Recommended object representation method, click-through rate prediction method and apparatus

Abstract

The present application provides a recommended object representation method, a click-through rate prediction method and an apparatus. The recommended object representation method in embodiments comprises: acquiring text data and first non-text data corresponding to each recommended object among a plurality of recommended objects, wherein the text data indicates a first feature of the recommended object described by means of text-type data, and the first non-text data indicates a second feature of the recommended object described by means of non-text-type data; for each recommended object, inputting the text data into a first representation network to obtain a text representation output by the first representation network, and inputting the first non-text data into a second representation network to obtain a non-text representation output by the second representation network; and training the first representation network to align a non-text representation and a text representation corresponding to a same recommended object, and predicting a target click-through rate of the recommended object on the basis of the output of the trained first representation network and the output of the second representation network. By fusing text-type and non-text-type features, the possibility of recommending to users articles in which the users are highly interested is improved.

Description

Recommendation object representation methods, click-through rate prediction methods and devices

[0001] This application claims priority to Chinese Patent Application No. 202410495354.1, filed on April 23, 2024, entitled “Recommendation Object Characterization Method, Click Rate Prediction Method and Apparatus”, the entire contents of which are incorporated herein by reference. Technical Field

[0002] This invention relates to the field of recommendation technology, and in particular to a method for representing recommendation objects, a method for predicting click-through rates, and an apparatus. Background Technology

[0003] Recommender systems analyze users' browsing history and interests to provide personalized item recommendations, presenting items that match their preferences. In recommender systems, ID-based collaborative filtering algorithms have become the most mainstream paradigm for modeling items. Currently, recommendation models based on IDs (userID, itemID, and various categorical IDs), such as DeepFM, PNN, and DCN, are common paradigms. The ID-based paradigm encodes user and item features into sparse one-hot embeddings, which are then transformed into dense embeddings through a lookup process. Through a feature interaction layer, potential relationships between features can be identified.

[0004] However, recommendation models based on the ID paradigm focus on category and numerical features, which may result in lower reference value for the click-through rates of items predicted by the recommendation model. This could lead to the model recommending items that users are less interested in, thus reducing the user experience. Summary of the Invention

[0005] This application provides a method for representing recommended items, a method for predicting click-through rates (CTR), and an apparatus. By aligning the textual and non-textual representations of recommended items, the textual representation of the recommended item incorporates features from different modalities, making the textual representation of the recommended item adaptable to the recommendation task. Subsequently, the CTR of the recommended item is predicted using the aligned textual representation and the non-textual representation of the recommended item, thereby improving the accuracy of CTR prediction and increasing the likelihood of recommending items of high interest to users, thus improving the user experience.

[0006] In a first aspect, embodiments of this application provide a method for representing a recommendation object, including:

[0007] For each of multiple recommended objects, obtain text data and first non-text data. The text data indicates a first feature of the recommended object described by text-type data; the first non-text data indicates a second feature of the recommended object described by non-text-type data. For each recommended object, input the text data corresponding to the recommended object into a first representation network to obtain the text representation of the recommended object output by the first representation network. Input the first non-text data corresponding to the recommended object into a second representation network to obtain the non-text representation of the recommended object output by the second representation network. Based on the non-text representation and text representation corresponding to each recommended object, train the first representation network to align the non-text representation and text representation corresponding to the same recommended object, so that the prediction network connected by the second representation network predicts the target click-through rate of the recommended object based at least on the output of the trained first representation network and the output of the second representation network.

[0008] In this solution, the text representations and non-text representations of recommended items are aligned. The text representations of the recommended items incorporate features from different modalities, making them adaptable to the recommendation task. Subsequently, the click-through rate (CTR) of the recommended items is predicted using the aligned text representations and the corresponding non-text representations. This improves the accuracy of CTR prediction, thereby increasing the likelihood of recommending items of high interest to users and ultimately enhancing the user experience.

[0009] In one possible implementation, the first representation network is a pre-trained language model; the second representation network is a neural network model.

[0010] In one possible implementation, the prediction network interacts with the first and second features, based at least on the outputs of the trained first and second representation networks, to predict the target click-through rate of the recommended object.

[0011] In this scheme, the prediction network can learn collaborative information through the interaction of the first and second features. In summary, by integrating the features of different modalities of the recommended object into the text representation of the recommended object, and combining collaborative information to predict the click-through rate of the recommended object, the accuracy of the click-through rate prediction can be improved, thereby increasing the likelihood of recommending items of high interest to users and thus improving the user experience.

[0012] In one possible implementation, based on the non-textual representation and textual representation corresponding to each recommendation object, a first representation network is trained to align the non-textual representation and textual representation corresponding to the same recommendation object, including: based on the error between the non-textual representation and textual representation corresponding to each recommendation object, training the first representation network to reduce the difference between the textual representation and non-textual representation corresponding to the same recommendation object, and to expand the difference between the textual representation and non-textual representation corresponding to different recommendation objects.

[0013] In this approach, comparative learning can more accurately align textual and non-textual representations.

[0014] In one possible implementation, the second representation network and the prediction network are pre-trained networks, and the training methods for the second representation network and the prediction network include:

[0015] Acquire the second non-text data corresponding to each user from multiple users. The second non-text data indicates the user's third feature. Each user has a corresponding label, which indicates the recommended object to be recommended. For each user, input the first non-text data corresponding to each recommended object and the second non-text data corresponding to the user into a second representation network. Input the output of the second representation network into a prediction network to obtain the target click-through rate (CTR) of each recommended object corresponding to the user predicted by the prediction network. Based on the error between the target CTR of each recommended object corresponding to each user and the label corresponding to each user, train the second representation network and the prediction network.

[0016] In this scheme, by pre-training the second representation network and prediction module of the recommendation model, the recommendation model can learn collaborative information more accurately in the scenario where the prediction network realizes feature interaction.

[0017] In one possible implementation, the method further includes: for each user, inputting the first non-text data corresponding to each recommended object and the second non-text data corresponding to the user into a trained second representation network, inputting the output of the trained second representation network into a prediction network to obtain the target click-through rate of each recommended object corresponding to the user predicted by the prediction network; and fine-tuning the trained second representation network and the prediction network based on the error between the target click-through rate of each recommended object corresponding to each user and the label corresponding to each user.

[0018] In one possible implementation, the prediction network is used to perform feature interactions on the first feature, the second feature, and the third feature.

[0019] Secondly, embodiments of this application provide a recommended object characterization device, the device comprising:

[0020] The acquisition module is used to acquire text data and first non-text data corresponding to each of the multiple recommendation objects. The text data indicates that the first feature of the recommendation object is described by text-type data; the first non-text data indicates that the second feature of the recommendation object is described by non-text-type data.

[0021] The representation module is used to input the text data corresponding to each recommendation object into the first representation network to obtain the text representation of the recommendation object output by the first representation network; and to input the first non-text data corresponding to the recommendation object into the second representation network to obtain the non-text representation of the recommendation object output by the second representation network.

[0022] The training module is used to train a first representation network to align the non-textual and textual representations corresponding to the same recommendation object based on the non-textual and textual representations corresponding to each recommendation object, so that the prediction network connected by the second representation network predicts the target click-through rate of the recommendation object based at least on the outputs of the trained first and second representation networks.

[0023] In this solution, the text representations and non-text representations of recommended items are aligned. The text representations of the recommended items incorporate features from different modalities, making them adaptable to the recommendation task. Subsequently, the click-through rate (CTR) of the recommended items is predicted using the aligned text representations and the corresponding non-text representations. This improves the accuracy of CTR prediction, thereby increasing the likelihood of recommending items of high interest to users and ultimately enhancing the user experience.

[0024] In one possible implementation, the first representation network is a pre-trained language model; the second representation network is a neural network model.

[0025] In one possible implementation, the prediction network interacts with the first and second features, based at least on the outputs of the trained first and second representation networks, to predict the target click-through rate of the recommended object.

[0026] In this scheme, the prediction network can learn collaborative information through the interaction of the first and second features. In summary, by integrating the features of different modalities of the recommended object into the text representation of the recommended object, and combining collaborative information to predict the click-through rate of the recommended object, the accuracy of the click-through rate prediction can be improved, thereby increasing the likelihood of recommending items of high interest to users and thus improving the user experience.

[0027] In one possible implementation, a training module is used to train a first representation network based on the error between the non-text representation and the text representation corresponding to each recommendation object, to reduce the difference between the text representation and the non-text representation corresponding to the same recommendation object, and to expand the difference between the text representation and the non-text representation corresponding to different recommendation objects.

[0028] In this approach, comparative learning can more accurately align textual and non-textual representations.

[0029] In one possible implementation, the second representation network and the prediction network are pre-trained networks. The device further includes: a click-through rate prediction module, used to acquire second non-textual data corresponding to each of the multiple users, the second non-textual data indicating the user's third feature; each user has a corresponding label, the label indicating the recommended object to be recommended; for each user, the first non-textual data corresponding to each recommended object and the second non-textual data corresponding to the user are input into the second representation network, and the output of the second representation network is input into the prediction network to obtain the target click-through rate of each recommended object corresponding to the user predicted by the prediction network; based on the error between the target click-through rate of each recommended object corresponding to each user and the label corresponding to each user, the second representation network and the prediction network are trained.

[0030] In this scheme, by pre-training the second representation network and prediction module of the recommendation model, the recommendation model can learn collaborative information more accurately in the scenario where the prediction network realizes feature interaction.

[0031] In one possible implementation, the device further includes: a fine-tuning module, configured to, for each user, input first non-text data corresponding to each recommended object and second non-text data corresponding to the user into a trained second representation network, input the output of the trained second representation network into a prediction network, and obtain the target click-through rate predicted by the prediction network for each recommended object corresponding to the user; and fine-tune the trained second representation network and the prediction network based on the error between the target click-through rate of each recommended object corresponding to each user and the label corresponding to each user.

[0032] In one possible implementation, the prediction network is used to perform feature interactions on the first feature, the second feature, and the third feature.

[0033] Thirdly, embodiments of this application provide a click-through rate prediction method, including:

[0034] The process involves obtaining the text representation and first non-text data corresponding to the recommended object. The text representation is the output of a first representation network, which is trained using a first method. The first non-text data indicates the second feature of the recommended object described by non-textual data. Next, the process involves obtaining the second non-text data corresponding to the target user, which indicates the third feature of the target user. The first non-text data corresponding to the recommended object and the second non-text data corresponding to the target user are then input into a second representation network to obtain the output of the second representation network: the first non-text representation of the recommended object and the second non-text representation of the target user. Finally, the text representation, the first non-text representation, and the second non-text representation of the recommended object are input into a prediction network to obtain the target click-through rate predicted by the prediction network.

