Method, apparatus, device, storage medium and product for generating content recommendation model

Abstract

The present application discloses a method, apparatus, device, storage medium and product for generating a content recommendation model, relating to the field of machine learning. The method includes: generating at least two candidate content recommendation models in a search space; obtaining model encodings respectively corresponding to the at least two candidate content recommendation models in the search space; performing performance prediction on the at least two candidate content recommendation models based on the model encodings respectively corresponding to the at least two candidate content recommendation models, to obtain performance prediction results respectively corresponding to the at least two candidate content recommendation models; and determining a target content recommendation model from the at least two candidate content recommendation models based on the performance prediction results, where the target content recommendation model is used to recommend content to an account. Through the above manner, it is possible to more quickly and automatically select a target content recommendation model with better recommendation performance from multiple candidate content recommendation models. The present application can be applied to various scenarios such as cloud technology, artificial intelligence, and intelligent transportation.

Description

Technical Field

[0001] Embodiments of the present application relate to the field of machine learning, and particularly to a method, apparatus, device, storage medium, and product for generating a content recommendation model. Background Art

[0002] With the development of network technology, the phenomenon of information overload has become increasingly obvious. It is difficult for traditional information recommendation methods to provide personalized recommendation processes for objects from a vast amount of information.

[0003] In related technologies, relevant data on object preferences and needs are usually collected, and with the help of the collected training data, a manually constructed personalized recommendation model is trained. Then, with the help of the trained recommendation model, information is recommended for the object. For example, according to the historical preference data of the object, information that meets its preferences is recommended for the object.

[0004] In the above process, although the trained recommendation model can provide a relatively effective recommendation process for the object, the manually constructed personalized recommendation model is still inevitably intervened by human cognition, and the trained recommendation model is highly correlated with the collection of training data. When this training model is used to analyze other relevant data, the recommendation prediction effect of the data will be greatly reduced. Summary of the Invention

[0005] Embodiments of the present application provide a method, apparatus, device, storage medium, and product for generating a content recommendation model, which can avoid excessive manual intervention, more quickly and automatically select a target content recommendation model with better recommendation performance, and thus, with the help of the target content recommendation model, more accurately recommend appropriate content for different accounts. The technical solutions are as follows.

[0006] On the one hand, a method for generating a content recommendation model is provided. The method includes:

[0007] Generating at least two candidate content recommendation models in a search space, where the search space includes a preset neural network composition structure, and the candidate content recommendation models are candidate models for content recommendation analysis;

[0008] Obtaining model encodings respectively corresponding to the at least two candidate content recommendation models in the search space, where the model encodings are used to indicate the composition patterns of the neural network composition structures in the candidate content recommendation models;

[0009] Performing performance prediction on the at least two candidate content recommendation models based on the model encodings respectively corresponding to the at least two candidate content recommendation models, to obtain performance prediction results respectively corresponding to the at least two candidate content recommendation models;

[0010] Determine a target content recommendation model from the at least two candidate content recommendation models based on the performance prediction result, where the target content recommendation model is used to recommend content to an account.

[0011] On the other hand, a device for generating a content recommendation model is provided. The device includes:

[0012] A generation module, configured to generate at least two candidate content recommendation models in a search space, where the search space includes a preset neural network composition structure, and the candidate content recommendation models are candidate models for performing content recommendation analysis;

[0013] An acquisition module, configured to acquire model encodings respectively corresponding to the at least two candidate content recommendation models in the search space, where the model encodings are used to indicate the composition patterns of the neural network composition structures in the candidate content recommendation models;

[0014] A prediction module, configured to perform performance prediction on the at least two candidate content recommendation models based on the model encodings respectively corresponding to the at least two candidate content recommendation models, to obtain performance prediction results respectively corresponding to the at least two candidate content recommendation models;

[0015] A determination module, configured to determine a target content recommendation model from the at least two candidate content recommendation models based on the performance prediction result, where the target content recommendation model is used to recommend content to an account.

[0016] On the other hand, a computer device is provided. The computer device includes a processor and a memory. At least one instruction, at least one program, a code set or an instruction set is stored in the memory. The at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the method for generating a content recommendation model as described in any one of the above embodiments of the present application.

[0017] On the other hand, a computer-readable storage medium is provided. At least one instruction, at least one program, a code set or an instruction set is stored in the storage medium. 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 the method for generating a content recommendation model as described in any one of the above embodiments of the present application.

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

[0019] The beneficial effects brought by the technical solutions provided in the embodiments of the present application at least include:

[0020] In the search space, a candidate content recommendation model is generated based on a preset neural network composition structure. Through the composition pattern of the neural network composition structure, the model codes corresponding to the candidate content recommendation models are obtained, so as to predict the performance of the candidate content recommendation models, and to implement the process of determining the target content recommendation model from the candidate content recommendation models. According to the preset neural network composition structure, not only the range of the candidate content recommendation models generated by the search space is limited, making it conform to the basic paradigm of the neural network of the content recommendation system, but also the limitation of only using the existing models for content recommendation can be avoided. In addition, according to the composition pattern in the neural network composition structure, the model codes of different candidate content recommendation models are determined, and the model performance of each candidate content recommendation model is determined based on the model codes, so that a candidate content recommendation model with better recommendation performance can be more quickly and automatically selected from multiple candidate content recommendation models as the target content recommendation model. When content is recommended to an account through the target content recommendation model, appropriate content can be recommended to different accounts more accurately and efficiently, improving the accuracy of content recommendation and enhancing the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0022] Figure 1 It is a schematic diagram of related technologies provided by an exemplary embodiment of the present application;

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

[0024] Figure 3 It is a flowchart of a method for generating a content recommendation model provided by an exemplary embodiment of the present application;

[0025] Figure 4 It is a flowchart of a method for generating a content recommendation model provided by another exemplary embodiment of the present application;

[0026] Figure 5 It is a flowchart of a method for generating a content recommendation model provided by still another exemplary embodiment of the present application;

[0027] Figure 6It is a schematic diagram of a substructure of a search space provided by an exemplary embodiment of the present application;

[0028] Figure 7 It is a training flowchart of a performance prediction model provided by an exemplary embodiment of the present application;

[0029] Figure 8 It is a structural block diagram of a method for generating a content recommendation model provided by an exemplary embodiment of the present application;

[0030] Figure 9 It is a structural block diagram of a server provided by an exemplary embodiment of the present application. Detailed implementation manners

[0031] To make the objectives, technical solutions and advantages of the present application clearer, the following will further describe the embodiments of the present application in detail with reference to the accompanying drawings.

[0032] First, a brief introduction to the nouns involved in the embodiments of the present application is given.

[0033] Artificial Intelligence (AI): It is to use a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, a theory, method, technology and application system that perceives the environment, acquires knowledge and uses knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science. It attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines to enable the machines to have the functions of perception, reasoning and decision-making.

[0034] Artificial intelligence technology is an interdisciplinary subject, involving a wide range of fields, including both hardware-level technologies and software-level technologies. The basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. The software technologies of artificial intelligence mainly include several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0035] Machine Learning (ML): It is an interdisciplinary subject involving multiple fields such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers simulate or implement human learning behaviors to acquire new knowledge or skills, and reorganize the existing knowledge structure to continuously improve their own performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence. Machine learning and deep learning usually include technologies such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

[0036] Recommendation systems are the basic information infrastructure in the era of information overload, used to select content that meets the needs and preferences of an object from a vast amount of information for display, and are one of the most successful application scenarios of machine learning and even artificial intelligence technologies. In the field of recommendation systems, the collaborative filtering task is the earliest and relatively basic recommendation task. Its core idea is to legally collect the behavior data of a batch of objects, conduct collaborative analysis on the behavior data, and then perform information filtering and information recommendation for the objects based on the results of the collaborative analysis. Schematically, the similarity between objects and between items is calculated through heuristic methods or machine learning model methods, so as to calculate the items that meet the needs of the target object and obtain an item recommendation list. In recent years, model-based collaborative filtering methods have become the mainstream choice, from the early matrix factorization models to the recent neural network models.

[0037] Optionally, when using the Neural Collaborative Filtering (NCF) method for collaborative filtering, the interaction relationship between the object and the item is learned through a multi-layer perceptron. The traditional linear model is replaced by a non-linear multi-layer perceptron, which has stronger modeling ability on large datasets.

[0038] Schematically, as Figure 1 shown, it is a framework schematic diagram of the NCF method, which includes an input layer 110, an embedding layer 120, a neural collaborative filtering layer 130, and an output layer 140.

[0039] In the input layer 110, the object (User) and the item (Item) are converted into sparse vectors through one-hot encoding.

[0040] In the embedding layer 120, the User vector and the Item vector are respectively embedded into a space with a smaller dimension, assumed to be K dimensions. Among them, the embedding matrix is represented as P, and the embedding layer 120 multiplies the input User vector u i by the embedding matrix P to obtain the embedded vector P of the User vector iSchematically, if there are a total of M objects and the embedding dimension is K - dimensional, the size of the embedding matrix of the objects is M×K, where the i - th row of the embedding matrix represents the embedding vector of the i - th object; similarly, if there are a total of N items, the size of the embedding matrix of the items is N×K.

[0041] In the neural collaborative filtering layer 130, after feeding the embedding vectors of the objects and the embedding vectors of the items into the NCF, a vector is obtained after passing through the internal processing layers (layer1, layer2... layerX) of the NCF. Among them, the internal processing layers of the NCF can be preset.

[0042] In the output layer 140, the vector output by the NCF is mapped through a fully - connected layer to obtain a predicted score. And gradient descent is performed through a loss function to update the model parameters of the NCF.

[0043] However, in terms of the dataset, the datasets used in different collaborative filtering tasks often have different properties. For example, there are differences in the form, scale, distribution, etc. of the datasets. Schematically, the datasets adopted by the collaborative filtering method are generally implemented in two forms, implicit and explicit, and the datasets may be different in scale (large or small) and distribution (dense or sparse). In addition, at the model level, on the one hand, the matrix factorization method is easy to train, but due to its limited model capacity, it cannot capture complex object - item interaction behaviors; on the other hand, when the amount of data is sufficient, the neural network model may obtain better performance, but it may perform poorly for relatively small datasets. For example, although the NCF method can stably outperform the matrix factorization model (MF) method on relatively large datasets, the NCF model may not necessarily achieve better performance on some datasets, and the above - mentioned methods rely too much on the selection of the dataset. Only after manually selecting a suitable dataset as the training dataset can a highly accurate recommendation model be trained.

[0044] In the embodiments of the present application, a method for generating a content recommendation model is provided, which can avoid excessive manual intervention and more quickly and automatically select a target content recommendation model with better recommendation performance from multiple generated candidate content recommendation models, so as to more accurately recommend suitable content for different accounts by means of the target content recommendation model. For the method for generating the content recommendation model trained in the present application, when applied, it includes at least one of multiple scenarios such as a music recommendation scenario, a news recommendation scenario, a video recommendation scenario, etc.

[0045] It should be noted that the above - mentioned application scenarios are only schematic examples, and the method for generating the content recommendation model provided in this embodiment can also be applied to other scenarios, and the embodiments of the present application do not limit this.

[0046] It should be noted that the information involved in this application (including but not limited to object device information, object personal information, etc.), data (including but not limited to data for analysis, stored data, displayed data, etc.), and signals are all authorized by the object or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions. For example, the account data, content data, etc. involved in this application are obtained under full authorization.

[0047] Secondly, the implementation environment involved in the embodiments of this application is described. Schematically, please refer to Figure 2 , in which the implementation environment involves a terminal 210 and a server 220, and the terminal 210 and the server 220 are connected through a communication network 230.

[0048] In some embodiments, the terminal 210 is used to send account data and content data to the server 220. Among them, the account data refers to the profile information stored in the terminal and related to the object, such as: operation data of the account, information data of the account, etc.; the content data refers to the content to be recommended to the object (account), such as: news, film and television songs, etc. Optionally, the content data can also be implemented as data obtained by the server 220. In some embodiments, application programs are installed in the terminal 210, and different application programs correspond to different account data and content data.

[0049] The method for generating the content recommendation model provided by the embodiments of this application can be implemented by the terminal 210 alone, or by the server 220, or by data interaction between the terminal 210 and the server 220. The embodiments of this application do not limit this. In this embodiment, after the terminal 210 obtains the account data through an application program with a data acquisition function, it sends the obtained account data to the server 220. Schematically, the server 220 recommends content data to different accounts according to the account data, where the server 220 performs the recommendation process through the target content recommendation model. Schematically, the acquisition method of the target content recommendation model is described.

