Content recommendation method and device, model training method and device, equipment and medium

Abstract

The invention discloses a content recommendation method and device, a model training method and device, equipment and a medium. The method comprises the steps that multiple pieces of first content click data output by at least two service types for a to-be-recommended user under a push service are acquired, and the first content click data comprise to-be-pushed content and an estimated click rate of the to-be-pushed content under the corresponding service type; the multiple pieces of first content click data are input into an order-preserving regression model used online for click rate correction, multiple pieces of second content click data output by the order-preserving regression model are obtained, and the second content click data comprise the to-be-pushed content and the corrected click rate of the to-be-pushed content; and determining target contents according to the corrected click rates of the plurality of to-be-pushed contents, and issuing the target contents to the to-be-recommended users. According to the method, the order-preserving regression model is adopted to correct the click rate, so that content recommendation is more accurate.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence technology, and in particular relates to a content recommendation, model training method, device, equipment and medium. Background Art

[0002] When social media pushes content that users may be interested in, they usually use a recommendation system to predict users' preferences for different content based on their personal characteristics, historical behaviors, social relationships, and other information, and display content to users according to certain strategies. A key issue in the recommendation system is how to sort multiple business contents, that is, the mixed sorting problem. There are multiple business pushes on business platforms (such as Weibo platforms), and different users have preferences for different business types. However, the push service cannot send too much content to disturb users. Therefore, how to select the content that the current user is most interested in from the push content of multiple services and maximize the content click efficiency on the basis of content quota issuance has become an urgent problem to be solved in the industry.

[0003] In the related art, a hybrid sorting model is trained based on the characteristics of users and content using the ranking learning method. The model can predict the user's click rate or preference for content as the content score; then, the content is sorted in descending order based on the content score to generate an initial mixed list; finally, the mixed list is adjusted based on some business rules or strategies to obtain the final mixed list. However, in the above scheme, the output result of the hybrid sorting model may be non-monotonic, that is, for the same user, two pieces of content with the same score sent by a certain business may have the opposite order of scores after being sorted by the ranking model, which will lead to unreasonable mixed results and inaccurate content recommendations. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide a content recommendation, model training method, device, electronic device and storage medium to solve the problem that the output results of the hybrid sorting model in the related art are non-monotonic and the content recommendation is not accurate enough.

[0005] To achieve the above objectives, the present application embodiment adopts the following technical solutions:

[0006] In a first aspect, an embodiment of the present application provides a content recommendation method, including: obtaining multiple first content click data output for at least two business types under a push business for a user to be recommended, the first content click data including the content to be pushed and the estimated click rate of the content to be pushed under the corresponding business type; inputting the multiple first content click data into an online rank-preserving regression model for click rate correction, and obtaining multiple second content click data output by the rank-preserving regression model, the second content click data including the content to be pushed and the corrected click rate of the content to be pushed, the corrected click rate maps the estimated click rates under different business types into a unified click rate range; determining the target content according to the corrected click rates of the multiple content to be pushed, and sending the target content to the user to be recommended.

[0007] In a second aspect, an embodiment of the present application provides a model training method, comprising: obtaining fifth content click data historically issued to sample users for at least two business types under a push business, the fifth content click data comprising sample content, an estimated click rate of the sample content under the corresponding business type, and a click result of whether the sample content is clicked by the sample user; performing offline training on a to-be-trained order-preserving regression model according to multiple pieces of the fifth content click data, and obtaining a trained order-preserving regression model for online use, the order-preserving regression model being used to correct the click rate of input content, so that the output corrected click rate maps the estimated click rates under different business types to a unified click rate range.

[0008] In a third aspect, an embodiment of the present application provides a content recommendation device, comprising: a first acquisition module, used to acquire multiple first content click data output for at least two business types under a push business for a user to be recommended, the first content click data including the content to be pushed and the estimated click rate of the content to be pushed under the corresponding business type; a correction module, used to input the multiple first content click data into an online rank-preserving regression model for click rate correction, and obtain multiple second content click data output by the rank-preserving regression model, the second content click data including the content to be pushed and the corrected click rate of the content to be pushed, the corrected click rate maps the estimated click rates under different business types into a unified click rate range; a sending module, used to determine the target content according to the corrected click rates of the multiple content to be pushed, and send the target content to the user to be recommended.

[0009] In a fourth aspect, an embodiment of the present application provides a model training device, comprising: a second acquisition module, used to obtain multiple fifth content click data historically issued to sample users for at least two business types under a push business, the fifth content click data including sample content, an estimated click rate of the sample content under the corresponding business type, and a click result of whether the sample content is clicked by the sample user; a training module, used to perform offline training on a to-be-trained order-preserving regression model based on the multiple fifth content click data, to obtain a trained order-preserving regression model for online use, the order-preserving regression model being used to perform click rate correction on input content, so that the output corrected click rate maps the estimated click rates under different business types to a unified click rate range.

[0010] In a fifth aspect, an embodiment of the present application provides an electronic device, comprising: a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the method described in the embodiment of the first aspect of the present application, or implements the steps of the method described in the embodiment of the second aspect of the present application.

[0011] In a sixth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the method described in the embodiment of the first aspect of the present application are implemented, or the steps of the method described in the embodiment of the second aspect of the present application are implemented.

