An information recommendation method, apparatus, and storage medium

By training the information recommendation model with data weighting and adjusting parameters according to the recommendation scenario and task type, the problem of insufficient information recommendation accuracy is solved, and high-accuracy information recommendation is achieved in different scenarios.

CN116383485BActive Publication Date: 2026-11-17BEIJING WODONG TIANJUN INFORMATION TECH CO LTD +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202310213805.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-03-07
Publication Date
2026-11-17
Estimated Expiration
2043-03-07

AI Technical Summary

Technical Problem

In existing technologies, recommending products based on users' historical purchase information suffers from insufficient accuracy, especially when there are significant differences in shopping preferences across different pages, leading to a decrease in the accuracy of information recommendations.

Method used

An information recommendation model is adopted. This model is trained by weighting the training data and according to at least one recommendation scenario type and at least one recommendation task type. It obtains the action information of the target object under the target recommendation scenario and target recommendation task, and adjusts the model parameters to improve the recommendation accuracy.

Benefits of technology

It improves the accuracy of information recommendation, especially in sparse scenarios, enhancing the effectiveness of information recommendation and the robustness of the model, and strengthening the predictive ability in medium and long-tail scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116383485B_ABST
    Figure CN116383485B_ABST
Patent Text Reader

Abstract

This application discloses an information recommendation method, apparatus, and storage medium, comprising: acquiring the operation information of a target object under a target recommendation scenario and a target recommendation task; inputting the target recommendation scenario, target recommendation task, and operation information into an information recommendation model to obtain the recommendation information corresponding to the target object; the information recommendation model is a model trained by weighting the training data according to the recommendation scenario type and recommendation task type; and recommending the recommendation information to the target object.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

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

[0002] With the development of internet technology, users have more convenient ways to obtain information. They are often recommended a lot of information, such as products that users may buy or news that users may be interested in.

[0003] In related technologies, a user's shopping preferences are determined based on their historical purchase information, and products are then recommended to the user. However, since a user's shopping preferences may differ on different pages, directly relying on historical purchase information to determine the products to be recommended reduces the accuracy of product recommendations, which in turn reduces the accuracy of information recommendations. Summary of the Invention

[0004] To address the aforementioned technical problems, embodiments of this application aim to provide an information recommendation method, apparatus, and storage medium that can improve the accuracy of information recommendations.

[0005] The technical solution of this application is implemented as follows:

[0006] This application provides an information recommendation method, which includes:

[0007] Obtain the action information of the target object in the target recommendation scenario and target recommendation task;

[0008] The target recommendation scenario, the target recommendation task, and the information to be operated on are input into the information recommendation model to obtain the information to be recommended corresponding to the target object; the information recommendation model is a model trained by weighting the training data according to at least one recommendation scenario type and at least one recommendation task type.

[0009] The information to be recommended is recommended to the target object.

[0010] This application provides an information recommendation device, the device comprising:

[0011] The acquisition unit is used to acquire the operation information of the target object in the target recommendation scenario and target recommendation task.

[0012] The input unit is used to input the target recommendation scenario, the target recommendation task, and the information to be operated into the information recommendation model to obtain the information to be recommended corresponding to the target object; the information recommendation model is a model trained by weighting the training data according to at least one recommendation scenario type and at least one recommendation task type.

[0013] The recommendation unit is used to recommend the information to be recommended to the target object.

[0014] This application provides an information recommendation device, the device comprising:

[0015] The system includes a memory, a processor, and a communication bus. The memory communicates with the processor via the communication bus. The memory stores information recommendation programs that the processor can execute. When the information recommendation programs are executed, the processor performs the information recommendation method described above.

[0016] This application provides a storage medium storing a computer program for use in an information recommendation device. The computer program, when executed by a processor, implements the information recommendation method described above.

[0017] This application provides an information recommendation method, apparatus, and storage medium. The information recommendation method includes: acquiring actionable information of a target object under a target recommendation scenario and a target recommendation task; inputting the target recommendation scenario, target recommendation task, and actionable information into an information recommendation model to obtain actionable information corresponding to the target object; the information recommendation model is a model trained by weighting training data according to at least one recommendation scenario type and at least one recommendation task type; and recommending actionable information to the target object. Using the above method, the information recommendation apparatus acquires actionable information of a target object under a target scenario and a target recommendation task, and uses the information recommendation model to determine the actionable information based on the target recommendation scenario, target recommendation task, and actionable information. The information recommendation model is a model trained by weighting training data according to at least one recommendation scenario type and at least one recommendation task type. That is, the model parameters of the information recommendation model are adjusted according to the recommendation scenario type and recommendation task type corresponding to the training data, resulting in high accuracy in information recommendation. Using this highly accurate information recommendation model, highly accurate actionable information can be determined, thus improving the accuracy of information recommendation. Attached Figure Description

[0018] Figure 1 This application provides a schematic diagram illustrating various business scenarios in the prior art.

[0019] Figure 2 A schematic diagram illustrating multi-scenario learning, multi-task learning, and multi-scenario multi-task learning in the prior art, provided for embodiments of this application;

[0020] Figure 3 A flowchart of an information recommendation method provided in an embodiment of this application;

[0021] Figure 4 An exemplary multi-scenario, multi-task model block diagram provided for embodiments of this application;

[0022] Figure 5 A schematic diagram of the composition structure of an information recommendation device provided in this application embodiment. Figure 1 ;

[0023] Figure 6 A schematic diagram of the composition structure of an information recommendation device provided in this application embodiment. Figure 2 . Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0025] To better serve users, many industrial recommendation systems have begun to segment their operations into different scenarios based on user departments, product categories, and traffic channels. For example, users can access recommendation services through PC browsers and mobile applications, or through the homepage or specific vertical pages. Figure 1 As shown, the business scenarios range from the featured brands on the e-commerce platform's homepage and the "Featured Recommendations" section on the homepage to recommendations for order details, claiming platform benefits, and pending payment scenarios.

