A method for determining a prediction model and related apparatuses

Abstract

The application discloses a prediction model determination method and related device, relates to the fields of artificial intelligence, machine learning and the like, and disperses the influence of scene preferences on prediction click probability into a click rate estimation main model and a bias step auxiliary model. In an actual prediction process, the click rate estimation main model carrying part of the scene preference information can be used for click probability calculation, and the prediction result of the bias capturing auxiliary model carrying another part of the scene preference information is not used. The obtained click probability retains the influence of part of the scene preference information, scene preference information correction is realized, better object-based content click probability prediction is realized, the influence of the scene preference information on the click probability is not ignored, a better matching user click intention prediction result can be obtained, and the accuracy of the click probability is improved.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a method and apparatus for determining a prediction model. Background Technology

[0002] With the rapid development of e-commerce and social media platforms based on internet technology, recommendation systems have become an indispensable tool for many enterprises. The core issue in object recommendation is how to extract content from a massive pool of potential recommendations that contains content that users are most likely to be interested in, thereby improving the accuracy of content recommendations. Potential recommendations can include articles, videos, and more.

[0003] Current recommendation systems can be based on user consumption behavior to determine the recommendation system and use it to recommend suitable objects to users, forming a closed loop of "system recommendation - user click behavior - backend data recording - system recommendation". In reality, different users have different preferences for the scene categories of recommended objects. For example, they have a greater interest in articles from the WeChat official account category and a less interest in videos from the video account category. This scene preference affects the object recommendation results to a certain extent, so that the scene category preferred by the user receives more recommendation traffic. In the aforementioned closed loop, this scene preference characteristic will be recorded and amplified, which can easily lead to the recommendation algorithm learning biased click prediction results, thereby reducing the impact of object content on the recommendation results. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a method and related apparatus for determining a prediction model. When making object recommendations, it retains some scene preference information, alleviates the problem of scene preference characteristics being accumulated and amplified, and improves the accuracy of object recommendations.

[0005] The embodiments of this application disclose the following technical solutions:

[0006] On the one hand, this application provides a method for determining a prediction model, the method comprising:

[0007] Acquire sample data and the labels of the sample data. The sample data includes the object characteristics of historical objects and the scene preference information of historical users. The labels of the sample data are used to indicate the click information of the historical users on the historical objects.

[0008] Using the sample data as input data, the first click probability of the historical user on the historical object is obtained through the click-through rate prediction main model. Using the scene preference information of the historical user as input data, the second click probability of the historical user on the historical object is obtained through the deviation capture auxiliary model.

[0009] Based on the first click probability and the second click probability, determine the predicted click probability of the historical user for the historical object;

[0010] A loss function is constructed based on the predicted click probability and the labels of the sample data;

[0011] The main model for click-through rate (CTR) prediction and the auxiliary model for bias capture are trained using the loss function, so that the main model for CTR prediction is transformed into a prediction model, which is used to calculate the click probability for target data.

[0012] On the other hand, this application provides a device for determining a prediction model, the device comprising:

[0013] A data acquisition unit is used to acquire sample data and labels for the sample data. The sample data includes object features of historical objects and scene preference information of historical users. The labels of the sample data are used to indicate the click information of the historical users on the historical objects.

[0014] The model calculation and control unit is used to take the sample data as input data, obtain the first click probability of the historical user on the historical object through the click rate prediction main model, and take the scene preference information of the historical user as input data to obtain the second click probability of the historical user on the historical object through the deviation capture auxiliary model.

[0015] A probability calculation unit is used to determine the predicted click probability of the historical user for the historical object based on the first click probability and the second click probability;

[0016] The loss function construction unit is used to construct a loss function based on the predicted click probability and the labels of the sample data;

[0017] The training unit is used to train the main click-through rate prediction model and the bias capture auxiliary model using the loss function, so that the main click-through rate prediction model is transformed into a prediction model, which is used to calculate the click probability for the target data.

[0018] Optionally, the deviation capture auxiliary model includes fully connected layers, and the total number of fully connected layers is less than the total number of layers in the click-through rate prediction main model.

[0019] Optionally, the deviation capture auxiliary model includes a fully connected layer.

[0020] Optionally, the scene preference information includes scene click information corresponding to multiple scene categories, the multiple scene categories include the scene category corresponding to the historical object, and the scene click information is determined based on the click information of historical users on objects corresponding to multiple scene categories within a first preset time period.

[0021] Optionally, the object features of the historical object include the scene category of the historical object, or the object features of the historical object include the scene category of the historical object and the content features of the historical object, wherein the scene category corresponding to the historical object is used to indicate the category of the content platform to which the historical object belongs;

[0022] The sample data also includes at least one of the following: the content preference information of the historical user, the comprehensive click information of the historical object, and the comprehensive click information of the content platform to which the historical object belongs; the comprehensive click information of the historical object is determined based on the click information of the historical object within a second preset time period, and the comprehensive click information of the content platform to which the historical object belongs is determined based on the click information of the object sent by the content platform to which the historical object belongs within a second preset time period.

[0023] Optionally, the sample data has uniform data bits;

[0024] The content features of the historical object are placed in the first data position corresponding to the content features in the unified data position. The scene category corresponding to the historical object is placed in the second data position corresponding to the scene category in the unified data position. The content preference information of the historical user is placed in the third data position corresponding to the content preference information in the unified data position. The comprehensive click information of the historical object is placed in the fourth data position corresponding to the object click information in the unified data position. The comprehensive click information of the content platform to which the historical object belongs is placed in the fifth data position, which is the sub-data position corresponding to the scene category corresponding to the historical object. The fifth data position is the data position corresponding to the platform click information in the unified data position. The scene preference information of the historical user is placed in the sixth data position corresponding to the scene preference information in the unified data position.

[0025] Optionally, the device further includes:

[0026] A prediction data acquisition unit is used to acquire multiple target data corresponding to multiple objects to be recommended;

[0027] A prediction probability calculation unit is used to obtain multiple click probabilities corresponding to multiple target data through the prediction model;

[0028] The display order determination unit is used to determine the display order of the objects to be recommended corresponding to the multiple target data based on the multiple click probabilities corresponding to the multiple target data.

[0029] Optionally, the device further includes:

[0030] The data acquisition unit is used to obtain updated sample data and updated sample data labels based on the target user's click behavior on the target object among the multiple objects to be recommended displayed.

[0031] The model update unit is used to update the prediction model based on the model structure, which includes the main click-through rate prediction model and the auxiliary deviation capture model, using the updated sample data and the labels of the updated sample data.

[0032] Optionally, the object characteristics of the historical object include the scene category corresponding to the historical object. The scene category corresponding to the historical object is used to indicate the category of the content platform to which the historical object belongs or the content performance category of the historical object. The category of the content platform to which the historical object belongs includes the public account category or the video account category. The content performance category of the historical object includes the article category or the video category.