[0012] At least one of the above technical solutions adopted in the embodiments of the present application can achieve the following beneficial effects:

[0013] When recommending content to a user to be recommended, the embodiment of the present application obtains multiple first content click data output for the user to be recommended by at least two business types under the push service, the first content click data includes the content to be pushed and the estimated click rate of the content to be pushed under the corresponding business type, the multiple first content click data are input into the online order-preserving regression model for click rate correction, and multiple second content click data output by the order-preserving regression model are obtained, the second content click data includes the content to be pushed and the corrected click rate of the content to be pushed, the corrected click rate maps the estimated click rate under different business types to a unified click rate range, determines the target content according to the corrected click rate of the multiple content to be pushed, and sends the target content to the user to be recommended. The embodiment of the present application uses the order-preserving regression model to correct the click rate of the estimated click rates of different dimensions of the multiple content to be pushed output by at least two business types under the push service, obtains the corrected click rate under the same dimension, and sends the target content determined according to the corrected click rate to the user to be recommended, so as to select the content that the current user is most interested in from the push content of multiple businesses, and maximize the content click efficiency on the basis of the quota of content issuance. And because the input and output of the rank-preserving regression model are monotonic regression, that is, the higher the estimated click-through rate of the input, the higher the corrected click-through rate of the output. Therefore, the corrected click-through rate output by the rank-preserving regression model can well reflect the actual preference of the recommended user for the content, making content recommendations more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0015] Figure 1 A schematic diagram of a Push service mixed arrangement recommendation system provided for an embodiment of the present application;

[0016] Figure 2 A processing flow chart of a multi-service mixed recommendation system under a Push service provided by an embodiment of the present application;

[0017] Figure 3 A flowchart of a content recommendation method provided for an embodiment of the present application;

[0018] Figure 4 A schematic diagram of a process of correcting click rate using an rank-preserving regression model provided for an embodiment of the present application;

[0019] Figure 5 A flowchart of a content recommendation method provided for another embodiment of the present application;

[0020] Figure 6A schematic diagram of an order-preserving regression function fitting provided for an embodiment of the present application;

[0021] Figure 7 A schematic diagram of a process for online training of an order-preserving regression model provided for an embodiment of the present application;

[0022] Figure 8 A schematic diagram of the overall flow of a content recommendation method provided for another embodiment of the present application;

[0023] Fig. 9 A flowchart of a content recommendation method provided for another embodiment of the present application;

[0024] Fig.10 A schematic diagram of the structure of a content recommendation device provided by an embodiment of the present application;

[0025] Fig.11 A schematic diagram of the structure of a model training device provided in one embodiment of the present application;

[0026] Fig.12 A schematic diagram of the structure of an electronic device provided for one embodiment of the present application. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0028] The terms "first", "second", etc. in this application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here. In addition, the "and / or" in this application represents at least one of the connected objects, and the character " / " generally represents that the front and back associated objects are in an "or" relationship. It should be noted that the data involved in this application are all obtained on the premise of obtaining user authorization.

[0029] Social media is an Internet-based application that allows users to create, share and communicate various content, such as text, pictures, videos, audio, etc. An important function of social media is to push content that users may be interested in to increase user participation and satisfaction. In order to achieve this function, social media usually needs to adopt a recommendation system, that is, based on the user's personal characteristics, historical behavior, social relationships and other information, predict the user's preference for different content, and display content to users according to certain strategies.

[0030] A key issue in the recommendation system is how to sort multiple business contents, that is, the mixed sorting problem. There are multiple business pushes in the business platform (such as the Weibo platform), for example: interest-based business push is to mine the content that users are interested in based on user interests; attention-based business push is to mine bloggers that users care about based on users' social relationships; hot-spot business push is to recommend hot spots to users. The business platform has a rich variety of business types, and different users have preferences for different business types. However, the push service cannot send too much content to disturb users. Therefore, how to select the content that the current user is most interested in from the push content of multiple businesses, and maximize the content click efficiency on the basis of content quota delivery, has become an urgent problem that the industry needs to solve.

[0031] In the related art, according to the characteristics of users and content, a hybrid sorting model such as a click-through rate (CTR) prediction model is trained using a sorting learning method. The model can predict the user's click rate or preference for the content as the score of the content; then, according to the content score, the content is sorted in descending order to generate an initial mixed list; finally, according to some business rules or strategies, the mixed list is adjusted to obtain the final mixed list. However, in the above scheme, the output result of the hybrid sorting model may be non-monotonic, that is, for the same user, two pieces of content with the same score sent by a certain business may have opposite scores after being sorted by the sorting model, which will lead to unreasonable mixed results and thus inaccurate content recommendations. To this end, the present application proposes a content recommendation, model training method, device, electronic device and storage medium to solve the problem that the output result of the hybrid sorting model in the related art is non-monotonic and the content recommendation is not accurate enough.

[0032] In order to clearly describe the content recommendation, model training method, device, electronic device and storage medium of this application, the following is combined with Figure 1 , Figure 2 The application scenarios of the technical solution of this application are described.

[0033] Figure 1 This is a schematic diagram of the Push service mixed recommendation system. Figure 1 As shown, taking the Weibo platform as an example, each business type under the Push business (such as interest, social, and hot business types) calculates the estimated click-through rate of each content (i.e., material) for the recommended user under the business type through the sorting model of each business, sorts according to the estimated click-through rate, and sends the top-ranked multiple contents to the mixed arrangement calculation module. The mixed arrangement calculation module uses a mixed sorting model to correct the estimated click-through rates of different dimensions of the content sent by each business to obtain the corrected click-through rate of the same dimension. The corrected click-through rate maps the estimated click-through rates under different business types to a unified click-through rate range, and performs mixed sorting of multiple business contents according to the corrected click-through rate to select the preferred content; finally, after the filtering logic of the master control such as sent and read, the target content determined according to the corrected click-through rate is sent to the recommended user through the sending channel.