[0026] These scenarios often share useful information about users and products, such as basic user attributes and product characteristics. However, the traffic varies significantly across different scenarios. For example, the "Featured Recommendations" section on the homepage of an e-commerce platform accounts for more than 50% of the total recommended traffic. Figure 2 As shown in (a), recommendation systems can be enriched through information sharing between relevant scenarios in a multi-scenario learning (MSL) setup. In each scenario, user satisfaction and user engagement are typically quantified using different objectives, such as click-through rate (CTR) and conversion rate (CVR). That is, as... Figure 2 As shown in (b), multi-task learning (MTL) settings can also be applied to leverage the inherent correlations between different tasks. For multi-scenario multi-task learning scenarios, such as... Figure 2 As shown in (c), M2M proposes a new multi-scenario multi-task paradigm to predict various tasks from different scenarios in an end-to-end manner.

[0027] This application provides an information recommendation method, which is applied to an information recommendation device. Figure 3A flowchart of an information recommendation method provided in this application embodiment is shown below. Figure 3 As shown, information recommendation methods may include:

[0028] S101. Obtain the operation information of the target object in the target recommendation scenario and target recommendation task;

[0029] The information recommendation method provided in this application embodiment is applicable to scenarios where information to be recommended is recommended to a target object.

[0030] In the embodiments of this application, the information recommendation device can be implemented in various forms. For example, the information recommendation device described in this application may include devices such as mobile phones, cameras, tablet computers, laptops, handheld computers, personal digital assistants (PDAs), portable media players (PMPs), navigation devices, wearable devices, smart bracelets, pedometers, etc., as well as devices such as digital TVs, desktop computers, servers, etc.

[0031] In this embodiment of the application, the target object can be a user.

[0032] In this embodiment, the target recommendation scenario can be a scenario where products are recommended on the homepage of an e-commerce platform, a scenario where products are recommended on the page where a user has successfully placed an order, or other product recommendation scenarios or information recommendation scenarios; the specific scenario can be determined according to the actual situation, and this embodiment does not limit it.

[0033] It should be noted that the number of target recommendation scenarios can be determined according to the actual situation, and this application embodiment does not limit this.

[0034] In this embodiment, the target recommendation task can be the task of a product being clicked by a user, the task of product conversion, or other tasks; the specific target recommendation task can be determined according to the actual situation, and this embodiment does not limit it.

[0035] It should be noted that the number of target recommendation tasks can be determined according to the actual situation, and this application embodiment does not limit this.

[0036] In this embodiment of the application, the information to be operated on may be the information of the exposed product.

[0037] In this embodiment, the information recommendation device can obtain the target object's pending operation information in the target recommendation scenario and target recommendation task from other devices; it can also obtain the target object's pending operation information in the target recommendation scenario and target recommendation task from the information input interface; or it can obtain the target object's pending operation information in the target recommendation scenario and target recommendation task from other places; the specific details can be determined according to the actual situation, and this embodiment does not limit it.

[0038] S102. Input the target recommendation scenario, target recommendation task, and information to be operated into the information recommendation model to obtain the information to be recommended corresponding to the target object; the information recommendation model is a model trained by weighting the training data according to at least one recommendation scenario type and at least one recommendation task type.

[0039] In this embodiment of the application, after the information recommendation device obtains the operation information of the target object in the target recommendation scenario and the target recommendation task, the information recommendation device inputs the target recommendation scenario, the target recommendation task and the operation information into the information recommendation model to obtain the recommendation information corresponding to the target object.

[0040] In this embodiment, the information recommendation device inputs the target recommendation scenario, target recommendation task, and information to be operated into the information recommendation model. Before obtaining the information to be recommended corresponding to the target object, it also acquires at least one set of training data and at least one set of sample labels corresponding to multiple sample objects. In the at least one set of training data, a preset number of training data are acquired to obtain at least one set of test data. In the at least one set of sample labels, labels corresponding to the at least one set of test data are acquired to obtain at least one set of test labels. The model is determined using sample weights. Based on the at least one set of training data and at least one set of sample labels, a first model loss of the initial information recommendation model under at least one recommendation scenario type is determined. Based on the at least one set of test data and at least one set of test labels, a second model loss of the initial information recommendation model under at least one recommendation scenario type is determined. The model parameters of the initial information recommendation model are adjusted based on the first model loss and the second model loss to obtain the information recommendation model.

[0041] It should be noted that at least one set of training data includes at least one set of operational information for multiple sample objects under at least one recommendation scenario type and at least one recommendation task type.

[0042] In the embodiments of this application, at least one set of operation information can be operation information of multiple sample objects on exposed products in at least one recommendation scenario type, such as operation information of clicking on products, or operation information of converting products, etc.

[0043] It should also be noted that at least one set of training data can include product information of the exposed products.

[0044] In this embodiment, the information recommendation device can obtain at least one set of training data and at least one set of sample labels corresponding to multiple sample objects from a database; or it can obtain at least one set of training data and at least one set of sample labels corresponding to multiple sample objects through an information transmission interface; the specific method can be determined according to the actual situation, and this embodiment does not limit it.

[0045] In this application embodiment, different recommendation task types correspond to different tags. For example, in the recommendation task type where the product is clicked, the corresponding tag is "product clicked" (tag value can be 1) or "product not clicked" (tag value can be 0); in the recommendation task type where the product is converted, the corresponding tag is "product converted" (tag value can be 1) or "product not converted" (tag value can be 0).

[0046] It should be noted that the preset quantity can be the quantity configured in the information recommendation device; the preset quantity can also be the quantity transmitted to the information recommendation device from other devices; or it can be the quantity information obtained by the information recommendation device in other ways; the specific way in which the information recommendation device obtains the preset quantity can be determined according to the actual situation, and this application embodiment does not limit it in this way.

[0047] For example, the preset quantity can be 1000, 3000, or other quantities; the specific preset quantity value can be determined according to the actual situation, and this application embodiment does not limit it.

[0048] In this embodiment of the application, the process by which the information recommendation device obtains a preset number of training data from at least one set of training data to obtain at least one set of test data can be that the information recommendation device obtains a preset number of training data that meet the data accuracy conditions from at least one set of training data to obtain at least one set of test data.

