An information processing method and apparatus

Abstract

This application discloses an information processing method and apparatus. The method includes: acquiring source information, which includes at least two data points; determining the click-through rate (CTR) of each data point in the source information based on a click-through rate (CTR) prediction model, wherein the CTR model is trained based on exposed and unexposed data, and during the training process, the prediction results of the CTR prediction model are corrected by lowering the contribution rate of exposed data or increasing the contribution rate of unexposed data; and sorting the source information according to the CTR of each data point to determine the target information to be displayed.

Description

Technical Field

[0001] This application relates to the field of information processing technology, and in particular to an information processing method and apparatus. Background Technology

[0002] Currently, click-through rate (CTR) prediction models are typically trained based on information from exposed data. However, because exposed and unexposed samples have different distributions, considering only exposed samples leads to non-random sample selection that fails to represent the population, resulting in selection bias. Summary of the Invention

[0003] This application provides an information processing method and apparatus.

[0004] According to a first aspect of the embodiments of this application, an information processing method is provided, comprising: acquiring source information, the source information including at least two data points; determining the click-through rate (CTR) of each data point in the source information based on a click-through rate (CTR) prediction model, wherein the CTR model is trained based on exposed data and unexposed data, and during the training of the CTR prediction model, the prediction results of the CTR prediction model are corrected by lowering the contribution rate of exposed data or increasing the contribution rate of unexposed data; and sorting the source information according to the CTR of each data point to determine the target information to be displayed.

[0005] According to one embodiment of this application, the method further includes: determining first sample data based on exposed data and the actual click value of the exposed data, wherein the click label value of the first sample data is the actual click value of the exposed data; predicting the click information of unexposed data to obtain the predicted click value of the unexposed data; determining second sample data based on the unexposed data and the predicted click value of the unexposed data, wherein the click label value of the second sample data is the predicted click value of the unexposed data; determining the click rate corresponding to each data point in the first sample data and the click rate corresponding to each data point in the second sample data based on the first sample data, the second sample data, and the click rate prediction model; adjusting the contribution rate of the first sample data or increasing the contribution rate of the second sample data during the process of determining the loss function value of the first sample data and the second sample data; and training the click rate prediction model based on the loss function value.

[0006] According to one embodiment of this application, predicting the click information of unexposed data to obtain the predicted click value of the unexposed data includes: predicting the click information of unexposed data based on a click tag prediction model to obtain the predicted click value of the unexposed data, wherein the click tag prediction model is trained based on the click information of random data, and the random data is data randomly selected from the exposure situation of ignored data.

[0007] According to one embodiment of this application, reducing the contribution rate of the first sample data or increasing the contribution rate of the second sample data includes: reducing the contribution rate of the first sample data based on the predicted exposure rate of the first sample data; or increasing the contribution rate of the second sample data based on the predicted exposure rate of the second sample data.

[0008] According to one embodiment of this application, the method further includes: determining the predicted exposure rate of the first sample data and the predicted exposure rate of the second sample data based on an exposure rate prediction model, wherein the predicted exposure rate of the first sample data determined by the exposure rate prediction model is greater than the predicted exposure rate of the second sample data.

[0009] According to one embodiment of this application, reducing the contribution rate of the first sample data based on the predicted exposure rate of the first sample data includes: using the predicted exposure rate as a divisor factor to reduce the contribution rate of the first sample data.

[0010] According to one embodiment of this application, the method further includes: amplifying the predicted exposure rate using hyperparameters to constrain the value of the loss function.

[0011] According to one embodiment of this application, the predicted exposure rate is amplified by hyperparameters, including: increasing the predicted exposure rate by a fixed value by hyperparameters; or amplifying the predicted exposure rate by a fixed factor by hyperparameters.

[0012] According to one embodiment of this application, it further includes: using the first N data points in the target information as training data to retrain the click-through rate prediction model, where N is a natural number greater than 1.