[0034] Figure 2 This is a processing flow chart of the multi-service mixed recommendation system under Push service. Figure 2 As shown in the figure, different business types send estimated click-through rates of different dimensions, such as the estimated click-through rate calculated by the algorithm sorting service through the CTR estimation model, the real click-through rate of pre-delivery sent by the pre-delivery service after pre-delivery and elimination, and the business segmentation click-through rate sent by the real-time delivery service. The estimated click-through rates of different dimensions cannot be compared directly, and need to be corrected by the hybrid sorting model in the mixed sorting system. The corrected estimated click-through rate (referred to as the corrected click-through rate) can be sorted and compared in the same dimension, and the top-ranked material (i.e., content) will be sent to the user.

[0035] This application adopts the rank-preserving regression model as a hybrid sorting model to correct the click-through rate of the estimated click-through rate of different dimensions of the input. Since the input and output of the rank-preserving regression model are monotonic regression, that is, the higher the estimated click-through rate of the input, the higher the corrected click-through rate of the output, the corrected click-through rate output by the rank-preserving regression model can well reflect the actual preference of the user to be recommended for the content, making the content recommendation more accurate.

[0036] The technical solutions provided by various embodiments of the present application are described in detail below in conjunction with the accompanying drawings.

[0037] Figure 3 A flowchart of a content recommendation method provided by an embodiment of the present application is shown below. Figure 3 As shown, the content recommendation method of the embodiment of the present application may specifically include the following steps:

[0038] S301, obtaining a plurality of first content click data output for to-be-recommended users by at least two service types under a push service, wherein the first content click data includes the to-be-pushed content and an estimated click rate of the to-be-pushed content under the corresponding service type.

[0039] In an embodiment of the present application, the executor of the content recommendation method of the embodiment of the present application is a content recommendation device, which can be arranged in an electronic device. The electronic device can be a terminal device or a server. Among them, the terminal device can be a mobile phone, a tablet computer, a desktop computer, a portable notebook, a vehicle-mounted device, etc.; the server can be an independent server or a server cluster composed of multiple servers. The electronic device has at least one business platform, such as a Weibo platform. There are multiple businesses on the business platform, such as push business, sharing business, etc. The embodiment of the present application is based on the Push business to realize the online recommendation of multiple business content under the Push business, that is, the online reasoning process of the order-preserving regression model.

[0040] The Push service includes push services of multiple types. For example, interest-based service push is to mine content that users are interested in based on their interests; attention-based service push is to mine bloggers that users care about based on their social relationships; and hotspot service push is to recommend hotspots to users.

[0041] The user to be recommended is the user for whom the Push service content is currently recommended.

[0042] For each type of business type under the Push business (e.g., interest-based, social-based, and hot-spot-based business types), the sorting model corresponding to each business is used to calculate the estimated click-through rate for each content (i.e., material) for the user to be recommended under this business type, that is, the estimated probability that each content under this type of business will be clicked by the user to be recommended, and the contents under this type of business are sorted according to the estimated click-through rate, and the top N (i.e., TopN) contents (recorded as the content to be pushed) and the corresponding estimated click-through rates, i.e., the TopN first content click data (including the content to be pushed and the estimated click-through rate of the content to be pushed under this business type) are output to the content recommendation device for mixed sorting. For example, the interest-based business type outputs the top three first content click data with the highest estimated click-through rate to the content recommendation device, the social-based business type outputs the top two first content click data with the highest estimated click-through rate to the content recommendation device, and the hot-spot-based business type outputs the top one first content click data with the highest estimated click-through rate to the content recommendation device.

[0043] S302, input multiple first content click data into the online rank-preserving regression model for click rate correction, and obtain multiple second content click data output by the rank-preserving regression model, the second content click data includes the content to be pushed and the corrected click rate of the content to be pushed, and the corrected click rate maps the estimated click rate under different business types into a unified click rate range.

[0044] In an embodiment of the present application, since the dimensions of the estimated click-through rates output by each type of business type under the Push business are different and cannot be compared, it is necessary to correct the estimated click-through rate to obtain a corrected click-through rate. The corrected click-through rate maps the estimated click-through rates under different business types into a unified click-through rate range, so as to unify the estimated click-through rates of different dimensions into the corrected click-through rates of the same dimension, which is convenient for comparison and sorting.

[0045] Obtain an order-preserving regression model, which may be an untrained order-preserving regression model with randomly set model parameters, or an order-preserving regression model obtained through offline training (see the relevant description in subsequent embodiments for the offline training process), or an order-preserving regression model obtained by online training of the order-preserving regression model obtained through offline training during an online reasoning process (see the relevant description in subsequent embodiments for the online training process).