[0035] In this solution, the text and non-text representations of recommended items are aligned. The text representation of the recommended item incorporates features from different modalities, making it adaptable to the recommendation task. Subsequently, the click-through rate (CTR) of the recommended item is predicted using the aligned text representation, the non-text data of the recommended item, and the user's non-text data. This improves the accuracy of the CTR prediction, thereby increasing the likelihood of recommending items of high interest to users and ultimately enhancing the user experience.

[0036] In one possible implementation, the prediction network is used to perform feature interactions on the first feature, the second feature, and the third feature based on the output of the second representation network and the text representation, and based on the result of the feature interactions, predicts the target click-through rate of the recommended object.

[0037] In this scheme, the prediction network obtains collaborative information through the interaction of the first, second, and third features. The text representation corresponding to the recommended object integrates features of different modalities of the recommended object. At the same time, by combining the collaborative information to predict the click-through rate of the recommended object, the accuracy of the click-through rate prediction can be improved, thereby increasing the likelihood of recommending items of high interest to users and thus improving the user experience.

[0038] In one possible implementation, there are multiple recommendation objects and multiple target users. The method also includes: determining the tags corresponding to the target users, where the tags indicate the recommendation objects to be recommended; and training a second representation network and a prediction network based on the error between the target click-through rate of the recommendation objects and the tags.

[0039] In one possible implementation, the prediction network is also connected to an embedding network, which processes the text representation so that the processed text representation can be concatenated with the output of the second representation network; the prediction network is used to predict the target click-through rate of the recommended object based on the output of the embedding network and the output of the second representation network.

[0040] In one possible implementation, there are multiple recommendation objects, and the method further includes: determining recommendation information based on the target click-through rate of the recommendation objects, wherein the recommendation information indicates the recommendation objects to be recommended, or the order in which the recommendation objects to be recommended; and sending the recommendation information to the target user.

[0041] In one possible implementation, the third feature includes the target user's historical click recommendations; the text representation, the first non-text representation, and the second non-text representation corresponding to the recommendations are input into the prediction network, including:

[0042] The text representations of the recommended objects, the text representations of the recommended objects based on the target user's historical clicks, the first non-text representation, and the second non-text representation are input into the prediction network.

[0043] In one possible implementation, determining the text representation corresponding to the recommended object includes: determining a set of text representations, which includes the identifiers and text representations corresponding to multiple recommended objects; and determining the text representation corresponding to the recommended object from the set of text representations based on the identifiers of the recommended object.

[0044] In this solution, by pre-constructing a set of text features, the text representation corresponding to the recommended object can be quickly obtained, meeting the needs of real-time processing.

[0045] Fourthly, embodiments of this application provide a click-through rate prediction device, comprising:

[0046] The first data acquisition module is used to acquire the text representation and the first non-text data corresponding to the recommended object. The text representation is the text representation output by the first representation network. The first representation network is trained by any of the methods in the first aspect. The first non-text data indicates the second feature of the recommended object described by non-text data.

[0047] The second data acquisition module is used to acquire the second non-text data corresponding to the target user, and the second non-text data indicates the third feature of the target user.

[0048] The representation network is used to input the first non-text data corresponding to the recommendation object and the second non-text data corresponding to the target user into the second representation network, and obtain the first non-text representation of the recommendation object and the second non-text representation of the target user output by the second representation network.

[0049] The prediction module is used to input the text representation, first non-text representation and second non-text representation of the recommended object into the prediction network to obtain the target click-through rate of the recommended object predicted by the prediction network.

[0050] In this solution, the text and non-text representations of recommended items are aligned. The text representation of the recommended item incorporates features from different modalities, making it adaptable to the recommendation task. Subsequently, the click-through rate (CTR) of the recommended item is predicted using the aligned text representation, the non-text data of the recommended item, and the user's non-text data. This improves the accuracy of the CTR prediction, thereby increasing the likelihood of recommending items of high interest to users and ultimately enhancing the user experience.

[0051] In one possible implementation, the prediction network is used to perform feature interactions on the first feature, the second feature, and the third feature based on the output of the second representation network and the text representation, and based on the result of the feature interactions, predicts the target click-through rate of the recommended object.

[0052] In this scheme, the prediction network obtains collaborative information through the interaction of the first, second, and third features. The text representation corresponding to the recommended object integrates features of different modalities of the recommended object. At the same time, by combining the collaborative information to predict the click-through rate of the recommended object, the accuracy of the click-through rate prediction can be improved, thereby increasing the likelihood of recommending items of high interest to users and thus improving the user experience.

[0053] In one possible implementation, there are multiple recommendation objects and multiple target users. The device further includes: a training module for determining the tags corresponding to the target users, where the tags indicate the recommendation objects to be recommended; and training a second representation network and a prediction network based on the error between the target click-through rate of the recommendation objects and the tags.

[0054] In one possible implementation, the prediction network is used to perform feature interactions on the first feature, the second feature, and the third feature based on the output of the second representation network and the text representation, and based on the result of the feature interactions, predicts the target click-through rate of the recommended object.

[0055] In one possible implementation, there are multiple recommended objects, and the device further includes: a recommendation module, used to determine recommendation information based on the target click-through rate of the recommended objects, the recommendation information indicating the recommended objects to be recommended, or the order in which the recommended objects to be recommended; and to send the recommendation information to the target user.

[0056] In one possible implementation, the third feature includes the target user's historical click recommendation objects;

[0057] The prediction module is used to input the text representation of the recommended object, the text representation of the target user's historical clicks on the recommended object, the first non-text representation, and the second non-text representation into the prediction network.

[0058] In one possible implementation, a first data acquisition module is used to determine a text representation set, which includes the identifiers and text representations corresponding to multiple recommendation objects; based on the identifiers of the recommendation objects, the text representations corresponding to the recommendation objects are determined from the text representation set.

[0059] In this solution, by pre-constructing a set of text features, the text representation corresponding to the recommended object can be quickly obtained, meeting the needs of real-time processing.

[0060] Fifthly, embodiments of this application provide a recommended object characterization apparatus, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory. When the program stored in the memory is executed, the processor is configured to execute the method provided in the first aspect, or the processor is configured to execute the method provided in the third aspect.

[0061] Sixthly, embodiments of this application provide a recommended object characterization apparatus, which executes computer program instructions to perform the method provided in the first aspect, or to perform the method provided in the third aspect. Exemplarily, the apparatus may be a chip or a processor.

[0062] In one example, the device may include a processor that may be coupled to memory, read instructions from the memory and execute the methods provided in the first aspect, or execute the methods provided in the third aspect. The memory may be integrated into the chip or processor, or it may be independent of the chip or processor.

[0063] In a seventh aspect, embodiments of this application provide a computing device cluster, including at least one computing device, each computing device including a processor and a memory; the processor of the at least one computing device is configured to execute instructions stored in the memory of the at least one computing device, such that the computing device cluster executes instructions stored in a computer storage medium, and when the instructions are executed on the computer, the computer executes the method provided in the first aspect, or executes the method provided in the third aspect.

[0064] Eighthly, embodiments of this application provide a computer storage medium storing instructions that, when executed on a computer, cause the computer to perform the method provided in the first aspect, or to perform the method provided in the third aspect.

[0065] Ninthly, embodiments of this application provide a computer program product containing instructions that, when executed on a computer, cause the computer to perform the method provided in the first aspect, or to perform the method provided in the third aspect. Attached Figure Description

[0066] Figure 1 is a schematic diagram of an artificial intelligence framework provided in an embodiment of this application;

[0067] Figure 2 is a schematic diagram of a recommendation framework provided in an embodiment of this application;

[0068] Figure 3 is a schematic diagram of the alignment of text representation and non-text representation provided in an embodiment of this application;

[0069] Figure 4 is a schematic diagram of the model management system provided in an embodiment of this application;

[0070] Figure 5 is a schematic diagram of the mask training first representation network 210 provided in an embodiment of this application;

[0071] Figure 6 is a flowchart illustrating the recommended object representation method provided in an embodiment of this application;

[0072] Figure 7a is a schematic diagram of the comparative learning training provided in an embodiment of this application;

[0073] Figure 7b is a schematic diagram of the structure of the model management system provided in an embodiment of this application;

[0074] Figure 8 is a schematic diagram of the application scenario of the recommendation model provided in the embodiments of this application;

[0075] Figure 9 is a schematic diagram of the recommendation model training process provided in the embodiments of this application;

[0076] Figure 10a is a schematic diagram of the structure of a model deployment system provided in an embodiment of this application;

[0077] Figure 10b is a schematic diagram of the structure of a model deployment system provided in an embodiment of this application;

[0078] Figure 11 is a flowchart illustrating a recommended method provided in an embodiment of this application;

[0079] Figure 12 is a schematic diagram of the structure of a recommended object characterization device provided in an embodiment of this application;

[0080] Figure 13 is a schematic diagram of a recommended device provided in an embodiment of this application;

[0081] Figure 14 is a schematic diagram of the structure of a computing device provided in an embodiment of this application. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be described below with reference to the accompanying drawings.

[0083] In the description of the embodiments of this application, the words "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the words "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a specific manner.

[0084] In the description of the embodiments in this application, the term "and / or" is merely a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, B existing alone, and A and B existing simultaneously. Furthermore, unless otherwise stated, the term "multiple" means two or more. For example, multiple systems refer to two or more systems, and multiple terminals refer to two or more terminals.

[0085] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and their variations all mean "including but not limited to," unless otherwise specifically emphasized.

[0086] The following explanations cover some of the terms used in this embodiment. It should be noted that these explanations are for the convenience of those skilled in the art and do not constitute a limitation on the scope of protection claimed in this application.

[0087] Artificial Intelligence (AI) is a branch of computer science that attempts to understand the nature of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems.

[0088] Multilayer Perceptron (MLP): Also known as Artificial Neural Network (ANN), a multilayer perceptron has fully connected layers. The bottom layer is the input layer, the middle layers are hidden layers, and the last layer is the output layer. The simplest MLP contains only one hidden layer.

[0089] Language models (LMs) play a crucial role in natural language processing. Their task is to predict the probability of a sentence appearing in a language. To date, the development of language models has progressed through grammatical rule-based language models, statistical language models, and neural network language models.

