Feedback indicator prediction model training method, object recommendation method and related equipment

Abstract

The present disclosure is about a feedback index prediction model training method, an object recommendation method and related equipment, wherein the training method includes: obtaining random delivery information of sample objects obtained based on random delivery experiments; the random delivery information represents the feedback of randomly delivered user accounts on sample objects; according to the random delivery information corresponding to the sample objects, determining the target feedback index corresponding to the sample objects; according to the first historical feedback information sequence of the sample objects in historical recommendations, determining the historical feedback features corresponding to the sample objects; training a preset machine learning model according to the historical feedback features corresponding to the sample objects and the target feedback index to obtain a feedback index prediction model; wherein the target feedback index is used as the expected prediction value corresponding to the historical feedback features. The present disclosure improves the accuracy of the feedback index prediction model for item value prediction, thereby facilitating the matching of the display opportunities of each item with its true value.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a feedback indicator prediction model training method, an object recommendation method, and related equipment. Background Art

[0002] In recent years, there has been an increasing number of studies on the fairness of recommendation systems. In existing schemes that use items (such as videos to be recommended) as a fairness perspective, it is believed that the exposure of each item is positively correlated with its utility. The widely used utility measurement is based on direct statistics of feedback such as click-through rate and like rate in the recommendation system log.

[0003] However, the above-mentioned utility in related technologies ignores the impact of exposure bias in the recommendation system during the measurement process, resulting in poor accuracy of the utility indicators used in existing fairness measurement technologies, which cannot reflect the true value of the items, and further leads to a mismatch between the display opportunities of each item and its true value. Summary of the invention

[0004] The present disclosure provides a feedback indicator prediction model training method, an object recommendation method and related devices to at least solve the problem of poor prediction accuracy of the machine learning model trained in the related art for predicting the value of items. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a feedback indicator prediction model training method is provided, comprising:

[0006] Obtaining random delivery information of a sample object obtained based on a random delivery experiment; the random delivery information represents feedback from a randomly delivered user account on the sample object;

[0007] Determine the target feedback index corresponding to the sample object according to the random delivery information corresponding to the sample object;

[0008] Determining a historical feedback feature corresponding to the sample object according to a first historical feedback information sequence of the sample object in the historical recommendation;

[0009] A preset machine learning model is trained according to the historical feedback features corresponding to the sample object and the target feedback index to obtain a feedback index prediction model; wherein the target feedback index serves as the expected prediction value corresponding to the historical feedback features.

[0010] In an exemplary embodiment, determining the historical feedback feature corresponding to the sample object according to the first historical feedback information sequence of the sample object in the historical recommendation includes:

[0011] Acquire a first historical feedback information sequence of the sample object in historical recommendations;

[0012] Dividing the first historical feedback information sequence into a first number of subsequences according to a preset aggregation number; each of the subsequences includes consecutive first historical feedback information of the preset aggregation number;

[0013] Aggregating the first historical feedback information in each subsequence to obtain an aggregate feedback information sequence corresponding to the sample object; the aggregate feedback information sequence includes aggregate feedback information corresponding to each of the subsequences;

[0014] According to the aggregated feedback information sequence, a historical feedback feature corresponding to the sample object is determined.

[0015] In an exemplary implementation, determining the historical feedback feature corresponding to the sample object according to the aggregated feedback information sequence includes:

[0016] Determining a second quantity according to the first quantity and a preset dimension;

[0017] The second number of historical feedback features are extracted from the aggregate feedback information sequence according to a preset dimension, and each of the historical feedback features includes continuous aggregate feedback information of the preset dimension.

[0018] In an exemplary embodiment, the method further comprises:

[0019] For each sampling period in the plurality of sampling periods, randomly selecting a first preset number of new objects from the new object pool corresponding to the sampling period, and placing the first preset number of new objects into the sample object pool;

[0020] For each sample object in the sample object pool, randomly delivering the sample object to a second preset number of user accounts in the entire user account space in each delivery cycle, and generating random delivery information of the sample object corresponding to the delivery cycle according to feedback information of the second preset number of user accounts;

[0021] Determine the random delivery times of each sample object in the delivery sample object pool, and remove the sample objects whose random delivery times reach a preset times threshold from the delivery sample object pool;

[0022] The random delivery information of the target sample object is generated according to the random delivery information of the removed sample object corresponding to each delivery period.

[0023] In an exemplary embodiment, the random placement information of the sample objects includes feedback information of at least one dimension;

[0024] The determining, according to the random delivery information corresponding to the sample object, the target feedback indicator corresponding to the sample object comprises:

[0025] According to the feedback information of each dimension, a dimension feedback indicator corresponding to each dimension is obtained;

[0026] Determine a weight coefficient corresponding to each dimension;

[0027] A weighted sum is performed according to the dimension feedback index and the weight coefficient corresponding to each dimension to obtain the target feedback index corresponding to the sample object.

[0028] According to a second aspect of an embodiment of the present disclosure, there is provided an object recommendation method, comprising:

[0029] Acquire the number of feedbacks of the candidate object and the second historical feedback information sequence corresponding to the candidate object;

[0030] Determining, according to the second historical feedback information sequence corresponding to the candidate object, a historical feedback feature corresponding to the candidate object;

[0031] Inputting the historical feedback features corresponding to the candidate object into a feedback index prediction model to perform feedback index prediction, thereby obtaining a first predicted feedback index corresponding to the candidate object;

[0032] Determining a first display opportunity measurement indicator corresponding to the candidate object according to the first predicted feedback indicator corresponding to the candidate object and the feedback quantity;

[0033] According to the first display opportunity measurement indicator corresponding to the candidate object, a first target candidate object is selected from the candidate object pool; the first target candidate object is used as the object to be recommended;

[0034] The feedback indicator prediction model is trained according to the feedback indicator prediction model training method provided in the first aspect.

[0035] In an exemplary embodiment, selecting a target candidate object from the candidate object pool according to the first display opportunity measurement indicator corresponding to each candidate object in the candidate object pool includes:

[0036] Obtaining the original correlation between each candidate object and a target user account; the target user account refers to a user account to be recommended;

[0037] According to the original relevance degree corresponding to each of the candidate objects and the first display opportunity measurement indicator, a first target candidate object is selected from the candidate object pool based on the target regulation rule in the target regulation strategy.

[0038] In an exemplary embodiment, before obtaining the second historical feedback information sequence corresponding to the candidate object, the method further includes:

[0039] When the target start intervention time information in the target regulation strategy is satisfied at the current moment, the step of acquiring the second historical feedback information sequence corresponding to the candidate object is performed.

[0040] In an exemplary embodiment, a plurality of different control strategies are determined, each of the control strategies comprising start intervention time information and control rules;

[0041] According to the recommendation data set corresponding to each of the regulation strategies, a system display opportunity measurement index and a system average click rate corresponding to each of the regulation strategies are determined; the recommendation data set corresponding to the regulation strategy is determined according to a plurality of recommendation logs sorted by time, each of the recommendation logs includes a recommendation record and corresponding feedback information, and the recommendation record includes a user account and a candidate object recommended to the user account;

[0042] A target regulation strategy among the multiple regulation strategies is determined according to a system display opportunity measurement indicator and a system average click rate corresponding to each regulation strategy among the multiple regulation strategies.

[0043] In an exemplary embodiment, determining the system display opportunity measurement index and the system average click rate corresponding to each regulation strategy according to the recommendation data set corresponding to each regulation strategy includes:

[0044] For each of the control strategies, the system display opportunity measurement index corresponding to the control strategy is determined according to the second display opportunity measurement index corresponding to each candidate object in the recommendation data set corresponding to the control strategy; the second display opportunity measurement index corresponding to each candidate object in the recommendation data set is determined according to the feedback indicator prediction model and the feedback information of each recommendation data in the recommendation data set corresponding to the control strategy;

[0045] The system average click rate corresponding to the regulation strategy is determined according to the ratio between the number of recommended data in the recommendation data set corresponding to the regulation strategy and the number of the plurality of recommendation logs.

[0046] In an exemplary embodiment, the method further comprises:

[0047] Obtaining the plurality of recommendation logs sorted by time;

[0048] Marking the plurality of recommended logs as not recommended, thereby obtaining a plurality of not recommended data;

[0049] For each of the control strategies, according to the starting intervention time information in the control strategy, target non-recommended data among the multiple non-recommended data are determined, and the target non-recommended data are marked as recommended data, so as to obtain a recommended data set corresponding to the control strategy;

[0050] Determine a second display opportunity measurement indicator corresponding to each candidate object in the recommendation data set according to the feedback indicator prediction model and the feedback information of each recommendation data in the recommendation data set corresponding to the regulation strategy;

[0051] Determine a second target candidate object from the candidate objects corresponding to the recommendation data set based on the second display opportunity measurement indicator and click-through rate corresponding to each candidate object in the recommendation data set and based on the regulation rules in the regulation strategy;

[0052] Based on the second target candidate object, target remaining non-recommended data is selected from the remaining non-recommended data, the target remaining non-recommended data is marked as recommended data, and the recommended data set corresponding to the regulation strategy is updated.

[0053] According to a third aspect of an embodiment of the present disclosure, a feedback indicator prediction model training device is provided, comprising:

[0054] A random delivery information acquisition unit is configured to acquire random delivery information of a sample object obtained based on a random delivery experiment; the random delivery information represents feedback from a randomly delivered user account on the sample object;

[0055] a target feedback indicator determination unit, configured to determine a target feedback indicator corresponding to the sample object according to the random delivery information corresponding to the sample object;

[0056] A first historical feedback feature determining unit is configured to determine a historical feedback feature corresponding to the sample object according to a first historical feedback information sequence of the sample object in the historical recommendation;

[0057] The training unit is configured to execute a preset machine learning model according to the historical feedback features corresponding to the target sample object and the target feedback index training to obtain a feedback index prediction model; wherein the target feedback index serves as the expected prediction value corresponding to the historical feedback features.