[0046] It should be noted here that after the training of the order-preserving regression model is completed, the trained order-preserving regression model can be saved to the Hadoop Distributed File System (HDFS) path, and the order-preserving regression model can be loaded at the same time. If the loading is successful, the order-preserving regression model is stored in the Remote Dictionary Server (Redis) database in the form of a key-value pair. The value (Value) in the key-value pair includes two parts. The first part Value1 is the name of the order-preserving regression model. After each training is completed, a name is set for the newly obtained order-preserving regression model. Value1 can also include the training time of the model. For example, during the online reasoning process, a new order-preserving regression model is trained every 1 hour, which can be recorded as "Value1: 3:00model1, Value1: 4:00model2". The second part Value2 is the model parameters of the order-preserving regression model. The order-preserving regression model includes the following three model parameters: boundaries, predictions and isotonic, which respectively represent the segmentation points of the input data, the segmentation values ​​of the output data and whether it is an increasing function. These three model parameters can be serialized into js object notation (javascript object notation, json) for storage. The key in the key-value pair is fixed, which makes it easy to find the corresponding value in the database through the key during online reasoning and build the corresponding order-preserving regression model. The value in the key-value pair does not overwrite the old value, which can avoid affecting the process of online reasoning that is currently using the old value. The key-value pair storage method is convenient for online training. On the one hand, it is convenient to use real-time data to train the latest order-preserving regression model, improve the real-time performance of model training, and thus improve the effect of online sorting. On the other hand, it can improve the reading efficiency of the order-preserving regression model and reduce storage space and computing overhead.

[0047] The estimated click rate of the content to be pushed is corrected using the obtained rank-preserving regression model to obtain the corrected click rate of the content to be pushed, and second content click data including the content to be pushed and the corrected click rate of the content to be pushed is output.

[0048] The rank-preserving regression model can be expressed as an rank-preserving regression function, which can be a piecewise constant or linear function as follows:

[0049]

[0050] Among them, x is the estimated click rate, f(x) is the corrected click rate, To estimate the click rate segment point increment sequence, is the corrected click rate segment point increasing sequence. The estimated click rate segment point increasing sequence and the corrected click rate segment point increasing sequence are the model parameters of the order-preserving regression model.

[0051] The isotonic regression model can fit a piecewise constant or linear function according to the order relationship of the input data, so that the output data also maintains the same order relationship, while minimizing the square error. That is, the input and output of the isotonic regression model are monotonic regression, that is, the higher the estimated click-through rate of the input, the higher the corrected click-through rate of the output. Therefore, the corrected click-through rate output by the isotonic regression model can well reflect the real preference of the recommended user for the content, avoiding the deviation between the corrected click-through rate output by the isotonic regression model and the actual click-through rate, improving the accuracy and stability of the mixed sorting, and making content recommendations more accurate.

[0052] In addition, the rank-preserving regression model can map the estimated click-through rates of multiple content in different dimensions into a unified click-through rate range, making the estimated click-through rates of different businesses comparable, avoiding the unfairness of the corrected click-through rate output by the rank-preserving regression model, and improving the fairness and diversity of mixed sorting.

[0053] It should be noted here that before inputting multiple pieces of first content click data into the rank-preserving regression model used online, these data need to be converted into a format suitable for the input of the rank-preserving regression model, such as the format of (label, rank_score, 1.0), where label indicates whether the content is clicked by the user, rank_score indicates the estimated click rate of the content, and 1.0 indicates the weight of the content.

[0054] S303: determining target content according to the corrected click rates of the plurality of content to be pushed, and sending the target content to the user to be recommended.

[0055] In an embodiment of the present application, multiple content to be pushed output by at least two business types are arranged in descending order according to the corrected click-through rate, and at least one content to be pushed with the highest order is determined as the target content, and the target content is sent to the user to be recommended. For example, if a quota is set to send one content each time, the content to be pushed with the highest order is sent to the user to be recommended. For example, if a corrected click-through rate threshold is set, the content to be pushed with a corrected click-through rate greater than the click-through rate threshold is sent to the user to be recommended.

[0056] Figure 4 The flowchart of the order-preserving regression model to correct the click rate is as follows: Figure 4As shown, a rank-preserving regression model is trained based on the historical content click data in the historical release opening log, and the estimated click rate (i.e., estimated CTR) of each content (i.e., material) output by each business type is input into the online rank-preserving regression model for click rate correction to obtain the corrected click rate (i.e., corrected CTR) of each content. The corrected CTRs are compared to select the optimal material.

[0057] Furthermore, since different users have different preferences for different types of business content, in order to achieve accurate content recommendation, when training the rank-preserving regression model, the corresponding rank-preserving regression model can be trained according to the different activity levels of the sample users. Activity is used to characterize whether the sample user is a monthly active user of the push service and / or the active frequency of the sample user on the business platform. Therefore, the corresponding rank-preserving regression model can be trained according to whether the sample user is a monthly active user of the push service (i.e., whether the user has used the push service in the last month) and / or according to the active frequency of the user to be recommended on the business platform, such as full frequency (the user comes every day in the last month), high frequency (the number of days the user has come in the last month exceeds 90%), medium frequency (the number of days the user has come in the last month exceeds 10% but does not exceed 90%), low frequency (the number of days the user has come in the last month does not exceed 10%), etc.

[0058] Correspondingly, in the online reasoning process, it is also necessary to use the rank-preserving regression model corresponding to the activity of the user to be recommended for click-through rate correction. Therefore, the "inputting multiple first content click data into the online rank-preserving regression model for click-through rate correction" in the above step S303 may specifically include the following steps: obtaining the user behavior characteristics of the user to be recommended; determining the activity of the user to be recommended based on the user behavior characteristics of the user to be recommended; obtaining the rank-preserving regression model corresponding to the activity based on the activity of the user to be recommended; the activity is used to characterize whether the user to be recommended is a monthly active user of the push service and / or the active frequency of the user to be recommended on the service platform; inputting multiple first content click data into the corresponding rank-preserving regression model for click-through rate correction.

[0059] Specifically, user behavior characteristics, such as user content clicking behavior, content sharing behavior, content liking behavior, content commenting behavior, content citing behavior, etc.

