Prediction method and device for content, training method and device for prediction model

Abstract

The present disclosure relates to a prediction method and device for content, and a training method and device for a prediction model. The prediction method for content includes: obtaining feature data of candidate recommended content for a candidate object; wherein the feature data includes content features corresponding to the candidate recommended content and object features corresponding to the candidate object; inputting the feature data into a trained prediction model, and obtaining a first conversion value and a second conversion value through the prediction model; wherein the first conversion value reflects the possibility that the candidate recommended content is triggered by the candidate object and converted, and the second conversion value reflects the possibility that the candidate recommended content is not triggered by the candidate object and converted; and determining predicted conversion information of the candidate recommended content corresponding to the candidate object based on the first conversion value and the second conversion value.

Description

Technical Field

[0001] The present disclosure relates to the technical field of information recommendation, and in particular to a method and device for predicting content, and a method and device for training a prediction model. Background Art

[0002] In the field of information recommendation, recommending content that better meets users' needs can improve recommendation efficiency, reduce interference to users, and help promote results. To this end, some prediction models have been proposed in the prior art to predict the conversion of certain recommended content after it is recommended to users.

[0003] Since users often need to trigger (for example, click) recommended content before they can make a conversion, a two-stage approach is often used in related technologies to predict the conversion of recommended content. Specifically, it is to predict the possibility of the recommended content being triggered by the user and the possibility of conversion after being triggered, and then combine the two to predict the possibility of conversion of the recommended content. However, the conversion data of recommended content is usually counted by the provider of the recommended content, rather than by the platform that displays the recommended content. Therefore, there is a possibility that the provider mistakenly attributes the conversion data that occurred on other platforms to this platform. At this time, since the platform has counted that the recommended content has not been triggered, it will count the recommended content as unconverted, resulting in the loss of a large number of positive samples, affecting the prediction accuracy. Summary of the invention

[0004] The present disclosure provides a content prediction method and device, and a prediction model training method and device to at least solve the problem of low prediction accuracy in the related art, and may not solve any of the above problems.

[0005] According to a first aspect of the present disclosure, a prediction method for content is provided, comprising: obtaining feature data of candidate recommended content for a candidate object; wherein the feature data comprises content features corresponding to the candidate recommended content and object features corresponding to the candidate object; inputting the feature data into a trained prediction model, and obtaining a first conversion value and a second conversion value through the prediction model; wherein the first conversion value reflects the possibility that the candidate recommended content is triggered by the candidate object and converted, and the second conversion value reflects the possibility that the candidate recommended content is not triggered by the candidate object and converted; and determining predicted conversion information of the candidate recommended content corresponding to the candidate object based on the first conversion value and the second conversion value.

[0006] Optionally, the prediction model includes a first prediction network and a second prediction network; wherein, inputting the feature data into the trained prediction model and obtaining the first conversion value and the second conversion value through the prediction model includes: inputting the feature data into the first prediction network and obtaining a prediction trigger value through the first prediction network; inputting the feature data into the second prediction network and obtaining a conversion parameter through the second prediction network, wherein the conversion parameter reflects the possibility of conversion of the candidate recommended content under different triggering situations; and determining the first conversion value and the second conversion value according to the prediction trigger value and the conversion parameter.

[0007] Optionally, the conversion parameters include a predicted triggered conversion value and a predicted untriggered conversion value, the predicted triggered conversion value reflecting the possibility of the candidate recommended content being converted when triggered by the candidate object, and the predicted untriggered conversion value reflecting the possibility of the candidate recommended content being converted when not triggered by the candidate object.

[0008] Optionally, the second prediction network includes a first subnetwork and a second subnetwork; wherein, inputting the feature data into the second prediction network and obtaining the conversion parameter through the second prediction network includes: inputting the feature data into the first subnetwork and obtaining the predicted triggered conversion value through the first subnetwork; inputting the feature data into the second subnetwork and obtaining the predicted untriggered conversion value through the second subnetwork.

[0009] Optionally, determining the first conversion value and the second conversion value according to the predicted trigger value and the conversion parameters includes: determining the product of the predicted trigger value and the predicted trigger conversion value as the first conversion value; determining a predicted no-trigger value according to the predicted trigger value; and determining the product of the predicted no-trigger value and the predicted no-trigger conversion value as the second conversion value.

[0010] Optionally, the first conversion value reflects the probability that the candidate recommended content is triggered by the candidate object and converted, and the second conversion value reflects the probability that the candidate recommended content is not triggered by the candidate object and converted; wherein, determining the predicted conversion information based on the first conversion value and the second conversion value includes: determining the sum of the first conversion value and the second conversion value, and determining the predicted conversion information based on the sum.

[0011] Optionally, the feature data further includes a cross feature, and the cross feature reflects the association between the candidate object and the candidate recommended content.

[0012] According to a second aspect of the present disclosure, a training method for a prediction model is provided, the prediction model comprising a first prediction network and a second prediction network, the training method comprising: obtaining feature data and sample tags of a recommended sample; wherein the recommended sample comprises sample recommended content and a sample object targeted by the sample recommended content, and the feature data comprises content features corresponding to the sample recommended content and object features corresponding to the sample object; if it is determined that the sample tag is a trigger tag, inputting the feature data into the first prediction network, obtaining a first predicted trigger value through the first prediction network, and training the first prediction network according to the trigger tag and the first predicted trigger value; wherein the trigger tag is used to reflect whether the sample recommended content is triggered by the sample object; if it is determined that the sample tag comprises a conversion tag, inputting the feature data into the first prediction network and the second prediction network, obtaining predicted conversion information of the sample recommended content corresponding to the sample object through the first prediction network and the second prediction network, and training the second prediction network according to the sample tag and the predicted conversion information; wherein the conversion tag is used to reflect whether the sample recommended content undergoes a conversion corresponding to the sample object.

[0013] Optionally, the inputting the feature data into the first prediction network and the second prediction network, and obtaining predicted conversion information of the sample recommended content corresponding to the sample object through the first prediction network and the second prediction network, includes: inputting the feature data into the first prediction network, and obtaining a second predicted trigger value through the first prediction network; inputting the feature data into the second prediction network, and obtaining a predicted triggered conversion value and a predicted untriggered conversion value through the second prediction network, wherein the predicted triggered conversion value reflects the possibility of the sample recommended content being converted when triggered by the sample object, and the predicted untriggered conversion value reflects the possibility of the sample recommended content being converted when not triggered by the sample object; determining a first conversion value and a second conversion value according to the second predicted trigger value, the predicted triggered conversion value, and the predicted untriggered conversion value; wherein the first conversion value reflects the possibility of the sample recommended content being triggered by the sample object and being converted, and the second conversion value reflects the possibility of the sample recommended content being not triggered by the sample object and being converted; determining the predicted conversion information according to the first conversion value and the second conversion value.

[0014] Optionally, the recommended sample with the conversion mark also has the trigger mark, and the training of the second prediction network according to the sample mark and the predicted conversion information includes: training the second prediction network according to the trigger mark, the conversion mark, the first conversion value and the second conversion value.

[0015] Optionally, obtaining the sample tag of the recommended sample includes: obtaining trigger data within a first time period after the sample recommended content is recommended to the sample object, and determining the trigger tag based on the trigger data; wherein the trigger data is data generated when the sample recommended content is triggered by the sample object; obtaining conversion data within a second time period after the sample recommended content is recommended to the sample object, and determining the conversion tag based on the conversion data; wherein the conversion data is data generated when the sample recommended content is converted, and the second time period is greater than the first time period.

[0016] According to a third aspect of the present disclosure, a prediction device for content is provided, comprising: an acquisition unit, configured to: acquire feature data of candidate recommended content for a candidate object; wherein the feature data comprises content features corresponding to the candidate recommended content and object features corresponding to the candidate object; a prediction unit, configured to: input the feature data into a trained prediction model, and obtain a first conversion value and a second conversion value through the prediction model; wherein the first conversion value reflects the possibility that the candidate recommended content is triggered by the candidate object and converted, and the second conversion value reflects the possibility that the candidate recommended content is not triggered by the candidate object and converted; a determination unit, configured to: determine the predicted conversion information of the candidate recommended content corresponding to the candidate object based on the first conversion value and the second conversion value.

[0017] Optionally, the prediction model includes a first prediction network and a second prediction network, and the prediction unit is further configured to: input the feature data into the first prediction network, and obtain a prediction trigger value through the first prediction network; input the feature data into the second prediction network, and obtain a conversion parameter through the second prediction network, wherein the conversion parameter reflects the possibility of conversion of the candidate recommended content under different triggering conditions; and determine the first conversion value and the second conversion value based on the prediction trigger value and the conversion parameter.

[0018] Optionally, the conversion parameters include a predicted triggered conversion value and a predicted untriggered conversion value, the predicted triggered conversion value reflecting the possibility of the candidate recommended content being converted when triggered by the candidate object, and the predicted untriggered conversion value reflecting the possibility of the candidate recommended content being converted when not triggered by the candidate object.

[0019] Optionally, the second prediction network includes a first subnetwork and a second subnetwork, and the prediction unit is further configured to: input the feature data into the first subnetwork, and obtain the predicted triggered conversion value through the first subnetwork; input the feature data into the second subnetwork, and obtain the predicted untriggered conversion value through the second subnetwork.

[0020] Optionally, the prediction unit is further configured to: determine the product of the predicted trigger value and the predicted trigger conversion value as the first conversion value; determine the predicted no-trigger value based on the predicted trigger value; determine the product of the predicted no-trigger value and the predicted no-trigger conversion value as the second conversion value.

[0021] Optionally, the first conversion value reflects the probability that the candidate recommended content is triggered by the candidate object and converted, and the second conversion value reflects the probability that the candidate recommended content is not triggered by the candidate object and converted, wherein the determination unit is further configured to: determine the sum of the first conversion value and the second conversion value, and determine the predicted conversion information based on the sum.