[0090] Large Language Models (LLMs) are language models containing over 10 billion parameters trained on massive amounts of text data, such as GPT-4 and LLaMA. LLMs have achieved remarkable results on various tasks. They are deep learning models trained on massive amounts of text data, incorporating various linguistic knowledge and rules. They can not only generate natural language text but also deeply understand its meaning, handling various natural language tasks such as text summarization, question answering, and translation.

[0091] Pre-training (or trained) refers to training a model in advance or the process of training a model in advance. Pre-training is typically done using a large non-textual dataset, which facilitates subsequent fine-tuning.

[0092] Pre-trained Language Model (PLM): By pre-training on large-scale text corpora, it can learn rich language knowledge and language patterns, thus achieving good results in various natural language processing tasks.

[0093] Fine-tuning refers to the process of applying a pre-trained model to a specific non-text dataset and adapting its parameters accordingly. In other words, fine-tuning involves retraining a pre-trained model (usually trained on a large non-text dataset) to suit a specific task or non-text dataset. Since real-world tasks typically have relatively small non-text datasets, fine-tuning a pre-trained model can yield good results. This technique allows the model to inherit existing knowledge and quickly adapt to new tasks or improve model performance through minimal customization. The fine-tuning process usually involves adjusting model parameters, updating weights, and may include modifying the model's architecture. This approach is highly effective in improving model performance on specific tasks and reducing training time and data requirements.

[0094] Contrastive learning aims to learn representations of data to capture the basic structure and relationships between different data points. The main idea of ​​contrastive learning is to bring similar samples closer together in the representation space while increasing the distance between different samples. This method can be applied to both supervised and unsupervised data and has shown good results in various visual and language tasks. In unsupervised learning, contrastive learning is an effective self-supervised learning method. Contrastive learning works by constructing a contrastive loss function that encourages sufficiently high similarity between positive samples and sufficiently low similarity between negative samples. This learning process helps the model better understand and generate representations of data, whether it is similar or dissimilar.

[0095] Recommender Systems: A recommender system is an information filtering system designed to predict and display products, services, or information that a user may be interested in. By analyzing users' historical behavior, preferences, social network interactions, and other relevant data, these systems can provide personalized content recommendations. Their purpose is to enhance user experience and increase the visibility of content or products, thereby helping businesses or service providers improve user satisfaction and business performance. Recommender systems are widely used in e-commerce, online video platforms, social media, and content aggregation websites.

[0096] Click-Through-Rate (CTR): A commonly used term in internet advertising, referring to the click-through rate of online ads (image ads / text ads / keyword ads / ranking ads / video ads, etc.), which is the actual number of clicks on the ad (strictly speaking, the number of times the ad reaches the target page) divided by the number of ad displays (show content).

[0097] Click-Through Rate Prediction (CTR) is the process of predicting the probability of a user clicking on an online ad or recommended content. This process uses machine learning algorithms to analyze user behavior data, contextual information, and ad characteristics to predict the likelihood of a specific ad or content being clicked. The purpose of CTR prediction is to optimize ad delivery and content recommendation by predicting click-through rates to determine which ads or content are more likely to attract user attention, thereby increasing ad revenue, enhancing user experience, and improving the effectiveness of personalized content recommendations. This technology plays a central role in digital marketing, search engine advertising, and social media promotion.

[0098] Collaborative signals, in recommender systems, refer to information extracted from the behavior or preferences of multiple users to infer and enhance predictions of items that a single user might be interested in. These signals are based on the principle of collaborative filtering, which analyzes the interaction history of a large number of users with items (such as ratings, browsing, and purchases) to discover similarities between users or associations between items. Collaborative signals can reveal common trends and preferences within user groups, helping recommender systems overcome the limitations of individual user data and more accurately personalize recommendations. For example, if multiple users like the same movie, the system might recommend that movie to other users with similar tastes, even if those users have never previously expressed interest in the film.

[0099] Hard prompts are handcrafted, predefined text or text templates with discrete input markers. Static prompts can be incorporated into programs for programming, storage, and reuse; large model-based applications can have multiple prompt templates available for their use.

[0100] Features: In machine learning and deep learning, features generally refer to useful information or attributes extracted from data. Features can be various aspects, attributes, or characteristics of the data, used to describe different facets of the data. Features play a crucial role in machine learning tasks because the performance of a model largely depends on the quality of the features used. Features can be various types of data, such as numerical, categorical, and textual data. In deep learning, neural networks automatically learn feature representations in data; these representations are typically high-level abstractions extracted from raw data through multi-layered neural networks. In image recognition tasks, such as identifying images of cats and dogs, features can be information such as edges, textures, and colors contained in the image. In natural language processing tasks, text data can be represented as word vectors or other forms of representation as features. In short, features are information describing key aspects or attributes of data and play a vital role in machine learning and deep learning tasks.

[0101] Cold start: This refers to the initial stage of model operation where no initial conditions are given to the model beforehand. The model needs to develop various indicator data from scratch throughout the entire operation. For example, simulated water level needs to start from 0 and gradually approach the actual water level, while simulated ocean currents need to start from 0 and gradually approach the actual flow velocity and direction. This initial stage is like warming up the model, establishing the necessary conditions. The initial data is inaccurate and cannot be used as research results. If the model simulation period is long enough, in the later stages of operation, the model will have reached conditions close to reality, and only then can the model truly begin to simulate your research content.

[0102] Figure 1 shows a schematic diagram of an artificial intelligence framework, which describes the overall workflow of an artificial intelligence system and is applicable to general artificial intelligence needs.

[0103] The above-mentioned artificial intelligence framework will be elaborated from two dimensions: "intelligent information chain" (horizontal axis) and "IT value chain" (vertical axis).

[0104] The "intelligent information chain" reflects a series of processes from data acquisition to processing. For example, it could be a general process of intelligent information perception, intelligent information representation and formation, intelligent reasoning, intelligent decision-making, and intelligent execution and output. In this process, data undergoes a condensation process of "data—information—knowledge—wisdom."

[0105] The "IT value chain" reflects the value that artificial intelligence brings to the information technology industry, from the underlying infrastructure of artificial intelligence, information (provided and processed by technology) to the industrial ecosystem of systems.

[0106] (1) Infrastructure:

[0107] Infrastructure provides computing power to support artificial intelligence systems, enabling communication with the external world and providing support through a basic platform. This communication occurs through sensors; computing power is provided by intelligent chips (hardware acceleration chips such as CPUs, NPUs, GPUs, ASICs, and FPGAs); and the basic platform includes distributed computing frameworks and related platform guarantees and support, which may include cloud storage and computing, interconnected networks, etc. For example, sensors communicate with the outside world to acquire data, and this data is provided to intelligent chips in the distributed computing system provided by the basic platform for computation.

[0108] (2) Data

[0109] The data at the next layer of infrastructure is used to represent the data sources in the field of artificial intelligence. The data involves graphics, images, voice, text, and IoT data from traditional devices, including business data from existing systems and sensor data such as force, displacement, liquid level, temperature, and humidity.

[0110] (3) Data processing

[0111] Data processing typically includes methods such as data training, machine learning, deep learning, search, reasoning, and decision-making.

[0112] Among them, machine learning and deep learning can perform intelligent information modeling, extraction, preprocessing, and training on data, including symbolization and formalization.

[0113] Reasoning refers to the process in which, in a computer or intelligent system, the machine thinks and solves problems by simulating human intelligent reasoning, based on reasoning control strategies and using formalized information. Typical functions include search and matching.

[0114] Decision-making refers to the process of making decisions based on intelligent information after reasoning, and it typically provides functions such as classification, sorting, and prediction.

[0115] (4) General ability

[0116] After the data processing mentioned above, the results of the data processing can be used to form some general capabilities, such as algorithms or a general system, for example, translation, text analysis, computer vision processing, speech recognition, image recognition, etc.

[0117] (5) Smart Products and Industry Applications

[0118] Intelligent products and industry applications refer to products and applications of artificial intelligence systems in various fields. They encapsulate overall artificial intelligence solutions, productize intelligent information decision-making, and realize practical applications. Their application areas mainly include: intelligent manufacturing, intelligent transportation, smart home, intelligent healthcare, intelligent security, autonomous driving, safe city, and intelligent terminals.

[0119] The application scenarios of the embodiments of this application will be introduced next.

[0120] Within the aforementioned artificial intelligence framework, the data in Figure 1 can be voice, text, or other similar data. Correspondingly, data processing can be performed using WizIran language processing, which can realize various functions such as translation, text analysis, speech recognition, question answering, and voice control.

[0121] Recommender systems analyze users' browsing history and interests to provide personalized item recommendations, presenting items that match their preferences. In recommender systems, ID-based collaborative filtering algorithms have become the most mainstream paradigm for modeling items. Currently, recommendation models based on IDs (userID, itemID, and various categorical IDs), such as DeepFM, PNN, and DCN, focus on high-order explicit feature interactions. The ID-based paradigm encodes user and item features into sparse one-hot embeddings, which are then transformed into dense embeddings through a lookup process. Through a feature interaction layer, potential relationships between features can be identified.

[0122] Recommendation models based on the ID paradigm are prevalent, but they exhibit the following limitations:

[0123] 1) Reliance on historical data: This reliance may lead to insufficient learning in scenarios such as cold start or long tail scenarios (scenarios with characteristics such as wide range, large differences, and low frequency). That is, when the corresponding data of items or features is limited or missing, it will result in insufficient learning.

[0124] 2) Limited Feature Range: ID-based models primarily focus on categorical and numerical features, neglecting potential untapped knowledge from other modalities, such as semantic knowledge contained in the text. This semantic knowledge could improve the model's effectiveness. For example, "bread" and "toast" may have similar semantic meanings, and related information could be used for more efficient recommendations.

[0125] Currently, PLM uses hard prompts to construct raw features in text format for each interaction, which are then input into PLM for recommendation. This approach represents features using natural language rather than one-hot vectors, preserving the semantic integrity of the features. This allows PLM to apply the external knowledge and reasoning abilities learned by the model, effectively capturing valuable information even in sparse scenarios. For example, ID-based recommendation models struggle to learn embeddings for data-sparse items like "toast," while PLM uses external knowledge to categorize bread as "bread," and the model includes semantic knowledge and descriptions related to bread, thus enhancing the effectiveness of the recommendation system.