[0013] According to a second aspect of the embodiments of this application, an information processing apparatus is provided, comprising: a source information acquisition module for acquiring source information, the source information including at least two data points; a click-through rate (CTR) prediction module for determining the CTR of each data point in the source information based on a CTR prediction model, wherein the CTR model is trained based on exposed data and unexposed data, and during the training of the CTR prediction model, the prediction result of the CTR prediction model is corrected by lowering the contribution rate of exposed data or increasing the contribution rate of unexposed data; and a data sorting module for sorting the source information according to the CTR of each data point to determine the target information to be displayed.

[0014] According to a third aspect of the present application, a computer storage medium is provided, the storage medium including a set of computer-executable instructions, which, when executed, are used to perform any of the above-described information processing methods.

[0015] This application provides an information processing method and apparatus. The method includes: acquiring source information, which includes at least two data points; determining the click-through rate (CTR) of each data point in the source information based on a click-through rate (CTR) prediction model, wherein the CTR model is trained based on exposed and unexposed data, and during the training process, the prediction results of the CTR prediction model are corrected by lowering the contribution rate of exposed data or increasing the contribution rate of unexposed data; and sorting the source information according to the CTR of each data point to determine the target information to be displayed. Because the prediction results of the CTR prediction model are corrected by lowering the contribution rate of exposed data or increasing the contribution rate of unexposed data during the training process, the prediction results of the CTR prediction model are more accurate.

[0016] It should be understood that the implementation of this application does not need to achieve all the beneficial effects described above. Rather, a specific technical solution can achieve a specific technical effect, and other embodiments of this application can also achieve beneficial effects not mentioned above. Attached Figure Description

[0017] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, wherein:

[0018] In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts.

[0019] Figure 1 This is a schematic diagram illustrating the basic flow of an information processing method according to an embodiment of this application;

[0020] Figure 2 This is a schematic diagram of the process of training a click-through rate prediction model according to an embodiment of this application;

[0021] Figure 3 This is a schematic flowchart illustrating the specific implementation of the information processing method according to another embodiment of this application;

[0022] Figure 4 This is a schematic diagram of the training data used in another embodiment of the click label prediction model of this application;

[0023] Figure 5 This is a schematic diagram of the composition of the information processing device according to an embodiment of this application. Detailed Implementation

[0024] To make the objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0026] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0027] The information processing method provided in this application is mainly applied in scenarios where data is sorted from highest to lowest based on its click-through rate, so that data with higher click-through rates is displayed first, thereby shortening the time users spend searching for information and significantly improving the efficiency of users obtaining information.

[0028] Figure 1 The implementation flow of an information processing method according to an embodiment of this application is shown. (Reference) Figure 1 The method includes:

[0029] Operate S110 to obtain source information, which includes at least two data points;

[0030] Source information refers to the searchable information provided by a data provider or data exchange platform where a user is looking for specific data. For example, searchable papers provided by a paper search platform; product information available for sale provided by a manufacturer.

[0031] Depending on the application scenario, the source information can be all information or a certain type of information that has been filtered according to specific conditions.

[0032] Each piece of data in the source information refers to a data unit that can be displayed independently in the source information, such as a paper, a sentence in a paper, or a product in a product library.

[0033] Operation S120 determines the click-through rate of each piece of data in the source information based on the click-through rate prediction model. The click-through rate model is trained based on exposed data and unexposed data. During the training of the click-through rate prediction model, the prediction results of the click-through rate prediction model are corrected by lowering the contribution rate of exposed data or increasing the contribution rate of unexposed data.

[0034] The click-through rate (CTR) of each data point represents the probability that the data will be clicked by a user. It is usually a decimal between 0 and 1, with the closer it is to 1, the greater the probability of it being clicked.

[0035] Click-through rate (CTR) prediction models are used to predict the CTR of each data point based on factors such as the user type of the current user, the attributes of the product itself, historical click data, and / or recent click data trends.

