WO2024012219A1 - 一种模型的训练方法的相关装置 - Google Patents

一种模型的训练方法的相关装置 Download PDF

Info

Publication number
WO2024012219A1
WO2024012219A1 PCT/CN2023/103919 CN2023103919W WO2024012219A1 WO 2024012219 A1 WO2024012219 A1 WO 2024012219A1 CN 2023103919 W CN2023103919 W CN 2023103919W WO 2024012219 A1 WO2024012219 A1 WO 2024012219A1
Authority
WO
WIPO (PCT)
Prior art keywords
error
model
sample data
loss function
observed
Prior art date
Application number
PCT/CN2023/103919
Other languages
English (en)
French (fr)
Inventor
戴全宇
吴鹏
李昊轩
董振华
唐睿明
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Publication of WO2024012219A1 publication Critical patent/WO2024012219A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning

Definitions

  • the present application relates to the field of artificial intelligence technology, and in particular to a device related to a model training method.
  • the workflow of the personalized recommendation system is: use sample data for training to obtain a prediction model, then use the prediction model to predict the user's preference for recommended items, and then recommend items to the user based on the prediction results.
  • the sample data includes user behavior, product characteristics and contextual information. These sample data all have labels (that is, the degree of preference for items), and generally come from the historical interaction logs of users and the recommendation system.
  • the sample data is affected by the selection bias problem, and the distribution of the sample data is biased. Specifically, it can be understood that the sample data only contains the preferences of some users for some items, rather than the preferences of all users for all items. , that is, some data without labels are not included in the sample data.
  • the selection bias includes user selection bias and system exposure bias.
  • the prediction model obtained is also biased, causing the user's preference for items output by the prediction model to be inaccurate, affecting the success rate of recommendations.
  • a prediction model is used to predict sample data, and then a loss function is constructed based on the error between the prediction result and the label, and then the parameters of the prediction model are updated based on the loss function.
  • the error imputation model can exactly solve this problem.
  • the error imputation model can output the predicted value of the error between the prediction result and the label. This prediction value can be used to replace the error between the prediction result and the label to construct the loss. function to complete the update of parameters of the prediction model.
  • Embodiments of the present application provide a model training method and related devices for improving the generalization ability of the prediction model.
  • this application provides a model training method, including: a first error of the prediction result of the first sample data according to the prediction model, a first output of the error attribution model and an observed value of the first sample data. Probability, calculate the loss function of the error imputation model.
  • the first error refers to the difference between the prediction result of the prediction model for the first sample data and the label of the first sample data.
  • the first output of the error imputation model is used to characterize the first sample data.
  • the loss function of the error attribution model includes a bias term and a variance term. The bias term is used to characterize the prediction model.
  • the deviation between the actual loss function and the theoretical loss function, the variance term is used to characterize the variance between the actual loss function and the theoretical loss function of the prediction model; the parameters of the error imputation model are updated according to the loss function of the error imputation model.
  • the loss function of the error attribution model includes a bias term and a variance term.
  • the bias term is used to characterize the deviation between the actual loss function and the theoretical loss function of the prediction model.
  • the variance term is used to characterize the variance between the actual loss function and the theoretical loss function of the prediction model. . Since the generalization boundary of the prediction model is mainly determined by the error, bias and variance of the prediction model, and the prediction model is trained using the error imputation model, this application uses a loss function including bias terms and variance terms to train the error imputation model. , can improve the generalization ability of the prediction model and improve the stability of the prediction model, making the trained prediction model close to the unbiased model, thus improving the prediction accuracy of the model.
  • the bias term is obtained based on the first error, the first output of the error imputation model and the probability of the first sample data being observed; and/or the variance term is obtained based on the first error, the error imputation model The probability of the first output and the first sample data being observed is obtained.
  • the probability of sample data being observed and the predicted value of the error are used to train the prediction model.
  • the deviation term can be attributed to the model based on the first error and error.
  • the probability of the first output and the first sample data being observed is obtained.
  • the variance term is also obtained based on the first error and the probability of the first output of the error imputation model and the first sample data being observed.
  • the loss function of the error imputation model is obtained by performing a weighted sum of the bias term and the variance term based on the weight of the bias term and the weight of the variance term.
  • the weight of the bias term and/or the weight of the variance term is obtained based on the first sample data.
  • D represents the full sample space
  • (u,i) represents the first sample data
  • u represents the user
  • i represents the target object
  • o u,i is used to characterize whether the first sample data has been observed
  • e u,i represents the first error
  • D represents the full sample space
  • (u,i) represents the first sample data
  • u represents the user
  • i represents the target object
  • o u,i is used to characterize whether the first sample data has been observed
  • e ui represents the first error
  • the probability of the first sample data being observed is output by the target probability model based on the input first sample data, and the target probability model can be obtained through pre-training.
  • the method after updating the parameters of the error imputation model according to the loss function of the error imputation model, the method also includes: according to the second error of the prediction result of the second sample data by the prediction model, the error of the error imputation model The second output, the probability of the second sample data being observed and the indication information of the second sample data are used to calculate the loss function of the prediction model.
  • the second output of the error attribution model is used to characterize the predicted value of the second error, and the indication information is used to indicate Whether the second sample data is observed, the second sample data can be the same as the first sample data, or it can be different from the first sample data; when the second sample data is different from the first sample data, the second sample data It can be data with labels or data without labels; the parameters of the prediction model are updated according to the loss function of the prediction model.
  • the loss function of the error attribution model includes a bias term and a variance term.
  • the bias term is used to characterize the deviation between the actual loss function and the theoretical loss function of the prediction model.
  • the variance term is used to characterize the variance between the actual loss function and the theoretical loss function of the prediction model. . Since the generalization boundary of the prediction model is mainly determined by the error, bias and variance of the prediction model, using the loss function of the error imputation model to train the error imputation model, and then using the output of the error imputation model to train the prediction model, can improve the generalization of the prediction model. ization ability, improve the stability of the prediction model, and make the trained prediction model close to the unbiased model, thus improving the prediction accuracy of the model.
  • this application provides a model training device, including: a computing unit configured to calculate the first number of samples according to the prediction model. Calculate the loss function of the error imputation model based on the first error of the prediction result, the first output of the error imputation model and the probability that the first sample data is observed, and the first output of the error imputation model is used to characterize the prediction of the first error value, the loss function of the error attribution model includes a deviation term and a variance term. The deviation term is used to characterize the deviation between the actual loss function and the theoretical loss function of the prediction model. The variance term is used to characterize the difference between the actual loss function and the theoretical loss function of the prediction model. The variance of; the update unit is used to update the parameters of the error imputation model according to the loss function of the error imputation model.
  • the bias term is obtained based on the first error, the first output of the error imputation model and the probability of the first sample data being observed; and/or the variance term is obtained based on the first error, the error imputation model The probability of the first output and the first sample data being observed is obtained.
  • the loss function of the error imputation model is obtained by performing a weighted sum of the bias term and the variance term based on the weight of the bias term and the weight of the variance term.
  • the weight of the bias term and/or the weight of the variance term is obtained based on the first sample data.
  • D represents the full sample space
  • (u,i) represents the first sample data
  • u represents the user
  • i represents the target object
  • o u,i is used to characterize whether the first sample data has been observed
  • e u,i represents the first error
  • D represents the full sample space
  • (u,i) represents the first sample data
  • u represents the user
  • i represents the target object
  • o u,i is used to characterize whether the first sample data has been observed
  • e u,i represents the first error
  • the probability of the first sample data being observed is output by the target probability model based on the input first sample data.
  • the calculation unit is also used to calculate the second error of the prediction result of the second sample data according to the prediction model, the second output of the error attribution model, the probability of the second sample data being observed, and the second sample
  • the label of the data is used to calculate the loss function of the prediction model.
  • the second output of the error attribution model is used to characterize the predicted value of the second error.
  • the label of the second sample data indicates whether the user's evaluation of the object has been observed; the update unit is also used
  • the parameters of the prediction model are updated according to the loss function of the prediction model.
  • this application provides a computer device.
  • the computer device includes a memory and a processor.
  • a processor configured to execute computer programs or instructions stored in the memory, so that the computer device performs the method of any one of the first aspects.
  • the present application provides a computer-readable storage medium.
  • the computer-readable storage medium has program instructions. When the program instructions are executed directly or indirectly, any method in the first aspect is implemented.
  • the present application provides a chip system.
  • the chip system includes at least one processor.
  • the processor is configured to execute computer programs or instructions stored in the memory.
  • the computer program or instructions are executed in at least one processor, the first Either method is implemented in one aspect.
  • the present application provides a computer program product, which includes instructions that, when executed on a computer, cause the computer to perform any of the methods of the first aspect.
  • Figure 1 is a schematic diagram of the architecture of the recommendation system
  • Figure 2 is a schematic diagram of movie scoring
  • Figure 3 is a schematic diagram of the principle of the advertising recommendation system
  • Figure 4 is a schematic diagram of an embodiment of the model training method provided by the embodiment of the present application.
  • Figure 5 is a general schematic diagram of the model training method provided by the embodiment of the present application.
  • Figure 6 is a schematic diagram of the prediction network architecture of CVR in the embodiment of this application.
  • Figure 7 is a schematic structural diagram of a model training device provided by an embodiment of the present application.
  • FIG. 8 is a schematic structural diagram of a computer device provided by an embodiment of the present application.
  • At least one of a, b, or c can mean: a, b, c, a-b, a-c, b-c, or a-b-c, where a, b, c can be single or multiple .
  • the recommendation system includes: log library, offline training module and online prediction module.
  • the log library stores logs, which are used to record a series of user behavior data in the front-end display list, such as browsing, clicks, comments, downloads, etc.
  • the offline training module is used to perform offline training based on logs to obtain a prediction model.
  • the prediction model is used to predict the user's preference for a certain object. It should be noted that the embodiment of the present application does not specifically limit the type of the object, for example , the object can be news, advertisements, movies, songs, commodities, etc.