[0126] CTR prediction can be directly performed using PLM fine-tuning. For example, user and item features can be concatenated into key:value format text data, and then the PLM can be fine-tuned using this text data. The PLM ultimately outputs the predicted CTR value. However, modeling based solely on plain text may not be suitable for capturing collaborative signals in recommender systems, which usually leads to performance degradation. Secondly, as the number of features increases, scalability issues may arise, i.e., the constructed feature text data exceeds the maximum character limit of the language model, and the excess portion is truncated, resulting in the loss of valuable recommendation-related information.

[0127] Based on this, this application provides a recommendation framework for CTR prediction. The recommendation object is an overview of the content to be recommended, such as music, videos, news, friends, products, etc., and can be designed according to actual circumstances.

[0128] As shown in Figure 2, the recommendation framework includes: a first representation network and a recommendation model; wherein, the recommendation model includes an embedding network, a second representation network and a prediction network.

[0129] The first representation network is used to perform semantic representation on the text data corresponding to the recommended object. The first representation network can be a PLM (Programmable Matrix Modeling). Its input is the text data corresponding to the recommended object, and its output is the text representation of the recommended object. The text data refers to the features of the recommended object described by text-type data. For ease of description and differentiation, this feature can be called the first feature.

[0130] The second representation network is used to represent the non-textual data corresponding to the recommended object (referred to as the first non-textual data for ease of description and differentiation) and the non-textual data corresponding to the user (referred to as the second non-textual data for ease of description and differentiation). The second representation network can be a neural network, such as an embedding table. Its inputs are the first non-textual data corresponding to the recommended object and the second non-textual data corresponding to the user, and its outputs are the non-textual representations corresponding to the recommended object (referred to as the first non-textual representation for ease of description and differentiation) and the non-textual representations corresponding to the user (referred to as the second non-textual representation for ease of description and differentiation). The first non-textual data represents the features of the recommended object described by non-textual data; for ease of description and differentiation, this feature can be called the second feature, and the non-textual type can be numerical and / or categorical. The second non-textual data represents the features of the user described by non-textual data; for ease of description and differentiation, this feature can be called the third feature, and the non-textual type can be numerical and / or categorical.

[0131] The embedding network is used to process the dimensions of the text representation output by the first representation network, ensuring that the dimensions of the text representation are consistent with the dimensions of the output of the second representation network, i.e., the dimensions of the text representation and the non-document representation are consistent. It should be noted that the embedding network is only one possible module, not a mandatory one. For example, the embedding network can be an MLP.

[0132] In scenarios where the recommendation model does not include an embedding network, the prediction network is used to predict the click-through rate (CTR) of the recommended object (which can be referred to as the target CTR for ease of description and distinction) based on the outputs of the first representation network and the second representation network. In scenarios where the recommendation model includes an embedding network, the prediction network is used to determine the target CTR of the recommended object based on the outputs of the embedding network, the first representation network, and the second representation network. In some possible implementations, the prediction network is used to perform feature interactions on the first, second, and third features to predict the target CTR of the recommended object. For example, the prediction network may include a feature interaction layer and an output layer. The feature interaction layer performs feature interactions on the first, second, and third features, and the output layer predicts the target CTR of the recommended object based on the results of the feature interactions. It should be noted that the structure of the prediction network shown in Figure 2 is merely an example and does not constitute a specific limitation; more layers, such as an attention layer, can be added.

[0133] During the training of the first representation network, as shown in Figure 3, the non-text representation corresponding to the recommended object needs to be aligned with the text representation corresponding to the recommended object, so that the text representation corresponding to the recommended object can adapt to the recommendation task. Subsequently, the click-through rate of the recommended object is predicted by the aligned text representation and non-text representation corresponding to the recommended object, which improves the accuracy of the click-through rate prediction.

[0134] The following describes the relevant content of the recommendation model under the recommendation framework provided in the embodiments of this application.

[0135] Specifically, the technical solutions provided in this application may include one or more of the following processing stages: data acquisition and preprocessing, training the recommendation model, deploying the recommendation model, and using the recommendation model.

[0136] The following sections will introduce each processing stage.

[0137] 1. Data acquisition and preprocessing.

[0138] In some embodiments, training the first representation model and the recommendation model requires collecting data and preprocessing the collected data to construct a corpus and a non-text dataset. The non-text dataset includes first non-text data corresponding to each of the multiple recommendation objects, second non-text data corresponding to each of the multiple users, and each user has a corresponding label. The corpus includes text data corresponding to each of the multiple recommendation objects. In this embodiment, the text data corresponding to each recommendation object, the first text data corresponding to each recommendation object, the second non-text data corresponding to a single user, and the label corresponding to the user can be called a training sample. The text data corresponding to a single recommendation object, the first text data, and the second non-text data are called a data pair.

[0139] Among them, text data indicates the first feature of the recommended object described by text-type data. The text data can be the feature value corresponding to the first feature (for ease of description and distinction, it can be called the first feature value, which is the feature value corresponding to the user in the first feature); the first feature can be multiple feature items, such as the name of the recommended object, the function description, etc.

[0140] The first non-text data is the second feature of the recommended object described by non-text data. The first non-text data can be the feature value corresponding to the second feature (for ease of description and distinction, it can be called the second feature value, which is the feature value corresponding to the user in the second feature). The second feature can be a numerical feature and / or a categorical feature. The second feature can be multiple feature items, such as the application developer, application category, application size, music composer, music style, etc. For numerical feature items in the second feature, the second feature value of the feature item can be the initial feature value of the recommended object corresponding to the feature item. For categorical feature items in the second feature, the second feature value of the feature item is used to represent the category. In addition, the second feature value needs to be vectorized, for example, it can be represented by hot encoding. For example, for each categorical feature feature item, a vector is constructed. Each element in the vector corresponds to the second feature value of a feature item. The order of the elements in the vector represents the arrangement of the feature values ​​of the feature items. If the feature value corresponding to the element matches the feature value corresponding to the recommended object, the element value is a preset value, such as 1; otherwise, it is 0.

[0141] The second non-text data is used to describe relevant user information. In this embodiment, the second non-text data indicates the user's third feature described by non-text data. There can be multiple third features, which can be numerical features and / or categorical features. In specific implementation, the second text data can be the feature value corresponding to the third feature (for ease of description and distinction, it can be called the third feature value, that is, the feature value corresponding to the user in the third feature); wherein, the third feature is multiple feature items. For example, multiple feature items can be age, gender, ethnicity, place of residence, job type, marital status, etc., and can also include the recommended objects clicked by the user in history, as well as contextual information such as time, weather, holidays, location, etc.; for feature items of numerical features, the third feature value of the feature item can be the initial feature value corresponding to the user in the feature item, or the feature value after normalization and standardization of the initial feature value; for feature items of categorical features, the target feature value of the feature item needs to be vectorized, for example, it can be represented by hot encoding. For details on the vectorization of the third feature value of the feature item, please refer to the above description and will not be repeated here.

[0142] It should be noted that the first feature and the second feature can be completely different or partially the same, and the embodiments of this application do not specifically limit this.

[0143] The tags are used to indicate several recommended items, such as one or more. It's important to note that user click-through rates for recommended items are difficult to quantify; therefore, tags generally refer to the recommended items that need to be recommended. It should also be noted that the recommended items can be those that users have actually clicked.

[0144] The aforementioned data can be collected based on user-related data such as browsing history, interests, and click records of recommended objects. When collecting this data, to ensure a comprehensive reflection of recommendations from various users, the data should include as many users and recommended objects as possible. Furthermore, the collected data can undergo processes such as anomaly removal (e.g., removing data with missing features), encoding, and normalization to obtain processed second non-text data corresponding to users, text data corresponding to recommended objects, and first non-text data corresponding to recommended objects. These are then labeled with user tags to obtain multiple training samples, serving as the data foundation for training the recommendation model. The user tags can be descriptions of the actual data collection or tags of similar users (users with similar or identical recommendation behavior). In some possible scenarios, after obtaining multiple training samples, they can be stored in a database for convenient subsequent maintenance, updates, and use.

[0145] 2. Training the recommendation model

[0146] The training process of the recommendation model in this embodiment consists of three steps, which are described in detail below:

[0147] Step 1: Domain Adaptation Pre-training. Pre-train the first representation network to better adapt to the unique characteristics of the recommendation scenario. Simultaneously, train the second representation network and the prediction network in a recommendation model.

[0148] Step 2: Modal Alignment for Recommendation. Align the text representations obtained from the pre-trained first representation network with the non-text representations corresponding to the recommended objects obtained from the second representation network. This alignment allows the text representations to adapt to the recommendation task, meaning the text representations can be embedded into the recommendation model.

[0149] Step 3: Multimodal Feature Fusion. The aligned text representation is treated as an additional text representation and then integrated with the output embedding of the second representation network. Subsequently, a new recommendation model is trained based on the additional text representation, the first non-text data (second features describing the recommended object), and the second non-text data (third features describing the user) to leverage the aligned text representation, the first non-text data (second features describing the recommended object), and the second non-text data (third features describing the user), thereby improving the overall effectiveness of the recommendation system.

[0150] Before detailing the three steps, we will first describe the model management system to which the method provided in this application embodiment may be applied. Figure 4 is a schematic diagram of the architecture of a model management system provided in this application embodiment. As shown in Figure 4, the system includes several computing devices 420 and terminals 410. Figure 4 shows three computing devices 420.

[0151] The terminal 410 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. Exemplary embodiments of the terminal 410 involved in this solution include, but are not limited to, electronic devices running iOS, Android, Windows, Harmony OS, or other operating systems. This application does not specifically limit the type of electronic device.

[0152] The computing device 420 can be configured as an independent physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server or cloud server cluster that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. The cloud server cluster is deployed in several cloud data centers. The software can be an application that implements object control methods, but is not limited to the above forms.

[0153] The terminal 410 and the computing device 420 are connected via a network, enabling the terminal 410 to access the cloud management platform 111 deployed on the cloud server cluster 100. The network can be a wired network or a wireless network. For example, a wired network can be a cable network, a fiber optic network, a Digital Data Network (DDN), etc., while a wireless network can be a telecommunications network, an internal network, the Internet, a Local Area Network (LAN), a Wide Area Network (WAN), a Wireless Local Area Network (WLAN), a Metropolitan Area Network (MAN), a Public Service Telephone Network (PSTN), a Bluetooth network, a ZigBee network, a Global System for Mobile Communications (GSM), a CDMA (Code Division Multiple Access) network, a CPRS (General Packet Radio Service) network, etc., or any combination thereof. Understandably, a network can use any known network communication protocol to enable communication between different client layers and gateways. These network communication protocols can be various wired or wireless communication protocols, such as Ethernet, Universal Serial Bus (USB), FireWire, Global System for Mobile Communications (GSM), General Packet Radio Service (GPRS), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), Time Division Code Division Multiple Access (TD-SCDMA), Long Term Evolution (LTE), New Radio (NR), Bluetooth, Wireless Fidelity (Wi-Fi), and other communication protocols.