[0036] Ideally, a click-through rate (CTR) prediction model should be trained using unbiased data, that is, all known data should be exposed to users, and the CTR prediction model should be trained based on the actual click values ​​of whether users clicked on these data. In this way, the CTR prediction model can obtain the user's true click intention for each data, thereby enabling the trained CTR prediction model to more accurately predict the user's click rate for unknown data.

[0037] However, in practical applications, due to the massive amount of existing data, the system usually cannot display all of it to the user. Instead, it can only display a portion of the data for the user to browse according to specific rules. The portion of data displayed to the user is called exposed data, while the remaining data not displayed to the user is called unexposed data.

[0038] Since click tags for only exposed data can be obtained from actual user click data, most technical solutions for training click-through rate (CTR) prediction models either use only exposed data as training samples or use a click tag prediction model trained on exposed data to predict click tags for unexposed data. Predicting click tags for unexposed data is also typically based on user click data on exposed data. For example, when displaying data like "puppy" and "kitten," if a user clicks on "kitten" but not "puppy," then for unexposed data, the similarity between the unexposed data and "kitten" or "puppy" is used to predict whether the user will click: if it's more similar to "kitten," it's predicted as "click"; if it's more similar to "puppy," it's predicted as "no click."

[0039] Therefore, the impact of exposed data on the prediction results of the click-through rate (CTR) prediction model is very significant. If the rules used to expose the data are biased, for example, making data with certain characteristics more likely to be exposed, the prediction results of the CTR prediction model will be skewed.

[0040] In this embodiment, correcting the prediction results of the click-through rate (CTR) prediction model mainly refers to correcting the deviation caused by the excessive influence of exposed data on the prediction results. Specifically, this is mainly achieved by reducing the contribution rate of exposed data or increasing the contribution rate of unexposed data, thereby minimizing the impact of exposed data on the prediction results or increasing the impact of unexposed data, thus avoiding or reducing the influence of exposure rules on the prediction results and making the prediction results of the CTR prediction model closer to the user's true intentions.

[0041] The contribution rate of exposed data mainly refers to the degree of influence of exposed data on the model training results. Lowering the contribution rate of exposed data can be achieved by reducing the proportion of exposed data in the total sample data; or by reducing the proportion of the loss function value of exposed data in the overall model loss function value.

[0042] Similarly, the contribution rate of unexposed data mainly refers to the degree of influence of unexposed data on the model training results. Increasing the contribution rate of unexposed data can be achieved by increasing the proportion of unexposed data in the total sample data; or by increasing the proportion of the loss function value of unexposed data in the total loss function value of the model.

[0043] Operation S130 sorts the source information according to the click rate of each piece of data to determine the target information to be displayed.

[0044] In this embodiment, since the click-through rate (CTR) of each data point is predicted using a corrected click model, the negative impact of excessive influence from already exposed data on the CTR model's prediction results can be mitigated. This makes the CTR model's prediction results more accurate and closer to the user's true intentions. Consequently, the information needed by the user can be pushed more precisely.

[0045] Figure 2 The main process of training a click-through rate prediction model according to an embodiment of this application is shown, including:

[0046] Operation S210: Based on the exposed data and the actual click value of the exposed data, determine the first sample data. The click tag value of the first sample data is the actual click value of the exposed data.

[0047] Typically, the clicked label value is a boolean, where 0 or false represents no click and 1 or true represents a click.

[0048] In other words, for a specific piece of data in the exposed data, if a user clicks on it, the click tag value of that data is set to 1 or true; correspondingly, if the user does not click on it, the click tag value of that data is set to 0 or false.

[0049] Operation S220 is used to predict the click information of the unexposed data and obtain the predicted click value of the unexposed data.

[0050] In the embodiments of this application, there is no limitation on the specific method of predicting the click information of unexposed data. It can be predicted by a click tag prediction model or by estimating the value through a statistical model and the distribution of click values.

[0051] Operation S230: Based on the unexposed data and the predicted click value of the unexposed data, determine the second sample data. The click label value of the second sample data is the predicted click value of the unexposed data.