[0022] Optionally, the feature data further includes a cross feature, and the cross feature reflects the association between the candidate object and the candidate recommended content.

[0023] According to a fourth aspect of the present disclosure, a training device for a prediction model is provided, the prediction model comprising a first prediction network and a second prediction network, the training device comprising: an acquisition unit, configured to: acquire feature data and sample tags of a recommended sample; wherein the recommended sample comprises sample recommended content and a sample object targeted by the sample recommended content, and the feature data comprises content features corresponding to the sample recommended content and object features corresponding to the sample object; the first training unit, configured to: if it is determined that the sample tag is a trigger tag, input the feature data into the first prediction network, obtain a first predicted trigger value through the first prediction network, and train the first prediction network according to the trigger tag and the first predicted trigger value; wherein the trigger tag is used to reflect whether the sample recommended content is triggered by the sample object; the second training unit, configured to: if it is determined that the sample tag includes a conversion tag, input the feature data into the first prediction network and the second prediction network, obtain predicted conversion information of the sample recommended content corresponding to the sample object through the first prediction network and the second prediction network, and train the second prediction network according to the sample tag and the predicted conversion information; wherein the conversion tag is used to reflect whether the sample recommended content undergoes a conversion corresponding to the sample object.

[0024] Optionally, the second training unit is further configured to: input the feature data into the first prediction network, and obtain a second predicted trigger value through the first prediction network; input the feature data into the second prediction network, and obtain a predicted triggered conversion value and a predicted untriggered conversion value through the second prediction network, wherein the predicted triggered conversion value reflects the possibility of the sample recommended content being converted when triggered by the sample object, and the predicted untriggered conversion value reflects the possibility of the sample recommended content being converted when not triggered by the sample object; determine a first conversion value and a second conversion value based on the second predicted trigger value, the predicted triggered conversion value, and the predicted untriggered conversion value; wherein the first conversion value reflects the possibility of the sample recommended content being triggered by the sample object and being converted, and the second conversion value reflects the possibility of the sample recommended content being not triggered by the sample object and being converted; determine the predicted conversion information based on the first conversion value and the second conversion value.

[0025] Optionally, the recommended sample with the conversion mark also has the trigger mark, and the second training unit is further configured to: train the second prediction network according to the trigger mark, the conversion mark, the first conversion value and the second conversion value.

[0026] Optionally, the acquisition unit is further configured to: acquire trigger data within a first time period after the sample recommended content is recommended to the sample object, and determine the trigger mark based on the trigger data; wherein the trigger data is data generated when the sample recommended content is triggered by the sample object; acquire conversion data within a second time period after the sample recommended content is recommended to the sample object, and determine the conversion mark based on the conversion data; wherein the conversion data is data generated when the sample recommended content is converted, and the second time period is greater than the first time period.

[0027] According to a fifth aspect of the present disclosure, an electronic device is provided, comprising: at least one processor; and at least one memory storing computer executable instructions, wherein the computer executable instructions, when executed by the at least one processor, cause the at least one processor to execute a content prediction method or a prediction model training method according to the present disclosure.

[0028] According to a sixth aspect of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by at least one processor, the at least one processor is prompted to execute the content prediction method or prediction model training method according to the present disclosure.

[0029] According to a seventh aspect of the present disclosure, a computer program product is provided, comprising computer instructions, which, when executed by at least one processor, implement the content prediction method or prediction model training method according to the present disclosure.

[0030] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0031] According to the content prediction method, content prediction device, prediction model training method and training device of the embodiments of the present disclosure, the prediction model can respectively predict the possibility of the recommended content being triggered and converted and the possibility of not being triggered and converted, and then predict the possibility of the recommended content being converted. It can count the situation of no triggering but conversion, solve the problem of positive sample loss, thereby compensating for the defects of the existing model and helping to improve the prediction accuracy.

[0032] 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

[0033] 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.

[0034] Figure 1 is a flowchart illustrating a prediction method for content according to an exemplary embodiment of the present disclosure.

[0035] Figure 2 Detailed description is given of the structure of a prediction model according to an exemplary embodiment of the present disclosure.

[0036] Figure 3 is a flowchart illustrating a training method of a prediction model according to an exemplary embodiment of the present disclosure.

[0037] Figure 4 is a schematic diagram illustrating a sample build logic according to an exemplary embodiment of the present disclosure.

[0038] Figure 5 is a block diagram illustrating a prediction apparatus for contents according to an exemplary embodiment of the present disclosure.

[0039] Figure 6 is a block diagram illustrating a training apparatus for a prediction model according to an exemplary embodiment of the present disclosure.

[0040] Figure 7 is a block diagram of an electronic device according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0041] 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.

[0042] 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 implementation methods described in the following examples do not represent all implementation methods 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 attached claims.

[0043] It should be noted that the phrase "at least one of the items" in the present disclosure includes three types of parallel situations: "any one of the items", "a combination of any number of the items", and "all of the items". For example, "including at least one of A and B" includes the following three parallel situations: (1) including A; (2) including B; (3) including A and B. Another example is "executing at least one of step 1 and step 2" which means the following three parallel situations: (1) executing step 1; (2) executing step 2; (3) executing step 1 and step 2.

[0044] For post-activation application recommendations, the user's behavior chain is: exposure → trigger (such as click, or issuing a triggering voice command) → conversion, where conversion for e-commerce recommendations can be purchase, and for application recommendations can be installation, opening, registration, activation, next-day opening / payment, etc. after downloading the application.

[0045] For exposure, when a user performs some predetermined operations (such as entering a webpage with recommended content or clicking on a video with recommended content), the client will send an exposure request to the server. In response to the request, the server will filter out suitable recommended content and feedback it to the corresponding client. When the client displays the received recommended content, an exposure is generated.

[0046] The triggering of downloads usually occurs within a few minutes after the display, and generally does not exceed 20 minutes. The actions after downloading are generally not immediate. Users may wait from a few minutes to a few hours, or even a few days after downloading before installing the corresponding application and registering and activating it.

[0047] The recommendation model system estimates the accurate conversion rate of an exposure request (the probability that an exposure request generates a conversion behavior, which refers to the behavior after downloading, such as registration, activation, etc.) based on a series of user behaviors, and estimates the revenue brought to the recommendation platform by different recommended content (such as eCPM, effective cost per mille), and determines which recommended content should be used for this exposure request through eCPM sorting. The model needs to be trained based on sample data obtained online.

[0048] In the related art, two methods are often used to calculate the conversion rate.

[0049] One is a two-stage method, and the reference behavior link is: exposure → trigger → conversion. In other words, it is to determine the probability of exposure being triggered, such as the click-through rate pctr, and the probability of conversion after triggering, such as the click-through conversion rate pcvr, thereby obtaining the conversion rate pscvr = pctr*pcvr. To this end, it is necessary to obtain the trigger data and conversion data corresponding to the exposure to understand whether the exposure is triggered and whether a conversion occurs. Since conversion data is usually counted by the provider of the recommended content rather than the platform that displays the recommended content, there is a possibility that the provider mistakenly attributes the conversion data that occurred on other platforms to this platform, resulting in a situation where there is no trigger but conversion. At this time, the corresponding exposure will be counted as unconverted, resulting in the conversion being attributed to the exposure but not to the trigger, and a large part of the samples are lost. The loss of positive samples caused by the defects of the model itself will cause the prediction accuracy of the trained model to be low.

[0050] The other is a single-stage method, which directly estimates the probability of exposure being converted, that is, the conversion rate pscvr. Although this method solves the problem of missing positive samples due to the inability to attribute triggers, it will bring another problem. The previous delay in triggering positive samples was basically within 20 minutes, but after switching to the single-stage model, the delay time was the same as the conversion mark, basically delayed for more than half a day. According to data analysis, after the delay in triggering positive samples exceeds 1 hour, the AUC (area under the curve, ROC (receiver operating characteristic curve, receiver operating characteristic) curve, reflecting the probability of ranking the positive example score ahead of the negative example when the model is scoring) will decrease by percentage points. Therefore, after switching to the single-stage method, the defects of the model itself will cause the original trigger part to be delayed along with the conversion part, which in turn affects the final prediction accuracy.

[0051] According to the exemplary embodiment of the present disclosure, a prediction strategy for content and a training strategy for the prediction model used are proposed. The prediction model can respectively predict the possibility of the recommended content being triggered and converted and the possibility of not being triggered and converted, and then predict the possibility of the recommended content being converted. It can count the situation of no trigger but conversion, solve the problem of positive sample loss, thereby making up for the defects of the existing model and helping to improve the prediction accuracy. In addition, the prediction model includes a first prediction network for obtaining a predicted trigger value, and a second prediction network for combining the first prediction network to obtain predicted conversion information. When training the model, by obtaining the trigger mark and conversion mark of the recommended sample online, on the one hand, the trigger mark just obtained and the corresponding recommended sample can be used to train the first prediction network in time, which helps to obtain a higher AUC, thereby ensuring the prediction accuracy of the first prediction network; on the other hand, the conversion mark and the corresponding recommended sample can be used to train the second prediction network, which can reduce the risk of positive sample loss and ensure the prediction accuracy of the second prediction network. The combination of these two aspects improves the final prediction accuracy of the prediction model. After testing, after applying the prediction model provided according to the exemplary embodiment of the present disclosure, the expected CPM can be increased by 11%.

[0052] Next, we will refer to Figures 1 to 7 A prediction method and a prediction device for content and a training method and a training device for a prediction model according to an exemplary embodiment of the present disclosure are specifically described. The prediction model includes a first prediction network and a second prediction network.

[0053] Figure 1 is a flowchart illustrating a prediction method for content according to an exemplary embodiment of the present disclosure. Figure 2 is a schematic diagram showing a structure of a prediction model according to an exemplary embodiment of the present disclosure. It should be understood that the prediction method for content according to the exemplary embodiment of the present disclosure can be implemented in a terminal device (as a client) such as a smart phone, a tablet computer, and a personal computer (PC), and can also be implemented in a device such as a server (as a server).