[0033] On the other hand, this application provides a computer device, the device including a processor and a memory:

[0034] The memory is used to store computer programs and to transfer the computer programs to the processor;

[0035] The processor is configured to execute the method for determining the prediction model described above, according to instructions in the computer program.

[0036] On the other hand, embodiments of this application provide a computer-readable storage medium for storing a computer program for executing the method for determining the prediction model described above.

[0037] On the other hand, embodiments of this application provide a computer program product including a computer program, which, when run on a computer device, causes the computer device to perform the method for determining the prediction model.

[0038] As can be seen from the above technical solution, by acquiring sample data and its labels, the sample data includes the object characteristics of historical objects and the scene preference information of historical users. The labels of the sample data are used to indicate the click information of historical users on historical objects. The sample data containing scene preference information is used as the input data of the main click-through rate prediction model. The main click-through rate prediction model determines the first click probability based on the sample data. The scene preference information is used as the input data of the deviation capture auxiliary model. The deviation capture auxiliary model determines the second click probability based on the information source preference of historical users. The predicted click probability is obtained by using the first click probability and the second click probability. Since the scene preference information of historical users is simultaneously input into the main click-through rate prediction model and the deviation capture auxiliary model, the influence of scene preference on the predicted click probability is distributed among the main click-through rate prediction model and the deviation capture auxiliary model. This allows the main click-through rate prediction model to bear part of the influence of scene preference information, and the deviation capture auxiliary model to bear part of the influence of scene preference information.

[0039] Subsequently, a loss function was constructed using the predicted click probability and the labels of the sample data. Based on this loss function, the main click-rate prediction model and the bias capture auxiliary model were trained. After training, the main click-rate prediction model was transformed into a prediction model, which has the ability to compute scene preference information. Thus, in the actual prediction process, the click probability can be calculated using the main click-rate prediction model that carries some of the scene preference information, instead of using the prediction results of the bias capture auxiliary model that carries another part of the scene preference information. The resulting click probability retains the influence of some scene preference information, which reduces the accumulation of scene preference information compared to retaining all the preference information. This achieves the correction of scene preference information and better realizes click probability prediction based on object content. Compared to not retaining the influence of scene preference information, it does not ignore the influence of scene preference information on click probability, and can obtain prediction results that better match user click intentions, thus improving the accuracy of click probability. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A schematic diagram illustrating an application scenario of a method for determining a prediction model provided in an embodiment of this application;

[0042] Figure 2 A flowchart illustrating a method for determining a prediction model provided in an embodiment of this application;

[0043] Figure 3 A schematic diagram of sample data provided in an embodiment of this application;

[0044] Figure 4 A schematic diagram of a model structure provided in an embodiment of this application;

[0045] Figure 5 This is a schematic diagram of an actual prediction process provided in an embodiment of this application;

[0046] Figure 6 A schematic diagram of a model update provided in an embodiment of this application;

[0047] Figure 7 A structural block diagram of a predictive model determination device provided in an embodiment of this application;

[0048] Figure 8 A structural diagram of a terminal device provided in an embodiment of this application;

[0049] Figure 9 This is a structural diagram of a server provided in an embodiment of this application. Detailed Implementation

[0050] The embodiments of this application will now be described with reference to the accompanying drawings.

[0051] Currently, user scenario preferences influence the recommendation results to some extent. In recommendation systems based on consumption behavior, user scenario preference characteristics are recorded and amplified, which can easily lead to the recommendation algorithm learning biased click probability predictions, thereby reducing the impact of the content of the object on the recommendation results.

[0052] To address the aforementioned technical problems, embodiments of this application provide a method and related apparatus for determining a prediction model. This method distributes the influence of scene preferences on the predicted click probability across a main click-through rate (CTR) prediction model and a bias capture auxiliary model. The main CTR prediction model incorporates some scene preference information, as do the bias capture auxiliary model. In actual prediction, the main CTR prediction model, which incorporates some scene preference information, can be used to calculate the predicted click probability, instead of using the prediction result from the bias capture auxiliary model, which incorporates the remaining scene preference information. This results in a predicted click probability that retains some of the scene preference information's influence, better enabling content-based click probability prediction for objects. Furthermore, the obtained click probability better matches user click intentions and has higher accuracy.

[0053] The predictive model determination method provided in this application is based on Artificial Intelligence (AI). 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 that knowledge to obtain optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce new intelligent machines that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0054] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, as well as machine learning / deep learning, autonomous driving, and intelligent transportation.

[0055] In the embodiments of this application, the main artificial intelligence software technologies involved include the aforementioned machine learning / deep learning directions. For example, it may involve deep learning in machine learning (ML), including various artificial neural networks (ANNs).

[0056] The method for determining the predictive model provided in this application can be implemented using a computer device, which can be a terminal device or a server. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Terminal devices include, but are not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, and in-vehicle terminals. Terminal devices and servers can be directly or indirectly connected via wired or wireless communication, and this application does not impose any limitations on this connection.

[0057] This computer device, equipped with data processing capabilities, possesses machine learning capabilities. Machine learning is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and many other disciplines. It specifically studies how computers can 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 endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0058] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, smart customer service, vehicle networking, and intelligent transportation. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0059] It is understood that in the specific implementation of this application, data related to user scenario preference information, content preference information, etc. are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0060] To facilitate understanding of the technical solutions provided in this application, the following section will introduce a method for determining a prediction model provided in an embodiment of this application, in conjunction with a practical application scenario.

[0061] See Figure 1 , Figure 1 This is a schematic diagram illustrating an application scenario for a method for determining a prediction model provided in an embodiment of this application. Figure 1 The application scenario shown includes a server 10, which is used to determine a prediction model, and the prediction model is used to calculate the click probability.

[0062] Server 10 can obtain sample data and sample data tags. The sample data includes the object features of historical objects and the scene preference information of historical users. The object features of historical objects are used to describe at least one of the content and attributes of historical objects. The tags of the sample data are used to indicate the click information of historical users on historical objects.

[0063] After acquiring sample data and its labels, server 10 can use the sample data containing scene preference information as input data for the main click-through rate (CTR) prediction model. The main CTR prediction model determines the first click probability based on the sample data and uses the scene preference information as input data for the deviation capture auxiliary model. The deviation capture auxiliary model determines the second click probability based on historical users' information source preferences. The predicted click probability can then be obtained using the first and second click probabilities. Because historical users' scene preference information is simultaneously input into both the main CTR prediction model and the deviation capture auxiliary model, the influence of scene preferences on the predicted click probability is distributed across both models. This allows the main CTR prediction model to bear some of the influence of scene preference information, and the deviation capture auxiliary model to bear some of the influence of scene preference information.

