sequence constructed by interaction data of the target account within a historical time period; Acquire target interaction data within a specified interaction interval in the historical interaction sequence, where the specified interaction interval is related to the interest time distribution of the target account; Acquire associated account data and associated interaction data corresponding to the associated account, wherein the associated account data is used to indicate basic account information of the associated account, and the associated interaction data includes historical interaction data corresponding to the associated account; Determine a target knowledge graph corresponding to the target account based on the account basic data, the historical interaction data, the associated account, and the associated interaction data; Based on the target knowledge graph, determine the account node corresponding to the account basic data and the interaction node corresponding to the historical interaction data; determining an account feature representation based on the account node, and determining an interaction vector representation based on the interaction node; Concatenate the account feature representation and the interaction vector representation to obtain an interaction concatenation vector; Performing feature extraction on the interactive splicing vector to obtain an interactive feature representation; fusing the account feature representation, the interaction feature representation, and the target feature representation corresponding to the target interaction data to obtain a fused feature; Predicting interest distribution of the content in the content library based on the fusion feature to obtain an interest distribution prediction result, wherein the interest distribution prediction result is used to determine the content to be recommended to the target account; The updating process of the target knowledge graph includes: Based on the target knowledge graph, determining the associated node information corresponding to the associated account data and the interactive node information corresponding to the historical interactive data; Constructing an association relationship graph with the associated node information and the account basic data; and constructing an interaction relationship graph with the interactive node information and the account basic data; When the associated account data is updated, the associated node information is updated accordingly, so that the associated relationship graph is updated accordingly; or, when the historical interaction data is updated, the interaction node information is updated accordingly, so that the interaction relationship graph is updated accordingly; The association relationship graph and the interaction relationship graph are stacked through stacking layers to obtain an updated knowledge graph.

2. The method according to claim 1, It is characterized in that The acquiring target interaction data in a specified interaction interval in the historical interaction sequence includes: The latest n interaction data in the historical interaction sequence are obtained as the target interaction data, where n is a positive integer.

3. The method according to claim 2, It is characterized in that The acquiring the most recent n interaction data in the historical interaction sequence as the target interaction data includes: The most recent interaction data in the historical interaction sequence is obtained as the target interaction data.

4. The method according to claim 1, It is characterized in that The fusing the account feature representation, the interaction feature representation, and the target feature representation corresponding to the target interaction data to obtain a fused feature includes: splicing the account feature representation corresponding to the account basic data, the interaction feature representation corresponding to the historical interaction data, and the target feature representation corresponding to the target interaction data to obtain a spliced ​​feature; Feature fusion extraction is performed on the splicing features to obtain the fusion features.

5. The method according to claim 4, It is characterized in that The performing interest distribution prediction on the content in the content library based on the fusion feature to obtain the interest distribution prediction result includes: The interest distribution prediction is performed on the content in the content library by using the splicing feature and the fusion feature to obtain the interest distribution prediction result.

6. The method according to claim 4, It is characterized in that The step of performing feature fusion extraction on the splicing features to obtain the fusion features includes: Inputting the concatenated features into a feedforward neural network to obtain a first feedforward feature vector; Determine a first fusion result based on the splicing feature and the first feedforward feature vector; Determine the attention vector corresponding to the first fusion result through a multi-head self-attention mechanism; Determine a second fusion result based on the first fusion result and the attention vector; Obtaining a third fusion result based on the second fusion result and the convolution vector corresponding to the second fusion result; The fusion feature is obtained based on the third fusion result and the second feedforward feature vector corresponding to the third fusion result.

7. The method according to claim 4, It is characterized in that The method further comprises: Obtaining a target stitching vector in the interactive stitching vector that is within the specified interactive interval; Perform feature extraction on the target splicing vector to obtain the target feature representation.

8. The method according to claim 1, It is characterized in that The determining, based on the target knowledge graph, the account node corresponding to the account basic data and the interaction node corresponding to the historical interaction data includes: Based on the target knowledge graph, determining the associated node information corresponding to the associated account data and the interactive node information corresponding to the historical interactive data; Based on the associated node information and the interactive node information, the account node corresponding to the account basic data and the interactive node corresponding to the historical interactive data are determined.

9. A content recommendation device, It is characterized in that The device comprises: An acquisition module, used to acquire account basic data and historical interaction data corresponding to a target account, wherein the account basic data is used to indicate basic account information of the target account, and the historical interaction data is used to indicate a historical interaction sequence constructed by interaction data of the target account within a historical time period; The acquisition module is further used to acquire target interaction data within a specified interaction interval in the historical interaction sequence, where the specified interaction interval is related to the interest time distribution of the target account; an acquisition unit, configured to acquire associated account data and associated interaction data corresponding to the associated account, wherein the associated account data is used to indicate basic account information of the associated account, and the associated interaction data includes historical interaction data corresponding to the associated account; a determining unit, configured to determine a target knowledge graph corresponding to the target account based on the account basic data, the historical interaction data, the associated account, and the associated interaction data; determine an account node corresponding to the account basic data and an interaction node corresponding to the historical interaction data based on the target knowledge graph; determine an account feature representation based on the account node, and determine an interaction vector representation based on the interaction node; A concatenation unit, configured to concatenate the account feature representation and the interaction vector representation to obtain an interaction concatenation vector; An extraction unit, used for performing feature extraction on the interactive splicing vector to obtain an interactive feature representation; A fusion module, used to fuse the account feature representation, the interaction feature representation and the target feature representation corresponding to the target interaction data to obtain a fusion feature; A prediction module, configured to predict the interest distribution of the content in the content library based on the fusion feature to obtain an interest distribution prediction result, wherein the interest distribution prediction result is used to determine the content to be recommended to the target account; The determination unit is further used to determine, based on the target knowledge graph, the associated node information corresponding to the associated account data and the interactive node information corresponding to the historical interaction data; construct an associated relationship graph with the associated node information and the account basic data; and construct an interactive relationship graph with the interactive node information and the account basic data; when the associated account data is updated, the associated node information is updated accordingly so that the associated relationship graph is updated accordingly; or, when the historical interaction data is updated, the interactive node information is updated accordingly so that the interactive relationship graph is updated accordingly; and the associated relationship graph and the interactive relationship graph are stacked through stacking layers to obtain an updated knowledge graph.

10. A computer device, It is characterized in that The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the content recommendation method as described in any one of claims 1 to 8.

11. A computer-readable storage medium, It is characterized in that The storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the content recommendation method as described in any one of claims 1 to 8.

12. A computer program product, It is characterized in that The method comprises a computer program or an instruction, which, when executed by a processor, implements the content recommendation method as claimed in any one of claims 1 to 8.

Images