[0049] It should be noted that the data accuracy condition can be data free from erroneous operations. For example, for products browsed by users, training data corresponding to products whose browsing time is greater than or equal to a preset time can be selected, which would satisfy the data accuracy condition.

[0050] In this embodiment of the application, the process of determining the first model loss of the initial information recommendation model under at least one recommendation scenario type by using sample weights to determine the model and by using at least one set of training data and at least one set of sample labels includes: determining at least one set of first model losses of the initial information recommendation model under at least one recommendation scenario type by using at least one set of training data and at least one set of sample labels; inputting at least one set of first model losses and at least one recommendation scenario identifier information corresponding to at least one recommendation scenario type into the sample weights to determine at least one set of weight parameters; and determining the first model loss of the initial information recommendation model by using at least one set of weight parameters and at least one set of first model losses.

[0051] In this embodiment of the application, the process by which the information recommendation device determines at least one set of first model losses of an initial information recommendation model under at least one recommendation scenario type based on at least one set of training data and at least one set of sample labels includes: inputting at least one set of training data into the initial information recommendation model to obtain at least one set of output sample labels; and determining at least one set of first model losses based on at least one set of output sample labels and at least one set of sample labels.

[0052] In this embodiment, the process by which the information recommendation device determines the first model loss of the initial information recommendation model based on at least one set of weight parameters and at least one set of first model losses can be as follows: determining the product between the first set of weight parameters and the first set of model losses in the at least one set of weight parameters to obtain a first product; determining the product between the second set of weight parameters and the second set of model losses in the at least one set of weight parameters to obtain a second product; ...; determining the product between the last set of weight parameters and the last set of model losses in the at least one set of weight parameters to obtain a last product; and then determining the sum of the first product, the second product, ..., the last product to obtain the first model loss.

[0053] Specifically, the process of determining the first model loss T(θ,w) of the initial information recommendation model based on at least one set of weight parameters and at least one set of first model losses is shown in formula (1):

[0054]

[0055] Where, p t (f i w) represents the sample weights (at least one set of weight parameters) that are adaptive to at least one recommendation scenario type; For at least one set of first model loss; T represents the number of recommendation task types, t represents the t-th recommendation task type, λ tIt is a hyperparameter that balances the loss function for multiple recommendation task types, where K represents the number of samples in a batch. From D train Obtained.

[0056] It should be noted that a batch may contain samples from multiple recommendation scenario types, and the quantities are unbalanced; scenarios with more samples are more likely to have more samples. The loss function L is determined based on the recommendation task type. If task t is a binary classification recommendation task, L... t Typically, the cross-entropy loss function is used.

[0057]

[0058] in It is the estimated probability (output label) of the i-th sample in task t. This is the true label (sample label) corresponding to the sample.

[0059] In this embodiment, the initial information recommendation model can be represented as f(θ), where θ is a model parameter that is not perceived by the model and can be any information recommendation model corresponding to multiple scenarios and multiple tasks, such as M2M or AESM. 2 The specific initial information recommendation model can be determined according to the actual situation, and this application does not limit it in this embodiment.

[0060] In this embodiment of the application, the process by which the information recommendation device inputs at least one set of first model losses and at least one recommendation scenario identifier information corresponding to at least one recommendation scenario type into a sample weight determination model to obtain at least one set of weight parameters includes: classifying at least one set of first model losses according to at least one recommendation task type to obtain at least one set of task losses; converting at least one set of task losses into vectors to obtain at least one set of first loss vectors; converting at least one recommendation scenario identifier information into at least one recommendation scenario identifier vector; concatenating at least one set of first loss vectors and at least one recommendation scenario identifier vector to obtain a concatenated vector; and inputting the concatenated vector into the sample weight determination model to obtain at least one set of weight parameters.

[0061] In this embodiment of the application, the process of the information recommendation device inputting the concatenated vector into the sample weight determination model to obtain at least one set of weight parameters includes: inputting the concatenated vector into the sample weight determination model to obtain at least one set of initial weight parameters; and classifying the at least one set of initial weight parameters according to at least one recommendation scenario type to obtain at least one set of weight parameters.

[0062] In this embodiment, the information recommendation device can convert at least one set of task losses into dense vectors to obtain at least one set of first loss vectors; or it can convert at least one set of task losses into other vectors to obtain at least one set of first loss vectors. The specific method of converting at least one set of task losses into vectors to obtain at least one set of first loss vectors can be determined according to the actual situation, and this embodiment does not limit it.

[0063] In this embodiment, the method of converting at least one recommendation scene identifier information into at least one recommendation scene identifier vector is the same as the method of converting at least one set of task losses into vectors to obtain at least one set of first loss vectors; the method of converting at least one recommendation scene identifier information into at least one recommendation scene identifier vector may also be different from the method of converting at least one set of task losses into vectors to obtain at least one set of first loss vectors; the specific method can be determined according to the actual situation, and this embodiment does not limit it.

[0064] It should be noted that the method of converting at least one recommended scene identifier information into at least one recommended scene identifier vector can be the vector conversion method in the prior art; the specific method can be determined according to the actual situation, and this application embodiment does not limit it.

[0065] In this embodiment, the information recommendation device can first concatenate at least one set of first loss vectors together to obtain an initial concatenated vector, and then concatenate at least one recommended scene identifier vector to the end of the initial concatenated vector to obtain the concatenated vector. Alternatively, it can concatenate the first recommended scene identifier vector from at least one recommended scene identifier vector after the first loss vector in the at least one set of first loss vectors; then concatenate the second recommended scene identifier vector from at least one recommended scene identifier vector after the second loss vector in the at least one set of first loss vectors; ...; and finally concatenate the last recommended scene identifier vector from at least one recommended scene identifier vector after the last loss vector in the at least one set of first loss vectors to obtain the concatenated vector. The specific method of concatenating at least one set of first loss vectors and at least one recommended scene identifier vector to obtain the concatenated vector can be determined according to the actual situation, and this embodiment does not limit this.