[0058] In an exemplary embodiment, the first historical feedback feature determination unit includes:

[0059] A first acquisition unit is configured to execute acquisition of a first historical feedback information sequence of the sample object in the historical recommendation;

[0060] A subsequence division unit is configured to divide the first historical feedback information sequence into a first number of subsequences according to a preset aggregation number; each of the subsequences includes consecutive first historical feedback information of the preset aggregation number;

[0061] an aggregation processing unit, configured to perform aggregation processing on the first historical feedback information in each subsequence to obtain an aggregate feedback information sequence corresponding to the sample object; the aggregate feedback information sequence includes aggregate feedback information corresponding to each of the subsequences;

[0062] The historical feedback feature determination subunit is configured to determine the historical feedback feature corresponding to the sample object according to the aggregate feedback information sequence.

[0063] In an exemplary embodiment, the historical feedback feature determination subunit includes:

[0064] A second quantity determining unit is configured to determine a second quantity according to the first quantity and a preset dimension;

[0065] The historical feedback feature extraction unit is configured to extract the second number of historical feedback features from the aggregate feedback information sequence according to the preset dimension; each of the historical feedback features includes continuous aggregate feedback information of the preset dimension.

[0066] In an exemplary embodiment, the apparatus further comprises:

[0067] The random sampling unit is configured to randomly select a first preset number of new objects from the new object pool corresponding to the sampling period for each sampling period in the plurality of sampling periods, and put the first preset number of new objects into the sample object pool;

[0068] A random delivery unit is configured to execute, for each sample object in the delivery sample object pool, randomly delivering the sample object to a second preset number of user accounts in the entire user account space in each delivery cycle, and generating random delivery information of the sample object corresponding to the delivery cycle according to feedback information of the second preset number of user accounts;

[0069] A sample object removing unit is configured to determine the random delivery times of each sample object in the delivery sample object pool, and remove the sample objects whose random delivery times reach a preset times threshold from the delivery sample object pool;

[0070] The random delivery information generating unit is configured to generate the random delivery information of the sample object according to the random delivery information of each delivery period corresponding to the removed sample object.

[0071] In an exemplary embodiment, the random delivery information of the sample objects includes feedback information of at least one dimension; the target feedback index determination unit includes:

[0072] A dimension feedback indicator determination unit is configured to obtain a dimension feedback indicator corresponding to each dimension according to the feedback information of each dimension;

[0073] A weight coefficient determination unit, configured to determine a weight coefficient corresponding to each dimension;

[0074] The weighted summation unit is configured to perform weighted summation according to the dimensional feedback indicators and weight coefficients corresponding to each dimension to obtain the target feedback indicators corresponding to the sample object.

[0075] According to a fourth aspect of an embodiment of the present disclosure, there is provided an object recommendation device, including:

[0076] A second acquisition unit is configured to execute acquisition of the number of feedbacks corresponding to the candidate object and a second historical feedback information sequence corresponding to the candidate object;

[0077] A second historical feedback feature determining unit is configured to determine the historical feedback feature corresponding to the candidate object according to the second historical feedback information sequence corresponding to the candidate object;

[0078] A first feedback index prediction unit is configured to input the historical feedback features corresponding to the candidate object into a feedback index prediction model to perform feedback index prediction, and obtain a first predicted feedback index corresponding to the candidate object;

[0079] A first display opportunity measurement index determining unit is configured to determine a first display opportunity measurement index corresponding to the candidate object according to the first predicted feedback index corresponding to the candidate object and the feedback quantity;

[0080] A first target candidate object selection unit is configured to select a first target candidate object from a candidate object pool according to a first display opportunity measurement indicator corresponding to the candidate object; the first target candidate object is used as an object to be recommended;

[0081] The feedback indicator prediction model is trained according to the feedback indicator prediction model training method provided in the first aspect.

[0082] In an exemplary embodiment, the first target candidate object selection unit includes:

[0083] a relevance determination unit configured to obtain an original relevance between each candidate object and a target user account; the target user account refers to a user account to be recommended;

[0084] The selection subunit is configured to select a first target candidate object from the candidate object pool based on the target regulation rules in the target regulation strategy according to the original relevance degree and the first display opportunity measurement indicator corresponding to each of the candidate objects.

[0085] In an exemplary embodiment, the apparatus further comprises:

[0086] The execution unit is configured to execute the step of obtaining the second historical feedback information sequence corresponding to the candidate object when the target starting intervention time information in the target regulation strategy is met at the current moment.

[0087] In an exemplary embodiment, the apparatus further comprises:

[0088] A control strategy determination unit is configured to execute and determine a plurality of different control strategies, each of which includes start intervention time information and control rules;

[0089] A first determination unit is configured to execute, according to the recommendation data set corresponding to each of the regulation strategies, to determine the system display opportunity measurement index and the system average click rate corresponding to each of the regulation strategies; the recommendation data set corresponding to the regulation strategy is determined according to a plurality of recommendation logs sorted by time, each of the recommendation logs includes a recommendation record and corresponding feedback information, and the recommendation record includes a user account and a candidate object recommended to the user account;

[0090] The target regulation strategy determination unit is configured to determine the target regulation strategy among the multiple regulation strategies according to the system display opportunity measurement index and the system average click rate corresponding to each regulation strategy among the multiple regulation strategies.

[0091] In an exemplary embodiment, the first determining unit includes:

[0092] The system display opportunity measurement index determination unit is configured to execute, for each of the control strategies, according to the second display opportunity measurement index corresponding to each candidate object in the recommendation data set corresponding to the control strategy, determine the system display opportunity measurement index corresponding to the control strategy; the second display opportunity measurement index corresponding to each candidate object in the recommendation data set is determined according to the feedback index prediction model and the feedback information of each recommendation data in the recommendation data set corresponding to the control strategy;

[0093] The system average click rate determination unit is configured to determine the system average click rate corresponding to the regulation strategy according to the ratio between the number of recommended data in the recommendation data set corresponding to the regulation strategy and the number of the multiple recommendation logs.

[0094] In an exemplary embodiment, the apparatus further comprises:

[0095] A recommendation log acquisition unit, configured to acquire the plurality of recommendation logs sorted by time;

[0096] A first marking unit is configured to mark the plurality of recommended logs as non-recommended to obtain a plurality of non-recommended data;

[0097] A second marking unit is configured to execute, for each of the control strategies, determining target non-recommended data among the multiple non-recommended data according to the starting intervention time information in the control strategy, marking the target non-recommended data as recommended data, and obtaining a recommended data set corresponding to the control strategy;

[0098] A second display opportunity measurement index determining unit is configured to determine a second display opportunity measurement index corresponding to each candidate object in the recommendation data set according to the feedback index prediction model and the feedback information of each recommendation data in the recommendation data set corresponding to the regulation strategy;

[0099] A second target candidate object determination unit is configured to determine a second target candidate object from the candidate objects corresponding to the recommendation data set based on a second display opportunity measurement indicator and a click-through rate corresponding to each candidate object in the recommendation data set and based on a control rule in the control strategy;

[0100] The recommended data set updating unit is configured to select target remaining non-recommended data from the remaining non-recommended data based on the second target candidate object, mark the target remaining non-recommended data as recommended data and update the recommended data set corresponding to the regulation strategy.

[0101] According to a fifth aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0102] processor;

[0103] a memory for storing instructions executable by the processor;

[0104] Among them, the processor is configured to execute the instructions to implement the feedback indicator prediction model training method of the first aspect or the object recommendation method of the second aspect.

[0105] According to the sixth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the feedback indicator prediction model training method of the first aspect or the object recommendation method of the second aspect.

[0106] According to a seventh aspect of an embodiment of the present disclosure, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the feedback indicator prediction model training method of the first aspect or the object recommendation method of the second aspect.

[0107] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0108] By obtaining random delivery information of sample objects obtained based on random delivery experiments, determining the target feedback index corresponding to the sample objects according to the random delivery information, and determining the historical feedback features corresponding to the sample objects according to the historical feedback information sequence of the sample objects in historical recommendations, and then using the above target feedback index as the expected prediction value corresponding to the historical feedback features, training a preset machine learning model based on the historical feedback features and the target feedback index to obtain a feedback index prediction model; in the above technical solution, since the random delivery information represents the feedback of randomly delivered user accounts on the sample objects, the target feedback index determined based on the random delivery information can represent the feedback of all user accounts on the sample objects, that is, the target feedback index is an unbiased value measurement of the sample objects, then using the target feedback index as the expected prediction value in model training can improve the accuracy of the feedback index prediction model obtained by training for the prediction of item value, and the prediction result of the feedback index prediction model can reflect the true value of the items, which is conducive to improving the matching between the display opportunities of each item and its true value.

[0109] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0111] Figure 1 is a schematic diagram of an application environment according to an exemplary embodiment;

[0112] Figure 2 is a flow chart of a feedback indicator prediction model training method according to an exemplary embodiment;

[0113] Figure 3 is a flow chart showing a method of generating random placement information of sample objects based on a random placement experiment according to an exemplary embodiment;

[0114] Figure 4is a flowchart showing a method of determining a historical feedback feature corresponding to a sample object according to a first historical feedback information sequence of the sample object in historical recommendation according to an exemplary embodiment;

[0115] Figure 5 is a flowchart of an object recommendation method according to an exemplary embodiment;

[0116] Figure 6 is a flowchart showing a method of selecting a first target candidate object from a candidate object pool according to a first fair value score corresponding to the candidate object according to an exemplary embodiment;

[0117] Figure 7 is a block diagram of a feedback indicator prediction model training device according to an exemplary embodiment;

[0118] Figure 8 is a block diagram of an object recommendation device according to an exemplary embodiment;

[0119] Fig. 9 It is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0120] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0121] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0122] See also Figure 1 , which is a schematic diagram of a reaction application environment according to an exemplary embodiment, the application environment may include a terminal 110 and a server 120, and the terminal 110 and the server 120 may be connected via a wired network or a wireless network.

[0123] The terminal 110 may be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc., but is not limited thereto. The terminal 110 may be installed with client software such as an application (Application, referred to as App) that provides a human-computer interaction function. The application may be an independent application or a subprogram in the application. Exemplarily, the application may include a video application, an information application, a live broadcast application, etc. The user account of the terminal 110 may log in to the application through the pre-registered user information, which may include an account number and a password.

