Recommendation index prediction method and device, equipment and storage medium

Abstract

The invention discloses a recommendation index prediction method and device, equipment and a storage medium, and relates to the technical field of artificial intelligence. The method comprises the steps that a plurality of training samples are obtained, each training sample comprises a first sequence feature of a recommended object and a second sequence feature of directional content, the first sequence feature comprises at least one industry feature related to the recommended object, and the second sequence feature comprises at least one industry feature related to the directional content; the industry feature is used for representing an orientation object corresponding to the orientation content and identification information in an industry corresponding to the orientation object; obtaining coding information corresponding to each training sample; according to the coding information corresponding to each training sample, obtaining a prediction recommendation index corresponding to each training sample; and training the index prediction model according to the coding information corresponding to each training sample, the prediction recommendation index and the label data. The method can improve the prediction accuracy of the recommendation index for the directional content.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular to a method, device, equipment and storage medium for predicting a recommendation indicator. Background Art

[0002] Click-through rate prediction is a very important task in the advertising field. It has important applications in achieving accurate advertising delivery (or recommendation) and estimating the value of advertising delivery. The click-through rate refers to the probability that the recommended object (such as a user) corresponding to the advertisement will click on the advertisement.

[0003] At present, artificial intelligence technology has begun to be applied to the prediction of click-through rate of advertisements. In the related art, a click-through rate prediction model is trained based on the feature representation of the recommended object itself and the feature representation of the advertisement itself to obtain a trained click-through rate prediction model, which can be used to predict the click-through rate of the recommended object for the advertisement to determine whether to recommend the advertisement to the recommended object.

[0004] However, the correlation between the feature representation of the recommended object itself and the feature representation of the advertisement itself is not high, which results in the low prediction accuracy of the click-through rate prediction model. Summary of the invention

[0005] The embodiments of the present application provide a method, apparatus, device and storage medium for predicting a recommendation index, which can improve the prediction accuracy of an index prediction model, thereby improving the prediction accuracy of recommendation indexes for targeted content. The technical solution may include the following contents.

[0006] According to one aspect of an embodiment of the present application, a method for predicting a recommendation index is provided, the method comprising:

[0007] Acquire a plurality of training samples, wherein the training samples include a first sequence feature of a recommended object and a second sequence feature of targeted content, wherein the first sequence feature includes at least one industry feature related to the recommended object, and the second sequence feature includes at least one industry feature related to the targeted content;

[0008] For each of the training samples, obtaining encoding information corresponding to the training sample according to the first sequence feature and the second sequence feature included in the training sample;

[0009] Determining the prediction recommendation index corresponding to each of the training samples according to the coding information corresponding to each of the training samples through the index prediction model, wherein the prediction recommendation index refers to the predicted value of the recommendation index;

[0010] According to the encoding information, the predicted recommendation index and the marked recommendation index corresponding to each of the training samples, the index prediction model is trained to obtain a trained index prediction model, wherein the trained index prediction model is used to obtain the predicted recommendation index of the recommendation object for the targeted content, and the marked recommendation index refers to the marked value of the recommendation index;

[0011] The industry feature is used to characterize the directional object corresponding to the directional content and the identification information in the industry corresponding to the directional object.

[0012] According to one aspect of an embodiment of the present application, a method for predicting a recommendation index is provided, the method comprising:

[0013] Acquire input data, the input data including a first sequence feature of a recommended object and a second sequence feature of targeted content, the first sequence feature including at least one industry feature related to the recommended object, and the second sequence feature including at least one industry feature related to the targeted content;

[0014] Acquire encoding information corresponding to the input data according to a first sequence feature and a second sequence feature included in the input data;

[0015] Determining, by an indicator prediction model, a predicted recommended indicator corresponding to the input data according to the coding information corresponding to the input data, wherein the predicted recommended indicator refers to a predicted value of a recommended indicator;

[0016] The industry feature is used to characterize the directional object corresponding to the directional content and the identification information in the industry corresponding to the directional object.

[0017] According to one aspect of an embodiment of the present application, a device for predicting a recommendation index is provided, the device comprising:

[0018] A training sample acquisition module, used to acquire a plurality of training samples, wherein the training samples include a first sequence feature of a recommended object and a second sequence feature of a targeted content, wherein the first sequence feature includes at least one industry feature related to the recommended object, and the second sequence feature includes at least one industry feature related to the targeted content;

[0019] A coding information acquisition module, used for acquiring, for each of the training samples, coding information corresponding to the training sample according to the first sequence feature and the second sequence feature included in the training sample;

[0020] A recommendation index prediction module, used to determine the predicted recommendation index corresponding to each of the training samples according to the encoding information corresponding to each of the training samples through an index prediction model, wherein the predicted recommendation index refers to the predicted value of the recommendation index;

[0021] The prediction model training module is used to train the indicator prediction model according to the coding information, prediction recommendation indicators and marked recommendation indicators corresponding to each of the training samples to obtain a trained indicator prediction model. The trained indicator prediction model is used to obtain the prediction recommendation indicator of the recommended object for the targeted content. The marked recommendation indicator refers to the label value of the recommendation indicator; wherein the industry characteristics are used to characterize the targeted object corresponding to the targeted content and the identification information in the industry corresponding to the targeted object.

[0022] According to one aspect of an embodiment of the present application, a device for predicting a recommendation index is provided, the device comprising:

[0023] An input data acquisition module, used to acquire input data, the input data including a first sequence feature of a recommended object and a second sequence feature of a targeted content, the first sequence feature including at least one industry feature related to the recommended object, and the second sequence feature including at least one industry feature related to the targeted content;

[0024] A coding information acquisition module, used to acquire coding information corresponding to the input data according to the first sequence feature and the second sequence feature included in the input data;

[0025] A recommendation index prediction module is used to determine the predicted recommendation index corresponding to the input data according to the coding information corresponding to the input data through an index prediction model, wherein the predicted recommendation index refers to the predicted value of the recommendation index; wherein the industry characteristics are used to characterize the targeted object corresponding to the targeted content and the identification information in the industry corresponding to the targeted object.

[0026] According to one aspect of an embodiment of the present application, a computer device is provided, comprising a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the above-mentioned method for predicting recommendation indicators.

[0027] According to one aspect of an embodiment of the present application, a computer-readable storage medium is provided, in which a computer program is stored. The computer program is loaded and executed by a processor to implement the above-mentioned method for predicting recommendation indicators.

[0028] According to one aspect of an embodiment of the present application, a computer program product is provided, the computer program product comprising a computer program, the computer program being stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device performs the above-mentioned method for predicting the recommendation index.

[0029] The technical solution provided by the embodiments of the present application includes at least the following beneficial effects.

[0030] By constructing the first sequence features of the recommended object based on the industry features related to the recommended object, and constructing the second sequence features of the targeted content based on the industry features related to the targeted content, it is possible to associate the recommended object and the targeted content with the industry features corresponding to the targeted object (such as goods) in the industry (such as the e-commerce industry), thereby improving the correlation between the first sequence features of the recommended object and the second sequence features of the targeted content, as well as the correlation between the first sequence features of the recommended object and the second sequence features of the targeted content and the industry corresponding to the targeted object. The first sequence features of the recommended object and the second sequence features of the targeted content are then used to train the indicator prediction model, which can improve the prediction accuracy of the indicator prediction model, thereby improving the prediction accuracy of the recommendation indicator of the recommended object for the targeted content in the industry corresponding to the targeted object. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0032] Figure 1 It is a schematic diagram of an implementation environment of a solution provided by an embodiment of the present application;

[0033] Figure 2 is a schematic diagram of an indicator prediction model provided by an embodiment of the present application;

[0034] Figure 3 is a flow chart of a method for predicting a recommendation indicator provided by an embodiment of the present application;

[0035] Figure 4 is a schematic diagram of a method for obtaining training samples provided by an embodiment of the present application;

[0036] Figure 5 It is a flowchart of a training method of an indicator prediction model provided by an embodiment of the present application;

[0037] Figure 6 is a schematic diagram of a cross-scattering method provided by an embodiment of the present application;

[0038] Figure 7 This is a schematic diagram comparing the effects of three contrast loss acquisition methods provided by an embodiment of the present application;

[0039] Figure 8 is a flow chart of a method for predicting a recommendation indicator provided by another embodiment of the present application;

[0040] Fig. 9 It is a schematic diagram of a training method of an indicator prediction model provided by an embodiment of the present application;

[0041] Fig.10 This is a schematic diagram comparing the effects of the embodiment of the present application and the related art provided by an embodiment of the present application;

[0042] Fig.11 It is a block diagram of a prediction device for recommendation indicators provided by an embodiment of the present application;

[0043] Fig.12 is a block diagram of a prediction device for recommendation indicators provided by another embodiment of the present application;

[0044] Fig.13 is a block diagram of a prediction device for recommendation indicators provided by another embodiment of the present application;

[0045] Fig.14 It is a schematic diagram of the structure of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0046] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0047] Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.

[0048] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0049] Machine Learning (ML) is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

[0050] The technical solution provided in the embodiment of the present application involves artificial intelligence machine learning technology, which uses machine learning technology to train an indicator prediction model, and then predicts the recommended indicator through the trained indicator prediction model. Optionally, the recommended indicator includes but is not limited to at least one of the following: click-through rate, conversion rate, duration, click-to-conversion rate, exposure and frequency.

[0051] In the method provided in the embodiment of the present application, the execution subject of each step may be a computer device, which refers to an electronic device with data calculation, processing and storage capabilities. The computer device may be a terminal device such as a PC (Personal Computer), a tablet computer, a smart phone, a wearable device, an intelligent robot, a vehicle-mounted device, etc.; it may also be a server. Among them, the server may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0052] The technical solution provided by the embodiment of the present application is applicable to any scenario that requires recommendation index prediction, such as targeted content (such as advertisements, news, videos) delivery scenarios, targeted content delivery value estimation scenarios, targeted content recommendation scenarios (such as advertisement precision ranking scenarios), etc. The technical solution provided by the embodiment of the present application can improve the prediction accuracy of the index prediction model, thereby improving the prediction accuracy of the recommendation index of the recommended object for the targeted content.

[0053] The following will explain the model architecture of the solution implementation environment and indicator prediction model provided in the embodiment of the present application.

[0054] Please refer to Figure 1 , which shows a schematic diagram of a solution implementation environment provided by an embodiment of the present application. The solution implementation environment may include a model training device 10 and a model using device 20.

[0055] The model training device 10 may be an electronic device such as a PC, a computer, a tablet computer, a server, an intelligent robot, a vehicle-mounted terminal, or other electronic devices with strong computing capabilities. The model training device 10 is used to train the indicator prediction model 30 .