  • the prediction model is deployed online to form an online prediction module.
  • the online prediction module is used to output recommendation results based on the user's requested access, object characteristics and contextual information. As shown in Figure 1, the recommendation results are Object 1, Object 2 and Object 3 in the display list.
  • User behavior data regarding recommended results constitutes user feedback data, which will be added to the logs in the log library.
  • sample data used to train the prediction model will be affected by the selection bias problem, where the selection bias includes user selection bias and system exposure bias.
  • the horizontal side of the table represents different movies
  • the vertical side of the table represents different users.
  • the data in the table represents the user's ratings of the movies. Since the users selectively rate the movies, some of the movie ratings are is unknown.
  • the question mark in the table indicates that the user's rating for the movie is unknown. This is the user selection bias.
  • Existing training methods often only select the part of the data for which the user's rating of the movie is known as sample data for training, so the prediction model obtained by training is also biased.
  • the ad recommendation process mainly includes three links: exposure, click and conversion.
  • Exposure refers to recommending ads to users
  • clicks refer to users clicking on recommended ads
  • conversion refers to users clicking on recommended ads.
  • ads are not recommended to users randomly, but selectively.
  • click link the ads users click are limited to recommended ads, so the user’s click behavior on non-recommended ads is unknown. This is the system exposure bias.
  • users' behavioral data on recommended advertisements are used as sample data for training, so the prediction model obtained by training is also biased.
  • One method is to, in addition to using labeled data to train the prediction model, also use unlabeled data to train the prediction model; taking Figure 2 as an example, the labeled data refers to the movies that the user has rated, and the data without labels Refers to movies that have not been rated by users.
  • Another approach is to predict the sample data by predicting what is observed The distribution of the observed data is adjusted with the probability, so that the biased observed data tends to be unbiased after adjustment.
  • the training method that uses the above two methods at the same time can be called the Doubly Robust Learning method.
  • a loss function needs to be constructed based on the error between the prediction result of the prediction model and the label, and then the parameters of the prediction model are updated based on the loss function.
  • the error cannot be calculated directly.
  • the error is predicted through the Error Imputation model, a loss function is constructed using the predicted value of the error output by the Error Imputation model, and then the parameters of the prediction model are updated based on the loss function.
  • the prediction model can be trained using the full sample data (including data with labels and data without labels), so that the trained prediction model is close to unbiased.
  • embodiments of the present application provide a model training method, which belongs to the Doubly Robust Learning method. This method is used to calculate the error imputation model.
  • the loss function includes a bias term and a variance term.
  • the bias term is used to characterize the deviation between the actual loss function and the theoretical loss function of the prediction model
  • the variance term is used to characterize the variance between the actual loss function and the theoretical loss function of the prediction model.
  • the embodiment of the present application uses a loss function including a bias term and a variance term to train the error imputation model. Training can indirectly improve the generalization ability of the prediction model.
  • embodiments of this application can be applied to various scenarios.
  • embodiments of the present application can be applied to conversion rate prediction scenarios, that is, the prediction model trained by the method provided by the embodiments of the present application can be used to predict the conversion rate, where the conversion rate can be understood as the user purchasing the product after clicking on it.
  • Product probability can be applied.
  • the embodiments of the present application can be applied to click-through rate prediction scenarios, that is, the prediction model trained by the method provided by the embodiments of the present application can be used to predict the click-through rate.
  • the click-through rate can be understood as the user clicking on an item when the item is recommended.
  • the probability can be applied to various scenarios.
  • conversion rate prediction scenarios that is, the prediction model trained by the method provided by the embodiments of the present application can be used to predict the conversion rate, where the conversion rate can be understood as the user purchasing the product after clicking on it.
  • Product probability can be applied to click-through rate prediction scenarios, that is, the prediction model trained by the method provided by the embodiments of the present application can
  • the embodiment of the present application provides an embodiment of a model training method, including:
  • Step 101 Use the target probability model to output the probability that the first sample data is observed based on the input first sample data.
  • the target probability model can also be called a propensity score prediction model.
  • the first sample data is input into the target probability model, and the target probability model will output the probability of the first sample data being observed, that is, the probability of the first sample data being observed is the target probability model according to the input first Sample data output.
  • the first sample data generally includes two parts: user and target object.
  • the target object can be movies, news, products, etc.
  • the probability that the first sample data is observed can be understood as the probability that the first sample data can be obtained. Taking the first sample data as a movie as an example, the probability that the first sample data is observed can be understood as the probability that the user rates a certain movie, or in other words, the probability that the user's rating for a certain movie can be obtained.
  • the target probability model can be obtained through pre-training. Since the target probability model is a relatively mature model, the training process of the target probability model will not be described in detail here. In addition, the target probability model can also be trained together with the error attribution model and the prediction model.
  • step 101 is optional.
  • Step 102 Calculate the loss function of the error imputation model based on the first error of the prediction result of the prediction model for the first sample data, the first output of the error imputation model and the probability that the first sample data is observed.
  • the loss function includes bias terms and variance terms.
  • the first error refers to the difference between the prediction result of the prediction model for the first sample data and the label of the first sample data.
  • the first output of the error imputation model is used to characterize the predicted value of the first error.
  • the deviation term is used to characterize the deviation between the actual loss function and the theoretical loss function of the prediction model
  • the variance term is used to characterize the variance between the actual loss function and the theoretical loss function of the prediction model.
  • the deviation term and variance term can be passed through the prediction model.
  • the actual loss function and the theoretical loss function are calculated.
  • the bias term can be obtained from at least one parameter of the first error, the first output of the error imputation model, and the probability of the first sample data being observed.
  • the bias term is based on the first error, the error imputation model The first output of the model and the probability of the first sample data being observed are obtained.
  • D represents the full sample space
  • (u,i) represents the first sample data
  • u represents the user
  • i represents the target object
  • o u,i is used to characterize whether the first sample data has been observed
  • e u,i represents the first error
  • o u,i When the value of o u,i is 1, o u,i means that the first sample data has been observed; when the value of o u,i is 0, o u,i means that the first sample data has not been observed. observed.
  • the variance term can be obtained from at least one parameter of the first error, the first output of the error imputation model and the probability of the first sample data being observed.
  • the variance term is based on the first error , the first output of the error imputation model and the probability of the first sample data being observed are obtained.
  • D represents the full sample space
  • (u,i) represents the first sample data
  • u represents the user
  • i represents the target object
  • o u,i is used to characterize whether the first sample data has been observed
  • e u,i represents the first error
  • o u,i When the value of o u,i is 1, o u,i means that the first sample data has been observed; when the value of o u,i is 0, o u,i means that the first sample data has not been observed. observed.
  • the deviation term and the variance term both contain the parameter o u,i , and when the value of o u,i is 1, o u,i means that the first sample data is Observed; when the value of o u,i is 0, o u,i means that the first sample data has not been observed.
  • the value of o u,i is 0, in the deviation term is 0, similarly, in the variance term is also 0. Therefore, it can be considered that the first sample data in the embodiment of the present application are all observed sample data, that is, labeled sample data.
  • the loss function of the error imputation model can be obtained by the weighted sum of the bias term and the variance term based on the weight of the bias term and the weight of the variance term.
  • the loss function of the error imputation model can be expressed as Among them, ⁇ ⁇ (x u,i ) represents the weight of the deviation term, and x u,i represents the first sample data.
  • the weight of the bias term and/or the weight of the variance term can be hyperparameters or obtained based on the first sample data.
  • Step 103 Update the parameters of the error imputation model according to the loss function of the error imputation model.
  • the stochastic gradient descent method can be used to update the parameters of the error imputation model. Since the stochastic gradient descent method is a relatively mature technology, it will not be described in detail here.
  • Step 104 Calculate the loss function of the prediction model based on the second error of the prediction result of the prediction model for the second sample data, the second output of the error attribution model, the probability of the second sample data being observed, and the indication information of the second sample data.
  • the second output of the error imputation model is used to characterize the predicted value of the second error
  • the indication information is used to indicate whether the second sample data is observed.
  • the second sample data may be the same as the first sample data, or may be different from the first sample data; when the second sample data is different from the first sample data, the second sample data may be data with a label, or may be is data without labels.
  • the second error can be calculated; when the second sample data is data without labels, the second error cannot be calculated and can only be calculated using the 2 error prediction value.
  • the loss function of the prediction model is obtained based on the second error and the second output of the error attribution model.
  • the loss function of the prediction model can be expressed as Among them, in this formula, o u,i is used to characterize whether the second sample data is observed, e u,i represents the second error, represents the second output of the error attribution model, Indicates the probability that the second sample data is observed.
  • o u,i When the value of o u,i is 1, o u,i indicates that the second sample data has been observed; when the value of o u,i is 0, o u,i indicates that the second sample data has not been observed. .
  • Step 105 Update the parameters of the prediction model according to the loss function of the prediction model.
  • the stochastic gradient descent method can be used to update the parameters of the prediction model. Since the stochastic gradient descent method is a relatively mature technology, it will not be described in detail here.
  • the prediction model can be any prediction model.
  • the prediction model can be a collaborative filtering model based on matrix decomposition (Matrix Factorization Collaborative Filtering Model, MF), a collaborative filtering model based on neural network (Neural Collaborative Filtering Model, NCF), a A Factorization-Machine based Neural Network for CTR Prediction (DeepFM) model based on factorization machine and a Deep&Cross Network for Ad Click Predictions (DCN) model for ad click prediction.
  • matrix decomposition Matrix Factorization Collaborative Filtering Model, MF
  • NCF neural network
  • DeepFM A Factorization-Machine based Neural Network for CTR Prediction
  • DCN Deep&Cross Network for Ad Click Predictions
  • the parameter update process of the error attribution model and the parameter update process of the prediction model are generally carried out alternately. That is, the parameters of the prediction model are first kept unchanged, and some sample data are used to pass steps 101 to 103. Update the parameters of the error attribution model, keep the parameters of the error attribution model unchanged, and then use a part of the sample data to update the parameters of the prediction model through step 104.