[0052] Operation S240: Based on the first sample data, the second sample data, and the click-through rate prediction model, determine the click-through rate corresponding to each piece of data in the first sample data and the click-through rate corresponding to each piece of data in the second sample data;

[0053] Operation S250: In the process of determining the loss function values ​​of the first sample data and the second sample data, adjust the contribution rate of the first sample data to decrease or the contribution rate of the second sample data to increase.

[0054] In this embodiment, reducing the contribution rate of the first sample data or increasing the contribution rate of the second sample data is mainly achieved by reducing the proportion of the loss function value of the exposed data in the entire model loss function value or increasing the proportion of the loss function value of the unexposed data in the entire model loss function value during the process of determining the loss function values ​​of the first and second sample data.

[0055] Operate S260 to train the click-through rate prediction model based on the loss function value.

[0056] Training the click-through rate (CTR) prediction model involves adjusting the algorithm used by the CTR prediction model or adjusting the parameter values ​​of the CTR prediction model to minimize the loss function value.

[0057] Reducing the proportion of exposure data in the overall sample data may decrease the proportion of data with actual click tags, potentially reducing the accuracy of click-through rate model predictions.

[0058] In this embodiment, the contribution rate of exposed data is reduced or the contribution rate of unexposed data is increased by lowering the proportion of the loss function value of exposed data in the total loss function value of the model or increasing the proportion of the loss function value of unexposed data in the total loss function value of the model. This not only makes fuller use of the real labels in the exposed data and makes the click-through rate model prediction results more reliable, but also corrects the result deviations that may be caused by the exposed data, thereby making the click-through rate model prediction results more accurate.

[0059] Figure 3 This illustration shows a specific implementation flow of an information processing method according to another embodiment of this application. This embodiment is applied to a scenario on a shopping website where products are ranked based on predicted click-through rates, and products with higher click-through rates are recommended to users.

[0060] In this embodiment, the click tag value of the unexposed data is obtained by predicting the click information of the unexposed data based on the click tag prediction model; and the contribution rate of the exposed data is reduced and the contribution rate of the unexposed data is increased by predicting the exposure rate. The predicted exposure rate is determined based on the exposure rate prediction model.

[0061] like Figure 3 As shown, the process of training the click-through rate (CTR) prediction model and using the CTR prediction model's exposure data in this embodiment mainly includes the following steps:

[0062] Step S3010: Determine the exposure sample;

[0063] Among them, the exposure sample, as the first sample data, is determined based on the information of the exposed products and whether the exposed products have been clicked by users.

[0064] The exposure label for all exposed samples is set to "1", indicating that the data has been exposed.

[0065] If an exposed product is clicked, the click label for that sample is set to "1"; if an exposed product is not clicked, the click label for that sample is set to "0".

[0066] Step S3020: Use the click tag prediction model to determine the predicted value of the click tag for the unexposed sample;

[0067] Since the unexposed samples are not shown to users, it is impossible to know the actual value of whether users will click on them; it can only be obtained through prediction.

[0068] In this embodiment, the predicted value of the click tag of the unexposed sample is predicted using a click tag prediction model.

[0069] Step S3030: Identify unexposed samples;

[0070] Among them, the unexposed samples, as the second sample data, are determined based on the unexposed products and the predicted values ​​obtained in step S3020.

[0071] The exposure label for unexposed samples is set to "0", indicating that the data has not been exposed.

[0072] If it is predicted that the unexposed product will be clicked, the click label for this sample is set to "1"; if it is predicted that the unexposed product will not be clicked, the click label for this sample is set to "0".

[0073] Subsequently, the exposed samples and the unexposed samples together form the samples used to train the exposure rate prediction model and the click-through rate prediction model, and are used to train the exposure rate prediction model and the click-through rate prediction model through two branches respectively.