[0064] Subsequently, server 10 can construct a loss function using the predicted click probability and the labels of the sample data, and train the model structure, including the main click rate prediction model and the bias capture auxiliary model, based on the loss function. After training, the main click rate prediction model is transformed into a prediction model, which has the ability to compute scene preference information. In the actual prediction process, the click probability can be calculated using the main click rate prediction model that carries part of the scene preference information, instead of using the prediction result of the bias capture auxiliary model that carries another part of the scene preference information. The resulting click probability retains the influence of some scene preference information, which reduces the accumulation of scene preference information compared to retaining all the preference information. This better realizes object click probability prediction based on object content. Compared to not retaining the influence of scene preference information, it does not ignore the influence of scene preference information on click probability, and can obtain prediction results that are more in line with user click intentions, thus improving the accuracy of click probability.

[0065] Next, with reference to the accompanying drawings, a method for determining a prediction model provided by an embodiment of this application will be described.

[0066] See Figure 2 , Figure 2 A flowchart illustrating a method for determining a prediction model provided in this application embodiment, the method comprising:

[0067] S101, Obtain sample data and the labels of the sample data.

[0068] In this embodiment, a predictive model can be used to calculate the click probability. Specifically, based on the information of the object to be recommended and the information of the target user, the click probability of the target user on the object to be recommended can be calculated. The higher the click probability, the greater the likelihood that the target user will click on the object to be recommended. Objects with a high click probability can then be recommended to the target user and may be placed at the top of the recommended object list. Objects with a low click probability may not be recommended to the target user or may be placed at the end of the recommended object list. The recommended object list can be displayed through subscription account message boxes, pushed message cards, etc.

[0069] The objects to be recommended can be articles or videos. Articles can contain at least one of the following: text, images, links, and videos. Videos can contain text and links. Each object to be recommended has a corresponding scene category. This scene category indicates the category of the content platform to which the object belongs or the content presentation category of the object. Specifically, the scene category is the category of the content platform to which the object belongs or the content presentation category of the object. For example, the content platform to which the object belongs might be a WeChat Official Account or a WeChat Video Account, and the corresponding category of the content platform might be a WeChat Official Account category or a WeChat Video Account category. The content presentation category might be an article category or a video category.

[0070] For example, WeChat Official Account A, a content platform belonging to the Official Account category, can publish an article named "aaa" as a candidate for recommendation. The content platform to which the candidate belongs is Official Account A, the category of the content platform to which the candidate belongs is the Official Account category, and the content presentation category of the candidate is the Article category. Therefore, the corresponding scenario category for the candidate is either the Official Account category or the Article category. Similarly, WeChat Video Account B, a content platform belonging to the Video Account category, can publish a video named "bbb" as a candidate for recommendation. The content platform to which the candidate belongs is Video Account B, the category of the content platform to which the candidate belongs is the Video Account category, and the content presentation category of the candidate is the Video category. Therefore, the corresponding scenario category for the candidate is either the Video Account category or the Video category.

[0071] Since different users may have different preferences for the same scenario category, and the same user may have different preferences for different scenario categories, the scenario category corresponding to the object to be recommended affects the calculation of the target user's click probability for the object to be recommended. Therefore, the prediction model used to predict the click probability of the object to be recommended can be trained based on the user's scenario preference information.

[0072] In this embodiment, sample data and its labels can be obtained. The sample data and labels can be used as training data to train a model and obtain a prediction model. The sample data can be used as model input, and the labels are used to construct a loss function. Specifically, the labels of the sample data are used to indicate historical user click information on historical objects. This click information includes whether the historical user clicked on the historical object or not. When a click occurred, the label of the sample data can be a first identifier; when no click occurred, the label of the sample data can be a second identifier. The first identifier can be, for example, "1", and the second identifier can be, for example, "0".

[0073] Specifically, the sample data includes the object characteristics of historical objects and the scene preference information of historical users. Historical objects can be articles or videos, etc. Articles can contain at least one of text, images, links, and videos, while videos can contain text, links, etc. The object characteristics of historical objects describe at least one of the content and attributes of the historical objects, and the scene preference information of historical users reflects the degree of preference of historical users for the scene category corresponding to the historical objects.

[0074] Historical objects can have corresponding scene categories. As an attribute of a historical object, its object characteristics can include the scene category it corresponds to. The scene category indicates the category of the content platform to which the historical object belongs or the content presentation category of the historical object. For example, the category of the content platform to which the historical object belongs is a WeChat Official Account or a WeChat Video Account. The content presentation category is, for example, an article or a video. Specifically, the scene category corresponding to a historical object is the category of the historical content platform to which it belongs or the content presentation category of the historical object. For example, WeChat Official Account C, as a content platform of the WeChat Official Account category, publishes an article named "ccc". This historical object belongs to WeChat Official Account C, the category of the content platform is WeChat Official Account, and the content presentation category is article. Therefore, the scene category corresponding to this historical object is either WeChat Official Account or article.

[0075] Historical user scenario preference information can reflect the degree of preference of historical users for the scenario categories corresponding to historical objects. For example, historical user scenario preference information includes the preference value of historical users for the scenario categories corresponding to historical objects, or historical user scenario preference information includes scenario click information corresponding to multiple scenario categories. Multiple scenario categories include the scenario categories corresponding to historical objects. The scenario click information is determined based on the click information of historical users on objects corresponding to multiple scenario categories within a first preset time period. In this way, the scenario click information corresponding to multiple scenario categories can reflect the scenario preferences of historical users.

[0076] Among these, at least one of the following click information—such as clicks, click-through rate (CTR), click days, and reading time—is considered. A higher number of clicks, a higher CTR, more click days, or a longer reading time indicates a greater historical user preference for that scenario category. The first preset time period can be the past 28 days, one week, one month, 90 days, etc. CTR can be reflected through clicks and impressions; specifically, CTR = clicks / impressions * 100%. CTR can also be reflected through click days and impression days; specifically, CTR = click days / impression days * 100%.

[0077] The degree of preference of historical users for a certain scenario category can be determined based on the click information of that scenario category and the click information of historical users for various scenario categories. For example, it can be the ratio of the number of clicks on that scenario category to the total number of clicks on all scenario categories by historical users. For instance, if the first preset time period is the past month, and historical users clicked 30 times on the WeChat Official Account category and 10 times on the WeChat Video Account category, then the degree of preference of historical users for the WeChat Official Account category can be considered as 30 / (30+10) = 75%, and the degree of preference of historical users for the WeChat Video Account category can be considered as 10 / (30+10) = 25%.