[0154] In one possible scenario, computing device 420 can serve as a cloud (a software platform employing application virtualization technology, integrating multiple functions such as software search, download, use, management, and backup); in practical use, computing device 420 deploys a cloud management platform and a data center, and then terminal 410 and the cloud interact through the cloud management platform; in addition, the data center can deploy nodes, where nodes in the data center can be virtual machine instances, container instances, physical servers, etc.; in the data center, the number of nodes is generally massive.

[0155] In this embodiment, nodes can be used for training the first representation network and the recommendation model. In one possible scenario, when the recommendation model is large, it can be distributed and trained in parallel across multiple nodes based on the idea of ​​model parallelism. In another possible scenario, the recommendation model can be trained on a single node. The first representation network is similar.

[0156] In one possible scenario, the model training scheme provided in this application embodiment can be implemented by software. For example, the software can be online model training software or a service. Furthermore, the software has a client and a server; the terminal 410 runs the software client, and the computing device 420 runs the software server. Subsequently, while the terminal 410 is running the software client, it can call the server running on the computing device 420 to train the first representation network and the recommendation model.

[0157] In addition, for the development of first representation networks and recommendation models, developers can install AI development frameworks on terminal 410 and then develop AI models locally, or they can use AI development frameworks on online platforms (such as online open-source framework platforms, public cloud AI infrastructure development platforms, etc.) to develop AI models. AI development frameworks in the industry are usually open source. Typical AI development frameworks used for developing deep student models, also known as deep learning frameworks, include: PaddlePaddle, Tensorflow, Caffe, Theano, MXNet, Torch, and PyTorch, etc.

[0158] The three training steps will be described in detail below.

[0159] 1. Domain Adaptation Pre-training: To enhance the first representation network's understanding of recommendation domain text, this embodiment constructs a corpus. The corpus includes text data corresponding to each recommendation object among multiple recommendation objects. The text data includes detailed explanations and descriptions of the recommendation objects, going beyond content expressed solely using the ID-based paradigm. In this training step, the first representation network is pre-trained using this corpus. As shown in Figure 5, in a specific implementation, the text data can be masked to obtain masked text data. Each word in the masked text data is input into the first representation network 210 to obtain the text representation of each word output by the first representation network 210. Based on the text representation of each word output by the first representation network 210, the masked object is predicted. Based on the error between the predicted masked object and the real masked object, the first representation network 210 is trained. In some possible implementations, Masked Language Model Loss (MLM) can be used. This method masks a portion of the characters in a sentence and calculates the loss for each masked character prediction. For example, the input text "This ballet skirt is suitable for a ballet dancer's dress-up game..." might be masked as "This ballet skirt is [MASK] suitable for a dress-up game [MASK]...", forcing the first representation network to predict the masked token and giving the MLM loss. as follows:

[0160] Where m represents the number of masked words; \m represents the number of unmasked words; Text data representing the mask; This represents unmasked text data.

[0161] To enhance the first representation network's ability to solve recommendation tasks, this embodiment simultaneously trains a second representation network and a prediction network in the recommendation model. The recommendation model is used to estimate CTR based on first non-text data corresponding to the recommended object and second non-text data corresponding to the user. Assuming that the text data corresponding to each recommended object, the first text data corresponding to each recommended object, the second non-text data corresponding to a single user, and the user's corresponding tag can be called a training sample, and the text data corresponding to a single recommended object, the first text data, and the second non-text data are called a data pair, the specific training method is as follows: For each data pair, the second non-text data corresponding to the user and the first non-text data corresponding to the recommended object are input into the second representation network, and the output of the second representation network is input into the prediction network. The prediction network outputs the target click-through rate of the recommended object described by the data pair. For each training sample, based on the target click-through rate of each recommended object described by the training sample and the tag corresponding to the training sample, the error corresponding to the training sample, such as cross-entropy loss, is obtained. Based on the error corresponding to each training sample, the second representation network and the prediction network are trained. Correspondingly, the trained prediction network is used to perform feature interaction on the first feature, the second feature, and the third feature.

[0162] 2. Modal Alignment for Recommendation: Although the first representation network achieves a deep understanding of the recommendation domain through pre-training, a significant gap still exists between its output space and the recommendation task. This difference arises because, in scenarios where the first representation network is suited for PLM (Programmable Memory Management), its capabilities far exceed the requirements of the recommendation task. Without targeted training and constraints, the output of the first representation network may not align with the needs of the recommendation system. Therefore, orienting the output of the first representation network towards the recommendation task is crucial; this process is called alignment. This necessary alignment aims to provide a textual representation of the textual data that can be seamlessly integrated with the first non-textual data (used to describe the second feature) corresponding to the recommendation object in the next stage. In scenarios where the prediction network is used to implement feature interactions between the first, second, and third features, to improve the alignment efficiency at the feature interaction level, this embodiment aligns the textual representation of the recommendation object output by the first representation network and the first non-textual representation of the recommendation object output by the second representation network, achieving alignment at the recommendation object level. Considering that the text representation and the first non-text representation of the recommended object are different modalities and meanings, but they are features of the same recommended object, the text representation and the first non-text representation of the same recommended object should be as close as possible in the representation space, while the text representation and the first non-text representation of different recommended objects should be as far apart as possible.

[0163] Figure 6 is a flowchart illustrating the recommendation object representation method provided in an embodiment of this application. This embodiment can be applied to computing device 420. As shown in Figure 6, the recommendation object representation method provided in this embodiment includes at least the following steps:

[0164] Step 601: The computing device 420 acquires text data and first non-text data corresponding to each of the multiple recommended objects. The text data indicates that the first feature of the recommended object is described by text-type data; the first non-text data indicates that the second feature of the recommended object is described by non-text-type data.

[0165] Step 602: For each recommended object, the computing device 420 inputs the text data corresponding to the recommended object into the first representation network to obtain the text representation of the recommended object output by the first representation network; and inputs the first non-text data corresponding to the recommended object into the second representation network to obtain the non-text representation of the recommended object output by the second representation network.

[0166] Step 603: The computing device 420 trains a first representation network to align the non-textual and textual representations corresponding to the same recommendation object based on the non-textual and textual representations corresponding to each recommendation object.

[0167] In step 603, the computing device 420 trains a first representation network based on the error between the non-text representation and the text representation corresponding to the recommended object to reduce the difference between the text representation and the non-text representation corresponding to the same recommended object, and to increase the difference between the text representation and the non-text representation corresponding to different recommended objects.

[0168] For example, as shown in Figure 7a, for each recommendation object, the text data corresponding to the recommendation object and the first non-text data x corresponding to the recommendation object are determined. i,n After the text data is input into the first representation network, it is processed by a dimension processing function to obtain the processed text representation (referred to as e). i,t ); The first non-text data x corresponding to the recommended object i,n The input is fed into the second representation network, and the first non-textual representation (referred to as e) corresponding to the recommended object output by the second representation network is obtained. i,n Based on text representation e i,t and the first non-textual representation e i,n Contrastive learning is performed to train the first representation network. The process of contrastive learning can be based on text representation e. i,t and the first non-textual representation e i,n Determine the loss Based on loss Train the first representation network. Among them, as follows:

[0169] Among them, among them, This represents the text representation of the i-th recommended object after the text data is input into the first representation network, such as PLM, and processed by the dimension processing function. This represents the first non-text representation output by the second representation network after the first non-text data corresponding to the i-th recommended object is input into the second representation network; This represents the first non-text representation output by the second representation network after the first non-text data corresponding to the k-th recommended object is input into the second representation network; This represents the text representation of the k-th recommended object after the text data is input into the first representation network, such as PLM, and processed by the dimension processing function; where i and k are different.

[0170] In the specific implementation, as shown in Figure 7b, the text representation e of the same recommendation object i,t and the first non-textual representation e i,n As positive sample pairs, the text representations of different recommendation objects e i,t and the first non-textual representation e i,n As negative sample pairs, positive sample pairs and negative sample pairs are compared and learned to train the first representation processing module.

[0171] It should be noted that the second representation network in the recommendation model described in step 602 and the prediction network described in step 603 are pre-trained modules; for example, the details of the training method are described in the relevant description of the training of the recommendation model in the domain adaptation pre-training, and will not be repeated here.

[0172] 3. Multimodal feature fusion: In order to alleviate the suboptimal performance problem caused by using text data, the embodiments of this application integrate a first representation network, such as PLM, as a text encoder, while retaining a recommendation model based on the ID paradigm to capture cooperative signals.

[0173] In some possible implementations, to reduce training time and online latency, as shown in Figure 8, this embodiment of the application creates a recommendation object text representation table (referred to as E). i,t Recommended object text representation table E i,t This includes an identifier and text representation for each recommended object. The text representation is the text data corresponding to the recommended object input into a first representation network, and the first representation network outputs the text representation. The recommended object text representation table E is used for storage. i,t Then, by using the identifier of the recommended object, the text representation table E of the recommended object can be accessed. i,tThis allows for the rapid acquisition of text representations of recommended objects, meeting the demands of real-time processing. Subsequently, when there is a dimensionality mismatch between the outputs of the text representation and the second representation network, the recommendation model can be adaptively integrated into the embedding network before further training. In this embodiment, the recommendation model improves the accuracy of click-through rate prediction by integrating the first, second, and third features described by the text representation, first non-textual data, and second non-textual data. In the prediction network, the first, second, and third features are treated uniformly. This uniformity across textual and non-textual modalities prevents significant modifications to the original network, thereby facilitating its application to any network architecture.

[0174] The input to the prediction network is... in, This indicates concatenation; 'e' represents the input, including 'x'. u x h x c x i,n x u This represents a user profile, such as gender, age, ethnicity, occupation, etc. h This represents the recommended items a user has clicked on in their history, such as music, downloaded apps, and purchased products such as books, clothing, shoes, bags, food, and fruit. c This refers to contextual features, which can also be understood as scene features. For example, these could be weather, location, whether it is a holiday, etc., where x u x h x c The second non-text data corresponding to the user; x i,n E represents the first non-text data corresponding to the i-th recommended object; i,t This represents the set of text representations output by a first representation network, such as PLM: a text representation table of recommendation objects; x i,n,id This indicates that the i-th recommended object is in E i,t The identifier in; x h,id Recommended items clicked by users in history are in E i,t The identifier in; x i,n,id E i,t Indicates from E i,t Determine x in i,id Corresponding text representation; x h,id E i,t Indicates from E i,t Determine x in h,id The text representation corresponding to the identified historical object.