[0050] Optionally, the server 220 generates at least two candidate content recommendation models (such as candidate content recommendation model 1, candidate content recommendation model 2, candidate content recommendation model n, etc.) in a search space composed of a pre-set neural network composition structure, and determines model encodings corresponding to the at least two candidate content recommendation models respectively (such as model encoding 1 corresponding to candidate content recommendation model 1, model encoding 2 corresponding to candidate content recommendation model 2, model encoding n corresponding to candidate content recommendation model n) based on the composition patterns of the neural network composition structures in the candidate content recommendation models. In addition, the server 220 performs performance prediction on the at least two candidate content recommendation models based on the model encodings corresponding to the at least two candidate content recommendation models respectively, obtains performance prediction results corresponding to the at least two candidate content recommendation models respectively (such as performance prediction result 1 corresponding to candidate content recommendation model 1, performance prediction result 2 corresponding to candidate content recommendation model 2, performance prediction result n corresponding to candidate content recommendation model n), and determines a target content recommendation model from the at least two candidate content recommendation models based on the performance prediction results.

[0051] Optionally, after receiving the account data and content data (or account data) sent by the terminal 210, the server 220 recommends content to different accounts based on the target content recommendation model, that is, recommends corresponding content data to different accounts. For example, it recommends news, videos, songs, etc. that match their preferences to different accounts.

[0052] It should be noted that the above-mentioned terminal includes, but is not limited to, mobile terminals such as mobile phones, tablet computers, portable laptops, intelligent voice interaction devices, smart home appliances, in-vehicle terminals, etc., and can also be implemented as a desktop computer, etc.; the above-mentioned server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, and can also be a cloud server that provides 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 Network (CDN), and big data and artificial intelligence platforms.

[0053] Among them, cloud technology refers to a hosting technology that unifies a series of resources such as hardware, application programs, and networks within a wide area network or local area network to achieve data calculation, storage, processing, and sharing. Cloud technology is the general term for network technology, information technology, integration technology, management platform technology, application technology, etc. based on the cloud computing business model, which can form a resource pool, be used on demand, and be flexible and convenient.

[0054] In some embodiments, the above-mentioned server can also be implemented as a node in a blockchain system.

[0055] In combination with the above-mentioned noun introduction and application scenarios, the method for generating the content recommendation model provided by the present application is described, and the method is applied to the server as an example. Figure 3 As shown, the method includes the following steps 310 to 340.

[0056] Step 310: Generate at least two candidate content recommendation models in the search space.

[0057] In principle, the proposal of a new network structure requires the proposal of some targeted improvement solutions based on the deficiencies of the previous network structure in some aspects. For example, adding some modules that are conducive to training, or improving the basic convolutional layer module, etc. However, the process of manually building a model is usually very time-consuming and labor-intensive, so the automatic search network came into being.

[0058] The process of automatically searching for network structures is called NAS (Neural Architecture Search, NAS), which usually involves three aspects: search space, search strategy, and model estimation. The search space determines the upper and lower limits of the searched model performance. A well-designed search space can avoid finding network structures with no potential.

[0059] The search space includes a preset neural network structure.

[0060] Optionally, at least two candidate content recommendation models are generated in the search space through a preset neural network composition structure. Schematically, the neural network composition structure is a structural prototype of the candidate content recommendation model, which includes at least one substructure, for example: the candidate content recommendation model is a model obtained based on at least one substructure in the neural network composition structure, and different candidate content recommendation models are obtained through different forms of substructures.

[0061] In an optional embodiment, each substructure corresponds to multiple algorithmic operations. In each substructure corresponding to the neural network composition structure, an algorithmic operation is selected from the multiple algorithmic operations as the algorithmic operation corresponding to the substructure. According to the distribution of different substructures in the neural network composition structure, different substructures are spliced together to obtain a candidate content recommendation model.

[0062] Optionally, in a neural network composition structure preset in the search space, at least one substructure is selected to form a candidate content recommendation model based on a substructure in the neural network composition structure.

[0063] Schematically, the preset neural network composition structure in the search space includes four sub-structures, namely sub-structure A, sub-structure B, sub-structure C, and sub-structure D. Among them, sub-structure A correspondingly includes algorithms A1 and A2, sub-structure B correspondingly includes algorithms B1 and B2, sub-structure C correspondingly includes algorithms C1, C2, and C3, and sub-structure D correspondingly includes algorithms D1 and D2.

[0064] Based on the distribution of different sub-structures in the neural network composition structure and the selection combinations of different algorithms in different sub-structures, different candidate content recommendation models are obtained. For example: a candidate content recommendation model 1 is composed of algorithm A1 in sub-structure A, algorithm B2 in sub-structure B, algorithm C2 in sub-structure C, and algorithm D1 in sub-structure D; a candidate content recommendation model 2 is composed of algorithm A1 in sub-structure A, algorithm B2 in sub-structure B, algorithm C2 in sub-structure C, and algorithm D2 in sub-structure D, etc.

[0065] Among them, the candidate content recommendation model is a candidate model for content recommendation analysis.

[0066] Content recommendation is used to indicate recommending appropriate content to the objects in need. For example: in the content recommendation scenario, there are multiple recommended contents to be recommended. When recommending content to object A, the recommended content to be recommended to object A is selected from multiple recommended contents to be recommended. Through the candidate content recommendation model, recommended content that meets the preferences of different objects can be recommended.

[0067] In an optional embodiment, the search space is not only used to generate candidate content recommendation models, but also can pre-store candidate content recommendation models. For example: the pre-stored candidate content recommendation model is a content recommendation model obtained through related technologies.

[0068] Step 320, obtain the model codes corresponding to at least two candidate content recommendation models in the search space.

[0069] Among them, the model code is used to indicate the composition pattern of the neural network composition structure in the candidate content recommendation model.

[0070] Schematically, when composing a candidate content recommendation model based on a preset neural network composition structure, according to the differences in the neural network composition structures of the candidate content recommendation models, at least two candidate content recommendation models each have their corresponding model codes.

[0071] Optionally, the model encodings corresponding to the candidate content recommendation models are obtained based on the algorithm operations of their neural network compositions. Schematically, in a candidate content recommendation model, the structure encoding corresponding to each sub-structure in the neural network composition is determined, and after splicing the structure encodings corresponding to different sub-structures, the model encoding corresponding to the candidate content recommendation model is obtained.

[0072] Schematically, the candidate content recommendation model 1 is a recommendation model composed of algorithm A1 in sub-structure A, algorithm B2 in sub-structure B, algorithm C2 in sub-structure C, and algorithm D1 in sub-structure D. Then, in the candidate content recommendation model 1, the structure encoding of sub-structure A is represented as (1, 0), the structure encoding of sub-structure B is represented as (0, 1), the structure encoding of sub-structure C is represented as (0, 1, 0), and the structure encoding of sub-structure D is represented as (0, 1). Splice the structure encodings corresponding to different sub-structures in the candidate content recommendation model 1 to obtain the model encoding corresponding to the candidate content recommendation model 1.

[0073] In an alternative embodiment, after obtaining the candidate content recommendation models composed of different sub-structures, the candidate content recommendation models are encoded to obtain the model encodings corresponding to the candidate content recommendation models.

[0074] Schematically, after obtaining the above candidate content recommendation model 1, the candidate content recommendation model 1 is encoded. For example, the encoding result is 001. Use the encoding result 001 corresponding to the candidate content recommendation model 1 as the model encoding corresponding to the candidate content recommendation model 1. When the machine analyzes the candidate content recommendation model 1, it can determine the parameter operation situation of the neural network composition structure that makes up the candidate content recommendation model 1 according to the "001" corresponding to the model encoding. Similarly, after obtaining the above candidate content recommendation model 2, the candidate content recommendation model 2 is encoded. For example, the encoding result is 010. Use the encoding result 010 corresponding to the candidate content recommendation model 2 as the model encoding corresponding to the candidate content recommendation model 2, and so on. Among them, the encoding results corresponding to different candidate content recommendation models are different and are used to distinguish different candidate content recommendation models.

[0075] Step 330, perform performance prediction on at least two candidate content recommendation models based on the model encodings respectively corresponding to the at least two candidate content recommendation models, and obtain the performance prediction results respectively corresponding to the at least two candidate content recommendation models.

[0076] Schematically, based on the model encodings respectively corresponding to at least two candidate content recommendation models, different candidate content recommendation models and the composition modes of the neural network composition structures respectively corresponding to different candidate content recommendation models are distinguished.

[0077] Optionally, performance prediction is used to indicate the prediction of the content recommendation performance of the candidate content recommendation models. Schematically, after obtaining at least two candidate content recommendation models, the content recommendation accuracy of the at least two candidate content recommendation models is determined, so as to determine the content recommendation performance of different candidate content recommendation models, that is: using the content recommendation accuracy as the performance prediction result; or, after obtaining at least two candidate content recommendation models, determining the content recommendation diversity of the at least two candidate content recommendation models, so as to determine the content recommendation performance of different candidate content recommendation models, that is: using the content recommendation diversity as the performance prediction result; or, after obtaining at least two candidate content recommendation models, determining the content recommendation novelty of the at least two candidate content recommendation models, so as to determine the content recommendation performance of different candidate content recommendation models, that is: using the content recommendation novelty as the performance prediction result, etc.

[0078] Step 340, determining a target content recommendation model from at least two candidate content recommendation models based on the performance prediction result.

[0079] Optionally, after obtaining the performance prediction results corresponding to at least two candidate content recommendation models respectively, comparing the performance prediction results corresponding to the at least two candidate content recommendation models respectively, and selecting at least one candidate content recommendation model from the at least two candidate content recommendation models as the target content recommendation model according to the performance prediction result.

[0080] For example: using one candidate content recommendation model with the best performance prediction result as the only target content recommendation model; or, using the n candidate content recommendation models with the best performance prediction results as the target content recommendation models, where n is a positive integer.

[0081] Schematically, when the performance prediction result is implemented as the content recommendation accuracy, using the n candidate content recommendation models with the highest content recommendation accuracy as the target content recommendation models; or, when the performance prediction result is implemented as the content recommendation diversity, using one candidate content recommendation model with the highest content recommendation diversity as the target content recommendation model, etc.

[0082] The target content recommendation model is used to recommend content to the account.

[0083] Optionally, after obtaining the target content recommendation model, using the account and the recommended content as the input of the model, and through the target content recommendation model, allocating corresponding recommended content to different accounts, so as to recommend the recommended content to the corresponding accounts.

[0084] The account is used to indicate the object receiving the recommended content; the recommended content is used to indicate the content information in the process of recommending content to the object. Schematically, the recommended content can be implemented in at least one of the following forms.

[0085] (1) The recommended content is implemented in text form

[0086] Schematically, taking a news application as an example, during the operation of the news application, in addition to pushing real-time updated news content to the object, it may also recommend different news content for different accounts according to the news browsing situations of different accounts.

[0087] For example: when providing news recommendation services for account X, the news browsing situation of account X in the historical time period is used as the account information corresponding to account X, the account information corresponding to account X is used as the input of the target content recommendation model, and the target content recommendation model recommends news text content that conforms to its news browsing situation for account X based on the account information corresponding to account X, etc.

[0088] (2) The recommended content is implemented in multimedia form

[0089] Schematically, taking a music application as an example, the music application can recommend corresponding music content for different accounts based on the music listening preferences corresponding to different accounts through the above-mentioned target content recommendation model.

[0090] For example: when providing news recommendation services for account Y, the music like situation of account Y in the historical time period is used as the account information corresponding to account Y, the account information corresponding to account Y is used as the input of the target content recommendation model, and the target content recommendation model recommends recommended music that conforms to its music like situation for account Y based on the account information corresponding to account Y, etc.

[0091] It should be noted that the above are only schematic examples, and the embodiments of the present application are not limited thereto.

[0092] In summary, in the search space, candidate content recommendation models are generated based on different pre-set neural network composition structures. Through the composition patterns of the neural network composition structures, the model encodings corresponding to the candidate content recommendation models are obtained respectively, so as to predict the performance of the candidate content recommendation models, thereby realizing the process of determining the target content recommendation model from the candidate content recommendation models. According to the pre-set neural network composition structure, not only the range of candidate content recommendation models generated by the search space is limited, but also the limitations of only using existing models for content recommendation can be avoided. In addition, according to the composition patterns in the neural network composition structure, the model encodings of different candidate content recommendation models are determined, and the model performance of each candidate content recommendation model is determined based on the model encodings, so that a candidate content recommendation model with better recommendation performance can be selected more quickly and automatically from multiple candidate content recommendation models as the target content recommendation model. When content is recommended to an account through the target content recommendation model, appropriate content can be recommended to different accounts more accurately and efficiently, improving the accuracy of content recommendation and enhancing the user experience of the object.