[0054] Reference Figure 1 In step 101, feature data of candidate recommendation content for the candidate object is obtained. When the operation of the candidate object triggers the recommendation request of the client (corresponding to the exposure request mentioned above), in response to the request, the recommendation system (which can be run on the client or the server) will first obtain a number of candidate recommendation content (the present disclosure does not limit the specific acquisition method). According to the prediction method of the exemplary embodiment of the present disclosure, the conversion status of these candidate recommendation content can be predicted respectively, so as to finally select the appropriate candidate recommendation content and display it to the candidate object on the client.

[0055] It should be understood that in order to achieve prediction, only the feature data of the candidate recommendation content for the candidate object can be obtained without obtaining the candidate recommendation content to reduce the amount of data obtained. Of course, the candidate recommendation content id (or recommendation request id) and the candidate object id can be obtained to distinguish different candidate recommendation content and candidate objects.

[0056] Reference Figure 2 , the feature data includes content features corresponding to the candidate recommended content and object features corresponding to the candidate object. Content features include but are not limited to: candidate recommended content ID, candidate recommended content provider name, candidate recommended content provider industry category, candidate recommended content size, candidate recommended content embedding features obtained based on visual algorithms, etc., which can fully reflect the characteristics of the candidate recommended content itself. Object features include but are not limited to the candidate object's ID, age, gender, region, Internet IP, terminal model used, historically triggered recommended content, historically converted recommended content (i.e., recommended content that has been converted in the past), and disliked recommended content. In the case where the recommendation platform is a multimedia platform, object features can also include embedded features obtained based on multimedia resources (such as videos) browsed by the candidate object, etc., which can fully reflect the characteristics of the candidate object to which the recommended content will be displayed. By extracting these features, it is possible to provide a basis for the matching prediction of candidate recommended content and candidate objects, thereby ensuring the reliable prediction. In addition, since the object features are obtained, the prediction model can naturally extract the personalized needs of the candidate objects, so that the model can be applied to any candidate object without having to train the model for each candidate object separately. This improves the universality of the model, improves its adaptability to newly emerging candidate objects, and improves the prediction response rate, reducing the storage pressure caused by storing massive models.

[0057] Optionally, refer to Figure 2 , the feature data also includes cross features, which reflect the association between the candidate object and the candidate recommended content. For example, it can be a feature that reflects the historical behavior of the candidate object on the candidate recommended content, such as the historical trigger features and historical conversion features of the candidate object on the candidate recommended content, which can provide further basis for the matching prediction of the candidate recommended content and the candidate object. Feature data can also include other applicable features extracted by other models, such as the features extracted by the aforementioned recommendation system, to provide richer information.

[0058] Reference Figure 2 For each of the above feature data, its features can be extracted by extracting the embedded features of the corresponding information (such as content information, object information, and cross information). For the other features mentioned above, since they have already been extracted by the model, no secondary extraction is required. Then these features are spliced ​​to obtain the feature data for input into the prediction model.

[0059] In step 102, the feature data is input into the trained prediction model, and the first conversion value and the second conversion value are obtained through the prediction model. The first conversion value reflects the possibility that the candidate recommended content is triggered by the candidate object and converted, and the second conversion value reflects the possibility that the candidate recommended content is not triggered by the candidate object and converted. By predicting the first conversion value and the second conversion value respectively, the situation of no trigger but conversion can be counted, which solves the problem of positive sample loss, thereby making up for the defects of the existing model and helping to improve the prediction accuracy.

[0060] Optionally, the prediction model includes a first prediction network and a second prediction network. Figure 2 , step 102 may specifically include the following three steps.

[0061] The first step is to input the feature data into the first prediction network (reference Figure 2 The predicted trigger value is used to reflect the possibility that a candidate recommendation content is triggered by the targeted candidate object. For example, the predicted trigger value can be in the form of probability or score, that is, it can be scored according to a certain scoring system. It can also be a Boolean value, with 0 representing no triggering and 1 representing triggering. By exclusively using the first prediction network to obtain the predicted trigger value, the first prediction network can be trained in time with the trigger data just obtained during the model training phase, reducing the risk of delay, and helping to obtain a higher AUC, thereby making up for the defects of the existing model, ensuring the prediction accuracy of the first prediction network, and ultimately improving the prediction accuracy of the prediction model.

[0062] In the second step, the feature data is input into the second prediction network, and the conversion parameters are obtained through the second prediction network. The conversion parameters reflect the possibility of conversion of the candidate recommended content under different triggering conditions. The second prediction network can make conversion predictions for different triggering conditions, so it can count the cases of conversion without triggering, solve the problem of positive sample loss, make up for the defects of the existing model, and help improve the prediction accuracy.

[0063] Optionally, the conversion parameters include a predicted triggered conversion value and a predicted untriggered conversion value. The predicted triggered conversion value reflects the possibility of conversion of the candidate recommended content when triggered by the candidate object, and the predicted untriggered conversion value reflects the possibility of conversion of the candidate recommended content when not triggered by the candidate object, which can cover different triggering situations.

[0064] Accordingly, the second prediction network includes the first subnetwork (refer to Figure 2 The dotted box on the left) and the second subnetwork (reference Figure 2The third step specifically includes: inputting the feature data into the first sub-network, and obtaining the predicted triggered conversion value through the first sub-network; inputting the feature data into the second sub-network, and obtaining the predicted untriggered conversion value through the second sub-network. By configuring independent sub-networks to obtain the predicted triggered conversion value and the predicted untriggered conversion value, the two sub-networks can perform their respective functions, reduce mutual interference, and help improve the prediction accuracy.

[0065] The third step is to determine the first conversion value and the second conversion value according to the prediction trigger value and the conversion parameter. The first conversion value and the second conversion value can be obtained by combining the output results of the first prediction network and the second prediction network to achieve data aggregation.

[0066] Optionally, based on the conversion parameters including the above-mentioned predicted trigger conversion value and predicted no-trigger conversion value, the third step may include: determining the product of the predicted trigger value and the predicted trigger conversion value as the first conversion value; determining the predicted no-trigger value according to the predicted trigger value; determining the product of the predicted no-trigger value and the predicted no-trigger conversion value as the second conversion value. The predicted trigger value reflects the possibility that the candidate recommended content is triggered by the candidate object, and the predicted trigger conversion value reflects the possibility that the candidate recommended content is converted when it is triggered by the candidate object. Combining the two, it can reflect the possibility that the candidate recommended content is triggered by the candidate object and converted, and obtain the first conversion value. The predicted no-trigger value determined by the predicted trigger value reflects the possibility that the candidate recommended content is not triggered by the candidate object, and the predicted no-trigger conversion value reflects the possibility that the candidate recommended content is converted when it is not triggered by the candidate object. Combining the two, it can reflect the possibility that the candidate recommended content is not triggered by the candidate object and converted, and obtain the second conversion value. By clarifying the corresponding relationship, the first conversion value and the second conversion value obtained can be guaranteed to be reliable. Taking the probability form and the triggering operation being click as an example, the product of the predicted click-through rate pctr and the predicted click-through conversion rate pcvr1 can be determined as the first conversion rate pscvr1; based on the second predicted click-through rate pctr, the predicted no-click rate 1-pctr can be determined; the product of the predicted no-click rate 1-pctr and the predicted no-click conversion rate pcvr0 can be determined as the second conversion rate pscvr0.

[0067] In step 103, predicted conversion information of the candidate recommended content corresponding to the candidate object is determined according to the first conversion value and the second conversion value. By finally summarizing the first conversion value and the second conversion value, the predicted conversion information can be obtained, and the conversion status of the candidate recommended content can be effectively predicted.

[0068] In some embodiments, a predicted conversion value may be first determined according to the first conversion value and the second conversion value, and then predicted conversion information may be determined based on the predicted conversion value.

[0069] The predicted conversion value is used to reflect the possibility of the candidate object to undergo conversion behavior based on the corresponding candidate recommended content. Similar to the predicted trigger value, the predicted conversion value can also be in the form of probability, score, Boolean value, etc. When the probability form is adopted, the predicted conversion value is, for example, the predicted conversion rate. Optionally, the first conversion value reflects the probability that the candidate recommended content is triggered by the candidate object and converted, and the second conversion value reflects the probability that the candidate recommended content is not triggered by the candidate object and converted. The predicted conversion value is the sum of the first conversion value and the second conversion value, which is easy to calculate and can effectively reflect the conversion of the candidate recommended content. Still taking the probability form and the triggering operation being a click as an example, the sum of the first conversion rate pscvr1 and the second conversion rate pscvr0 can be determined as the predicted conversion rate pscvr, that is, pscvr=pscvr1+pscvr0.

[0070] For the determination of the predicted conversion information, the predicted conversion value can be directly used as the predicted conversion information, or further processed to obtain the predicted conversion information. For example, the predicted conversion value is the predicted conversion rate pscvr, and the predicted conversion information is a predicted Boolean value. A conversion rate threshold can be configured. If the predicted conversion rate pscvr is greater than or equal to the conversion rate threshold, the predicted result is considered to be a conversion, and the predicted Boolean value is 1. If the predicted conversion rate pscvr is less than the conversion rate threshold, the predicted result is considered to be no conversion, and the predicted Boolean value is 0.

[0071] In other embodiments, the first conversion value and the second conversion value may be used together as predicted conversion information without further calculation processing.

[0072] Figure 3 1 is a flowchart illustrating a method for training a prediction model according to an exemplary embodiment of the present disclosure. The prediction model includes a first prediction network and a second prediction network. Figure 4 is a sample construction logic diagram showing an exemplary embodiment of the present disclosure. It should be understood that the training method of the prediction model according to the exemplary embodiment of the present disclosure can be implemented in a terminal device such as a smart phone, a tablet computer, a personal computer (PC), or can also be implemented in a device such as a server.