[0074] The branch used for training the exposure prediction model includes:

[0075] Step S3040: Train the exposure rate prediction model;

[0076] Step S3050: Predict the exposure rate of each sample using the trained exposure rate prediction model;

[0077] The branch used to train the click-through rate prediction model includes:

[0078] Step S3060: Train the click-through rate prediction model;

[0079] Step S3070: Predict the click-through rate for each sample using the click-through rate prediction model;

[0080] Step S3080: Calculate the value of the loss function using the all-proportional dual robust method. During the calculation of the loss function, the exposure rate of the sample is used for bias correction.

[0081] Among them, the all-biased dual robust method is a new method created by the inventor of this application based on the dual robust method, which uses the exposure rate of the sample to correct bias in the process of calculating the value of the loss function.

[0082] Specifically, when training the click-through rate (CTR) prediction model, the objective definition of the CTR prediction model does not solve an optimization problem of the following form:

[0083]

[0084] in,

[0085] The 1st, 2nd, ..., Lth samples are exposed samples and have labels;

[0086] L+1, L+2, ..., L ^These samples are unexposed samples and have no labels;

[0087] y l This is the true label of the l-th sample;

[0088] y l ^ It is the predicted value of the l-th sample;

[0089] σ l The label is generated for the l-th sample using a model derived from random flow learning;

[0090] z l It is the exposure rate of the l-th sample;

[0091] θ is a hyperparameter;

[0092] loss(a,b)=(ab) 2 Or binary cross-entropy, etc.

[0093] In this embodiment, z l The exposure rate is predicted using a trained exposure rate prediction model. In other embodiments, z l It can also be the exposure rate predicted based on a binary distribution (e.g., Bernoulli distribution).

[0094] Since the exposure rate of exposed data is usually greater than that of unexposed data, when using exposure rate as a divisor, samples with higher exposure rates actually contribute less to the CTR model's prediction results, while samples with lower exposure rates contribute more. This allows us to adjust the contribution rate of exposed sample data and increase the contribution rate of unexposed sample data by adjusting the exposure rate, thus preventing the CTR prediction model from being overly influenced by exposed data and correcting any biases in the model's predictions.

[0095] Furthermore, considering that most samples, especially unexposed samples, have very low exposure rates, placing the exposure rate of the sample in the denominator of the corresponding term of the loss function for bias correction may lead to an extremely large value of the loss function, making it difficult for the model to converge. Therefore, the same hyperparameter θ is introduced to amplify the predicted exposure rate in order to constrain the value of the loss function.

[0096] In this embodiment, a fixed value (z) is added to the predicted exposure rate using hyperparameters. l The predicted exposure rate is amplified using a +) method. In other embodiments, the predicted exposure rate can also be amplified by a fixed factor (z) using hyperparameters. l *) Amplify the predicted exposure rate.

[0097] The value of the hyperparameter θ can be determined by proposing some candidate values ​​and then using the experimental results of each candidate value to determine the final value.

[0098] Because of different exposure algorithms or rules, the same data may be exposed in some scenarios but not in others. Therefore, the exposure rate prediction models trained in different scenarios will also be different. In this case, correcting only the exposed data without correcting the unexposed data will have certain limitations.

[0099] In this embodiment, exposure rate is used for correction regardless of whether the sample has been exposed or not. This allows for precise correction based on the different exposure rates of each sample under different environments, making it applicable to various scenarios with different exposure rates, better achieving no selection bias, and having a wider range of applications.

[0100] However, using exposure rate for bias correction regardless of whether the sample has been exposed may result in high time and space complexity during the computation. To address this, this embodiment further reduces the time and space complexity during computation through data fusion and batch training.

[0101] Step S3090: Obtain the trained click-through rate prediction model;

[0102] After steps S3010 to S3090, through multiple rounds of training, the model algorithm and parameters are continuously optimized based on the loss function until the value of the loss function reaches the expected range, resulting in a well-trained click-through rate (CTR) prediction model. This trained CTR prediction model can then be put into practical application to predict the CTR of each product.