[0093] In an alternative embodiment, in the pre-set neural network composition structure included in the search space, there are multiple sub-structures. Different candidate content recommendation models are generated in the search space through the combination of the multiple sub-structures. Schematically, as Figure 4 shown, the above Figure 3 illustrated embodiment can also be implemented as steps 410 to 460 below.

[0094] Step 410, determine the operator selection results corresponding to the multiple sub-structures respectively.

[0095] Schematically, the different candidate content recommendation models generated in the search space are determined based on the differences of the multiple sub-structures in the neural network composition structure. Among them, different sub-structures have their corresponding operator selection results respectively.

[0096] Among them, the operator selection result is used to indicate the operator adopted when performing parameter operations inside the sub-structure.

[0097] Schematically, for the same sub-structure, the function corresponding to the sub-structure is pre-set, and the selection of the operator is used to indicate different methods for implementing the function.

[0098] Optionally, the operator includes both the method adopted when performing parameter operations and the elements involved when performing parameter operations. Schematically, for different sub-structures, the operators adopted when constructing the sub-structure are pre-determined, and there are multiple choices for the operators corresponding to different sub-structures. The operator selection result is used to indicate the result determined after the operator is selected.

[0099] For example, for sub-structure A, the pre-set operators corresponding to sub-structure A include operator A1 and operator A2. When operator A1 is selected in sub-structure A, it represents that the operator selection result corresponding to sub-structure A is operator A1. Similarly, for sub-structure B, the pre-set operators corresponding to sub-structure B include operator B1, operator B2, and operator B3. When operator B2 is selected in sub-structure B, it represents that the operator selection result corresponding to sub-structure B is operator B2. And so on, to determine the operator selection results corresponding to different sub-structures.

[0100] In an alternative embodiment, the search space generates at least two candidate content recommendation models through a historical interaction data set and a pre-set neural network composition structure.

[0101] Among them, the historical interaction data set stores at least one of account data, content data, and interaction data pairs.

[0102] Schematically, the account data mainly refers to the operation data of the object within a historical time period. For example: the account data is the operation data of the object within the past month, such as the like data corresponding to the like operation, the comment data corresponding to the comment operation, the deletion data corresponding to the deletion operation, etc. That is, the operation data includes not only positive data but also negative data. It should be noted that the above account data, operation data, and other data are all obtained with the authorization of the object.

[0103] Optionally, the content data is used to indicate various recommended contents, such as: text content, image content, audio-visual content, etc. Schematically, the content data is related to the content recommendation scenario. When the content recommendation scenario is implemented as a music recommendation scenario, the content data is usually implemented as audio content; or, when the content recommendation scenario is implemented as a video recommendation scenario, the content data is usually implemented as video data, etc.

[0104] Among them, the interaction data pair is used to indicate that there is a historical interaction relationship between at least one account data and one content data. For example: if account data M has liked content data N, then there is a historical interaction relationship between account data M and content data N. Taking account data M and content data N as an interaction data pair, it is expressed as "account data M - content data N"; or, if account data L has commented on content data Q, then there is a historical interaction relationship between account data L and content data Q. Taking account data L and content data Q as an interaction data pair, it is expressed as "account data L - content data Q".

[0105] In an alternative embodiment, taking the pre-set neural network composition structure including an input encoding sub-structure, an embedding function sub-structure, an interaction function sub-structure, and a prediction function sub-structure as an example, the operators within different sub-structures are described.

[0106] (1) Input encoding sub-structure

[0107] In an optional embodiment, in response to the input encoding sub-structure being included in a plurality of sub-structures, obtain a first matrix representation corresponding to the account data and a second matrix representation corresponding to the content data; or, obtain an account interaction matrix representation corresponding to the account data and a content interaction matrix representation corresponding to the content data in the interaction data pair.

[0108] Among them, the input encoding sub-structure is a model structure that constitutes a candidate content recommendation model and is used to perform matrix transformation on the input data in the search space.

[0109] Optionally, input the historical interaction data set into the search space, and through the input encoding sub-structure in the search space, perform matrix transformation on the data in the historical interaction data set. Based on the matrix transformation performed on different data in the historical interaction data set, determine different operator selection results, that is, determine the operator selection result of the input encoding sub-structure based on the method for obtaining the matrix representation.

[0110] Illustratively, in the historical interaction data set, according to the encoding representations (IDs) corresponding to the account data and the content data respectively, represent different account data and content data, that is: adopt the one-hot encoding method to determine the account encoding representations corresponding to different account data and the content encoding representations corresponding to different content data. For example: the account encoding representation corresponding to the account data ID1 is 1000, the account encoding representation corresponding to the account data ID2 is 0100, etc.; similarly, the account encoding representation corresponding to the content data CON1 is 1000, the account encoding representation corresponding to the content data CON2 is 0100, etc.

[0111] Optionally, based on the fact that the account data and the content data are data encoding representations based on one-hot encoding, convert the account encoding representation corresponding to the account data into a first matrix representation through the input encoding sub-structure; convert the content encoding representation corresponding to the content data into a second matrix representation through the input encoding sub-structure.

[0112] Or, in the historical interaction data set, according to the historical interaction situation of the account data and the content data, represent different account data and content data, that is: adopt the multi-hot encoding method to determine the account interaction encoding representation corresponding to the account data and the content interaction encoding representation corresponding to the content data in the interaction data pair. Illustratively, when encoding and representing the account data and the content data respectively, use the interaction situation between the account data and the content data as the account interaction encoding representation corresponding to the account data, and use the interaction situation between the content data and the account data as the content interaction encoding representation corresponding to the content data.

[0113] For example, if there are historical interactions between account data ID1 and content data CON1 and content data CON2, then based on the historical interaction situations between account data ID1 and content data CON1 and content data CON2, the account interaction code corresponding to account data ID1 is determined to be represented as 1100; if there is a historical interaction between account data ID2 and content data CON1 and there is no historical interaction between account data ID2 and content data CON2, then based on the historical interaction situations between account data ID2 and content data CON1 and content data CON2, the account interaction code corresponding to account data ID2 is determined to be represented as 1000, etc.

[0114] Optionally, based on the historical interaction code representations of the account data and the content data being multi-hot encoding representations, the account interaction code representation corresponding to the account data is converted into an account interaction matrix representation through the input encoding sub-structure; the content interaction code representation corresponding to the content data is converted into a content interaction matrix representation through the input encoding sub-structure.

[0115] Schematically, the encoding representations of the account data and the content data obtained by the above one-hot encoding method are used as one type of operator; the encoding representations of the interaction data pairs obtained by the above multi-hot encoding method are used as another type of operator, that is: in the input encoding sub-structure, there are two types of operators corresponding thereto, and based on the selection of the two types of operators in the input encoding sub-structure, the two types of operator selection results corresponding to the input encoding sub-structure are determined.

[0116] (2) Embedding function sub-structure

[0117] In an optional embodiment, in response to the embedding function sub-structure being included in a plurality of sub-structures, the first matrix representation and the second matrix representation are projected into a vector space to obtain a first embedding vector corresponding to the first matrix representation and a second embedding vector corresponding to the second matrix representation; alternatively, the account interaction matrix representation corresponding to the account data in the interaction data pair and the content interaction matrix representation corresponding to the content data are projected into a vector space to obtain a third embedding vector corresponding to the account interaction matrix representation and a fourth embedding vector corresponding to the content interaction matrix representation.

[0118] Among them, the embedding function sub-structure is a model structure that constitutes a candidate content recommendation model and is used to project high-dimensional encoding into a low-dimensional space in a search space to obtain a low-dimensional vector representation that is convenient for analysis.

[0119] Schematically, the input of the embedding function sub-structure is closely related to the input of the input encoding sub-structure. After the historical interaction data set is input into the input encoding sub-structure in the search space, based on the selection of different operators in the input encoding sub-structure, different operators are selected in the embedding function sub-structure to perform parameter operations inside the sub-structure.

[0120] Optionally, when the operator selection result in the input encoding sub-structure is "the encoded representations of the account data and the content data obtained based on the one-hot encoding method", in the embedding function sub-structure, the operator compatible with this operator selection result is implemented as an embedding matrix look-up function (ID-look-up), and this embedding matrix look-up function is represented in the form of a matrix (Matrix, MAT).

[0121] Optionally, when the operator selection result in the input encoding sub-structure is "the encoded representations of the interaction data pairs obtained based on the multi-hot encoding method", in the embedding function sub-structure, the operator compatible with this operator selection result is implemented in at least the following two forms.

[0122] (1) Embedding matrix look-up function and mean pooling operation

[0123] Among them, this embedding matrix look-up function is represented as MAT and is used to perform matrix operation with the matrix representation corresponding to the interaction data pair. Optionally, a mean-pooling operation is performed on the result after the matrix operation to reduce the computational amount.

[0124] (2) Multilayer Perceptron (MLP) model

[0125] Schematically, after obtaining the matrix representation corresponding to the interaction data pair, passing this matrix representation through the multilayer perceptron model to convert the matrix representation corresponding to the interaction data pair into a dense vector, that is, no longer representing the interaction data pair in the form of a matrix representation, but projecting the core feature corresponding to each dimension in the matrix representation into a low-dimensional space, so as to represent the above matrix representation with a dense vector. For example: most of the elements in the matrix representation are equal to 0, and the element that is 1 in the matrix representation is the core feature in this dimension. Passing the matrix representation through the MLP model to obtain the dense vector corresponding to this matrix representation.

[0126] Among them, the structural encoding of the embedding function structure is determined based on the method for obtaining the embedding vector. Optionally, based on the above analysis, "MAT" and "MLP" are used to represent the corresponding two operators in the embedding function sub-structure.

[0127] Schematically, when the operator selection result in the input encoding sub-structure is "the encoded representations of the account data and the content data obtained based on the one-hot encoding method", in the embedding function sub-structure, the operator "MAT" is selected as the operator selection result, and the first matrix representation and the second matrix representation are projected into the vector space to obtain the first embedding vector corresponding to the first matrix representation and the second embedding vector corresponding to the second matrix representation.

[0128] Schematically, when the operator selection result in the input encoding substructure is "the encoded representation of the interaction data pair obtained based on the multi-hot encoding method", in the embedding function substructure, the operator "MAT" or the operator "MLP" is selected as the operator selection result, and the account interaction matrix representation corresponding to the account data and the content interaction matrix representation corresponding to the content data in the interaction data pair are projected into the vector space to obtain a third embedding vector corresponding to the account interaction matrix representation and a fourth embedding vector corresponding to the content interaction matrix.

[0129] It should be noted that the above is only a schematic example, and the embodiments of the present application are not limited thereto.

[0130] (III) Interaction function substructure

[0131] In an alternative embodiment, in response to the interaction function substructure being included in the multiple substructures, the first embedding vector and the second embedding vector are interacted to obtain a combined vector; or, the third embedding vector and the fourth embedding vector are interacted to obtain a combined vector.

[0132] Among them, the interaction function substructure is a model structure that constitutes the candidate content recommendation model, and is used to interact between the account data and the content data in the search space, so as to realize the matching process between the account data and the content data with the help of the combined vector after interaction.

[0133] Among them, the combined vector is used to indicate the predicted interaction relationship between the account data and the content data.

[0134] Schematically, taking the output of the embedding function substructure as the input of the interaction function substructure, that is, taking the first embedding vector corresponding to the account data and the second embedding vector corresponding to the content data as the input of the interaction function substructure, and in the interaction function substructure, the first embedding vector and the second embedding vector are interacted.

[0135] Optionally, the interaction process is implemented as at least one of the following processing methods: (1) Multiplication processing (MUL), which is used to indicate that the vector product of the first embedding vector and the second embedding vector is used as the combined vector; (2) Subtraction processing (MINUS), which is used to indicate that the vector difference between the first embedding vector and the second embedding vector is used as the combined vector; (3) Maximum value processing (MAX), which is used to indicate that the maximum vector in the first embedding vector and the second embedding vector is used as the combined vector; (4) Minimum value processing (MIN), which is used to indicate that the minimum vector in the first embedding vector and the second embedding vector is used as the combined vector; (5) Concatenation processing (CONCAT), which is used to indicate that the concatenated vector after concatenating the first embedding vector and the second embedding vector is used as the combined vector, etc.