[0066] For example, the process of concatenating at least one set of first loss vectors and at least one recommendation scene identifier vector to obtain a concatenated vector is shown in formula (3):

[0067] d = concat(F1(Loss1), ..., F T (Loss T ),S emb (3)

[0068] Among them, at least one recommended scene identifier (scene ID) is transformed into a vector S through the embedding layer. emb (At least one recommended scene identifier vector) can be introduced by introducing T functions F1, F2, ... F, implemented by a multi-layer perceptron (MLP). T Each loss function is transformed into a dense vector, that is, at least one set of task losses Loss1, Loss2, ... Loss is transformed using MLP. T Convert to vectors to obtain at least one set of first loss vectors (F1(Loss1),…F…). T (Loss T )), where T is the number of tasks; concatenate at least one set of first loss vectors and at least one recommendation scene identifier vector to obtain concatenated vector d.

[0069] After that, d i T adaptive sample weights are obtained from T MLPs, for example, d i The adaptive sample weights p for task t are obtained through the MLP network corresponding to task t. t (w).

[0070] definition The feedforward process of a fully connected neural network is shown in equation (4):

[0071]

[0072] Here, l identifies the specific layer, and σ is the activation function. These represent the weight parameters, biases, and output of the l-th fully connected neural network for task t, respectively. Subsequently, the dense feature vectors... The adaptive sample weights for this task are obtained by inputting the sigmoid activation function, as shown in formula (5):

[0073]

[0074] It should be noted that |L| is the number of layers in the subnetwork.

[0075] In this embodiment, the process by which an information recommendation device determines a second model loss of an initial information recommendation model under at least one recommendation scenario type based on at least one set of test data and at least one set of test labels includes: inputting at least one set of test data into the initial information recommendation model to obtain at least one set of output test labels; determining at least one set of second model losses of the initial information recommendation model based on at least one set of output test labels and at least one set of test labels; obtaining model parameters of the initial information recommendation model; adjusting the model parameters using the first model loss to obtain adjustment parameters; and determining a second model loss based on at least one set of second model losses and adjustment parameters.

[0076] In this embodiment, the process of training an initial information recommendation model using at least one set of test data and at least one set of test labels is called the meta-testing process, which uses the second derivative of the model parameters. To achieve this process, gradient descent can be used to virtually update the model parameters (θ), that is, the model parameters (θ) are adjusted using the first model loss to obtain the adjusted parameters (θ′), as shown in formula (6):

[0077]

[0078] Where α is the learning rate in the meta-training stage (the process of training an initial information recommendation model by determining the model using sample weights and based on at least one set of training data and at least one set of sample labels is called meta-training); T(θ,w) is the first model loss. This indicates that the gradient is calculated with respect to the parameter θ based on T(θ,w), yielding the specific gradient value. Meta-test set From D test The obtained samples consist of fair scene samples. Through validation on a fair dataset, the information recommendation model can prioritize scenes with sparse samples, thereby improving the model's fairness. Meanwhile, the meta-test samples high-quality samples, increasing the model's robustness. The loss function for the meta-test can be determined using formula (7), i.e., the second model loss (M(θ′,w)) can be determined:

[0079]

[0080] In this embodiment, the process by which the information recommendation device adjusts the model parameters of an initial information recommendation model based on a first model loss and a second model loss to obtain an information recommendation model includes: determining a model training loss based on the first model loss and the second model loss; adjusting the model parameters of the initial information recommendation model to obtain a first adjusted model when the model training loss is greater than or equal to a preset model loss; adjusting the sample weights to determine the model parameters to obtain a second adjusted model; continuing to train the first adjusted model using multiple training data, multiple sample labels, and the second adjusted model until the adjusted model loss corresponding to the first adjusted model is less than or equal to the preset model loss, and using the first adjusted model as the information recommendation model; and using the initial information recommendation model as the information recommendation model when the model training loss is less than the preset model loss.

[0081] In this embodiment of the application, the process by which the information recommendation device determines the model training loss based on the first model loss and the second model loss can be as follows: obtaining the loss fusion coefficient; determining the product of the loss fusion coefficient and the second model loss; and determining the sum between the product and the first model loss, thereby obtaining the model training loss.

[0082] In this embodiment, the preset model loss can be the loss configured in the information recommendation device, the loss transmitted to the information recommendation device from other devices, or the loss obtained by the information recommendation device in other ways. The specific way in which the information recommendation device obtains the preset model loss can be determined according to the actual situation, and this embodiment does not limit it.

[0083] In this embodiment of the application, during the training phase of the initial information recommendation model, training data from N scenarios (at least one type of recommendation scenario) are used. Training is performed, where d si Let represent the training data for the i-th scene. Then, model f(θ) is tested on a fair test set. The evaluation is conducted on a fair test set, which consists of a small sample fairly sampled from the training set (i.e., from at least one set of training data, a predetermined number of training data points are obtained to get at least one set of test data; from at least one set of sample labels, labels corresponding to at least one set of test data are obtained to get at least one set of test labels, thus obtaining a fair test set), approximately 1% of the training set, and is called the meta-test set. The meta-test set has the same number of samples for each scenario, and the behavior is of relatively high quality (i.e., the data in the test set is training data that meets the data accuracy requirements). Figure 4As shown, based on at least one set of training data (at least one set of training data in the training set) and at least one set of sample labels (at least one set of sample labels in the training set), at least one set of first model losses (Loss1, ..., Loss) of the initial information recommendation model (f(θ)) is determined for at least one recommendation scenario type. n The process involves inputting at least one set of first model losses and at least one recommendation scenario identifier corresponding to at least one recommendation scenario type into the sample weight determination model (p(w)) to obtain at least one set of weight parameters; determining the first model loss (training Loss) of the initial information recommendation model based on the at least one set of weight parameters and at least one set of first model losses; inputting at least one set of test data (test data in the scenario fairness test set) into the initial information recommendation model (f(θ′)) to obtain at least one set of output test labels; and determining at least one set of second model losses (Loss1, ..., Loss) of the initial information recommendation model based on the at least one set of output test labels and at least one set of test labels (test labels in the scenario fairness test set). n The process involves: obtaining the initial information recommendation model parameters (θ); adjusting the model parameters using the first model loss to obtain the adjusted parameters; determining the second model loss (Meta-test Loss) based on at least one set of second model losses and the adjusted parameters; determining the model training loss based on the first and second model losses; adjusting the model parameters of the initial information recommendation model when the model training loss is greater than or equal to the preset model loss to obtain the first adjusted model (f(θ*)); adjusting the sample weights to determine the model parameters to obtain the second adjusted model (p(w*)); continuing to train the first adjusted model using multiple training data, multiple sample labels, and the second adjusted model until the adjusted model loss corresponding to the first adjusted model is less than or equal to the preset model loss, and using the first adjusted model as the information recommendation model; and using the initial information recommendation model as the information recommendation model when the model training loss is less than the preset model loss. This process is then applied to the training set. Train a multi-scenario, multi-task model f(θ) to achieve better recommendation results for long-tail products without upgrading the model structure, improve model fairness, and thus improve overall efficiency.