[0056] The indicator prediction model 30 is a neural network model for predicting the recommendation indicators of targeted content. The recommendation indicators include but are not limited to at least one of the following: click-through rate, conversion rate, duration, click-through conversion rate, exposure and frequency. When the recommendation indicators meet certain threshold conditions, targeted content can be recommended to the recommended object. The targeted content includes but is not limited to at least one of the following: advertising, news, videos, games, music, goods and services. The recommended object refers to the object to which the targeted content is to be recommended, which includes but is not limited to at least one of the following: users, companies and schools. The above-mentioned recommendation indicators, recommended objects and targeted content are only exemplary, and the embodiments of the present application do not limit the recommendation indicators, recommended objects and targeted content. Exemplarily, in an advertising precision ranking scenario, the indicator prediction model 30 can be used to predict the user's click-through rate on an advertisement to determine whether to recommend the advertisement to the user; in a news delivery scenario, the indicator prediction model 30 can predict the time a user spends on the news to determine whether to recommend the news to the user; in a video delivery scenario, the indicator prediction model 30 can predict the conversion rate for a video to determine whether to recommend the video to the user. The embodiments of the present application do not limit the industries to which the indicator prediction model 30 can be applied.

[0057] Optionally, the model training device 10 can use machine learning to train the indicator prediction model 30 so that it has better performance.

[0058] Exemplarily, the model training device 10 obtains the predicted recommendation index of the training sample through the index prediction model 30, and then obtains the main task loss according to the predicted recommendation index and the labeled recommendation index of the training sample. At the same time, the model training device 10 constructs the positive sample encoding and the negative sample encoding corresponding to the training sample according to the encoding information of the training sample (such as word embedding), and then obtains the contrast loss according to the positive sample encoding and the negative sample encoding, and finally trains the index prediction model 30 according to the main task loss and the contrast loss.

[0059] The training process is an iterative process, such as using multiple batches of training data to iteratively adjust the parameters of the indicator prediction model 30. Optionally, the termination conditions of the iteration include: minimizing the total loss of the indicator prediction model 30, the number of iterations is greater than or equal to a threshold, the total loss of the indicator prediction model 30 is less than or equal to a threshold, etc., which is not limited in the embodiment of the present application.

[0060] The above-mentioned trained indicator prediction model 30 can be deployed in the model using device 20 for use to provide the recommended indicator prediction service. The model using device 20 can be a terminal device such as a mobile phone, a computer, a smart TV, a multimedia playback device, a wearable device, a medical device, etc., or a server, which is not limited in this application.

[0061] In some embodiments, Figure 2 As shown, the above-mentioned indicator prediction model 30 can be a multi-layer neural network, which may include an input layer (not shown), an embedding layer 301 and a prediction network 302. The prediction network 302 may include a domain factor decomposition layer, a feature cross layer, a task layer, a hidden layer, a connection layer, a normalization layer, a pooling layer and an output layer.

[0062] Among them, the input layer is used to obtain training samples.

[0063] The embedding layer 301 is used to encode the training samples to obtain the encoding information of the training samples. The encoding information is used to characterize the training samples. Optionally, the encoding information can be constructed using vectors, strings, digital sequences, etc. In some examples, the encoding information may also refer to word embedding, i.e., embedding, which is not limited in the embodiments of the present application. Exemplarily, multiple vocabularies or multiple encoders may be deployed in the embedding layer 301 to obtain the encoding information of the training samples. The embedding layer 301 may be deployed separately from the indicator prediction model 30, or it may be deployed in the indicator prediction model 30, which is not limited in the embodiments of the present application.

[0064] The prediction network 302 is used to determine the predicted recommendation index of the training sample according to the encoding information of the training sample, and the predicted recommendation index refers to the predicted value of the recommendation index. Exemplarily, taking the recommendation index as the click-through rate as an example, the range of the value of the predicted recommendation index can be [0, 1] to indicate the possibility of the recommended object clicking on the targeted content. Among them, the recommended object can be a user, consumer, browser, etc. of the directional object corresponding to the content, and the embodiment of the present application is not limited to this. For example, in an e-commerce scenario, the directional object can be a commodity (such as food, daily necessities, home appliances, etc.), then the recommended object can refer to the consumer corresponding to the commodity, and the directional content can be an advertisement related to the commodity. The embodiment of the present application does not limit the directional object. For example, when the directional content is an advertisement, it can be a commodity in different industries (such as the furniture industry, the automobile industry, the clothing industry, etc.); when the directional content is news, it can be the content corresponding to the news.

[0065] Optionally, the domain factor decomposition layer in the prediction network 302 is used to obtain the latent vector representation corresponding to the training sample according to the encoding information of the training sample. Among them, the domain factor decomposition layer can be constructed using a structure such as FwFFM (Field-Weighted Field-Ware Factorization Machines, field weighted domain factor decomposition machine) structure, FFM structure, etc., which can fully cross the features and learn more domain knowledge information. The latent vector representation can be used to characterize the cross-features between the various features corresponding to the training sample, that is, the interaction between the various features.

[0066] For example, taking FwFFM as an example, the processing process of the domain factor decomposition layer can be expressed by the following formula:

[0067]

[0068] Where m is the total number of features, x is 1 or 0, i refers to the encoding information of the i-th feature (such as word embedding), j refers to the encoding information of the j-th feature, and w i is the weight value of the i-th feature, which is a scalar, w0 is the initial weight parameter, F(i) refers to the feature domain corresponding to the encoded information i, v(i, F(j)) refers to the interaction vector of the encoded information i to the feature domain F(j), and v(j, F(i)) refers to the interaction vector of the encoded information j to the feature domain F(i).

[0069] The first hidden layer (from bottom to top) in the prediction network 302 is used to perform feature learning (such as densification) according to the encoding information corresponding to the training sample. The hidden layer can be constructed using MLP (Multi-Layer Perceptron), CNN (Convolutional Neural Networks), DNN (Deep Neural Networks), etc.

[0070] The feature cross layer in the prediction network 302 is used to process the encoding information of the training sample to cross each feature corresponding to the training sample. Optionally, the feature cross layer can be constructed using a PNN (Product Neural Network), an IPNN (Inner Product Neural Network), or the like.

[0071] The second hidden layer in the prediction network 302 is used to perform feature learning on the output of the feature cross layer.

[0072] The first connection layer in the prediction network 302 is used to splice the output of the domain factor decomposition layer, the output of the first hidden layer, and the output of the second hidden layer, such as splicing the output of the first hidden layer after the output of the domain factor decomposition layer, and splicing the output of the second hidden layer after the output of the first hidden layer.

[0073] The normalization layer (ie, Layer Normal) in the prediction network 302 is used to normalize the output of the first connection layer. The third hidden layer in the prediction network 302 is used to perform feature learning on the output of the normalization layer.

[0074] The task layer in the prediction network 302 is used to process the output of the normalization layer according to the encoding information of the training sample. Optionally, the task layer can be constructed using MMoE (Multi-gate Mixture-of-Experts), MoE (Mixture-of-Experts), etc.

[0075] The second connection layer in the prediction network 302 is used to concatenate the output of the third hidden layer and the output of the task layer. The pooling layer in the prediction network 302 is used to pool the encoding information of the training sample and the output of the second connection layer, such as mean pooling, sum pooling, etc.

[0076] The output layer in the prediction network 302 is used to obtain the prediction recommendation index according to the output of the pooling layer. Optionally, a sigmoid function can be deployed in the output layer to perform logistic regression on the output of the pooling layer to obtain the prediction recommendation index.

[0077] In an embodiment of the present application, the coding information input by the pooling layer, the coding information input by the task layer, the coding information input by the domain factor decomposition layer, and the coding information input by the feature cross-layer can be constructed using different vocabularies. This can make the coding information corresponding to the training samples richer and more comprehensive, which is conducive to improving the accuracy of the prediction recommendation indicators.

[0078] Exemplarily, taking the encoding information as word embedding as an example, the embedding layer 301 includes four vocabularies: vocabulary 1, vocabulary 2, vocabulary 3 and vocabulary 4. Vocabulary 1 is used to obtain the word embedding input by the pooling layer, vocabulary 2 is used to obtain the word embedding input by the domain factor decomposition layer, vocabulary 3 is used to obtain the word embedding input by the feature cross layer, and vocabulary 4 is used to obtain the word embedding input by the task layer.

[0079] Optionally, each vocabulary can obtain two types of encoding information corresponding to the training samples, and the two types of encoding information can be spliced ​​together as the input of the next layer, which can further improve the richness and comprehensiveness of the encoding information, thereby facilitating improving the accuracy of the prediction recommendation index.

[0080] The embodiments of the present application do not limit the model architecture of the indicator prediction model. The indicator prediction model in the above embodiments is only exemplary and explanatory. Any model that can achieve the prediction of the recommended indicator should be within the protection scope of the embodiments of the present application.

[0081] The following is an embodiment of the method of the present application. For details not disclosed in the embodiment of the method of the present application, please refer to the above embodiment.

[0082] Please refer to Figure 3 , which shows a flow chart of a method for predicting a recommendation index provided by an embodiment of the present application. The execution subject of each step of the method may be the model training device introduced above. The method may include the following steps (301-304).

[0083] Step 301, obtain multiple training samples, which include a first sequence feature of a recommended object and a second sequence feature of targeted content. The first sequence feature includes at least one industry feature related to the recommended object, and the second sequence feature includes at least one industry feature related to the targeted content. The industry feature is used to characterize the targeted object corresponding to the targeted content, and the identification information in the industry corresponding to the targeted object.

[0084] The training samples are used to train the indicator prediction model. In the embodiment of the present application, multiple batches of training samples can be used to iteratively train the indicator prediction model. Since each iteration process of the indicator prediction model is the same, the embodiment of the present application uses a certain batch of training samples to illustrate the training process of the indicator prediction model.

[0085] In an embodiment of the present application, the industry feature is related to the directional object and the industry corresponding to the directional object, which can reflect the characteristics of the directional object and the characteristics of the industry in which the directional object is located. Exemplarily, the industry feature can be constructed in the form of type-subtype of the directional object in the industry to which it corresponds, such as shoes-sneakers, cosmetics-lipsticks, clothes-dresses, etc. The industry feature can be constructed in the form of type-brand of the directional object in the industry to which it corresponds, for example, clothes-a certain clothes brand A, shoes-a certain shoe brand B, etc. The embodiment of the present application does not limit the method for constructing industry features. Optionally, in the field of shopping (such as the field of e-commerce), the industry feature can be referred to as a product word feature.

[0086] The first sequence feature is used to characterize the recommended object. Optionally, the first sequence feature is a multi-valued sequence feature, and its corresponding industry features are sorted in chronological order. Exemplarily, each industry feature in the first sequence feature is an industry feature corresponding to a directional object having an interactive relationship with the recommended object, and the industry features are sorted in sequence according to the interaction time. For example, the first sequence feature may include: shoes-sports shoes, cosmetics-lipstick, clothes-dresses, and clothes-a certain clothing brand A.

[0087] In one example, the first sequence feature includes a first subsequence feature and a second subsequence feature, the industry feature in the first subsequence feature has a binding relationship with the recommended object, the second subsequence feature includes the first subsequence feature, and the second subsequence feature contains industry features that have a non-binding relationship with the recommended object.

[0088] Among them, if there is a binding relationship between the industry feature and the recommended object, then the industry feature is strongly associated with the recommended object, that is, the recommended object is more interested in the directional object corresponding to the industry feature, and there is a high probability of interactive behavior. Optionally, the binding relationship can be determined by manual annotation, which is a formal and recognized relationship. If there is a non-binding relationship between the industry feature and the recommended object, then the industry feature is weakly associated or unassociated with the recommended object.