[0073] Reference Figure 3 In step 301, feature data and sample tags of the recommended samples are obtained. The recommended samples include sample recommendation content and sample objects targeted by the sample recommendation content. When the client of the sample object initiates a recommendation request, the client will display the corresponding sample recommendation content and generate recommendation data representing the display behavior. The recommended samples can be obtained online, that is, during the actual operation of the platform, based on the behavior of the sample object on the sample recommendation content (or the recommended data), sample tags are generated in real time to obtain recommended samples, so as to realize online training of the prediction model.

[0074] It should be understood that in order to implement model training, only the feature data and sample labels of the recommended samples can be obtained without obtaining the sample recommendation content in the recommended samples to reduce the amount of data obtained. Of course, the identification of the recommended samples can be obtained, such as the sample recommendation content id (or recommendation request id) and the targeted sample object id, to distinguish different recommended samples.

[0075] The feature data includes content features corresponding to the sample recommended content and object features corresponding to the sample object. Content features include but are not limited to: sample recommended content ID, sample recommended content provider name, sample recommended content provider industry category, sample recommended content size, sample recommended content embedding features obtained based on visual algorithms, etc., which can fully reflect the characteristics of the sample recommended content itself. Object features include but are not limited to the sample object's ID, age, gender, region, Internet IP, terminal model used, historically triggered recommended content, historically converted recommended content (i.e., recommended content that has historically made conversion behaviors), and disliked recommended content. For the case where the recommendation platform is a multimedia platform, object features can also include embedded features obtained based on multimedia resources (such as videos) browsed by the sample object, etc., which can fully reflect the characteristics of the sample object to which the recommended content will be displayed. By extracting these features, a basis can be provided for the matching prediction of the sample recommended content and the sample object, thereby ensuring the reliable prediction. In addition, since the object features are obtained, the prediction model can naturally extract the personalized needs of the sample object, so that the model can be applied to any sample object without having to train the model separately for each sample object, which improves the universality of the model and helps to reduce the cost of model training.

[0076] Optionally, the feature data also includes cross-features, which reflect the association between the sample object and the sample recommended content. For example, the cross-features may be features that reflect the historical behaviors of the sample object on the sample recommended content, such as the historical trigger features and historical conversion features of the sample object in the sample recommended content, which can provide further basis for the matching prediction of the sample recommended content and the sample object. The feature data may also include other applicable features extracted by other models to provide richer information. For each of the above feature data, its features can be extracted in the form of embedded features of the corresponding information. For the other aforementioned features, since they are already features extracted by the model, no secondary extraction is required. These features are then concatenated to obtain the feature data for the input prediction model.

[0077] Optionally, in step 301, the operation of obtaining the sample mark of the recommended sample includes: obtaining trigger data within the first time period after the sample recommended content is recommended to the sample object, and determining the trigger mark according to the trigger data. The trigger data is the data generated when the sample recommended content is triggered by the sample object, and the trigger mark is used to reflect whether the sample recommended content is triggered by the sample object. The moment when the sample recommended content is recommended is also the moment when the recommended data is generated. By setting aside the first time period and counting the trigger data within the first time period, the delay waiting of the first time period can be provided for the statistics of the triggering behavior. If the trigger data is generated within the first time period, that is, the sample object triggers the sample recommended content, then the trigger mark can be determined to be a positive mark; if the trigger data is not generated within the first time period, then the trigger mark can be determined to be a negative mark. Thus, the trigger mark is obtained. During online training, multiple sample recommended contents may be counted at one time, and the sample recommended contents in which the sample mark is a trigger mark can be summarized and recorded as a recommendation trigger flow. Specifically, the first time period can be, for example, 10 minutes, 20 minutes, 30 minutes, or 40 minutes. Taking 20 minutes as an example, for each piece of recommended data, wait for 20 minutes. If there is a triggering behavior within 20 minutes, trigger data is generated and the corresponding recommended data is marked as a trigger positive sample, otherwise it is marked as a trigger negative sample, thereby constructing the recommended trigger flow sample data with a 20-minute delay wait.

[0078] Optionally, in step 301, the operation of obtaining the sample mark of the recommended sample also includes: obtaining conversion data within the second time period after the sample recommended content is recommended to the sample object, and determining the conversion mark according to the conversion data. Among them, the conversion data is the data generated when the sample recommended content is converted, and the conversion mark is used to reflect whether the sample recommended content is converted, and the second time period is greater than the first time period. The conversion often occurs after the trigger. By setting aside a second time period greater than the first time period and counting the conversion data within the second time period, a delayed wait of the second time period can be provided for the statistics of the conversion behavior. If the conversion data is generated within the second time period, that is, the sample object has made a conversion behavior on the sample recommended content, the conversion mark can be determined as a positive mark; if the conversion data is not generated within the second time period, the conversion mark can be determined as a negative mark. Thus, the conversion mark is obtained. During online training, for the sample recommended content whose sample marks include conversion marks, they can be summarized and recorded as recommendation trigger conversion flow. Furthermore, the sample tag of the recommendation triggered conversion flow may also include a trigger tag, which, combined with the conversion tag, can reflect whether the corresponding sample recommendation content has been triggered and converted, and whether the converted sample recommendation content has been triggered, so as to distinguish whether miscounting of conversion data has occurred. Specifically, the length of the second time period can be adaptively set according to the different recommended content, for example, it can be 6 hours, 8 hours, 12 hours, 24 hours, or 72 hours. Taking 6 hours as an example, for each piece of recommendation data, if there is a conversion behavior within 6 hours, conversion data is generated, and the corresponding recommendation data is marked as a positive conversion sample, otherwise it is marked as a negative conversion sample, thereby constructing the recommendation trigger conversion flow sample data with a 6-hour delayed wait.

[0079] During specific execution, all recommendation data, trigger data and conversion data on the platform can be obtained online. These data are all associated with the recommendation request. Therefore, based on the recommendation request id, the trigger data and conversion data can be spliced ​​onto the recommendation data. If a recommendation data is spliced ​​with trigger data generated within the first time period, it is determined to be a trigger positive sample; if a recommendation data is not spliced ​​with trigger data, but has been generated for more than the first time period, it is determined to be a trigger negative sample. In this way, a recommendation trigger flow is constructed. If a recommendation data is spliced ​​with conversion data generated within the second time period, it is determined to be a conversion positive sample; if a recommendation data is not spliced ​​with conversion data, but has been generated for more than the second time period, it is determined to be a conversion negative sample. In this way, a recommendation trigger conversion flow is constructed. In addition, for the recommendation samples corresponding to the recommendation trigger conversion flow, it can be determined whether the recommendation samples with conversions are triggered based on whether their recommendation data is spliced ​​with trigger data. Refer to Figure 4, merge the two data streams (for example, using a merge sort algorithm) to obtain a mixed stream, and add a field to distinguish which stream the data comes from. The feature data of the mixed stream can be directly input into the prediction model without having to differentiate between the two data streams, which can improve training efficiency. It should be understood that since the mixed stream is obtained online, it is not necessarily the case that each mixed stream obtained contains both the recommendation trigger stream and the recommendation trigger conversion stream. In extreme cases, it may only contain the recommendation trigger stream, or only contain the recommendation trigger conversion stream. In addition, although the mixed stream involves multiple recommendation samples, only the feature data of one recommendation sample is input into the prediction model for each calculation.

[0080] Return to reference Figure 3 In step 302, if the sample mark is determined to be a trigger mark, the feature data is input into the first prediction network, a first prediction trigger value is obtained through the first prediction network, and the first prediction network is trained according to the trigger mark and the first prediction trigger value.

[0081] The sample is marked as a trigger mark, which means that the currently calculated recommendation sample comes from the recommendation trigger stream, and its feature data can be input into the first prediction network (refer to Figure 2 The predicted trigger value is used to reflect the possibility of a sample recommendation content being triggered by the sample object. For example, the predicted trigger value can be in the form of probability, such as the predicted click-through rate pctr, or in the form of a score, that is, scoring according to a certain score system, or in the form of a Boolean value, with 0 representing no trigger and 1 representing trigger.

[0082] By comparing the trigger mark of the recommended sample from the recommended trigger stream with the first predicted trigger value, the corresponding first loss function can be calculated. Based on the first loss function, the parameters of the first prediction network can be adjusted to achieve its training. It should be understood that since the feature data is not input into the second prediction network in step 302, the corresponding network output value will not be generated, and the corresponding loss function will not be obtained, so the second prediction network will not be adjusted, which not only reduces the amount of training calculations, but also can timely use the newly obtained trigger mark and the corresponding recommended sample to train the first prediction network, reduce the risk of delay, and help to obtain a higher AUC, thereby ensuring the prediction accuracy of the first prediction network, and ultimately improving the prediction accuracy of the prediction model.

[0083] Return to reference Figure 3 In step 303, if it is determined that the sample tag includes a conversion tag, the feature data is input into the first prediction network and the second prediction network, and the predicted conversion information of the sample recommendation content corresponding to the sample object is obtained through the first prediction network and the second prediction network. The second prediction network is trained according to the sample tag and the predicted conversion information.

[0084] The sample is marked as a conversion mark, which means that the sample recommendation content currently calculated comes from the recommendation trigger conversion flow, and its feature data can be input into the first prediction network and the second prediction network to obtain the predicted conversion information. When the prediction model obtained by training is applied, the first prediction network and the second prediction network are integrated to obtain the predicted conversion information according to this process. Therefore, the feature data is also input into these two networks at the same time during training, which can ensure that the training is consistent with the later application and ensure the prediction accuracy of the later application model. Step 303 can use the recommendation trigger conversion flow to train the second prediction network, which can reduce the risk of positive sample loss and ensure the prediction accuracy of the second prediction network. At the same time, step 303 only trains the second prediction network, which can ensure that the first prediction network trained by the recommendation trigger flow is not affected, and realizes the independent training of the first prediction network and the second prediction network.