[0124] The server 120 may be a server that provides background services for the application in the terminal 110. Specifically, the service provided by the server 120 may be an object recommendation service, and the object may be determined according to the specific application scenario, for example, including but not limited to video, news information, advertisement, etc. The server 120 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0125] The method of the embodiment of the present disclosure may be executed by an electronic device, which may be a terminal or a server, or may be executed by a terminal and a server in cooperation with each other, which is not specifically limited in the embodiment of the present disclosure.

[0126] Figure 2 is a flowchart of a feedback indicator prediction model training method according to an exemplary embodiment. Figure 2 As shown, the following steps are included:

[0127] In step S201, random placement information of sample objects obtained based on a random placement experiment is obtained.

[0128] The random delivery information represents the feedback of the randomly delivered user account on the sample object, and the randomly delivered user account is a user account randomly selected from the entire user account space. The sample object is an object randomly collected in the random delivery experiment.

[0129] Based on this, in an exemplary implementation, the embodiment of the present disclosure may further include a step of generating random placement information of sample objects based on a random placement experiment, such as Figure 3 As shown, when generating random placement information of sample objects based on a random placement experiment, the following may be included:

[0130] In step S301, for each sampling period in a plurality of sampling periods, a first preset number of new objects are randomly selected from a new object pool corresponding to the sampling period, and the first preset number of new objects are placed in a sample object pool.

[0131] Among them, the sampling period can be set according to actual needs, for example, it can be measured in hours, with sampling once every hour. The number of sampling periods can be set according to actual needs, such as 24 consecutive sampling periods. The newly added object pool includes the newly uploaded objects corresponding to the current sampling period. Taking the sampling period of 1 hour as an example, the sampling is once every hour. When the sampling period of 9:00 arrives, the first preset number of newly uploaded objects can be randomly selected from the newly uploaded objects (such as newly uploaded videos) between 8:00 and 9:00. The first preset number can be an empirical value set according to the actual application situation, or it can be a preset ratio of the number of newly uploaded objects corresponding to the sampling period.

[0132] In an exemplary embodiment, the newly added object may be a newly added object or a newly added class of objects. Accordingly, each sample object may indicate one, for example, a video, or may indicate multiple, for example, videos of the same category. When the newly added objects indicate multiple, for example, videos of the same type, each newly added object may indicate the same number of the same category, so that the newly added objects can maintain a balance in number, thereby improving the accuracy of subsequent unbiased value measurement.

[0133] In step S303, for each sample object in the sample object pool, the sample object is randomly delivered to a second preset number of user accounts in the entire user account space in each delivery cycle, and random delivery information of the sample object corresponding to the current sampling cycle is generated based on feedback information from the second preset number of user accounts.

[0134] The second preset number may be an empirical value set according to actual application conditions, or a value determined according to the number of objects in the system or the number of newly added objects corresponding to each sampling period.

[0135] In the embodiment of the present disclosure, a plurality of continuous delivery cycles may be pre-set, and in each delivery cycle, each sample object in the delivery sample object pool is randomly delivered to a second preset number of user accounts in the entire user account space.

[0136] In step S305, the random delivery times of each sample object in the delivery sample object pool are determined, and the sample objects whose random delivery times reach a preset times threshold are removed from the delivery sample object pool.

[0137] The preset number threshold can be determined by multiplying the number of delivery cycles by the second preset number. For example, if there are b delivery cycles, and each delivery sample in the delivery sample object pool is randomly delivered to Y user accounts in the total user account space in each delivery cycle, then the number of random delivery times that each sample object in the delivery sample object pool needs to reach is b*Y, that is, the preset number threshold is b*Y.

[0138] In step S307, random delivery information of the sample object is generated according to the random delivery information of each delivery period corresponding to the removed sample object.

[0139] It can be understood that when there are multiple sample objects to be removed, random placement information of each removed sample object can be generated.

[0140] For example, assuming the sampling period is 1 hour, X new objects are collected every hour and put into the sample object pool. There are a total of 3 consecutive sampling periods, namely 1 o'clock, 2 o'clock and 3 o'clock. Each sample object is randomly delivered to Y user accounts in the entire user account space. The delivery period is 2, then the preset number threshold is 2*Y, and delivery can be carried out while collection is in progress.

[0141] When the sampling period corresponding to point 1 arrives, X (1 point) objects are randomly selected from the newly added objects before point 1 and put into the sample object pool. That is, there are a total of X (1 point) sample objects in the sample object pool at this time. Each sample object X in the X (1 point) sample objects is randomly selected and put into the sample object pool. i Randomly distribute it to Y user accounts in the entire user account space, and generate X according to the feedback information of the Y user accounts. i Corresponding to the random placement information of 1 point. It should be noted that each sample object X i The Y user accounts randomly placed are selected completely randomly, that is, the Y user accounts corresponding to different sample objects may have some of the same user accounts, or may be completely different user accounts. At this time, the number of random placements for each sample object in the sample object pool is Y times. If the preset number threshold is not reached, the sample object will not be removed.

[0142] When the sampling period corresponding to point 2 arrives, X (2 points) are randomly selected from the newly added objects between point 1 and point 2 and put into the sample object pool. That is, at this time, there are a total of X (1 point) + X (2 points) sample objects in the sample object pool. Each sample object X in the 2X sample objects is randomly selected from the newly added objects between point 1 and point 2 and put into the sample object pool. i Randomly distribute it to Y user accounts in the entire user account space, and generate X according to the feedback information of the Y user accounts. iCorresponding to the random delivery information of 2 points. At this time, the number of random delivery times of each of the X (1 point) sample objects in the delivery sample object pool is 2Y, and the number of random delivery times of each of the X (2 points) sample objects is Y. Then the X (1 point) sample objects can be removed from the delivery sample object pool, and each of the X (1 point) sample objects is a removed sample object. The random delivery information of the removed sample object can be generated according to the random delivery information corresponding to 1 point and 2 points.

[0143] Since there are X (2 points) sample objects left in the delivery object pool at this time, and since the X (2 points) sample objects have only experienced one delivery cycle, each sample object X (2 points) of the X (2 points) sample objects in the delivery object pool can be delivered at 3 o'clock in the next delivery cycle. i Randomly distribute it to Y user accounts in the entire user account space, and generate X according to the feedback information of the Y user accounts. i Corresponding to the random delivery information of 3 points. At this time, the number of random delivery of each sample object in the delivery sample object pool of X (2 points) sample objects is 2Y, then the X (2 points) sample objects can be removed from the delivery sample object pool, and each of the X (2 points) sample objects is a removed sample object, and the random delivery information of the removed sample object can be generated according to the random delivery information corresponding to 2 points and 3 points.

[0144] At this point, the sample object pool is empty and the random delivery experiment can be ended. Then 2X removed sample objects can be obtained. Each removed sample object corresponds to 2*Y random delivery records. Each random delivery record represents the correspondence between the user account, sample object and feedback information of the random delivery. That is to say, in the above example, the random delivery information of each sample object is composed of 2*Y random delivery records.

[0145] The disclosed embodiment can obtain the feedback of the sample object in the random delivery user account through the above-mentioned random delivery experiment. The feedback can be considered as the feedback of all user accounts to the sample object in terms of probability, so that the random delivery information can eliminate the exposure bias problem existing in the system recommendation log.

[0146] In step S203, the target feedback index corresponding to the sample object is determined according to the random delivery information corresponding to the sample object.

[0147] Specifically, the target feedback indicator may be an indicator calculated based on feedback information in the random delivery information, for example, the target feedback indicator may be a click-through rate.

[0148] Since the random delivery information of the sample object can be regarded as the feedback of all user accounts on the sample object, the target feedback index determined based on the random delivery information can reflect the true value of the sample object and is an unbiased value measurement of the sample object.

[0149] In an exemplary embodiment, in order to more accurately measure the unbiased value of the sample object, the random delivery information of the sample object may include feedback information of at least one dimension, for example, it may include a click dimension, a like dimension, a comment dimension, etc. Then, when determining the target feedback index corresponding to the sample object according to the random delivery information corresponding to the sample object, step S203 may include the following steps:

[0150] (1) According to the feedback information of each dimension, the dimension feedback index corresponding to each dimension is obtained. For example, for the click dimension, the click rate can be calculated based on the feedback information of the click dimension in the random delivery information; for the like dimension, the like rate can be calculated based on the feedback information of the like dimension in the random delivery information; for the comment dimension, the comment rate can be calculated based on the feedback information of the comment dimension in the random delivery information, then the click rate, like rate and comment rate are the dimension feedback indexes of the corresponding dimension.

[0151] (2) Determine a weight coefficient corresponding to each dimension. The weight coefficient can be set based on experience in actual applications.

[0152] (3) Performing a weighted summation based on the dimension feedback indicators and weight coefficients corresponding to each dimension to obtain the target feedback indicator corresponding to the sample object.

[0153] Specifically, the target feedback index F = ∑w j *f j , where j represents the dimension and w j represents the weight coefficient corresponding to dimension j, f j Represents the dimension feedback indicator corresponding to dimension j.

[0154] Taking the feedback information of the three dimensions of click dimension, like dimension and comment dimension as an example, the target feedback index F = w 点击率 *f 点击率 +w 点赞率 *f 点赞率 +w 评论率 *f 评论率 .

[0155] When the random placement information of the sample object includes feedback information of at least one dimension, the embodiment of the present disclosure determines the target feedback index corresponding to the target sample object based on the dimensional feedback index of each dimension, thereby improving the accuracy of unbiased value measurement of the sample object.

[0156] In step S205, the historical feedback feature corresponding to the sample object is determined according to the first historical feedback information sequence of the sample object in the historical recommendation.

[0157] Specifically, the first historical feedback information sequence of the sample object in the historical recommendation can be obtained according to the recommendation log of the recommendation system, and the first historical feedback information in the first historical feedback information sequence is arranged in the time sequence of recommendation or feedback.

[0158] In step S207, a preset machine learning model is trained according to the historical feedback features corresponding to the sample object and the target feedback index to obtain a feedback index prediction model; wherein the target feedback index serves as the expected prediction value corresponding to the historical feedback features.