[0089] Exemplarily, if the first subsequence feature includes industry feature 1 and industry feature 2, the second subsequence feature may include industry feature 1, industry feature 2, industry feature 3, and industry feature 4. Industry feature 3 and industry feature 4 may be obtained by machine selection, such as by MLP. Compared with the first subsequence feature, the second subsequence feature is more comprehensive, which is conducive to improving the comprehensiveness of the first sequence feature to solve the problem of feature sparsity, thereby improving the prediction accuracy of the indicator prediction model. Among them, feature sparsity refers to insufficient coverage of the feature, that is, insufficient samples covered by the feature.

[0090] Since the second sequence feature characterizes the targeted content, it is different from the related art that characterizes the targeted content with the characteristics of the targeted content itself (such as the name and number of the targeted content). The embodiment of the present application uses industry characteristics related to the targeted content to characterize the targeted content, so that the recommendation of the targeted content can be extended to the recommendation of the targeted object to form a strong association with the recommended object, which is conducive to improving the prediction accuracy of the indicator prediction model. For example, the second sequence feature can include: clothes-a certain clothing brand A. Optionally, the second sequence feature can also include the consumption of the targeted content, such as whether the targeted content is consumed.

[0091] In one example, the second sequence feature includes a third subsequence feature and a fourth subsequence feature, the industry feature in the third subsequence feature has a binding relationship with the targeted content, and the fourth subsequence feature is used to characterize the consumption of the targeted content.

[0092] If there is a binding relationship between the industry feature and the targeted content, then the industry feature is strongly correlated with the targeted content. For example, regarding an advertisement for clothing, the corresponding third subsequence feature may include industry features such as clothing-a certain clothing brand A, clothing-dresses, etc. Optionally, the third subsequence feature can also be constructed based on the industry features corresponding to the consumed object bound to the targeted content. For example, if advertisement A is bound to product B, the industry features corresponding to product B can be determined as the third subsequence feature. This embodiment of the application is not limited to this.

[0093] Taking advertisement as an example, the fourth subsequence feature is used to characterize that the advertisement exposure is not clicked, or is used to characterize that the advertisement exposure is clicked. Optionally, if the advertisement exposure is not clicked, the corresponding training sample can be determined as a negative sample of the indicator prediction model, and the label data is 0; if the advertisement exposure is clicked, the corresponding training sample can be determined as a positive sample of the indicator prediction model, and the label data is 1.

[0094] In one example, the identification information of the recommended object and the first sequence feature are stored in the database in the form of a key-value pair, and the identification information of the targeted content and the second sequence feature are stored in the database in the form of a key-value pair.

[0095] The identification information of the recommended object is used to identify the recommended object, and the identification information of the targeted content is used to identify the targeted content. The identification information may refer to, for example, ID (Identity Document), name, number, etc. Optionally, the identification information of the targeted object is used to identify the targeted object, and its corresponding identification information may be a product word sequence in its corresponding industry.

[0096] Optionally, refer to Figure 4 , the process of obtaining training samples can be as follows: according to the identification information of the recommended object (such as User-ID), obtain the first sequence feature from the database 401; according to the identification information of the targeted content (such as AD-ID), obtain the second sequence feature from the database 401; combine the first sequence feature and the second sequence feature to obtain the training sample 402.

[0097] In some embodiments, the identification information of the recommended object can be incorporated into the first sequence feature, and the identification information of the targeted content can be incorporated into the second sequence feature. In some embodiments, the identification information of the recommended object and the identification information of the targeted content can be incorporated into the training sample in a separate form. In some embodiments, the fourth subsequence feature may not be stored in the database. When obtaining the training sample, the fourth subsequence feature is obtained separately and then combined to form the training sample. This embodiment of the application is not limited to this.

[0098] Step 302: For each training sample, according to the first sequence feature and the second sequence feature included in the training sample, obtain the encoding information corresponding to the training sample.

[0099] The coding information is used to characterize the training sample. In the embodiment of the present application, each industry feature in the training sample corresponds to coding information. After obtaining the coding information of each industry feature, the coding information of each industry feature is combined to obtain the coding information corresponding to the training sample. Taking the coding information as word embedding as an example, after obtaining the word embedding of each industry feature, the word embedding of each industry feature is combined to obtain the word embedding corresponding to the training sample.

[0100] In one example, the encoding information of the first sequence feature and the encoding information of the second sequence feature can be obtained respectively, and then the encoding information of the first sequence feature and the encoding information of the second sequence feature can be combined to obtain the encoding information corresponding to the training sample, that is, the encoding information corresponding to the training sample includes the encoding information of the first sequence feature and the encoding information of the second sequence feature.

[0101] For example, taking the case where the encoded information is word embedding, the word embedding corresponding to the recommended object can be combined according to the word embedding of each industry feature in the first sequence feature, that is, the word embedding corresponding to the first sequence feature, which can be used to characterize the recommended object. The word embedding corresponding to the targeted content can be combined according to the word embedding of each industry feature in the second sequence feature, that is, the word embedding corresponding to the second sequence feature, which can be used to characterize the targeted content. The first sequence feature and the second sequence feature share the same vocabulary, which allows the word embedding corresponding to the first sequence feature and the word embedding corresponding to the second sequence feature to remain in the same feature space.

[0102] In one example, the encoding information corresponding to the training sample can be obtained according to a vocabulary, and the vocabulary is used to characterize the correspondence between industry characteristics and encoding information. For example, the vocabulary can include multiple sets of relationship pairs consisting of industry characteristics and encoding information. Optionally, in the case where the encoding information is word embedding, the process of obtaining the encoding information can also be referred to as word embedding lookup. In some examples, encoding methods such as one-hot encoding and Word2vec can also be used to obtain the encoding information of each industry feature to generate the encoding information corresponding to the training sample.

[0103] In one example, the encoding information corresponding to the training sample includes first sub-encoding information and second sub-encoding information, both of which are used to represent the training sample, and the first sub-encoding information and the second sub-encoding information share the same vocabulary. The representation form of the first sub-encoding information is different from the representation form of the second sub-encoding information, but the two can share the same feature space.

[0104] In this vocabulary, an industry feature can correspond to two types of coding information, one type of coding information constitutes the first sub-coding information, and the other type of coding information constitutes the second sub-coding information. Optionally, the above vocabulary can be the vocabulary corresponding to the domain factor decomposition layer in the indicator prediction model, or it can be any vocabulary in the multiple vocabularies corresponding to the indicator prediction model. By jointly representing the training samples by the first sub-coding information and the second sub-coding information, the accuracy of the coding information corresponding to the training samples can be improved, which is conducive to improving the prediction accuracy of the indicator prediction model.

[0105] It should be noted that in the training process of the indicator prediction model, the training samples include the first sub-coding information and the second sub-coding information under each vocabulary. In the construction process of the contrast loss, only the first sub-coding information and the second sub-coding information under a vocabulary need to be obtained.

[0106] Optionally, each vocabulary may correspond to only one type of encoding information, that is, one type of encoding information is used to represent the training sample, which is not limited in the embodiments of the present application.

[0107] Step 303: Determine the predicted recommendation index corresponding to each training sample according to the coding information corresponding to each training sample through the index prediction model. The predicted recommendation index refers to the predicted value of the recommendation index.

[0108] Optionally, feature extraction is performed on the encoded information of each training sample through the indicator prediction model to obtain the feature representation of each training sample, and then logistic regression is performed on the feature representation of each training sample to obtain the prediction recommendation indicator corresponding to each training sample.

[0109] The predicted recommendation index refers to a predicted recommendation index, which is used to determine whether to recommend targeted content to the recommended object. Exemplarily, the predicted recommendation index may be a predicted click-through rate, conversion rate, duration, click-through rate, etc. Optionally, the predicted recommendation index may be a predicted probability value. For example, the predicted click-through rate is a probability value that can be used to indicate the possibility of the recommended object clicking on the targeted content.

[0110] In the embodiment of the present application, for each training sample, the encoding information of the training sample under multiple vocabularies can be obtained, and then the prediction recommendation index corresponding to the training sample can be obtained according to the encoding information of the training sample under multiple vocabularies through the index prediction model. Among them, the encoding information under each vocabulary can be composed of two different types of sub-encoding information. This is conducive to improving the comprehensiveness and richness of the encoding information, and then improving the accuracy of the prediction recommendation index.

[0111] For example, Figure 2 Taking the indicator prediction model in as an example, the process of obtaining the prediction and recommendation indicators can be as follows:

[0112] The embedding layer 301 obtains the first input of the pooling layer (sum) according to vocabulary 1 (i.e., the encoding information of the training sample under vocabulary 1), obtains the input of the domain factor decomposition layer (FwFFM) (i.e., the encoding information of the training sample under vocabulary 2) and the input of the first hidden layer (MLP) according to vocabulary 2, obtains the input of the feature cross layer (IPNN) according to vocabulary 3 (i.e., the encoding information of the training sample under vocabulary 3), and obtains the first input of the task layer (MoE) according to vocabulary 4 (i.e., the encoding information of the training sample under vocabulary 4).

[0113] The second hidden layer (MLP) performs feature learning on the output of the feature cross layer to obtain the output of the second hidden layer. Then, the output of the domain factorization layer, the output of the first hidden layer, and the output of the second hidden layer are concatenated through the first connection layer (concat) to obtain the output of the first connection layer.

[0114] The normalization layer (Layer Normal) normalizes the output of the first connection layer to obtain the output of the normalization layer. The third hidden layer (MLP) performs feature learning on the output of the normalization layer to obtain the output of the third hidden layer. The task layer processes the output of the normalization layer according to the first input of the task layer to obtain the output of the task layer.

[0115] The second connection layer (concat) concatenates the output of the third hidden layer and the output of the task layer to obtain the output of the second connection layer. The pooling layer processes the output of the second connection layer according to the first input of the pooling layer to obtain the output of the pooling layer. Finally, the output layer (sigmoid) performs logistic regression on the output of the pooling layer to obtain the prediction recommendation index corresponding to the training sample.

[0116] Step 304, train the indicator prediction model according to the encoding information, predicted recommendation indicators and marked recommendation indicators corresponding to each training sample, and obtain a trained indicator prediction model. The trained indicator prediction model is used to obtain the predicted recommendation indicator of the recommended object for the targeted content. The marked recommendation indicator refers to the marked value of the recommendation indicator.

[0117] Among them, the marked recommendation indicator can be used to represent the real behavior of the recommended object on the targeted content. For example, taking the recommendation indicator as click-through rate, if the marked recommendation indicator is 1, then the marked recommendation indicator can be used to represent that the recommended object has clicked on the targeted content; if the marked recommendation indicator is 0, then the marked recommendation indicator can be used to represent that the recommended object has not clicked on the targeted content.