[0103] Step S3100: Use the trained click-through rate prediction model to predict click-through rate, ranking, and recommendation.

[0104] Specifically, the top N data points from the sorted target information are recommended to the user, becoming new exposure data, and then used as exposure samples in the next training data.

[0105] Step S3110: Obtain exposure samples to replace random traffic and start repeated training.

[0106] Since the exposure samples obtained through the corrected click-through rate prediction model have already undergone correction processing, after exposure, the feedback from users who click or not click online can be directly used as training samples for the click-through rate prediction model without further correction. This greatly simplifies the subsequent repeated training process of the click-through rate prediction model.

[0107] Furthermore, using exposed samples instead of random traffic to initiate repeated training can enable personalized recommendations, which is more conducive to increasing the probability of user clicks and thus improving merchants' revenue.

[0108] Figure 4 The training data used by a click label prediction model according to another embodiment of this application is shown. For example... Figure 4 As shown, this embodiment uses random traffic 401 instead of exposed data to train the click tag prediction model 40.

[0109] Random traffic 401 involves randomly selecting a subset of products from a pool of available items, ignoring exposure rules and the product's existing exposure status. These randomly selected products are then presented to the user for selection, thus creating random traffic. Users online provide feedback by clicking or not clicking on these random traffic items, which in turn adds corresponding click tags to the products: click tags for clicked products are set to "1," while click tags for unclicked products are set to "0."

[0110] Because the distribution of exposed and unexposed samples in random traffic is the same, and the exposure rate of the samples is the same, there is no selection bias. Using an unbiased model to generate labels will not propagate bias, which is conducive to obtaining high-quality samples.

[0111] This can further reduce the potential impact of exposed data on the prediction results of the click tag prediction model 40, making the prediction results of the click tag prediction model 40 more impartial. Consequently, the click tags generated based on the prediction results of the click tag prediction model 40 for unexposed data are more credible and closer to the user's true intentions.

[0112] To verify the practical application effect of the information processing method of this application, the inventors randomly exposed products to users, using clicked products as positive samples and unclicked products as negative samples, and learned an unbiased click label prediction model through a first neural network. During the random traffic period, 3291 products were exposed to 70563 users. Each "user-product" pair was considered a sample, resulting in a sample space ω of 232,222,833 samples. Of these, 952,414 samples in the sample space ω were exposed in the online non-random traffic, with an exposure label of 1 and a click label of 0 or 1. Other samples not exposed in the non-random traffic had an exposure label of 0, and their click labels were missing but could be generated using the unbiased click label prediction model.

[0113] Next, an exposure rate prediction model was learned using the second neural network and exposure labels. This model was applied to the sample space ω and the result was probabilized using the sigmoid function. The exposure rate of each sample was calculated for correction. By taking multiple values ​​and comparing the effects, the hyperparameter θ was set to 1e-4.

[0114] Then, using the all-proportional dual robust method mentioned earlier in this application, a click-through rate prediction model is learned with a third neural network and click labels. Using the sample space ω, and probabilizing the results with the sigmoid function, the 20 products with the highest click-through rates are recommended to each user.

[0115] The following results were obtained through experimental comparison:

[0116] Table 1

[0117]

[0118] As shown in Table 1, the products recommended in this application embodiment have a higher click-through rate than popular products, which shows that the products recommended in this application embodiment are more in line with user needs and are more popular with users.

[0119] Table 2

[0120]

[0121] The data in Table 2 represents the daily average metrics obtained from the tests. It is evident that products recommended using the all-preference dual robustness method described in this embodiment show significant improvements in click volume, number of clicks, click-through rate, and conversion rate. This not only makes them more popular with users but also brings greater profits to merchants.

[0122] Furthermore, embodiments of this application also provide an information processing apparatus. For example... Figure 5 As shown, the device 50 includes: a source information acquisition module 501, used to acquire source information, which includes at least two data points; a click-through rate (CTR) prediction module 502, used to determine the CTR of each data point in the source information based on a CTR prediction model, wherein the CTR model is trained based on exposed data and unexposed data, and during the training of the CTR prediction model, the prediction results of the CTR prediction model are corrected by lowering the contribution rate of exposed data or increasing the contribution rate of unexposed data; and a data sorting module 503, used to sort the source information according to the CTR of each data point to determine the target information to be displayed.