  • the loss function of the error attribution model includes a deviation term and a variance term.
  • the deviation term is used to characterize the deviation between the actual loss function and the theoretical loss function of the prediction model
  • the variance term is used to characterize the actual loss function of the prediction model. and the variance between the theoretical loss function. Since the generalization boundary of the prediction model is mainly determined by the error, bias and variance of the prediction model, and the prediction model is trained using the error imputation model, the embodiment of the present application uses a loss function including a bias term and a variance term to train the error imputation model. Training can improve the generalization ability of the prediction model and improve the stability of the prediction model, making the trained prediction model close to the unbiased model, thus improving the prediction accuracy of the model.
  • the prediction model can be trained by using data without labels to make the prediction model close to unbiased, or the prediction model can be made close to unbiased by adjusting the distribution of the observed data, and this application
  • the embodiment simultaneously uses the probability of sample data being observed and the predicted value of the error to train the prediction model and establish the loss function of the error attribution model, which can make the prediction model after training closer to unbiased and further improve the prediction accuracy of the model.
  • the method provided by the embodiment of the present application can be summarized as: feature extraction and prediction of sample data in the log, where prediction refers to prediction of sample data and prediction of sample data.
  • Predict the error of the result then calculate the bias term and variance term based on the predicted value of the prediction result and the error of the prediction result, and then perform a weighted sum of the bias term and variance term to obtain the loss function of the error attribution model; you can also use the prediction
  • the loss function of the prediction model is calculated based on the predicted value of the error of the result and the prediction result; finally, the loss function of the error imputation model and the loss function of the prediction model are used to alternately train the error imputation model and the prediction model to perform the parameters of the error imputation model and the prediction model. renew.
  • the method provided by the embodiment of this application is applied to the CVR prediction network architecture shown in Figure 6.
  • the network architecture includes three parts.
  • the first part of the network architecture is the click through rate (CTR).
  • the prediction model is used to estimate CTR;
  • the second part of the network architecture is the prediction model of conversion rate (Post-click Conversion Rate, CVR), which is used to estimate CVR;
  • the third part of the network architecture is the error attribution model.
  • Predictions for error attribution are all a Factorization-Machine based Neural Network for CTR Prediction (DeepFM) used for click rate prediction. It should be noted that DeepFM can also Use other neural networks instead.
  • DeepFM Factorization-Machine based Neural Network for CTR Prediction
  • the CTR prediction model, the CVR prediction model, and the error attribution model all include: a representation layer, a deep model tower, and a decomposition machine model tower. Among them, the CTR prediction model and the CVR prediction model share the representation layer.
  • the representation layer is used to convert discrete features (including user features and item features) into sample representations (i.e., low-dimensional vectors).
  • the deep model tower and decomposition machine model tower are used to output estimated results based on the input sample representations.
  • the prediction result of the CTR prediction model is CTR, which can be used as the probability of sample data being observed in the CVR prediction process;
  • the prediction result of the CVR prediction model is CVR, and
  • the prediction result of the error attribution model is Error prediction of CVR.
  • the error attribution model loss function can be established based on the error predictions of CTR and CVR, and the loss function can be established based on the error predictions of CTR, CVR and CVR.
  • CTR estimates a dual-robust loss function, and then trains an error attribution model and a CVR prediction model based on these two loss functions.
  • a specific example is provided below. In this example, the method provided by the embodiment of the present application is compared with existing training methods.
  • CTR refers to the click through rate (Click Through Rate)
  • CVR refers to the conversion rate (Post-click Conversion Rate)
  • CTCVR refers to the click conversion rate (Post-view Click- through&Conversion Rate).
  • DCN Deep&Cross Network for Ad Click Predictions
  • ESMM Entire Space Multi-Task Model
  • DR-JL Double Robust Joint Learning.
  • Multi-IPW refers to the multi-task Inverse Propensity Weighting CVR Estimator (Multi-task Inverse Propensity Weighting CVR Estimator)
  • Multi-DR refers to the multi-task doubly robust CVR estimator (Multi-task Doubly Robust CVR Estimator)
  • MRDR Multi-task Doubly Robust CVR Estimator
  • MRDR Multi-task Doubly Robust CVR Estimator
  • MRDR Multi-task Doubly Robust CVR Estimator
  • MRDR Multi-task Doubly Robust CVR Estimator
  • MRDR More Robust Doubly Robust
  • DR-BIAS refers to the method of training the deviation term in the embodiment of this application as the loss function of the error attribution model
  • DR-MSE is Refers to the training method shown in Figure
  • this application also provides an embodiment of a model training device, including: a calculation unit 201, configured to calculate the first error of the prediction result of the first sample data according to the prediction model and the error attribution model.
  • the probability of the first output and the first sample data being observed is used to calculate the loss function of the error imputation model.
  • the first output of the error imputation model is used to represent the predicted value of the first error.
  • the loss function of the error imputation model includes the bias term and variance.
  • the deviation term is used to characterize the deviation between the actual loss function and the theoretical loss function of the prediction model
  • the variance term is used to characterize the variance between the actual loss function and the theoretical loss function of the prediction model
  • the update unit 202 is used to blame the model according to the error
  • the loss function updates the parameters of the error attribution model.
  • the bias term is obtained based on the first error, the first output of the error imputation model and the probability of the first sample data being observed; and/or the variance term is obtained based on the first error, the error imputation model The probability of the first output and the first sample data being observed is obtained.
  • the loss function of the error imputation model is obtained by performing a weighted sum of the bias term and the variance term based on the weight of the bias term and the weight of the variance term.
  • the weight of the bias term and/or the weight of the variance term is obtained based on the first sample data.
  • D represents the full sample space
  • (u,i) represents the first sample data
  • u represents the user
  • i represents the target object
  • o u,i is used to characterize whether the first sample data has been observed
  • e u,i represents the first error
  • D represents the full sample space
  • (u,i) represents the first sample data
  • u represents the user
  • i represents the target object
  • o u,i is used to characterize whether the first sample data has been observed
  • e u,i represents the first error
  • the probability of the first sample data being observed is output by the target probability model based on the input first sample data.
  • the calculation unit 201 is also used to calculate the second error of the prediction result of the second sample data according to the prediction model, the second output of the error attribution model, the probability of the second sample data being observed, and the second
  • the label of the sample data is used to calculate the loss function of the prediction model, the second output of the error attribution model is used to characterize the predicted value of the second error, and the label of the second sample data indicates whether the user's evaluation of the object has been observed; update unit 202, It is also used to update the parameters of the prediction model based on the loss function of the prediction model.
  • FIG. 8 is a schematic structural diagram of a computer device provided by an embodiment of the present application.
  • the computer device 900 is usually a server equipped with the above-mentioned model training device.
  • Computer device 900 is implemented by a general bus architecture.
  • Computer device 900 includes at least one processor 901, a communication bus 902, a memory 903, and at least one communication interface 904.
  • the processor 901 is a general central processing unit (CPU), a network processor (NP), a microprocessor, or one or more integrated circuits used to implement the solution of the present application. , for example, application-specific integrated circuit (ASIC), programmable logic device (PLD) or a combination thereof.
  • ASIC application-specific integrated circuit
  • PLD programmable logic device
  • the above-mentioned PLD is a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a general array logic (GAL) or any combination thereof.
  • Communication bus 902 is used to transfer information between the above-mentioned components.
  • the communication bus 902 is divided into an address bus, a data bus, a control bus, etc.
  • address bus a data bus
  • control bus a control bus
  • only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.
  • the memory 903 is a read-only memory (ROM) or other type of static storage device that can store static information and instructions.
  • memory 903 is random access memory (RAM) or other types of dynamic storage devices that can store information and instructions.
  • the memory 903 is electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disk storage, optical disc storage ( Including compressed optical discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or can be used to carry or store desired program code in the form of instructions or data structures and can be used by a computer Any other medium for access, but not limited to this.
  • the memory 903 exists independently and is connected to the processor 901 through the communication bus 902.
  • the memory 903 and the processor 901 are integrated together.
  • Communication interface 904 uses any transceiver-like device for communicating with other devices or communication networks.
  • Communication interface 904 includes wired Communication Interface.
  • the communication interface 904 also includes a wireless communication interface.
  • the wired communication interface is, for example, an Ethernet interface.
  • the Ethernet interface is an optical interface, an electrical interface, or a combination thereof.
  • the wireless communication interface is a wireless local area networks (WLAN) interface, a cellular network communication interface or a combination thereof, etc.
  • WLAN wireless local area networks
  • the processor 901 includes one or more CPUs, such as CPU0 and CPU1 as shown in FIG. 8 .
  • the computer device 900 includes multiple processors, such as the processor 901 and the processor 905 shown in FIG. 8 .
  • processors are a single-core processor (single-CPU), or a multi-core processor (multi-CPU).
  • a processor here refers to one or more devices, circuits, and/or processing cores for processing data (such as computer program instructions).
  • the memory 903 is used to store the program code 99 for executing the solution of the present application, and the processor 901 executes the program code 99 stored in the memory 903. That is to say, the computer device 900 implements the above method embodiments through the processor 901 and the program code 99 in the memory 903 .
  • An embodiment of the present application also provides a chip including one or more processors. Some or all of the processor is used to read and execute the computer program stored in the memory to perform the methods of the aforementioned embodiments.
  • the chip should include a memory, and the memory and the processor are connected to the memory through circuits or wires. Further optionally, the chip also includes a communication interface, and the processor is connected to the communication interface.
  • the communication interface is used to receive data and/or information that needs to be processed.
  • the processor obtains the data and/or information from the communication interface, processes the data and/or information, and outputs the processing results through the communication interface.
  • the communication interface may be an input-output interface.
  • some of the one or more processors may implement part of the steps in the above method through dedicated hardware. For example, processing involving a neural network model may be performed by a dedicated neural network processor. Or graphics processor to achieve.
  • the method provided by the embodiments of this application can be implemented by one chip, or can be implemented by multiple chips collaboratively.
  • Embodiments of the present application also provide a computer storage medium, which is used to store computer software instructions used for the above-mentioned computer equipment, including for executing programs designed for the computer equipment.
  • the computer device may correspond to the functions of the model training device in the embodiment as shown in the aforementioned FIG. 7 .
  • Embodiments of the present application also provide a computer program product.
  • the computer program product includes computer software instructions.
  • the computer software instructions can be loaded by a processor to implement the processes in the methods shown in the foregoing embodiments.

