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

Abstract

The present application discloses a content recommendation method, apparatus, device, storage medium, and program product, relating to the field of machine learning. The method includes: obtaining target account attribute data of a target object account and target content attribute data of a target content; extracting account feature representation and content feature representation; obtaining a multi-order interaction feature representation based on a model parameter search space, wherein the model parameter search space includes architectural parameters corresponding to the multi-order interaction relationships between the account attribute data and the content attribute data; performing predictive analysis on the multi-order interaction feature representation to obtain a recommendation probability between the target object account and the target content; and recommending content to the target object account based on the recommendation probability. That is, by searching the model parameter search space to obtain the multi-order interaction feature representation, and then performing predictive analysis and recommendation, the method can reduce space storage and improve the efficiency and 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 program product. Background Art

[0002] Content recommendation usually refers to displaying recommended content on the terminal to achieve the purpose of recommending information to users, such as: recommending product images, recommending video content, recommending website links, etc. Taking recommended product images as an example, the target product image is displayed on the terminal to recommend the target product to users, thereby achieving product promotion effects.

[0003] In related technologies, content recommendation methods are generally implemented by constructing a click-through rate prediction task, using a deep sparse network to perform feature interaction between the user's account attribute data and the candidate content attribute data, performing content recommendation analysis on the obtained interaction features, and selecting target content from the candidate content attribute data for recommendation based on the analysis results.

[0004] However, in the process of using deep sparse networks for content recommendation analysis, due to the diversity of recommended content, different deep sparse networks need to be used to analyze different types of candidate content, resulting in the consumption of a large amount of storage space to store different deep sparse networks, which reduces the content recommendation efficiency and the recommendation accuracy. Summary of the Invention

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

[0006] In one aspect, a content recommendation method is provided, the method comprising:

[0007] Obtain target account attribute data of the target object account and target content attribute data of the target content;

[0008] extracting an account feature representation corresponding to the target account attribute data, and extracting a content feature representation corresponding to the target content attribute data;

[0009] Obtaining a multi-order interaction feature representation of the interaction between the account feature representation and the content feature representation based on a model parameter search space, wherein the model parameter search space includes architectural parameters corresponding to the multi-order interaction relationships between the account attribute data and the content attribute data, the architectural parameters being used to represent the effectiveness of the interaction relationships in the recommendation prediction process;

[0010] Performing predictive analysis on the multi-order interaction feature representation to obtain a recommendation probability between the target object account and the target content;

[0011] Content is recommended to the target account based on the recommendation probability.

[0012] In another aspect, a content recommendation device is provided, the device comprising:

[0013] An acquisition module, used to acquire target account attribute data of a target object account and target content attribute data of a target content;

[0014] an extraction module, configured to extract an account feature representation corresponding to the target account attribute data, and to extract a content feature representation corresponding to the target content attribute data;

[0015] The acquisition module is further configured to acquire, based on a model parameter search space, a multi-order interaction feature representation of the interaction between the account feature representation and the content feature representation, wherein the model parameter search space includes architectural parameters corresponding to the multi-order interaction relationships between the account attribute data and the content attribute data, the architectural parameters being used to represent the effectiveness of the interaction relationship in the recommendation prediction process;

[0016] An analysis module, configured to perform predictive analysis on the multi-order interaction feature representation to obtain a recommendation probability between the target object account and the target content;

[0017] A recommendation module is used to recommend content to the target account based on the recommendation probability.

[0018] 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.

[0019] On the other hand, a computer-readable storage medium is provided, wherein 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 a processor to implement a content recommendation method as described in any of the above-mentioned embodiments of the present application.

[0020] In another aspect, a computer program product is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the content recommendation method described in any of the above embodiments.

[0021] The beneficial effects of the technical solutions provided in the embodiments of the present application include at least:

[0022] By storing the architecture parameters corresponding to the multi-order interaction relationship between the account attribute data and the content attribute data in the model parameter search space, the effectiveness of the relationship in the recommendation prediction process is represented. The multi-order interaction feature representation between the account feature representation of the target object account and the content feature representation of the target content is obtained according to the model parameter search space, and a predictive analysis is performed on them to determine the recommendation probability of the target content. A more effective multi-order interaction feature representation can be obtained based on the architecture parameters, and a predictive analysis can be performed on it, so that the accuracy of the prediction probability is higher. In addition, storing the architecture parameters in the search space instead of storing the deep sparse network can reduce the storage amount, thereby improving the efficiency of content recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. 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 any creative work.

[0024] Figure 1 This is a schematic diagram of a content recommendation method provided by an exemplary embodiment of the present application;

[0025] Figure 2 This is a schematic diagram of an implementation environment provided by an exemplary embodiment of the present application;

[0026] Figure 3 is a flow chart of a content recommendation method provided by an exemplary embodiment of the present application;

[0027] Figure 4 is a flow chart of a content recommendation method provided by another exemplary embodiment of the present application;

[0028] Figure 5 is a flow chart of a content recommendation method provided by another exemplary embodiment of the present application;

[0029] Figure 6 This is a schematic diagram of a content recommendation method provided by an exemplary embodiment of the present application;

[0030] Figure 7 is a structural block diagram of a content recommendation device provided by an exemplary embodiment of the present application;

[0031] Figure 8 is a structural block diagram of a content recommendation device provided by another exemplary embodiment of the present application;

[0032] Figure 9It is a structural diagram of a server provided by an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0034] First, a brief introduction is given to the terms involved in the embodiments of this application.

[0035] Artificial Intelligence (AI) is the theory, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0036] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0037] Machine Learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning through demonstration.

[0038] For illustration purposes, the technical solutions provided in this application are introduced. For illustration purposes, 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 1As shown, taking the case where the target object account is implemented as a consumer account as an example, target account attribute data 110 of the target object account and target content attribute data 120 of the target content are obtained, wherein the target account attribute data 110 includes account information, historical consumption records, historical browsing records and other data, and the target content attribute data 120 is implemented as product attribute data, and the product attribute data includes product name, product sales, product historical recommendation records and other data.

[0039] The account feature representation 111 corresponding to the target account attribute data 110 is extracted, and the content feature representation 121 corresponding to the target content attribute data 120 is extracted. A multi-order interaction feature representation 131 of the interaction between the account feature representation 111 and the content feature representation 121 is obtained based on the model parameter search space 130. The model parameter search space 130 includes architecture parameters 132 corresponding to the multi-order interaction relationships between the multiple account attribute data and the multiple content attribute data. The architecture parameters 132 represent the effectiveness of the multi-order interaction relationship in the content recommendation prediction process. For example, the model parameter search space 130 includes architecture parameters a (0.5), architecture parameters b (0.7), and architecture parameters c (0.9). The architecture parameter a is implemented as the effectiveness of the multi-order interaction between account attribute data 1, account number 2, and content 1, which is 0.5.

[0040] Predictive analysis is performed on the multi-order interaction feature representation 131 to obtain a recommendation probability 140 between the target account and the target content, which is used to represent the probability value of recommending the target content to the target account. Content is recommended to the target account based on the recommendation probability 140.

[0041] The content recommendation method provided in the present application represents the effectiveness of the relationship in the recommendation prediction process by using the architecture parameters corresponding to the multi-order interaction relationship between the account attribute data and the content attribute data stored in the model parameter search space. The multi-order interaction feature representation between the account feature representation of the target object account and the content feature representation of the target content is obtained according to the model parameter search space, and a predictive analysis is performed on them to determine the recommendation probability of the target content. The multi-order interaction feature representation with higher effectiveness can be obtained based on the architecture parameters, and a predictive analysis can be performed on them, so that the accuracy of the prediction probability is higher. In addition, storing the architecture parameters in the search space instead of storing the deep sparse network can reduce the storage amount, thereby improving the efficiency of content recommendation.

[0042] The content recommendation method provided in the embodiment of the present application can be implemented by a terminal or a server alone, or by a terminal and a server together. Taking the terminal and a server together implementing the content recommendation method as an example, the implementation environment involved in the embodiment of the present application is described. For schematic illustration, please refer to Figure 2The implementation environment involves a terminal 210 and a server 220 , and the terminal 210 and the server 220 are connected via a communication network 230 .

[0043] In some embodiments, an application 211 with a content recommendation function is run in the terminal 210, wherein the application 211 can be implemented as at least one of a shopping program, a social program, a game program, a functional program, and the like.

[0044] In one feasible scenario, when the terminal 210 runs the application 211 using the target object account, the terminal 210 generates a recommendation request using the target account attribute data of the target object account and automatically sends the request to the server 220 .

[0045] In a feasible scenario, during the process of running the application 211 in the target object account in the terminal 210, when the terminal 210 receives a content recommendation trigger operation, a recommendation request is generated based on the content recommendation trigger operation and sent to the server 220, wherein the content recommendation trigger operation is used to trigger the display of recommended content, and the recommendation request includes the target account attribute data 221 corresponding to the target object account.