[0078] In addition, sample data can also include content preference information of historical users, which reflects the degree of preference of historical users for the content of historical objects. Content preference information can include interest tags that historical users are interested in, such as "sports" and "news," or content preference information can be represented by interest tags and preference levels. For example, the preference level corresponding to the interest tag "news" is 80%, and the preference level corresponding to the interest tag "sports" is 20%. The object features of historical objects can also include the content features of historical objects, which characterize the content of historical objects. That is, the object features of historical objects can include the scene category corresponding to the historical object, or the object features of historical objects can include the scene category of the historical object and the content features of the historical object, such as "sports" and "news." When the content preference information of historical users and the content features of historical objects match, historical users have a higher interest in historical objects, and historical users have a higher probability of clicking on historical objects.

[0079] The sample data may also include at least one of the following: comprehensive click information for historical objects and comprehensive click information for the content platforms to which the historical objects belong. The comprehensive click information for historical objects reflects their popularity and can, to some extent, their value. The comprehensive click information for the content platforms to which the historical objects belong reflects their popularity and can, to some extent, their popularity. The category of the content platform to which the historical object belongs is indicated by the corresponding scenario category. The comprehensive click information for historical objects is determined based on the click information of the historical objects within a second preset time period. The comprehensive click information for the content platforms to which the historical objects belong is determined based on the click information of objects posted by the content platforms to which the historical objects belong within the second preset time period.

[0080] The click information includes at least one of the following: number of clicks, click-through rate (CTR), number of click days, and reading time. The second preset time period can be the past 28 days, one week, one month, 90 days, etc. CTR can be reflected through the number of clicks and impressions; specifically, CTR = number of clicks / impressions * 100%. CTR can also be reflected through the number of click days and impression days; specifically, CTR = number of click days / impression days * 100%. Taking the article "ccc" published by public account C as an example, where the content platform is public account C and the second preset time period is the past month, the overall click information for article "ccc" is determined based on the click-through rate of article "ccc" within the past month, and the overall click information for public account C is determined based on the click-through rate of the content published by public account C within the past month.

[0081] In addition, the sample data may include at least one of the user identifier of historical users and the object identifier of historical objects. The user identifier of historical users is used to uniquely identify historical users, and the object identifier of historical objects is used to uniquely identify historical objects. For example, the user ID of historical users and the object ID of historical objects are examples of the object ID of historical objects.

[0082] In this embodiment of the application, the sample data may have a unified data bit. This unified data bit is applicable to multiple sample data. Each item of the sample data has a corresponding data bit, so that the content can be filled into the corresponding data bit in the unified data bit, which is beneficial for the unified processing of sample data corresponding to historical objects from different content platforms. When the object characteristics of a historical object include its content characteristics, the content characteristics can be placed in the first data position corresponding to the content characteristics in the unified data position. When the object characteristics of a historical object include its scene category, the scene category can be placed in the second data position corresponding to the scene category in the unified data position. The content preference information of historical users can be placed in the third data position corresponding to the content preference information in the unified data position. When the sample data includes the comprehensive click information of historical objects, the comprehensive click information of historical objects can be placed in the fourth data position corresponding to the object click information in the unified data position. When the sample data includes the comprehensive click information of the content platform to which the historical object belongs, the comprehensive click information of the content platform to which the historical object belongs can be placed in the fifth data position, which is the sub-data position corresponding to the scene category of the historical object. The fifth data position is the data position corresponding to the platform click information in the unified data position. The scene preference information of historical users can be placed in the sixth data position corresponding to the scene preference information in the unified data position. In addition, the user ID of historical users can be placed in the seventh data position corresponding to the user ID in the unified data position, and the object ID of historical objects can be placed in the eighth data position corresponding to the object ID in the unified data position.

[0083] When the historical user's scenario preference information includes scenario click information corresponding to multiple scenario categories, the scenario click information corresponding to each scenario category can be placed in the sixth data position along with the sub-data positions corresponding to each scenario category. Furthermore, when the sample data includes comprehensive click information from the content platform to which the historical object belongs, the fifth data position also includes sub-data positions corresponding to other scenario categories. These other scenario categories are scenario categories different from the scenario categories corresponding to the historical object. The sub-data positions corresponding to other scenario categories can have preset fill data, such as all zeros.

[0084] It should be noted that the order of the “first data bit”, “second data bit”, “third data bit”, “fourth data bit”, “fifth data bit” and “sixth data bit” can be arbitrary. For example, the third data bit can be before the first data bit or after the first data bit, and so on.

[0085] For example, the sample data includes the object characteristics of historical objects, the content preference information of historical users, the scene preference information of historical users, and the comprehensive click information of the content platform to which the historical objects belong. The object characteristics of historical objects include the scene category corresponding to the historical object. Therefore, the unified data bits of the sample data include the second data bit corresponding to the scene category, the third data bit corresponding to the content preference information, the fifth data bit corresponding to the platform click information, and the sixth data bit corresponding to the scene preference information. These are used to store relevant data for the scene category, content preference information, comprehensive click information of the platform to which the historical objects belong, and scene preference information, respectively. Taking the scene category as including WeChat Official Accounts and Video Accounts as an example, the fifth data bit includes the first sub-data bit corresponding to the WeChat Official Accounts category and the second sub-data bit corresponding to the Video Accounts category, and the sixth data bit includes the third sub-data bit corresponding to the WeChat Official Accounts category and the fourth sub-data bit corresponding to the Video Accounts category.

[0086] refer to Figure 3 The diagram illustrates sample data provided in an embodiment of this application. It sequentially includes a third data bit corresponding to content preference information, a fifth data bit corresponding to platform click information, a sixth data bit corresponding to scene preference information, and a second data bit corresponding to scene category. The fifth data bit contains a first sub-data bit corresponding to the WeChat Official Account category and a second sub-data bit corresponding to the WeChat Video Account category. The sixth data bit includes a third sub-data bit corresponding to the WeChat Official Account category and a fourth sub-data bit corresponding to the WeChat Video Account category. For example, if a historical user's content preference information is "sports," the data corresponding to "sports" can be placed in the third data bit. In the historical user's scene preference information, the number of clicks for the WeChat Official Account category, for example, "20," can be placed in the third sub-data bit, and the number of clicks for the WeChat Video Account category, for example, "5," can be placed in the fourth sub-data bit. The identifier data of the historical object's scene category, "WeChat Official Account category," can be placed in the second data bit. When the historical object's scene category is the WeChat Official Account category, the click-through rate of the content platform to which the historical object belongs (e.g., WeChat Official Account C), for example, 80%, can be placed in the first sub-data bit, and the second sub-data bit is set to zero. Of course, Figure 3 The content of each data bit is only an example; in reality, its form and content can be adjusted according to the actual situation.

[0087] S102, using sample data as input data, the first click probability of historical users on historical objects is obtained through the click-through rate prediction main model, and the second click probability of historical users on historical objects is obtained through the deviation capture auxiliary model, using the scenario preference information of historical users as input data.