Abstract

本申请实施例公开了一种模型的训练方法和相关装置，用于提高预测模型的泛化能力。本申请实施例方法包括：根据预测模型对于第一样本数据的预测结果的第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率，计算错误归咎模型的损失函数，然后根据错误归咎模型的损失函数对错误归咎模型的参数进行更新，其中，错误归咎模型的第一输出用于表征第一误差的预测值，错误归咎模型的损失函数包括偏差项和方差项，偏差项用于表征预测模型的实际损失函数与理论损失函数间的偏差，方差项用于表征预测模型的实际损失函数与理论损失函数间的方差。

Description

一种模型的训练方法的相关装置
本申请要求于2022年7月15日提交中国专利局、申请号为202210836072.4、发明名称为“一种模型的训练方法的相关装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能技术领域，尤其涉及一种模型的训练方法的相关装置。
背景技术
随着人工智能技术的发展，个性化推荐系统得到了越来越广泛的应用。
个性化推荐系统的工作流程是：利用样本数据进行训练，以得到预测模型，然后利用预测模型预测用户对待推荐的物品的喜好程度，然后根据预测的结果为用户推荐物品。其中，样本数据包括用户行为、商品特征和上下文信息。这些样本数据都是具有标签(即对物品的喜好程度)的，一般来自于用户和推荐系统的历史交互日志中。
然而，样本数据受到选择偏置问题的影响，样本数据的分布是有偏的，具体可以理解为，样本数据仅包含部分用户对部分物品的喜好程度，而不是包含所有用户对所有物品的喜好程度，即部分不具有标签的数据未包含在样本数据中。其中，选择偏置包括用户选择偏置和系统曝光偏置。
以用户选择偏置为例，在电影评分网站中，用户倾向于对自己特别喜欢或不喜欢的电影进行评分，所以采集到的样本数据大多是用户对自己特别喜欢或不喜欢的电影的评分，而缺少用户对喜欢程度一般的电影的评分。
而在有偏的样本数据上训练，得到的预测模型也是有偏的，导致预测模型输出的用户对物品的喜好程度不准确，影响推荐的成功率。
为了解决上述问题，除了具有标签的样本数据外，还需要利用不具有标签的数据对预测模型进行训练。目前，能够利用具有标签的数据对预测模型进行训练的方法主要是引入错误归咎(Error Imputation)模型。
在普通的学习方法中，都是利用预测模型对样本数据进行预测，然后根据预测结果和标签之间的误差构建损失函数，再根据损失函数对预测模型的参数进行更新。而对于不具有标签的数据来说，预测结果和标签之间的误差是无法直接计算的。而错误归咎(Error Imputat ion)模型恰好能够解决这个问题，错误归咎模型可以输出预测结果和标签之间的误差的预测值，可以利用该预测值代替预测结果和标签之间的误差，以构建损失函数，从而完成对预测模型的参数的更新。
然而，利用已有的错误归咎模型训练得到的预测模型的泛化能力较弱，这导致预测模型对于某些输入数据的预测结果的准确率较低。因此，如何通过训练错误归咎模型，以提高预测模型的泛化能力是亟需解决的问题。
发明内容
本申请实施例提供了一种模型的训练方法和相关装置，用于提高预测模型的泛化能力。
第一方面，本申请提供了一种模型的训练方法，包括：根据预测模型对于第一样本数据的预测结果的第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率，计算错误归咎模型的损失函数，第一误差是指预测模型对于第一样本数据的预测结果与第一样本数据的标签间的差值，错误归咎模型的第一输出用于表征第一误差的预测值，错误归咎模型的损失函数包括偏差项和方差项，偏差项用于表征预测模型 的实际损失函数与理论损失函数间的偏差，方差项用于表征预测模型的实际损失函数与理论损失函数间的方差；根据错误归咎模型的损失函数对错误归咎模型的参数进行更新。
错误归咎模型的损失函数包括偏差项和方差项，偏差项用于表征预测模型的实际损失函数与理论损失函数间的偏差，方差项用于表征预测模型的实际损失函数与理论损失函数间的方差。由于预测模型的泛化边界主要由预测模型的误差、偏差和方差决定，而预测模型是利用错误归咎模型训练得到的，所以本申请利用包括偏差项和方差项的损失函数对错误归咎模型进行训练，能够提高预测模型的泛化能力，提高预测模型的稳定性，使得训练后的预测模型接近无偏模型，从而提升模型的预测准确度。
作为一种可实现的方式，偏差项是根据第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率得到的；和/或方差项是根据第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率得到的。
在双稳健学习(Doubly Robust Learning)方法中，会同时利用样本数据被观测的概率以及误差的预测值来训练预测模型，而在本申请中，偏差项是可根据第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率得到的，方差项也是根据第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率得到的，根据双稳健学习方法的优势可知，本申请的方法能够进一步提高预测模型的泛化能力，使得训练后的预测模型更接近无偏模型，进一步地提升模型的预测准确度。
作为一种可实现的方式，错误归咎模型的损失函数是根据偏差项的权重和方差项的权重，对偏差项和方差项进行加权求和得到的。
作为一种可实现的方式，偏差项的权重和/或方差项的权重是根据第一样本数据得到的。
根据第一样本数据动态计算偏差项的权重和方差项的权重，使得偏差项的权重和方差项的权重能够适应第一样本数据，提高偏差项的权重和方差项的权重的准确性，从而提升错误归咎模型的损失函数的准确性，提升对错误归咎模型的训练效果，从而提升预测模型的广泛性。
作为一种可实现的方式，偏差项为其中，D表示全样本空间,(u,i)表示第一样本数据，u表示用户，i表示目标对象，ou,i用于表征第一样本数据是否被观测到，表示错误归咎模型的第一输出，eu,i表示第一误差，表示第一样本数据被观测的概率；当ou,i的取值为1时，ou,i表示第一样本数据被观测到；当ou,i的取值为0时，ou,i表示第一样本数据未被观测到。
作为一种可实现的方式，方差项为其中，D表示全样本空间,(u,i)表示第一样本数据，u表示用户，i表示目标对象，ou,i用于表征第一样本数据是否被观测到，表示错误归咎模型的第一输出，eui表示第一误差，表示第一样本数据被观测的概率；当ou,i的取值为1时，ou,i表示第一样本数据被观测到；当ou,i的取值为0时，ou,i表示所述第一样本数据未被观测到。
作为一种可实现的方式，第一样本数据被观测的概率是目标概率模型根据输入的第一样本数据输出的，目标概率模型可以通过预先的训练得到。
作为一种可实现的方式，在根据错误归咎模型的损失函数对错误归咎模型的参数进行更新之后，方法还包括：根据预测模型对于第二样本数据的预测结果的第二误差、错误归咎模型的第二输出、第二样本数据被观测的概率以及第二样本数据的指示信息，计算预测模型的损失函数，错误归咎模型的第二输出用于表征第二误差的预测值，指示信息用于指示第二样本数据是否被观测到，第二样本数据可以与第一样本数据相同，也可以与第一样本数据不同；当第二样本数据与第一样本数据不同时，第二样本数据可以是具有标签的数据，也可以是不具有标签的数据；根据预测模型的损失函数对预测模型的参数进行更新。
错误归咎模型的损失函数包括偏差项和方差项，偏差项用于表征预测模型的实际损失函数与理论损失函数间的偏差，方差项用于表征预测模型的实际损失函数与理论损失函数间的方差。由于预测模型的泛化边界主要由预测模型的误差、偏差和方差决定，所以利用错误归咎模型的损失函数训练得到错误归咎模型，再利用错误归咎模型的输出训练预测模型，能够提高预测模型的泛化能力，提高预测模型的稳定性，使得训练后的预测模型接近无偏模型，从而提升模型的预测准确度。
第二方面，本申请提供了一种模型的训练装置，包括：计算单元，用于根据预测模型对于第一样本数 据的预测结果的第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率，计算错误归咎模型的损失函数，错误归咎模型的第一输出用于表征第一误差的预测值，错误归咎模型的损失函数包括偏差项和方差项，偏差项用于表征预测模型的实际损失函数与理论损失函数间的偏差，方差项用于表征预测模型的实际损失函数与理论损失函数间的方差；更新单元，用于根据错误归咎模型的损失函数对错误归咎模型的参数进行更新。
作为一种可实现的方式，偏差项是根据第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率得到的；和/或方差项是根据第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率得到的。
作为一种可实现的方式，错误归咎模型的损失函数是根据偏差项的权重和方差项的权重，对偏差项和方差项进行加权求和得到的。
作为一种可实现的方式，偏差项的权重和/或方差项的权重是根据第一样本数据得到的。