[0159] Specifically, the historical feedback features corresponding to the sample objects can be input into the preset machine learning model in the form of a vector to perform feedback index prediction processing to obtain an output result, and then the loss value is determined based on the difference between the output result and the target feedback index. The model parameters of the preset machine learning model are reversely adjusted based on the loss value and the iterative training is continued until the preset training end conditions are met and the training is ended. The machine learning model corresponding to the model parameters at the end of the training is used as the feedback index prediction model.

[0160] In order to improve the accuracy of the feedback index prediction model for unbiased value prediction of items, in an exemplary embodiment, the preset machine learning model can be a sequence model, such as a GRU model or a Transformer model, wherein GRU (Gate Recurrent Unit) is a type of recurrent neural network (RNN). The loss value in training can be, but is not limited to, calculated using the mean square error MSE as the loss function. Based on this, in an exemplary embodiment, the above step S205 may include, when determining the historical feedback features corresponding to the sample object based on the first historical feedback information sequence of the sample object in the historical recommendation, Figure 4 The following steps:

[0161] In step S401, a first historical feedback information sequence of the sample object in historical recommendation is obtained.

[0162] Specifically, the first historical feedback information of the sample object within a preset historical time period can be obtained according to the recommendation log of the recommendation system, and the first historical feedback information is sorted according to the recommendation time or feedback time corresponding to each first historical feedback information to obtain the first historical feedback information sequence of the sample object in the historical recommendation.

[0163] In step S403, the first historical feedback information sequence is divided into a first number of subsequences according to a preset aggregation quantity.

[0164] Each of the subsequences includes the preset aggregation number of consecutive historical feedback information.

[0165] The preset aggregation number can be set according to the data size of the first historical feedback information sequence in actual application, for example, any integer can be selected in the range of 10 to 50. The first historical feedback information sequence of the sample object includes 500 first historical feedback information arranged in order from the earliest to the latest recommendation time, which is represented as {FB1, FB2, ..., FB 500}, where FB represents the first historical feedback information, and the preset aggregation number is 10, for example, then 500 / 10=50 subsequences can be obtained: {FB1, FB2, ..., FB 10}1、{FB 11 , FB 12 ,…,FB 20}2、……、{FB 401 , FB 402 ,…,FB 500 ) 50 .

[0166] In step S405, the first historical feedback information in each subsequence is aggregated to obtain an aggregated feedback information sequence corresponding to the sample object.

[0167] The aggregate feedback information sequence includes aggregate feedback information corresponding to each of the subsequences. Specifically, the aggregation process may be to calculate an average value of the first historical feedback information in the subsequence, and use the average value as the aggregate feedback information corresponding to the subsequence.

[0168] Still taking the above 50 subsequences as an example, for the subsequences {FB1, FB2, ..., FB 10}1, its corresponding aggregate feedback information It can be expressed as (FB1+FB2+…+FB 10 ) / 10, so that the subsequence can be represented by the calculated value, then the aggregate feedback information sequence corresponding to the sample object can be expressed as It includes 50 data points, that is, the number of data points in the aggregated feedback information sequence of the sample object m=[N / x], where N is the number of first historical feedback information in the first historical feedback information sequence of the sample object, and x is the preset aggregation number.

[0169] In step S407, the historical feedback features corresponding to the sample object are determined according to the aggregated feedback information sequence.

[0170] In the above embodiment, since the aggregated feedback information in the aggregated feedback information sequence is determined based on multiple consecutive first historical feedback information in the corresponding subsequence, the accuracy of the historical feedback features corresponding to the sample object can be improved, thereby improving the training effect of the feedback index prediction model.

[0171] In an exemplary implementation, the above step S407 may include:

[0172] Determining a second quantity according to the first quantity and a preset dimension;

[0173] The second number of historical feedback features are extracted from the aggregated feedback information sequence according to the preset dimension, each of the historical feedback features including consecutive aggregated feedback information of the preset dimension.

[0174] The second number is determined according to the first number and the preset dimension. Specifically, the second number can be the difference between the first number and the preset dimension plus the interval number used in the extraction. The interval number can be set according to actual experience, for example, it can be 1 or 2. It should be noted that the interval number here refers to the number of data points between two adjacent historical feedback features. In practical applications, in order to avoid data waste, the preferred interval number is set to 1.

[0175] For example, assuming that the preset dimension is y, the aggregate feedback information sequence corresponding to the sample object is expressed as You can extract The historical feedback features that constitute the first y dimension can be extracted if the number of intervals is 1 The historical feedback features that constitute the second y dimension, and so on, until the extraction This constitutes the m-y+1th historical feedback feature of y dimension, so that m-y+1 historical feedback features of y dimension can be obtained for the sample object.

[0176] When performing model training, the second number of historical feedback features corresponding to the sample objects may be respectively input into the sequence model for training, and the expected prediction value corresponding to each historical feedback feature is the target feedback indicator.

[0177] In the above embodiment, since the second number is determined based on the first number and the preset dimension, when the second number of historical feedback features are extracted from the aggregated feedback information sequence according to the preset dimension, the number of extracted historical feedback features and the length of each historical feedback feature can be regulated by the preset dimension, so that the training input data can be controlled more flexibly, which can improve the training efficiency while improving the model training effect.

[0178] The disclosed embodiment determines the target feedback index based on the random delivery information. Since the random delivery information is obtained based on random experiments, it can probabilistically characterize the feedback of all user accounts on the corresponding sample objects, so that the target feedback index can measure the unbiased value of the sample objects, that is, reflect the true value of the sample objects. When the model is trained in combination with the historical feedback information sequence of the sample objects, the target feedback index is used as the expected prediction value, which can improve the accuracy of the feedback index prediction model obtained by training for the prediction of the value of items. The prediction result of the feedback index prediction model can reflect the true value of the items, which is beneficial to improve the matching between the display opportunities of each item and its true value.

[0179] In an exemplary embodiment, see Figure 5 , which is a flowchart of an object recommendation method according to an exemplary embodiment. Figure 5 As shown, including:

[0180] In step S501, the number of feedbacks corresponding to the candidate object and the second historical feedback information sequence corresponding to the candidate object are obtained.

[0181] Among them, the candidate object is any candidate object in the candidate object pool. In actual applications, each candidate object in the candidate object pool can indicate one or multiple candidates of the same category. For example, each candidate object can be a video or a group of videos of the same category.

[0182] In a specific implementation, the terminal may generate and send a recommendation request for the target user account in response to a refresh operation of the target user account. The recommendation request may include information such as a user identifier of the target user account.

[0183] The number of feedbacks corresponding to the candidate object can be determined based on the recommendation log corresponding to the candidate object in the recommendation system. The number of feedbacks can be the total number of recommendations corresponding to the candidate object or the number of clicks corresponding to the candidate object.

[0184] The second historical feedback information sequence corresponding to the candidate object may be a preset number of second historical feedback information of the candidate object closest to the current moment, and the second historical feedback information in the second historical feedback information sequence may be arranged in descending order according to recommendation time or feedback time.

[0185] In step S503, the historical feedback features corresponding to the candidate object are determined according to the second historical feedback information sequence corresponding to the candidate object.

[0186] Specifically, the historical feedback feature corresponding to the candidate object can be determined according to the second historical feedback information sequence corresponding to the candidate object. Figure 4 The method embodiment shown will not be described in detail here.

[0187] In step S505, the historical feedback features corresponding to the candidate object are input into a feedback index prediction model to perform feedback index prediction, so as to obtain a first predicted feedback index corresponding to the candidate object.

[0188] Among them, the feedback indicator prediction model is based on the aforementioned embodiment of the present disclosure Figure 2 The method embodiment shown is obtained by training.

[0189] In step S507, a first display opportunity measurement indicator corresponding to the candidate object is determined according to the first predicted feedback indicator corresponding to the candidate object and the feedback quantity.

[0190] It can be understood that the display opportunity measurement index is used to measure the display opportunity of the corresponding candidate object. Generally, the larger the display opportunity measurement index is, the greater the display opportunity of the corresponding candidate object is, and the more likely it is to be recommended for display.

[0191] In a specific implementation, the first display opportunity measurement indicator corresponding to the candidate object can be the ratio of the first predicted feedback indicator corresponding to the candidate object to the number of feedbacks. In this case, the larger the ratio, the greater the display opportunity of the corresponding candidate object, and the more likely it is to be recommended for display.

[0192] In step S509, a first target candidate object is selected from the candidate object pool according to the first display opportunity measurement indicator corresponding to the candidate object.

[0193] Among them, the first target candidate object is used as the object to be recommended. Exemplarily, the object to be recommended can be used to recommend to the target user account in response to the recommendation request of the target user account. Specifically, the candidate object with the largest first display opportunity measurement index in the candidate object pool can be used as the first target candidate object; of course, the candidate objects in the candidate object pool can also be sorted in descending order according to the first display opportunity measurement index, and the candidate object ranked first in the sorting order is selected as the first target candidate object.

[0194] The disclosed embodiment obtains the number of feedbacks corresponding to the candidate object and the corresponding second historical feedback information sequence, and determines the historical feedback features based on the second historical feedback information sequence, and then inputs the historical feedback features into the trained feedback index prediction model to predict the feedback index to obtain a first predicted feedback index, and determines the first fair value share corresponding to the candidate object according to the first predicted feedback index and the number of feedbacks, and selects the first target candidate object from the candidate object pool based on the first fair value share and recommends it to the target, thereby improving the fairness of item recommendation from the perspective of fairness of item recommendation in the recommendation scenario, so that the recommended target candidate object matches its true value.

[0195] In an exemplary embodiment, in order to enable the recommendation system to have both fair recommendations of items and benefits for the recommendation system when making recommendations, candidate objects to be recommended can be selected from two dimensions: fairness and quality. Among them, the quality dimension can be measured in an actual recommendation scenario by the correlation between the candidate objects provided by the recommendation system itself and the target user account, because generally, the more relevant the candidate objects are, the higher the probability of being clicked, and thus the benefits of the recommendation system are also higher.