[0084] It should be noted that the training set consists of at least one set of training data and at least one set of sample labels.

[0085] In this embodiment of the application, the process of adjusting the model parameters of the initial information recommendation model to obtain the first adjusted model by the information recommendation device includes: determining the first updated model parameters based on the model parameters of the initial information recommendation model, the first model loss, and the second model loss; and replacing the model parameters of the initial information recommendation model with the first updated model parameters to obtain the first adjusted model.

[0086] In this embodiment, the final optimization objective of the initial information recommendation model is shown in formula (8):

[0087] arg θ,w min(T(θ,w)+βM(θ′,w))(8)

[0088] Where β represents the importance of the meta-test loss (also known as the loss fusion coefficient); arg θ,w The parameters that need to be updated during minimization are w (the model parameters of the sample weight-determined model) and θ (the model parameters of the initial information recommendation model). After training the initial information recommendation model once using the training set and determining the first model loss using the sample weight-determined model, a second model loss is determined by testing the initial information recommendation model once using the test set. This determines the first model training loss. When determining multiple model training losses, the model parameters (θ) corresponding to the minimum model training loss are used as the target model parameters, thereby replacing the model parameters in the initial information recommendation model, thus obtaining the information recommendation model.

[0089] It should be noted that the condition for minimizing the model training loss can be determined based on the preset model loss.

[0090] In this embodiment, by minimizing formula (8), the parameter θ can be updated using gradient descent as shown in formula (9), that is, the first updated model parameter is determined using formula (9):

[0091]

[0092] It should be noted that θ represents the model parameters of the initial information recommendation model, α is the learning rate during the meta-training stage, and T(θ,w) is the first model loss. This indicates that for the parameter θ, the gradient is calculated based on T(θ,w) to obtain the specific gradient value; β is the loss fusion coefficient; θ′ is the adjustment parameter; M(θ′,w) is the second model loss; and the hyperparameter γ is the learning rate for updating the parameter θ. To obtain the specific gradient value, we calculate the gradient based on M(θ′,w) for the parameter θ′.

[0093] In this embodiment of the application, the process of adjusting the model parameters of the sample weight determination model by the information recommendation device to obtain the second adjusted model includes: obtaining the model parameters of the sample weight determination model; determining the second updated model parameters based on the second model loss and the model parameters of the sample weight determination model; and replacing the model parameters of the sample weight determination model with the second updated model parameters to obtain the second adjusted model.

[0094] In this embodiment, by minimizing formula (8), the parameter w can be updated using gradient descent as shown in formula (10):

[0095]

[0096] It should be noted that w is the model parameter that determines the sample weights; the hyperparameter φ is the learning rate that updates the parameter w; and M(θ′,w) is the second model loss. This indicates that with respect to the parameter w, the gradient is calculated based on M(θ′,w) to obtain the specific gradient value.

[0097] In this embodiment, the meta-test phase is equivalent to performing the second derivative on the meta-test set. The loss of the meta-test phase is used to update the parameter w of the weight function p(w) and to calculate the gradient of the parameters of the multi-scenario multi-task model.

[0098] It is understandable that if the gradient of a training sample has a decreasing effect on the meta-test loss M(θ′,w), it is considered beneficial to the fairness of the information recommendation model and effective for long-tail scenarios. Weight-aware networks (sample weight determination models) tend to assign it higher weights. The second derivative of the meta-test step guides and modifies the update direction of parameter θ, so the information recommendation model can treat different scenarios fairly. This application guides f(θ) and p(w) to achieve better performance on the training set and a fair meta-test set through formulas (9) and (10).

[0099] In this embodiment of the application, the process of the information recommendation device acquiring at least one set of training data and at least one set of sample labels corresponding to multiple sample objects according to the recommendation scenario type includes: acquiring multiple training data and multiple sample labels corresponding to multiple sample objects; classifying the multiple training data and multiple sample labels according to at least one recommendation scenario type to obtain at least one set of training data and at least one set of sample labels.

[0100] In this embodiment, multiple training data and multiple sample labels corresponding to multiple sample objects can be obtained from a database; multiple training data and multiple sample labels can also be obtained through an information transmission interface; multiple training data and multiple sample labels can also be obtained through other means; the specific way to obtain multiple training data and multiple sample labels can be determined according to the actual situation, and this embodiment does not limit it.

[0101] S103. Recommend the information to be recommended to the target audience.

[0102] In this embodiment, the information recommendation device inputs the target recommendation scenario, the target recommendation task, and the information to be operated into the information recommendation model. After obtaining the information to be recommended corresponding to the target object, the information recommendation device can recommend the information to be recommended to the target object.

[0103] In this embodiment of the application, the information recommendation device recommends the information to be recommended to the target object in such a way that the information to be recommended is recommended to the target object when the target object is in the target recommendation scenario and the target recommendation task.