作为一种可实现的方式，偏差项为其中，D表示全样本空间,(u,i)表示第一样本数据，u表示用户，i表示目标对象，ou,i用于表征第一样本数据是否被观测到，表示错误归咎模型的第一输出，eu,i表示第一误差，表示第一样本数据被观测的概率；当ou,i的取值为1时，ou,i表示第一样本数据被观测到；当ou,i的取值为0时，ou,i表示第一样本数据未被观测到。
作为一种可实现的方式，方差项为其中，D表示全样本空间,(u,i)表示第一样本数据，u表示用户，i表示目标对象，ou,i用于表征第一样本数据是否被观测到，表示错误归咎模型的第一输出，eu,i表示第一误差，表示第一样本数据被观测的概率；当ou,i的取值为1时，ou,i表示第一样本数据被观测到；当ou,i的取值为0时，ou,i表示所述第一样本数据未被观测到。
作为一种可实现的方式，第一样本数据被观测的概率是目标概率模型根据输入的第一样本数据输出的。
作为一种可实现的方式，计算单元，还用于根据预测模型对于第二样本数据的预测结果的第二误差、错误归咎模型的第二输出、第二样本数据被观测的概率以及第二样本数据的标签，计算预测模型的损失函数，错误归咎模型的第二输出用于表征第二误差的预测值，第二样本数据的标签指示用户对对象的评价是否被观测到；更新单元，还用于根据预测模型的损失函数对预测模型的参数进行更新。
第三方面，本申请提供了一种计算机设备，计算机设备包括：存储器和处理器。处理器，用于执行存储器中存储的计算机程序或指令，以使计算机设备执行如第一方面中任一项的方法。
第四方面，本申请提供了一种计算机可读存储介质，计算机可读存储介质具有程序指令，当程序指令被直接或者间接执行时，使得第一方面中任一的方法被实现。
第五方面，本申请提供了一种芯片系统，芯片系统包括至少一个处理器，处理器用于执行存储器中存储的计算机程序或指令，当计算机程序或指令在至少一个处理器中执行时，使得第一方面中任一项的方法被实现。
第六方面，本申请提供了一种计算机程序产品，包括指令，当指令在计算机上运行时，使得计算机执行第一方面中任一项的方法。
附图说明
图1为推荐系统的架构示意图；
图2为电影评分的示意图；
图3为广告推荐系统的原理示意图；
图4为本申请实施例提供的模型的训练方法的一个实施例的示意图；
图5为本申请实施例提供的模型的训练方法的概括示意图；
图6为本申请实施例中的CVR的预测网络架构示意图；
图7为本申请实施例提供的模型的训练装置的结构示意图；
图8为本申请实施例提供的一种计算机设备的结构示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述。显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
本申请的说明书和权利要求书及上述附图中的术语“第一”、第二”以及相应术语标号等是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。应该理解这样使用的术语在适当情况下可以互换，这仅仅是描述本申请的实施例中对相同属性的对象在描述时所采用的区分方式。此外，术语“包括”和“具有”以及他们的任何变形，意图在于覆盖不排他的包含，以便包含一系列单元的过程、方法、系统、产品或设备不必限于那些单元，而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它单元。
在本申请的描述中，除非另有说明，“/”表示或的意思，例如，A/B可以表示A或B；本申请中的“和/或”仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。另外，在本申请的描述中，“至少一项”是指一项或者多项，“多项”是指两项或两项以上。“以下至少一项(个)”或其类似表达，是指的这些项中的任意组合，包括单项(个)或复数项(个)的任意组合。例如，a,b,或c中的至少一项(个)，可以表示：a,b,c,a-b,a-c,b-c,或a-b-c，其中a,b,c可以是单个，也可以是多个。
请参阅图1，图1为推荐系统的架构示意图。该推荐系统包括：日志库、离线训练模块以及线上预测模块。
日志库中存储有日志，该日志用于记录用户在前端展示列表中的一系列的行为数据，如浏览、点击、评论、下载等。
离线训练模块用于根据日志进行离线训练，以得到预测模型，预测模型用于预测用户对某一对象的喜好程度，需要说明的是，本申请实施例对该对象的种类不做具体限定，例如，该对象可以是新闻、广告、电影、歌曲、商品等。预测模型部署于在线上，形成线上预测模块。
线上预测模块用于基于用户的请求访问、对象特征和上下文信息输出推荐结果，如图1所示，推荐结果为展示列表中的对象1、对象2和对象3。
用户对于推荐结果的行为数据构成用户反馈数据，又会被添加到日志库中的日志。
本申请实施例可以应用于图1中的离线训练模块。
应理解，用于训练预测模型的样本数据会受到选择偏置问题的影响，其中，选择偏置包括用户选择偏置和系统曝光偏置。
为了便于理解，下面结合图2和图3对用户选择偏置和系统曝光偏置进行说明。
如图2所示，表格的横向表示不同的电影，表格的纵向表示不同的用户，表格中的数据表示用户对电影的评分，由于用户是有选择地对电影进行评分，所以导致部分电影的评分是未知的，表格中的问号表示该用户对该电影的评分是未知的，这便是用户选择偏置。已有的训练方法往往只选择用户对电影的评分为已知的这部分数据作为样本数据进行训练，所以训练得到的预测模型也是有偏的。
如图3所示，在广告推荐过程中，主要包括曝光、点击和转化三个环节，曝光是指将广告推荐给用户，点击是指用户点击了推荐的广告，转换是指用户在点击推荐的广告后购买了广告中的商品。在曝光环节，广告并不是随机推荐给用户，而是有选择性地推荐给用户，在点击环节，用户点击的广告仅限于推荐的广告，所以用户对未推荐的广告的点击行为是未知的，这便是系统曝光偏置。在已有的训练方法中，会利用用户对推荐的广告的行为数据作为样本数据进行训练，所以训练得到的预测模型也是有偏的。
由此可见，若要实现预测模型无偏，可以采用两种方法。一种方法是，除了利用具有标签的数据训练预测模型外，还利用不具有标签的数据训练预测模型；以图2为例，具有标签的数据是指用户已评分的电影，不具有标签的数据是指用户未评分的电影。另一种方法是，通过预测样本数据被观测到 的概率来调整被观测到的数据的分布，使得有偏的被观测到的数据在经过调整后趋向无偏。同时采用上述两种方法的训练方法可以称为双稳健学习(Doubly Robust Learning)方法。
应理解，在训练过程中，需要根据预测模型的预测结果和标签之间的误差构建损失函数，再根据损失函数对预测模型的参数进行更新。而对于不具有标签的数据来说，该误差是无法直接计算的。为此，通过错误归咎(Error Imputation)模型对该误差进行预测，并利用将错误归咎模型输出的误差的预测值构建损失函数，再根据损失函数对预测模型的参数进行更新。这样，便可以利用全样本数据(包括具有标签的数据和不具有标签的数据)对预测模型进行训练，从而使得训练得到的预测模型接近无偏的。
为了提高错误归咎模型的性能，以提高预测模型的泛化能力，本申请实施例提供了一种模型的训练方法，该方法属于双稳健学习(Doubly Robust Learning)方法，使用该方法计算错误归咎模型的损失函数包括偏差项和方差项，偏差项用于表征预测模型的实际损失函数与理论损失函数间的偏差，方差项用于表征预测模型的实际损失函数与理论损失函数间的方差。由于预测模型的泛化边界主要由预测模型的误差、偏差和方差决定，而预测模型是利用错误归咎模型训练得到的，所以本申请实施例利用包括偏差项和方差项的损失函数对错误归咎模型进行训练，能够间接提高预测模型的泛化能力。
本申请实施例可以应用于多种场景。例如，本申请实施例可以应用于转化率预测场景，即可以利用本申请实施例提供的方法训练出的预测模型预测转化率，其中，转化率可以理解为用户在点击了商品的情况下购买该商品的概率。本申请实施例可以应用于点击率预测场景，即可以利用本申请实施例提供的方法训练出的预测模型预测点击率，其中，点击率可以理解为在推荐了某物品的情况下用户点击该物品的概率。
下面对本申请实施例提供的模型的训练方法进行具体介绍。
如图4所示，本申请实施例提供了一种模型的训练方法的一个实施例，包括：
步骤101，通过目标概率模型根据输入的第一样本数据输出第一样本数据被观测的概率。
目标概率模型也可以称为倾向性得分预测模型。
具体地，将第一样本数据输入到目标概率模型中，目标概率模型会输出第一样本数据被观测的概率，即第一样本数据被观测的概率是目标概率模型根据输入的第一样本数据输出的。
第一样本数据一般包括用户和目标对象两部分，目标对象可以是电影、新闻、商品等。
第一样本数据被观测的概率可以理解为能够获取到第一样本数据的概率。以第一样本数据是电影为例，第一样本数据被观测的概率可以理解为，用户对某一电影评分的概率，或者说，能够获取到用户对某一电影的评分的概率。
需要说明的是，目标概率模型可以通过预先训练得到，由于目标概率模型是较为成熟的模型，故在此不对目标概率模型的训练过程进行详细说明。除此之外，目标概率模型也可以与错误归咎模型、预测模型一同训练。
由于第一样本数据被观测的概率也可以是预先设定好的，所以步骤101是可选的。