[0175] The output of the prediction network is: In this context, FC stands for fully connected layer, which is an example of an output layer, and FI stands for feature interaction layer.

[0176] Here, we assume that the text data corresponding to each recommended object, the first text data corresponding to each recommended object, the second non-text data corresponding to a single user, and the user's corresponding label can be called a training sample. The text data corresponding to a single recommended object, the first text data, and the second non-text data are called a data pair. The specific process of training the recommendation model can be as follows: For each data pair, the text data corresponding to the recommended data in the data pair is input into the trained first representation network to obtain the text representation of the recommended data output by the trained first representation network; the second non-text data corresponding to the user and the first non-text data corresponding to the recommended object in the data pair are input into the second representation network; the text representation of the recommended object and the output of the second representation network are input into the prediction network, and the prediction network outputs the target click-through rate of the recommended object described by the data pair. For each training sample, the target click-through rate is based on the description of each recommended object in that training sample. The error between the label and the training sample is used to obtain the error corresponding to the training sample (for ease of description and distinction, it can be called sample error), such as cross-entropy loss. Based on the error corresponding to each training sample, the recommendation model is trained. Here, sample error can be represented by cross-entropy loss. Cross-entropy loss as follows:

[0177] Where N is the number of recommended objects in the training samples; It is a binary indicator that indicates whether the recommended object j is a recommended object; It represents the target click-through rate for the recommended object j.

[0178] In summary, this application provides a flowchart illustrating a recommendation model training method. This embodiment can be applied to computing device 420. As shown in Figure 9, the recommendation object representation method provided in this application includes at least the following steps:

[0179] Step 901: The computing device 420 acquires text data and first non-text data corresponding to each of the multiple recommended objects, and second non-text data corresponding to each of the multiple users. The text data indicates that the first feature of the recommended object is described by text-type data; the first non-text data indicates that the second feature of the recommended object is described by non-text-type data; the second non-text data indicates that the third feature of the user is described by non-text-type data; each user has a corresponding tag, and the tag indicates the recommended object to be recommended.

[0180] Step 902: The computing device 420 trains the network based on the first non-text data corresponding to each recommended object, the second non-text data corresponding to each user, and the label corresponding to each user, to obtain the second representation network and the prediction network.

[0181] For details, please refer to the above description of the second representation network and prediction network in domain adaptation pre-training.

[0182] In specific implementation, for each user, the computing device 420 inputs the first non-text data corresponding to each recommended object and the second non-text data corresponding to the user into the second representation network, and inputs the output of the second representation network into the prediction network to obtain the target click-through rate of each recommended object corresponding to the user predicted by the prediction network; based on the error between the target click-through rate of each recommended object corresponding to each user and the label corresponding to each user, the second representation network and the prediction network are trained.

[0183] The method for obtaining the target click-through rate (CTR) of each recommended object corresponding to the user predicted by the prediction network can be as follows: For any user, the second non-text data corresponding to the user and the first non-text data corresponding to a recommended object are simultaneously input into the second representation network to obtain the first non-text representation of the recommended object and the second non-text representation of the user as output by the second representation network; the first non-text representation and the second non-text representation are input into the prediction network to obtain the target CTR of the user corresponding to the recommended object predicted by the prediction network; by processing each recommended object in the above manner, the target CTR of each recommended object corresponding to the user predicted by the prediction network can be obtained.

[0184] The computing device 420 determines the error between the target click-through rate of each recommended object for each user and the label corresponding to each user to train the second representation network and the prediction network in the following way: For each user, based on the target click-through rate of each recommended object for that user and the label corresponding to that user, the error corresponding to that user, such as cross-entropy loss, is obtained; based on the error corresponding to each user, the second representation network and the prediction network are trained.

[0185] In this embodiment of the application, the trained prediction network is used to perform feature interaction on the first feature, the second feature, and the third feature in order to predict the click-through rate of the recommended object.

[0186] Step 903: The computing device 420 trains the network based on the text data corresponding to each recommended object to obtain the first representation network.

[0187] For details, please refer to the above description of the pre-training of the first representation network in domain adaptation pre-training.

[0188] Step 904: For each recommended object, the computing device 420 inputs the text data corresponding to the recommended object into the first representation network to obtain the text representation of the recommended object output by the first representation network; and inputs the first non-text data corresponding to the recommended object into the second representation network to obtain the first non-text representation of the recommended object output by the second representation network.

[0189] Step 905: The computing device 420 trains a first representation network to align the first non-text representation and the text representation corresponding to the same recommended object based on the text representation and the first non-text representation corresponding to each recommended object.

[0190] For details, please refer to the description of steps 602 and 603 above, which will not be repeated here.

[0191] Step 906: For each recommended object, the computing device 420 inputs the text data corresponding to the recommended object into the trained first representation network to obtain the text representation of the recommended object output by the trained first representation network.

[0192] Step 907: For each user, the computing device 420 inputs the first non-text data corresponding to each recommended object and the second non-text data corresponding to the user into the second representation network; the output of the second representation network and the text representation corresponding to each recommended object output by the trained first representation network are input into the prediction network to obtain the target click-through rate of each recommended object corresponding to the user predicted by the prediction network.

[0193] For details on determining the target click-through rate for each recommended object for a user, please refer to the description of step 902, which will not be repeated here.

[0194] Step 908: The computing device 420 trains a second representation network and a prediction network based on the error between the target click-through rate of each recommended object for each user and the tag for each user.

[0195] For example, cross-entropy loss can be described by multimodal feature fusion. Train the second representation network and the prediction network.

[0196] In this solution, the text and non-text representations of recommended items are aligned. The text representation of the recommended item incorporates features from different modalities, making it adaptable to the recommendation task. Subsequently, the click-through rate (CTR) of the recommended item is predicted using the aligned text representation, the non-text representation of the recommended item, and the non-text representation of the user. This improves the accuracy of CTR prediction, thereby increasing the likelihood of recommending items of high interest to users and ultimately enhancing the user experience.

[0197] 3. Deploy the recommendation model

[0198] In some possible implementations, the trained recommendation model and the text representation table E of the recommendation object are... i,t It can be deployed on computing device 420.

[0199] In some possible scenarios, computing device 420 deploys a cloud management platform and a data center. The data center can deploy nodes, which can be used to deploy trained recommendation models and text representation tables of recommendation objects. i,t Once deployed, the recommendation model can become an AI application or a part of one. Users can access this AI application online via a webpage or a client app. When the AI ​​application is used, it can invoke the recommendation model deployed on the nodes to provide a response via online invocation. Thus, through the cloud management platform and the trained recommendation model, inference can be performed on online request data, and inference results can be returned. During the process of providing online services using the trained recommendation model, the cloud management platform can charge based on the number of times the recommendation model is invoked or the resource consumption of the recommendation model's inference.

[0200] In one possible scenario, the model deployment scheme provided in this application embodiment can be implemented through software. For example, the software can be online model deployment software or a service. Furthermore, the software has a client and a server; the terminal 410 runs the software client, and the computing device 420 runs the software server. Subsequently, during the process of the terminal 410 running the software client, it can call the server running on the computing device 420 to deploy the trained recommendation model and the text representation table E of the recommendation objects on the computing device 420. i,t .

[0201] For example, computing device 420 includes execution device 422 and training device 421; referring to Figure 10a, an embodiment of this application provides a system architecture.

[0202] The data acquisition device 710 is used to collect task data (generally raw data) under the recommendation task and store it in the database 720. The terminal 410 can process the task data in the database 720 to obtain multiple training samples under the recommendation task and store them in the database 720. The training device 421 can obtain multiple training samples under the recommendation task from the database 720, train the first representation model and the recommendation model, such as pre-training the first representation network, pre-training the first representation network and the prediction network in the recommendation model, aligning the outputs of the first representation network and the second representation network, and training the recommendation model after aligning the first representation network and the second representation network. For details, please refer to the description of steps 901 to 908 above.

[0203] The recommended model 401 obtained by training device 421 can be applied to different systems or devices. For example, as shown in Figure 10a, it can be deployed in execution device 422, and in some other possible scenarios, it can also be deployed in terminal 410.

[0204] The execution device 422 is equipped with an I / O interface 4222 for data interaction with external devices. The "user" can input data to the I / O interface 4222 through the terminal 410.

[0205] The execution device 422 can call data, code, etc. in the data storage system 1030, and can also store data, instructions, etc. in the data storage system 630.

[0206] The calculation module 4221 uses the recommendation model 401 to infer the input data, such as the second text data corresponding to the user, and the text data and first non-text data corresponding to each recommended object, to obtain a prediction result. Specifically, the prediction result is the target click-through rate for each recommended object. Subsequently, the calculation module 4221 can use the prediction result as the output result, or determine recommendation information based on the prediction result. The recommendation information indicates the target recommended objects to be recommended, and / or the order in which the target recommended objects to be recommended, and uses the recommendation information as the output result.

[0207] Finally, the I / O interface 4222 returns the output to the terminal 410 for the user.

[0208] At a deeper level, the training device 421 can generate corresponding recommendation models 401 based on multiple training samples under different recommendation tasks to provide users with better results.

[0209] In the scenario shown in Figure 10a, the user can manually specify the data to be input into the execution device 422, for example, by operating through the interface provided by the I / O interface 4222. Alternatively, the terminal 410 can automatically input data into the I / O interface 4222 and obtain the output results. If the terminal 410 needs user authorization to automatically input data, the user can set the corresponding permissions in the terminal 410. The user can view the output results of the execution device 422 on the terminal 410; the specific presentation format can be display, sound, animation, etc. The terminal 410 can also act as a data acquisition terminal to store the acquired task data into the database 730.

[0210] It is worth noting that Figure 10a is only a schematic diagram of a system architecture provided by an embodiment of this application. The positional relationship between the devices, components, modules, etc. shown in the figure does not constitute any limitation. For example, in Figure 10a, the data storage system 1030 is an external memory relative to the execution device 422. In other cases, the data storage system 1030 can also be placed in the execution device 422.

[0211] For example, computing device 420 includes execution device 422, and terminal 410 includes terminal 411 and terminal 412; referring to Figure 10b, this application embodiment provides a system architecture.