[0085] Optionally, refer to Figure 2 In step 303, the feature data is input into the first prediction network and the second prediction network, and the operation of obtaining the predicted conversion information of the sample recommendation content corresponding to the sample object through the first prediction network and the second prediction network specifically includes four steps.

[0086] In the first step, the feature data is input into the first prediction network, and the second prediction trigger value is obtained through the first prediction network. It should be understood that the second prediction trigger value here is to distinguish it from the first prediction trigger value output when training the first prediction network. The difference between the two lies in which data stream the sample recommendation content used comes from, but the final results are all prediction trigger values. When applying the network, it can be understood that only the second prediction trigger value is calculated, and only one first prediction trigger value is not output, that is, the first prediction trigger value is only used for the training of the second prediction network.

[0087] In the second step, the feature data is input into the second prediction network, and the predicted triggered conversion value and the predicted untriggered conversion value are obtained through the second prediction network. The predicted triggered conversion value reflects the possibility of conversion of the sample recommended content when it is triggered by the sample object, and the predicted untriggered conversion value reflects the possibility of conversion of the sample recommended content when it is not triggered by the sample object. Specifically, the second prediction network includes a first subnetwork (refer to Figure 2 The dotted box on the left) and the second subnetwork (reference Figure 2 The feature data can be input into these two sub-networks at the same time to obtain the predicted triggered conversion value and the predicted untriggered conversion value respectively.

[0088] The third step is to determine the first conversion value and the second conversion value based on the second predicted trigger value, the predicted triggered conversion value and the predicted non-triggered conversion value. The first conversion value reflects the possibility that the sample recommended content is triggered by the sample object and a conversion occurs, and the second conversion value reflects the possibility that the sample recommended content is not triggered by the sample object and a conversion occurs. Taking the probability form and the triggering operation being a click as an example, the product of the second predicted click rate pctr and the predicted click conversion rate pcvr1 can be determined as the first conversion rate pscvr1; based on the second predicted click rate pctr, the predicted non-click rate 1-pctr is determined; the product of the predicted non-click rate 1-pctr and the predicted non-click conversion rate pcvr0 is determined as the second conversion rate pscvr0.

[0089] The fourth step is to determine predicted conversion information based on the first conversion value and the second conversion value.

[0090] In some embodiments, the predicted conversion value may be first determined according to the first conversion value and the second conversion value, and then the predicted conversion information may be determined based on the predicted conversion value. For the determination of the predicted conversion value, still taking the probability form and the trigger operation being click as an example, the sum of the first conversion rate pscvr1 and the second conversion rate pscvr0 may be determined as the predicted conversion rate pscvr, that is, pscvr=pscvr1+pscvr0. For the determination of the predicted conversion information, the predicted conversion value may be directly used as the predicted conversion information, or may be further processed to obtain the predicted conversion information. For example, the predicted conversion value is the predicted conversion rate pscvr, and the predicted conversion information is a predicted Boolean value. A conversion rate threshold may be configured. If the predicted conversion rate pscvr is greater than or equal to the conversion rate threshold, it is considered that the predicted result is a conversion, and the predicted Boolean value is 1. If the predicted conversion rate pscvr is less than the conversion rate threshold, it is considered that the predicted result is no conversion, and the predicted Boolean value is 0. When the model is applied, the conversion information may be predicted based on this as the output value of the prediction model.

[0091] In other embodiments, the first conversion value and the second conversion value may be used together as predicted conversion information without further calculation processing.

[0092] Optionally, the recommended sample with the conversion mark also has a trigger mark. In step 303, the operation of training the second prediction network according to the sample mark and the predicted conversion information includes: training the second prediction network according to the trigger mark, the conversion mark, the first conversion value and the second conversion value. The training operation can specifically be, according to the conversion mark and the trigger mark, determining the recommended trigger conversion mark and the recommended non-trigger conversion mark; training the second prediction network according to the recommended trigger conversion mark, the recommended non-trigger conversion mark, the first conversion value and the second conversion value. As an example, referring to Table 1 and Table 2, when the sample recommendation content is triggered by the sample object and conversion occurs, the recommended trigger conversion mark is positive, otherwise it is negative; when the sample recommendation content is not triggered by the sample object and conversion occurs, the recommended non-trigger conversion mark is positive, otherwise it is negative.

[0093] Table 1 Recommended trigger conversion tags

[0094] Trigger mark is positive Trigger mark is negative Conversion marked as positive just burden Conversion flag negative burden burden

[0095] Table 2 Recommended non-trigger conversion tags

[0096] Trigger mark is positive Trigger mark is negative Conversion marked as positive burden just Conversion flag negative burden burden

[0097] When training the second prediction network, the recommended trigger conversion mark can be compared with the first conversion value to obtain a corresponding second loss function, and then the parameters of the first sub-network in the second prediction network can be adjusted according to the second loss function; the recommended non-trigger conversion mark can be compared with the second conversion value to obtain a corresponding third loss function, and then the parameters of the second sub-network in the second prediction network can be adjusted according to the third loss function. In addition, to avoid adjusting the parameters of the first prediction network, corresponding isolation instructions can be added.

[0098] Figure 5 is a block diagram showing a prediction device for content according to an exemplary embodiment of the present disclosure. It should be understood that the prediction device for content according to the exemplary embodiment of the present disclosure can be implemented in a terminal device such as a smart phone, a tablet computer, a personal computer in the form of software, hardware, or a combination of software and hardware, and can also be implemented in a device such as a server.

[0099] Reference Figure 5 The content prediction device 500 includes an acquisition unit 501, a prediction unit 502, and a determination unit 503.

[0100] The acquisition unit 501 can acquire feature data of candidate recommended content for the candidate object. When the operation of the candidate object triggers the recommendation request of the client (corresponding to the exposure request in the previous text), in response to the request, the recommendation system (which can be run on the client or the server) will first obtain a number of candidate recommended content (the present disclosure does not limit the specific acquisition method). According to the exemplary embodiment of the present disclosure, the prediction device 500 for content can predict the conversion status of these candidate recommended content respectively, so as to finally select the appropriate candidate recommended content and display it to the candidate object on the client.

[0101] It should be understood that in order to achieve prediction, only the feature data of the candidate recommendation content for the candidate object can be obtained without obtaining the candidate recommendation content to reduce the amount of data obtained. Of course, the candidate recommendation content id (or recommendation request id) and the candidate object id can be obtained to distinguish different candidate recommendation content and candidate objects.

[0102] The feature data includes content features corresponding to the candidate recommendation content and object features corresponding to the candidate object. Since the object features are obtained, the prediction model can naturally extract the personalized needs of the candidate object, so that the model can be applied to any candidate object without having to train the model for each candidate object separately, which improves the universality of the model, improves the adaptability to newly emerging candidate objects, and improves the prediction response rate, reducing the storage pressure caused by storing massive models.

[0103] Optionally, the feature data also includes cross features, which reflect the association between the candidate object and the candidate recommended content, and can provide further basis for the matching prediction between the candidate recommended content and the candidate object. The feature data may also include other applicable features extracted by other models to provide richer information. For each of the above feature data, its features can be extracted in the form of embedded features of the corresponding information, and no secondary extraction is required for the other aforementioned features. These features are then concatenated to obtain the feature data for the input prediction model.

[0104] The prediction unit 502 can input the feature data into a trained prediction model, and obtain a first conversion value and a second conversion value through the prediction model. The first conversion value reflects the possibility that the candidate recommended content is triggered by the candidate object and converted, and the second conversion value reflects the possibility that the candidate recommended content is not triggered by the candidate object and converted. By predicting the first conversion value and the second conversion value respectively, the situation of no trigger but conversion can be counted, which solves the problem of positive sample loss, thereby making up for the defects of the existing model and helping to improve the prediction accuracy.

[0105] Optionally, the prediction model includes a first prediction network and a second prediction network. Accordingly, the operation performed by the prediction unit 502 may specifically include the following three steps.

[0106] In the first step, the feature data is input into the first prediction network, and the prediction trigger value is obtained through the first prediction network. The prediction trigger value is used to reflect the possibility that a candidate recommendation content is triggered by the candidate object being targeted. By exclusively using the first prediction network to obtain the prediction trigger value, the first prediction network can be trained in a timely manner using the newly acquired trigger data during the model training phase, reducing the risk of delay and helping to obtain a higher AUC, thereby making up for the defects of the existing model, ensuring the prediction accuracy of the first prediction network, and ultimately improving the prediction accuracy of the prediction model.

[0107] In the second step, the feature data is input into the second prediction network, and the conversion parameters are obtained through the second prediction network. The conversion parameters reflect the possibility of conversion of the candidate recommended content under different triggering conditions. The second prediction network can make conversion predictions for different triggering conditions, so it can count the cases of conversion without triggering, solve the problem of positive sample loss, make up for the defects of the existing model, and help improve the prediction accuracy.

[0108] Optionally, the conversion parameters include a predicted triggered conversion value and a predicted untriggered conversion value. The predicted triggered conversion value reflects the possibility of conversion of the candidate recommended content when triggered by the candidate object, and the predicted untriggered conversion value reflects the possibility of conversion of the candidate recommended content when not triggered by the candidate object, which can cover different triggering situations.

[0109] Accordingly, the second prediction network includes a first sub-network and a second sub-network, and the third step specifically includes: inputting the feature data into the first sub-network, and obtaining the predicted triggered conversion value through the first sub-network; inputting the feature data into the second sub-network, and obtaining the predicted untriggered conversion value through the second sub-network. By configuring independent sub-networks to obtain the predicted triggered conversion value and the predicted untriggered conversion value, the two sub-networks can perform their respective functions, reduce mutual interference, and help improve the prediction accuracy.

[0110] The third step is to determine the first conversion value and the second conversion value according to the prediction trigger value and the conversion parameter. The first conversion value and the second conversion value can be obtained by combining the output results of the first prediction network and the second prediction network to achieve data aggregation.