步骤102，根据预测模型对于第一样本数据的预测结果的第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率，计算错误归咎模型的损失函数，错误归咎模型的损失函数包括偏差项和方差项。
其中，第一误差是指预测模型对于第一样本数据的预测结果与第一样本数据的标签间的差值。
错误归咎模型的第一输出用于表征第一误差的预测值。
偏差项用于表征预测模型的实际损失函数与理论损失函数间的偏差，方差项用于表征预测模型的实际损失函数与理论损失函数间的方差，其中，该偏差项和方差项可以通过预测模型的实际损失函数与理论损失函数计算得到。
偏差项可以由第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率中的至少一个参数得到，作为一种可实现的方式，偏差项是根据第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率得到的。
例如，偏差项为其中，D表示全样本空间,(u,i)表示第一样本数据，u表示用户，i表示目标对象，ou,i用于表征第一样本数据是否被观测到，表示错误归咎模型的第一输出，eu,i表示第一误差，表示第一样本数据被观测的概率；
当ou,i的取值为1时，ou,i表示第一样本数据被观测到；当ou,i的取值为0时，ou,i表示第一样本数据未被观测到。
同样地，方差项可以由第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率中的至少一个参数得到，作为一种可实现的方式，方差项是根据第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率得到的。
例如，方差项为其中，D表示全样本空间,(u,i)表示第一样本数据，u表示用户，i表示目标对象，ou,i用于表征第一样本数据是否被观测到，表示错误归咎模型的第一输出，eu,i表示第一误差，表示第一样本数据被观测的概率；
当ou,i的取值为1时，ou,i表示第一样本数据被观测到；当ou,i的取值为0时，ou,i表示第一样本数据未被观测到。
根据偏差项和方差项的公式中可以看出，偏差项和方差项都包含参数ou,i，且当ou,i的取值为1时，ou,i表示第一样本数据被观测到；当ou,i的取值为0时，ou,i表示所述第一样本数据未被观测到。当ou,i的取值为0时，偏差项中的为0，同样地，方差项中的也为0，因此，可以认为本申请实施例中的第一样本数据都是被观测到的样本数据，即带标签的样本数据。
错误归咎模型的损失函数可以是根据偏差项的权重和方差项的权重，对偏差项和方差项进行加权求和得到的。
相应地，错误归咎模型的损失函数可以表示为其中，λξ(xu,i)表示偏差项的权重，xu,i表示第一样本数据。
偏差项的权重和/或方差项的权重可以是超参数，也可以是根据第一样本数据得到的。
步骤103，根据错误归咎模型的损失函数对错误归咎模型的参数进行更新。
具体地，利用错误归咎模型的损失函数，可以采用随机梯度下降的方法更新错误归咎模型的参数，由于随机梯度下降的方法是较成熟的技术，故在此不做详述。
步骤104，根据预测模型对于第二样本数据的预测结果的第二误差、错误归咎模型的第二输出、第二样本数据被观测的概率以及第二样本数据的指示信息，计算预测模型的损失函数，错误归咎模型的第二输出用于表征第二误差的预测值，指示信息用于指示第二样本数据是否被观测到。
第二样本数据可以与第一样本数据相同，也可以与第一样本数据不同；当第二样本数据与第一样本数据不同时，第二样本数据可以是具有标签的数据，也可以是不具有标签的数据。
需要说明的是，当第二样本数据是具有标签的数据时，则可以计算得到第二误差；当第二样本数据是不具有标签的数据时，第二误差是无法计算的，只能利用第二误差的预测值。而对于全样本空间来说，既包含具有标签的数据，也包含不具有标签的数据，所以认为预测模型的损失函数是根据第二误差和错误归咎模型的第二输出得到的。
具体地，预测模型的损失函数可以表示为 其中，在该公式中，ou,i用于表征第二样本数据是否被观测到，eu,i表示第二误差，表示错误归咎模型的第二输出，表示第二样本数据被观测的概率。
当ou,i的取值为1时，ou,i表示第二样本数据被观测到；当ou,i的取值为0时，ou,i表示第二样本数据未被观测到。
从公式中可以看出，当ou,i的取值为0时，第二样本数据未被观测到(即不具有标签)，公式中的的取值为0，此时不需要第二误差eu,i，利用第二误差的预测值即可完成计算；当ou,i的取值为1时，第 二样本数据被观测到(即具有标签)，公式中的的取值为0，此时不需要第二误差的预测值利用第二误差eu,i即可完成计算。
步骤105，根据预测模型的损失函数对预测模型的参数进行更新。
具体地，根据预测模型的损失函数，可以采用随机梯度下降的方法更新预测模型的参数，由于随机梯度下降的方法是较成熟的技术，故在此不做详述。
其中，预测模型可以是任意预测模型，例如，预测模型可以是基于矩阵分解的协同过滤模型(Matrix Factorization Collaborative Filtering Model，MF)、基于神经网络的协同过滤模型(Neural Collaborative Filtering Model，NCF)、一种基于分解机的神经网络用于点击率预测(A Factorization-Machine based Neural Network for CTR Prediction，DeepFM)模型和用于广告点击预测的深度交叉网络(Deep&Cross Network for Ad Click Predictions，DCN)模型。
需要说明的是，在训练过程中，错误归咎模型的参数更新过程和预测模型的参数更新过程一般是交替进行的，即先保持预测模型的参数不变，利用部分样本数据通过步骤101至步骤103对错误归咎模型的参数进行更新，然后保持错误归咎模型的参数不变，再利用一部分样本数据通过步骤104对预测模型的参数进行更新。
在本申请实施例中，错误归咎模型的损失函数包括偏差项和方差项，偏差项用于表征预测模型的实际损失函数与理论损失函数间的偏差，方差项用于表征预测模型的实际损失函数与理论损失函数间的方差。由于预测模型的泛化边界主要由预测模型的误差、偏差和方差决定，而预测模型是利用错误归咎模型训练得到的，所以本申请实施例利用包括偏差项和方差项的损失函数对错误归咎模型进行训练，能够提高预测模型的泛化能力，提高预测模型的稳定性，使得训练后的预测模型接近无偏模型，从而提升模型的预测准确度。
并且，基于前文说明可知，可以通过利用不具有标签的数据训练预测模型进行训练，以使得预测模型接近无偏，也可以通过调整被观测到的数据的分布使得预测模型接近无偏，而本申请实施例同时使用了样本数据被观测的概率以及误差的预测值来训练预测模型建立错误归咎模型的损失函数，能够使得训练后预测模型更接近无偏，进一步地提升模型的预测准确度。基于上文的说明，如图5所示，本申请实施例提供的方法可以概括为：对日志中样本数据进行特征提取和预测，这里的预测是指对样本数据进行预测以及对样本数据的预测结果的误差进行预测；然后根据预测结果和预测结果的误差的预测值计算偏差项和方差项，再对偏差项和方差项进行加权求和，以得到错误归咎模型的损失函数；还可以根据预测结果和预测结果的误差的预测值计算预测模型的损失函数；最后利用错误归咎模型的损失函数和预测模型的损失函数交替训练错误归咎模型和预测模型，以对错误归咎模型和预测模型的参数进行更新。
为了便于理解，下面提供了一个应用例。
在该应用例中，将本申请实施例提供的方法应用于图6所示的CVR的预测网络架构中，该网络架构包括三部分，网络架构的第一部分是点击率(Click Through Rate，CTR)预测模型，用于CTR的预估；网络架构的第二部分是转化率(Post-click Conversion Rate，CVR)的预测模型，用于CVR的预估；网络架构的第三部分是错误归咎模型，用于错误归咎的预估。CTR预测模型、CVR预测模型以及错误归咎模型都是用于点击率预测的一种基于分解机的神经网络(A Factorization-Machine based Neural Network for CTR Prediction，DeepFM)，需要说明的是，DeepFM也可以采用其他神经网络替换。
CTR预测模型、CVR预测模型以及错误归咎模型都包括：表征层、深度模型塔和分解机模型塔，其中，CTR预测模型和CVR预测模型共享表征层。表征层用于将离散的特征(包括用户特征和物品特征)转换成样本表征(即低维的向量)，深度模型塔和分解机模型塔用于根据输入的样本表征输出预估结果。
从图6中可以看出，CTR预测模型的预估结果为CTR，CTR可以作为CVR预测过程中样本数据被观测的概率；CVR预测模型的预估结果为CVR，错误归咎模型的预估结果为CVR的误差预测。
采用本申请实施例提供的方法，由于将CTR作为CVR预测过程中样本数据被观测的概率，所以可以根据CTR和CVR的误差预测建立错误归咎模型损失函数，根据CTR、CVR和CVR的误差预测建立CTR预估双稳健损失函数，然后根据这两个损失函数训练错误归咎模型和CVR预测模型。为了说明本申请实施例提供的方法的效果，下面提供了一具体示例，在该示例中，将本申请实施例提供的方法与已有的训练方法进行比较。
具体地，使用某一浏览器广告的数据进行试验，数据集如下表1所示。