[0212] The execution device 422 is implemented by one or more servers, optionally in conjunction with other computing devices, such as data storage, routers, load balancers, etc. The execution device 422 can be deployed on a single physical site or distributed across multiple physical sites. In this embodiment, each server in the execution device 422 can be deployed with a recommendation model.

[0213] Users can interact with execution device 422 by operating their respective local devices (such as terminals 411 and 412). Each local device can represent any computing device, such as a personal computer, computer workstation, smartphone, tablet, smart camera, smart car or other type of cellular phone, media consumption device, wearable device, set-top box, game console, etc.

[0214] Each user's local device can interact with the execution device 422 through a communication network of any communication mechanism / standard. The communication network can be a wide area network, a local area network, a point-to-point connection, or any combination thereof.

[0215] In another implementation, one or more aspects of the execution device 422 can be implemented by each local device. For example, terminal 411 can provide local data to the execution device 422, or terminal 411 can deploy a recommendation model and a text representation table E of the recommendation objects. i,t The execution device 422 provides feedback on the recommended object clicked by the user; the execution device 422 can update the parameter values ​​of the model parameters of the recommendation model based on the recommended object clicked by the user.

[0216] It should be noted that all the functions of execution device 422 can also be implemented by a local device. For example, terminal 411 can deploy an inference model and provide inference services to its own users, or provide inference services to users of terminal 412.

[0217] 4. Use recommendation models

[0218] The recommended model usage scheme provided in this application embodiment can be applied to the aforementioned terminal 410 and also to the computing device 420. In specific implementation, it can run on the terminal 410 or computing device 420 in a software manner. For example, the software can be a service or an application. Furthermore, the technical solutions provided in this application embodiment can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network personal computers (PCs), minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc. This application embodiment can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application embodiment can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can reside in local and remote computer storage media, including storage devices.

[0219] Next, based on the model management system provided above, a recommended method provided in this application embodiment will be described in detail.

[0220] In some possible scenarios, the recommendation model is deployed on computing device 420. Figure 11 is a flowchart illustrating the recommendation method provided in an embodiment of this application. This embodiment can be applied to computing device 420. As shown in Figure 11, the recommendation method provided in this embodiment includes at least the following steps:

[0221] Step 1101: Terminal 410 sends the target user's information to computing device 420.

[0222] In some possible scenarios, user information may include the feature values ​​of the target user's third characteristic.

[0223] In some possible scenarios, terminal 410 can install software with recommendation requirements. This software can call the recommendation model deployed on computing device 420. Terminal 410 can collect information of target users based on recommendation requirements and send it to computing device 420.

[0224] Step 1102: The computing device 420 obtains second non-textual data corresponding to the target user based on the target user's information. The second non-textual data indicates the target user's third feature.

[0225] In some possible implementations, the computing device 420 stores a non-text data template corresponding to the second non-text data. For example, the template defines the location and meaning of the third feature. The computing device 420 extracts the feature value corresponding to the third feature from the target user's information and fills it into the non-text data template to obtain the second non-text data. For example, the third feature may include a description of the user profile, recommended objects based on the user's historical clicks, and contextual features (i.e., scene features). The contextual features are used to describe some characteristics of the user's environment, such as date, day of the week, weather, etc.

[0226] Step 1103: The computing device 420 determines the text representation and first non-text data corresponding to each of the multiple recommended objects. The text representation is the text representation output by the first representation network. The text representation is used to represent the first feature of the recommended object. The first non-text data indicates the second feature of the recommended object described by non-text data.

[0227] In some possible implementations, the computing device 420 acquires text features in the following ways:

[0228] Determine a set of text representations, which includes the identifier and text representation of each of the multiple recommendation objects; based on the identifier of the recommendation object, determine the text representation corresponding to the recommendation object from the set of text representations.

[0229] For example, the text representation set can be the recommendation object text representation table E shown in Figure 8. i,t By identifying the recommended object, the textual representation table E of the recommended object can be accessed. i,t From the text representation table E of the recommended object i,t The text representation corresponding to the recommended object is determined in the middle.

[0230] In addition, the first non-text data corresponding to the recommended object is also the first non-text data corresponding to the recommended object from the non-text dataset determined during the data collection and preprocessing process.

[0231] Step 1104: For each recommended object, the computing device 420 inputs the first non-text data corresponding to the recommended object and the second non-text data corresponding to the target user into the second representation network to obtain the first non-text representation of the recommended object and the second non-text representation of the target user output by the second representation network; the first non-text representation, the second non-text representation, and the text features corresponding to the recommended object are input into the prediction network to obtain the target click-through rate of the recommended object predicted by the prediction network.

[0232] In this embodiment, the prediction network is used to perform feature interactions on the first feature corresponding to the recommended object, the second feature corresponding to the recommended object, and the third feature corresponding to the user, based on the output of the second representation network and the text representation. Based on the result of the feature interactions, the network predicts the target click-through rate of the recommended object. For example, the prediction network may consist of a feature interaction layer and an output layer. The feature interaction layer is used to perform feature interactions on the first feature corresponding to the recommended object, the second feature corresponding to the recommended object, and the third feature corresponding to the user, based on the output of the second representation network and the text representation. The output layer is used to predict the target click-through rate of the recommended object based on the result of the feature interactions.

[0233] In some possible scenarios, the dimensions of the text representation and the output of the second representation network do not match. In such cases, the recommendation model may also include an embedding network, which is used to process the dimensions of the text representation; and a prediction network, which is used to predict the target click-through rate of the recommended object based on the output of the embedding network and the output of the second representation network.

[0234] In some possible implementations, where the second non-textual feature includes recommended objects clicked in the user's history, the computing device 420 needs to determine the textual features corresponding to the recommended objects clicked in the user's history; for example, the textual representation table E of the recommended objects can be accessed through the identifier of the recommended objects clicked in the user's history. i,t From the text representation table E of the recommended object i,t The process involves determining the text representations corresponding to the recommended objects clicked in the user's history. Then, the computing device 420 inputs the text representations of the recommended objects, the text representations of the recommended objects clicked in the user's history, and the non-text representations into the prediction network of the recommendation model to obtain the target click-through rate (CTR) of the recommended objects predicted by the prediction network. In scenarios where the prediction network includes an embedding network, the embedding network performs dimensionality processing on the text representations of the recommended objects and the text representations of the recommended objects clicked in the user's history, respectively. The prediction network then predicts the target CTR of the recommended objects based on the outputs of the embedding network and the second representation network.

[0235] Step 1105: The computing device 420 determines recommendation information based on the target click-through rate of each recommended object. The recommendation information indicates the target recommended objects to be recommended and / or the order in which the target recommended objects to be recommended.

[0236] In some possible implementations, the target recommendation objects to be recommended are the several recommendation objects with the highest target click-through rate. After determining the target click-through rate corresponding to each recommendation object, the computing device 420 then selects the several recommendation objects with the highest target click-through rate as the target recommendation objects.

[0237] In some other possible implementations, the target recommendation objects to be recommended can be all recommendation objects; correspondingly, the recommendation information indicates the order in which the target recommendation objects to be recommended need to be; here, after determining the target click rate of each recommendation object to the target user, the computing device 420 sorts the recommendation objects in descending order of the target click rate to obtain the recommendation information.

[0238] Step 1106: The computing device 420 sends recommendation information to the terminal 410.

[0239] In some possible scenarios, terminal 410 can access software with recommendation needs. This software can call the recommendation model deployed on computing device 420. Terminal 410 can collect information about the target user based on the recommendation needs and send it to computing device 420. Subsequently, computing device 420 sends recommendation information to terminal 410. When terminal 410 accesses software with recommendation needs, it displays the recommendation information, such as displaying the target recommended objects and the order in which they should be recommended. Afterward, the user can click on the target recommended object according to their own needs.

[0240] In this embodiment of the application, the text representation and non-text representation of the recommended items are aligned. The text representation of the recommended object incorporates features from different modalities of the recommended object, making the text representation of the recommended object adaptable to the recommendation task. Subsequently, the click-through rate of the recommended object is predicted by the aligned text representation, the non-text representation of the recommended object, and the non-text representation of the user. This improves the accuracy of the click-through rate prediction, thereby increasing the likelihood of recommending items of high interest to the user and thus improving the user experience.

[0241] Based on the same concept as the method embodiments of this application, this application also provides a recommendation object characterization device. The recommendation object characterization device includes several modules, each module being used to execute various steps in the personal information identification method provided in the embodiments of this application. The division of modules is not limited here. Those skilled in the art will clearly understand that in practical applications, the various steps in the recommendation object characterization method provided in the embodiments of this application can be assigned to different modules as needed, that is, the internal structure of the device can be divided into different modules to complete all or part of the functions described above. The modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the modules in the above device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0242] For example, the recommendation object representation device is used to execute the personal information recognition method provided in the embodiments of this application. FIG12 is a schematic diagram of the structure of the recommendation object representation device provided in the embodiments of this application. As shown in FIG12, the recommendation object representation device provided in the embodiments of this application includes:

[0243] The acquisition module 1201 is used to acquire text data and first non-text data corresponding to each of the multiple recommendation objects. The text data indicates that the first feature of the recommendation object is described by text-type data; the first non-text data indicates that the second feature of the recommendation object is described by non-text-type data.

[0244] The representation module 1202 is used to input the text data corresponding to the recommendation object into the first representation network for each recommendation object, and obtain the text representation of the recommendation object output by the first representation network; and input the first non-text data corresponding to the recommendation object into the second representation network, and obtain the non-text representation of the recommendation object output by the second representation network.

[0245] Training module 1203 is used to train a first representation network to align the non-textual and textual representations corresponding to the same recommendation object based on the non-textual and textual representations corresponding to each recommendation object, so that the prediction network connected by the second representation network predicts the target click-through rate of the recommendation object based at least on the outputs of the trained first and second representation networks.

[0246] Based on the same concept as the method embodiments of this application, this application also provides a click-through rate prediction device. As shown in FIG13, the click-through rate prediction device provided in this application includes:

[0247] The first acquisition module 1301 is used to acquire the text representation and the first non-text data corresponding to the recommended object. The text representation is the text representation output by the first representation network, and the first non-text data indicates the second feature of the recommended object described by non-text data.

[0248] The second acquisition module 1302 is used to acquire the second non-text data corresponding to the target user, the second non-text data indicating the third feature of the target user;

[0249] The representation module 1303 is used to input the first non-text data corresponding to the recommendation object and the second non-text data corresponding to the target user into the second representation network to obtain the first non-text representation of the recommendation object and the second non-text representation of the target user output by the second representation network.