[0046] After receiving the recommendation request, server 220 performs feature extraction on the target content attribute data 222 of the target content and the target account attribute data 221 of the target object account stored in server 220, obtaining an account feature representation corresponding to target account attribute data 221 and a content feature representation corresponding to target content attribute data 222. Server 220 also includes a model parameter search space 223. Based on this model parameter search space, a multi-order interaction feature representation 224 is obtained, which represents the interaction between the account feature representation and the content feature representation. Furthermore, model parameter search space 223 includes architectural parameters corresponding to each of the multi-order interaction feature representations 224, indicating the effectiveness of the multi-order interaction relationships. Predictive analysis is performed on the multi-order interaction feature representation 224 to obtain a recommendation probability corresponding to the target content. This recommendation probability is fed back to terminal 210 as a recommendation prediction result, and terminal 210 displays the target content based on the recommendation prediction result.

[0047] The above-mentioned terminal can be a mobile phone, tablet computer, desktop computer, portable notebook computer, smart TV, vehicle terminal, smart home device and other terminal devices in various forms, and the embodiments of the present application are not limited to this.

[0048] It is worth noting that the above-mentioned servers can be independent physical servers, or they can be server clusters or distributed systems composed of multiple physical servers. They can also be cloud servers that provide basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), as well as big data and artificial intelligence platforms.

[0049] In some embodiments, the above-mentioned server can also be implemented as a node in a blockchain system.

[0050] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, account attribute data, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions.

[0051] In combination with the above-mentioned noun introduction and application scenarios, the content recommendation method provided by this application is described. This method can be executed by a server or a terminal, or by both a server and a terminal. In the embodiment of this application, the method is described as being executed by a server. Figure 3 As shown, the method includes the following steps.

[0052] Step 310: Acquire target account attribute data of the target object account and target content attribute data of the target content.

[0053] Illustratively, the target account attribute data is used to represent different types of attribute data of the target object account.

[0054] Optionally, the target account attribute data includes account profile data corresponding to the target object account, wherein the account profile data includes account information filled in by the user during the application process for the target object account, such as: account name, account gender, account age, account preferences, etc. At least one of the information. It is worth noting that the account profile data involved in this embodiment is data authorized by the user for public use.

[0055] Optionally, the target account attribute data includes device data corresponding to the target object account, such as: a device model corresponding to the target object account, an application version corresponding to the target object account, etc.

[0056] Optionally, the target account attribute data includes historical content attribute data of the target object account, such as at least one of the target object account's content browsing records, content click records (click-through rate for specified types of content), content browsing time, etc. within a historical time period.

[0057] In some embodiments, the target content refers to recommended content to be recommended to the target object account, and is used to be displayed to the target user in the terminal to which the target object account belongs.

[0058] Illustratively, the target content attribute data is used to represent different types of attribute data of the target content.

[0059] Optionally, the target content attribute data includes content information data of the target content. For example, taking the target content as music, the target content attribute data includes at least one of the data such as the music name, music type (such as folk, rock, lyrical, light music, etc.), singer, composer, and release year.

[0060] Optionally, the target content attribute data includes historical recommendation data of the target content, wherein the historical recommendation data includes at least one of historical click-through rate, historical browsing time, historical recommendation times, historical recommendation accounts, and the like.

[0061] Among them, the historical click-through rate refers to the click-through rate corresponding to the target content within the historical time period; the historical browsing time refers to the browsing time of the target content within the historical time period; the historical recommendation times refers to the number of times the target content is recommended to the target account (including the target object account or other object accounts) within the historical time period; the historical recommendation account refers to the corresponding object account to which the target content is recommended within the historical time period.

[0062] In some embodiments, the target object account is a corresponding object account in a target application running in the terminal in advance.

[0063] Optionally, the target content is implemented as a single content, or the target content is implemented as at least one candidate content in the candidate content list, that is, the target content is implemented as multiple contents, which is not limited.

[0064] Step 320 : extracting the account feature representation corresponding to the target account attribute data, and extracting the content feature representation corresponding to the target content attribute data.

[0065] Schematically, extracting the account feature representation corresponding to the target account attribute data refers to projecting the target account attribute data into a low-dimensional representation space and representing it with a dense account representation vector (ie, an account embedding vector), which serves as the account feature representation.

[0066] Schematically, extracting the content feature representation corresponding to the target content attribute data refers to projecting the target content attribute data into a low-dimensional representation space and representing it with a dense content representation vector (ie, a content embedding vector), which serves as the content feature representation.

[0067] Optionally, the above two low-dimensional representation spaces can be implemented as representation spaces of the same dimension; or, the above two low-dimensional representation spaces can be implemented as representation spaces of different dimensions, which is not limited.

[0068] Optionally, the account feature representation and the content feature representation are extracted simultaneously; or, the account feature representation and the content feature representation are extracted successively according to a specified order.

[0069] Step 330 : Acquire a multi-order interactive feature representation of the interaction between the account feature representation and the content feature representation based on the model parameter search space.

[0070] The model parameter search space includes architectural parameters corresponding to the multi-order interaction relationships between account attribute data and content attribute data. The architectural parameters are used to represent the effectiveness of the multi-order interaction relationships in the recommendation prediction process.

[0071] Schematically, a multi-order interactive feature representation refers to an interactive feature representation obtained by performing feature interaction between at least one account feature representation and at least one content feature representation. For example, the account feature representation 1 corresponding to the account age data, the account feature representation 2 corresponding to the account preference data, and the content feature representation 1 corresponding to the content name data are subjected to feature interaction to obtain a multi-order interactive feature representation which is implemented as a three-order feature representation. That is, the feature order of the multi-order interactive feature representation corresponds to the number of features that are interacted.

[0072] Illustratively, the model parameter search space stores architectural parameters corresponding to an interactive relationship between at least one account attribute data (including target account attribute data) and at least one content attribute data (including target content attribute data).

[0073] The interactive relationship includes feature fusion between feature representations; or, the interactive relationship includes feature splicing between feature representations.

[0074] Schematically, a single multi-order interaction relationship corresponds to one architectural parameter, that is, a single multi-order interaction relationship corresponds to an architectural parameter one-to-one.

[0075] In some embodiments, the architecture parameter refers to the weight value corresponding to the multi-order interaction feature, which indicates the importance of the multi-order interaction feature. That is, the higher the parameter value of the architecture parameter, the higher the importance of the corresponding multi-order interaction relationship in the recommendation prediction process.

[0076] The recommendation prediction refers to the predicted probability of recommending the target content to the target account.

[0077] Optionally, the multi-order interaction feature representation is obtained by at least one of the following methods:

[0078] 1. The model parameter search space stores architectural parameters corresponding to various multi-order interaction relationships. Based on the multi-order interaction relationships between the account feature representation and the content feature representation, the architectural parameters corresponding to the account feature representation and the content feature representation are determined. The account feature representation and the content feature representation are weightedly combined based on the architectural parameters to obtain the combined result as the multi-order interaction feature representation.

[0079] 2. The model parameter search space stores multiple architectural parameters. Each architectural parameter is annotated with a multi-order interaction relationship. Each architectural parameter corresponds to a multi-order interaction feature representation. Once the architectural parameter corresponding to the multi-order interaction between the account feature representation and the content feature representation is determined, the multi-order interaction feature representation corresponding to that architectural parameter is obtained.

[0080] 3. Multiple interaction feature representations are stored in the model parameter search space. Each interaction feature representation is annotated with a corresponding interaction relationship. The corresponding interaction feature representation in the model parameter search space is determined based on the interaction relationship between the account feature representation and the content feature representation as a multi-order interaction feature representation.

[0081] It is worth noting that the above-mentioned method for obtaining the multi-order interaction feature representation is only an illustrative example and is not limited thereto.

[0082] Optionally, the interaction between the account feature representation and the content feature representation corresponds to one multi-order interaction feature representation; or, corresponds to multiple different multi-order interaction feature representations, which is not limited to this.

[0083] Optionally, the model parameter search space is a vector search space pre-set by the server; or, the model parameter search space is a vector search space automatically generated based on the multi-order interaction relationship between the acquired account attribute data and content attribute data; or, the model parameter search space is a vector search space generated after the user selects the specified account attribute data and content attribute data for interaction according to actual needs, and there is no limitation on this.

[0084] Step 340 : Perform predictive analysis on the multi-order interaction feature representation to obtain the recommendation probability between the target object account and the target content.

[0085] Optionally, the predictive analysis refers to determining a probability value of recommending the target content to the target object account based on multi-order interaction features.

[0086] Optionally, predictive analysis refers to determining the probability value of the target object account clicking on the target content based on multi-order interaction features as the recommendation probability, that is, the higher the probability value of the target click rate, the greater the recommendation probability.

[0087] Optionally, the predictive analysis includes at least one of the following analysis methods:

[0088] 1. A click-through rate prediction model is pre-set. Multi-order interaction feature representations are input into the click-through rate prediction model, and the output result is used as the recommendation probability between the target account and the target content.