[0196] Based on this, in an exemplary embodiment, Figure 6 As shown, step S509, when selecting the first target candidate object from the candidate object pool according to the first display opportunity measurement indicator corresponding to the candidate object, may include:

[0197] In step S601, the original correlation degree between each candidate object and the target user account is obtained.

[0198] The target user account refers to the user account to be recommended, and specifically may be the user account corresponding to the recommendation request. The original correlation between each candidate object in the candidate object pool and the target user may be obtained by any method of determining the correlation between the candidate object and the target user provided in the prior art, and the embodiments of the present disclosure do not specifically limit this.

[0199] In step S603, according to the original relevance degree corresponding to each of the candidate objects and the first display opportunity measurement indicator, a first target candidate object is selected from the candidate object pool based on the target regulation rule in the target regulation strategy.

[0200] Among them, the target regulation strategy can be determined based on offline simulation experiments before online recommendation, and the target adjustment rule can be one of the following determined based on offline model experiments.

[0201] In an exemplary embodiment, the target regulation rule may be the first regulation rule, that is, the original relevance degree corresponding to the candidate object and the first display opportunity measurement index are weighted and summed to obtain the first regulation rule score corresponding to each candidate object, and then the candidate object with the highest first regulation rule score is selected as the first target candidate object. Specifically, the first regulation rule score can be expressed as the following formula (1):

[0202] S(i)=β*S1(i)+(1-β)*S2(i) (1)

[0203] Among them, S(i) represents the first regulation rule score of candidate object i, S1(i) represents the original relevance of candidate object i, S2(i) represents the first display opportunity measurement index of candidate object i, and β is a coefficient determined based on offline simulation experiments.

[0204] In another exemplary embodiment, the target regulation rule may be a second regulation rule, that is, finding the first candidate object whose original relevance exceeds the first target threshold from the candidate object pool, and then selecting the first candidate object with the highest first display opportunity measurement index from the first candidate object as the first target candidate object. In a specific implementation, the second regulation rule score of each candidate object may be calculated according to the following formula (2), and then the candidate object with the highest second regulation rule score may be selected as the first target candidate object.

[0205]

[0206] Among them, S′(i) represents the second regulation rule score of candidate object i, S1(i) represents the original relevance of candidate object i, S2(i) represents the first display opportunity measurement index of candidate object i, and β is a coefficient determined based on offline simulation experiments. It shows the average of the original correlation degree of each candidate object in the candidate object pool. represents the first target threshold, I(x) represents the conditional judgment function, if x is true, the function value is 1, otherwise it is 0.

[0207] In another exemplary embodiment, the target regulation rule may be a third regulation rule, that is, finding a second candidate object whose first display opportunity measurement index exceeds the second target threshold from the candidate object pool, and then selecting the second candidate object with the highest original correlation from the second candidate objects as the first target candidate object. In a specific implementation, the third regulation rule score of each candidate object may be calculated according to the following formula (3), and then the candidate object with the highest third regulation rule score may be selected as the first target candidate object.

[0208]

[0209] Where S″(i) represents the third regulation rule score of candidate object i, S1(i) represents the original relevance of candidate object i, S2(i) represents the first display opportunity measurement index of candidate object i, and β is a coefficient determined based on offline simulation experiments. Represents the average of the first impression opportunity measurement indicators corresponding to each candidate in the candidate pool. represents the second target threshold, I(x) represents the conditional judgment function, if x is true, the function value is 1, otherwise it is 0.

[0210] The disclosed embodiment adopts a combination of the fairness dimension and the quality dimension to select candidate objects to be recommended, which can improve the fairness of item recommendation while ensuring the high benefit of the recommendation system, thus achieving a win-win situation.

[0211] In an exemplary embodiment, see Figure 6 In order to further ensure the fairness of item recommendation and the win-win situation of recommendation system benefits, before obtaining the second historical feedback information sequence corresponding to the candidate object, it is also possible to determine whether the target start intervention time information in the target regulation strategy is met at the current moment. Only when the target start intervention time information in the target regulation strategy is met at the current moment, the step of obtaining the second historical feedback information sequence corresponding to the candidate object is performed. For example, if the target start intervention time information is expressed as the system's recommendation amount, it is possible to determine whether to perform recommendation intervention based on the recommendation amount of the recommendation system at the current moment. If the recommendation amount of the recommendation system at the current moment reaches the recommendation amount indicated by the target start intervention time information, the second historical feedback information sequence corresponding to the candidate object can be obtained, and recommendation intervention can be performed based on this.

[0212] Based on this, in an exemplary embodiment, before online recommendation, the method may further include an offline simulation experiment phase to determine the target regulation strategy, specifically including the following steps:

[0213] (1) Determine a plurality of different control strategies, each of which includes start intervention time information and control rules.

[0214] The starting intervention time information may be the time length from the first recommendation or the number of recommendations that have been recommended. The control rules may include the first control rule, the second control rule and the third control rule. Parameters such as β in each control rule may be set. Different parameter settings in the same control rule result in different corresponding control strategies.

[0215] (2) Determine the system display opportunity measurement index and system average click rate corresponding to each of the regulation strategies based on the recommendation data set corresponding to each of the regulation strategies.

[0216] The recommended data set corresponding to the control strategy is determined according to a plurality of recommendation logs sorted by time, each of which includes a recommendation record and corresponding feedback information, and the recommendation record includes a user account and a candidate object recommended to the user account. The process of determining the recommended data set corresponding to each control strategy will be described in detail later.

[0217] In an exemplary embodiment, for each control strategy, the system display opportunity measurement index corresponding to the control strategy may be determined according to the second display opportunity measurement index corresponding to each candidate object in the recommendation data set corresponding to the control strategy;

[0218] Among them, the second display opportunity measurement index corresponding to each candidate object in the recommended data set corresponding to the control strategy can be determined according to the feedback index prediction model and the feedback information of each recommended data in the recommended data set corresponding to the control strategy. Specifically, according to the feedback information of each recommended data in the recommended data set corresponding to the control strategy, the feedback information sequence corresponding to each candidate object in the recommended data set can be obtained. Then, using the aforementioned embodiment of the present disclosure Figure 4 The method embodiment shown can obtain the historical feedback features corresponding to each candidate object in the recommendation data set based on the feedback information sequence corresponding to each candidate object in the recommendation data set, and obtain the second predicted feedback index corresponding to each candidate object in the recommendation data set by inputting the historical feedback features corresponding to each candidate object in the recommendation data set into the feedback index prediction model for feedback index prediction. Then, the second display opportunity measurement index corresponding to each candidate object in the recommendation data set can be obtained by the ratio of the second predicted feedback index corresponding to each candidate object in the recommendation data set to the number of feedbacks corresponding to the corresponding candidate object in the recommendation data set.

[0219] Specifically, the system display opportunity measurement index corresponding to the control strategy may be the variance of the second display opportunity measurement index corresponding to each candidate object in the recommendation data set corresponding to the control strategy, or may be the Herfindahl-Hirschman index.

[0220] For each control strategy, the system average click rate corresponding to the control strategy may be determined according to the ratio between the number of recommended data in the recommendation data set corresponding to the control strategy and the number of the plurality of recommendation logs.

[0221] The above embodiment provides a specific implementation method for determining the system display opportunity measurement index and the system average click-through rate corresponding to each regulation strategy, ensuring the accuracy of the determination of the system display opportunity measurement index and the system average click-through rate, and further ensuring the accuracy of the subsequent determination of the target regulation strategy, which is conducive to achieving a win-win situation of fairness in item recommendation and benefits of the recommendation system during online recommendation.

[0222] (3) Determine a target control strategy among the multiple control strategies based on a system display opportunity measurement indicator and a system average click-through rate corresponding to each control strategy among the multiple control strategies.

[0223] For example, the system display opportunity measurement index and the system average click rate corresponding to each control strategy may be comprehensively compared, so as to select the control strategy with the highest system display opportunity measurement index and the highest system average click rate as the target control strategy.

[0224] The disclosed embodiment determines a target regulation strategy through the system display opportunity measurement index and the system average click-through rate corresponding to each regulation strategy, thereby ensuring that subsequent online recommendations based on the target regulation strategy can achieve a win-win situation of item recommendation fairness and recommendation system benefits.

[0225] In an exemplary embodiment, the method may further include the step of determining a recommended data set corresponding to each control strategy, the step specifically including:

[0226] (1) Obtain multiple recommendation logs sorted by time. Specifically, the multiple recommendation logs can be sorted by recommendation time or feedback time.

[0227] (2) Marking the multiple recommendation logs as not recommended to obtain multiple non-recommended data.

[0228] Since the multiple recommendation logs are marked as not recommended, the feedback information in the multiple recommendation logs cannot be used here.

[0229] (3) For each of the control strategies, determine the target non-recommended data among the multiple non-recommended data according to the starting intervention time information in the control strategy, mark the target non-recommended data as recommended data, and obtain a recommended data set corresponding to the control strategy.

[0230] For example, the starting intervention time information in the current control strategy is that the number of recommendations is 2, which means that intervention starts from the second non-recommended data among multiple non-recommended data sorted by time, that is, the two non-recommended data sorted in front are marked as recommended data, thereby obtaining the recommended data set corresponding to the control strategy at the current moment.

[0231] (4) Determine a second display opportunity measurement indicator corresponding to each candidate object in the recommendation data set according to the feedback indicator prediction model and the feedback information of each recommendation data in the recommendation data set corresponding to the regulation strategy.

[0232] Since it is marked as recommended data, the feedback information of each recommended data in the recommended data set can be used at this time. Specifically, according to the feedback information of each recommended data in the recommended data set corresponding to the current control strategy, the feedback information sequence corresponding to each candidate object in the recommended data set can be obtained. Figure 4 The method embodiment shown can obtain the historical feedback features corresponding to each candidate object in the recommendation data set based on the feedback information sequence corresponding to each candidate object in the recommendation data set, and obtain the second predicted feedback index corresponding to each candidate object in the recommendation data set by inputting the historical feedback features corresponding to each candidate object in the recommendation data set into the feedback index prediction model for feedback index prediction. Then, the second display opportunity measurement index corresponding to each candidate object in the recommendation data set can be obtained by the ratio of the second predicted feedback index corresponding to each candidate object in the recommendation data set to the number of feedbacks corresponding to the corresponding candidate object in the recommendation data set.