[0136] Among them, the operator selection result of the interaction function sub-structure is determined based on the method for obtaining the combined vector, that is, the operator selection result of the interaction function sub-structure is determined based on the above-mentioned interaction processing method. Optionally, five corresponding operators in the interaction function sub-structure are represented by "MUL", "MINUS", "MAX", "MIN", and "CONCAT".

[0137] Illustratively, when selecting to process the first embedding vector and the second embedding vector, if the operator "MUL" is used as the operator selection result, it means that the combined vector is obtained by multiplying the first embedding vector and the second embedding vector; or, if the operator "CONCAT" is used as the operator selection result, it means that the combined vector is obtained by concatenating the first embedding vector and the second embedding vector, etc.

[0138] It should be noted that the above is only an illustrative example, and the embodiments of the present application do not limit this.

[0139] Optionally, taking the example of obtaining the combined vector by multiplying the first embedding vector and the second embedding vector for illustration. Based on multiple account data and multiple content data stored in the historical interaction dataset, there are also multiple first embedding vectors corresponding to the account data and multiple second embedding vectors corresponding to the content data. The first embedding vector and the second embedding vector are multiplied item by item to generate the combined vector.

[0140] Illustratively, the historical interaction dataset stores account data 1, account data 2, content data 1, and content data 2. Account data 1 corresponds to the first embedding vector 1, account data 2 corresponds to the first embedding vector 2, content data 3 corresponds to the second embedding vector 1, and content data 2 corresponds to the second embedding vector 2. After the embedding vectors are interacted through the interaction function sub-structure, combined vector 1 representing the interaction between the first embedding vector 1 and the second embedding vector 1 is obtained; combined vector 2 representing the interaction between the first embedding vector 1 and the second embedding vector 2; combined vector 3 representing the interaction between the first embedding vector 2 and the second embedding vector 1; combined vector 4 representing the interaction between the first embedding vector 2 and the second embedding vector 2, etc.

[0141] Optionally, when the inputs of the interaction function sub-structure are the third embedding vector and the fourth embedding vector, the above method is used to perform interaction analysis on the third embedding vector and the fourth embedding vector, so as to obtain the combined vector corresponding to the third embedding vector and the fourth embedding vector.

[0142] It should be noted that the above is only an illustrative example, and the embodiments of the present application do not limit this.

[0143] (IV) Prediction function sub-structure

[0144] In an optional embodiment, in response to the prediction function sub-structure being included in multiple sub-structures, an interactive prediction analysis is performed on the combined vector to obtain a prediction result.

[0145] The prediction function sub-structure is a model structure that constitutes a candidate content recommendation model and is used to perform an interactive prediction analysis on the combined vector between account data and content data in the search space, that is, to analyze the interaction between account data and content data.

[0146] The prediction result is used to indicate the difference between the predicted interaction relationship and the historical interaction relationship.

[0147] Schematically, the output of the interaction function sub-structure is used as the input of the prediction function sub-structure. That is, the combined vector after the interaction of account data and content data is used as the input of the interaction function sub-structure. In the interaction function sub-structure, an interactive prediction analysis is performed on the combined vectors after the interaction between different account data and different content data.

[0148] Optionally, the result of the interactive prediction analysis is represented as the prediction score of the combined vector. The higher the score, the higher the probability of the interaction between the account data and content data corresponding to the combined vector.

[0149] Schematically, the interactive prediction analysis is implemented as at least one of the following analysis methods: (1) Summation processing (SUM), which is used to indicate the combined vector; (2) Weight (vector, VEC) assignment processing, which is used to indicate different weights are assigned to different dimensions of the combined vector, where the weight is represented as the inner product of the weighted vector; (3) Multi-layer perceptron processing (MLP), which is used to indicate that the combined vector passes through a multi-layer perceptron, so as to perform an interactive prediction on the combined vector by the multi-layer perceptron.

[0150] The operator selection result of the prediction function structure is determined based on the interactive prediction analysis method, that is, the operator selection result of the prediction function sub-structure is determined based on the above interactive prediction analysis method. Optionally, the three corresponding operators in the prediction function sub-structure are represented by "SUM", "VEC", and "MLP".

[0151] Schematically, when it is selected to process the combined vector through the prediction function sub-structure, if the operator "SUM" is used as the operator selection result, it means that the prediction result is obtained by adding the combined vector; or, if the operator "VEC" is used as the operator selection result, it means that different weights are assigned to different dimensions of the combined vector to obtain the prediction result; or, if the operator "MLP" is used as the operator selection result, it means that the prediction result is obtained through the multi-layer perceptron processing model, etc.

[0152] It should be noted that the above is only a schematic example, and the embodiments of the present application do not limit this.

[0153] Step 420: Generate at least two candidate content recommendation models in the search space based on the operator selection results corresponding to multiple sub-structures respectively.

[0154] Schematically, in each sub-structure, when different operators are selected as the operator selection results, at least two candidate content recommendation models will be obtained when obtaining the candidate content recommendation models composed of multiple sub-structures.

[0155] In an optional embodiment, the operator selection results corresponding to different sub-structures are represented in encoded form.

[0156] Schematically, for each sub-structure, there is a corresponding operator. Based on the selection of the operator in the sub-structure, the operator selection results corresponding to different sub-structures are determined in encoded form.

[0157] Optionally, the operators corresponding to different sub-structures are preset, and different operators have corresponding encoding bits. Based on the encoding bit results corresponding to different operators in the sub-structure, the operator selection result corresponding to the sub-structure is determined. Schematically, before the operator in the sub-structure is selected, the encoding bit results corresponding to different operators in the sub-structure are all 0; after the operator in the sub-structure is selected, the encoding bit corresponding to the selected operator in the sub-structure is set to 1, that is: the encoding bit result corresponding to the selected operator in the sub-structure is 1.

[0158] For example: taking the input encoding sub-structure as an example, the input encoding sub-structure corresponds to 2 operators, namely "encoding representation of account data and content data obtained based on the one-hot encoding method" and "encoding representation of interaction data pairs obtained based on the multi-hot encoding method". After the operator "encoding representation of account data and content data obtained based on the one-hot encoding method" is selected, the encoding bit corresponding to this operator is set to 1, and the operator selection result corresponding to the input encoding sub-structure is represented in encoded form (1, 0).

[0159] Or, taking the embedding function sub-structure as an example, the embedding function sub-structure corresponds to 2 operators, namely "MAT" and "MLP". After the operator "MLP" is selected, the encoding bit corresponding to this operator is set to 1, and the operator selection result corresponding to the embedding function sub-structure is represented in encoded form (0, 1).

[0160] Alternatively, taking the interaction function sub-structure as an example, the interaction function sub-structure corresponds to 5 operators, namely "MUL", "MINUS", "MAX", "MIN" and "CONCAT". After the operator "MINUS" is selected, the corresponding encoding bit of this operator is set to 1, and the operator selection result corresponding to the interaction function sub-structure is represented in the encoding form (0, 1, 0, 0, 0).

[0161] Alternatively, taking the prediction function sub-structure as an example, the prediction function sub-structure corresponds to 3 operators, namely "SUM", "VEC" and "MLP". After the operator "MLP" is selected, the corresponding encoding bit of this operator is set to 1, and the operator selection result corresponding to the prediction function sub-structure is represented in the encoding form (0, 0, 1).

[0162] It should be noted that the above are only illustrative examples of the preset operators and the encoding bits corresponding to the preset operators, and the embodiments of the present application are not limited thereto.

[0163] In an alternative embodiment, the operator selection results corresponding to the input encoding sub-structure, the embedding function sub-structure, the interaction function sub-structure and the prediction function sub-structure are combined to generate at least two candidate content recommendation models in the search space.

[0164] Optionally, after determining the operator selection results corresponding to different sub-structures, the operator selection results corresponding to different sub-structures are combined. Since the candidate content recommendation model is composed of different sub-structures, at least two candidate content recommendation models are obtained after combining the operator selection results corresponding to the sub-structures respectively.

[0165] Schematically, a candidate content recommendation model composed of the above input encoding substructure, embedding function substructure, interaction function substructure, and prediction function substructure will be used as an example for illustration. After selecting the operator "encoding representation of interaction data pairs obtained based on the multi-hot encoding method" in the encoding substructure, the operator "MLP" in the embedding function substructure, the operator "MUL" in the interaction function substructure, and the operator "MLP" in the prediction function substructure, a candidate content recommendation model M1 is obtained; after selecting the operator "encoding representation of interaction data pairs obtained based on the multi-hot encoding method" in the encoding substructure, the operator "MAT" in the embedding function substructure, the operator "MUL" in the interaction function substructure, and the operator "MLP" in the prediction function substructure, another candidate content recommendation model M2 is obtained. That is, compared with candidate content recommendation model M1, there are differences in the operator selection in the embedding function substructure. The operator in the embedding function substructure that makes up candidate content recommendation model M1 is "MLP", but the operator in the embedding function substructure that makes up candidate content recommendation model M2 is "MAT".

[0166] Based on the operator selection results obtained by the above operator selection in different substructures and the combined processing of different substructures, at least two candidate content recommendation models are generated in the search space. The above is only a schematic example, and the embodiments of the present application are not limited thereto.

[0167] Step 430, in the specified candidate content recommendation model, perform a concatenation operation on the operator selection results corresponding to each substructure to obtain the model encoding corresponding to the specified candidate content recommendation model.

[0168] Among them, the specified candidate content recommendation model is any one of the at least two generated candidate content recommendation models. That is, the composition of the model encoding of any one candidate content recommendation model among the at least two candidate content recommendation models will be used as an example for illustration.

[0169] Optionally, when the operator selection result is implemented in the form of an encoding bit as shown above, perform a concatenation operation on the operator selection results corresponding to different substructures, thereby generating at least two candidate content recommendation models in the search space.

[0170] For example, take the candidate content recommendation model composed of the above input encoding sub-structure, embedding function sub-structure, interaction function sub-structure, and prediction function sub-structure as an example for illustration. After selecting the operator "encoding representation of interaction data pairs obtained based on the multi-hot encoding method" in the input encoding sub-structure, the operator selection result corresponding to the input encoding sub-structure is (0, 1); after selecting the operator "MLP" in the embedding function sub-structure, the operator selection result corresponding to the embedding function sub-structure is (0, 1); after selecting the operator "MUL" in the interaction function sub-structure, the operator selection result corresponding to the interaction function sub-structure is (1, 0, 0, 0, 0); after selecting the operator "MLP" in the prediction function sub-structure, the operator selection result corresponding to the prediction function sub-structure is (0, 0, 1). After performing a splicing operation on the operator selection result (0, 1) corresponding to the input encoding sub-structure, the operator selection result (0, 1) corresponding to the embedding function sub-structure, the operator selection result (1, 0, 0, 0, 0) corresponding to the interaction function sub-structure, and the operator selection result (0, 0, 1) corresponding to the prediction function sub-structure, a candidate content recommendation model is obtained.

[0171] Step 440, perform a splicing operation to obtain model encodings corresponding to at least two candidate content recommendation models in the search space.

[0172] Similarly, after selecting the operator "encoding representation of interaction data pairs obtained based on the multi-hot encoding method" in the input encoding sub-structure, the operator selection result corresponding to the input encoding sub-structure is (0, 1); after selecting the operator "MLP" in the embedding function sub-structure, the operator selection result corresponding to the embedding function sub-structure is (0, 1); after selecting the operator "MAX" in the interaction function sub-structure, the operator selection result corresponding to the interaction function sub-structure is (0, 0, 1, 0, 0); after selecting the operator "VEC" in the prediction function sub-structure, the operator selection result corresponding to the prediction function sub-structure is (0, 1, 0). After performing a splicing operation on the operator selection result (0, 1) corresponding to the input encoding sub-structure, the operator selection result (0, 1) corresponding to the embedding function sub-structure, the operator selection result (0, 0, 1, 0, 0) corresponding to the interaction function sub-structure, and the operator selection result (0, 1, 0) corresponding to the prediction function sub-structure, another candidate content recommendation model is obtained.

[0173] Based on the operator selection results obtained through the above operator selection in different sub-structures, as well as processing operations such as splicing and combination on the operator selection results corresponding to different sub-structures, at least two candidate content recommendation models are generated in the search space. The above is only a schematic example, and the embodiments of the present application are not limited thereto.