[0088] When a recommendation system is established based on user consumption behavior, and suitable objects are recommended to users using this system, a closed loop can be formed: "system recommendation - user click behavior - backend data recording - system recommendation". If the recommendation system includes a single model that uses sample data and its labels as training data, then the scene preference information in this system will be recorded and amplified. This can easily lead to the recommendation algorithm learning biased click prediction results, thus reducing the impact of the content of the objects to be recommended. Specifically, user preferences (or aversions) to specific scene categories are recorded by data and further captured by the model, causing objects of that specific scene category to be assigned more (or less) recommendation traffic. Moreover, this bias will intensify over time. In the long run, when a better object appears in another scene, the model's predicted click-through rate score for that object will still be very low, thus losing the opportunity to recommend more diverse and higher-quality content to users in the entire system.

[0089] Based on this, a click-through rate (CTR) prediction master model can be used to process data other than scene preference information in the sample data to obtain click probability prediction results. An additional bias-capturing auxiliary model is set up to separately process scene preference information, outputting click probability prediction results based on this information. This ensures that the CTR prediction master model is unaffected by scene preference information, while only the bias-capturing auxiliary model is. The click probability prediction results from both models are then added to obtain the final prediction result. A loss function is constructed using the final prediction structure and the labels of the sample data, and both the CTR prediction master model and the bias-capturing auxiliary model are trained based on this loss function. In actual prediction, only the CTR prediction master model is used for click probability prediction, excluding the prediction results from the bias-capturing auxiliary model. This removes scene preference information, eliminating scene bias caused by its accumulation, and allows for better content-based click probability prediction. However, simply removing the prediction results from the bias-capturing auxiliary model, which carries scene preference information, and retaining only the prediction results from the CTR prediction master model, which is unaffected by scene preference information, leads to a lack of scene preference information, resulting in a failure to better match user preferences and reduced accuracy in click probability prediction.

[0090] In this embodiment of the application, a model structure including a main model for click-through rate prediction and an auxiliary model for deviation capture can be constructed, as shown in the following reference. Figure 4This is a schematic diagram of a model structure provided in an embodiment of this application. The main model for click-through rate prediction can be a machine learning model, specifically a neural network model, such as a Factorization Machine (FM), Field-aware Factorization Machines (FFM), Deep Neural Networks (DNN), and their evolved and improved models.

[0091] In this embodiment, sample data can be used as input data to obtain the first click probability of a historical user on a historical object through the click-through rate prediction main model. Historical user scenario preference information can be used as input data to obtain the second click probability of a historical user on a historical object through the bias capture auxiliary model. (Refer to...) Figure 4 Since the scenario preference information of historical users is simultaneously input into the main model of click-through rate prediction and the auxiliary model of deviation capture, the influence of scenario preference on the predicted click probability is distributed between the main model of click-through rate prediction and the auxiliary model of deviation capture. This allows the main model of click-through rate prediction to bear part of the influence of scenario preference information, and the auxiliary model of deviation capture to bear part of the influence of scenario preference information.

[0092] Using sample data as input to the main click-through rate (CTR) prediction model, the model processes the sample data to obtain the first click probability, denoted as pCTR. This pCTR represents the likelihood that a historical user will click on a historical object based on its content and contextual category. Specifically, using sample data as input, the main CTR prediction model embeds sparse features, including those of historical objects, to obtain first dense embeddings. These first dense embeddings are then concatenated with second dense features, including contextual preference information from historical users, to obtain complete dense features. Finally, the first click probability of a historical user on a historical object is calculated based on these complete dense features. Figure 4 Sparse features can include, for example, user ID, object ID, content features of historical objects, scene categories corresponding to historical objects, and content preference information of historical users. Second continuous features can also include comprehensive click information of historical objects, comprehensive click information of the content platform to which the historical objects belong, and scene preference information of historical users. Both the first and second continuous features are numerical features in the real number domain, with the first continuous feature being a low-dimensional feature. The main click-through rate prediction model can include multiple hidden layers. The input to multiple hidden layers can be complete continuous features, and the input to multiple hidden layers can be the first click probability. (Refer to...) Figure 4 .

[0093] Using historical users' scene preference information as input data, the bias capture auxiliary model can process this information to obtain the second click probability (biasCTR) of historical users on historical objects. This second click probability represents the likelihood that a historical user clicks on a historical object solely due to scene preference (regardless of the object's content). Specifically, the bias capture auxiliary model can include fully connected layers, with the total number of fully connected layers being less than the total number of layers in the main click-through rate prediction model. Because the number of fully connected layers is smaller, the weighting of the input scene preference information on the final prediction probability can be increased, thus biasing the influence of scene preference information more towards the bias capture auxiliary model and reducing the risk of overfitting the overall model to this feature. In practice, the number of layers in the bias capture auxiliary model can be less than or equal to one-third of the total number of layers in the main click-through rate prediction model to better distribute the influence of scene preference information. For example, the bias capture auxiliary model can include one fully connected layer, as shown in the reference... Figure 4 Furthermore, the influence weight of input scene preference information on the final prediction probability is increased, reducing the risk of overfitting the overall model to these features.

[0094] S103, Based on the first click probability and the second click probability, determine the predicted click probability of historical users for historical objects.

[0095] Since the first click probability is the likelihood that a historical user clicks on a historical object based on its content and context category, and the second click probability is the likelihood that a historical user clicks on a historical object solely due to context preference, the predicted click probability for a historical user on a historical object can be determined based on the first and second click probabilities. (Refer to...) Figure 4 The predicted click probability is denoted as CTR. The predicted click probability can be the sum of the first click probability and the second click probability, i.e., predicted click probability CTR = first click probability pCTR + second click probability biasCTR; or, a first weight can be determined for the first click probability, a second weight can be determined for the second click probability, the product of the first click probability and the first weight is taken as the first result, the product of the second click probability and the second weight is taken as the second result, and the sum of the first result and the second result is taken as the predicted click probability.

[0096] S104, construct the loss function based on the predicted click probability and the labels of the sample data.

[0097] Since the predicted click probability of historical users for historical objects is obtained by the model structure based on sample data, and the labels of the sample data indicate the click information of historical users for historical objects, the closer the predicted click probability matches the click information indicated by the labels of the sample data, the more accurate the click probability prediction of the model structure is. Therefore, a loss function can be constructed based on the predicted click probability of historical users for historical objects and the labels of the sample data. The loss function reflects the difference between the predicted click probability and the click information. Specifically, the cross-entropy of the predicted click probability of historical users for historical objects and the labels of the sample data can be used as the loss function.

[0098] S105 uses a loss function to train the main model for click-through rate (CTR) prediction and the auxiliary model for bias capture, transforming the main CTR prediction model into a prediction model. The prediction model is used to calculate the click probability for the target data.