[0089] 2. Perform a matching analysis based on the multi-order interaction features and the target account features corresponding to the target object account. As a predictive analysis, the corresponding matching results are used to represent the recommendation probability between the target object account and the target content.

[0090] It is worth noting that the above-mentioned analysis method for predictive analysis is only an illustrative example and is not limited to this embodiment of the present application.

[0091] Step 350: Recommend content to the target account based on the recommendation probability.

[0092] Illustratively, content recommendation refers to pushing content information corresponding to target content to a target object account.

[0093] Optionally, the content recommendation method includes at least one of the following methods:

[0094] 1. Preset a recommendation probability threshold and recommend target content that has reached the recommendation probability threshold to the target account;

[0095] 2. Sort the target content based on its corresponding recommendation probability, and recommend the target content with a specified proportion in the ranking to the target account;

[0096] 3. Sort the target content based on its corresponding recommendation probability and recommend it to the target account in the order of the ranking results. That is, the higher the ranking of the target content, the earlier it will be recommended to the target account.

[0097] It is worth noting that the above-mentioned content recommendation method is only an illustrative example and is not limited to this embodiment of the present application.

[0098] In summary, the content recommendation method provided in the embodiment of the present application represents the effectiveness of the relationship in the recommendation prediction process through the architecture parameters corresponding to the multi-order interaction relationship between the account attribute data and the content attribute data stored in the model parameter search space. The multi-order interaction feature representation between the account feature representation of the target object account and the content feature representation of the target content is obtained according to the model parameter search space, and a predictive analysis is performed on them to determine the recommendation probability of the target content. A more effective multi-order interaction feature representation can be obtained based on the architecture parameters, and a predictive analysis can be performed on it, so that the accuracy of the prediction probability is higher. In addition, storing architecture parameters in the search space instead of storing deep sparse networks can reduce the storage amount, thereby improving the efficiency of content recommendation.

[0099] In an optional embodiment, the multi-order interaction feature representation is determined by the architecture parameters, for example, please refer to Figure 4 , which shows a flow chart of a content recommendation method provided by an exemplary embodiment of the present application, that is, step 330 includes steps 331 and 332, as shown in FIG. Figure 4 As shown, the method includes the following steps:

[0100] Step 310: Acquire target account attribute data of the target object account and target content attribute data of the target content.

[0101] Optionally, the target account attribute data includes at least one of account information data, device data corresponding to the account, and historical content attribute data.

[0102] Optionally, the target content attribute data includes at least one of content attribute data, historical recommendation data, and the like.

[0103] In some embodiments, the target content is content that has not been recommended to the target object account; or the target content is content that has been recommended to the target object account in a historical time period and is newly recommended to the target object account in the current time period.

[0104] Optionally, the target account attribute data and the target content attribute data are obtained simultaneously; or, first obtain a content candidate list corresponding to the target content, and determine the object account for which content recommendation is to be made based on the content candidate list as the target object account; or, first determine the target object account for which content recommendation is to be made, and then determine a candidate content list to be recommended from the content set based on the target object account, and obtain the target content attribute data corresponding to the target content to be recommended based on the candidate content list. There is no limitation on this.

[0105] Step 320 : extracting the account feature representation corresponding to the target account attribute data, and extracting the content feature representation corresponding to the target content attribute data.

[0106] In some embodiments, a feature extraction model is preset, and the target account attribute data and the target content attribute data are simultaneously input into the feature extraction model, and the feature extraction result corresponding to the target account attribute data is output as the account feature representation, and the feature extraction result corresponding to the target content attribute data is output as the content feature representation. That is, the target account attribute data and the target content attribute data are simultaneously mapped to the same low-dimensional representation space, and the account embedding vector corresponding to the target object account is output as the account feature representation, and the content embedding vector corresponding to the target content is output as the content feature representation.

[0107] In some embodiments, an account feature extraction model and a content feature extraction model are predefined for the target account and target content, respectively. The target account attribute data corresponding to the target account is input into the account feature extraction model, and the model output serves as the account feature representation corresponding to the target account attribute data. The target content attribute data corresponding to the target content is input into the content feature extraction model, and the model output serves as the content feature representation corresponding to the target content attribute data. Specifically, the target account attribute data and target content attribute data are input into different low-dimensional representation spaces, and the corresponding account embedding vector and content embedding vector are output, respectively. The account embedding vector and the content embedding vector may be of the same or different dimensions, without limitation.

[0108] Step 331: Obtain architecture parameters corresponding to multi-order interaction relationships in the model parameter search space.

[0109] The multi-order interaction relationship refers to the interaction relationship between at least one account feature representation and at least one content feature representation.

[0110] Illustratively, an account feature representation is used as a first-order feature, and a content feature representation is used as a first-order feature. When the account feature representation and the content feature representation interact, a second-order feature is obtained. Similarly, third-order features, fourth-order features, and so on can be obtained. The order corresponding to the highest-order interaction feature in the current multi-order interaction relationship is fixed or can be set according to actual needs, and this is not limited.

[0111] Optionally, during the feature interaction, the same feature representation exists (eg, two account feature representations of the same type interact); or, during the feature interaction, all feature representations belong to different types of feature representations, which is not limited.

[0112] In some embodiments, each multi-order interaction relationship corresponds to an architectural parameter, which is used to represent the effectiveness of the multi-order interaction relationship in the subsequent recommendation prediction process. For example, when the content is implemented as commodity content, the content attribute data corresponding to the content includes content attribute data such as historical sales volume, historical collection volume, historical click-through rate, and age distribution of buyers corresponding to the commodity. The account attribute data corresponding to the object account includes account age, account gender, and account historical purchase records. The account feature representation a corresponding to the account age, the account feature representation b corresponding to the account historical purchase records, and the content feature representation A corresponding to the historical sales volume corresponding to the commodity and the content feature representation B corresponding to the age distribution of buyers are interacted as a multi-order interaction relationship. That is, the current multi-order interaction relationship includes feature interaction between the account feature representation a, the account feature representation b, the content feature representation A, and the content feature representation B.

[0113] Among them, when the value of the architecture parameter is larger, it indicates that the effectiveness of the multi-order interaction relationship is higher, so a higher weight should be assigned in the subsequent recommendation prediction process.

[0114] Optionally, the multi-order interaction relationship includes different interaction relationships corresponding to the same order, that is, interaction relationships between feature representations of the same number but different types; or, different interaction relationships corresponding to different orders, such as: second-order interaction relationships (two feature representations interacting), third-order interaction relationships (three feature representations interacting), etc.

[0115] In this embodiment, the architecture parameters are pre-trained parameters. The training process of the architecture parameters will be described in detail in subsequent embodiments and will not be described in this embodiment for the time being.

[0116] Optionally, the same architectural parameters may exist for the multi-order interaction relationships; or, each multi-order interaction relationship may correspond to a different architectural parameter, which is not limited.

[0117] Illustratively, a search is performed in the model parameter search space based on the multi-order interaction relationship corresponding to the account feature representation and the content feature representation to determine the architecture parameters corresponding to the multi-order interaction relationship.

[0118] Step 332 : performing weighted combination on the account feature representation of the target account attribute data and the content feature representation of the target content attribute data based on the architecture parameters in the model parameter search space to obtain a multi-order interactive feature representation.

[0119] Illustratively, based on the architectural parameters corresponding to the multi-order interaction relationship in the model parameter search space, the account feature representation and the content feature representation are weightedly combined, and the weighted combination result is used as the multi-order interaction feature representation corresponding to the account feature representation and the content feature representation. For example, a multi-order interaction relationship refers to the interaction between account feature representation A and content feature representation B. If the architectural parameter corresponding to the multi-order interaction relationship is determined to be 0.6 in the model parameter search space, then the multi-order interaction feature representation = 0.6 * (account feature representation A + content feature representation). In other words, the architectural parameters are used to represent the importance of the multi-order interaction feature identifier.

[0120] Step 340 : Perform predictive analysis on the multi-order interaction feature representation to obtain the recommendation probability between the target object account and the target content.

[0121] Schematically, a click-through rate prediction model is preset, and the multi-order interaction feature representation is input into the click-through rate prediction model. The click-through rate of the target object account for the target content is predicted and analyzed, and the result is output as the probability value of the target object account clicking the target content, as the recommendation probability between the target object account and the target content.

[0122] Step 350: Recommend content to the target account based on the recommendation probability.

[0123] The content recommendation in step 350 has been described in detail in the above step 350 and will not be repeated here.

[0124] In summary, the content recommendation method provided in the embodiment of the present application represents the effectiveness of the relationship in the recommendation prediction process through the architecture parameters corresponding to the multi-order interaction relationship between the account attribute data and the content attribute data stored in the model parameter search space. The multi-order interaction feature representation between the account feature representation of the target object account and the content feature representation of the target content is obtained according to the model parameter search space, and a predictive analysis is performed on them to determine the recommendation probability of the target content. A more effective multi-order interaction feature representation can be obtained based on the architecture parameters, and a predictive analysis can be performed on it, so that the accuracy of the prediction probability is higher. In addition, storing architecture parameters in the search space instead of storing deep sparse networks can reduce the storage amount, thereby improving the efficiency of content recommendation.