[0174] Step 450: Based on the model encodings respectively corresponding to at least two candidate content recommendation models, perform performance prediction on the at least two candidate content recommendation models to obtain the performance prediction results respectively corresponding to the at least two candidate content recommendation models.

[0175] Among them, based on the model encoding, the at least two candidate content recommendation models can be distinguished. Optionally, when the model encoding of the candidate content recommendation model is implemented as the above-mentioned spliced model encoding, based on this model encoding, it is possible to know the operator selection results corresponding to different sub-structures in different candidate content recommendation models, that is: know the parameter operation methods adopted by different sub-structures in different candidate content recommendation models.

[0176] Illustratively, after obtaining the model encodings respectively corresponding to at least two candidate content recommendation models, perform performance prediction on the at least two candidate content recommendation models. For example: predict the analysis results of different sub-structures that make up the candidate content recommendation model, and comprehensively analyze the results of multiple sub-structures to determine the performance prediction result of the candidate content recommendation model; or, arbitrarily select a training data set, input the training data in the training data set into at least two candidate content recommendation models respectively, analyze the training data through the at least two candidate content recommendation models respectively, and determine the performance prediction results respectively corresponding to the at least two candidate content recommendation models according to the difference between the data analysis result and the label marked by the training data. For example: if the prediction difference of the candidate content recommendation model M1 is small, the performance prediction result is good; if the prediction difference of the candidate content recommendation model M2 is large, the performance prediction result is poor, etc.

[0177] It should be noted that the above are only illustrative examples, and the embodiments of the present application are not limited thereto.

[0178] Step 460: Determine a target content recommendation model from the at least two candidate content recommendation models based on the performance prediction results.

[0179] Among them, the target content recommendation model is used to recommend content to the account.

[0180] Optionally, the performance prediction result corresponding to the candidate content recommendation model is represented by a prediction score; or, the performance prediction result corresponding to the candidate content recommendation model is represented by a prediction accuracy rate, etc.

[0181] Illustratively, after obtaining the performance prediction results respectively corresponding to at least two candidate content recommendation models, determine a target content recommendation model from the at least two candidate content recommendation models according to the quality of the performance prediction results. For example: use the n candidate content recommendation models with the best performance prediction results as the target content recommendation models; or, use the 1 candidate content recommendation model with the best performance prediction result as the only target content recommendation model, etc.

[0182] Schematically, when the performance prediction result corresponding to the candidate content recommendation model is represented by a prediction score, the top n candidate content recommendation models with the highest prediction scores are selected therefrom as the target content recommendation models; or, when the performance prediction result corresponding to the candidate content recommendation model is represented by a prediction accuracy rate, the top n candidate content recommendation models with the best prediction accuracy rates are selected therefrom as the target content recommendation models, etc.

[0183] It should be noted that the above are only schematic examples, and the embodiments of the present application are not limited thereto.

[0184] In summary, according to the preset neural network composition structure, not only the range of candidate content recommendation models generated by the search space is limited, but also the limitation of only using existing models for content recommendation can be avoided. In addition, according to the composition mode in the neural network composition structure, the model codes of different candidate content recommendation models are determined, and based on the model codes, the model performance of each candidate content recommendation model is determined, so that a candidate content recommendation model with better recommendation performance can be more quickly and automatically selected from multiple candidate content recommendation models as the target content recommendation model. When content is recommended to an account through the target content recommendation model, appropriate content can be recommended to different accounts more accurately and efficiently, improving the accuracy of content recommendation and enhancing the usage experience of the object.

[0185] In the embodiments of the present application, based on the automated machine learning technology, the operators corresponding to multiple substructures in the search space are described. Different candidate content recommendation models are obtained according to the selection operations of different operators in the substructures and the combination operations of different substructures, and the model codes of different candidate content recommendation models are determined according to the different operator selection results corresponding to the substructures. Thus, the performance prediction model is used to perform performance prediction on the candidate content recommendation models, and a candidate content recommendation model with better recommendation performance is selected as the target content recommendation model, so as to recommend appropriate content to different accounts more accurately and efficiently.

[0186] In an optional embodiment, when performing performance prediction on at least two candidate content recommendation models based on the model codes respectively corresponding to the at least two candidate content recommendation models, the model codes respectively corresponding to the at least two candidate content models are input into the performance prediction model to implement the performance prediction process. Schematically, as Figure 5 shown, step 330 in the above Figure 3 shown embodiment can also be implemented as the following steps 510 to step 560.

[0187] Step 510, input the model codes respectively corresponding to at least two candidate content recommendation models into the performance prediction model to obtain candidate performance prediction results respectively corresponding to the at least two candidate content recommendation models.

[0188] Among them, the performance prediction model is used to indicate a model for predicting at least two candidate content recommendation models. That is, through the performance prediction model, the model performance of at least two candidate content recommendation models is predicted.

[0189] Schematically, the performance prediction model performs a performance prediction process on at least two candidate content recommendation models through the model encodings respectively corresponding to the at least two candidate content recommendation models. For example: taking the model encoding corresponding to the candidate content recommendation model as the input of the performance prediction model, and the performance prediction model predicts the model performance of the candidate content recommendation model based on the different neural network composition structures in the candidate content recommendation model.

[0190] Optionally, the performance prediction model is a prediction model with basic performance prediction capabilities. For example: using a multi-layer perceptron (MLP) as the performance prediction model; or using a tree-based model as the performance prediction model, such as a random forest.

[0191] In an optional embodiment, after obtaining the model encodings respectively corresponding to at least two candidate content recommendation models, the model encodings respectively corresponding to the at least two candidate content recommendation models are input into the performance prediction model, and the performance prediction model respectively predicts the model performance of the at least two candidate content recommendation models, so as to obtain candidate performance prediction results respectively corresponding to the at least two candidate content recommendation models.

[0192] Schematically, taking the performance analysis of the above candidate content recommendation model 1 by the performance prediction model as an example for illustration. The model encoding corresponding to the candidate content recommendation model 1 is input into the performance prediction model, and the performance prediction model determines the model encoding of the candidate content recommendation model 1 based on the operator selection result in the sub-structure constituting the candidate content recommendation model 1.

[0193] For example: the model encoding of the candidate content recommendation model 1 is the splicing processing result of the operator selection results of four sub-structures.

[0194] A recommendation model composed of algorithm A1 in A, algorithm B2 in neural network composition structure B, algorithm C2 in neural network composition structure C, and algorithm D1 in neural network composition structure D. In candidate content recommendation model 1, the structure encoding representation of neural network composition structure A is (1, 0), the structure encoding representation of neural network composition structure B is (0, 1), the structure encoding representation of neural network composition structure C is (0, 1, 0), the structure encoding representation of neural network composition structure D is (0, 1), and the structure encoding representation of neural network composition structure D is (0, 1). Concatenate the structure encodings corresponding to different neural network composition structures in candidate content recommendation model 1 to obtain the model encoding corresponding to candidate content recommendation model 1.

[0195] In an alternative embodiment, after obtaining the candidate content recommendation model composed of different neural network composition structures, encode the candidate content recommendation model to obtain the model encoding corresponding to the candidate content recommendation model.

[0196] Illustratively, after obtaining the above candidate content recommendation model 1, encode candidate content recommendation model 1. For example, the encoding result is 001, and use the encoding result 001 corresponding to candidate content recommendation model 1 as the model encoding corresponding to candidate content recommendation model 1. Similarly, after obtaining the above candidate content recommendation model 2, encode candidate content recommendation model 2. For example, the encoding result is 010, and use the encoding result 010 corresponding to candidate content recommendation model 2 as the model encoding corresponding to candidate content recommendation model 2, etc. Among them, the encoding results corresponding to different candidate content recommendation models are different and are used to distinguish different candidate content recommendation models.

[0197] In an alternative embodiment, the candidate performance prediction results are used to train the performance prediction model to obtain the target performance prediction model after training the performance prediction model. That is: based on the candidate performance prediction results corresponding to at least two candidate content recommendation models respectively, train the performance prediction model to obtain the target performance prediction model.

[0198] Among them, the target performance prediction model is used to determine the target content recommendation model from at least two candidate content recommendation models.

[0199] Step 520, based on the candidate performance prediction results corresponding to at least two candidate content recommendation models respectively, determine the content recommendation model to be trained from at least two candidate content recommendation models.

[0200] Optionally, after obtaining the candidate performance prediction results corresponding to at least two candidate content recommendation models, compare the candidate performance prediction results corresponding to the at least two candidate content recommendation models, and based on the comparison results, determine the content recommendation model to be trained from the at least two candidate content recommendation models.

[0201] Illustratively, the candidate performance prediction result is implemented as a performance prediction score, and the top n candidate content recommendation models with the highest performance prediction scores are used as the content recommendation models to be trained; alternatively, the candidate performance prediction result is implemented as a performance prediction accuracy rate, and the top n candidate content recommendation models with the highest performance prediction accuracy rates are used as the content recommendation models to be trained.

[0202] That is: after determining the candidate performance prediction results corresponding to at least two candidate content recommendation models, differentially select whether to train the candidate content recommendation models according to the performance prediction situations corresponding to different candidate content recommendation models. When the candidate performance prediction result corresponding to a candidate content recommendation model is good, use this candidate content recommendation model as the content recommendation model to be trained, so that in the subsequent process, a training process is performed on this candidate content recommendation model; when the candidate performance prediction result corresponding to a candidate content recommendation model is poor, do not perform a training process on this candidate content recommendation model, thereby avoiding an additional training process for a candidate content recommendation model with poor prediction performance and effectively reducing the computational amount during the training process.

[0203] Step 530, train the content recommendation model to be trained through the sample interaction dataset to obtain a training analysis model corresponding to the content recommendation model to be trained.

[0204] Among them, the sample interaction dataset stores sample interaction data, and the sample interaction data is labeled with a sample interaction data label.

[0205] Optionally, the sample interaction data is a data pair composed of sample account data and sample content data, and the sample interaction data label is used to indicate the historical interaction situation between the sample account data and the sample content data during the historical interaction process.

[0206] For example: in the sample interaction dataset, the sample interaction data S1 indicates that there is a historical interaction between the sample account data ID1 and the sample content data CON1, and the sample interaction data S2 indicates that there is a historical interaction between the sample account data ID1 and the sample content data CON2, etc.

[0207] In an optional embodiment, obtain sample interaction data from the sample interaction dataset; perform interaction analysis on the sample interaction data through the content recommendation model to be trained to determine a data interaction prediction result corresponding to the content recommendation model to be trained.

[0208] Schematically, from the sample interaction dataset, at least one sample interaction data is arbitrarily selected, and the sample interaction data is input into the content recommendation model to be trained, and the sample interaction data is analyzed through the content recommendation model to be trained.

[0209] Optionally, when analyzing the sample interaction data through the content recommendation model to be trained, the content recommendation model to be trained analyzes the sample account data and the sample content data corresponding to the input sample interaction data respectively, so as to predict the interaction situation between the sample account data and the sample content data, and obtain a data interaction prediction result corresponding to the sample interaction data.

[0210] Among them, the data interaction prediction result is used to indicate the interaction prediction situation of the sample account data and the sample content data in the sample interaction data by the content recommendation model to be trained.

[0211] For example: Taking the sample interaction data S1 obtained from the sample interaction dataset as an example of the input of the content recommendation model to be trained for analysis. When the content recommendation model to be trained analyzes the sample interaction data S1, it determines the sample account data ID1 and the sample content data CON1 corresponding to the sample interaction data S1, and predicts the interaction situation between the sample account data ID1 and the sample content data CON1, and obtains a data interaction prediction result corresponding to the sample interaction data S1; Similarly, when the content recommendation model to be trained analyzes the sample interaction data S2, it determines the sample account data ID1 and the sample content data CON2 corresponding to the sample interaction data S2, and predicts the interaction situation between the sample account data ID1 and the sample content data CON2, and obtains a data interaction prediction result corresponding to the sample interaction data S2, etc.

[0212] It should be noted that the above is only a schematic example, and the embodiments of the present application are not limited thereto.

[0213] In an optional embodiment, based on the difference between the sample interaction data label and the data interaction prediction results respectively corresponding to at least two candidate content recommendation models, the content recommendation model to be trained is trained to obtain a training analysis model corresponding to the content recommendation model to be trained.

[0214] Among them, the sample interaction data label is an identifier corresponding to the sample interaction data. Through the sample interaction data label, the historical interaction situation between the sample account data and the sample content data in the sample interaction data can be determined.