[0099] Since the loss function reflects the difference between predicted click probability and click information, it can be used to train the main click-rate prediction model and the bias capture auxiliary model to minimize the loss function. This training process can be synchronous training based on the backpropagation algorithm. After training, it shows that the model structure has a relatively accurate click probability prediction capability. At this point, the main click-rate prediction model is transformed into a prediction model, which has the ability to calculate scene preference information. Thus, in the actual prediction process, the click probability can be calculated using the main click-rate prediction model that carries some scene preference information, instead of using the prediction results of the bias capture auxiliary model that carries another part of scene preference information. The resulting click probability retains some of the influence of scene preference information, better balancing the relationship between "removing scene bias" and "remembering user scene preferences." Compared to retaining all preference information, it reduces the accumulation of scene preference information and achieves correction of scene preference information, better realizing content-based click probability prediction. Compared to not retaining the influence of scene preference information, it does not ignore the impact of scene preference information on click probability, and can obtain prediction results that better match user click intentions, improving the accuracy of click probability.

[0100] Specifically, in the actual prediction process, target data can be input into the prediction model to obtain the click probability of the target user for the recommended object. The target data can include the scene category corresponding to the recommended object and the scene preference information of the target user. (Refer to...) Figure 5 This is a schematic diagram of an actual prediction process provided in an embodiment of this application. The scene preference information of the target user can be obtained in the actual prediction process, and therefore can be used as part of the target data.

[0101] In this embodiment, multiple target data points corresponding to multiple objects to be recommended can be obtained. Multiple click probabilities corresponding to each target data point are obtained through a prediction model. Based on these click probabilities, the display order of the objects to be recommended corresponding to the target data points is determined. The display order of the objects to be recommended allows users to prioritize the objects displayed earlier, as those displayed first have the highest probability of being clicked, thus better meeting the needs of the target users. The display of the objects to be recommended can be achieved through a terminal device connected to a server. The terminal device obtains the display order of the objects to be recommended and then displays them. The terminal device can also obtain the target user's click behavior on the displayed objects to be recommended.

[0102] In this embodiment, updated sample data and updated sample data labels are obtained based on the target user's click behavior on a target object among multiple displayed recommended objects. Using the updated sample data and updated sample data labels, and based on a model structure including a main model for click-through rate prediction and an auxiliary model for bias capture, the prediction model is updated. Thus, the prediction model is obtained by training the model using sample data, the click probability of the target user on the recommended object is obtained using the prediction model, the recommended object is displayed based on the target user's click probability on the recommended object, the click behavior of the target user on the displayed recommended object is obtained, the sample data is updated based on the click behavior, and then the prediction model is updated. Figure 6 This is a schematic diagram of a model update provided in an embodiment of this application, which enables the prediction system to form a closed loop of "system recommendation - user click behavior - background data recording - system recommendation". After the prediction model is updated, it can better match the needs of the target user.

[0103] In this embodiment, when the probability calculation method based on the prediction model is applied to object sorting in subscription account messages, better performance is achieved compared to the best existing model. After the model went live, in the experimental group A / control group B test with 1.4 million experimental users, the number of effective clicks for the video account category increased by 1.38%, and the overall information flow exposure click-through rate, including the video account category, increased by 0.49%.

[0104] Based on the method for determining a prediction model provided in the embodiments of this application, the embodiments of this application also provide an apparatus for determining a prediction model, referencing... Figure 7 The diagram shown is a structural block diagram of a predictive model determination device provided in an embodiment of this application. The predictive model determination device 1300 includes:

[0105] The data acquisition unit 1301 is used to acquire sample data and the labels of the sample data. The sample data includes the object characteristics of historical objects and the scene preference information of historical users. The labels of the sample data are used to indicate the click information of the historical users on the historical objects.

[0106] The model prediction unit 1302 is used to take the sample data as input data, obtain the first click probability of the historical user on the historical object through the click rate prediction main model, and take the scene preference information of the historical user as input data to obtain the second click probability of the historical user on the historical object through the deviation capture auxiliary model.

[0107] The probability calculation unit 1303 is used to determine the predicted click probability of the historical user for the historical object based on the first click probability and the second click probability;

[0108] The loss function construction unit 1304 is used to construct a loss function based on the predicted click probability and the labels of the sample data;

[0109] Training unit 1305 is used to train the main model for click-through rate prediction and the auxiliary model for bias capture using the loss function, so that the main model for click-through rate prediction is transformed into a prediction model, and the prediction model is used to calculate the click probability for target data.

[0110] Optionally, the model prediction unit 1304 includes:

[0111] The first model prediction subunit is used to take the sample data as input data, and through the click-through rate prediction main model, embed and map the sparse features of the object features including the historical object to obtain the first continuous feature; and concatenate the first continuous feature and the second continuous feature including the scene preference information of the historical user to obtain the complete continuous feature; and obtain the first click probability of the historical user on the historical object based on the complete continuous feature.

[0112] Optionally, the deviation capture auxiliary model includes fully connected layers, and the total number of fully connected layers is less than the total number of layers in the click-through rate prediction main model.

[0113] Optionally, the deviation capture auxiliary model includes a fully connected layer.

[0114] Optionally, the scene preference information includes scene click information corresponding to multiple scene categories, the multiple scene categories include the scene category corresponding to the historical object, and the scene click information is determined based on the click information of historical users on objects corresponding to multiple scene categories within a first preset time period.

[0115] Optionally, the object features of the historical object include the scene category of the historical object, or the object features of the historical object include the scene category of the historical object and the content features of the historical object, wherein the scene category corresponding to the historical object is used to indicate the category of the content platform to which the historical object belongs;

[0116] The sample data also includes at least one of the following: the content preference information of the historical user, the comprehensive click information of the historical object, and the comprehensive click information of the content platform to which the historical object belongs; the comprehensive click information of the historical object is determined based on the click information of the historical object within a second preset time period, and the comprehensive click information of the content platform to which the historical object belongs is determined based on the click information of the object sent by the content platform to which the historical object belongs within a second preset time period.

[0117] Optionally, the sample data has uniform data bits;

[0118] The content features of the historical object are placed in the first data position corresponding to the content features in the unified data position. The scene category corresponding to the historical object is placed in the second data position corresponding to the scene category in the unified data position. The content preference information of the historical user is placed in the third data position corresponding to the content preference information in the unified data position. The comprehensive click information of the historical object is placed in the fourth data position corresponding to the object click information in the unified data position. The comprehensive click information of the content platform to which the historical object belongs is placed in the fifth data position, which is the sub-data position corresponding to the scene category corresponding to the historical object. The fifth data position is the data position corresponding to the platform click information in the unified data position. The scene preference information of the historical user is placed in the sixth data position corresponding to the scene preference information in the unified data position.