[0118] In the embodiment of the present application, the indicator prediction model mainly performs the prediction task of the recommended indicator, and takes the training sample as input. However, considering that there may be deviations between the second sequence features in the training sample and the corresponding second sequence features when the indicator prediction model is used online, for example, the advertisement exposure corresponding to the training sample is not clicked, while the advertisement corresponding to the online input is not exposed. There is a deviation between the two. Therefore, the embodiment of the present application also supports the use of a contrastive learning algorithm, so that the industry features corresponding to the training sample are used to train their own word embeddings to make up for the shortcomings of the training sample, thereby reducing the impact of the sample selectivity bias and further improving the prediction accuracy of the indicator prediction model.

[0119] In one example, if Figure 5 As shown, step 304 may also include the following sub-steps:

[0120] Step 304a, constructing positive sample encoding and negative sample encoding corresponding to the training samples according to the encoding information of the first sequence features and the encoding information of the second sequence features corresponding to each training sample, wherein the positive sample encoding refers to the encoding information of the positive sample corresponding to the recommended object, and the negative sample encoding refers to the encoding information of the negative sample corresponding to the recommended object.

[0121] Optionally, positive samples corresponding to the recommended objects may be constructed based on the first sequence features, and then negative samples may be performed on the positive samples corresponding to the recommended objects to obtain negative samples corresponding to the recommended objects.

[0122] Exemplarily, the combination of the first subsequence feature and the second sequence feature is the positive sample corresponding to the recommended object. Alternatively, when the training sample is a positive sample (such as an advertisement is exposed and clicked), the combination of the first subsequence feature and the third sequence feature in the training sample is the positive sample corresponding to the recommended object. Optionally, the first subsequence feature and the second sequence feature are cross-dispersed to obtain a negative sample corresponding to the recommended object. Alternatively, when the training sample is a positive sample, the first subsequence feature and the third sequence feature in the training sample are cross-dispersed to obtain a negative sample corresponding to the recommended object.

[0123] In an embodiment of the present application, after obtaining the encoding information of the training samples, the encoding information of the training samples can be directly used to construct positive sample encoding and negative sample encoding, which can reduce the complexity of the training samples and reduce the workload of obtaining the encoding information.

[0124] In one example, for each training sample, a method for obtaining the corresponding positive sample code includes but is not limited to at least one of the following:

[0125] 1. Construct a positive sample encoding based on the encoding information of the first subsequence feature and the encoding information of the second subsequence feature corresponding to the training sample.

[0126] Optionally, the encoding information of the first subsequence feature and the encoding information of the second subsequence feature are combined to obtain the positive sample encoding.

[0127] 2. Construct a positive sample encoding based on the encoding information of the first subsequence feature and the encoding information of the third subsequence feature corresponding to the training sample.

[0128] Optionally, when the training sample is a positive sample, the encoding information of the first subsequence feature and the encoding information of the third subsequence feature are combined to obtain the positive sample encoding. The negative sample of the indicator prediction model is not used to construct the positive sample encoding corresponding to the recommended object.

[0129] For each training sample, the method for obtaining the corresponding negative sample encoding includes but is not limited to at least one of the following:

[0130] 1. The encoding information of the first subsequence feature and the encoding information of the second subsequence feature corresponding to the training sample are cross-scattered to construct the negative sample encoding.

[0131] Optionally, the cross-breaking may refer to diagonal crossing and element interchange, or may refer to random element interchange, or may refer to element interchange at a specified position, which is not limited in the embodiments of the present application.

[0132] For example, refer to Figure 6, each line in the first subsequence feature 601 and the second subsequence feature 602 is the encoding information of an industry feature. By replacing the elements in the first subsequence feature 601 (the encoding information of the industry feature) and the elements in the second subsequence feature 602 in the form of diagonal crossing, the negative sample encoding can be obtained.

[0133] 2. The encoding information of the first subsequence feature and the encoding information of the third subsequence feature corresponding to the training sample are cross-scattered to construct the negative sample encoding.

[0134] Exemplarily, the elements in the first subsequence feature (encoded information of the industry feature) and the elements in the third subsequence feature are replaced in the form of diagonal crossing to obtain the negative sample encoding. Optionally, both the positive samples and negative samples of the indicator prediction model can be used to construct the negative sample encoding corresponding to the recommended object.

[0135] The embodiment of the present application constructs negative sample encoding in the form of cross-scattering, which can make the indicator prediction model learn the industry characteristics more fully, and also make the encoding information of the recommended object more comprehensive and accurate.

[0136] In one example, the above-mentioned method for obtaining positive sample codes and the method for obtaining negative sample codes can be freely combined and applied. The embodiment of the present application does not limit the number of positive sample codes or the number of negative sample codes, and the ratio of positive sample codes to negative sample codes. It can be set and adjusted according to actual usage requirements.

[0137] Optionally, in the case where the training sample has the first sub-coding information and the second sub-coding information, the positive sample encoding and the negative sample encoding of the training sample under the first sub-coding information can be constructed according to the first sub-coding information, or the positive sample encoding and the negative sample encoding of the training sample under the second sub-coding information can be constructed according to the second sub-coding information. The use methods of the positive sample encoding and the negative sample encoding in the two forms will be described in detail below and will not be repeated here.

[0138] In one example, before obtaining the encoding information of the training sample, the negative sample corresponding to the training sample can be obtained. Here, the negative sample corresponding to the training sample is the negative sample corresponding to the recommended object. If the hollowing method is adopted, the negative sample corresponding to the training sample is obtained according to the first sequence feature. The process may include the following contents:

[0139] For each training sample, obtain the first industry feature in the first sequence of features corresponding to the training sample; move the first industry feature to the end of the first sequence of features to obtain the negative sample of the recommended object corresponding to the training sample. The first industry feature may refer to any industry feature in the first sequence of features. While obtaining the encoding information corresponding to each training sample, it is also necessary to obtain the negative sample encoding corresponding to each training sample according to the negative sample of the recommended object corresponding to each training sample.

[0140] Step 304b, according to the positive sample codes and negative sample codes corresponding to each training sample, respectively, the contrast loss of the indicator prediction model is obtained, and the contrast loss is used to enable the indicator prediction model to have the ability to learn the features corresponding to the negative sample codes.

[0141] Contrastive loss can be used to cluster positive sample encodings and negative sample encodings so that the indicator prediction model can distinguish between positive and negative samples of recommended objects. At the same time, contrastive loss can enable the indicator prediction model to learn industry characteristics in negative samples that positive samples do not have, thereby ensuring the accuracy and comprehensiveness of the encoded information. In this way, the indicator prediction model can understand the encoded information outside the training samples, thereby solving the sample bias problem and improving the prediction accuracy of the indicator prediction model.

[0142] In addition, after the industry characteristics of the recommended objects are learned more fully, the accuracy of the encoded information of the recommended objects will also be improved. Since the second-sequence characteristics of the targeted content and the first-sequence characteristics of the recommended objects are used as inputs of the indicator prediction model together, the accuracy of the corresponding predicted recommendation indicators will also be improved.

[0143] In the embodiment of the present application, the encoding information corresponding to the training sample includes the first sub-encoding information and the second sub-encoding information, and the method for obtaining the contrast loss may include but is not limited to at least one of the following:

[0144] 1. According to the positive sample encoding and the negative sample encoding constructed by the first sub-encoding information or the second sub-encoding information corresponding to each training sample, the contrast loss is obtained.

[0145] Optionally, for each training sample, the positive sample encoding and negative sample encoding constructed by the first sub-encoding information can be uniformly used to calculate the contrast loss, or the positive sample encoding and negative sample encoding constructed by the second sub-encoding information can be uniformly used to calculate the contrast loss. The embodiments of the present application are not limited to this.

[0146] 2. For each training sample, the positive sample code constructed by the first sub-coding information of the training sample is concatenated with the positive sample code constructed by the second sub-coding information of the training sample to obtain the final positive sample code corresponding to the training sample, and the negative sample code constructed by the first sub-coding information of the training sample is concatenated with the negative sample code constructed by the second sub-coding information of the training sample to obtain the final negative sample code corresponding to the training sample; the contrast loss is obtained according to the final positive sample code and the final negative sample code corresponding to each training sample.

[0147] Exemplarily, for a certain training sample, the training sample corresponds to: positive sample code 1 constructed by the first sub-coding information, positive sample code 2 constructed by the second sub-coding information, negative sample code 3 constructed by the first sub-coding information, and negative sample code 4 constructed by the second sub-coding information. The positive sample code 2 constructed by the second sub-coding information is spliced ​​after the positive sample code 1 constructed by the first sub-coding information to obtain the final positive sample code 5, and the negative sample code 4 constructed by the second sub-coding information is spliced ​​after the negative sample code 3 constructed by the first sub-coding information to obtain the final negative sample code 6. According to the final positive sample code 5 and the final negative sample code 6, the contrast loss is calculated.

[0148] 3. According to the positive sample encoding and negative sample encoding constructed by the first sub-encoding information corresponding to each training sample, a first sub-contrast loss is obtained, and the positive sample encoding and negative sample encoding constructed by the second sub-encoding information corresponding to each training sample, a second sub-contrast loss is obtained; the first sub-contrast loss and the second sub-contrast loss are summed to obtain the contrast loss.

[0149] That is, the first sub-coding information corresponds to a contrast loss, the second sub-coding information corresponds to a contrast loss, and the sum of the contrast losses of the two is the final contrast loss.

[0150] Exemplarily, based on the above embodiment, a first sub-contrast loss is obtained based on the positive sample encoding 1 and the negative sample encoding 3 constructed by the first sub-encoding information, and a second sub-contrast loss is obtained based on the positive sample encoding 2 and the negative sample encoding 4 constructed by the second sub-encoding information. The first sub-contrast loss and the second sub-contrast loss are summed to obtain the contrast loss.

[0151] For example, reference Figure 7Curve 701 is the AUC (Area Under Curve, ROC (Receiver Operating Characteristic)) change curve corresponding to the above-mentioned contrast loss acquisition method 1, curve 702 is the AUC change curve corresponding to the above-mentioned contrast loss acquisition method 2, and curve 703 is the AUC change curve corresponding to the above-mentioned contrast loss acquisition method 3. It can be seen that the effect of the contrast loss acquisition method 3 is better than that of the contrast loss acquisition method 2, and the effect of the contrast loss acquisition method 2 is better than that of the contrast loss acquisition method 1.

[0152] In the embodiment of the present application, the contrast loss can be constructed using the loss function corresponding to contrast learning, and the process can be as follows:

[0153] 1. For each training sample, encode the positive sample corresponding to the training sample and multiply it by the encoding information of the third subsequence feature corresponding to the training sample to obtain a first intermediate parameter; divide the first intermediate parameter and the learnable temperature coefficient to obtain a second intermediate parameter.

[0154] Optionally, the second intermediate parameter can be expressed as follows:; wherein q is the encoding information of the third subsequence feature, is the positive sample encoding, and is the learnable temperature coefficient.

[0155] 2. Perform an exponential operation using the second intermediate parameter as an exponent to obtain a third intermediate parameter; and obtain a first sum value according to the third intermediate parameters corresponding to each training sample.

[0156] Optionally, an exponential operation may be performed with the natural constant e as the base and the second intermediate parameter as the exponent to obtain a third intermediate parameter, and then the third intermediate parameters corresponding to each training sample are summed to obtain the first sum value.