[0125] In this embodiment, by obtaining the architectural parameters corresponding to the multi-order interaction relationship in the model adoption number search space, and performing weighted combination of the account feature representation and the content feature representation according to the architectural parameters to obtain the multi-order interaction feature representation, the weight corresponding to the multi-order interaction feature representation corresponding to the multi-order interaction relationship with a lower degree of effectiveness can be reduced, and the weight corresponding to the multi-order interaction relationship with a higher degree of effectiveness can be increased, thereby improving the recommendation accuracy of subsequent recommendation predictions and reducing the negative impact of less important interaction relationships on recommendation predictions.

[0126] In an optional embodiment, there is a corresponding training process for the architecture parameters. For example, please refer to Figure 5 , which shows a flow chart of a content recommendation method provided by an exemplary embodiment of the present application, such as Figure 5 As shown, the method includes the following steps.

[0127] Step 510: Obtain a sample data set.

[0128] The sample data set includes sample account attribute data of the sample object account and sample content attribute data of the sample content. The sample object account and the sample content are annotated with historical interaction events.

[0129] Optionally, the sample dataset is obtained from data stored in a local file; or, the sample dataset is obtained from a public open source dataset, which is not limited.

[0130] Illustratively, the sample account attribute data is used to represent different types of account attribute data.

[0131] Optionally, the sample account attribute data includes sample account profile data corresponding to the sample object account, wherein the sample account profile data includes publicly available account information corresponding to the sample object account during the account application process, such as at least one of: account name, account gender, account age, account preferences, etc.

[0132] Optionally, the sample account attribute data includes device data corresponding to the sample object account, such as the device model corresponding to the sample object account, the application version corresponding to the sample object account, and the like.

[0133] Optionally, the sample account attribute data includes historical content attribute data of the sample object account, such as at least one of content browsing records, content click records, content browsing duration, etc. of the sample object account in a historical time period.

[0134] Illustratively, the sample content attribute data is used to represent different types of content attribute data corresponding to the sample content.

[0135] Optionally, the sample content attribute data includes content information data. For example, taking the sample content as a commodity, the sample content data includes at least one of the following data: commodity name, commodity type (such as daily necessities, food, furniture, etc.), commodity price, commodity origin, etc.

[0136] Optionally, the sample content data includes historical recommendation data of the sample content. For example, taking a product as an example, the historical recommendation data includes the product's historical sales volume, historical recommendation times (the number of times it was recommended within a historical time period), historical recommendation accounts (the account types corresponding to the accounts recommended within a historical time period), etc.

[0137] Illustratively, the historical interaction events marked between the sample object account and the sample content refer to the corresponding events in the interaction process between the sample object account and the sample content within the historical time period. For example, if the sample content is a product and the sample object account has purchased the product within the historical time period, the consumption record generated by the purchase behavior will be regarded as a historical interaction event.

[0138] In some embodiments, the sample data set includes a training set and a test set, wherein the data in the training set is used to train the sample architecture parameters, the sample extraction model and the sample click-through rate prediction model; the data in the test set is used to determine the offline training effects corresponding to the sample architecture parameters, the sample extraction model and the sample click-through rate prediction model respectively.

[0139] Optionally, the test set and the training set contain the same data; or, the test set and the training set are implemented as different data sets, which is not limited.

[0140] Step 520 : extracting the sample account feature representation corresponding to the sample account attribute data, and extracting the sample content feature representation corresponding to the sample content attribute data.

[0141] Schematically, a sample extraction model is preset, and the sample extraction model is used to extract features from sample account attribute data and sample content attribute data.

[0142] In this embodiment, the sample account attribute data and sample content attribute data in the training set are input into the sample extraction model, and the embedding parameters (Embedding Parameter) corresponding to the sample account attribute data are output as the sample account feature representation, and the embedding parameters corresponding to the sample content attribute data are output as the sample content feature representation.

[0143] In this embodiment, the total number of feature fields corresponding to the sample account feature representation and the sample content feature representation is N.

[0144] Step 530 : Acquire a multi-order interactive sample representation of the interaction between the sample account feature representation and the sample content feature representation based on the sample parameter search space.

[0145] The sample parameter search space includes sample architecture parameters corresponding to the multi-order interaction relationships between the sample account attribute data and the sample content attribute data.

[0146] Schematically, a multi-order interaction relationship refers to the interaction between at least one sample account feature representation and at least one sample content feature representation in the training set, resulting in a multi-order feature representation. For example, if the total number of feature fields is N and the highest order in the interaction relationship is O, then the set size of the multi-order feature representation set corresponding to the multi-order interaction relationship between the sample account feature representation and the sample content feature representation is Typically, N is much larger than O, so the size of the set is approximately equal to N raised to the power of O. The set is implemented as a discrete set.

[0147] Schematically, if the set corresponding to the above multi-order feature representation is stored as a search space, it is difficult to select a suitable multi-order feature representation for the search space. Therefore, a sample parameter search space is preset, and the sample parameter search space stores the sample architecture parameters corresponding to the multi-order interaction relationships. That is, the sample architecture parameters correspond to the multi-order feature representation one-to-one. Therefore, the corresponding sample architecture parameters in the sample parameter search space can be realized as a length of A vector is used to indicate the effectiveness of each multi-order feature representation in the recommendation prediction process. The larger the value of a vector position, the higher the effectiveness of the multi-order feature representation corresponding to the vector position.

[0148] In this embodiment, the definition of the sample parameter search space can be implemented as a differentiable search method, such as the Differentiable Architecture Search (Darts algorithm). The Darts algorithm trains the sample architecture parameters by calculating gradients and backpropagation to achieve architecture parameter updates.

[0149] In this embodiment, since the vector dimension of the above sample architecture parameters is too high in actual application, a large amount of computing resources is required to search for the sample architecture parameters corresponding to the multi-order interaction relationship in the sample parameter search space, which not only slows down the search process but also reduces the search adaptability of the sample architecture parameters. Therefore, the original vector length is set to The initial architecture parameters are tensor-decomposed to obtain tensor product results corresponding to the initial architecture parameters as the sample architecture parameters. That is, a sample parameter search space is obtained, where the sample parameter search space includes the initial architecture parameters; the initial architecture parameters are tensor-decomposed to obtain tensor product results corresponding to the initial architecture parameters as the sample architecture parameters.

[0150] Among them, the sample parameter search space stores a length of Tensor decomposition refers to reducing the initial architecture parameters to architecture parameters of size O×N as sample architecture parameters. The sample parameter search space corresponding to the current sample architecture parameters is implemented as a low-rank approximation search space, which is used to reduce the number of parameters corresponding to the architecture parameters.

[0151] In an optional case, the above-mentioned tensor decomposition process is implemented before determining the multi-order interactive sample representation that interacts between the sample account feature representation and the sample content feature representation, that is, the multi-order interactive sample representation is currently determined using the sample architecture parameters after tensor decomposition.

[0152] In an optional case, the above-mentioned tensor decomposition process is implemented by performing tensor decomposition on the multi-order interactive sample representation that interacts between the sample account feature representation and the sample content feature representation after determining the multi-order interactive sample representation, that is, after determining the initial architecture parameters corresponding to the multi-order interactive sample representation, to obtain the sample architecture parameters.

[0153] Schematically, the multi-order interactive sample representation is obtained by determining the sample architecture parameters corresponding to the interaction between the sample account feature representation and the sample content feature representation in the sample parameter search space, and then weightedly combining the sample account feature representation and the sample content feature representation according to the sample architecture parameters.

[0154] Step 540 : Perform prediction analysis on the multi-order interaction sample representation to obtain the recommendation prediction probability between the sample object account and the sample content.

[0155] In some embodiments, the multi-order interaction sample representation is input into the sample click rate prediction model, and the recommendation prediction probability of the sample object account corresponding to the sample content is obtained as output.

[0156] Schematically, a sample click rate prediction model is preset to perform prediction analysis on the probability of a sample object account in a training set clicking on a sample content, and the result is output as the recommendation prediction probability of the sample object account corresponding to the sample content.

[0157] Step 550 : Based on the difference between the recommendation prediction probability and the historical interaction events, the sample architecture parameters are trained to obtain the architecture parameters.

[0158] For example, in the CTR prediction task, low-order feature representations are more important than high-order feature representations in the aforementioned preset sample CTR prediction model. This means that the sample CTR prediction model must first effectively model low-order feature representations before it can model higher-order feature representations. Therefore, during the training of the sample architecture parameters, progressive training is required. Progressive training involves adjusting parameters step by step, from low to high order.

[0159] In some embodiments, a first learning rate is obtained; and based on the difference between the recommendation prediction probability and the historical interaction events, and the first learning rate, the sample architecture parameters are progressively adjusted to obtain the architecture parameters.