[0233] (5) According to the second display opportunity measurement indicator and click-through rate corresponding to each candidate object in the recommendation data set, based on the regulation rules in the regulation strategy, determine the second target candidate object from the candidate objects corresponding to the recommendation data set.

[0234] Specifically, the click rate corresponding to each candidate object in the recommended data set can be obtained according to the feedback information of each recommended data in the recommended data set, and then combined with the control rules in the current control strategy, the second target candidate object can be determined from the candidate objects corresponding to the recommended data set.

[0235] (6) Based on the second target candidate object, select target remaining non-recommended data from the remaining non-recommended data, mark the target remaining non-recommended data as recommended data, and update the recommended data set corresponding to the control strategy.

[0236] Specifically, the remaining unrecommended data whose candidate object is the second target candidate object can be found from the remaining unrecommended data, and then the top remaining unrecommended data whose candidate object is the second target candidate object is determined as the target remaining unrecommended data, and the target remaining unrecommended data is marked as recommended data, and the recommended data is added to the recommended data set corresponding to the current control strategy so that the recommended data set corresponding to the control strategy is updated. The remaining unrecommended data refers to the data that is not marked as recommended data among the multiple unrecommended data.

[0237] It should be noted that the recommended data in the recommended data set corresponding to the control strategy are sorted according to the time when they are updated to the recommended data set. Therefore, the aforementioned steps (4) to (6) can be repeatedly executed for the updated recommended data set corresponding to the control strategy until the intervention is stopped when the remaining non-recommended data is empty, so that the recommended data set corresponding to the current control strategy can be obtained. The order of the recommended data in the recommended data set is different from the order of the original recommendation log provided by the recommendation system.

[0238] The disclosed embodiment conducts offline intervention experiments according to each control strategy, thereby obtaining a recommendation data set corresponding to each control strategy. The ordering of the recommendation data in the recommendation data set is different from the ordering of the original recommendation log provided by the recommendation system, thereby ensuring the accuracy of the target adjustment strategy determination during the offline simulation process.

[0239] It should be noted that the object (including sample objects and candidate objects) in the embodiments of the present disclosure may indicate one or multiple objects of the same category. When the object indicates multiple objects of the same category, the processing result corresponding to the object mentioned in the embodiments of the present disclosure may be the average of the processing results of the multiple objects in the same category.

[0240] Figure 7 is a block diagram of a feedback indicator prediction model training device according to an exemplary embodiment. Figure 7 The feedback index prediction model training device 700 includes a random delivery information acquisition unit 710, a target feedback index determination unit 720, a first historical feedback feature determination unit 730 and a training unit 740, wherein:

[0241] The random delivery information acquisition unit 710 is configured to acquire random delivery information of the sample object obtained based on the random delivery experiment; the random delivery information represents feedback of the randomly delivered user account on the sample object;

[0242] The target feedback indicator determination unit 720 is configured to determine the target feedback indicator corresponding to the sample object according to the random delivery information corresponding to the sample object;

[0243] A first historical feedback feature determining unit 730 is configured to determine a historical feedback feature corresponding to the sample object according to a first historical feedback information sequence of the sample object in the historical recommendation;

[0244] The training unit 740 is configured to execute a preset machine learning model according to the historical feedback features corresponding to the sample object and the target feedback index to obtain a feedback index prediction model; wherein the target feedback index serves as the expected prediction value corresponding to the historical feedback features.

[0245] In an exemplary embodiment, the first historical feedback feature determination unit 730 includes:

[0246] A first acquisition unit is configured to execute acquisition of a first historical feedback information sequence of the sample object in the historical recommendation;

[0247] A subsequence division unit is configured to divide the first historical feedback information sequence into a first number of subsequences according to a preset aggregation number; each of the subsequences includes consecutive first historical feedback information of the preset aggregation number;

[0248] an aggregation processing unit, configured to perform aggregation processing on the first historical feedback information in each subsequence to obtain an aggregate feedback information sequence corresponding to the sample object; the aggregate feedback information sequence includes aggregate feedback information corresponding to each of the subsequences;

[0249] The historical feedback feature determination subunit is configured to determine the historical feedback feature corresponding to the sample object according to the aggregate feedback information sequence.

[0250] In an exemplary embodiment, the historical feedback feature determination subunit includes:

[0251] A second quantity determining unit is configured to determine a second quantity according to the first quantity and a preset dimension;

[0252] The historical feedback feature extraction unit is configured to extract the second number of historical feedback features from the aggregate feedback information sequence according to the preset dimension; each of the historical feedback features includes continuous aggregate feedback information of the preset dimension.

[0253] In an exemplary embodiment, the apparatus 700 further includes:

[0254] The random sampling unit is configured to randomly select a first preset number of new objects from the new object pool corresponding to the sampling period for each sampling period in the plurality of sampling periods, and put the first preset number of new objects into the sample object pool;

[0255] A random delivery unit is configured to execute, for each sample object in the delivery sample object pool, randomly delivering the sample object to a second preset number of user accounts in the entire user account space in each delivery cycle, and generating random delivery information of the sample object corresponding to the delivery cycle according to feedback information of the second preset number of user accounts;

[0256] A sample object removing unit is configured to determine the random delivery times of each sample object in the delivery sample object pool, and remove the sample objects whose random delivery times reach a preset times threshold from the delivery sample object pool;

[0257] The random delivery information generating unit is configured to generate the random delivery information of the sample object according to the random delivery information of each delivery period corresponding to the removed sample object.

[0258] In an exemplary embodiment, the random delivery information of the sample objects includes feedback information of at least one dimension; the target feedback index determination unit 720 includes:

[0259] A dimension feedback indicator determination unit is configured to obtain a dimension feedback indicator corresponding to each dimension according to the feedback information of each dimension;

[0260] A weight coefficient determination unit, configured to determine a weight coefficient corresponding to each dimension;

[0261] The weighted summation unit is configured to perform weighted summation according to the dimensional feedback indicators and weight coefficients corresponding to each dimension to obtain the target feedback indicators corresponding to the sample object.

[0262] Figure 8 is a block diagram of an object recommendation device according to an exemplary embodiment. Figure 8 The object recommendation device 800 includes a second acquisition unit 810, a second historical feedback feature determination unit 820, a first feedback index prediction unit 830, a first display opportunity measurement index determination unit 840 and a first target candidate object selection unit 850, wherein:

[0263] The second acquisition unit 810 is configured to execute acquisition of the number of feedbacks corresponding to the candidate object and the second historical feedback information sequence corresponding to the candidate object;

[0264] A second historical feedback feature determining unit 820 is configured to determine the historical feedback feature corresponding to the candidate object according to the second historical feedback information sequence corresponding to the candidate object;

[0265] The first feedback index prediction unit 830 is configured to perform feedback index prediction by inputting the historical feedback features corresponding to the candidate object into the feedback index prediction model to obtain a first predicted feedback index corresponding to the candidate object;

[0266] A first display opportunity measurement index determining unit 840 is configured to determine a first display opportunity measurement index corresponding to the candidate object according to the first predicted feedback index corresponding to the candidate object and the feedback quantity;

[0267] The first target candidate object selection unit 850 is configured to select a first target candidate object from the candidate object pool according to the first fairness value score corresponding to the candidate object; the first target candidate object is used as the object to be recommended;

[0268] The feedback indicator prediction model is trained according to any one of the feedback indicator prediction model training methods provided in the embodiments of the present disclosure.

[0269] In an exemplary embodiment, the first target candidate object selection unit 850 includes:

[0270] a relevance determination unit configured to obtain an original relevance between each candidate object and a target user account; the target user account refers to a user account to be recommended;

[0271] The selection subunit is configured to select a first target candidate object from the candidate object pool based on the target regulation rules in the target regulation strategy according to the original relevance degree and the first display opportunity measurement indicator corresponding to each of the candidate objects.

[0272] In an exemplary embodiment, the apparatus 800 further includes:

[0273] The execution unit is configured to execute the step of obtaining the second historical feedback information sequence corresponding to the candidate object when the target starting intervention time information in the target regulation strategy is met at the current moment.

[0274] In an exemplary embodiment, the apparatus 800 further includes:

[0275] A control strategy determination unit is configured to execute and determine a plurality of different control strategies, each of which includes start intervention time information and control rules;

[0276] A first determination unit is configured to execute, according to the recommendation data set corresponding to each of the regulation strategies, to determine the system display opportunity measurement index and the system average click rate corresponding to each of the regulation strategies; the recommendation data set corresponding to the regulation strategy is determined according to a plurality of recommendation logs sorted by time, each of the recommendation logs includes a recommendation record and corresponding feedback information, and the recommendation record includes a user account and a candidate object recommended to the user account;

[0277] The target regulation strategy determination unit is configured to determine the target regulation strategy among the multiple regulation strategies according to the system display opportunity measurement index and the system average click rate corresponding to each regulation strategy among the multiple regulation strategies.

[0278] In an exemplary embodiment, the first determining unit includes:

[0279] The system display opportunity measurement index determination unit is configured to execute, for each of the control strategies, according to the second display opportunity measurement index corresponding to each candidate object in the recommendation data set corresponding to the control strategy, determine the system display opportunity measurement index corresponding to the control strategy; the second display opportunity measurement index corresponding to each candidate object in the recommendation data set is determined according to the feedback index prediction model and the feedback information of each recommendation data in the recommendation data set corresponding to the control strategy;

[0280] The system average click rate determination unit is configured to determine the system average click rate corresponding to the regulation strategy according to the ratio between the number of recommended data in the recommendation data set corresponding to the regulation strategy and the number of the multiple recommendation logs.