[0250] The prediction module 1304 is used to input the text representation, the first non-text representation and the second non-text representation corresponding to the recommended object into the prediction network to obtain the target click-through rate of the recommended object predicted by the prediction network.

[0251] Based on the same concept as the embodiments of the method in this application, this application also provides a computing device. This computing device can be computing device 420. Figure 14 is a schematic diagram of the structure of a computing device provided in an embodiment of this application.

[0252] As shown in Figure 14, the computing device 420 includes a processor 421, a memory 422, and a network interface 423.

[0253] Processor 421 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0254] The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM).

[0255] For example, a computer program may be stored in the memory 422. When the processor 421 executes the computer program, it implements the steps in the above-described embodiment of the recommended object representation method, such as steps 601 to 603 shown in FIG. 6, steps 901 to 908 shown in FIG. 9, and steps 1102 to 1106 shown in FIG. 11. Alternatively, when the processor 421 executes the computer program, it implements the functions of each module in the above-described device embodiment. Exemplarily, the computer program may be divided into one or more modules / units. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions. The one or more modules / units are stored in the memory 422 and executed by the processor 421 to complete this application. For example, the computer program may be divided into an acquisition module 1201, a representation module 1202, and a training module 1203. The specific functions of each module are described above. For example, the computer program may be divided into a first acquisition module 1301, a second acquisition module 1302, a representation module 1303, and a prediction module 1304. The specific functions of each module are described above.

[0256] Network interface 423 is used to send and receive data, for example, to send data processed by processor 421 to other electronic devices, or to receive data sent by other electronic devices.

[0257] Of course, for simplicity, Figure 14 only shows some of the components of the computing device 420 relevant to this application, omitting components such as buses, input / output interfaces, etc. In addition, the computing device 420 may include any other suitable components depending on the specific application.

[0258] It should be noted that in some possible scenarios, there may be multiple computing devices 420, and the acquisition module 1201, the representation module 1202, and the training module 1203 may be distributed across multiple computing devices 420; and / or, the first acquisition module 1301, the second acquisition module 1302, the representation module 1303, and the prediction module 1304 may be distributed across multiple computing devices 420.

[0259] In addition to the methods, apparatus, and computing devices described above, embodiments of this application may also provide a computer program product comprising computer program instructions. When executed by a processor, these computer program instructions cause the processor to perform the steps of the recommended object representation methods of the various embodiments of this application described in the "Methods" section of this specification. The computer program product may be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The computer program code may be in source code form, object code form, executable file, or some intermediate form. The computer program code may be executed entirely on a user's computing device, partially on a user's device, as a standalone software package, partially on a user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0260] Furthermore, embodiments of this application may also provide a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps of the display control method according to various embodiments of this disclosure as described in the "Method" section above. The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may include, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. It should be noted that the content contained in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, a computer-readable medium may not include electrical carrier signals and telecommunication signals.

[0261] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0262] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0263] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of the various embodiments of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the specific details described above.

[0264] The block diagrams of devices, apparatuses, devices, and systems disclosed herein are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0265] It should also be noted that in the apparatus, devices, and methods of this disclosure, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions to this disclosure.

[0266] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

[0267] It is understood that the various numerical designations used in the embodiments of this application are merely for descriptive convenience and are not intended to limit the scope of the embodiments of this application.

Claims

1. A method for representing recommendation objects, characterized in that, The method includes: For each of the multiple recommended objects, obtain text data and first non-text data, wherein the text data indicates a first feature of the recommended object described by text-type data; and the first non-text data indicates a second feature of the recommended object described by non-text-type data. For each recommended object, the text data corresponding to the recommended object is input into a first representation network to obtain the text representation of the recommended object output by the first representation network; the first non-text data corresponding to the recommended object is input into a second representation network to obtain the non-text representation of the recommended object output by the second representation network. Based on the non-textual and textual representations corresponding to each recommended object, the first representation network is trained to align the non-textual and textual representations corresponding to the same recommended object, so that the prediction network connected to the second representation network predicts the target click-through rate of the recommended object based at least on the outputs of the trained first representation network and the second representation network.

2. The method according to claim 1, characterized in that, The first representation network is a pre-trained language model; and / or, The second representation network is a neural network model; and / or, The prediction network, based at least on the outputs of the trained first representation network and the second representation network, interacts with the first feature and the second feature to predict the target click-through rate of the recommended object.

3. The method according to claim 1 or 2, characterized in that, The step of training the first representation network to align the non-textual and textual representations corresponding to the same recommendation object based on the non-textual and textual representations of each recommendation object includes: Based on the error between the non-text representation and the text representation corresponding to each recommended object, the first representation network is trained to reduce the difference between the text representation and the non-text representation corresponding to the same recommended object, and to increase the difference between the text representation and the non-text representation corresponding to different recommended objects.

4. The method according to any one of claims 1 to 3, characterized in that, The second representation network and the prediction network are pre-trained networks, and the training methods for the second representation network and the prediction network include: Obtain second non-text data corresponding to each of the multiple users, the second non-text data indicating the third feature of the user; each user has a corresponding tag, the tag indicating the recommendation object to be recommended; For each user, the first non-text data corresponding to each recommended object and the second non-text data corresponding to the user are input into the second representation network, and the output of the second representation network is input into the prediction network to obtain the target click-through rate of each recommended object corresponding to the user predicted by the prediction network. The second representation network and prediction network are trained based on the error between the target click-through rate of each recommended object for each user and the tag for each user.

5. The method according to claim 4, characterized in that, The method further includes: For each user, the first non-text data corresponding to each recommended object and the second non-text data corresponding to the user are input into the trained second representation network, and the output of the trained second representation network is input into the prediction network to obtain the target click-through rate of each recommended object corresponding to the user predicted by the prediction network. Based on the error between the target click-through rate of each recommended object for each user and the label for each user, the trained second representation network and prediction network are fine-tuned.

6. The method according to claim 4 or 5, characterized in that, The prediction network is used to perform feature interaction on the first feature, the second feature, and the third feature.

7. A click-through rate prediction method, characterized in that, The method includes: Obtain the text representation and first non-text data corresponding to the recommended object, wherein the text representation is the text representation output by the first representation network, the first representation network is trained by any one of claims 1 to 5, and the first non-text data indicates the second feature of the recommended object described by non-text data. Obtain the second non-text data corresponding to the target user, where the second non-text data indicates the third feature of the target user; The first non-text data corresponding to the recommended object and the second non-text data corresponding to the target user are input into the second representation network to obtain the first non-text representation of the recommended object and the second non-text representation of the target user output by the second representation network. The text representation, the first non-text representation, and the second non-text representation corresponding to the recommended object are input into the prediction network to obtain the target click-through rate of the recommended object predicted by the prediction network.

8. The method according to claim 7, characterized in that, The prediction network and the second characterization network are the fine-tuned prediction networks described in claim 5; and / or The prediction network is used to perform feature interaction on the first feature, the second feature, and the third feature based on the output of the second representation network and the text representation, and to predict the target click-through rate of the recommended object based on the result of the feature interaction.

9. The method according to any one of claims 7 or 8, characterized in that, The prediction network is also connected to an embedding network, which processes the text representation so that the processed text representation can be concatenated with the output of the second representation network; the prediction network is used to predict the target click-through rate of the recommended object based on the output of the embedding network and the output of the second representation network.

10. The method according to any one of claims 7 to 9, characterized in that, The third feature includes the target user's historical click recommendation objects; The step of inputting the text representation, the first non-text representation, and the second non-text representation corresponding to the recommended object into the prediction network includes: The text representations corresponding to the recommended objects, the text representations corresponding to the historical clicks of the target user on recommended objects, the first non-text representation, and the second non-text representation are input into the prediction network.

11. The method according to any one of claims 7 to 10, characterized in that, The determination of the text representation corresponding to the recommended object includes: Determine a set of text representations, which includes the identifiers and text representations corresponding to each of the multiple recommendation objects; Based on the identifier of the recommended object, the text representation corresponding to the recommended object is determined from the text representation set.

12. The method according to any one of claims 7 to 11, characterized in that, The method further includes: Based on the target click-through rate of the recommended objects, recommendation information is determined, which indicates the recommended objects to be recommended, or the order in which the recommended objects to be recommended. The recommendation information is sent to the target user.

13. A device for representing a recommended object, characterized in that, The device includes: The acquisition module is used to acquire text data and first non-text data corresponding to each of the multiple recommendation objects. The text data indicates a first feature of the recommendation object described by text-type data, and the first non-text data indicates a second feature of the recommendation object described by non-text-type data. The representation module is used to, for each recommended object, input the text data corresponding to the recommended object into a first representation network to obtain the text representation of the recommended object output by the first representation network; and input the first non-text data corresponding to the recommended object into a second representation network to obtain the non-text representation of the recommended object output by the second representation network. The training module is used to train the first representation network to align the non-text representation and text representation corresponding to the same recommendation object based on the non-text representation and text representation corresponding to each recommendation object, so that the prediction network connected to the second representation network can predict the target click-through rate of the recommendation object based at least on the output of the trained first representation network and the output of the second representation network.

14. A click-through rate prediction device, characterized in that, The device includes: The first acquisition module is used to acquire the text representation and the first non-text data corresponding to the recommended object. The text representation is the text representation output by the first representation network. The first representation network is trained by any one of the methods in claims 1 to 5. The first non-text data indicates the second feature of the recommended object described by non-text data. The second acquisition module is used to acquire the second non-text data corresponding to the target user, wherein the second non-text data indicates the third feature of the target user; The representation module is used to input the first non-text data corresponding to the recommendation object and the second non-text data corresponding to the target user into the second representation network to obtain the first non-text representation of the recommendation object and the second non-text representation of the target user output by the second representation network. The prediction module is used to input the text representation, the first non-text representation, and the second non-text representation corresponding to the recommended object into the prediction network to obtain the target click-through rate of the recommended object predicted by the prediction network.

15. A computing device cluster, characterized in that, It includes at least one computing device, each computing device including a processor and memory; The processor of the at least one computing device is configured to execute instructions stored in the memory of the at least one computing device to cause the cluster of computing devices to perform the method as described in any one of claims 1-12.

16. A computer storage medium, characterized in that, The computer storage medium stores instructions that, when executed on the computer, cause the computer to perform the method as described in any one of claims 1-12.

17. A computer program product containing instructions, characterized in that, When the instructions are executed on a computer, the computer causes the computer to perform the method as described in any one of claims 1-12.

Images