[0215] Optionally, after obtaining the data interaction prediction results corresponding to at least two candidate content recommendation models respectively, since the data interaction prediction results correspond to the sample interaction data, and the sample interaction data has its corresponding sample interaction data label, the differences between at least two data interaction prediction results and the corresponding sample interaction data labels are determined.

[0216] Illustratively, the sample interaction data input into the content recommendation model to be trained is sample interaction data S1, and the sample interaction data label corresponding to the sample interaction data S1 is L1. By analyzing the interaction relationship between the sample account data ID1 and the sample content data CON1 corresponding to the sample interaction data S1 through the content recommendation model to be trained, the data interaction prediction result P1 corresponding to the sample interaction data S1 is obtained. Then, the difference between the sample interaction data label L1 and the data interaction prediction result P1 is determined. For example: using the cross-entropy loss function, the loss value between the sample interaction data label being L1 and the data interaction prediction result P1 is determined.

[0217] Optionally, based on the above prediction method, the data interaction prediction results corresponding to different sample interaction data after being input into the content recommendation model to be trained are determined, and the loss values corresponding to different sample interaction data are determined according to the data interaction prediction results and the corresponding sample interaction data labels.

[0218] Optionally, based on the loss value between the sample interaction data label and at least two data interaction prediction results, the content recommendation model to be trained is trained to obtain a training analysis model corresponding to the content recommendation model to be trained.

[0219] Illustratively, after obtaining the loss value, with the goal of reducing the loss value, the content recommendation model to be trained is trained. In response to the training of the content recommendation model to be trained reaching the training goal, a training analysis model is obtained. For example: in response to the loss value reaching the convergence state, the content recommendation model to be trained obtained from the last iterative training is used as the training analysis model.

[0220] Illustratively, the loss value reaching the convergence state is used to indicate that the value of the loss value obtained through the loss function no longer changes or the change amplitude is less than the preset threshold. For example: the content recommendation model to be trained reaches the state where the loss function no longer continues to decrease during the training process, etc.

[0221] Step 540, perform interaction analysis on the sample interaction data in the sample interaction dataset respectively through the training analysis model to obtain a prediction analysis result corresponding to the training analysis model.

[0222] Schematically, after the training of the content recommendation model to be trained reaches the training objective, a training analysis model is obtained. By performing interaction analysis on at least one sample interaction data in the sample interaction dataset through this training analysis model, a prediction analysis result corresponding to the training analysis model is obtained.

[0223] Among them, the prediction analysis result is used to indicate the prediction result of the training analysis model for the sample account data and sample content data in the sample interaction data.

[0224] Schematically, taking the interaction analysis of sample interaction data S1 and sample interaction data S2 by a training analysis model as an example for illustration. This training analysis model predicts the interaction relationship between the sample account data ID1 and the sample content data CON1 in the sample interaction data S1, and obtains a prediction analysis result PA1 corresponding to the sample interaction data S1; this training analysis model predicts the interaction relationship between the sample account data ID1 and the sample content data CON2 in the sample interaction data S2, and obtains a prediction analysis result PA2 corresponding to the sample interaction data S2. The prediction analysis result PA1 corresponding to the sample interaction data S1 and the prediction analysis result PA2 corresponding to the sample interaction data S2 are used as the prediction analysis result corresponding to the training analysis model.

[0225] Optionally, after obtaining the prediction analysis result PA1 corresponding to the sample interaction data S1 and the prediction analysis result PA2 corresponding to the sample interaction data S2, comprehensively analyze the above prediction analysis results PA1 and PA2, and use the comprehensively analyzed result as the prediction analysis result corresponding to the training analysis model. For example: taking the average value of the prediction analysis result PA1 and the prediction analysis result PA2 as the prediction analysis result corresponding to this training analysis model; or taking the maximum value of the prediction analysis result PA1 and the prediction analysis result PA2 as the prediction analysis result corresponding to this training analysis model, etc.

[0226] Step 550, training the performance prediction model with the prediction analysis results respectively corresponding to at least two training analysis models to obtain the target performance prediction model.

[0227] Schematically, after obtaining the prediction analysis result corresponding to the training analysis model, select at least two training analysis models and determine the prediction analysis results respectively corresponding to the at least two training analysis models.

[0228] Optionally, taking the prediction analysis result corresponding to the training analysis model being implemented as the comprehensively analyzed result as an example for illustration. In order to efficiently search for the target content recommendation model that can achieve good recommendation performance from the search space, it is not necessary to estimate the numerical values of the feature indicators very precisely, but to distinguish the relative advantages and disadvantages of multiple candidate content recommendation models.

[0229] Schematically, after obtaining the prediction analysis results corresponding to at least two training analysis models, a pairwise loss function is used to optimize and train the performance prediction model, that is: the goal of this performance prediction model is to rank the recommendation performances of at least two training analysis models obtained by training, such as: ranking from good to bad. Among them, the pairwise loss function is as follows.

[0230]

[0231] Among them, L P is used to indicate the pairwise loss value of the pairwise loss function; x + is used to indicate the training analysis model with a better prediction analysis result; P(x + ) is used to indicate the better prediction analysis result; x - is used to indicate the training analysis model with a worse prediction analysis result; P(x - ) is used to indicate the worse prediction analysis result; O is used to indicate at least two training analysis models; σ is used to indicate the activation function.

[0232] Schematically, after ranking the prediction analysis results corresponding to at least two training analysis models from good to bad, the above pairwise loss function is used to determine the pairwise loss values corresponding to each pair among at least two training analysis models, and the performance prediction model is trained through the pairwise loss values to obtain the target performance prediction model.

[0233] Optionally, after obtaining the pairwise loss value, the performance prediction model is trained with the goal of reducing the pairwise loss value. In response to the pairwise loss value reaching the convergence state, the performance prediction model obtained from the most recent iterative training is used as the target performance prediction model.

[0234] Schematically, the pairwise loss value reaching the convergence state is used to indicate that the numerical value of the pairwise loss value obtained through the pairwise loss function no longer changes or the change amplitude is less than a preset threshold. For example: k pairwise loss values no longer continue to decrease; or, the change amplitude of k pairwise loss values is less than a preset threshold, where k is a positive integer.

[0235] Among them, the target performance prediction model is used to select a candidate content recommendation model with better prediction performance from at least two candidate content recommendation models corresponding to the search space.

[0236] Step 560, perform performance prediction on at least two candidate content recommendation models through the target performance prediction model to obtain the performance prediction results corresponding to at least two candidate content recommendation models respectively.

[0237] Schematically, after obtaining the target performance prediction model, the performance of at least two candidate content recommendation models corresponding to the search space is predicted respectively by the target performance prediction model, so as to determine the performance prediction results corresponding to the at least two candidate content recommendation models respectively.

[0238] In an optional embodiment, a target content recommendation model is predicted by the target performance prediction model.

[0239] Schematically, the n candidate content recommendation models with better performance prediction results are used as the target content recommendation models, so as to realize the content recommendation process with the help of the target content recommendation models, that is: to realize the process of recommending content data to an account.

[0240] For example: when n is 1, the candidate content recommendation model with the best performance prediction result is used as the target content recommendation model, and according to the analysis results of the account data to be analyzed and the content data to be analyzed by the target content recommendation model, the corresponding content data is recommended to different accounts, such as: recommending movies that match their preferences to different accounts; when n is a positive integer greater than 1, the n candidate content recommendation models with better performance prediction results are used as the target content recommendation models, the account data to be analyzed and the content data to be analyzed are analyzed respectively by the n target content recommendation models, and the analysis results corresponding to the n target content recommendation models are determined, and the analysis results of the n target content recommendation models are synthesized to recommend the corresponding content data to different accounts.

[0241] It should be noted that the above is only a schematic example, and the embodiments of the present application are not limited thereto.

[0242] To sum up, according to the preset neural network composition structure, not only the range of candidate content recommendation models generated by the search space is limited, but also the limitation of only using the existing models for content recommendation can be avoided. In addition, according to the composition mode in the neural network composition structure, the model codes of different candidate content recommendation models are determined, and the model performance of each candidate content recommendation model is determined based on the model codes, so that a candidate content recommendation model with better recommendation performance can be selected more quickly and automatically from multiple candidate content recommendation models as the target content recommendation model. When recommending content to an account through the target content recommendation model, appropriate content can be recommended to different accounts more accurately and efficiently, improving the accuracy of content recommendation and enhancing the user experience of the object.

[0243] In the embodiment of the present application, the training process of the performance prediction model is introduced. After obtaining the candidate content recommendation model, the model encoding corresponding to the candidate content recommendation model is input into the performance prediction model to obtain candidate performance prediction results corresponding to different candidate content recommendation models. According to the candidate performance prediction results, it is differentially determined whether to train the candidate content recommendation model, thereby avoiding additional training processes for candidate content recommendation models with poor prediction performance and effectively reducing the computational amount during the training process. After obtaining the trained training analysis model, the performance prediction model is trained through the sample interaction dataset to obtain the target performance prediction model. Through the target performance prediction model, a target content recommendation model with better recommendation performance can be obtained more quickly, avoiding the situation where a model can only correspond to one training dataset through the separate training method. Through the target performance prediction model, multiple models can be determined simultaneously.

[0244] In an alternative embodiment, the above-mentioned candidate content recommendation model is referred to as a collaborative filtering model, and an example is given to illustrate the prediction effect between objects and items. Schematically, the generation method of the content recommendation model includes two processes: the generation and selection of the collaborative filtering model, and the above-mentioned generation method of the content recommendation model can also be implemented as the following two parts, namely: (1) search space design; (2) search strategy design.

[0245] (1) Search space design

[0246] (1) Operator Selection

[0247] a. Input encoding: Schematically, for the encoding representation of objects and items, an intuitive way is to use the identity document (ID), that is, the ID is encoded using the one-hot method to obtain two embedding matrices, namely the embedding matrix corresponding to the object and the embedding matrix corresponding to the item; alternatively, the historical interaction method is adopted, and the object is encoded using the Multi-hot method of the object, that is, the object is represented by the items that have interacted with the object. Similarly, the item can be represented in a similar way. The encoding (one-hot or Multi-hot) here is used to represent the object and the item for the subsequent model.

[0248] b. Embedding function: Schematically, the embedding function is used to project the high-dimensional encoding obtained by encoding the input into a low-dimensional space to obtain an embedding vector, and its selection is closely related to the input encoding in the previous step. For the input encoding of the ID type, the embedding function compatible with it is the object / item embedding matrix lookup function (ID-look-up), denoted as MAT; for the input encoding of the historical interaction type, there are two embedding functions compatible with it. The first type is based on the embedding matrix lookup and mean pooling operations, also denoted as MAT; the second type is to convert the multi-hot history into a dense vector based on a multi-layer perceptron (MLP), called MLP. The design of MLP can also have various structures. Schematically, the settings in the neural network collaborative filtering model NeuMF are adopted.

[0249] c. Interaction function: The interaction function takes the embedding vectors of objects and items as inputs and is used for subsequent prediction. Optionally, the adopted interaction function is: multiplying the two vectors item by item to generate a combined vector. Among them, the physical meaning of the interaction function is to match objects with items. Schematically, operations similar to multiplying item by item also include subtraction, taking the maximum value, taking the minimum value, and concatenation, etc.

[0250] d. Prediction function: The prediction function is used to convert the output of the above interaction function into a final prediction score. The higher the score, the higher the interaction probability between the object and the item. Optionally, the summation method (denoted as SUM) is adopted to determine the prediction score. In addition, the inner product of the weighted vector can also be used to assign weights to different dimensions, denoted as VEC; the prediction result can also be obtained using a multi-layer perceptron, denoted as MLP.

[0251] (2) Complete Model Encoding

[0252] Schematically, the above operator selections at each stage are represented by encoding. To make the space more general, in addition to the process of operator selection using the above four stages, consider selecting the embedding dimension of the model from a predefined set S dim and selecting the optimizer learning rate from a predefined set S lr as part of the encoding. Therefore, the size of the entire space has to be multiplied by |S dim *S lr |.

[0253] Optionally, the encoding here adopts the one-hot encoding form. For example: the encoding (0, 1, 0, 0, 0) means that among the 5 operator selections for a certain function, the 2nd operator is selected. Schematically, the operator selection at each stage, plus the embedding dimension selection and the optimizer learning rate selection, thus obtains the encoding representation of the collaborative filtering model, and the subsequent model search process is based on this encoding representation.

[0254] In an optional embodiment, as shown in Table 1 below, the operators in the above search space are represented.