[0281] In an exemplary embodiment, the apparatus 800 further includes:

[0282] A recommendation log acquisition unit, configured to acquire the plurality of recommendation logs sorted by time;

[0283] A first marking unit is configured to mark the plurality of recommended logs as non-recommended to obtain a plurality of non-recommended data;

[0284] A second marking unit is configured to execute, for each of the control strategies, determining target non-recommended data among the multiple non-recommended data according to the starting intervention time information in the control strategy, marking the target non-recommended data as recommended data, and obtaining a recommended data set corresponding to the control strategy;

[0285] A second display opportunity measurement index determining unit is configured to determine a second display opportunity measurement index corresponding to each candidate object in the recommendation data set according to the feedback index prediction model and the feedback information of each recommendation data in the recommendation data set corresponding to the regulation strategy;

[0286] A second target candidate object determination unit is configured to determine a second target candidate object from the candidate objects corresponding to the recommendation data set based on a second display opportunity measurement indicator and a click-through rate corresponding to each candidate object in the recommendation data set and based on a control rule in the control strategy;

[0287] The recommended data set updating unit is configured to select target remaining non-recommended data from the remaining non-recommended data based on the second target candidate object, mark the target remaining non-recommended data as recommended data and update the recommended data set corresponding to the regulation strategy.

[0288] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0289] In an exemplary embodiment, an electronic device is also provided, including a processor; a memory for storing instructions executable by the processor; wherein, when the processor is configured to execute the instructions stored in the memory, it implements any one of the feedback indicator prediction model training methods or object recommendation methods provided in the embodiments of the present disclosure.

[0290] The electronic device may be a terminal, a server or a similar computing device. For example, the electronic device is a server. Fig. 9 is a block diagram of an electronic device according to an exemplary embodiment. Fig. 9 As shown, the server 900 may have relatively large differences due to different configurations or performances, and may include one or more central processing units (CPU) 910 (the processor 910 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 930 for storing data, and one or more storage media 920 (such as one or more mass storage devices) for storing application programs 923 or data 922. Among them, the memory 930 and the storage medium 920 can be short-term storage or permanent storage. The program stored in the storage medium 920 may include one or more modules, each of which may include a series of instruction operations on the server. Furthermore, the central processing unit 910 can be configured to communicate with the storage medium 920 and execute a series of instruction operations in the storage medium 920 on the server 900. The server 900 may also include one or more power supplies 960, one or more wired or wireless network interfaces 950, one or more input and output interfaces 940, and / or one or more operating systems 921, such as Windows ServerTM, Mac OSXTM, UnixTM, LinuxTM, FreeBSDTM, etc.

[0291] The input / output interface 940 may be used to receive or send data via a network. The specific example of the network may include a wireless network provided by a communication provider of the server 900. In one example, the input / output interface 940 includes a network adapter (Network Interface Controller, NIC), which may be connected to other network devices via a base station so as to communicate with the Internet. In one example, the input / output interface 940 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0292] It can be understood by those skilled in the art that Fig. 9 The structure shown is only for illustration and does not limit the structure of the above electronic device. Fig. 9 More or fewer components as shown, or with Fig. 9 Different configurations shown.

[0293] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 930 including instructions, and the instructions can be executed by the processor 910 of the device 900 to complete the above method. Alternatively, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0294] In an exemplary embodiment, a computer program product is also provided, including a computer program, which, when executed by a processor, implements any one of the feedback indicator prediction model training methods or object recommendation methods provided in the embodiments of the present disclosure.

[0295] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0296] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A feedback indicator prediction model training method, characterized in that: include: Obtaining random delivery information of a sample object obtained based on a random delivery experiment; the random delivery information represents feedback from a randomly delivered user account on the sample object, the randomly delivered user account is a user account randomly selected from the entire user account space of an object recommendation system, the object recommended by the object recommendation system includes multimedia information, and the feedback dimension includes one or more of clicks, likes, and comments; Determine the target feedback index corresponding to the sample object according to the feedback of each dimension in the random delivery information corresponding to the sample object; Determine the historical feedback feature corresponding to the sample object according to the first historical feedback information sequence of the sample object in the historical recommendation; the first historical feedback information in the first historical feedback information sequence is the feedback of the sample object within a preset historical time period obtained according to the recommendation log of the object recommendation system, and the historical feedback feature includes an average value of a plurality of consecutive first historical feedback information; Based on the historical feedback features of the sample objects as model input and the target feedback indicators of the sample objects as output targets, a machine learning model is trained to obtain a target feedback indicator prediction model, which is used to predict the target feedback indicators of all subsequent objects in the system; wherein the target feedback indicator prediction model is trained using the corresponding data of the sample objects, and is used to predict the target feedback indicators of all subsequent candidate objects in the object recommendation system, so as to use the target feedback indicators of all candidate objects to determine the display opportunity measurement indicator of the object, and the display opportunity measurement indicator is used to measure the subsequent display opportunities of all candidate objects in the system.

2. The feedback indicator prediction model training method according to claim 1, characterized in that: The determining, according to the first historical feedback information sequence of the sample object in the historical recommendation, the historical feedback feature corresponding to the sample object comprises: Acquire a first historical feedback information sequence of the sample object in historical recommendations; Dividing the first historical feedback information sequence into a first number of subsequences according to a preset aggregation number; each of the subsequences includes consecutive first historical feedback information of the preset aggregation number; Aggregating the first historical feedback information in each subsequence to obtain an aggregate feedback information sequence corresponding to the sample object; the aggregate feedback information sequence includes aggregate feedback information corresponding to each of the subsequences; According to the aggregated feedback information sequence, a historical feedback feature corresponding to the sample object is determined.

3. The feedback indicator prediction model training method according to claim 2, characterized in that: The determining, according to the aggregated feedback information sequence, the historical feedback feature corresponding to the sample object comprises: Determining a second quantity according to the first quantity and a preset dimension; The second number of historical feedback features are extracted from the aggregate feedback information sequence according to the preset dimension; each of the historical feedback features includes consecutive aggregate feedback information of the preset dimension.

4. The feedback indicator prediction model training method according to any one of claims 1 to 3, characterized in that: The method further comprises: For each sampling period in the plurality of sampling periods, randomly selecting a first preset number of new objects from the new object pool corresponding to the sampling period, and placing the first preset number of new objects into the sample object pool; For each sample object in the sample object pool, randomly delivering the sample object to a second preset number of user accounts in the entire user account space in each delivery cycle, and generating random delivery information of the sample object corresponding to the delivery cycle according to feedback information of the second preset number of user accounts; Determine the random delivery times of each sample object in the delivery sample object pool, and remove the sample objects whose random delivery times reach a preset times threshold from the delivery sample object pool; The random delivery information of the sample object is generated according to the random delivery information of each delivery period corresponding to the removed sample object.

5. The feedback indicator prediction model training method according to any one of claims 1 to 3, characterized in that: The random placement information of the sample objects includes feedback information of at least one dimension; The determining, according to the random delivery information corresponding to the sample object, the target feedback indicator corresponding to the sample object comprises: According to the feedback information of each dimension, a dimension feedback indicator corresponding to each dimension is obtained; Determine a weight coefficient corresponding to each dimension; A weighted sum is performed according to the dimension feedback index and the weight coefficient corresponding to each dimension to obtain the target feedback index corresponding to the sample object.

6. An object recommendation method, characterized in that: include: Acquire the number of feedbacks corresponding to the candidate object and the second historical feedback information sequence corresponding to the candidate object; Determining, according to the second historical feedback information sequence corresponding to the candidate object, a historical feedback feature corresponding to the candidate object; Inputting the historical feedback features corresponding to the candidate object into a feedback index prediction model to perform feedback index prediction, thereby obtaining a first predicted feedback index corresponding to the candidate object; Determining a first display opportunity measurement indicator corresponding to the candidate object according to the first predicted feedback indicator corresponding to the candidate object and the feedback quantity; According to the first display opportunity measurement indicator corresponding to the candidate object, a first target candidate object is selected from the candidate object pool; the first target candidate object is used as the object to be recommended; The feedback indicator prediction model is trained according to the feedback indicator prediction model training method according to any one of claims 1 to 5.

7. The object recommendation method according to claim 6, characterized in that: The selecting a first target candidate object from the candidate object pool according to the first display opportunity measurement indicator corresponding to each candidate object in the candidate object pool includes: Obtaining the original correlation between each candidate object and a target user account; the target user account refers to a user account to be recommended; According to the original relevance degree corresponding to each of the candidate objects and the first display opportunity measurement indicator, a first target candidate object is selected from the candidate object pool based on the target regulation rule in the target regulation strategy.

8. The object recommendation method according to claim 7, characterized in that: Before acquiring the second historical feedback information sequence corresponding to the candidate object, the method further includes: When the target start intervention time information in the target regulation strategy is satisfied at the current moment, the step of acquiring the second historical feedback information sequence corresponding to the candidate object is performed.

9. The object recommendation method according to claim 7 or 8, characterized in that: The method further comprises: Determining a plurality of different control strategies, each of the control strategies comprising start intervention time information and control rules; According to the recommendation data set corresponding to each of the regulation strategies, a system display opportunity measurement index and a system average click rate corresponding to each of the regulation strategies are determined; the recommendation data set corresponding to the regulation strategy is determined according to a plurality of recommendation logs sorted by time, each of the recommendation logs includes a recommendation record and corresponding feedback information, and the recommendation record includes a user account and a candidate object recommended to the user account; A target regulation strategy among the multiple regulation strategies is determined according to a system display opportunity measurement indicator and a system average click rate corresponding to each regulation strategy among the multiple regulation strategies.

10. The object recommendation method according to claim 9, characterized in that: Determining the system display opportunity measurement index and the system average click rate corresponding to each of the regulation strategies according to the recommendation data set corresponding to each of the regulation strategies includes: For each of the control strategies, the system display opportunity measurement index corresponding to the control strategy is determined according to the second display opportunity measurement index corresponding to each candidate object in the recommendation data set corresponding to the control strategy; the second display opportunity measurement index corresponding to each candidate object in the recommendation data set is determined according to the feedback indicator prediction model and the feedback information of each recommendation data in the recommendation data set corresponding to the control strategy; The system average click rate corresponding to the regulation strategy is determined according to the ratio between the number of recommended data in the recommendation data set corresponding to the regulation strategy and the number of the plurality of recommendation logs.