[0119] Optionally, the device further includes:

[0120] A prediction data acquisition unit is used to acquire multiple target data corresponding to multiple objects to be recommended;

[0121] A prediction probability calculation unit is used to obtain multiple click probabilities corresponding to multiple target data through the prediction model;

[0122] The display order determination unit is used to determine the display order of the objects to be recommended corresponding to the multiple target data based on the multiple click probabilities corresponding to the multiple target data.

[0123] Optionally, the device further includes:

[0124] The data acquisition unit is used to obtain updated sample data and updated sample data labels based on the target user's click behavior on the target object among the multiple objects to be recommended displayed.

[0125] The model update unit is used to update the prediction model based on the model structure, which includes the main click-through rate prediction model and the auxiliary deviation capture model, using the updated sample data and the labels of the updated sample data.

[0126] Optionally, the object characteristics of the historical object include the scene category corresponding to the historical object. The scene category corresponding to the historical object is used to indicate the category of the content platform to which the historical object belongs or the content performance category of the historical object. The category of the content platform to which the historical object belongs includes the public account category or the video account category. The content performance category of the historical object includes the article category or the video category.

[0127] Therefore, by acquiring sample data and its labels, the sample data includes the object characteristics of historical objects and the scene preference information of historical users. The labels of the sample data are used to indicate the click information of historical users on historical objects. The sample data containing scene preference information is used as the input data of the main click-through rate (CTR) prediction model. The main CTR prediction model determines the first click probability based on the sample data. The scene preference information is used as the input data of the auxiliary model for bias capture. The auxiliary model for bias capture determines the second click probability based on the information source preferences of historical users. The predicted click probability is obtained by using the first click probability and the second click probability. Since the scene preference information of historical users is simultaneously input into the main CTR prediction model and the auxiliary model for bias capture, the influence of scene preference on the predicted click probability is distributed among the main CTR prediction model and the auxiliary model for bias capture. This allows the main CTR prediction model to bear part of the influence of scene preference information, and the auxiliary model for bias capture to bear part of the influence of scene preference information.

[0128] Subsequently, a loss function was constructed using the predicted click probability and the labels of the sample data. Based on this loss function, the main click-rate prediction model and the bias capture auxiliary model were trained. After training, the main click-rate prediction model was transformed into a prediction model, which has the ability to compute scene preference information. Thus, in the actual prediction process, the click probability can be calculated using the main click-rate prediction model that carries some of the scene preference information, instead of using the prediction results of the bias capture auxiliary model that carries another part of the scene preference information. The resulting click probability retains the influence of some scene preference information, which reduces the accumulation of scene preference information compared to retaining all the preference information. This achieves the correction of scene preference information and better realizes click probability prediction based on object content. Compared to not retaining the influence of scene preference information, it does not ignore the influence of scene preference information on click probability, and can obtain prediction results that better match user click intentions, thus improving the accuracy of click probability.

[0129] This application also provides a computer device, which is the computer device described above, and may include a terminal device or a server. The aforementioned predictive model determination device may be configured in the computer device. The computer device will now be described in conjunction with the accompanying drawings.

[0130] If the computer device is a terminal device, please refer to Figure 8 As shown, this application provides a terminal device, taking a mobile phone as an example:

[0131] Figure 8 This diagram illustrates a partial structural representation of a mobile phone related to the terminal device provided in this embodiment. (Reference) Figure 8 The mobile phone includes components such as a radio frequency (RF) circuit 1410, a memory 1420, an input unit 1430, a display unit 1440, a sensor 1450, an audio circuit 1460, a Wi-Fi module 1470, a processor 1480, and a power supply 1490. Those skilled in the art will understand that... Figure 8 The mobile phone structure shown does not constitute a limitation on the mobile phone and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0132] The following is combined with Figure 8 A detailed introduction to each component of a mobile phone:

[0133] The RF circuit 1410 can be used to receive and transmit signals during information transmission or calls. In particular, it receives downlink information from the base station and processes it with the processor 1480; in addition, it transmits uplink data to the base station.

[0134] The memory 1420 can be used to store software programs and modules. The processor 1480 executes various mobile phone functions and data processing by running the software programs and modules stored in the memory 1420. The memory 1420 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 1420 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0135] The input unit 1430 can be used to receive input numeric or character information, and to generate key signal inputs related to user settings and function control of the mobile phone. Specifically, the input unit 1430 may include a touch panel 1431 and other input devices 1432.

[0136] The display unit 1440 can be used to display information input by the user or information provided to the user, as well as various menus of the mobile phone. The display unit 1440 may include a display panel 1441.

[0137] The mobile phone may also include at least one sensor 1450, such as a light sensor, a motion sensor, and other sensors.

[0138] Audio circuitry 1460, speaker 1461, and microphone 1462 provide an audio interface between the user and the mobile phone.

[0139] WiFi is a short-range wireless transmission technology. Through the WiFi module 1470, mobile phones can help users send and receive emails, browse web pages, and access streaming media, providing users with wireless broadband internet access.

[0140] The processor 1480 is the control center of the mobile phone. It connects to various parts of the mobile phone through various interfaces and lines. It performs various functions of the mobile phone and processes data by running or executing software programs and / or modules stored in the memory 1420 and calling data stored in the memory 1420.

[0141] The phone also includes a power supply 1490 (such as a battery) that powers the various components.

[0142] In this embodiment, the processor 1480 included in the terminal device also has the following functions:

[0143] Acquire sample data and the labels of the sample data. The sample data includes the object characteristics of historical objects and the scene preference information of historical users. The labels of the sample data are used to indicate the click information of the historical users on the historical objects.

[0144] Using the sample data as input data, the first click probability of the historical user on the historical object is obtained through the click-through rate prediction main model. Using the scene preference information of the historical user as input data, the second click probability of the historical user on the historical object is obtained through the deviation capture auxiliary model.

[0145] Based on the first click probability and the second click probability, determine the predicted click probability of the historical user for the historical object;

[0146] A loss function is constructed based on the predicted click probability and the labels of the sample data;

[0147] The main model for click-through rate (CTR) prediction and the auxiliary model for bias capture are trained using the loss function, so that the main model for CTR prediction is transformed into a prediction model, which is used to calculate the click probability for target data.

[0148] If the computer device is a server, this application embodiment also provides a server; please refer to [link to relevant documentation]. Figure 9 As shown, Figure 9The structural diagram of the server provided in this application embodiment shows that the server 1500 can vary significantly due to different configurations or performance. It may include one or more processors 1522, such as a central processing unit (CPU), memory 1532, and one or more storage media 1530 (e.g., one or more mass storage devices) for storing application programs 1542 or data 1544. The memory 1532 and storage media 1530 can be temporary or persistent storage. The program stored in the storage media 1530 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the server. Furthermore, the processor 1522 may be configured to communicate with the storage media 1530 and execute the series of instruction operations in the storage media 1530 on the server 1500.