[0060] The user's behavioral characteristics can be assisted in determining through the third content click data in the click log sent by all services of the business platform (not limited to Push services) for the user to be recommended, specifically including the following steps: obtaining multiple third content click data historically sent by the business platform for the user to be recommended, the third content click data including historical recommended content and the click results of whether the historical recommended content was clicked by the user to be recommended; determining the user's behavioral characteristics of the user to be recommended based on the multiple third content click data.

[0061] Different rank-preserving regression models are trained based on whether there are monthly active users and / or active frequency. During online inference, the rank-preserving regression model corresponding to the activity of the recommended user is used to correct the click-through rate. This allows the rank-preserving regression model to adapt to the preferences and behaviors of different types of users, making the corrected click-through rate more accurate and improving the personalization and accuracy of the sorting.

[0062] Since the target content to be sent may contain content that has been sent to the user to be recommended or has been read by the user to be recommended, before step S304 "sending the target content to the user to be recommended", the following steps may be included: filtering the sent and / or read content in the target content. If there is still content to be pushed that has not been filtered out after filtering, the content to be pushed that has not been filtered out is sent to the user to be recommended.

[0063] The filtering process before content delivery avoids delivering the delivered and / or read content to the user again.

[0064] Furthermore, the content recommendation method of the embodiment of the present application may also include the following online training process of the rank-preserving regression model. Figure 5 As shown, after step S303, the content recommendation method of the embodiment of the present application may further include the following steps:

[0065] S501, obtaining fourth content click data historically issued to sample users of at least two business types under the push business, where the sample users at least include users to be recommended, and the fourth content click data includes sample content, an estimated click rate of the sample content under the corresponding business type, and a click result of whether the sample content is clicked by the sample user.

[0066] In the embodiment of the present application, the sample users are the users to be trained online for the model. When the offline trained rank-preserving regression model is trained separately according to whether the user is a monthly active user and / or the user's active frequency on the business platform, these sample users need to correspond to the same grouping features as the online inferenced users to be recommended: whether they are monthly active users and / or the active frequency on the business platform, that is, the models obtained from different offline training are trained online using the data of the corresponding users.

[0067] It should be noted here that, assuming that today is the 16th, the model obtained by offline training during the first inference is obtained by offline training using the fourth content click data of the last 15 days (i.e., the 1st to the 15th). When the model is trained online on the 16th (for example, once every hour), the fourth content click data of the 15 days during offline training (i.e., the 1st to the 15th) and before the current time on the 16th can be used for online training. When tomorrow is the 17th, when the model is trained online on the 17th (for example, once every hour), the fourth content click data of the 15 days during offline training (i.e., the 1st to the 15th) and before the current time on the 17th can be used for online training.

[0068] S502: Perform online training on the online rank-preserving regression model according to the fourth content click data to update the online rank-preserving regression model.

[0069] In an embodiment of the present application, various existing training methods can be used to perform online training on the rank-preserving regression model currently used online to achieve an update of the rank-preserving regression model used online. For example, the fourth content click data is input into the rank-preserving regression model currently used online for click-through rate correction. The rank-preserving regression model currently used online outputs the corrected click-through rate of the sample content. The model parameters of the rank-preserving regression model are adjusted based on the click result of whether the sample content is clicked by the sample user. The fourth content click data is again input into the rank-preserving regression model with adjusted parameters for click-through rate correction. The cycle is continued until the training end condition is met (for example, the number of training times reaches a threshold or the loss function reaches a minimum, etc.). The training is terminated to obtain an updated rank-preserving regression model. The latest rank-preserving regression model is used for click-through rate correction in subsequent online inference.

[0070] Specifically, rank-preserving regression is a regression algorithm, which is similar to conventional regression in form. Given a set of observations (i.e., the click results of whether the sample content is clicked by the sample user) y1, y2, ... and original values ​​(i.e., the estimated click rate of the sample content under the corresponding business type) x1, x2, ..., find a mapping function f(x) that minimizes the following loss function:

[0071]

[0072] Among them, x i ≤x j …≤x n , w1 and w2 are the weights corresponding to the model and are positive values, and the order of the output data is required to be consistent with the order of the input data, that is, x i ≤x j = >f(x i )≤f(x j ).

[0073] The Pool Adjacent Violators (PAV) algorithm is generally used to solve order-preserving regression. The algorithm starts with an unordered sequence of numbers, observes the disorder in the sequence and absorbs elements one by one to form a sequence until the average value of all elements is less than or equal to the next element to be absorbed.

[0074] Given an observation sequence y = (y1, y2, ..., y k ) and the original sequence x = (x1, x2, ..., x k ) under the condition of selecting the input data (i.e. the estimated click rate) segment point increasing sequence According to the above loss function training, the output data (i.e., corrected click rate) is obtained as a segmented point increasing sequence Then we get the following order-preserving regression function:

[0075]

[0076] It can be seen that the function is divided into n segments, each segment is a constant or linear function, and the corresponding function diagram is as follows Figure 6 As shown, the horizontal axis is the estimated click rate and the vertical axis is the corrected click rate.

[0077] When training the rank-preserving regression model, you only need to estimate a click-through rate for each content, and then fit a piecewise function based on the order of the estimated click-through rates. This reduces the time and space overhead of training. Especially in social media scenarios, due to the large amount and diversity of content, the rank-preserving regression model has a high training efficiency, making the entire training process simpler and faster. Online training of the rank-preserving regression model can better adapt to the dynamic changes of users and content, and improve the timeliness and flexibility of sorting.