[0111] Optionally, based on the conversion parameters including the above-mentioned predicted trigger conversion value and predicted no-trigger conversion value, the third step may include: determining the product of the predicted trigger value and the predicted trigger conversion value as the first conversion value; determining the predicted no-trigger value according to the predicted trigger value; determining the product of the predicted no-trigger value and the predicted no-trigger conversion value as the second conversion value. The predicted trigger value reflects the possibility that the candidate recommended content is triggered by the candidate object, and the predicted trigger conversion value reflects the possibility that the candidate recommended content is converted when it is triggered by the candidate object. Combining the two, it can reflect the possibility that the candidate recommended content is triggered by the candidate object and converted, and obtain the first conversion value. The predicted no-trigger value determined by the predicted trigger value reflects the possibility that the candidate recommended content is not triggered by the candidate object, and the predicted no-trigger conversion value reflects the possibility that the candidate recommended content is converted when it is not triggered by the candidate object. Combining the two, it can reflect the possibility that the candidate recommended content is not triggered by the candidate object and converted, and obtain the second conversion value. By clarifying the corresponding relationship, the first conversion value and the second conversion value obtained can be guaranteed to be reliable.

[0112] The determination unit 503 may determine the predicted conversion information of the candidate recommended content corresponding to the candidate object according to the first conversion value and the second conversion value. By finally summarizing the first conversion value and the second conversion value, the predicted conversion information can be obtained to effectively predict the conversion of the candidate recommended content.

[0113] In some embodiments, the predicted conversion value can be first determined based on the first conversion value and the second conversion value, and then the predicted conversion information can be determined based on the predicted conversion value. The predicted conversion value is used to reflect the possibility of the candidate object undergoing conversion behavior based on the corresponding candidate recommended content. Optionally, the first conversion value reflects the probability that the candidate recommended content is triggered by the candidate object and converted, and the second conversion value reflects the probability that the candidate recommended content is not triggered by the candidate object and converted. The predicted conversion value is the sum of the first conversion value and the second conversion value, which is easy to calculate and can effectively reflect the conversion status of the candidate recommended content. For the determination of the predicted conversion information, the predicted conversion value can be directly used as the predicted conversion information, or it can be further processed to obtain the predicted conversion information.

[0114] In other embodiments, the first conversion value and the second conversion value may be used together as predicted conversion information without further calculation processing.

[0115] Figure 6 is a block diagram showing a training device for a prediction model according to an exemplary embodiment of the present disclosure. It should be understood that the training device for a prediction model according to an exemplary embodiment of the present disclosure can be implemented in a terminal device such as a smart phone, a tablet computer, a personal computer in the form of software, hardware, or a combination of software and hardware, and can also be implemented in a device such as a server.

[0116] Reference Figure 6 The prediction model training device 600 includes an acquisition unit 601, a first training unit 602, and a second training unit 603.

[0117] The acquisition unit 601 can obtain the feature data and sample tags of the recommended samples. The recommended samples include the sample recommendation content and the sample object for which the sample recommendation content is targeted. When the client of the sample object initiates a recommendation request, the client will display the corresponding sample recommendation content and generate recommendation data representing this display behavior. The recommended samples can be obtained online, that is, during the actual operation of the platform, based on the behavior of the sample object on the sample recommendation content (or on the recommended data), the sample tags are generated in real time to obtain the recommended samples, so as to realize the online training of the prediction model.

[0118] It should be understood that in order to implement model training, only the feature data and sample labels of the recommended samples can be obtained without obtaining the sample recommendation content in the recommended samples to reduce the amount of data obtained. Of course, the identification of the recommended samples can be obtained, such as the sample recommendation content id (or recommendation request id) and the targeted sample object id, to distinguish different recommended samples.

[0119] The feature data includes content features corresponding to the sample recommendation content and object features corresponding to the sample object. Since the object features are obtained, the prediction model can naturally extract the personalized needs of the sample object, so that the model can be applied to any sample object without having to train the model for each sample object separately, which improves the universality of the model and helps reduce the cost of model training.

[0120] Optionally, the feature data also includes cross features, which reflect the association between the sample object and the sample recommended content, and can provide further basis for the matching prediction between the sample recommended content and the sample object. The feature data may also include other applicable features extracted by other models to provide richer information. For each of the above feature data, its features can be extracted in the form of embedded features of the corresponding information, and no secondary extraction is required for the other aforementioned features. These features are then concatenated to obtain the feature data for the input prediction model.

[0121] Optionally, the operation of obtaining the sample tag of the recommended sample by the acquisition unit 601 includes: obtaining trigger data within the first time period after the sample recommended content is recommended to the sample object, and determining the trigger tag according to the trigger data. Among them, the trigger data is the data generated when the sample recommended content is triggered by the sample object, and the trigger tag is used to reflect whether the sample recommended content is triggered by the sample object. The moment when the sample recommended content is recommended is also the moment when the recommendation data is generated. By setting aside the first time period and counting the trigger data within the first time period, the delay waiting of the first time period can be provided for the statistics of the triggering behavior. If the trigger data is generated within the first time period, that is, the sample object triggers the sample recommended content, then the trigger tag can be determined to be a positive tag; if the trigger data is not generated within the first time period, then the trigger tag can be determined to be a negative tag. Thus, the trigger tag is obtained. During online training, multiple sample recommended contents may be counted at one time. For the sample recommended contents in which the sample tag is a trigger tag, they can be summarized and recorded as a recommendation trigger flow.

[0122] Optionally, the operation of obtaining the sample tag of the recommended sample by the acquisition unit 601 also includes: obtaining conversion data within the second time period after the sample recommended content is recommended to the sample object, and determining the conversion tag according to the conversion data. Among them, the conversion data is the data generated when the sample recommended content is converted, and the conversion tag is used to reflect whether the sample recommended content is converted, and the second time period is greater than the first time period. The conversion often occurs after the trigger. By setting aside a second time period greater than the first time period and counting the conversion data within the second time period, a delayed waiting time of the second time period can be provided for the statistics of the conversion behavior. If the conversion data is generated within the second time period, that is, the sample object has made a conversion behavior on the sample recommended content, the conversion tag can be determined as a positive tag; if the conversion data is not generated within the second time period, the conversion tag can be determined as a negative tag. Thus, the conversion tag is obtained. During online training, for the sample recommended content whose sample tags include conversion tags, they can be summarized and recorded as recommendation trigger conversion flow. Furthermore, the sample tag of the recommended triggered conversion flow may also include a trigger tag, which, combined with the conversion tag, can reflect whether the corresponding sample recommended content has been triggered and converted, and can also reflect whether the converted sample recommended content has been triggered, so as to distinguish whether miscounting of conversion data has occurred.

[0123] The first training unit 602 can be executed as follows: if it is determined that the sample mark is a trigger mark, the feature data is input into the first prediction network, a first prediction trigger value is obtained through the first prediction network, and the first prediction network is trained according to the trigger mark and the first prediction trigger value.

[0124] The sample is marked as a trigger mark, which means that the currently calculated recommended sample comes from the recommendation trigger stream, and its feature data can be input into the first prediction network to obtain the first prediction trigger value. The prediction trigger value is used to reflect the possibility that a sample recommendation content is triggered by the sample object.

[0125] By comparing the trigger mark of the recommended sample from the recommended trigger stream with the first predicted trigger value, the corresponding first loss function can be calculated. Based on the first loss function, the parameters of the first prediction network can be adjusted to achieve its training. It should be understood that since the first training unit 602 does not input the feature data into the second prediction network, the corresponding network output value will not be generated, and the corresponding loss function will not be obtained, so the second prediction network will not be adjusted, which not only reduces the amount of training calculations, but also can timely use the newly obtained trigger mark and the corresponding recommended sample to train the first prediction network, reduce the risk of delay, and help to obtain a higher AUC, thereby ensuring the prediction accuracy of the first prediction network, and ultimately improving the prediction accuracy of the prediction model.

[0126] The second training unit 603 can be executed as follows: if it is determined that the sample tag includes a conversion tag, the feature data is input into the first prediction network and the second prediction network, the predicted conversion information of the sample recommendation content corresponding to the sample object is obtained through the first prediction network and the second prediction network, and the second prediction network is trained according to the sample tag and the predicted conversion information.

[0127] The sample is marked as a conversion mark, which means that the sample recommendation content currently calculated comes from the recommendation trigger conversion flow, and its feature data can be input into the first prediction network and the second prediction network to obtain the predicted conversion information. When the prediction model obtained by training is applied, the first prediction network and the second prediction network are integrated to obtain the predicted conversion information according to this process. Therefore, the feature data is also input into these two networks at the same time during training, which can ensure that the training is consistent with the later application and ensure the prediction accuracy of the later application model. The second training unit 603 can use the recommendation trigger conversion flow to train the second prediction network, which can reduce the risk of positive sample loss and ensure the prediction accuracy of the second prediction network. At the same time, the second training unit 603 only trains the second prediction network, which can ensure that the first prediction network trained by the recommendation trigger flow is not affected, and realizes the independent training of the first prediction network and the second prediction network.

[0128] Optionally, the second training unit 603 inputs the feature data into the first prediction network and the second prediction network, and obtains the predicted conversion information of the sample recommendation content corresponding to the sample object through the first prediction network and the second prediction network, which specifically includes four steps.

[0129] In the first step, the feature data is input into the first prediction network, and the second prediction trigger value is obtained through the first prediction network. It should be understood that the second prediction trigger value here is to distinguish it from the first prediction trigger value output when training the first prediction network. The difference between the two lies in which data stream the sample recommendation content used comes from, but the final results are all prediction trigger values. When applying the network, it can be understood that only the second prediction trigger value is calculated, and only one first prediction trigger value is not output, that is, the first prediction trigger value is only used for the training of the second prediction network.