[0123] According to one embodiment of this application, the system further includes: a first sample data determination module, configured to determine first sample data based on exposed data and the actual click value of the exposed data, wherein the click label value of the first sample data is the actual click value of the exposed data; a predicted click value determination module, configured to predict the click information of unexposed data to obtain the predicted click value of the unexposed data; a second sample data determination module, configured to determine second sample data based on unexposed data and the predicted click value of the unexposed data, wherein the click label value of the second sample data is the predicted click value of the unexposed data; a click-through rate prediction module 502, further configured to determine the click-through rate corresponding to each piece of data in the first sample data and the click-through rate corresponding to each piece of data in the second sample data based on the first sample data, the second sample data, and the click-through rate prediction model; a loss function value determination module, configured to lower the contribution rate of the first sample data or increase the contribution rate of the second sample data during the process of determining the loss function values ​​of the first sample data and the second sample data; and a click-through rate prediction model training module, configured to train the click-through rate prediction model based on the loss function values.

[0124] According to one embodiment of this application, the predicted click value determination module is specifically used to predict the click information of unexposed data based on the click tag prediction model to obtain the predicted click value of the unexposed data. The click tag prediction model is trained based on the click information of random data, and the random data is data randomly selected from the exposure situation of ignored data.

[0125] According to one embodiment of this application, the loss function value determination module is specifically used to reduce the contribution rate of the first sample data based on the predicted exposure rate of the first sample data; or, to increase the contribution rate of the second sample data based on the predicted exposure rate of the second sample data.

[0126] According to one embodiment of this application, the device 50 further includes: an exposure rate prediction module, configured to determine the predicted exposure rate of the first sample data and the predicted exposure rate of the second sample data based on an exposure rate prediction model, wherein the predicted exposure rate of the first sample data determined according to the exposure rate prediction model is greater than the predicted exposure rate of the second sample data.

[0127] According to one embodiment of this application, the loss function value determination module is specifically used to use the predicted exposure rate as a divisor factor to reduce the contribution rate of the first sample data.

[0128] According to one embodiment of this application, the device 50 further includes a loss function value constraint module, which amplifies the predicted exposure rate through hyperparameters to constrain the value of the loss function.

[0129] According to one embodiment of this application, the loss function value constraint module is specifically used to increase the predicted exposure rate by a fixed value through hyperparameters; or, to amplify the predicted exposure rate by a fixed factor through hyperparameters.

[0130] According to one embodiment of this application, the device 50 further includes: a click-through rate prediction model training module, used to retrain the click-through rate prediction model by taking the first N data points in the target information as training data, where N is a natural number greater than 1.

[0131] According to a third aspect of the present application, a computer storage medium is provided, the storage medium including a set of computer-executable instructions, which, when executed, are used to perform any of the above-described information processing methods.

[0132] It should be noted that the above descriptions of the information processing device embodiments and the computer storage medium embodiments are similar to the descriptions of the foregoing method embodiments, and have similar beneficial effects, therefore, they will not be repeated. For technical details not disclosed in the descriptions of the information processing device embodiments and the computer storage medium embodiments of this application, please refer to the descriptions of the foregoing method embodiments of this application for understanding; to save space, they will not be repeated here.

[0133] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0134] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another device, or some features may be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0135] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0136] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0137] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage media, read-only memory (ROM), magnetic disks, or optical disks.

[0138] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage media, ROM, magnetic disks, or optical disks.