[0078] Figure 7 This is a flow chart of online training of the rank-preserving regression model. Figure 7 As shown, the user behavior characteristics of the sample user are obtained according to the user characteristic table, and the user behavior characteristics of the sample user are assisted in determining according to the sixth content click data in the click log historically issued to the sample user by all businesses of the business platform, and training samples are generated according to the fourth content click data in the click log historically issued to the sample user by at least two business types under the Push business, and an rank-preserving regression model corresponding to the user behavior characteristics is trained, and the trained rank-preserving regression model is saved.

[0079] To clearly describe the process of the content recommendation method of the present application embodiment, the following is combined with Figure 8 The overall process of the content recommendation method of the embodiment of the present application (including online reasoning and online training process) is described in detail. Figure 8As shown, the business platform conducts content mining; each business type under the Push business conducts content recommendation: each business type uses its own sorting model to calculate the estimated click-through rate of the content to be pushed for the recommended user; the online rank-preserving regression model is used to perform mixed sorting on the content to be pushed: the online rank-preserving regression model is used to correct the estimated click-through rate, and multiple content to be pushed output by each business type is mixedly sorted according to the corrected click-through rate to select the optimal content; the optimal content is filtered as having been sent and / or read; the filtered optimal content is sent to the recommended user; according to the third content historically sent by the sample user and the corresponding estimated click-through rate and the click result of whether it is clicked by the sample user for at least two business types under the Push business, the rank-preserving regression model is trained online, and the previously stored online rank-preserving regression model is updated.

[0080] In summary, the content recommendation method of the embodiment of the present application, when recommending content to the user to be recommended, uses the rank-preserving regression model to correct the click-through rate of the estimated click-through rates of different dimensions of multiple contents to be pushed output by at least two business types under the push service, obtains the corrected click-through rate under the same dimension, and sends the target content determined according to the corrected click-through rate to the user to be recommended, so as to select the content that the current user is most interested in from the push content of multiple services, and maximize the content click efficiency on the basis of the content quota issuance. And because the input and output of the rank-preserving regression model are monotonic regression, that is, the higher the estimated click-through rate of the input, the higher the corrected click-through rate of the output, so the corrected click-through rate output by the rank-preserving regression model can well reflect the real preference of the user to be recommended for the content, making the content recommendation more accurate. According to whether there are monthly active users and / or active frequencies, different rank-preserving regression models are trained, and the rank-preserving regression model corresponding to the user to be recommended is used to correct the click-through rate during online reasoning, so that the rank-preserving regression model can adapt to the preferences and behaviors of different types of users, making the corrected click-through rate more accurate, and improving the personalization and accuracy of the sorting. Filtering before content delivery avoids delivering already delivered and / or read content to users. The third-party content click data of all services on the business platform (not limited to Push services) for the recommended user can assist in determining the user's behavior characteristics, making the obtained user behavior characteristics more accurate. Online training can better adapt to the dynamic changes of users and content, and improve the timeliness and flexibility of sorting.

[0081] Fig. 9 A flow chart of a model training method provided in one embodiment of the present application. Fig. 9 As shown, the training method of the model in the embodiment of the present application may specifically include the following steps:

[0082] S901, obtaining multiple fifth content click data historically sent to sample users by at least two service types under the push service, wherein the fifth content click data includes sample content, an estimated click rate of the sample content under the corresponding service type, and a click result of whether the sample content is clicked by the sample user.

[0083] In an embodiment of the present application, the executor of the content recommendation method of the embodiment of the present application is a model training device, which can be set in an electronic device. The electronic device can be a terminal device or a server. Among them, the terminal device can be a mobile phone, a tablet computer, a desktop computer, a portable notebook, a vehicle-mounted device, etc.; the server can be an independent server or a server cluster composed of multiple servers. The embodiment of the present application is based on the historical content click data of multiple business types under the Push business of the business platform to train the order-preserving regression model, that is, the offline training process of the order-preserving regression model.

[0084] Sample users, that is, users whose historical content data is needed for training the current rank-preserving regression model.

[0085] Referring to the description in the above embodiment, when the rank-preserving regression model is trained offline, the corresponding rank-preserving regression model can be trained according to whether the sample user is a monthly active user of the push service and / or according to the active frequency of the sample user on the service platform. Therefore, the sample users here can also be a group of users divided according to whether they are monthly active users and active frequency, which are used to train the rank-preserving regression model corresponding to the group of users.

[0086] For each type of business type under the Push business (such as interest-based, social, and hot business types), the sorting model corresponding to each business is used to calculate the estimated click rate of the content that has been historically sent by the sample user (recorded as sample content), and together with the click results of whether the sample content is actually clicked, it is output as the fifth content click data to the model training device.

[0087] S902, offline training is performed on the rank-preserving regression model to be trained according to multiple fifth content click data, and the trained rank-preserving regression model is obtained for online use. The rank-preserving regression model is used to correct the click rate of the input content, so that the output corrected click rate maps the estimated click rates under different business types to a unified click rate range.

[0088] In the embodiment of the present application, the offline training process of this step is the same as the above Figure 5 The online training process in the illustrated embodiment is similar and will not be described in detail here.

[0089] Before saving the trained rank-preserving regression model, the trained rank-preserving regression model can also be evaluated for its effect. Specifically, the click rate of some content is predicted, and the evaluation index is calculated based on the click results of whether these contents are actually clicked, so as to evaluate the effect of the model. The evaluation index may include but is not limited to the mean square error (MSE) index. The MSE index represents the difference between the predicted value and the true value. The smaller the difference, the better the effect of the model.