[0130] In the second step, the feature data is input into the second prediction network, and the predicted triggered conversion value and the predicted untriggered conversion value are obtained through the second prediction network. The predicted triggered conversion value reflects the possibility of conversion of the sample recommended content when it is triggered by the sample object, and the predicted untriggered conversion value reflects the possibility of conversion of the sample recommended content when it is not triggered by the sample object. Specifically, the second prediction network includes a first sub-network and a second sub-network, and the feature data can be input into the two sub-networks at the same time to obtain the predicted triggered conversion value and the predicted untriggered conversion value respectively.

[0131] The third step is to determine the first conversion value and the second conversion value based on the second predicted trigger value, the predicted trigger conversion value and the predicted no-trigger conversion value. The first conversion value reflects the possibility that the sample recommended content is triggered by the sample object and a conversion occurs, and the second conversion value reflects the possibility that the sample recommended content is not triggered by the sample object and a conversion occurs. As an example, the product of the second predicted trigger value and the predicted trigger conversion value can be determined as the first conversion value; the predicted no-trigger value can be determined based on the second predicted trigger value; and the product of the predicted no-trigger value and the predicted no-trigger conversion value can be determined as the second conversion value.

[0132] The fourth step is to determine predicted conversion information based on the first conversion value and the second conversion value.

[0133] In some embodiments, the predicted conversion value may be first determined based on the first conversion value and the second conversion value, for example, the sum of the two may be determined as the predicted conversion value, and then the predicted conversion information may be determined based on the predicted conversion value. For determining the predicted conversion information, the predicted conversion value may be directly used as the predicted conversion information, or may be further processed to obtain the predicted conversion information.

[0134] In other embodiments, the first conversion value and the second conversion value may be used together as predicted conversion information without further calculation processing.

[0135] Optionally, the recommended sample with the conversion mark also has a trigger mark, and the operation of the second training unit 603 training the second prediction network according to the sample mark and the predicted conversion information includes: training the second prediction network according to the trigger mark, the conversion mark, the first conversion value and the second conversion value. The training operation can be specifically to determine the recommended trigger conversion mark and the recommended non-trigger conversion mark according to the conversion mark and the trigger mark; and train the second prediction network according to the recommended trigger conversion mark, the recommended non-trigger conversion mark, the first conversion value and the second conversion value. When training the second prediction network, the recommended trigger conversion mark can be compared with the first conversion value to obtain the corresponding second loss function, and then the parameters of the first sub-network in the second prediction network can be adjusted according to the second loss function; the recommended non-trigger conversion mark can be compared with the second conversion value to obtain the corresponding third loss function, and then the parameters of the second sub-network in the second prediction network can be adjusted according to the third loss function. In addition, to avoid adjusting the parameters of the first prediction network, corresponding isolation instructions can be added.

[0136] Figure 7 is a block diagram of an electronic device according to an exemplary embodiment of the present disclosure.

[0137] Reference Figure 7 The electronic device 700 includes at least one memory 701 and at least one processor 702, wherein the at least one memory 701 stores a set of computer executable instructions. When the computer executable instruction set is executed by the at least one processor 702, a prediction method for content or a training method for a prediction model according to an exemplary embodiment of the present disclosure is executed.

[0138] As an example, the electronic device 700 may be a PC, a tablet device, a personal digital assistant, a smart phone, or other device capable of executing the above instruction set. Here, the electronic device 700 is not necessarily a single electronic device, but may also be any device or circuit capable of executing the above instruction (or instruction set) individually or in combination. The electronic device 700 may also be part of an integrated control system or system manager, or may be configured as a portable electronic device interconnected with a local or remote (e.g., via wireless transmission) interface.

[0139] In the electronic device 700, the processor 702 may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller or a microprocessor. As an example and not limitation, the processor may also include an analog processor, a digital processor, a microprocessor, a multi-core processor, a processor array, a network processor, etc.

[0140] The processor 702 may execute instructions or codes stored in the memory 701, which may also store data. Instructions and data may also be sent and received over a network via a network interface device, which may employ any known transmission protocol.

[0141] The memory 701 may be integrated with the processor 702, for example, by placing RAM or flash memory within an integrated circuit microprocessor or the like. In addition, the memory 701 may include a separate device, such as an external disk drive, a storage array, or any other storage device that can be used by a database system. The memory 701 and the processor 702 may be operatively coupled, or may communicate with each other, such as through an I / O port, a network connection, etc., so that the processor 702 can read files stored in the memory.

[0142] In addition, the electronic device 700 may further include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, a mouse, a touch input device, etc.) All components of the electronic device 700 may be connected to each other via a bus and / or a network.

[0143] According to an exemplary embodiment of the present disclosure, a computer-readable storage medium may also be provided, and when the instructions in the computer-readable storage medium are executed by at least one processor, the at least one processor is prompted to execute the prediction method for content or the training method of the prediction model according to the exemplary embodiment of the present disclosure. Examples of computer-readable storage media here include: read-only memory (ROM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-R LTH, BD-RE, Blu-ray or optical disk storage, hard disk drive (HDD), solid state drive (SSD), card storage (such as, multimedia card, secure digital (SD) card or extreme digital (XD) card), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid state disk and any other device, any other device is configured to store computer programs and any associated data, data files and data structures in a non-transitory manner and provide the computer programs and any associated data, data files and data structures to a processor or computer so that the processor or computer can execute the computer program. The computer program in the above-mentioned computer-readable storage medium can be run in an environment deployed in a computer device such as a client, a host, an agent device, a server, etc. In addition, in one example, the computer program and any associated data, data files and data structures are distributed on a networked computer system, so that the computer program and any associated data, data files and data structures are stored, accessed and executed in a distributed manner by one or more processors or computers.

[0144] According to an exemplary embodiment of the present disclosure, a computer program product may also be provided. The computer program product includes computer instructions. When the computer instructions are executed by at least one processor, the at least one processor is prompted to perform the content prediction method or the prediction model training method according to the exemplary embodiment of the present disclosure.

[0145] According to the content prediction method, content prediction device, prediction model training method and training device of the embodiments of the present disclosure, the prediction model can respectively predict the possibility of the recommended content being triggered and converted and the possibility of not being triggered and converted, and then predict the possibility of the recommended content being converted. It can count the situation of no triggering but conversion, solve the problem of positive sample loss, thereby compensating for the defects of the existing model and helping to improve the prediction accuracy.

[0146] 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.

[0147] 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 method for predicting content, characterized in that: include: Acquire feature data of candidate recommended content for the candidate object; wherein the feature data includes content features corresponding to the candidate recommended content and object features corresponding to the candidate object, the content features include at least one of the following: candidate recommended content ID, candidate recommended content provider name, candidate recommended content provider industry category, candidate recommended content size, candidate recommended content embedding features obtained based on a visual algorithm, and the object features include at least one of the following: candidate object ID, age, gender, region, Internet IP, model of terminal used, historically triggered recommended content, historically converted recommended content, disliked recommended content, and embedding features obtained based on multimedia resources browsed by the candidate object; Input the feature data into a trained prediction model, and obtain a first conversion value and a second conversion value through the prediction model; wherein the first conversion value reflects the possibility that the candidate recommended content is triggered by the candidate object and is converted, and the second conversion value reflects the possibility that the candidate recommended content is not triggered by the candidate object and is converted; Determine predicted conversion information of the candidate recommended content corresponding to the candidate object according to the first conversion value and the second conversion value.

2. The content prediction method according to claim 1, characterized in that: The prediction model includes a first prediction network and a second prediction network; wherein, inputting the feature data into a trained prediction model and obtaining a first conversion value and a second conversion value through the prediction model comprises: Inputting the characteristic data into the first prediction network, and obtaining a prediction trigger value through the first prediction network; Inputting the feature data into the second prediction network, and obtaining a conversion parameter through the second prediction network, wherein the conversion parameter reflects the possibility of conversion of the candidate recommendation content under different triggering situations; The first conversion value and the second conversion value are determined according to the predicted trigger value and the conversion parameter.

3. The content prediction method according to claim 2, characterized in that: The conversion parameters include a predicted triggered conversion value and a predicted untriggered conversion value. The predicted triggered conversion value reflects the possibility of the candidate recommended content being converted when triggered by the candidate object, and the predicted untriggered conversion value reflects the possibility of the candidate recommended content being converted when not triggered by the candidate object.

4. The content prediction method according to claim 3, characterized in that: The second prediction network includes a first sub-network and a second sub-network; wherein the step of inputting the feature data into the second prediction network and obtaining the conversion parameter through the second prediction network includes: Inputting the feature data into the first sub-network, and obtaining the predicted trigger conversion value through the first sub-network; The feature data is input into the second sub-network, and the predicted untriggered conversion value is obtained through the second sub-network.

5. The content prediction method according to claim 3, characterized in that: The determining the first conversion value and the second conversion value according to the predicted trigger value and the conversion parameter includes: determining a product of the predicted trigger value and the predicted trigger conversion value as the first conversion value; Determining a predicted no trigger value according to the predicted trigger value; A product of the predicted no-trigger value and the predicted no-trigger conversion value is determined as the second conversion value.

6. The content prediction method according to any one of claims 1 to 5, characterized in that: The first conversion value reflects the probability that the candidate recommended content is triggered by the candidate object and converted, and the second conversion value reflects the probability that the candidate recommended content is not triggered by the candidate object and converted; wherein determining the predicted conversion information according to the first conversion value and the second conversion value includes: A sum of the first conversion value and the second conversion value is determined, and the predicted conversion information is determined based on the sum.

7. The content prediction method according to any one of claims 1 to 5, characterized in that: The feature data further includes a cross feature, and the cross feature reflects the association between the candidate object and the candidate recommendation content.