[0104] In this embodiment, the invention was verified both offline and online. All models used Adam with a learning rate of 0.001 and a batch size of 256 for each scene. Offline evaluation was performed on six collected mid-to-long-tail scenes. AUC (Area under the ROC curve) is a commonly used metric for evaluating CTR and CVR predictions. The comparison method is shown in Table (1):

[0105] Table (1)

[0106]

[0107]

[0108] It should be noted that Table 1 shows the click-through rate (CTR) and conversion rate results for different methods on offline industrial datasets. Here, DNN (single) refers to training a single DNN for each scene using only its own data. Hard Sharing refers to a multi-layer perceptron (MLP) with hard sharing at the bottom layer. Multi-gate Mixture-Of-Experts (MMOE) implicitly models the relationships between experts using multiple mixed experts; the merged representations from multiple gates can be transformed into multiple task prediction layers. Based on multiple scenes, each tower adds multiple tasks for multi-task prediction. PLE is another MMOE variant that divides experts into task-specific groups and task-sharing groups to avoid negative transfer and seesaw effects. Based on multiple scenes, each tower adds multiple tasks for multi-task prediction. STAR: The CTR prediction task uses a star topology structure with a shared center network and scene network. Since it does not support multi-task, CTR and conversion rate are trained and evaluated separately. M2M: Utilizes meta-learning to dynamically generate scene-specific gate network and tower network parameters; it is a multi-scene, multi-task model. AESM2: Stacks multi-task layers on multi-scene layers, enabling flexible establishment of intrinsic connections between different scenes. It also supports high-level feature extraction for different tasks and proposes a novel expert selection algorithm to automatically identify scene / task-specific and shared experts for each input. FSWNML: This application presents a meta-learning-based sample weighting method that is model-aware and can be used in any multi-scene, multi-task model. For example, FSWNML (HardSharing) represents applying FSWNML to a Hard Sharing model.

[0109] As shown in Table 1, the offline evaluation results demonstrate that training with only the data from the current scene generally does not yield good results, especially in sparse scenes such as Scene 1. Directly fusing data from multiple scenes, considering richer data information, can improve performance in most scenes, such as Scene 1 and Scene 2. MMOE, PLE, and STAR specifically implement multi-scene modeling, further improving performance compared to Hard Sharing. M2M and AESM2 comprehensively consider multi-scene, multi-task modeling, achieving better results. However, none of the above models consider model fairness, resulting in limited improvement in sparse scenes, such as Scene 1. The FSWNML (Hard Sharing) proposed in this application achieves near-optimal performance compared to all the comparative models in the sparse Scene 1, with parameters not dominated by scenes with abundant samples, enhancing prediction performance in mid-to-long-tail scenes. Meanwhile, since this application is model-agnostic (i.e., it can use any multi-scenario, multi-task model as the initial information recommendation model for training), its combination with M2M and AESM2 further improves performance on both tasks. This is because the high-quality samples in the meta-test enhance the model's robustness and mitigate the interference of noisy samples. It should be noted that the improvement in sparse scenario 1 is very significant, demonstrating the effectiveness of the sample weighting method in this application in sparse scenarios.

[0110] It is understandable that the information recommendation device obtains the information to be operated under the target scene and target recommendation task of the target object, and uses the information recommendation model to determine the information to be recommended based on the target recommendation scene, target recommendation task and information to be operated. The information recommendation model is a model trained by weighting the training data according to at least one recommendation scene type and at least one recommendation task type. That is, the model parameters of the information recommendation model are adjusted according to the recommendation scene type and recommendation task type corresponding to the training data, so that the information recommendation model has high accuracy when recommending information. Using this high-accuracy information recommendation model, highly accurate information to be recommended can be determined, thus improving the accuracy of information recommendation.

[0111] Based on the same inventive concept as the above-mentioned information recommendation method, this application provides an information recommendation device 1, corresponding to an information recommendation method; Figure 5 A schematic diagram of the composition structure of an information recommendation device provided in this application embodiment. Figure 1 The information recommendation device 1 may include:

[0112] Acquisition unit 11 is used to acquire the operation information of the target object in the target recommendation scenario and target recommendation task;

[0113] Input unit 12 is used to input the target recommendation scenario, the target recommendation task, and the information to be operated into the information recommendation model to obtain the information to be recommended corresponding to the target object; the information recommendation model is a model trained by weighting the training data according to at least one recommendation scenario type and at least one recommendation task type;

[0114] Recommendation unit 13 is used to recommend the information to be recommended to the target object.

[0115] In some embodiments of this application, the apparatus further includes a determining unit and an adjusting unit;

[0116] The acquisition unit 11 is used to acquire at least one set of training data and at least one set of sample labels corresponding to multiple sample objects; the at least one set of training data includes at least one set of operation information of the multiple sample objects under at least one recommendation scenario type and at least one recommendation task type; in the at least one set of training data, a preset number of training data are acquired respectively to obtain at least one set of test data; in the at least one set of sample labels, the labels corresponding to the at least one set of test data are acquired to obtain at least one set of test labels;

[0117] The determining unit is configured to determine the model using sample weights, determine the first model loss of the initial information recommendation model under the at least one recommendation scenario type based on the at least one set of training data and the at least one set of sample labels, and determine the second model loss of the initial information recommendation model under the at least one recommendation scenario type based on the at least one set of test data and the at least one set of test labels.

[0118] The adjustment unit is used to adjust the model parameters of the initial information recommendation model according to the first model loss and the second model loss to obtain the information recommendation model.

[0119] In some embodiments of this application, the determining unit is configured to determine at least one set of first model losses of the initial information recommendation model under the at least one recommendation scenario type based on the at least one set of training data and the at least one set of sample labels; and to determine the first model loss of the initial information recommendation model based on the at least one set of weight parameters and the at least one set of first model losses.

[0120] The input unit 12 is used to input the at least one set of first model losses and at least one recommendation scene identifier information corresponding to the at least one recommendation scene type into the sample weight determination model to obtain at least one set of weight parameters.