11. The object recommendation method according to claim 10, characterized in that: The method further comprises: Obtaining the plurality of recommendation logs sorted by time; Marking the plurality of recommended logs as not recommended, thereby obtaining a plurality of not recommended data; For each of the control strategies, according to the starting intervention time information in the control strategy, target non-recommended data among the multiple non-recommended data are determined, and the target non-recommended data are marked as recommended data, so as to obtain a recommended data set corresponding to the control strategy; Determine a second display opportunity measurement indicator corresponding to each candidate object in the recommendation data set according to the feedback indicator prediction model and the feedback information of each recommendation data in the recommendation data set corresponding to the regulation strategy; Determine a second target candidate object from the candidate objects corresponding to the recommendation data set based on the second display opportunity measurement indicator and click-through rate corresponding to each candidate object in the recommendation data set and based on the regulation rules in the regulation strategy; Based on the second target candidate object, target remaining non-recommended data is selected from the remaining non-recommended data, the target remaining non-recommended data is marked as recommended data, and the recommended data set corresponding to the regulation strategy is updated.

12. A feedback indicator prediction model training device, characterized in that: include: A random placement information acquisition unit, configured to acquire random placement information of sample objects obtained based on the random placement experiment; The random delivery information represents feedback from a randomly delivered user account on the sample object, the randomly delivered user account is a user account randomly selected from the entire user account space of the object recommendation system, the object recommended by the object recommendation system includes multimedia information, and the feedback dimension includes one or more of clicks, likes, and comments; A target feedback indicator determination unit is configured to determine a target feedback indicator corresponding to the sample object according to feedback of each dimension in the random delivery information corresponding to the sample object; A first historical feedback feature determination unit is configured to determine a historical feedback feature corresponding to the sample object according to a first historical feedback information sequence of the sample object in historical recommendations; the first historical feedback information in the first historical feedback information sequence is the feedback of the sample object within a preset historical time period obtained according to the recommendation log of the object recommendation system, and the historical feedback feature includes an average value of a plurality of consecutive first historical feedback information; The training unit is configured to execute a machine learning model based on the historical feedback characteristics of the sample object as a model input and the target feedback index of the sample object as an output target to obtain a target feedback index prediction model, and the model is used to predict the target feedback index of all subsequent objects in the system; wherein the target feedback index prediction model is trained using the corresponding data of the sample object, and is used to predict the target feedback index of all subsequent candidate objects in the object recommendation system, so as to use the target feedback indicators of all candidate objects to determine the display opportunity measurement indicator of the object, and the display opportunity measurement indicator is used to measure the subsequent display opportunities of all candidate objects in the system.

13. The feedback indicator prediction model training device according to claim 12, characterized in that: The first historical feedback feature determination unit includes: A first acquisition unit is configured to execute acquisition of a first historical feedback information sequence of the sample object in the historical recommendation; A subsequence division unit is configured to divide the first historical feedback information sequence into a first number of subsequences according to a preset aggregation number; each of the subsequences includes consecutive first historical feedback information of the preset aggregation number; an aggregation processing unit, configured to perform aggregation processing on the first historical feedback information in each subsequence to obtain an aggregate feedback information sequence corresponding to the sample object; the aggregate feedback information sequence includes aggregate feedback information corresponding to each of the subsequences; The historical feedback feature determination subunit is configured to determine the historical feedback feature corresponding to the sample object according to the aggregate feedback information sequence.

14. The feedback indicator prediction model training device according to claim 13, characterized in that: The historical feedback feature determination subunit includes: A second quantity determining unit is configured to determine a second quantity according to the first quantity and a preset dimension; The historical feedback feature extraction unit is configured to extract the second number of historical feedback features from the aggregate feedback information sequence according to the preset dimension; each of the historical feedback features includes continuous aggregate feedback information of the preset dimension.

15. The feedback indicator prediction model training device according to any one of claims 12 to 14, characterized in that: The device also includes: The random sampling unit is configured to randomly select a first preset number of new objects from the new object pool corresponding to the sampling period for each sampling period in the plurality of sampling periods, and put the first preset number of new objects into the sample object pool; A random delivery unit is configured to execute, for each sample object in the delivery sample object pool, randomly delivering the sample object to a second preset number of user accounts in the entire user account space in each delivery cycle, and generating random delivery information of the sample object corresponding to the delivery cycle according to feedback information of the second preset number of user accounts; A sample object removing unit is configured to determine the random delivery times of each sample object in the delivery sample object pool, and remove the sample objects whose random delivery times reach a preset times threshold from the delivery sample object pool; The random delivery information generating unit is configured to generate the random delivery information of the sample object according to the random delivery information of each delivery period corresponding to the removed sample object.

16. The feedback indicator prediction model training device according to any one of claims 12 to 14, characterized in that: The random placement information of the sample objects includes feedback information of at least one dimension; The target feedback indicator determination unit includes: A dimension feedback indicator determination unit is configured to obtain a dimension feedback indicator corresponding to each dimension according to the feedback information of each dimension; A weight coefficient determination unit, configured to determine a weight coefficient corresponding to each dimension; The weighted summation unit is configured to perform weighted summation according to the dimensional feedback indicators and weight coefficients corresponding to each dimension to obtain the target feedback indicators corresponding to the sample object.

17. An object recommendation device, characterized in that: include: A second acquisition unit is configured to execute acquisition of the number of feedbacks corresponding to the candidate object and a second historical feedback information sequence corresponding to the candidate object; A second historical feedback feature determining unit is configured to determine the historical feedback feature corresponding to the candidate object according to the second historical feedback information sequence corresponding to the candidate object; A first feedback index prediction unit is configured to input the historical feedback features corresponding to the candidate object into a feedback index prediction model to perform feedback index prediction, and obtain a first predicted feedback index corresponding to the candidate object; A first display opportunity measurement index determining unit is configured to determine a first display opportunity measurement index corresponding to the candidate object according to the first predicted feedback index corresponding to the candidate object and the feedback quantity; A first target candidate object selection unit is configured to select a first target candidate object from a candidate object pool according to a first display opportunity measurement indicator corresponding to the candidate object; the first target candidate object is used as an object to be recommended; The feedback indicator prediction model is trained according to the feedback indicator prediction model training method according to any one of claims 1 to 5.

18. The object recommendation device according to claim 17, characterized in that: The first target candidate object selection unit includes: a relevance determination unit configured to obtain an original relevance between each candidate object and a target user account; the target user account refers to a user account to be recommended; The selection subunit is configured to select a first target candidate object from the candidate object pool based on the target regulation rules in the target regulation strategy according to the original relevance degree and the first display opportunity measurement indicator corresponding to each of the candidate objects.

19. The object recommendation device according to claim 18, characterized in that: The device also includes: The execution unit is configured to execute the step of obtaining the second historical feedback information sequence corresponding to the candidate object when the target starting intervention time information in the target regulation strategy is met at the current moment.

20. The object recommendation device according to claim 18 or 19, characterized in that: The device also includes: A control strategy determination unit is configured to execute and determine a plurality of different control strategies, each of which includes start intervention time information and control rules; A first determination unit is configured to execute, according to the recommendation data set corresponding to each of the regulation strategies, to determine the system display opportunity measurement index and the system average click rate corresponding to each of the regulation strategies; the recommendation data set corresponding to the regulation strategy is determined according to a plurality of recommendation logs sorted by time, each of the recommendation logs includes a recommendation record and corresponding feedback information, and the recommendation record includes a user account and a candidate object recommended to the user account; The target regulation strategy determination unit is configured to determine the target regulation strategy among the multiple regulation strategies according to the system display opportunity measurement index and the system average click rate corresponding to each regulation strategy among the multiple regulation strategies.

21. The object recommendation device according to claim 20, characterized in that: The first determining unit includes: The system display opportunity measurement index determination unit is configured to execute, for each of the control strategies, according to the second display opportunity measurement index corresponding to each candidate object in the recommendation data set corresponding to the control strategy, determine the system display opportunity measurement index corresponding to the control strategy; the second display opportunity measurement index corresponding to each candidate object in the recommendation data set is determined according to the feedback index prediction model and the feedback information of each recommendation data in the recommendation data set corresponding to the control strategy; The system average click rate determination unit is configured to determine the system average click rate corresponding to the regulation strategy according to the ratio between the number of recommended data in the recommendation data set corresponding to the regulation strategy and the number of the multiple recommendation logs.

22. The object recommendation device according to claim 21, characterized in that: The device also includes: A recommendation log acquisition unit, configured to acquire the plurality of recommendation logs sorted by time; A first marking unit is configured to mark the plurality of recommended logs as non-recommended to obtain a plurality of non-recommended data; A second marking unit is configured to execute, for each of the control strategies, determining target non-recommended data among the multiple non-recommended data according to the starting intervention time information in the control strategy, marking the target non-recommended data as recommended data, and obtaining a recommended data set corresponding to the control strategy; A second display opportunity measurement index determining unit is configured to determine a second display opportunity measurement index corresponding to each candidate object in the recommendation data set according to the feedback index prediction model and the feedback information of each recommendation data in the recommendation data set corresponding to the regulation strategy; A second target candidate object determination unit is configured to determine a second target candidate object from the candidate objects corresponding to the recommendation data set based on a second display opportunity measurement indicator and a click-through rate corresponding to each candidate object in the recommendation data set and based on a control rule in the control strategy; The recommended data set updating unit is configured to select target remaining non-recommended data from the remaining non-recommended data based on the second target candidate object, mark the target remaining non-recommended data as recommended data and update the recommended data set corresponding to the regulation strategy.

23. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the feedback indicator prediction model training method according to any one of claims 1 to 5, or the object recommendation method according to any one of claims 6 to 11.

24. A computer-readable storage medium, when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to execute the feedback indicator prediction model training method described in any one of claims 1 to 5, or the object recommendation method described in any one of claims 6 to 11.

25. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the feedback indicator prediction model training method described in any one of claims 1 to 5 or the object recommendation method described in any one of claims 6 to 11 is implemented.

Images