[0255] Table 1

[0256]

[0257] Among them, the search space includes four stages, and the search range of operators needs to be defined for each stage. Therefore, the search space can be understood as a combination of different operators. As mentioned above, in practical applications, a set of hyperparameters for the embedding size and learning rate can also be defined, so that the search space is further expanded within a certain range.

[0258] As Figure 6 shown, it is a schematic diagram of the search space of the collaborative filtering model. The data is input into the search space, and the objects i and items j in the data are analyzed respectively. Through the input encoding 610 in the search space, the matrix c corresponding to the object i is determined i and the matrix c corresponding to the item j j , where the operator includes two cases: ID or history; in addition, through the embedding function 620, the vector e corresponding to the object i is determined user and the vector e corresponding to the item j item ; then, through the interaction function 630, the interaction analysis is performed on the vector corresponding to the object and the vector corresponding to the item. For example: through the interaction function 630, the vector e corresponding to the object i i and the vector e corresponding to the item j j are subjected to interaction analysis; finally, through the prediction function 640, the output vector is obtained.

[0259] (2) Search Strategy Design

[0260] After defining the search space, an effective and appropriate search strategy is designed to find a collaborative filtering model with excellent performance.

[0261] Schematically, to address the challenge of search efficiency, an easy-to-use and robust search strategy is adopted, which combines the random search algorithm with a model performance predictor (performance prediction model).

[0262] Among them, the reason for adopting the random search method is to consider the discreteness of the search space, and the random search method has a certain effectiveness. In addition, to make the search process more efficient, the random search is combined with performance prediction.

[0263] Schematically, the model performance predictor is used to distinguish the quality of the collaborative filtering model, that is: whether the collaborative filtering model to be analyzed is a good model or a bad model. Among them, the input of the model performance predictor is the model encoding representation shown above - x o, where o ranges from 1 to 6.

[0264] Optionally, a multi-layer perceptron or a tree-based model (such as a random forest) is used as the model performance predictor. Among them, the multi-layer perceptron supports parameter updates based on gradient descent, is more compatible with random search, and has a stronger ability to learn from complex data.

[0265] P(x0) = MLP(Concat(x0))

[0266] Among them, P(x0) is used to represent the model prediction performance; MLP is used to indicate the multi-layer perceptron; Concat is used to indicate the concatenation operation; x0 is used to indicate the model encoding representation. Among them, the predicted model prediction performance can be any metric for a given task, including common regression tasks for explicit data and ranking tasks for implicit data.

[0267] In order to efficiently search for a collaborative filtering model that can achieve good recommendation performance, it is not necessary to estimate the numerical values of the feature metrics very precisely, and it is only necessary to be able to distinguish the relative advantages and disadvantages of multiple models. Therefore, a pairwise loss function is used to optimize and train the model performance predictor, that is: the goal of this predictor is to rank the recommendation performance of a given number of recommendation models from best to worst.

[0268] In short, as Figure 7 shown, it is a schematic diagram of the workflow for obtaining the target model performance predictor. First, the random search 710 method is used to sample multiple collaborative filtering models from the complete search space; then, the model performance predictor 720 is used to predict the performance of multiple collaborative filtering models, and the model selection 730 is performed through the predicted performance, and several collaborative filtering models with the best estimated performance are selected from multiple collaborative filtering models for training data evaluation, that is: the publicly available and pre-obtained training data set is used to train and evaluate 740 several collaborative filtering models until the collaborative filtering models reach a convergence state, so as to obtain the recommendation effects of several collaborative filtering models on the training data set; finally, using this recommendation result, the parameters of the model performance predictor are updated 750 according to the above loss function. This workflow is repeated until a collaborative filtering model with sufficient recommendation effect on the given task is searched. When the termination condition (for example: the stop condition is set as a hyperparameter) is reached, the repetition of the workflow stops, that is, an effective target model performance predictor (the required model) is obtained. Through this target model performance predictor, the model performance of various collaborative filtering models in the search space can be predicted with high precision, and an effective collaborative filtering model suitable for the content recommendation task can be found from the search space.

[0269] Schematically, first, two hyperparameters, K1 and K2, are defined. The model performance predictor estimates and ranks K1 + K2 collaborative filtering models, and selects the top K1 collaborative filtering models in the sorted list to participate in the training, that is: obtain the recommendation performance results of the K1 collaborative filtering models, and update the parameters of the model performance predictor through this result.

[0270] Optionally, the algorithm implementation steps of the automated collaborative filtering model search are as follows.

[0271] The algorithm inputs are: search space F, learnable model performance predictor P, model recommendation evaluation metric M, empty set H, search set hyperparameters K1 and K2, and training data S.

[0272] (1) Perform a random initialization operation on the model performance predictor P; (2) Repeat the workflow as Figure 7 shown; (3) Randomly select a model set of K1 + K2 collaborative filtering models from the set F; (4) Generate model encodings for all collaborative filtering models in this model set; (5) Input the model encodings into the model performance predictor to obtain the model performances corresponding to different collaborative filtering models; (6) Select the Top-K1 (first K1) collaborative filtering models and use the training data S to train the K1 collaborative filtering models; (7) Evaluate the model performances of the K1 collaborative filtering models through the model recommendation evaluation metric M; (8) Update the empty set H, that is: add the newly evaluated K1 collaborative filtering models and the model performances corresponding to the K1 collaborative filtering models respectively; (9) Update the parameters of the predictor P through the model encodings of the collaborative filtering models stored in the set H and the model performances corresponding to different collaborative filtering models; (10) Until the termination condition is met; (11) Return the H set, and obtain at least one collaborative filtering model as the target collaborative filtering model from the K1 + K2 collaborative filtering models, so as to perform the content recommendation process through the target collaborative filtering model.

[0273] In summary, according to the preset neural network composition structure, not only the range of candidate content recommendation models generated by the search space is limited to conform to the network model paradigm of the content recommendation system, but also the limitations of only using existing models for content recommendation can be avoided. In addition, according to the composition pattern in the neural network composition structure, determine the model encodings of different candidate content recommendation models, and based on the model encodings, determine the model performances of each candidate content recommendation model, so as to be able to more quickly and automatically select a candidate content recommendation model with better recommendation performance as the target content recommendation model from multiple candidate content recommendation models. When performing content recommendation to an account through the target content recommendation model, it is possible to more accurately and efficiently recommend suitable content for different accounts, improve the accuracy of content recommendation, and enhance the usage experience of the object.

[0274] Figure 8 It is a generating device for a content recommendation model provided by an exemplary embodiment of the present application. As Figure 8 shown, the device includes the following parts:

[0275] A generating module 810, configured to generate at least two candidate content recommendation models in a search space, where the search space includes a preset neural network composition structure, and the candidate content recommendation models are candidate models for content recommendation analysis;

[0276] An obtaining module 820, configured to obtain model codes respectively corresponding to the at least two candidate content recommendation models in the search space, where the model codes are used to indicate the composition modes of the neural network composition structures in the candidate content recommendation models;

[0277] A predicting module 830, configured to perform performance prediction on the at least two candidate content recommendation models based on the model codes respectively corresponding to the at least two candidate content recommendation models, and obtain performance prediction results respectively corresponding to the at least two candidate content recommendation models;

[0278] A determining module 840, configured to determine a target content recommendation model from the at least two candidate content recommendation models based on the performance prediction results, where the target content recommendation model is used to recommend content to an account.

[0279] In an optional embodiment, the predicting module 830 is further configured to input the model codes respectively corresponding to the at least two candidate content recommendation models into a performance prediction model to obtain candidate performance prediction results respectively corresponding to the at least two candidate content recommendation models; based on the candidate performance prediction results respectively corresponding to the at least two candidate content recommendation models, train the performance prediction model to obtain a target performance prediction model, where the target performance prediction model is used to determine the target content recommendation model from the at least two candidate content recommendation models; perform performance prediction on the at least two candidate content recommendation models through the target performance prediction model, and obtain performance prediction results respectively corresponding to the at least two candidate content recommendation models.

[0280] In an alternative embodiment, the prediction module 830 is further configured to determine a content recommendation model to be trained from the at least two candidate content recommendation models based on the candidate performance prediction results respectively corresponding to the at least two candidate content recommendation models; train the content recommendation model to be trained with a sample interaction data set to obtain a training analysis model corresponding to the content recommendation model to be trained, where the sample interaction data set stores sample interaction data, and the sample interaction data is labeled with a sample interaction data label; perform interaction analysis on the sample interaction data in the sample interaction data set through the training analysis model to obtain prediction analysis results respectively corresponding to the at least two candidate content recommendation models; and train the performance prediction model with the prediction analysis results respectively corresponding to the at least two candidate content recommendation models to obtain the target performance prediction model.

[0281] In an alternative embodiment, the prediction module 830 is further configured to obtain sample interaction data from the sample interaction data set; perform interaction analysis on the sample interaction data through the content recommendation model to be trained to determine a data interaction prediction result corresponding to the content recommendation model to be trained; and train the content recommendation model to be trained based on the difference between the sample interaction data label and the data interaction prediction results respectively corresponding to the at least two candidate content recommendation models to obtain the training analysis model corresponding to the content recommendation model to be trained.

[0282] In an alternative embodiment, the sample interaction data is a data pair composed of sample account data and sample content data, and the sample interaction data label is used to indicate the historical interaction situation between the sample account data and the sample content data during the historical interaction process.

[0283] In an alternative embodiment, the neural network composition structure includes a plurality of sub-structures;

[0284] The generation module 810 is further configured to determine operator selection results respectively corresponding to the plurality of sub-structures, where the operator selection results are used to indicate the operators to be used when performing parameter operations inside the sub-structures; and generate at least two candidate content recommendation models in the search space based on the operator selection results respectively corresponding to the plurality of sub-structures.

[0285] In an alternative embodiment, the search space generates the at least two candidate content recommendation models through a historical interaction data set and the preset neural network composition structure;

[0286] The historical interaction data set stores at least one of account data, content data, and interaction data pairs, where the interaction data pair is used to indicate that there is a historical interaction relationship between at least one account data and one content data.

[0287] In an optional embodiment, the generating module 810 is further configured to, in response to the input encoding sub-structure being included in the multiple sub-structures, obtain a first matrix representation corresponding to the first type of data and a second matrix representation corresponding to the second type of data; or, obtain an account interaction matrix representation corresponding to the account data and a content interaction matrix representation corresponding to the content data in the interaction data pair, wherein the operator selection result of the input encoding sub-structure is determined based on the matrix representation obtaining method; in response to the embedding function sub-structure being included in the multiple sub-structures, project the first matrix representation and the second matrix representation into a vector space to obtain a first embedding vector corresponding to the first matrix representation and a second embedding vector corresponding to the second matrix representation; or, project the account interaction matrix representation corresponding to the account data and the content interaction matrix representation corresponding to the content data in the interaction data pair into a vector space to obtain a third embedding vector corresponding to the account interaction matrix representation and a fourth embedding vector corresponding to the content interaction matrix representation, wherein the operator selection result of the embedding function structure is determined based on the embedding vector obtaining method; in response to the interaction function sub-structure being included in the multiple sub-structures, multiply the first embedding vector and the second embedding vector to obtain a combined vector; or, perform an interaction process on the third embedding vector and the fourth embedding vector to obtain a combined vector, where the combined vector is used to indicate the predicted interaction relationship between the account data and the content data, and the operator selection result of the interaction function structure is determined based on the combined vector obtaining method; in response to the prediction function sub-structure being included in the multiple sub-structures, perform an interaction prediction analysis on the combined vector to obtain a prediction result, where the prediction result is used to indicate the difference between the predicted interaction relationship and the historical interaction relationship, and the operator selection result of the prediction function structure is determined based on the interaction prediction analysis method.

[0288] In an optional embodiment, the multiple sub-structures include an input encoding sub-structure, an embedding function sub-structure, an interaction function sub-structure, and a prediction function sub-structure;

[0289] The generating module 810 is further configured to perform a combination operation on the operator selection results corresponding to the input encoding sub-structure, the embedding function sub-structure, the interaction function sub-structure, and the prediction function sub-structure respectively, and generate at least two candidate content recommendation models in the search space.

[0290] In an optional embodiment, the obtaining module 820 is further configured to perform a splicing operation on the operator selection results corresponding to each sub-structure in a specified candidate content recommendation model to obtain a model encoding corresponding to the specified candidate content recommendation model; and use the splicing operation to obtain the model encodings corresponding to the at least two candidate content recommendation models in the search space.