表1数据集(M：百万，K：千)
利用表1所示的训练集，使用本申请实施例提供的方法和已有的训练方法分别对CTR预测模型和CVR预测模型进行训练，然后利用训练后的模型和测试集进行预测，最后对预测结果进行统计，统计结果如下表2所示，其中，CTR是指点击率(Click Through Rate)，CVR是指转化率(Post-click Conversion Rate)，CTCVR是指点击转化率(Post-view Click-through&Conversion Rate)。
表2统计结果
其中，DCN是指用于广告点击预测的深度交叉网络(Deep&Cross Network for Ad Click Predictions)，ESMM是指全空间多任务学习模型(Entire Space Multi-Task Model)，DR-JL是指双稳健联合学习(Doubly Robust Joint Learning)，Multi-IPW是指多任务逆倾向性得分加权CVR估计器(Multi-task Inverse Propensity Weighting CVR Estimator)，Multi-DR是指多任务双稳健CVR估计器(Multi-task Doubly Robust CVR Estimator)，MRDR是指更稳健的双稳健(More Robust Doubly Robust)，DR-BIAS是指将本申请实施例中的偏差项作为错误归咎模型的损失函数进行训练的方法，DR-MSE是指本申请实施例图4所示的训练方法，AUC(Area Under Curve)是预测结果的准确率的评价指标，具体可以理解为，如果随机挑选一个正样本和一个负样本,分类算法将这个正样本排在负样本前面的概率。
从表2可以看出，使用本申请实施例提供的方法训练得到的CTR预测模型、CVR预测模型和CTCVR预测模型的预测准确率更高。
如图7所示，本申请还提供了一种模型的训练装置的实施例，包括：计算单元201，用于根据预测模型对于第一样本数据的预测结果的第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率，计算错误归咎模型的损失函数，错误归咎模型的第一输出用于表征第一误差的预测值，错误归咎模型的损失函数包括偏差项和方差项，偏差项用于表征预测模型的实际损失函数与理论损失函数间的偏差，方差项用于表征预测模型的实际损失函数与理论损失函数间的方差；更新单元202，用于根据错误归咎模型的损失函数对错误归咎模型的参数进行更新。
作为一种可实现的方式，偏差项是根据第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率得到的；和/或方差项是根据第一误差、错误归咎模型的第一输出和第一样本数据被观测的概率得到的。
作为一种可实现的方式，错误归咎模型的损失函数是根据偏差项的权重和方差项的权重，对偏差项和方差项进行加权求和得到的。
作为一种可实现的方式，偏差项的权重和/或方差项的权重是根据第一样本数据得到的。
作为一种可实现的方式，偏差项为其中，D表示全样本空间,(u,i)表示第一样本数据，u表示用户，i表示目标对象，ou,i用于表征第一样本数据是否被观测到，表示错误归咎模型的第一输出，eu,i表示第一误差，表示第一样本数据被观测的概率；当ou,i的取值为1时，ou,i表示第一样本数据被观测到；当ou,i的取值为0时，ou,i表示第一样本数据未被观测到。
作为一种可实现的方式，方差项为其中，D表示全样本空间,(u,i)表示第一样本数据，u表示用户，i表示目标对象，ou,i用于表征第一样本数据是否被观测到，表示错误归咎模型的第一输出，eu,i表示第一误差，表示第一样本数据被观测的概率；当ou,i的取值为1时，ou,i表示第一样本数据被观测到；当ou,i的取值为0时，ou,i表示所述第一样本数据未被观测到。
作为一种可实现的方式，第一样本数据被观测的概率是目标概率模型根据输入的第一样本数据输出的。
作为一种可实现的方式，计算单元201，还用于根据预测模型对于第二样本数据的预测结果的第二误差、错误归咎模型的第二输出、第二样本数据被观测的概率以及第二样本数据的标签，计算预测模型的损失函数，错误归咎模型的第二输出用于表征第二误差的预测值，第二样本数据的标签指示用户对对象的评价是否被观测到；更新单元202，还用于根据预测模型的损失函数对预测模型的参数进行更新。
图8为本申请实施例提供的一种计算机设备的结构示意图。如图8所示，计算机设备900通常为服务器，搭载有上述的模型的训练装置。计算机设备900由一般性的总线体系结构来实现。
计算机设备900包括至少一个处理器901、通信总线902、存储器903以及至少一个通信接口904。
可选地，处理器901是一个通用中央处理器(central processing unit,CPU)、网络处理器(network processor，NP)、微处理器、或者是一个或多个用于实现本申请方案的集成电路，例如，专用集成电路(application-specific integrated circuit，ASIC)，可编程逻辑器件(programmable logic device，PLD)或其组合。上述PLD是复杂可编程逻辑器件(complex programmable logic device，CPLD)，现场可编程逻辑门阵列(field-programmable gate array，FPGA)，通用阵列逻辑(generic array logic，GAL)或其任意组合。
通信总线902用于在上述组件之间传送信息。通信总线902分为地址总线、数据总线、控制总线等。为便于表示，图中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
可选地，存储器903是只读存储器(read-only memory，ROM)或可存储静态信息和指令的其它类型的静态存储设备。可替换的，存储器903是随机存取存储器(random access memory，RAM)或者可存储信息和指令的其它类型的动态存储设备。可替换的，存储器903是电可擦可编程只读存储器(electrically erasable programmable read-only Memory，EEPROM)、只读光盘(compact disc read-only memory，CD-ROM)或其它光盘存储、光碟存储(包括压缩光碟、激光碟、光碟、数字通用光碟、蓝光光碟等)、磁盘存储介质或者其它磁存储设备，或者是能够用于携带或存储具有指令或数据结构形式的期望的程序代码并能够由计算机存取的任何其它介质，但不限于此。可选地，存储器903是独立存在的，并通过通信总线902与处理器901相连接。可选地，存储器903和处理器901集成在一起。
通信接口904使用任何收发器一类的装置，用于与其它设备或通信网络通信。通信接口904包括有线 通信接口。可选地，通信接口904还包括无线通信接口。其中，有线通信接口例如为以太网接口。以太网接口是光接口，电接口或其组合。无线通信接口为无线局域网(wireless local area networks，WLAN)接口，蜂窝网络通信接口或其组合等。
在具体实现中，作为一种实施例，处理器901包括一个或多个CPU，如图8中所示的CPU0和CPU1。
在具体实现中，作为一种实施例，计算机设备900包括多个处理器，如图8中所示的处理器901和处理器905。这些处理器中的每一个是一个单核处理器(single-CPU)，或者是一个多核处理器(multi-CPU)。这里的处理器指一个或多个设备、电路、和/或用于处理数据(如计算机程序指令)的处理核。
在一些实施例中，存储器903用于存储执行本申请方案的程序代码99，处理器901执行存储器903中存储的程序代码99。也就是说，计算机设备900通过处理器901以及存储器903中的程序代码99，来实现上述的方法实施例。
本申请实施例还提供一种芯片，包括一个或多个处理器。所述处理器中的部分或全部用于读取并执行存储器中存储的计算机程序，以执行前述各实施例的方法。
可选地，该芯片该包括存储器，该存储器与该处理器通过电路或电线与存储器连接。进一步可选地，该芯片还包括通信接口，处理器与该通信接口连接。通信接口用于接收需要处理的数据和/或信息，处理器从该通信接口获取该数据和/或信息，并对该数据和/或信息进行处理，并通过该通信接口输出处理结果。该通信接口可以是输入输出接口。
在一些实现方式中，所述一个或多个处理器中还可以有部分处理器是通过专用硬件的方式来实现以上方法中的部分步骤，例如涉及神经网络模型的处理可以由专用神经网络处理器或图形处理器来实现。
本申请实施例提供的方法可以由一个芯片实现，也可以由多个芯片协同实现。
本申请实施例还提供了一种计算机存储介质，该计算机存储介质用于储存为上述计算机设备所用的计算机软件指令，其包括用于执行为计算机设备所设计的程序。
该计算机设备可以如前述图7对应实施例中模型的训练装置的功能。
本申请实施例还提供了一种计算机程序产品，该计算机程序产品包括计算机软件指令，该计算机软件指令可通过处理器进行加载来实现前述各个实施例所示的方法中的流程。
以上，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。