8. A method for training a prediction model, characterized in that: The prediction model includes a first prediction network and a second prediction network, and the training method includes: Acquire feature data and sample tags of recommended samples; wherein, the recommended samples include sample recommended content and sample objects targeted by the sample recommended content, and the feature data include content features corresponding to the sample recommended content and object features corresponding to the sample objects, wherein the content features include at least one of the following: sample recommended content ID, sample recommended content provider name, sample recommended content provider industry category, sample recommended content size, and sample recommended content embedded features obtained based on a visual algorithm; and the object features include at least one of the following: sample object ID, age, gender, region, Internet IP, model of terminal used, recommended content triggered in the past, recommended content converted in the past, disliked recommended content, and embedded features obtained based on multimedia resources browsed by the sample object; If it is determined that the sample mark is a trigger mark, the feature data is input into the first prediction network, a first prediction trigger value is obtained through the first prediction network, and the first prediction network is trained according to the trigger mark and the first prediction trigger value; wherein the trigger mark is used to reflect whether the sample recommendation content is triggered by the sample object; If it is determined that the sample tag includes a conversion tag, the feature data is input into the first prediction network and the second prediction network, a first conversion value and a second conversion value are obtained through the first prediction network and the second prediction network, predicted conversion information of the sample recommended content corresponding to the sample object is determined based on the first conversion value and the second conversion value, and the second prediction network is trained based on the sample tag and the predicted conversion information; wherein the conversion tag is used to reflect whether the sample recommended content undergoes a conversion corresponding to the sample object, the first conversion value reflects the possibility that the sample recommended content is triggered by the sample object and undergoes a conversion, and the second conversion value reflects the possibility that the sample recommended content is not triggered by the sample object and undergoes a conversion.

9. The training method according to claim 8, characterized in that: The step of inputting the feature data into the first prediction network and the second prediction network, and obtaining the first conversion value and the second conversion value through the first prediction network and the second prediction network includes: Inputting the characteristic data into the first prediction network, and obtaining a second prediction trigger value through the first prediction network; Inputting the feature data into the second prediction network, and obtaining a predicted triggered conversion value and a predicted untriggered conversion value through the second prediction network, wherein the predicted triggered conversion value reflects the possibility of conversion of the sample recommended content when triggered by the sample object, and the predicted untriggered conversion value reflects the possibility of conversion of the sample recommended content when not triggered by the sample object; A first conversion value and a second conversion value are determined according to the second predicted trigger value, the predicted triggered conversion value, and the predicted untriggered conversion value.

10. The training method according to claim 9, characterized in that: The recommended sample with the conversion mark also has the trigger mark, and the training of the second prediction network according to the sample mark and the predicted conversion information includes: The second prediction network is trained according to the trigger mark, the conversion mark, the first conversion value, and the second conversion value.

11. The training method according to any one of claims 8 to 10, characterized in that: Acquiring the sample label of the recommended sample includes: Acquire trigger data within a first time period after the sample recommended content is recommended to the sample object, and determine the trigger mark according to the trigger data; wherein the trigger data is data generated when the sample recommended content is triggered by the sample object; Acquire conversion data within a second time period after the sample recommended content is recommended to the sample object, and determine the conversion mark based on the conversion data; wherein the conversion data is data generated when the sample recommended content is converted, and the second time period is greater than the first time period.

12. A content prediction device, characterized in that: include: The acquisition unit is configured to: acquire feature data of candidate recommended content for the candidate object; wherein the feature data includes content features corresponding to the candidate recommended content and object features corresponding to the candidate object, the content features include at least one of the following: candidate recommended content ID, candidate recommended content provider name, candidate recommended content provider industry category, candidate recommended content size, candidate recommended content embedding features obtained based on a visual algorithm, and the object features include at least one of the following: candidate object ID, age, gender, region, Internet IP, model of terminal used, historically triggered recommended content, historically converted recommended content, disliked recommended content, and embedding features obtained based on multimedia resources browsed by the candidate object; A prediction unit is configured to: input the feature data into a trained prediction model, and obtain a first conversion value and a second conversion value through the prediction model; wherein the first conversion value reflects the possibility that the candidate recommended content is triggered by the candidate object and is converted, and the second conversion value reflects the possibility that the candidate recommended content is not triggered by the candidate object and is converted; The determination unit is configured to determine, according to the first conversion value and the second conversion value, predicted conversion information of the candidate recommended content corresponding to the candidate object.

13. The content prediction device according to claim 12, characterized in that: The prediction model includes a first prediction network and a second prediction network, and the prediction unit is further configured to: Inputting the characteristic data into the first prediction network, and obtaining a prediction trigger value through the first prediction network; Inputting the feature data into the second prediction network, and obtaining a conversion parameter through the second prediction network, wherein the conversion parameter reflects the possibility of conversion of the candidate recommendation content under different triggering situations; The first conversion value and the second conversion value are determined according to the predicted trigger value and the conversion parameter.

14. The content prediction device according to claim 13, characterized in that: The conversion parameters include a predicted triggered conversion value and a predicted untriggered conversion value. The predicted triggered conversion value reflects the possibility of the candidate recommended content being converted when triggered by the candidate object, and the predicted untriggered conversion value reflects the possibility of the candidate recommended content being converted when not triggered by the candidate object.

15. The content prediction device according to claim 14, characterized in that: The second prediction network includes a first sub-network and a second sub-network, and the prediction unit is further configured as follows: Inputting the feature data into the first sub-network, and obtaining the predicted trigger conversion value through the first sub-network; The feature data is input into the second sub-network, and the predicted untriggered conversion value is obtained through the second sub-network.

16. The content prediction device according to claim 14, characterized in that: The prediction unit is further configured to: determining a product of the predicted trigger value and the predicted trigger conversion value as the first conversion value; Determining a predicted no trigger value according to the predicted trigger value; A product of the predicted no-trigger value and the predicted no-trigger conversion value is determined as the second conversion value.

17. The content-specific prediction device according to any one of claims 12 to 16, characterized in that: The first conversion value reflects the probability that the candidate recommended content is triggered by the candidate object and converted, and the second conversion value reflects the probability that the candidate recommended content is not triggered by the candidate object and converted, wherein the determining unit is further configured to: A sum of the first conversion value and the second conversion value is determined, and the predicted conversion information is determined based on the sum.

18. The content-specific prediction device according to any one of claims 12 to 16, characterized in that: The feature data further includes a cross feature, and the cross feature reflects the association between the candidate object and the candidate recommendation content.

19. A training device for a prediction model, characterized in that: The prediction model includes a first prediction network and a second prediction network, and the training device includes: The acquisition unit is configured to: acquire feature data and sample tags of the recommendation sample; wherein the recommendation sample includes sample recommendation content and sample objects targeted by the sample recommendation content, the feature data includes content features corresponding to the sample recommendation content and object features corresponding to the sample objects, the content features include at least one of the following: sample recommendation content ID, sample recommendation content provider name, sample recommendation content provider industry category, sample recommendation content size, sample recommendation content embedding features obtained based on a visual algorithm, and the object features include at least one of the following: sample object ID, age, gender, region, Internet IP, model of terminal used, historically triggered recommendation content, historically converted recommendation content, disliked recommendation content, and embedding features obtained based on multimedia resources browsed by the sample object; A first training unit is configured to: if it is determined that the sample tag is a trigger tag, input the feature data into the first prediction network, obtain a first prediction trigger value through the first prediction network, and train the first prediction network according to the trigger tag and the first prediction trigger value; wherein the trigger tag is used to reflect whether the sample recommendation content is triggered by the sample object; The second training unit is configured to: if it is determined that the sample tag includes a conversion tag, input the feature data into the first prediction network and the second prediction network, obtain a first conversion value and a second conversion value through the first prediction network and the second prediction network, determine the predicted conversion information of the sample recommended content corresponding to the sample object according to the first conversion value and the second conversion value, and train the second prediction network according to the sample tag and the predicted conversion information; wherein the conversion tag is used to reflect whether the sample recommended content undergoes a conversion corresponding to the sample object, the first conversion value reflects the possibility that the sample recommended content is triggered by the sample object and undergoes a conversion, and the second conversion value reflects the possibility that the sample recommended content is not triggered by the sample object and undergoes a conversion.

20. The training device according to claim 19, characterized in that The second training unit is further configured to: Inputting the characteristic data into the first prediction network, and obtaining a second prediction trigger value through the first prediction network; Inputting the feature data into the second prediction network, and obtaining a predicted triggered conversion value and a predicted untriggered conversion value through the second prediction network, wherein the predicted triggered conversion value reflects the possibility of conversion of the sample recommended content when triggered by the sample object, and the predicted untriggered conversion value reflects the possibility of conversion of the sample recommended content when not triggered by the sample object; A first conversion value and a second conversion value are determined according to the second predicted trigger value, the predicted triggered conversion value, and the predicted untriggered conversion value.

21. The training device according to claim 20, characterized in that The recommended sample with the conversion mark also has the trigger mark, and the second training unit is further configured to: The second prediction network is trained according to the trigger mark, the conversion mark, the first conversion value, and the second conversion value.

22. A training device according to any one of claims 19 to 21, characterised in that The acquisition unit is further configured to: Acquire trigger data within a first time period after the sample recommended content is recommended to the sample object, and determine the trigger mark according to the trigger data; wherein the trigger data is data generated when the sample recommended content is triggered by the sample object; Acquire conversion data within a second time period after the sample recommended content is recommended to the sample object, and determine the conversion mark based on the conversion data; wherein the conversion data is data generated when the sample recommended content is converted, and the second time period is greater than the first time period.

23. An electronic device, characterized in that: include: at least one processor; at least one memory storing computer executable instructions, When the computer executable instructions are executed by the at least one processor, the at least one processor is prompted to execute the content prediction method as described in any one of claims 1 to 7 or the prediction model training method as described in any one of claims 8 to 11.

24. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by at least one processor, the at least one processor is prompted to perform the content prediction method according to any one of claims 1 to 7 or the prediction model training method according to any one of claims 8 to 11.

25. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by at least one processor, the content prediction method according to any one of claims 1 to 7 or the prediction model training method according to any one of claims 8 to 11 is implemented.

Images