[0160] Illustratively, the first learning rate is a hyperparameter that can be manually adjusted by the user according to training needs, and is used to control the adjustment amount during the process of progressively adjusting the sample architecture parameters to ultimately obtain the architecture parameters. In this embodiment, the first learning rate is implemented as r1.

[0161] Among them, the progressive adjustment can be implemented by adjusting the sample architecture parameters corresponding to each order of the multi-order interaction relationship. After the sample architecture parameters corresponding to the current order are adjusted, the sample architecture parameters corresponding to the next order are adjusted. That is, the sample architecture parameters include n-order sample architecture parameters, where n is a positive integer; based on the difference between the recommended prediction probability and the historical interaction events, and the first learning rate, the sample architecture parameters of other orders except the i-th order sample architecture parameters are fixed, and the i-th order sample architecture parameters are gradient-adjusted, where 0<i<n; in response to the difference between the recommended prediction probability and the historical interaction events during the training process of the i-th order sample architecture parameters meeting the preset convergence condition, the sample architecture parameters of other orders except the i+1-th order sample architecture parameters are fixed, and the i+1-th order sample architecture parameters are gradient-adjusted.

[0162] In this embodiment, starting from order 1, the order corresponding to the sample architecture parameter is increased successively until the highest order n. Taking the training of the i-th order sample architecture parameter as an example, during the training process of the i-th order sample architecture parameter, the corresponding loss function is determined by the difference between the recommended prediction probability corresponding to the data on the validation set and the historical interaction events, and the parameter update amount corresponding to the i-th order sample architecture parameter is calculated according to the Darts algorithm. The gradient value corresponding to the i-th order sample architecture parameter is calculated according to the parameter update amount and the first learning rate, which is used to adjust the i-th order sample architecture parameter.

[0163] When the difference between the recommended prediction probability and the historical interaction events corresponding to the late stage of the i-th order sample architecture parameter training meets the preset convergence condition, that is, the parameter error corresponding to the current i-th order sample architecture parameter has reached convergence, the training process of the i-th order sample architecture parameter is completed, and the training of the i+1-th order sample architecture parameter begins.

[0164] Among them, during the training process for the i-th order sample architecture parameters, it is necessary to fix the sample architecture parameters of other orders except the i-th order sample architecture parameters. When the training of the i+1-th order sample architecture parameters is completed, it is necessary to fix the sample architecture parameters of other orders except the i+1-th order sample architecture parameters. That is, each order sample architecture parameter is trained independently until the n-th order sample architecture parameter is trained. In the training process for the 1st to n-1th order sample architecture parameters, when the loss function on the training set and the loss function on the validation set both meet the preset convergence conditions, the training process of the higher-order sample architecture parameters can be entered. In the training process for the n-th order sample architecture parameters, when the loss function on the training set and the loss function on the validation set both meet the preset convergence conditions, it is considered that the training of the n-th order sample architecture parameters is completed and the n-th order architecture parameters are obtained.

[0165] In some embodiments, after the n-order sample architecture parameters obtained by the above-mentioned progressive adjustment are completed, there is still a partial retention process, that is, based on the difference between the recommended prediction probability and the historical interaction events, the sample architecture parameters are trained to obtain candidate architecture parameters; the candidate architecture parameters are sorted according to the parameter weights to obtain the parameter sorting results corresponding to the candidate architecture parameters; and the candidate architecture parameters with a specified proportion in the parameter sorting results are used as architecture parameters.

[0166] In this embodiment, the candidate architecture parameters refer to the n-order sample architecture parameters obtained through the above-mentioned progressive adjustment. The candidate architecture parameters are sorted, and the sorting method can be implemented by sorting according to the vector value size of the candidate architecture parameters (that is, the parameter weight). The candidate architecture parameters with higher rankings have higher corresponding effectiveness. Therefore, in the parameter sorting results obtained by sorting the candidate architecture parameters, a specified proportion of candidate architecture parameters are selected as architecture parameters, such as: selecting the top R% of candidate architecture parameters as the final architecture parameters.

[0167] Optionally, the user can customize the ratio value to perform a grid search on the candidate architecture parameters to determine the architecture parameters corresponding to the specified ratio; or, a preset parameter selection model is used to input all candidate architecture parameters into the parameter selection model to output the architecture parameters of the specified ratio.

[0168] The multi-order interaction sample representation corresponding to the final architecture parameters is input into the sample click-through rate prediction model, and the corresponding recommendation prediction probability is output. A probability loss function is determined based on the difference between the recommendation prediction probability and historical interaction events. The parameters of the sample click-through rate prediction model are adjusted based on the probability loss function, which is implemented as a retraining process for the sample click-through rate prediction model. Finally, the retrained sample click-through rate prediction model is tested using the test data in the test set. If the test result meets the preset test result, the sample click-through rate prediction model is used as the click-through rate prediction model. In other words, based on the difference between the recommendation prediction probability and historical interaction events, the sample click-through rate prediction model is trained to obtain a click-through rate prediction model.

[0169] In addition, the embodiment of the present application also includes a sample extraction model for extracting features from sample account attribute data and sample content attribute data.

[0170] During the training of the sample extraction model, a second learning rate (r2) is set, and the training parameters of the sample extraction model are adjusted based on the loss function on the training set to obtain a feature extraction model. In other words, the second learning rate is obtained; based on the difference between the recommendation prediction probability and the historical interaction events, and the second learning rate, the model parameters of the sample extraction model are gradient-adjusted to obtain a feature extraction model.

[0171] In summary, the content recommendation method provided in the embodiment of the present application represents the effectiveness of the relationship in the recommendation prediction process through the architecture parameters corresponding to the multi-order interaction relationship between the account attribute data and the content attribute data stored in the model parameter search space. The multi-order interaction feature representation between the account feature representation of the target object account and the content feature representation of the target content is obtained according to the model parameter search space, and a predictive analysis is performed on them to determine the recommendation probability of the target content. A more effective multi-order interaction feature representation can be obtained based on the architecture parameters, and a predictive analysis can be performed on it, so that the accuracy of the prediction probability is higher. In addition, storing architecture parameters in the search space instead of storing deep sparse networks can reduce the storage amount, thereby improving the efficiency of content recommendation.

[0172] In this embodiment, during the training process, the effectiveness of multi-order feature representation in the recommendation prediction process is represented by using sample parameter search space and sample architecture parameters, which can improve the search efficiency. In addition, the sample architecture parameters are obtained by performing tensor decomposition on the initial architecture parameters, which can reduce the storage capacity of the search space and improve the space storage efficiency.

[0173] In this embodiment, by gradually adjusting the n-order sample architecture parameters, adjustments can be made layer by layer starting from low-order features, thereby making the parameter accuracy of the sample architecture parameters higher.

[0174] In this embodiment, selecting a specified proportion of candidate architecture parameters as the final architecture parameters can reduce the negative impact of less effective candidate architecture parameters on the training of the sample click-through rate prediction model, thereby improving the training accuracy of the sample click-through rate prediction model.

[0175] The following describes a content recommendation system according to an embodiment of the present application.

[0176] The content recommendation system provided in the embodiments of this application generally consists of four parts, including a data collection system, a feature processing system, a recall system, and a click-through rate prediction system. The click-through rate prediction model proposed in the embodiments of this application acts on the fourth part, that is, the click-through rate prediction system.

[0177] Data Collection System

[0178] In this embodiment, the data collection system collects and updates in real time the account behavior data, account attribute data, historical content attribute data, and other data of the target account.

[0179] The account behavior data includes historical account behaviors corresponding to the target object account. For example, if the target object account is a tourist account, the account behavior data may be a list of historical travel destinations corresponding to the tourist account.

[0180] Feature processing system

[0181] The feature processing system performs certain feature processing and construction based on the target account attribute data collected by the data collection system, removes data with low relevance or errors, and obtains input features that can be used for subsequent recall and click-through rate prediction.

[0182] Recall System

[0183] The recall system is used to determine the target content to be recommended. Since it only roughly generates a candidate content recommendation list (usually dozens to hundreds) for input into the subsequent click-through rate prediction system, the recall system is generally also called a coarse ranking system.

[0184] In this embodiment, the input to the click rate prediction system includes: target content set C to be recommended, target account attribute data U and target content attribute data P, wherein the target account attribute data U and the target content attribute data P can be implemented as attribute tables respectively.

[0185] Click-through rate prediction system

[0186] After obtaining the target content attribute data from the recall system, the CTR prediction model calculates the target content's predicted click-through rate, predicted conversion rate, and other parameters. The predicted recommendations are then ranked based on their probability values, and the top ranking results (typically ten to several dozen) are recommended to the target account. Therefore, the CTR prediction system is also known as a refined ranking system.

[0187] In a feasible scenario, the content recommendation system will also deploy a re-ranking system after the click-through rate prediction system in the fourth part. The system can re-rank the sorting results of the click-through rate prediction system based on the content strategy of the current target content and the recommendation needs of the content publisher. For example, for the diversified recommendation strategy, the target content will be rearranged according to its content category to ensure that the content list recommended to the target account covers as diverse recommended content as possible, thereby improving the category diversity of the recommendation list, thereby improving the user experience and avoiding redundant and highly similar recommended content.