[0157] Exemplarily, the first sum value can be expressed as follows:.

[0158] 3. Encode the negative sample corresponding to the training sample, and multiply it by the encoding information of the third subsequence feature corresponding to the training sample to obtain a fourth intermediate parameter; divide the fourth intermediate parameter and the temperature coefficient to obtain a fifth intermediate parameter.

[0159] Optionally, the fifth intermediate parameter can be expressed as follows:; wherein q is the encoding information of the third subsequence feature, which is the negative sample encoding.

[0160] 4. Perform an exponential operation using the fifth intermediate parameter as an exponent to obtain a sixth intermediate parameter; and obtain a second sum value based on the sixth intermediate parameters corresponding to each training sample.

[0161] Exemplarily, the second sum can be expressed as follows: .

[0162] 5. Sum the first sum and the second sum to obtain the third sum.

[0163] Exemplarily, the third sum may be expressed as follows:

[0164]

[0165] Where n is the total number of positive sample encodings and negative sample encodings.

[0166] 6. Divide the first sum and the third sum to obtain a seventh intermediate parameter; based on the seventh intermediate parameter, obtain the contrast loss of the indicator prediction model.

[0167] Exemplarily, the contrast loss can be expressed by the following formula:

[0168]

[0169] Optionally, for the above-mentioned method 1 for obtaining the contrast loss, the positive sample encoding and the negative sample encoding of each training sample under the first sub-coding information or the second sub-coding information are substituted into the above formula to obtain the contrast loss; for the above-mentioned method 2 for obtaining the contrast loss, the final positive sample encoding and the final negative sample encoding corresponding to each training sample are substituted into the above formula to obtain the contrast loss; for the above-mentioned method 3 for obtaining the contrast loss, the positive sample encoding and the negative sample encoding of each training sample under the first sub-coding information are substituted into the above formula to obtain the first sub-contrast loss, and the positive sample encoding and the negative sample encoding of each training sample under the second sub-coding information are substituted into the above formula to obtain the second sub-contrast loss.

[0170] Among them, the encoding information of the third subsequence feature corresponds to the positive sample code and the negative sample code. If the positive sample code and the negative sample code correspond to the first sub-coding information, the encoding information of the third subsequence feature is obtained based on the first sub-coding information; if the positive sample code and the negative sample code correspond to the second sub-coding information, the encoding information of the third subsequence feature is obtained based on the second sub-coding information.

[0171] Step 304c, according to the predicted recommendation index and the marked recommendation index corresponding to each training sample, the main task loss of the index prediction model is obtained, and the main task loss is used to indicate the prediction accuracy of the index prediction model for the recommended index.

[0172] Optionally, a cross entropy loss function, a focal loss function, a mean square error loss function, etc. may be used to calculate the main task loss.

[0173] Taking the cross entropy loss function as an example, the main task loss can be expressed as follows:

[0174]

[0175] Among them, m is the total number of training samples, is the label recommendation index of the i-th training sample, is the prediction recommendation indicator for the i-th training sample.

[0176] Step 304d: train the indicator prediction model according to the contrast loss and the main task loss of the indicator prediction model to obtain a trained indicator prediction model.

[0177] Optionally, the contrast loss of the indicator prediction model and the main task loss are summed to obtain the total loss of the indicator prediction model; with the goal of minimizing the total loss, the parameters of the indicator prediction model are iteratively adjusted to obtain the trained indicator prediction model. The indicator prediction model after iterative training can also be referred to as a trained indicator prediction model.

[0178] In summary, the technical solution provided by the embodiment of the present application, by constructing the first sequence features of the recommended object based on the industry features related to the recommended object, and constructing the second sequence features of the targeted content based on the industry features related to the targeted content, realizes the association of the recommended object and the targeted content with the industry features corresponding to the targeted object (such as goods) in the industry (such as the e-commerce industry), thereby improving the correlation between the first sequence features of the recommended object and the second sequence features of the targeted content, as well as the correlation between the first sequence features of the recommended object and the second sequence features of the targeted content and the industry corresponding to the targeted object. Then, the first sequence features of the recommended object and the second sequence features of the targeted content are used to train the indicator prediction model, which can improve the prediction accuracy of the indicator prediction model, thereby improving the prediction accuracy of the recommendation indicator of the recommended object for the targeted content in the industry corresponding to the targeted object.

[0179] In addition, by adopting the contrastive learning method, the industry characteristics corresponding to the negative sample encoding are used to train the encoding information of the recommended object to make up for the shortcomings of the training samples. After the industry characteristics of the recommended object are learned more fully, the accuracy of the encoding information of the recommended object will also be improved. Since the second sequence characteristics of the targeted content and the first sequence characteristics of the recommended object are used as training samples of the indicator prediction model together, the accuracy of the encoding information of the training sample will also be improved, which is conducive to reducing the impact of the sample selectivity bias and thereby improving the prediction accuracy of the indicator prediction model.

[0180] In addition, two types of sub-coding information are generated for each training sample, and the first sub-contrast loss and the second sub-contrast loss are constructed according to the positive sample coding and negative sample coding under the two types of sub-coding information, respectively, so as to obtain the contrast loss, and then the contrast loss is used to assist the indicator prediction model for training, which can effectively improve the prediction accuracy of the standard prediction model.

[0181] The trained indicator prediction model can be deployed in the model using device to provide recommended indicator prediction services. The following describes the method for using the trained indicator prediction model. For details not disclosed in the method embodiments of the present application, please refer to the above embodiments.

[0182] Please refer to Figure 8 , which shows a flow chart of a method for predicting a recommendation index provided by another embodiment of the present application. The execution subject of each step of the method may be the model using device introduced above. The method may include the following steps (801-803).

[0183] Step 801, obtaining input data, which includes a first sequence feature of a recommended object and a second sequence feature of targeted content, wherein the first sequence feature includes at least one industry feature related to the recommended object, and the second sequence feature includes at least one industry feature related to the targeted content, and the industry feature is used to characterize the targeted object corresponding to the targeted content, and identification information in the industry corresponding to the targeted object.

[0184] In the application embodiment, the recommended object corresponding to the input data can be a specified object to which the content is to be recommended, such as in an advertisement recommendation scenario, an advertisement can be recommended to the recommended object. The first sequence feature is used to characterize the recommended object. Optionally, the industry feature in the first sequence feature has a binding relationship with the recommended object, that is, the industry feature in the first sequence feature is strongly associated with the recommended object, and the binding relationship can be determined by manual annotation.

[0185] Targeted content may refer to content to be recommended to a recommendation object. For example, in an advertisement recommendation scenario, the targeted content may be a candidate advertisement recalled based on the first sequence feature. The second sequence feature is used to characterize the targeted content. Optionally, the industry feature in the second sequence feature has a binding relationship with the targeted content, and the binding relationship may also be determined by manual annotation.

[0186] Among them, the industry characteristics are related to the targeted object and the industry corresponding to the targeted object, which can reflect the characteristics of the targeted object and the characteristics of the industry in which the targeted object is located.

[0187] In one example, the process of acquiring input data can be as follows: acquiring at least one industry feature that has a binding relationship with the recommended object to construct a first sequence feature; acquiring at least one industry feature that has a binding relationship with the targeted content to construct a second sequence feature; acquiring input data based on the first sequence feature and the second sequence feature.

[0188] Among them, each industry feature in the first sequence of features can be a manually labeled feature, and each industry feature in the second sequence of features can also be a manually labeled feature, which is not limited in the embodiments of the present application.

[0189] Step 802: Obtain encoding information corresponding to the input data according to the first sequence feature and the second sequence feature included in the input data.

[0190] The coding information in the embodiment of the present application is used to characterize the input data. Optionally, the coding information corresponding to each industry feature in the input data is first obtained through the vocabulary, and then the coding information corresponding to each industry feature in the first sequence feature is combined to obtain the coding information of the first sequence feature, which can be used to characterize the recommended object, and the coding information corresponding to each industry feature in the second sequence feature is combined to obtain the coding information of the second sequence feature, which can be used to characterize the targeted content, and finally the coding information of the first sequence feature and the coding information of the second sequence feature are combined to obtain the coding information corresponding to the input data.

[0191] In one example, the encoding information corresponding to the input data includes at least one encoding group, each encoding group includes first sub-encoding information and second sub-encoding information, the first sub-encoding information and the second sub-encoding information are both used to represent the input data, and the first sub-encoding information and the second sub-encoding information share the same vocabulary.

[0192] For example, reference Figure 2 , the input data has a coding group under each vocabulary, that is, there are 4 coding groups corresponding to the input data, each coding group can include two sub-coding information (namely the first sub-coding information and the second sub-coding information), and the two sub-coding information are used to characterize the input data. In this way, by obtaining the prediction recommendation index through multiple coding groups, the features extracted by the index prediction model can be enriched, thereby improving the accuracy of the prediction recommendation index.

[0193] Step 803, determining the predicted recommended index corresponding to the input data according to the coding information corresponding to the input data through the index prediction model, where the predicted recommended index refers to the predicted value of the recommended index.

[0194] When the predicted recommendation index is greater than or equal to the index threshold, the targeted content can be recommended to the recommended object; when the predicted recommendation index is less than the index threshold, the targeted content is not recommended to the recommended object; the index threshold can be set and adjusted according to actual usage needs. Optionally, in the case of multiple targeted contents, after obtaining the predicted recommendation index corresponding to each targeted content, the multiple targeted contents can be sorted according to the predicted recommendation index to obtain the targeted content with the highest predicted recommendation index and recommend it to the user.

[0195] In one example, the process of obtaining the prediction recommendation index can be as follows: for each coding group, multiple sub-coding information in the coding group is spliced ​​to obtain the spliced ​​coding information corresponding to the coding group; through the indicator prediction model, the prediction recommendation index corresponding to the input data is determined according to the spliced ​​coding information corresponding to at least one coding group.

[0196] For example, refer to Figure 2 , respectively concatenate the two sub-coding information corresponding to each vocabulary, and obtain the concatenated coding information of the input data under vocabulary 1, vocabulary 2, vocabulary 3 and vocabulary 4, and then use the concatenated coding information of the input data under vocabulary 1, vocabulary 2, vocabulary 3 and vocabulary 4 as the input of the prediction network 302, and the prediction network 302 can output the prediction recommendation index corresponding to the input data.

[0197] In summary, the technical solution provided in the embodiment of the present application, by constructing the first sequence characteristics of the recommended object based on the industry characteristics related to the recommended object, and constructing the second sequence characteristics of the targeted content based on the industry characteristics related to the targeted content, realizes the association of the recommended object and the targeted content with the industry characteristics corresponding to the targeted object (such as the commodity) in the industry (such as the e-commerce industry), thereby improving the correlation between the first sequence characteristics of the recommended object and the second sequence characteristics of the targeted content, as well as the correlation between the first sequence characteristics of the recommended object and the second sequence characteristics of the targeted content and the industry corresponding to the targeted object, and then using the first sequence characteristics of the recommended object and the second sequence characteristics of the targeted content to predict the recommendation index, which can improve the prediction accuracy of the recommendation index.