[0291] In summary, in the search space, candidate content recommendation models are generated based on different preset neural network composition structures. Through the composition mode of the neural network composition structure, the model encodings corresponding to the candidate content recommendation models are obtained, so as to predict the performance of the candidate content recommendation models, and to implement the process of determining the target content recommendation model from the candidate content recommendation models. According to the preset neural network composition structure, not only the range of candidate content recommendation models generated by the search space is limited, but also the limitation of only using existing models for content recommendation can be avoided. In addition, according to the composition mode in the neural network composition structure, the model encodings of different candidate content recommendation models are determined, and the model performance of each candidate content recommendation model is determined based on the model encodings, so that a candidate content recommendation model with better recommendation performance can be more quickly and automatically selected from multiple candidate content recommendation models as the target content recommendation model. When content is recommended to an account through the target content recommendation model, appropriate content can be recommended to different accounts more accurately and efficiently, improving the accuracy of content recommendation and enhancing the user experience of the object.

[0292] It should be noted that: the content recommendation model generation device provided in the above embodiment is only illustrated by the division of the above functional modules. In actual application, the above functions can be allocated to different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the content recommendation model generation device provided in the above embodiment and the content recommendation model generation method embodiment belong to the same concept, and the specific implementation process can be seen in the method embodiment, which will not be repeated here.

[0293] Figure 9 FIG. shows a schematic structural diagram of a server provided by an exemplary embodiment of the present application. 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 central processing unit 901. The server 900 also includes a mass storage device 906 for storing an operating system 913, an application program 914, and other program modules 915.

[0294] The mass storage device 906 is connected to the central processing unit 901 through a mass storage controller (not shown) connected to the system bus 905. The mass storage device 906 and its associated computer-readable medium provide non-volatile storage for the server 900. That is to say, 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.

[0295] Without loss of generality, computer-readable media may include computer storage media and communication media. Computer storage media includes 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 includes RAM, ROM, erasable programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), flash memory or other solid state storage technologies, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape cartridges, tapes, disk storage or other magnetic storage devices. Of course, those skilled in the art will appreciate that computer storage media is not limited to the above several types. The above-mentioned system memory 904 and mass storage device 906 may be collectively referred to as memory.

[0296] According to various embodiments of the present application, the server 900 may also run by connecting to a remote computer on the network through a network such as the Internet. That is, the server 900 may be connected to the network 912 through the network interface unit 911 connected to the system bus 905, or in other words, the network interface unit 911 may also be used to connect to other types of networks or remote computer systems (not shown).

[0297] The above-mentioned memory further includes one or more programs, and the one or more programs are stored in the memory and configured to be executed by the CPU.

[0298] Embodiments of the present application further provide a computer device, which includes a processor and a memory. The memory stores at least one instruction, at least one segment of program, code set or instruction set, and the at least one instruction, at least one segment of program, code set or instruction set is loaded and executed by the processor to implement the method for generating a content recommendation model provided in the above-mentioned method embodiments.

[0299] An embodiment of the present application further provides a computer-readable storage medium, on which at least one instruction, at least one program, a code set or an instruction set is stored, and the at least one instruction, at least one program, the code set or the instruction set is loaded and executed by a processor to implement the method for generating a content recommendation model provided in each of the above method embodiments.

[0300] An embodiment of the present application further provides a computer program product or a computer program, the computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method for generating a content recommendation model described in any one of the above embodiments.

[0301] Optionally, the computer-readable storage medium may include: read-only memory (ROM, Read Only Memory), random access memory (RAM, Random Access Memory), solid state drive (SSD, Solid State Drives), or optical disc, etc. Among them, the random access memory may include resistive random access memory (ReRAM, Resistance RandomAccess Memory) and dynamic random access memory (DRAM, Dynamic Random Access Memory). The serial numbers of the above embodiments of the present application are only for description and do not represent the advantages and disadvantages of the embodiments.

[0302] Those of ordinary skill in the art can understand that all or part of the steps for implementing the above embodiments can be completed by hardware, or can be completed by a program instructing relevant hardware. The program can be stored in a computer-readable storage medium, and the above-mentioned storage medium can be a read-only memory, a magnetic disk or an optical disc, etc.

[0303] The above are only optional embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for generating a content recommendation model, characterized in that The method includes: Obtain a search space, where the search space includes a preset neural network composition structure; Determine the operator selection results corresponding to multiple substructures in the neural network composition structure. The substructures include an input encoding substructure, an embedding function substructure, an interaction function substructure, and a prediction function substructure. The operator selection result is used to indicate the operator adopted during parameter operation within the substructure; Based on the operator selection results corresponding to the multiple substructures respectively, generate at least two candidate content recommendation models in the search space. The candidate content recommendation models are candidate models for content recommendation analysis, and the candidate content recommendation models are models obtained based on at least one substructure in the neural network composition structure; Obtain the model encodings corresponding to the at least two candidate content recommendation models in the search space. The model encoding is used to indicate the composition mode of the neural network composition structure in the candidate content recommendation model. Among them, the model encoding is obtained by splicing the structure encodings corresponding to each substructure in a single candidate content recommendation model, and the structure encoding represents the operator selection result corresponding to the substructure in the form of one-hot encoding; Perform performance prediction on the at least two candidate content recommendation models based on the model encodings corresponding to the at least two candidate content recommendation models respectively, and obtain the performance prediction results corresponding to the at least two candidate content recommendation models respectively; Determine a target content recommendation model from the at least two candidate content recommendation models based on the performance prediction results. The target content recommendation model is used to recommend content to an account.

2. The method according to claim 1, wherein The performing performance prediction on the at least two candidate content recommendation models based on the model encodings corresponding to the at least two candidate content recommendation models respectively, and obtaining the performance prediction results corresponding to the at least two candidate content recommendation models respectively includes: Input the model encodings corresponding to the at least two candidate content recommendation models respectively into a performance prediction model, and obtain candidate performance prediction results corresponding to the at least two candidate content recommendation models respectively; Based on the candidate performance prediction results corresponding to the at least two candidate content recommendation models respectively, train the performance prediction model to obtain a target performance prediction model. The target performance prediction model is used to determine the target content recommendation model from the at least two candidate content recommendation models; Perform performance prediction on the at least two candidate content recommendation models through the target performance prediction model, and obtain the performance prediction results corresponding to the at least two candidate content recommendation models respectively.

3. The method according to claim 2, characterized in that, The training the performance prediction model based on the candidate performance prediction results corresponding to the at least two candidate content recommendation models respectively to obtain a target performance prediction model includes: Determine the content recommendation model to be trained from the at least two candidate content recommendation models based on the candidate performance prediction results corresponding to the at least two candidate content recommendation models respectively; Train the content recommendation model to be trained through a sample interaction data set to obtain a training analysis model corresponding to the content recommendation model to be trained. The sample interaction data set stores sample interaction data, and the sample interaction data is labeled with sample interaction data labels; Perform interaction analysis on the sample interaction data in the sample interaction data set through the training analysis model to obtain prediction analysis results corresponding to at least two candidate content recommendation models respectively; Train the performance prediction model with the prediction analysis results corresponding to at least two candidate content recommendation models respectively to obtain the target performance prediction model.

4. The method according to claim 3, characterized in that, The training of the content recommendation model to be trained through the sample interaction data set to obtain a training analysis model corresponding to the content recommendation model to be trained includes: Obtain sample interaction data from the sample interaction data set; Perform interaction analysis on the sample interaction data through the content recommendation model to be trained, and determine a data interaction prediction result corresponding to the content recommendation model to be trained; Based on the difference between the sample interaction data label and the data interaction prediction results respectively corresponding to at least two candidate content recommendation models, train the content recommendation model to be trained to obtain the training analysis model corresponding to the content recommendation model to be trained.

5. The method according to claim 4, characterized in that, The sample interaction data is a data pair composed of sample account data and sample content data. The sample interaction data label is used to indicate the historical interaction situation between the sample account data and the sample content data during the historical interaction process.

6. The method according to any one of claims 1 to 5, characterized in that, The search space generates at least two candidate content recommendation models through a historical interaction data set and the preset neural network composition structure; The historical interaction data set stores at least one of account data, content data, and interaction data pairs, where the interaction data pair is used to indicate that there is a historical interaction relationship between at least one account data and one content data.

7. The method according to claim 6, characterized in that Determining the operator selection results corresponding to multiple sub-structures in the neural network composition structure includes: In response to the input coding sub-structure being included in the multiple sub-structures, obtain a first matrix representation corresponding to the first type of data and a second matrix representation corresponding to the second type of data; or, obtain an account interaction matrix representation corresponding to the account data and a content interaction matrix representation corresponding to the content data in the interaction data pair, where the operator selection result of the input coding sub-structure is determined based on the method for obtaining the matrix representation; In response to the embedding function substructure being included in the multiple substructures, project the first matrix representation and the second matrix representation into a vector space to obtain a first embedding vector corresponding to the first matrix representation and a second embedding vector corresponding to the second matrix representation; or project the account interaction matrix representation corresponding to the account data and the content interaction matrix representation corresponding to the content data in the interaction data pair into a vector space to obtain a third embedding vector corresponding to the account interaction matrix representation and a fourth embedding vector corresponding to the content interaction matrix representation, wherein the operator selection result of the embedding function substructure is determined based on the method for obtaining the embedding vector; In response to the interaction function substructure being included in the multiple substructures, multiply the first embedding vector and the second embedding vector to obtain a combined vector; or perform interaction processing on the third embedding vector and the fourth embedding vector to obtain a combined vector, where the combined vector is used to indicate the predicted interaction relationship between the account data and the content data, and the operator selection result of the interaction function substructure is determined based on the method for obtaining the combined vector; In response to the prediction function substructure being included in the multiple substructures, perform interaction prediction analysis on the combined vector to obtain a prediction result, where the prediction result is used to indicate the difference between the predicted interaction relationship and the historical interaction relationship, and the operator selection result of the prediction function substructure is determined based on the interaction prediction analysis method.

8. The method according to any one of claims 1 to 5, characterized in that Based on the operator selection results respectively corresponding to the multiple substructures, generate at least two candidate content recommendation models in the search space, including: Perform a combination operation on the operator selection results respectively corresponding to the input encoding substructure, the embedding function substructure, the interaction function substructure, and the prediction function substructure, and generate at least two candidate content recommendation models in the search space.

9. The method according to claim 8, wherein Obtain the model encodings respectively corresponding to the at least two candidate content recommendation models in the search space, including: In a specified candidate content recommendation model, perform a concatenation operation on the operator selection results respectively corresponding to each substructure to obtain the model encoding corresponding to the specified candidate content recommendation model; Use the concatenation operation to obtain the model encodings respectively corresponding to the at least two candidate content recommendation models in the search space.

10. A generating device for a content recommendation model, characterized in that, The device includes: A generation module, configured to obtain a search space, where the search space includes a preset neural network composition structure; determine operator selection results corresponding to multiple sub-structures in the neural network composition structure, where the sub-structures include an input encoding sub-structure, an embedding function sub-structure, an interaction function sub-structure, and a prediction function sub-structure; the operator selection result is used to indicate an operator to be used when performing parameter operations inside the sub-structure; based on the operator selection results corresponding to the multiple sub-structures, generate at least two candidate content recommendation models in the search space, where the candidate content recommendation models are candidate models for content recommendation analysis, and the candidate content recommendation models are models obtained based on at least one sub-structure in the neural network composition structure; An acquisition module, configured to obtain model encodings corresponding to the at least two candidate content recommendation models in the search space, where the model encodings are used to indicate the composition mode of the neural network composition structure in the candidate content recommendation models; wherein, the model encoding is obtained by splicing structure encodings corresponding to each sub-structure in a single candidate content recommendation model, and the structure encoding represents the operator selection result corresponding to the sub-structure in the form of a one-hot encoding; A prediction module, configured to perform performance prediction on the at least two candidate content recommendation models based on the model encodings corresponding to the at least two candidate content recommendation models, and obtain performance prediction results corresponding to the at least two candidate content recommendation models; A determination module, configured to determine a target content recommendation model from the at least two candidate content recommendation models based on the performance prediction results, where the target content recommendation model is used to recommend content to an account.

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

12. A computer-readable storage medium, characterized in that, At least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement the content recommendation model generation method according to any one of claims 1 to 9.

13. A computer program product, characterized in that, It includes computer instructions, and when the computer instructions are executed by a processor, the content recommendation model generation method according to any one of claims 1 to 9 is implemented.

Images