[0188] In this embodiment, the click-through rate prediction system includes multiple parts, mainly including an offline search part, a model retraining part, and an online service part.

[0189] Among them, the offline search part is based on the low-rank approximation model parameter search space and progressive search strategy to obtain the architectural parameters corresponding to the multi-order interaction relationship.

[0190] In the model retraining part, by retaining the candidate architecture parameters with high importance (that is, the candidate architecture parameters in a specified proportion), building the corresponding feature interaction layer, and training all the embedding vectors involved in the sample click-through rate prediction model, we can obtain the representation embedding vectors corresponding to all feature representations and obtain the final click-through rate prediction model.

[0191] In the online service part, the structure and parameters of the entire click-through rate prediction model remain unchanged, only the real-time features on the product side are changed. Since feature calculation is generally very fast, while updating model parameters takes more time, the approach of using dynamic features in static models can achieve recommendation updates at the second level.

[0192] In summary, the content recommendation method provided in the embodiment of the present application represents the effectiveness of the relationship in the recommendation prediction process through the architecture parameters corresponding to the multi-order interaction relationship between the account attribute data and the content attribute data stored in the model parameter search space. The multi-order interaction feature representation between the account feature representation of the target object account and the content feature representation of the target content is obtained according to the model parameter search space, and a predictive analysis is performed on them to determine the recommendation probability of the target content. A more effective multi-order interaction feature representation can be obtained based on the architecture parameters, and a predictive analysis can be performed on it, so that the accuracy of the prediction probability is higher. In addition, storing architecture parameters in the search space instead of storing deep sparse networks can reduce the storage amount, thereby improving the efficiency of content recommendation.

[0193] For illustration, please refer to Figure 6 , which shows a schematic diagram of a content recommendation method provided by an exemplary embodiment of the present application, such as Figure 6 As shown, the method includes the following parts:

[0194] Raw data (Sparse Features) 610.

[0195] Illustratively, a click rate prediction task is established to obtain target account attribute data corresponding to the target object account and target content attribute data corresponding to the target content. The target content is implemented as a target product as an example for illustration.

[0196] In this embodiment, the target product is determined from the list of products to be recommended. In addition, the product attribute data corresponding to the target product is determined, wherein the product attribute data includes at least one of the data such as historical sales volume of the product, product type, product name, product price, and distribution of product consumers.

[0197] In this embodiment, the target object account is implemented as a consumer account as an example for explanation, wherein the target account attribute data corresponding to the consumer account includes at least one of the following data: the consumer name, the consumer gender, the historical consumption records corresponding to the consumer account, the favorites corresponding to the consumer account, etc.

[0198] In this embodiment, the above-mentioned product attribute data and target account attribute data are used as original input data for the click rate prediction task, such as data A, data B, data C, etc.

[0199] Embedding layer 620.

[0200] Schematically, a feature extraction model is preset to extract features from target account attribute data and output an account feature representation corresponding to the target account attribute data, and product attribute data is input into the feature extraction model to output a product feature representation corresponding to the product attribute data.

[0201] The feature extraction model is implemented by mapping the target account attribute data and product attribute data into a low-dimensional representation space to obtain the corresponding account representation vector as the account feature representation and the corresponding product representation vector as the product feature representation.

[0202] Feature Interaction Layer 630.

[0203] Illustratively, a model parameter search space is constructed, wherein the model parameter search space includes architectural parameters corresponding to multi-order interaction relationships between account attribute data and product attribute data.

[0204] In this embodiment, the multi-order interaction relationship between account attribute data and product attribute data refers to feature interaction between at least one account feature representation and at least one product feature representation. For example, the feature interaction layer includes feature A, feature B, feature C, and feature D. These four features are implemented as four first-order features. If the first-order feature A and the first-order feature B interact with each other, a second-order feature AB is obtained. If the first-order feature B and the first-order feature D interact with each other, a second-order feature BD is obtained. If the first-order feature A, the first-order feature B, and the first-order feature C interact with each other, a third-order feature ABC is obtained. If the first-order feature B, the first-order feature C, and the first-order feature D interact with each other, a third-order feature BCD is obtained.

[0205] In this embodiment, each multi-order interaction relationship corresponds to an architecture parameter, which is used to indicate the effectiveness of the multi-order interaction relationship in the recommendation prediction process. That is, the higher the value of the architecture parameter, the more important the multi-order interaction relationship is.

[0206] In this embodiment, the model parameter search space includes architectural parameters corresponding to multi-order interaction relationships. Based on the multi-order interaction relationship between the account feature representation and the product feature representation, the model parameter search space is searched for architectural parameters corresponding to the multi-order interaction relationship. The account feature representation and the product feature representation are weightedly combined according to the architectural parameters, and the weighted combination result is used as the multi-order interaction feature representation corresponding to the account feature representation and the product feature representation.

[0207] Output layer (Output Layer) 640.

[0208] The multi-order interaction feature representation is converted into a scalar result through the output layer, wherein the conversion process can be implemented as a linear operation on the multi-order interaction feature representation; or, the conversion process can be implemented as a preset click-through rate prediction model, and the multi-order interaction feature representation is input into the click-through rate prediction model for click-through rate prediction analysis, and the output result is used as the click prediction probability of the target product, wherein the click-through rate prediction model is a machine learning model trained in advance.

[0209] Prediction Sore: 650.

[0210] In this embodiment, the predicted click probability corresponding to the target product is used as the output of the click-through rate prediction model. Based on this output, products are selected for recommendation to the target account. Optionally, a click probability threshold is preset, and only target products with a click probability that reaches the threshold are recommended to the consumer account. Alternatively, the predicted click probabilities corresponding to multiple target products are sorted from high to low, and a specified number of target products are selected based on the sorting results and recommended to the consumer account.

[0211] In this embodiment, during the training phase, the closer the click prediction probability is to 1, the higher the probability that the target product will be clicked by the consumer account.

[0212] The content recommendation method in this embodiment is applied to multiple open source datasets for verification, and the results of the click-through rate prediction model obtained by searching are compared with the existing models in multiple click-through rate prediction benchmarks, as shown in Table 1.

[0213] Table 1

[0214]

[0215]

[0216] As can be seen from Table 1, the click-through rate prediction model obtained through the model parameter search control has higher accuracy in most indicators and achieves better results in some indicators.

[0217] In summary, the content recommendation method provided in the embodiment of the present application represents the effectiveness of the relationship in the recommendation prediction process through the architecture parameters corresponding to the multi-order interaction relationship between the account attribute data and the content attribute data stored in the model parameter search space. The multi-order interaction feature representation between the account feature representation of the target object account and the content feature representation of the target content is obtained according to the model parameter search space, and a predictive analysis is performed on them to determine the recommendation probability of the target content. A more effective multi-order interaction feature representation can be obtained based on the architecture parameters, and a predictive analysis can be performed on it, so that the accuracy of the prediction probability is higher. In addition, storing architecture parameters in the search space instead of storing deep sparse networks can reduce the storage amount, thereby improving the efficiency of content recommendation.

[0218] Schematically, the application scenario of the content recommendation method provided by this application is described, taking the application to the song recommendation scenario as an example, combined with Figure 3 The corresponding steps are explained in detail.

[0219] Step 310: Acquire target account attribute data of the target object account and target content attribute data of the target content.

[0220] Indicatively, the target object account is the target audience account, and the target account attribute data includes at least one of the account information data of the target audience account, the device data corresponding to the target audience account (such as: listening device model, song playback program version, etc.), historical listening records, historical song collection records, historical search records, etc.

[0221] Illustratively, the target content is at least one candidate song for recommendation to the target audience account, and the target content attribute data includes at least one of the candidate song's title, singer, song type, release time, historical playback volume, historical collection volume, historical click-through rate, and historical recommendation times.

[0222] Step 320 : extracting the account feature representation corresponding to the target account attribute data, and extracting the content feature representation corresponding to the target content attribute data.

[0223] Schematically, the target account attribute data is projected into a low-dimensional representation space, and the account embedding vector is output as the account feature representation.

[0224] Schematically, the target content attribute data is projected into a low-dimensional representation space, and the output is a content embedding vector as a content feature representation.

[0225] Step 330 : Acquire a multi-order interactive feature representation of the interaction between the account feature representation and the content feature representation based on the model parameter search space.

[0226] The model parameter search space includes architectural parameters corresponding to the multi-order interaction relationships between account attribute data and content attribute data. The architectural parameters are used to represent the effectiveness of the multi-order interaction relationships in the recommendation prediction process.