[0149] Server 1500 may also include one or more power supplies 1526, one or more wired or wireless network interfaces 1550, one or more input / output interfaces 1558, and / or one or more operating systems 1541, such as Windows Server. TM Mac OS X TM Unix TM Linux TM FreeBSD TM etc.

[0150] The steps performed by the server in the above embodiments can be based on Figure 9 The server structure shown.

[0151] In addition, this application embodiment also provides a storage medium for storing a computer program for executing the method provided in the above embodiment.

[0152] This application also provides a computer program product including instructions that, when run on a computer, cause the computer to perform the methods provided in the above embodiments.

[0153] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium can be at least one of the following media: read-only memory (ROM), RAM, magnetic disk or optical disk, and other media that can store program code.

[0154] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0155] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Moreover, based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for determining a prediction model, characterized in that, The method includes: Acquire sample data and the labels of the sample data. The sample data includes the object characteristics of historical objects and the scene preference information of historical users. The labels of the sample data are used to indicate the click information of the historical users on the historical objects. Using the sample data as input data, the first click probability of the historical user on the historical object is obtained through the click-through rate prediction main model. Using the scene preference information of the historical user as input data, the second click probability of the historical user on the historical object is obtained through the deviation capture auxiliary model. Based on the first click probability and the second click probability, determine the predicted click probability of the historical user for the historical object; A loss function is constructed based on the predicted click probability and the labels of the sample data; The main model for click-through rate (CTR) prediction and the auxiliary model for bias capture are trained using the loss function, so that the main model for CTR prediction is transformed into a prediction model, which is used to calculate the click probability for target data.

2. The method according to claim 1, characterized in that, The step of using the sample data as input data and obtaining the probability of the historical user's first click on the historical object through the click-through rate prediction main model includes: Using the sample data as input data, the sparse features of the object features, including the historical objects, are embedded and mapped through the main click-through rate prediction model to obtain the first continuous features. By using the main click-through rate prediction model, the first continuous feature and the second continuous feature, which includes the scene preference information of the historical users, are concatenated to obtain a complete continuous feature. The first click probability of a historical user on a historical object is obtained using the main click rate prediction model based on the complete and continuous features.

3. The method according to claim 1, characterized in that, The deviation capture auxiliary model includes fully connected layers, and the total number of fully connected layers is less than the total number of layers in the click-through rate prediction main model.

4. The method according to claim 3, characterized in that, The deviation capture auxiliary model includes a fully connected layer.

5. The method according to any one of claims 1-4, characterized in that, The scene preference information includes scene click information corresponding to multiple scene categories. The multiple scene categories include the scene categories corresponding to the historical objects. The scene click information is determined based on the click information of historical users on objects corresponding to multiple scene categories within a first preset time period.

6. The method according to any one of claims 1-4, characterized in that, The object features of the historical object include the scene category of the historical object, or the object features of the historical object include the scene category of the historical object and the content features of the historical object, wherein the scene category corresponding to the historical object is used to indicate the category of the content platform to which the historical object belongs; The sample data also includes at least one of the following: the content preference information of the historical user, the comprehensive click information of the historical object, and the comprehensive click information of the content platform to which the historical object belongs; the comprehensive click information of the historical object is determined based on the click information of the historical object within a second preset time period, and the comprehensive click information of the content platform to which the historical object belongs is determined based on the click information of the object sent by the content platform to which the historical object belongs within a second preset time period.

7. The method according to claim 6, characterized in that, The sample data has a uniform data bit; The content features of the historical object are placed in the first data position corresponding to the content features in the unified data position. The scene category corresponding to the historical object is placed in the second data position corresponding to the scene category in the unified data position. The content preference information of the historical user is placed in the third data position corresponding to the content preference information in the unified data position. The comprehensive click information of the historical object is placed in the fourth data position corresponding to the object click information in the unified data position. The comprehensive click information of the content platform to which the historical object belongs is placed in the fifth data position, which is the sub-data position corresponding to the scene category corresponding to the historical object. The fifth data position is the data position corresponding to the platform click information in the unified data position. The scene preference information of the historical user is placed in the sixth data position corresponding to the scene preference information in the unified data position.

8. The method according to any one of claims 1-4, characterized in that, The method further includes: Obtain multiple target data corresponding to multiple objects to be recommended; The prediction model is used to obtain multiple click probabilities corresponding to multiple target data. Based on multiple click probabilities corresponding to multiple target data, the display order of the recommended objects corresponding to multiple target data is determined.

9. The method according to claim 8, characterized in that, The method further includes: Based on the target user's click behavior on the target object among the multiple objects to be recommended, updated sample data and updated sample data labels are obtained; The prediction model is updated using the updated sample data and the updated sample data labels, based on the model structure including the main click-through rate prediction model and the deviation capture auxiliary model.

10. The method according to any one of claims 1-4, characterized in that, The object characteristics of the historical object include the scene category corresponding to the historical object. The scene category corresponding to the historical object is used to indicate the category of the content platform to which the historical object belongs or the content performance category of the historical object. The category of the content platform to which the historical object belongs includes the public account category or the video account category. The content performance category of the historical object includes the article category or the video category.

11. A device for determining a prediction model, characterized in that, The device includes: A data acquisition unit is used to acquire sample data and labels for the sample data. The sample data includes object features of historical objects and scene preference information of historical users. The labels of the sample data are used to indicate the click information of the historical users on the historical objects. The model prediction unit is used to take the sample data as input data, obtain the first click probability of the historical user on the historical object through the click rate prediction main model, and take the scene preference information of the historical user as input data to obtain the second click probability of the historical user on the historical object through the deviation capture auxiliary model. A probability calculation unit is used to determine the predicted click probability of the historical user for the historical object based on the first click probability and the second click probability; The loss function construction unit is used to construct a loss function based on the predicted click probability and the labels of the sample data; The training unit is used to train the main click-through rate prediction model and the bias capture auxiliary model using the loss function, so that the main click-through rate prediction model is transformed into a prediction model, which is used to calculate the click probability for the target data.

12. The apparatus according to claim 11, characterized in that, The model prediction unit includes: The first model prediction subunit is used to take the sample data as input data, and through the click-through rate prediction main model, embed and map the sparse features of the object features including the historical object to obtain the first continuous feature; and concatenate the first continuous feature and the second continuous feature including the scene preference information of the historical user to obtain the complete continuous feature; and obtain the first click probability of the historical user on the historical object based on the complete continuous feature.

13. A computer device, characterized in that, The computer device includes a processor and memory: The memory is used to store computer programs and to transfer the computer programs to the processor; The processor is configured to execute the method for determining the prediction model according to any one of claims 1-10, based on instructions in the computer program.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method for determining the prediction model according to any one of claims 1-10.

15. A computer program product comprising a computer program, characterized in that, When it is run on a computer device, it causes the computer device to perform the method for determining the prediction model as described in any one of claims 1-10.

Images