[0198] In some embodiments, the indicator prediction model is used to obtain the predicted click rate, the targeted content is an advertisement, and the encoded information is a word embedding. Fig. 9 , the technical solution provided by the embodiment of the present application is described, which may include the following contents. Among them, the contents not described in the embodiment of the present application can refer to the above embodiment.

[0199] The model training device obtains training data of the indicator prediction model, the training data includes a plurality of training samples with marked click-through rates, the training samples may include a first feature sequence of a sample user and a second feature sequence of a sample advertisement. Optionally, the sample user and the sample advertisement may have an interactive relationship, or may not have an interactive relationship, and the marked click-through rate refers to a marked value of the click-through rate.

[0200] The first feature sequence includes industry characteristics of commodities associated with sample users (such as those with interactive behaviors, those of interest to users, and those bound to them) within a certain period of time, and the second feature sequence includes industry characteristics of commodities associated with sample advertisements (such as those bound to them).

[0201] The model training device obtains the first word embedding (i.e., first sub-coding information) and the second word embedding (i.e., second sub-coding information) of each training sample in each vocabulary (vocabulary 1-vocabulary 4) through the embedding layer 801 in the indicator prediction model.

[0202] The model training device obtains the predicted click rate corresponding to each training sample according to the first word embedding and the second word embedding of each training sample in each vocabulary through the prediction network 802 in the indicator prediction model. The predicted click rate refers to the predicted value of the click rate.

[0203] The model training device constructs the positive sample word embedding (i.e., positive sample encoding) and negative sample word embedding (i.e., negative sample encoding) of each training sample under the first word embedding, the positive sample word embedding and negative sample word embedding of each training sample under the second word embedding, and the word embedding of the third sub-feature sequence corresponding to each training sample under the first word embedding and the second word embedding through the first word embedding and the second word embedding of each training sample under the vocabulary 2. Optionally, the vocabulary 2 corresponds to the domain factor decomposition layer in the prediction network 802.

[0204] The model training device obtains a first sub-contrast loss according to the positive sample word embedding and the negative sample word embedding of each training sample under the first word embedding, and the word embedding of the third sub-feature sequence corresponding to each training sample under the first word embedding.

[0205] The model training device obtains a second sub-contrast loss according to the positive sample word embedding and the negative sample word embedding of each training sample under the second word embedding, and the word embedding of the third sub-feature sequence corresponding to each training sample under the second word embedding.

[0206] The model training device adds the first sub-contrast loss and the second sub-contrast loss to obtain the contrast loss of the indicator prediction model.

[0207] The model training device obtains the main task loss of the indicator prediction model according to the predicted click rate and the marked click rate corresponding to each training sample.

[0208] The model training device iteratively trains the indicator prediction model according to the contrast loss and the main task loss to obtain a trained indicator prediction model.

[0209] The model uses the indicator prediction model that has been trained and deployed on the device.

[0210] In the e-commerce advertising ranking scenario, the model uses the indicator prediction model trained by the device to obtain the predicted click-through rate between the target user and each candidate advertisement based on the first sequence characteristics of the target user and the second sequence characteristics of each candidate advertisement, and sorts each candidate advertisement in descending order of predicted click-through rate to obtain a candidate advertisement sequence. The target user refers to the user to whom the advertisement is to be recommended.

[0211] The model uses a device to recommend the first candidate advertisement in the candidate advertisement sequence to the target user, or recommends the first n candidate advertisements in the candidate advertisement sequence to the target user.

[0212] In some embodiments, the technical solution provided in the embodiments of the present application can bring a 2.21% increase in GMV (Gross Merchandise Volume) in the e-commerce industry, and a 1.47% increase in consumption in the e-commerce industry. At the same time, the AUC corresponding to the indicator prediction model using the technical solution provided in the embodiments of the present application has also increased by one thousandth, with a significant improvement effect.

[0213] In some embodiments, reference Fig.10 Curve 1001 is the AUC change curve corresponding to the related technology, curve 1002 is the AUC change curve corresponding to the technical solution provided in the embodiment of the present application, and area 1003 is the improvement degree of the technical solution provided in the embodiment of the present application relative to the related technology. It can be seen that, relative to the related technology, the performance of the indicator prediction model using the technical solution provided in the embodiment of the present application is significantly improved.

[0214] In summary, by adopting the technical solution provided in the embodiments of the present application, the prediction accuracy of the indicator prediction model can be effectively improved, thereby increasing the click-through rate of advertisements, which in turn is conducive to improving the conversion rate of advertisements.

[0215] The following is an embodiment of the device of the present application, which can be used to execute the embodiment of the method of the present application. For details not disclosed in the embodiment of the device of the present application, please refer to the embodiment of the method of the present application.

[0216] Please refer to Fig.11, which shows a block diagram of a device for predicting a recommendation index provided by an embodiment of the present application. The device can be used to implement the above-mentioned method for predicting a recommendation index. The device 1100 may include: a training sample acquisition module 1101, a coding information acquisition module 1102, a recommendation index prediction module 1103 and a prediction model training module 1104.

[0217] The training sample acquisition module 1101 is used to acquire multiple training samples, wherein the training samples include a first sequence feature of a recommended object and a second sequence feature of targeted content, wherein the first sequence feature includes at least one industry feature related to the recommended object, and the second sequence feature includes at least one industry feature related to the targeted content.

[0218] The coding information acquisition module 1102 is used to acquire, for each of the training samples, coding information corresponding to the training sample according to the first sequence feature and the second sequence feature included in the training sample.

[0219] The recommendation index prediction module 1103 is used to determine the predicted recommendation index corresponding to each of the training samples according to the encoding information corresponding to each of the training samples through the index prediction model, and the predicted recommendation index refers to the predicted value of the recommendation index.

[0220] The prediction model training module 1104 is used to train the indicator prediction model according to the coding information, prediction recommendation indicators and marked recommendation indicators corresponding to each of the training samples, so as to obtain a trained indicator prediction model. The trained indicator prediction model is used to obtain the prediction recommendation indicator of the recommended object for the targeted content. The marked recommendation indicator refers to the mark value of the recommendation indicator; wherein the industry characteristics are used to characterize the targeted object corresponding to the targeted content, and the identification information in the industry corresponding to the targeted object.

[0221] In some embodiments, the coding information corresponding to the training sample includes the coding information of the first sequence feature and the coding information of the second sequence feature; Fig.12 As shown, the prediction model training module 1104 includes: a positive and negative sample construction submodule 1104a, a contrast loss acquisition submodule 1104b, a task loss acquisition submodule 1104c and a prediction model training submodule 1104d.

[0222] The positive and negative sample construction submodule 1104a is used to construct the positive sample coding and negative sample coding corresponding to the training samples according to the coding information of the first sequence features and the coding information of the second sequence features corresponding to each of the training samples, wherein the positive sample coding refers to the coding information of the positive sample corresponding to the recommended object, and the negative sample coding refers to the coding information of the negative sample corresponding to the recommended object.

[0223] The contrast loss acquisition submodule 1104b is used to obtain the contrast loss of the indicator prediction model according to the positive sample encoding and the negative sample encoding corresponding to each of the training samples, and the contrast loss is used to enable the indicator prediction model to have the ability to learn the features corresponding to the negative sample encoding.

[0224] The task loss acquisition submodule 1104c is used to obtain the main task loss of the indicator prediction model according to the predicted recommendation indicators and marked recommendation indicators corresponding to each of the training samples, and the main task loss is used to indicate the prediction accuracy of the indicator prediction model for the recommended indicator.

[0225] The prediction model training submodule 1104d is used to train the indicator prediction model according to the contrast loss and the main task loss of the indicator prediction model to obtain the trained indicator prediction model.

[0226] In some embodiments, the first sequence feature includes a first subsequence feature and a second subsequence feature, the industry feature in the first subsequence feature has a binding relationship with the recommended object, the second subsequence feature includes the first subsequence feature, and the second subsequence feature includes an industry feature that has a non-binding relationship with the recommended object; the second sequence feature includes a third subsequence feature and a fourth subsequence feature, the industry feature in the third subsequence feature has a binding relationship with the targeted content, and the fourth subsequence feature is used to characterize the consumption of the targeted content;

[0227] In some embodiments, the positive and negative sample construction submodule 1104a is used to:

[0228] For each of the training samples, construct the positive sample code according to the coding information of the first subsequence feature and the coding information of the second subsequence feature corresponding to the training sample; or construct the positive sample code according to the coding information of the first subsequence feature and the coding information of the third subsequence feature corresponding to the training sample;

[0229] The coding information of the first subsequence feature and the coding information of the second subsequence feature corresponding to the training sample are cross-dispersed to construct the negative sample coding; or the coding information of the first subsequence feature and the coding information of the third subsequence feature corresponding to the training sample are cross-dispersed to construct the negative sample coding.

[0230] In some embodiments, the encoding information corresponding to the training sample includes first sub-encoding information and second sub-encoding information, the first sub-encoding information and the second sub-encoding information are both used to represent the training sample, and the first sub-encoding information and the second sub-encoding information share the same vocabulary;

[0231] The contrast loss acquisition submodule 1104b is used to obtain the contrast loss according to the positive sample encoding and the negative sample encoding constructed by the first sub-encoding information or the second sub-encoding information corresponding to each of the training samples.

[0232] Alternatively, the contrast loss acquisition submodule 1104b is also used to, for each of the training samples, concatenate the positive sample code constructed by the first sub-coding information of the training sample with the positive sample code constructed by the second sub-coding information of the training sample to obtain the final positive sample code corresponding to the training sample, and concatenate the negative sample code constructed by the first sub-coding information of the training sample with the negative sample code constructed by the second sub-coding information of the training sample to obtain the final negative sample code corresponding to the training sample; and obtain the contrast loss according to the final positive sample code and the final negative sample code corresponding to each of the training samples.

[0233] Alternatively, the contrast loss acquisition submodule 1104b is also used to obtain a first sub-contrast loss based on the positive sample encoding and the negative sample encoding constructed by the first sub-encoding information corresponding to each of the training samples, and to obtain a second sub-contrast loss based on the positive sample encoding and the negative sample encoding constructed by the second sub-encoding information corresponding to each of the training samples; the first sub-contrast loss and the second sub-contrast loss are summed to obtain the contrast loss.

[0234] In some embodiments, the contrast loss acquisition submodule 1104b is further used to:

[0235] For each of the training samples, encode the positive sample corresponding to the training sample, and multiply the encoded information of the third subsequence feature corresponding to the training sample to obtain a first intermediate parameter;

[0236] Dividing the first intermediate parameter and the learnable temperature coefficient to obtain a second intermediate parameter;

[0237] Performing an exponential operation using the second intermediate parameter as an exponent to obtain a third intermediate parameter;

[0238] Obtaining a first sum value according to the third intermediate parameters respectively corresponding to the training samples;

[0239] Encoding the negative sample corresponding to the training sample, and multiplying the encoded information of the third subsequence feature corresponding to the training sample to obtain a fourth intermediate parameter;

[0240] Dividing the fourth intermediate parameter and the temperature coefficient to obtain a fifth intermediate parameter;

[0241] Performing an exponential operation using the fifth intermediate parameter as an exponent to obtain a sixth intermediate parameter;

[0242] Obtaining a second sum value according to the sixth intermediate parameters corresponding to the training samples;

[0243] Summing the first sum value and the second sum value to obtain a third sum value;

[0244] Dividing the first sum value and the third sum value to obtain a seventh intermediate parameter;

[0245] According to the seventh intermediate parameter, the contrast loss of the indicator prediction model is obtained.