[0227] Illustratively, the model parameter search space stores architectural parameters corresponding to the interaction between an account feature representation of at least one account attribute data and a content feature representation of at least one content attribute data. For example, the account feature representation corresponding to the historical listening history of target listener account A is subjected to feature interaction with the content feature representations corresponding to the song title, singer, and song type of candidate song a to obtain an interactive feature representation. Corresponding architectural parameters are then added to indicate the importance of the interactive feature representation in the subsequent recommendation prediction process. For example, if candidate song a is a popular song and candidate song b is an unpopular song, and candidate song a is more likely to be recommended than candidate song b, then the architectural parameters corresponding to the interactive feature representation obtained by the interaction between candidate song a and the target account attribute data corresponding to the target listener account should be greater than the architectural parameters corresponding to the interactive feature representation obtained by the interaction between candidate song b and the target account attribute data corresponding to the target listener account.

[0228] Illustratively, an automatic search is performed in the model parameter search space based on the interactive relationship between the account feature representation and the content feature representation to determine a multi-order interactive feature representation corresponding to the interaction between the account feature representation and the content feature representation. For example, if account feature representation 1 represents the listener's age, account feature representation 2 represents historical listening history, content feature representation 1 represents the song era, and content feature representation 2 represents the singer, then the model parameter search space is used to determine a multi-order interactive feature representation 1 corresponding to account feature representation 1, account feature representation, content feature representation 1, and content feature representation 2. This is used to represent the interactive feature representation resulting from the interaction of these four feature representations.

[0229] Step 340 : Perform predictive analysis on the multi-order interaction feature representation to obtain the recommendation probability between the target object account and the target content.

[0230] Schematically, a predictive analysis is performed based on the multi-order interactive feature representation to determine the probability value of the target audience account listening to the candidate song as the recommendation probability.

[0231] In this embodiment, the multi-order interaction feature representation is input into a preset click rate prediction model, and the result obtained is output as the recommendation probability between the target listener account and the candidate song.

[0232] Step 350: Recommend content to the target account based on the recommendation probability.

[0233] The recommendation probabilities corresponding to multiple candidate songs are sorted from high to low to obtain a recommendation sequence, and songs are recommended to the target listener account according to the arrangement order of the candidate songs on the recommendation sequence.

[0234] In summary, the content recommendation method provided in the embodiment of the present application represents the effectiveness of the relationship in the recommendation prediction process through the architecture parameters corresponding to the multi-order interaction relationship between the account attribute data and the content attribute data stored in the model parameter search space. The multi-order interaction feature representation between the account feature representation of the target object account and the content feature representation of the target content is obtained according to the model parameter search space, and a predictive analysis is performed on them to determine the recommendation probability of the target content. A more effective multi-order interaction feature representation can be obtained based on the architecture parameters, and a predictive analysis can be performed on it, so that the accuracy of the prediction probability is higher. In addition, storing architecture parameters in the search space instead of storing deep sparse networks can reduce the storage amount, thereby improving the efficiency of content recommendation.

[0235] Figure 7 is a structural block diagram of a content recommendation device provided by an exemplary embodiment of the present application. Figure 7 As shown, the device includes:

[0236] An acquisition module 710 is used to acquire target account attribute data of a target object account and target content attribute data of a target content;

[0237] An extraction module 720 is configured to extract an account feature representation corresponding to the target account attribute data, and to extract a content feature representation corresponding to the target content attribute data;

[0238] The acquisition module 710 is further configured to acquire, based on a model parameter search space, a multi-order interaction feature representation of the interaction between the account feature representation and the content feature representation, wherein the model parameter search space includes architectural parameters corresponding to the multi-order interaction relationships between the account attribute data and the content attribute data, the architectural parameters being used to represent the effectiveness of the interaction relationships in the recommendation prediction process;

[0239] An analysis module 730 is configured to perform predictive analysis on the multi-order interaction feature representation to obtain a recommendation probability between the target object account and the target content;

[0240] The recommendation module 740 is configured to recommend content to the target account based on the recommendation probability.

[0241] In an optional embodiment, the acquisition module 710 is further used to obtain the architectural parameters corresponding to the multi-order interaction relationship in the model parameter search space, where the multi-order interaction relationship refers to the interaction relationship between at least one account feature representation and at least one content feature representation; and based on the architectural parameters in the model parameter search space, the account feature representation of the target account attribute data and the content feature representation of the target content attribute data are weightedly combined to obtain the multi-order interaction feature representation.

[0242] In an optional embodiment, the acquisition module 710 is further configured to acquire a sample data set, wherein the sample data set includes sample account attribute data of a sample object account and sample content attribute data of a sample content, wherein the sample object account and the sample content are annotated with historical interaction events;

[0243] The extraction module 720 is further configured to extract the sample account feature representation corresponding to the sample account attribute data, and extract the sample content feature representation corresponding to the sample content attribute data;

[0244] The acquisition module 710 is further configured to acquire, based on a sample parameter search space, a multi-order interaction sample representation of the interaction between the sample account feature representation and the sample content feature representation, wherein the sample parameter search space includes sample architecture parameters corresponding to the multi-order interaction relationships between the sample account attribute data and the sample content attribute data;

[0245] The analysis module 730 is further configured to perform a prediction analysis on the multi-order interaction sample representation to obtain a recommendation prediction probability between the sample object account and the sample content;

[0246] The device further comprises:

[0247] The training module 750 is further configured to train the sample architecture parameters based on the difference between the recommendation prediction probability and the historical interaction events to obtain the architecture parameters.

[0248] In an optional embodiment, the acquisition module 710 is further configured to acquire a sample parameter search space, wherein the sample parameter search space includes initial architecture parameters;

[0249] The device further comprises:

[0250] The decomposition module 760 is configured to perform tensor decomposition on the initial architecture parameters to obtain tensor product results corresponding to the initial architecture parameters as the sample architecture parameters.

[0251] In an optional embodiment, the training module 710 includes:

[0252] An acquiring unit 711 is configured to acquire a first learning rate;

[0253] An adjusting unit 712 is configured to perform a gradient adjustment on the sample architecture parameters based on the difference between the recommendation prediction probability and the historical interaction event, and the first learning rate, to obtain the architecture parameters.

[0254] In an optional embodiment, the sample architecture parameters include n-order sample architecture parameters, where n is a positive integer;

[0255] The adjustment unit 712 is also used to fix the sample architecture parameters of other orders except the i-th order sample architecture parameters based on the difference between the recommended prediction probability and the historical interaction events, and the first learning rate, and perform gradient adjustment on the i-th order sample architecture parameters, wherein 0<i<n; in response to the difference between the recommended prediction probability and the historical interaction events meeting the preset convergence condition during the training process of the i-th order sample architecture parameters, fix the sample architecture parameters of other orders except the i+1-th order sample architecture parameters, and perform gradient adjustment on the i+1-th order sample architecture parameters.

[0256] In an optional embodiment, the method further includes a sample extraction model, wherein the sample extraction model is used to extract features from the sample account attribute data and the sample content attribute data;

[0257] The acquiring unit 711 is further configured to acquire a second learning rate;

[0258] The adjustment unit 712 is further configured to perform gradient adjustment on the model parameters of the sample extraction model based on the difference between the recommendation prediction probability and the historical interaction event, and the second learning rate, to obtain a feature extraction model.

[0259] In an optional embodiment, the training module 750 is further used to train the sample architecture parameters based on the difference between the recommended prediction probability and the historical interaction event to obtain candidate architecture parameters; sort the candidate architecture parameters according to parameter weights to obtain parameter sorting results corresponding to the candidate architecture parameters; and use a specified proportion of candidate architecture parameters in the parameter sorting results as the architecture parameters.

[0260] In an optional embodiment, the analysis module 730 is further configured to input the multi-stage interaction sample representation into a sample click rate prediction model, and output the recommendation prediction probability of the sample object account corresponding to the sample content.

[0261] In an optional embodiment, the training module 750 is further configured to train the sample click-through rate prediction model based on the difference between the recommendation prediction probability and the historical interaction event to obtain a click-through rate prediction model.

[0262] In summary, the content recommendation device provided in the embodiment of the present application represents the effectiveness of the relationship in the recommendation prediction process through the architecture parameters corresponding to the multi-order interaction relationship between the account attribute data and the content attribute data stored in the model parameter search space, obtains the account feature representation of the target object account and the multi-order interaction feature representation between the content feature representation of the target content according to the model parameter search space, performs predictive analysis on them, and thus determines the recommendation probability of the target content. It can obtain a more effective multi-order interaction feature representation based on the architecture parameters and perform predictive analysis on them, so that the accuracy of the prediction probability is higher. In addition, storing the architecture parameters in the search space instead of storing the deep sparse network can reduce the storage amount, thereby improving the efficiency of content recommendation.

[0263] It should be noted that the content recommendation device provided in the above embodiment is merely an example of the division of the aforementioned functional modules. In actual applications, the aforementioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the content recommendation device provided in the above embodiment and the content recommendation method embodiment are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0264] Figure 9 The following is a schematic diagram showing the structure of a server provided by an exemplary embodiment of the present application. Specifically:

[0265] The server 900 includes a central processing unit (CPU) 901, a system memory 904 including a random access memory (RAM) 902 and a read-only memory (ROM) 903, and a system bus 905 connecting the system memory 904 and the CPU 901. The server 900 also includes a mass storage device 906 for storing an operating system 913, application programs 99, and other program modules 915.