[0121] In some embodiments of this application, the device further includes a classification unit, a conversion unit, and a splicing unit;

[0122] The classification unit is used to classify the at least one set of first model losses according to the at least one recommendation task type, so as to obtain at least one set of task losses;

[0123] The transformation unit is used to convert the at least one set of task losses into vectors to obtain at least one set of first loss vectors; and to convert the at least one recommendation scene identification information into at least one recommendation scene identification vector.

[0124] The splicing unit is used to splice the at least one set of first loss vectors and the at least one recommendation scene identifier vector to obtain a spliced ​​vector;

[0125] The input unit 12 is used to input the concatenated vector into the sample weight determination model to obtain the at least one set of weight parameters.

[0126] In some embodiments of this application, the input unit 12 is used to input the at least one set of test data into the initial information recommendation model to obtain at least one set of output test labels;

[0127] The determining unit is configured to determine at least one set of second model losses for the initial information recommendation model based on the at least one set of output test labels and the at least one set of test labels; and to determine the second model loss based on the at least one set of second model losses and the adjustment parameters.

[0128] The acquisition unit 11 is used to acquire the model parameters of the initial information recommendation model;

[0129] The adjustment unit is used to adjust the model parameters using the first model loss to obtain the adjustment parameters.

[0130] In some embodiments of this application, the apparatus further includes a training unit;

[0131] The determining unit is configured to determine the model training loss based on the first model loss and the second model loss;

[0132] The adjustment unit is used to adjust the model parameters of the initial information recommendation model to obtain a first adjusted model when the model training loss is greater than or equal to the preset model loss; and to adjust the model parameters of the sample weight determination model to obtain a second adjusted model.

[0133] The training unit is used to continue training the first adjustment model using the multiple training data, the multiple sample labels, and the second adjustment model until the adjustment model loss corresponding to the first adjustment model is less than or equal to the preset model loss, and then the first adjustment model is used as the information recommendation model; if the model training loss is less than the preset model loss, the initial information recommendation model is used as the information recommendation model.

[0134] In some embodiments of this application, the apparatus further includes a replacement unit;

[0135] The determining unit is used to determine the first updated model parameters based on the model parameters of the recommendation model, the first model loss, and the second model loss of the initial information.

[0136] The replacement unit is used to replace the model parameters of the initial information recommendation model with the first updated model parameters to obtain the first adjusted model.

[0137] In some embodiments of this application, the acquisition unit 11 is used to acquire the model parameters of the sample weight determination model;

[0138] The determining unit is used to determine the model parameters of the model based on the second model loss and the sample weights, and to determine the second updated model parameters;

[0139] The replacement unit is used to replace the model parameters of the sample weights with the second updated model parameters to obtain the second adjusted model.

[0140] In some embodiments of this application, the acquisition unit 11 is used to acquire multiple training data and multiple sample labels corresponding to multiple sample objects;

[0141] The classification unit is used to classify the plurality of training data and the plurality of sample labels according to the at least one recommended scenario type, so as to obtain at least one set of training data and at least one set of sample labels.

[0142] It should be noted that, in practical applications, the aforementioned acquisition unit 11, input unit 12, and recommendation unit 13 can be implemented by the processor 14 on the information recommendation device 1, specifically by a CPU (Central Processing Unit), MPU (Microprocessor Unit), DSP (Digital Signal Processor), or Field Programmable Gate Array (FPGA), etc.; the aforementioned data storage can be implemented by the memory 15 on the information recommendation device 1.

[0143] This application embodiment also provides an information recommendation device 1, such as... Figure 6 As shown, the information recommendation device 1 includes a processor 14, a memory 15, and a communication bus 16. The memory 15 communicates with the processor 14 through the communication bus 16. The memory 15 stores programs executable by the processor 14. When the program is executed, the information recommendation method described above is executed by the processor 14.

[0144] In practical applications, the aforementioned memory 15 can be volatile memory, such as random-access memory (RAM); or non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD); or a combination of the above types of memory, and provide instructions and data to the processor 14.

[0145] This application provides a computer-readable storage medium having a computer program thereon, which, when executed by a processor 14, implements the information recommendation method as described above.

[0146] It is understandable that the information recommendation device obtains the information to be operated under the target scene and target recommendation task of the target object, and uses the information recommendation model to determine the information to be recommended based on the target recommendation scene, target recommendation task and information to be operated. The information recommendation model is a model trained by weighting the training data according to at least one recommendation scene type and at least one recommendation task type. That is, the model parameters of the information recommendation model are adjusted according to the recommendation scene type and recommendation task type corresponding to the training data, so that the information recommendation model has high accuracy when recommending information. Using this high-accuracy information recommendation model, highly accurate information to be recommended can be determined, thus improving the accuracy of information recommendation.

[0147] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0148] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0149] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0151] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application.

Claims

1. An information recommendation method, characterized in that, The method includes: Obtain the actionable information of the target object in the target recommendation scenario and target recommendation task; the actionable information is the information of the exposed product. The target recommendation scenario, the target recommendation task, and the information to be operated are input into the information recommendation model to obtain the information to be recommended corresponding to the target object; the information recommendation model is a model trained by weighting the training data according to at least one recommendation scenario type and at least one recommendation task type. The information to be recommended is recommended to the target object; Before inputting the target recommendation scenario, the target recommendation task, and the information to be operated into the information recommendation model to obtain the information to be recommended corresponding to the target object, the method further includes: Obtain at least one set of training data and at least one set of sample labels corresponding to multiple sample objects; the at least one set of training data includes at least one set of operation information of the multiple sample objects under at least one recommendation scenario type and at least one recommendation task type; From the at least one set of training data, a preset number of training data are obtained to obtain at least one set of test data; from the at least one set of sample labels, the labels corresponding to the at least one set of test data are obtained to obtain at least one set of test labels; The model is determined by using sample weights, and the first model loss of the initial information recommendation model is determined under the at least one recommendation scenario type based on the at least one set of training data and the at least one set of sample labels; the sample weight determination model is used to determine the weight parameters corresponding to the first model loss of the initial information recommendation model under the at least one recommendation scenario type. Determine the second model loss of the initial information recommendation model under the at least one recommendation scenario type based on the at least one set of test data and the at least one set of test labels; The model parameters of the initial information recommendation model are adjusted based on the first model loss and the second model loss to obtain the information recommendation model.