Claims (13)

  1. 一种模型的训练方法，其特征在于，包括：
    根据预测模型对于第一样本数据的预测结果的第一误差、错误归咎模型的第一输出和所述第一样本数据被观测的概率，计算所述错误归咎模型的损失函数，所述错误归咎模型的第一输出用于表征所述第一误差的预测值，所述错误归咎模型的损失函数包括偏差项和方差项，所述偏差项用于表征所述预测模型的实际损失函数与理论损失函数间的偏差，所述方差项用于表征所述预测模型的实际损失函数与理论损失函数间的方差；
    根据所述错误归咎模型的损失函数对所述错误归咎模型的参数进行更新。
  2. 根据权利要求1所述的方法，其特征在于，所述偏差项是根据所述第一误差、所述错误归咎模型的第一输出和所述第一样本数据被观测的概率得到的；和/或
    所述方差项是根据所述第一误差、所述错误归咎模型的第一输出和所述第一样本数据被观测的概率得到的。
  3. 根据权利要求1或2所述的方法，其特征在于，
    所述错误归咎模型的损失函数是根据所述偏差项的权重和所述方差项的权重，对所述偏差项和所述方差项进行加权求和得到的。
  4. 根据权利要求3所述的方法，其特征在于，所述偏差项的权重和/或所述方差项的权重是根据所述第一样本数据得到的。
  5. 根据权利要求1至4中任意一项所述的方法，其特征在于，所述偏差项为 其中，D表示全样本空间,(u,i)表示所述第一样本数据，u表示用户，i表示目标对象，ou,i用于表征所述第一样本数据是否被观测到，表示错误归咎模型的第一输出，eu,i表示所述第一误差，表示所述第一样本数据被观测的概率；
    当ou,i的取值为1时，ou,i表示所述第一样本数据被观测到；当ou,i的取值为0时，ou,i表示所述第一样本数据未被观测到。
  6. 根据权利要求1至5中任意一项所述的方法，其特征在于，所述方差项为 其中，D表示全样本空间,(u,i)表示所述第一样本数据，u表示用户，i表示目标对象，ou,i用于表征所述第一样本数据是否被观测到，表示错误归咎模型的第一输出，eu,i表示所述第一误差，表示所述第一样本数据被观测的概率；
    当ou,i的取值为1时，ou,i表示所述第一样本数据被观测到；当ou,i的取值为0时，ou,i表示所述第一样本数据未被观测到。
  7. 根据权利要求1至6中任意一项所述的方法，其特征在于，所述第一样本数据被观测的概率是目标概率模型根据输入的所述第一样本数据输出的。
  8. 根据权利要求1至7中任意一项所述的方法，其特征在于，在所述根据所述错误归咎模型的损失函数对所述错误归咎模型的参数进行更新之后，所述方法还包括：
    根据所述预测模型对于第二样本数据的预测结果的第二误差、所述错误归咎模型的第二输出、所述第二样本数据被观测的概率以及所述第二样本数据的指示信息，计算所述预测模型的损失函数，所述错误归咎模型的第二输出用于表征所述第二误差的预测值，所述指示信息用于指示所述第二样本数据是否被观测到；
    根据所述预测模型的损失函数对所述预测模型的参数进行更新。
  9. 一种模型的训练装置，其特征在于，包括：
    计算单元，用于根据预测模型对于第一样本数据的预测结果的第一误差、错误归咎模型的第一输出和所述第一样本数据被观测的概率，计算所述错误归咎模型的损失函数，所述错误归咎模型的第一输出用于表征所述第一误差的预测值，所述错误归咎模型的损失函数包括偏差项和方差项，所述偏差项用于表征所述预测模型的实际损失函数与理论损失函数间的偏差，所述方差项用于表征所述预测模型的实际损失函数与理论损失函数间的方差；
    更新单元，用于根据所述错误归咎模型的损失函数对所述错误归咎模型的参数进行更新。
  10. 一种计算机设备，其特征在于，所述计算机设备包括：存储器和处理器；
    所述处理器，用于执行存储器中存储的计算机程序或指令，以使所述计算机设备执行如权利要求1-8中任一项所述的方法。
  11. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质具有程序指令，当所述程序指令被直接或者间接执行时，使得如权利要求1至8中任一所述的方法被实现。
  12. 一种芯片系统，其特征在于，所述芯片系统包括至少一个处理器，所述处理器用于执行存储器中存储的计算机程序或指令，当所述计算机程序或所述指令在所述至少一个处理器中执行时，使得如权利要求1至8中任一所述的方法被实现。
  13. 一种计算机程序产品，其特征在于，包括指令，当所述指令在计算机上运行时，使得计算机执行权利要求1至8中任一项所述的方法。
PCT/CN2023/103919 2022-07-15 2023-06-29 一种模型的训练方法的相关装置 WO2024012219A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210836072.4 2022-07-15
CN202210836072.4A CN117436540A (zh) 2022-07-15 2022-07-15 一种模型的训练方法的相关装置

Publications (1)

Publication Number Publication Date
WO2024012219A1 true WO2024012219A1 (zh) 2024-01-18

Family

ID=89535523

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/103919 WO2024012219A1 (zh) 2022-07-15 2023-06-29 一种模型的训练方法的相关装置

Country Status (2)

Country Link
CN (1) CN117436540A (zh)
WO (1) WO2024012219A1 (zh)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113508378A (zh) * 2019-10-31 2021-10-15 华为技术有限公司 推荐模型的训练方法、推荐方法、装置及计算机可读介质
US20220093271A1 (en) * 2019-01-25 2022-03-24 Children's Hospital Medical Center Bayesian causal inference models for healthcare treatment using real world patient data
US20220108334A1 (en) * 2020-10-01 2022-04-07 Adobe Inc. Inferring unobserved event probabilities

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220093271A1 (en) * 2019-01-25 2022-03-24 Children's Hospital Medical Center Bayesian causal inference models for healthcare treatment using real world patient data
CN113508378A (zh) * 2019-10-31 2021-10-15 华为技术有限公司 推荐模型的训练方法、推荐方法、装置及计算机可读介质
US20220108334A1 (en) * 2020-10-01 2022-04-07 Adobe Inc. Inferring unobserved event probabilities

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
GUO SIYUAN, LIXIN ZOU, YIDING LIU, WENWEN YE, SUQI CHENG, SHUAIQIANG WANG, HECHANG CHEN, DAWEI YIN, YI CHANG: "Enhanced Doubly Robust Learning for Debiasing Post-Click Conversion Rate Estimation", SIGIR ’21, JULY 11–15, 2021, VIRTUAL EVENT, CANADA, 1 July 2021 (2021-07-01) - 15 July 2021 (2021-07-15), pages 275 - 284, XP093128383 *
WANG XIAOJIE, ZHANG RUI, SUN YU, QI JIANZHONG: "Doubly Robust Joint Learning for Recommendation on Data Missing Not at Random", PROCEEDINGS OF THE 36TH INTERNATIONAL CONFERENCE ON MACHINE LEARNING, vol. 97, 1 January 2019 (2019-01-01), pages 6638 - 6647, XP093128380 *

Also Published As

Publication number Publication date
CN117436540A (zh) 2024-01-23

Similar Documents

Publication Publication Date Title
US11711447B2 (en) Method and apparatus for real-time personalization
Liu et al. Risk assessment in system FMEA combining fuzzy weighted average with fuzzy decision-making trial and evaluation laboratory
Gao et al. Evaluating user reputation in online rating systems via an iterative group-based ranking method
US20180341898A1 (en) Demand forecast
EP4181026A1 (en) Recommendation model training method and apparatus, recommendation method and apparatus, and computer-readable medium
Mirghafoori et al. Analysis of the barriers affecting the quality of electronic services of libraries by VIKOR, FMEA and entropy combined approach in an intuitionistic-fuzzy environment
CN105159910A (zh) 信息推荐方法和装置
EP4242955A1 (en) User profile-based object recommendation method and device
CN111798280B (zh) 多媒体信息推荐方法、装置和设备及存储介质
Majd et al. A trust model for recommender agent systems
CN105761154A (zh) 一种社会化推荐方法及装置
CN102163304A (zh) 带有优化的域间信息质量评定的协作联网的方法和系统
CN111340522B (zh) 资源推荐方法、装置、服务器及存储介质
CN111861679A (zh) 一种基于人工智能的商品推荐方法
CN112016000B (zh) 一种基于卷积协同过滤的电影组推荐方法及系统
JP2019036173A (ja) 算出装置、算出方法、算出プログラム及び第２モデル
WO2024012219A1 (zh) 一种模型的训练方法的相关装置
CN116385076A (zh) 广告推荐方法、装置、终端设备和计算机可读存储介质
JP7231322B2 (ja) 学習装置、学習方法、学習プログラム及びプログラム
CN113515704B (zh) 推荐效果评价方法、装置、系统及计算机程序产品
US11334935B2 (en) Method, system, and manufacture for light hypergraph based recommendation
US11868429B1 (en) Taxonomization of features used in prediction models according to sub-categories in a list of ranked categories
CN113159877A (zh) 数据处理方法、装置、系统、计算机可读存储介质
TWI714213B (zh) 用戶型態預測系統及其方法
CN116089722B (zh) 基于图产出标签的实现方法、装置、计算设备和存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23838712

Country of ref document: EP

Kind code of ref document: A1