[0090] In summary, the training method of the model in the embodiment of the present application obtains multiple fifth content click data historically issued to sample users by at least two business types under the push service, and the fifth content click data includes sample content, the estimated click rate of the sample content under the corresponding business type, and the click result of whether it is clicked by the sample user. The order-preserving regression model to be trained is trained offline based on the multiple fifth content click data to obtain the trained order-preserving regression model for online use. The embodiment of the present application implements offline training of the order-preserving regression model based on multiple fifth content click data historically issued to sample users by at least two business types under the push service, which is used for subsequent online reasoning calls.

[0091] Fig.10 This is a schematic diagram of the structure of a content recommendation device provided by an embodiment of the present application. Fig.10 As shown, the content recommendation device 1000 of the embodiment of the present application may specifically include: a first acquisition module 1001, a correction module 1002 and a delivery module 1003. Among them:

[0092] The first acquisition module 1001 is used to acquire multiple pieces of first content click data output for the to-be-recommended user by at least two business types under the push business, wherein the first content click data includes the to-be-pushed content and the estimated click rate of the to-be-pushed content under the corresponding business type.

[0093] The correction module 1002 is used to input multiple first content click data into the online rank-preserving regression model for click rate correction, and obtain multiple second content click data output by the rank-preserving regression model. The second content click data includes the content to be pushed and the corrected click rate of the content to be pushed. The corrected click rate maps the estimated click rate under different business types into a unified click rate range.

[0094] The sending module 1003 is used to determine the target content according to the corrected click rates of the multiple pieces of content to be pushed, and send the target content to the user to be recommended.

[0095] In the embodiment of the present application, the specific process of each module and unit in the content recommendation device of the embodiment of the present application realizing its function can refer to the relevant description in the above-mentioned content recommendation method embodiment, which will not be repeated here.

[0096] In summary, when recommending content to a user to be recommended, the content recommendation device of the embodiment of the present application uses an rank-preserving regression model to correct the click-through rate of the estimated click-through rates of different dimensions of multiple content to be pushed output by at least two types of services under the push service, obtains the corrected click-through rate under the same dimension, and sends the target content determined according to the corrected click-through rate to the user to be recommended, thereby selecting the content that the current user is most interested in from the push content of multiple services, and maximizing the content click efficiency on the basis of the quota content delivery. And because the input and output of the rank-preserving regression model are monotonic regressions, that is, the higher the estimated click-through rate of the input, the higher the corrected click-through rate of the output, so the corrected click-through rate output by the rank-preserving regression model can well reflect the real preference of the user to be recommended for the content, making the content recommendation more accurate.

[0097] Fig.11 A schematic diagram of a model training device provided in one embodiment of the present application. Fig.11 As shown, the model training device 1000 of the embodiment of the present application may specifically include: a second acquisition module 1101 and a training module 1102. Among them:

[0098] The second acquisition module 1101 is used to obtain multiple fifth content click data historically sent to sample users by at least two business types under the push business, and the fifth content click data includes sample content, estimated click rate of the sample content under the corresponding business type, and click results of whether it is clicked by the sample user.

[0099] The training module 1102 is used to perform offline training on the rank-preserving regression model to be trained based on multiple fifth content click data to obtain the trained rank-preserving regression model for online use. The rank-preserving regression model is used to correct the click rate of the input content so that the output corrected click rate maps the estimated click rates under different business types to a unified click rate range.

[0100] In the embodiment of the present application, the specific process of each module and unit in the training device of the model of the embodiment of the present application to realize its function can be found in the relevant description in the embodiment of the training method of the above-mentioned model, which will not be repeated here.

[0101] In summary, the training device of the model in the embodiment of the present application obtains multiple fifth content click data issued historically for sample users for at least two business types under the push service, and the fifth content click data includes sample content, the estimated click rate of the sample content under the corresponding business type, and the click result of whether it is clicked. The order-preserving regression model to be trained is trained offline based on the multiple fifth content click data to obtain the trained order-preserving regression model for online use. The embodiment of the present application implements offline training of the order-preserving regression model based on multiple fifth content click data issued historically for sample users for at least two business types under the push service, which is used for subsequent online reasoning calls.

[0102] The present application also provides an electronic device. Fig.12 As shown, the electronic device 1200 includes: a processor 1201, a memory 1202, and a program or instruction stored in the memory 1202 and executable on the processor 1201. When the program or instruction is executed by the processor 1201, the steps of the content recommendation method as in any of the above embodiments are implemented, or the steps of the model training method as in any of the above embodiments are implemented.

[0103] The electronic device of the embodiment of the present application, when recommending content to the user to be recommended, uses the rank-preserving regression model to correct the click rate of the estimated click rates of different dimensions of the multiple contents to be pushed output by at least two business types under the push service, obtains the corrected click rate under the same dimension, and sends the target content determined according to the corrected click rate to the user to be recommended, so as to select the content that the current user is most interested in from the push content of multiple services, and maximize the content click efficiency on the basis of the quota distribution of content. And because the input and output of the rank-preserving regression model are monotonic regression, that is, the higher the estimated click rate of the input, the higher the corrected click rate of the output, so the corrected click rate output by the rank-preserving regression model can well reflect the real preference of the user to be recommended for the content, making the content recommendation more accurate. When training the rank-preserving regression model offline, the rank-preserving regression model is offline trained according to the multiple fifth content click data historically issued by the sample user of at least two business types under the push service, which is used for subsequent online reasoning.