2. The method according to claim 1, characterized in that, The step of determining the model using sample weights and determining the first model loss of the initial information recommendation model under the at least one recommendation scenario type based on the at least one set of training data and the at least one set of sample labels includes: Based on the at least one set of training data and the at least one set of sample labels, determine at least one set of first model losses for the initial information recommendation model under the at least one recommendation scenario type; The at least one set of first model losses and at least one recommendation scenario identifier information corresponding to the at least one recommendation scenario type are input into the sample weight determination model to obtain at least one set of weight parameters; The first model loss of the initial information recommendation model is determined based on the at least one set of weight parameters and the at least one set of first model losses.

3. The method according to claim 2, characterized in that, The step of inputting the at least one set of first model losses and at least one recommendation scenario identifier information corresponding to the at least one recommendation scenario type into the sample weight determination model yields at least one set of weight parameters, including: According to the at least one recommendation task type, classify the at least one set of first model losses to obtain at least one set of task losses; Transform the at least one set of task losses into vectors to obtain at least one set of first loss vectors; The at least one recommended scene identifier information is converted into at least one recommended scene identifier vector; The at least one set of first loss vectors and the at least one recommended scene identifier vector are concatenated to obtain a concatenated vector; The concatenated vector is input into the sample weight determination model to obtain the at least one set of weight parameters.

4. The method according to claim 1, characterized in that, Determining the second model loss of the initial information recommendation model under the at least one recommendation scenario type based on the at least one set of test data and the at least one set of test labels includes: Input the at least one set of test data into the initial information recommendation model to obtain at least one set of output test labels; Based on the at least one set of output test labels and the at least one set of test labels, determine at least one set of second model losses for the initial information recommendation model; Obtain the model parameters of the initial information recommendation model; The model parameters are adjusted using the first model loss to obtain the adjusted parameters; The second model loss is determined based on the at least one set of second model losses and the adjustment parameters.

5. The method according to claim 1, characterized in that, The step of adjusting the model parameters of the initial information recommendation model based on the first model loss and the second model loss to obtain the information recommendation model includes: The model training loss is determined based on the first model loss and the second model loss; If the training loss of the model is greater than or equal to the preset model loss, the model parameters of the initial information recommendation model are adjusted to obtain a first adjusted model; the model parameters of the sample weight determination model are adjusted to obtain a second adjusted model. The first adjustment model is trained using multiple training data, multiple sample labels, and the second adjustment model until the adjustment model loss corresponding to the first adjustment model is less than or equal to the preset model loss. The first adjustment model is then used as the information recommendation model. If the training loss of the model is less than the preset model loss, the initial information recommendation model will be used as the information recommendation model.

6. The method according to claim 5, characterized in that, The step of adjusting the model parameters of the initial information recommendation model to obtain the first adjusted model includes: The first updated model parameters are determined based on the model parameters of the recommendation model, the first model loss, and the second model loss, according to the initial information. The first adjusted model is obtained by replacing the model parameters of the initial information recommendation model with the first updated model parameters.

7. The method according to claim 5, characterized in that, The process of adjusting the sample weights to determine the model parameters and obtaining the second adjusted model includes: Obtain the sample weights to determine the model parameters; The model parameters are determined based on the second model loss and the sample weights, and the second updated model parameters are determined accordingly. The model parameters of the determined model are obtained by replacing the sample weights with the second updated model parameters, thus obtaining the second adjusted model.

8. The method according to claim 1, characterized in that, The step of obtaining at least one set of training data and at least one set of sample labels corresponding to multiple sample objects includes: Obtain multiple training data and multiple sample labels corresponding to multiple sample objects; The plurality of training data and the plurality of sample labels are classified according to at least one recommended scenario type to obtain at least one set of training data and at least one set of sample labels.

9. An information recommendation device, characterized in that, The device includes: The acquisition unit is used to acquire the operation information of the target object in the target recommendation scenario and target recommendation task; the operation information is the information of the exposed product. The input unit is used to input the target recommendation scenario, the target recommendation task, and the information to be operated into the information recommendation model to obtain the information to be recommended corresponding to the target object; the information recommendation model is a model trained by weighting the training data according to at least one recommendation scenario type and at least one recommendation task type. The recommendation unit is used to recommend the information to be recommended to the target object; The device further includes a determining unit and an adjusting unit; The acquisition unit is used to acquire at least one set of training data and at least one set of sample labels corresponding to multiple sample objects; the at least one set of training data includes at least one set of operation information of the multiple sample objects under at least one recommendation scenario type and at least one recommendation task type; in the at least one set of training data, a preset number of training data are acquired respectively to obtain at least one set of test data; in the at least one set of sample labels, the labels corresponding to the at least one set of test data are acquired to obtain at least one set of test labels; The determining unit is used to determine the model using sample weights, and to determine the first model loss of the initial information recommendation model under the at least one recommendation scenario type based on the at least one set of training data and the at least one set of sample labels; and to determine the second model loss of the initial information recommendation model under the at least one recommendation scenario type based on the at least one set of test data and the at least one set of test labels; the sample weight determining model is used to determine the weight parameters corresponding to the first model loss of the initial information recommendation model under the at least one recommendation scenario type. The adjustment unit is used to adjust the model parameters of the initial information recommendation model according to the first model loss and the second model loss to obtain the information recommendation model.

10. An information recommendation device, characterized in that, The device includes: The system includes a memory, a processor, and a communication bus, wherein the memory communicates with the processor via the communication bus, and the memory stores information-recommended programs executable by the processor. When the information-recommended programs are executed, the processor performs the method as described in any one of claims 1 to 8.

11. A storage medium storing a computer program thereon for use in an information recommendation device, characterized in that, When executed by a processor, the computer program implements the method described in any one of claims 1 to 8.

Citation Information

Patent Citations

  • Recommendation method and device, medium, device and computing equipment

    CN113672807A

  • Information pushing method and device, equipment and storage medium

    CN113761352A