[0246] In some embodiments, the prediction model training submodule 1104d is used to:

[0247] Summing the contrast loss of the indicator prediction model and the main task loss to obtain the total loss of the indicator prediction model;

[0248] With the goal of minimizing the total loss, the parameters of the indicator prediction model are iteratively adjusted to obtain the trained indicator prediction model.

[0249] In some embodiments, the identification information of the recommended object and the first sequence feature are stored in a database in the form of a key-value pair, and the identification information of the targeted content and the second sequence feature are stored in the database in the form of a key-value pair; the training sample acquisition module 1101 is used to:

[0250] Acquiring the first sequence feature from the database according to the identification information of the recommended object;

[0251] Acquiring the first sequence feature from the database according to the identification information of the recommended object;

[0252] acquiring the second sequence feature from the database according to the identification information of the targeted content;

[0253] The first sequence feature and the second sequence feature are combined to obtain the training sample.

[0254] In some embodiments, Fig.12As shown, the device 1100 also includes: an industry feature selection module 1105 and a negative sample construction module 1106.

[0255] The industry feature selection module 1105 is used to obtain, for each of the training samples, the first industry feature in the first sequence features included in the training sample.

[0256] The negative sample construction module 1106 is used to move the first industry feature to the end of the first sequence feature to obtain a negative sample of the recommended object corresponding to the training sample.

[0257] The coding information acquisition module 1102 is further configured to acquire negative sample codes corresponding to each training sample according to the negative samples of the recommended object corresponding to each training sample.

[0258] In summary, the technical solution provided by the embodiment of the present application, by constructing the first sequence features of the recommended object based on the industry features related to the recommended object, and constructing the second sequence features of the targeted content based on the industry features related to the targeted content, realizes the association of the recommended object and the targeted content with the industry features corresponding to the targeted object (such as goods) in the industry (such as the e-commerce industry), thereby improving the correlation between the first sequence features of the recommended object and the second sequence features of the targeted content, as well as the correlation between the first sequence features of the recommended object and the second sequence features of the targeted content and the industry corresponding to the targeted object. Then, the first sequence features of the recommended object and the second sequence features of the targeted content are used to train the indicator prediction model, which can improve the prediction accuracy of the indicator prediction model, thereby improving the prediction accuracy of the recommendation indicator of the recommended object for the targeted content in the industry corresponding to the targeted object.

[0259] Please refer to Fig.13 , which shows a block diagram of a device for predicting a recommendation index provided by another embodiment of the present application. The device can be used to implement the above-mentioned method for predicting a recommendation index. The device 1300 may include: an input data acquisition module 1301, a coding information acquisition module 1302, and a recommendation index prediction module 1303.

[0260] The input data acquisition module 1301 is used to acquire input data, wherein the input data includes a first sequence feature of a recommended object and a second sequence feature of targeted content, wherein the first sequence feature includes at least one industry feature related to the recommended object, and the second sequence feature includes at least one industry feature related to the targeted content.

[0261] The coding information acquisition module 1302 is used to acquire the coding information corresponding to the input data according to the first sequence feature and the second sequence feature included in the input data.

[0262] The recommendation index prediction module 1303 is used to determine the predicted recommendation index corresponding to the input data according to the encoding information corresponding to the input data through the index prediction model, and the predicted recommendation index refers to the predicted value of the recommendation index; wherein the industry characteristics are used to characterize the targeted object corresponding to the targeted content, and the identification information in the industry corresponding to the targeted object.

[0263] In some embodiments, the input data acquisition module 1301 is used to:

[0264] Acquire at least one industry feature having a binding relationship with the recommended object to construct the first sequence feature;

[0265] Acquire at least one industry feature having a binding relationship with the targeted content to construct the second sequence feature;

[0266] The input data is obtained according to the first sequence feature and the second sequence feature.

[0267] In some embodiments, the coding information corresponding to the input data includes at least one coding group, each of the coding groups includes first sub-coding information and second sub-coding information, the first sub-coding information and the second sub-coding information are both used to represent the input data, and the first sub-coding information and the second sub-coding information share the same vocabulary; the recommendation index prediction module 1303 is used to:

[0268] For each of the coding groups, concatenating the first sub-coding information and the second sub-coding information in the coding group to obtain concatenated coding information corresponding to the coding group;

[0269] The prediction recommendation index corresponding to the input data is determined by the index prediction model according to the splicing coding information respectively corresponding to the at least one coding group.

[0270] In summary, the technical solution provided in the embodiment of the present application, by constructing the first sequence characteristics of the recommended object based on the industry characteristics related to the recommended object, and constructing the second sequence characteristics of the targeted content based on the industry characteristics related to the targeted content, realizes the association of the recommended object and the targeted content with the industry characteristics corresponding to the targeted object (such as the commodity) in the industry (such as the e-commerce industry), thereby improving the correlation between the first sequence characteristics of the recommended object and the second sequence characteristics of the targeted content, as well as the correlation between the first sequence characteristics of the recommended object and the second sequence characteristics of the targeted content and the industry corresponding to the targeted object, and then using the first sequence characteristics of the recommended object and the second sequence characteristics of the targeted content to predict the recommendation index, which can improve the prediction accuracy of the recommendation index.

[0271] It should be noted that the device provided in the above embodiment, when implementing its functions, is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiment belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0272] Please refer to Fig.14 , which shows a schematic diagram of the structure of a computer device provided by an embodiment of the present application. The computer device can be any electronic device with data calculation, processing and storage functions, and the computer device can be implemented as Figure 1 The model training device 10 or the model using device 20 in the implementation environment of the illustrated solution may specifically include the following contents.

[0273] The computer device 1400 includes a central processing unit (such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit) and an FPGA (Field Programmable Gate Array)) 1401, a system memory 1404 including a RAM (Random-Access Memory) 1402 and a ROM (Read-Only Memory) 1403, and a system bus 1405 connecting the system memory 1404 and the central processing unit 1401. The computer device 1400 also includes a basic input / output system (Input Output System, I / O system) 1406 for helping various devices in the server to transmit information, and a large-capacity storage device 1407 for storing an operating system 1413, application programs 1414 and other program modules 1415.

[0274] In some embodiments, the basic input / output system 1406 includes a display 1408 for displaying information and an input device 1409 such as a mouse and a keyboard for user inputting information. The display 1408 and the input device 1409 are connected to the central processing unit 1401 through an input / output controller 1410 connected to the system bus 1405. The basic input / output system 1406 may also include an input / output controller 1410 for receiving and processing inputs from a plurality of other devices such as a keyboard, a mouse, or an electronic stylus. Similarly, the input / output controller 1410 also provides output to a display screen, a printer, or other types of output devices.

[0275] The mass storage device 1407 is connected to the central processing unit 1401 via a mass storage controller (not shown) connected to the system bus 1405. The mass storage device 1407 and its associated computer readable medium provide non-volatile storage for the computer device 1400. That is, the mass storage device 1407 may include a computer readable medium (not shown) such as a hard disk or a CD-ROM (Compact Disc Read-Only Memory) drive.

[0276] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer storage media include RAM, ROM, EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other solid-state storage technology, CD-ROM, DVD (Digital Video Disc) or other optical storage, cassettes, tapes, disk storage or other magnetic storage devices. Of course, those skilled in the art will know that the computer storage medium is not limited to the above. The above-mentioned system memory 1404 and mass storage device 1407 can be collectively referred to as memory.

[0277] According to the embodiment of the present application, the computer device 1400 can also be connected to a remote computer on the network through a network such as the Internet. That is, the computer device 1400 can be connected to the network 1412 through the network interface unit 1411 connected to the system bus 1405, or the network interface unit 1411 can be used to connect to other types of networks or remote computer systems (not shown).

[0278] The memory also includes a computer program, which is stored in the memory and configured to be executed by one or more processors to implement the above-mentioned prediction method of the recommendation index.

[0279] In some embodiments, a computer-readable storage medium is further provided, wherein a computer program is stored in the storage medium, and when the computer program is executed by a processor, the method for predicting the recommended indicator described above is implemented.

[0280] Optionally, the computer readable storage medium may include: ROM (Read-Only Memory), RAM (Random-Access Memory), SSD (Solid State Drives) or optical disks, etc. Among them, the random access memory may include ReRAM (Resistance Random Access Memory) and DRAM (Dynamic Random Access Memory).

[0281] In some embodiments, a computer program product is further provided, the computer program product comprising a computer program, the computer program being stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device performs the above-mentioned method for predicting the recommendation index.

[0282] It should be noted that, before collecting the relevant data of the user and during the process of collecting the relevant data of the user, the embodiment of the present application can display a prompt interface, a pop-up window or output a voice prompt information, and the prompt interface, pop-up window or voice prompt information is used to prompt the user that the relevant data is currently being collected, so that the present application only starts to execute the relevant steps of obtaining the relevant data of the user after obtaining the confirmation operation issued by the user to the prompt interface or pop-up window, otherwise (that is, when the confirmation operation issued by the user to the prompt interface or pop-up window is not obtained), the relevant steps of obtaining the relevant data of the user are terminated, that is, the relevant data of the user is not obtained. In other words, all user data collected by this application are processed strictly in accordance with the requirements of the laws and regulations of the relevant countries, and the informed consent or separate consent of the subject of personal information is obtained with the consent and authorization of the user. The subsequent data use and processing behavior is carried out within the scope of authorization of the laws and regulations and the subject of personal information, and the collection, use and processing of relevant user data need to comply with the relevant laws, regulations and standards of the relevant countries and regions. For example, the directional content, industry characteristics, sequence characteristics, training samples, input data, etc. involved in this application are all obtained with full authorization.

[0283] It should be understood that the "multiple" mentioned in this article refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. In addition, the step numbers described in this article only illustrate a possible execution sequence between the steps. In some other embodiments, the above steps may not be executed in the order of the numbers, such as two steps with different numbers are executed at the same time, or two steps with different numbers are executed in the opposite order to the diagram. The embodiments of the present application are not limited to this.

[0284] The above description is only an exemplary embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting a recommendation index, characterized in that: The method comprises: Acquire a plurality of training samples, wherein the training samples include a first sequence feature of a recommended object and a second sequence feature of targeted content, wherein the first sequence feature includes at least one industry feature related to the recommended object, and the second sequence feature includes at least one industry feature related to the targeted content; For each of the training samples, obtaining encoding information corresponding to the training sample according to the first sequence feature and the second sequence feature included in the training sample; Determining the prediction recommendation index corresponding to each of the training samples according to the coding information corresponding to each of the training samples through the index prediction model, wherein the prediction recommendation index refers to the predicted value of the recommendation index; According to the encoding information, the predicted recommendation index and the marked recommendation index corresponding to each of the training samples, the index prediction model is trained to obtain a trained index prediction model, wherein the trained index prediction model is used to obtain the predicted recommendation index of the recommendation object for the targeted content, and the marked recommendation index refers to the marked value of the recommendation index; The industry feature is used to characterize the directional object corresponding to the directional content and the identification information in the industry corresponding to the directional object.