[0139] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An information processing method, comprising: Obtain source information, which includes at least two data points; Based on the click-through rate prediction model, the click-through rate of each piece of data in the source information is determined, wherein, The click-through rate (CTR) prediction model is trained based on exposed and unexposed data. During the training process, the prediction results of the CTR prediction model are corrected by lowering the contribution rate of exposed data or increasing the contribution rate of unexposed data. Lowering the contribution rate of exposed data includes reducing the proportion of exposed data in the total sample data or reducing the proportion of the loss function value of exposed data in the total model loss function value. Increasing the contribution rate of unexposed data includes increasing the proportion of unexposed data in the total sample data or increasing the proportion of the loss function value of unexposed data in the total model loss function value. Based on the click-through rate of each piece of data, the source information is sorted to determine the target information to be displayed.

2. The method according to claim 1, further comprising: Based on the exposed data and the actual click value of the exposed data, a first sample data is determined, wherein the click tag value of the first sample data is the actual click value of the exposed data; Predict the click information of unexposed data to obtain the predicted click value of the unexposed data; Based on the unexposed data and the predicted click value of the unexposed data, a second sample data is determined, wherein the click tag value of the second sample data is the predicted click value of the unexposed data. Based on the first sample data, the second sample data, and the click-through rate prediction model, determine the click-through rate corresponding to each piece of data in the first sample data and the click-through rate corresponding to each piece of data in the second sample data; In the process of determining the loss function values ​​of the first sample data and the second sample data, the contribution rate of the first sample data is reduced or the contribution rate of the second sample data is increased. The click-through rate prediction model is trained based on the loss function value.

3. The method according to claim 2, wherein predicting the click information of the unexposed data to obtain the predicted click value of the unexposed data includes: Based on a click tag prediction model, click information of unexposed data is predicted to obtain the predicted click value of the unexposed data. The click tag prediction model is trained based on click information from random data. The random data is data randomly selected based on the exposure of ignored data.

4. The method according to claim 2, wherein lowering the contribution rate of the first sample data or increasing the contribution rate of the second sample data comprises: The contribution rate of the first sample data is reduced based on the predicted exposure rate of the first sample data. Alternatively, the contribution rate of the second sample data can be increased by using the predicted exposure rate of the second sample data.

5. The method according to claim 4, further comprising: Based on the exposure rate prediction model, the predicted exposure rate of the first sample data and the predicted exposure rate of the second sample data are determined, wherein the predicted exposure rate of the first sample data determined according to the exposure rate prediction model is greater than the predicted exposure rate of the second sample data.

6. The method according to claim 4, wherein reducing the contribution rate of the first sample data based on the predicted exposure rate of the first sample data comprises: Using the predicted exposure rate as a divisor factor, the contribution rate of the first sample data is reduced.

7. The method according to claim 4, further comprising: The predicted exposure rate is amplified using hyperparameters to constrain the value of the loss function.

8. The method according to claim 7, wherein the step of amplifying the predicted exposure rate using hyperparameters comprises: The predicted exposure rate is increased by a fixed value using hyperparameters; Alternatively, the predicted exposure rate can be amplified by a fixed factor using hyperparameters.

9. The method according to claim 1, further comprising: The first N data points in the target information are used as training data to retrain the click-through rate prediction model, where N is a natural number greater than 1.

10. A data recommendation device, comprising: A source information acquisition module is used to acquire source information, which includes at least two data points. The click-through rate (CTR) prediction module is used to determine the CTR of each piece of data in the source information based on a CTR prediction model. The CTR prediction model is trained based on exposed data and unexposed data. During the training of the CTR prediction model, the prediction results of the CTR prediction model are corrected by lowering the contribution rate of exposed data or increasing the contribution rate of unexposed data. Lowering the contribution rate of exposed data includes reducing the proportion of exposed data in the total sample data or reducing the proportion of the loss function value of exposed data in the total model loss function value. Increasing the contribution rate of unexposed data includes increasing the proportion of unexposed data in the total sample data or increasing the proportion of the loss function value of unexposed data in the total model loss function value. The data sorting module is used to sort the source information according to the click rate of each piece of data in order to determine the target information to be displayed.

Images