[0266] The mass storage device 906 is connected to the central processing unit 901 via a mass storage controller (not shown) connected to the system bus 905. The mass storage device 906 and its associated computer-readable media provide non-volatile storage for the server 900. That is, the mass storage device 906 may include a computer-readable medium (not shown) such as a hard disk or a compact disc read-only memory (CD-ROM) drive.

[0267] 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, tape cassettes, magnetic tape, 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-mentioned ones. The above-mentioned system memory 904 and mass storage device 906 can be collectively referred to as memory.

[0268] According to various embodiments of the present application, the server 900 may also be connected to a remote computer on a network such as the Internet for operation. That is, the server 900 may be connected to the network 912 via the network interface unit 911 connected to the system bus 905, or the network interface unit 911 may be used to connect to other types of networks or remote computer systems (not shown).

[0269] The memory also includes one or more programs, which are stored in the memory and configured to be executed by the CPU.

[0270] The embodiment of the present application further provides a computer device, which can be implemented as follows: Figure 3The terminal or server shown. The computer device includes a processor and a memory, wherein the memory stores 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 the processor to implement the content recommendation method provided by each of the above method embodiments.

[0271] An embodiment of the present application also provides a computer-readable storage medium, which stores 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.

[0272] Embodiments of the present application also provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the content recommendation method described in any of the above embodiments.

[0273] 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 for description only and do not represent the advantages or disadvantages of the embodiments.

[0274] Those 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 a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0275] The above description is merely an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A content recommendation method, characterized in that: The method comprises: Obtaining, based on a sample parameter search space, a multi-order interactive sample representation of the interaction between a sample account feature representation and a sample content feature representation, wherein the sample account feature representation is a feature representation extracted from the sample account attribute data, and the sample content feature representation is a feature representation extracted from the sample content attribute data; the sample parameter search space includes sample architecture parameters corresponding to the multi-order interactive relationships between the sample account attribute data and the sample content attribute data; Performing predictive analysis on the multi-order interaction sample representation to obtain a recommendation prediction probability between the sample object account and the sample content; Based on the difference between the recommendation prediction probability and the historical interaction events, the sample architecture parameters are trained to obtain candidate architecture parameters; Sort the candidate architecture parameters according to parameter weights to obtain parameter sorting results corresponding to the candidate architecture parameters; The candidate architecture parameters of a specified proportion in the parameter ranking result are used as the architecture parameters corresponding to the multi-order interaction relationships in the model parameter search space; the architecture parameters are used to indicate the effectiveness of the multi-order interaction relationships in the recommendation prediction process; Obtain target account attribute data of the target object account and target content attribute data of the target content; extracting an account feature representation corresponding to the target account attribute data, and extracting a content feature representation corresponding to the target content attribute data; obtaining, based on the model parameter search space, a multi-order interaction feature representation of the interaction between at least one of the account feature representations and at least one of the content feature representations, wherein the feature order of the multi-order interaction feature representation corresponds to the number of interacting features; Performing predictive analysis on the multi-order interaction feature representation to obtain a recommendation probability between the target object account and the target content; Content is recommended to the target account based on the recommendation probability.

2. The method according to claim 1, characterized in that The acquiring, based on the model parameter search space, a multi-order interaction feature representation of the interaction between at least one of the account feature representations and at least one of the content feature representations includes: Obtaining architecture parameters corresponding to the multi-order interaction relationship in the model parameter search space, where the multi-order interaction relationship refers to an interaction relationship between at least one account feature representation and at least one content feature representation; The account feature representation of the target account attribute data and the content feature representation of the target content attribute data are weightedly combined based on the architecture parameters in the model parameter search space to obtain the multi-order interaction feature representation.

3. The method according to claim 1, characterized in that Before acquiring a multi-order interactive sample representation based on the sample parameter search space for interaction between the sample account feature representation and the sample content feature representation, the method further includes: Acquire a sample data set, wherein the sample data set includes the sample account attribute data of the sample object account and the sample content attribute data of the sample content, wherein the sample object account and the sample content are annotated with the historical interaction event; The sample account feature representation corresponding to the sample account attribute data is extracted, and the sample content feature representation corresponding to the sample content attribute data is extracted.

4. The method according to claim 1, wherein Before acquiring a multi-order interactive sample representation based on the sample parameter search space for interaction between the sample account feature representation and the sample content feature representation, the method further includes: Acquire a sample parameter search space, wherein the sample parameter search space includes initial architecture parameters; Performing tensor decomposition on the initial architecture parameters to obtain tensor product results corresponding to the initial architecture parameters as the sample architecture parameters.

5. The method according to claim 1, wherein The training of the sample architecture parameters based on the difference between the recommendation prediction probability and the historical interaction events includes: Get the first learning rate; Based on the difference between the recommendation prediction probability and the historical interaction event, and the first learning rate, a gradient adjustment is performed on the sample architecture parameters.

6. The method according to claim 5, characterized in that The sample frame parameters include n-order sample frame parameters, where n is a positive integer; The gradient adjustment of the sample architecture parameters based on the difference between the recommendation prediction probability and the historical interaction event, and the first learning rate, includes: Based on the difference between the recommendation prediction probability and the historical interaction event, and the first learning rate, fixing the other order sample architecture parameters except the i-th order sample architecture parameter, and performing gradient adjustment on the i-th order sample architecture parameter, where 0<i<n; In response to the training process of the i-th order sample architecture parameters, the difference between the recommended prediction probability and the historical interaction event meets the preset convergence condition, the sample architecture parameters of other orders except the i+1-th order sample architecture parameters are fixed, and the i+1-th order sample architecture parameters are gradient adjusted.

7. The method according to claim 1, characterized in that The method further includes a sample extraction model, wherein the sample extraction model is used to extract features from the sample account attribute data and the sample content attribute data; The method further comprises: Get the second learning rate; Based on the difference between the recommendation prediction probability and the historical interaction event, and the second learning rate, gradient adjustment is performed on the model parameters of the sample extraction model to obtain a feature extraction model.

8. The method according to any one of claims 1 to 7, characterized in that: The performing predictive analysis on the multi-order interaction sample representation to obtain the recommendation prediction probability between the sample object account and the sample content includes: The multi-order interaction sample representation is input into a sample click rate prediction model, and the recommendation prediction probability of the sample object account corresponding to the sample content is obtained as an output.

9. The method according to claim 8, characterized in that After inputting the multi-order interaction sample representation into the sample click rate prediction model and outputting the recommendation prediction probability of the sample object account corresponding to the sample content, the method further includes: Based on the difference between the recommendation prediction probability and the historical interaction events, the sample click-through rate prediction model is trained to obtain a click-through rate prediction model.

10. A content recommendation device, characterized in that: The device comprises: An acquisition module is configured to acquire, based on a sample parameter search space, a multi-order interaction sample representation of the interaction between a sample account feature representation and a sample content feature representation, wherein the sample account feature representation is a feature representation extracted from sample account attribute data, and the sample content feature representation is a feature representation extracted from sample content attribute data; the sample parameter search space includes sample architecture parameters corresponding to the multi-order interaction relationships between the sample account attribute data and the sample content attribute data; perform predictive analysis on the multi-order interaction sample representation to obtain a recommendation prediction probability between the sample object account and the sample content; train the sample architecture parameters based on the difference between the recommendation prediction probability and historical interaction events to obtain candidate architecture parameters; sort the candidate architecture parameters according to parameter weights to obtain a parameter sorting result corresponding to the candidate architecture parameters; use a specified proportion of candidate architecture parameters in the parameter sorting result as the architecture parameters corresponding to the multi-order interaction relationships in the model parameter search space; the architecture parameters are used to indicate the effectiveness of the multi-order interaction relationships in the recommendation prediction process; obtain target account attribute data of the target object account and target content attribute data of the target content; an extraction module, configured to extract an account feature representation corresponding to the target account attribute data, and to extract a content feature representation corresponding to the target content attribute data; The acquisition module is further configured to acquire, based on the model parameter search space, a multi-order interaction feature representation of the interaction between at least one of the account feature representations and at least one of the content feature representations, wherein the feature order of the multi-order interaction feature representation corresponds to the number of interacting features; An analysis module, configured to perform predictive analysis on the multi-order interaction feature representation to obtain a recommendation probability between the target object account and the target content; A recommendation module is used to recommend content to the target account based on the recommendation probability.

11. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one program, and the at least one program is loaded and executed by the processor to implement the content recommendation method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that The storage medium stores at least one program, and the at least one program is loaded and executed by the processor to implement the content recommendation method according to any one of claims 1 to 9.

13. A computer program product, characterized in that The method comprises computer instructions, which, when executed by a processor, implements the content recommendation method according to any one of claims 1 to 9.

Images