2. The method according to claim 1, characterized in that The coding information corresponding to the training sample includes coding information of the first sequence feature and coding information of the second sequence feature; The indicator prediction model is trained according to the coding information, the prediction recommendation indicator and the label recommendation indicator respectively corresponding to each of the training samples to obtain the trained indicator prediction model, including: According to the encoding information of the first sequence features and the encoding information of the second sequence features respectively corresponding to each of the training samples, construct the positive sample encoding and the negative sample encoding corresponding to the training samples, wherein the positive sample encoding refers to the encoding information of the positive sample corresponding to the recommended object, and the negative sample encoding refers to the encoding information of the negative sample corresponding to the recommended object; According to the positive sample codes and the negative sample codes respectively corresponding to the training samples, a contrast loss of the indicator prediction model is obtained, wherein the contrast loss is used to enable the indicator prediction model to have the ability to learn the features corresponding to the negative sample codes; According to the prediction recommendation index and the label recommendation index corresponding to each of the training samples, a main task loss of the index prediction model is obtained, wherein the main task loss is used to indicate the prediction accuracy of the index prediction model for the recommendation index; The indicator prediction model is trained according to the contrast loss and the main task loss of the indicator prediction model to obtain the trained indicator prediction model.

3. The method according to claim 2, characterized in that The first sequence feature includes a first subsequence feature and a second subsequence feature, the industry feature in the first subsequence feature is in a binding relationship with the recommended object, the second subsequence feature includes the first subsequence feature, and the second subsequence feature includes an industry feature that is not in a binding relationship with the recommended object; the second sequence feature includes a third subsequence feature and a fourth subsequence feature, the industry feature in the third subsequence feature is in a binding relationship with the targeted content, and the fourth subsequence feature is used to characterize the consumption of the targeted content; The constructing the positive sample encoding and the negative sample encoding corresponding to the training samples according to the encoding information of the first sequence features and the encoding information of the second sequence features respectively corresponding to the training samples includes: For each of the training samples, construct the positive sample code according to the coding information of the first subsequence feature and the coding information of the second subsequence feature corresponding to the training sample; or construct the positive sample code according to the coding information of the first subsequence feature and the coding information of the third subsequence feature corresponding to the training sample; The coding information of the first subsequence feature and the coding information of the second subsequence feature corresponding to the training sample are cross-dispersed to construct the negative sample coding; or the coding information of the first subsequence feature and the coding information of the third subsequence feature corresponding to the training sample are cross-dispersed to construct the negative sample coding.

4. The method according to claim 2, characterized in that: The encoding information corresponding to the training sample includes first sub-encoding information and second sub-encoding information, the first sub-encoding information and the second sub-encoding information are both used to represent the training sample, and the first sub-encoding information and the second sub-encoding information share the same vocabulary; The step of obtaining the contrast loss of the indicator prediction model according to the positive sample codes and the negative sample codes respectively corresponding to the training samples includes: Obtaining the contrast loss according to the positive sample encoding and the negative sample encoding constructed by the first sub-encoding information or the second sub-encoding information respectively corresponding to each of the training samples; or, For each of the training samples, concatenate the positive sample code constructed by the first sub-coding information of the training sample with the positive sample code constructed by the second sub-coding information of the training sample to obtain the final positive sample code corresponding to the training sample, and concatenate the negative sample code constructed by the first sub-coding information of the training sample with the negative sample code constructed by the second sub-coding information of the training sample to obtain the final negative sample code corresponding to the training sample; obtain the contrast loss according to the final positive sample code and the final negative sample code corresponding to each of the training samples; or, A first sub-contrast loss is obtained according to the positive sample encoding and the negative sample encoding constructed by the first sub-encoding information corresponding to each of the training samples, and a second sub-contrast loss is obtained according to the positive sample encoding and the negative sample encoding constructed by the second sub-encoding information corresponding to each of the training samples; the first sub-contrast loss and the second sub-contrast loss are summed to obtain the contrast loss.

5. The method according to claim 2, characterized in that: The step of obtaining the contrast loss of the indicator prediction model according to the positive sample codes and the negative sample codes respectively corresponding to the training samples includes: For each of the training samples, encode the positive sample corresponding to the training sample, and multiply the encoded information of the third subsequence feature corresponding to the training sample to obtain a first intermediate parameter; Dividing the first intermediate parameter and the learnable temperature coefficient to obtain a second intermediate parameter; Performing an exponential operation using the second intermediate parameter as an exponent to obtain a third intermediate parameter; Obtaining a first sum value according to the third intermediate parameters respectively corresponding to the training samples; Encoding the negative sample corresponding to the training sample, and multiplying the encoded information of the third subsequence feature corresponding to the training sample to obtain a fourth intermediate parameter; Dividing the fourth intermediate parameter and the temperature coefficient to obtain a fifth intermediate parameter; Performing an exponential operation using the fifth intermediate parameter as an exponent to obtain a sixth intermediate parameter; Obtaining a second sum value according to the sixth intermediate parameters corresponding to the training samples; Summing the first sum value and the second sum value to obtain a third sum value; Dividing the first sum value and the third sum value to obtain a seventh intermediate parameter; According to the seventh intermediate parameter, the contrast loss of the indicator prediction model is obtained.

6. The method according to claim 2, characterized in that The step of training the indicator prediction model according to the contrast loss and the main task loss of the indicator prediction model to obtain the trained indicator prediction model includes: Summing the contrast loss of the indicator prediction model and the main task loss to obtain the total loss of the indicator prediction model; With the goal of minimizing the total loss, the parameters of the indicator prediction model are iteratively adjusted to obtain the trained indicator prediction model.

7. The method according to claim 1, characterized in that The identification information of the recommended object and the first sequence feature are stored in a database in the form of a key-value pair, and the identification information of the targeted content and the second sequence feature are stored in the database in the form of a key-value pair; The obtaining of multiple training samples comprises: Acquiring the first sequence feature from the database according to the identification information of the recommended object; acquiring the second sequence feature from the database according to the identification information of the targeted content; The first sequence feature and the second sequence feature are combined to obtain the training sample.

8. The method according to claim 1, characterized in that The method further comprises: For each of the training samples, obtaining a first industry feature in a first sequence feature included in the training sample; Moving the first industry feature to the end of the first sequence feature to obtain a negative sample of the recommended object corresponding to the training sample; According to the negative samples of the recommended objects respectively corresponding to the training samples, the negative sample codes respectively corresponding to the training samples are obtained.

9. A method for predicting a recommendation index, characterized in that: The method comprises: Acquire input data, the input data including a first sequence feature of a recommended object and a second sequence feature of targeted content, the first sequence feature including at least one industry feature related to the recommended object, and the second sequence feature including at least one industry feature related to the targeted content; Acquire encoding information corresponding to the input data according to a first sequence feature and a second sequence feature included in the input data; Determining, by an indicator prediction model, a predicted recommended indicator corresponding to the input data according to the coding information corresponding to the input data, wherein the predicted recommended indicator refers to a predicted value of a recommended indicator; The industry feature is used to characterize the directional object corresponding to the directional content and the identification information in the industry corresponding to the directional object.

10. The method according to claim 9, characterized in that The obtaining of input data comprises: Acquire at least one industry feature having a binding relationship with the recommended object to construct the first sequence feature; Acquire at least one industry feature having a binding relationship with the targeted content to construct the second sequence feature; The input data is obtained according to the first sequence feature and the second sequence feature.

11. The method according to claim 9, characterized in that The encoding information corresponding to the input data includes at least one encoding group, each of the encoding groups includes first sub-encoding information and second sub-encoding information, the first sub-encoding information and the second sub-encoding information are both used to represent the input data, and the first sub-encoding information and the second sub-encoding information share the same vocabulary; The step of determining the prediction recommendation index corresponding to the input data according to the encoding information corresponding to the input data by the index prediction model includes: For each of the coding groups, concatenating the first sub-coding information and the second sub-coding information in the coding group to obtain concatenated coding information corresponding to the coding group; The prediction recommendation index corresponding to the input data is determined by the index prediction model according to the splicing coding information respectively corresponding to the at least one coding group.

12. A device for predicting a recommendation index, characterized in that: The device comprises: A training sample acquisition module, used to acquire a plurality of training samples, wherein the training samples include a first sequence feature of a recommended object and a second sequence feature of a targeted content, wherein the first sequence feature includes at least one industry feature related to the recommended object, and the second sequence feature includes at least one industry feature related to the targeted content; A coding information acquisition module, used for acquiring, for each of the training samples, coding information corresponding to the training sample according to the first sequence feature and the second sequence feature included in the training sample; A recommendation index prediction module, used to determine the predicted recommendation index corresponding to each of the training samples according to the encoding information corresponding to each of the training samples through an index prediction model, wherein the predicted recommendation index refers to the predicted value of the recommendation index; The prediction model training module is used to train the indicator prediction model according to the coding information, prediction recommendation indicators and marked recommendation indicators corresponding to each of the training samples to obtain a trained indicator prediction model. The trained indicator prediction model is used to obtain the prediction recommendation indicator of the recommended object for the targeted content. The marked recommendation indicator refers to the label value of the recommendation indicator; wherein the industry characteristics are used to characterize the targeted object corresponding to the targeted content and the identification information in the industry corresponding to the targeted object.

13. A prediction device for recommendation index, characterized in that: The device comprises: An input data acquisition module, used to acquire input data, the input data including a first sequence feature of a recommended object and a second sequence feature of a targeted content, the first sequence feature including at least one industry feature related to the recommended object, and the second sequence feature including at least one industry feature related to the targeted content; A coding information acquisition module, used to acquire coding information corresponding to the input data according to the first sequence feature and the second sequence feature included in the input data; A recommendation index prediction module is used to determine the predicted recommendation index corresponding to the input data according to the coding information corresponding to the input data through an index prediction model, wherein the predicted recommendation index refers to the predicted value of the recommendation index; wherein the industry characteristics are used to characterize the targeted object corresponding to the targeted content and the identification information in the industry corresponding to the targeted object.

14. A computer device, characterized in that: The computer device includes a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the method for predicting a recommendation indicator as described in any one of claims 1 to 8, or to implement the method for predicting a recommendation indicator as described in any one of claims 9 to 11.

15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the method for predicting a recommendation indicator as described in any one of claims 1 to 8, or to implement the method for predicting a recommendation indicator as described in any one of claims 9 to 11.

16. A computer program product, characterized in that The computer program product includes a computer program, which is stored in a computer-readable storage medium. The processor reads and executes the computer program from the computer-readable storage medium to implement the method for predicting a recommendation indicator as described in any one of claims 1 to 8, or to implement the method for predicting a recommendation indicator as described in any one of claims 9 to 11.

Images