[0104] An embodiment of the present application also proposes a readable storage medium, on which one or more computer programs are stored. The one or more computer programs include instructions. When the program or instructions are executed by a processor in an electronic device that includes multiple applications, the processor in the electronic device can execute each step of the above-mentioned content recommendation method embodiment or model training method embodiment.

[0105] The readable storage medium of the embodiment of the present application, when recommending content to the user to be recommended, uses the rank-preserving regression model to correct the click rate of the estimated click rates of different dimensions of the multiple contents to be pushed output by at least two business types under the push service, obtains the corrected click rate under the same dimension, and sends the target content determined according to the corrected click rate to the user to be recommended, so as to select the content that the current user is most interested in from the push content of multiple services, and maximize the content click efficiency on the basis of the quota distribution of content. And because the input and output of the rank-preserving regression model are monotonic regression, that is, the higher the estimated click rate of the input, the higher the corrected click rate of the output, so the corrected click rate output by the rank-preserving regression model can well reflect the real preference of the user to be recommended for the content, making the content recommendation more accurate. When training the rank-preserving regression model offline, the rank-preserving regression model is offline trained according to the multiple fifth content click data historically issued by the sample user of at least two business types under the push service, which is used for subsequent online reasoning.

[0106] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0107] For the convenience of description, the above device is described in various units according to their functions. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0108] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0109] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0110] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0111] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0112] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0113] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0114] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0115] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0116] The present application may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0117] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0118] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A content recommendation method, characterized in that: include: Acquire a plurality of first content click data output for to-be-recommended users of at least two service types under the push service, wherein the first content click data includes to-be-pushed content and an estimated click rate of the to-be-pushed content under the corresponding service type; Inputting the plurality of first content click data into an online rank-preserving regression model for click rate correction, obtaining a plurality of second content click data output by the rank-preserving regression model, wherein the second content click data includes the content to be pushed and the corrected click rate of the content to be pushed, and the corrected click rate maps the estimated click rates under different business types into a unified click rate range; The target content is determined according to the corrected click rates of the plurality of contents to be pushed, and the target content is sent to the user to be recommended.

2. The method according to claim 1, characterized in that The rank-preserving regression model is the following piecewise constant or linear function: Wherein, x is the estimated click rate, f(x) is the corrected click rate, To estimate the click rate segment point increment sequence, An increasing sequence of segment points for correcting click rate.

3. The method according to claim 1, characterized in that The step of inputting the plurality of first content click data into an online rank-preserving regression model for click rate correction includes: Acquire user behavior characteristics of the user to be recommended; Determining the activity of the user to be recommended according to the user behavior characteristics of the user to be recommended; According to the activity of the user to be recommended, obtaining an rank-preserving regression model corresponding to the activity; the activity is used to characterize whether the user to be recommended is a monthly active user of the push service and / or the activity frequency of the user to be recommended on the service platform; The plurality of first content click data are input into a corresponding rank-preserving regression model for click rate correction.

4. The method according to claim 1, characterized in that: Before sending the target content to the user to be recommended, the method further includes: Filter the delivered and / or read content in the target content.

5. The method according to claim 3, characterized in that: The obtaining of the user behavior characteristics of the user to be recommended includes: Acquire multiple pieces of third content click data historically issued by the business platform to the user to be recommended, wherein the third content click data includes historical recommended content and a click result indicating whether the historical recommended content is clicked by the user to be recommended; Determine the user behavior characteristics of the to-be-recommended user according to the plurality of pieces of third content click data.

6. The method according to claim 1, characterized in that After sending the target content to the user to be recommended, the method further includes: Acquire fourth content click data historically issued to sample users by the at least two service types under the push service, the sample users at least including the to-be-recommended users, the fourth content click data including sample content, an estimated click rate of the sample content under the corresponding service type, and a click result of whether the sample content is clicked by the sample user; The online rank-preserving regression model used online is trained online according to the fourth content click data to update the online rank-preserving regression model used online.

7. A model training method, characterized in that: include: Acquire multiple fifth content click data historically issued to sample users by at least two service types under the push service, wherein the fifth content click data includes sample content, an estimated click rate of the sample content under the corresponding service type, and a click result of whether the sample content is clicked by the sample user; The rank-preserving regression model to be trained is trained offline based on the multiple fifth content click data to obtain a trained rank-preserving regression model for online use. The rank-preserving regression model is used to correct the click rate of the input content so that the output corrected click rate maps the estimated click rates under different business types to a unified click rate range.

8. A content recommendation device, characterized in that: include: A first acquisition module is used to acquire a plurality of first content click data output for a to-be-recommended user in at least two service types under a push service, wherein the first content click data includes the to-be-recommended content and an estimated click rate of the to-be-recommended content under the corresponding service type; a correction module, configured to input the plurality of first content click data into an online rank-preserving regression model for click rate correction, and obtain a plurality of second content click data output by the rank-preserving regression model, wherein the second content click data includes the content to be recommended and a corrected click rate of the content to be recommended, wherein the corrected click rate maps the estimated click rates under different business types into a unified click rate range; The sending module is used to determine the target content according to the corrected click rates of the multiple pieces of content to be pushed, and send the target content to the user to be recommended.

9. An electronic device, characterized in that: The method comprises a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the method according to any one of claims 1 to 6, or implements the steps of the method according to claim 7.

10. A readable storage medium, characterized in that: The readable storage medium stores a program or instruction, and when the program or instruction is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented, or the steps of the method according to claim 7 are implemented.

Images