Model training methods, information recommendation methods, devices, equipment and storage media

Abstract

This application proposes a model training method, information recommendation method, device, electronic device, and computer storage medium. The method includes: obtaining a training dataset based on the historical behavior data of multiple users; the training dataset includes multiple item samples and label information for each item sample, the label information being used to characterize whether the item sample is clicked after each user browses; processing the training dataset using an initial item ranking model to obtain a first processing result; processing the training dataset and the first processing result using an initial creative selection model to obtain a second processing result; and jointly training the initial item ranking model and the initial creative selection model based on the first processing result and the second processing result to obtain an item ranking model and a creative selection model.

Description

Technical Field

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

[0002] The primary function of a product recommendation system is to address the problem of an overabundance of products. Within a short timeframe, it filters out a small number of items that users are likely to be interested in. Currently, recommendation systems mainly consist of two phases: recall and ranking. The recall phase typically retrieves thousands of candidate products from a massive pool, while the ranking phase scores these candidates to determine the final products displayed to the user. Furthermore, each product has numerous images and creative titles to display; selecting suitable creative titles for the right users can significantly improve click-through rates and the platform's traffic value. Currently, such as... Figure 1A and Figure 1B As shown, recommendation systems typically place the creative model before or after the ranking model. That is, after recall, the creative for each candidate product is determined first and then ranked to complete the product recommendation, or the products to be displayed to the user are determined after ranking and then creatives are selected to complete the product recommendation. However, in the implementation process, the inventors found that the existing technology has at least the following problems: the ranking model and the creative model need to be trained separately, which results in long model training time and consumes a lot of computing resources. Summary of the Invention

[0003] This application provides a model training method, an information recommendation method, an apparatus, an electronic device, and a computer storage medium.

[0004] The technical solution of this application is implemented as follows:

[0005] This application provides a model training method, the method comprising:

[0006] A training dataset is obtained based on the historical behavior data of multiple users; the training dataset includes multiple item samples and label information for each item sample, the label information being used to characterize whether the item sample was clicked after each user browsed it;

[0007] The training dataset is processed using an initial item sorting model to obtain a first processing result; the training dataset and the first processing result are then processed using an initial creative selection model to obtain a second processing result.

[0008] Based on the first processing result and the second processing result, the initial item sorting model and the initial creative selection model are jointly trained to obtain the item sorting model and the creative selection model.

[0009] This application provides an information recommendation method, the method comprising:

[0010] Based on the target user's historical behavior data, determine the target user's user feature vector, the item feature vectors of multiple candidate items, and the creative feature vectors of each candidate idea corresponding to the multiple candidate items;

[0011] The user feature vector and the item feature vector are processed using an item ranking model to obtain a first prediction result; the user feature vector and the creative feature vector are processed using a creative selection model to obtain a second prediction result.

[0012] Based on the first prediction result and the second prediction result, recommendation information is determined; the recommendation information includes at least one target item among the plurality of candidate items and a target idea corresponding to each target item; wherein, the item ranking model and the idea selection model are obtained according to the model training method provided by one or more of the aforementioned technical solutions.

[0013] This application provides a model training device, which includes an acquisition module, a processing module, and a training module.

[0014] The acquisition module is used to acquire a training dataset based on the historical behavior data of multiple users; the training dataset includes multiple item samples and label information for each item sample, the label information being used to characterize whether the item sample was clicked after each user browsed it;

[0015] The processing module is used to process the training dataset using an initial item sorting model to obtain a first processing result; and to process the training dataset and the first processing result using an initial creative selection model to obtain a second processing result.

[0016] The training module is used to jointly train the initial item sorting model and the initial creative selection model based on the first processing result and the second processing result, so as to obtain the item sorting model and the creative selection model.

[0017] This application also proposes an information recommendation device, which includes a first determining module, a prediction module, and a second determining module, wherein...

[0018] The first determining module is used to determine the user feature vector of the target user, the item feature vectors of multiple candidate items, and the creative feature vectors of each candidate idea corresponding to the multiple candidate items based on the target user's historical behavior data.

[0019] The prediction module is used to process the user feature vector and the item feature vector using an item ranking model to obtain a first prediction result; and to process the user feature vector and the creative feature vector using a creative selection model to obtain a second prediction result.

[0020] The second determining module is used to determine recommendation information based on the first prediction result and the second prediction result; the recommendation information includes at least one target item among the plurality of candidate items and a target idea corresponding to each target item; wherein the item ranking model and the idea selection model are obtained according to the model training method provided by one or more of the aforementioned technical solutions.

[0021] This application provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the model training method or information recommendation method provided by one or more of the aforementioned technical solutions.

[0022] This application provides a computer storage medium storing a computer program; when the computer program is executed, it can implement the model training method or information recommendation method provided by one or more of the aforementioned technical solutions.

[0023] This application proposes a model training method, information recommendation method, device, electronic device, and computer storage medium. The method includes: obtaining a training dataset based on historical behavior data of multiple users; the training dataset includes multiple item samples and label information for each item sample, the label information being used to characterize whether the item sample is clicked after each user browses; processing the training dataset using an initial item ranking model to obtain a first processing result; processing the training dataset and the first processing result using an initial creative selection model to obtain a second processing result; and jointly training the initial item ranking model and the initial creative selection model based on the first processing result and the second processing result to obtain an item ranking model and a creative selection model.

[0024] As can be seen, in this embodiment, after obtaining training datasets for multiple users, during the processing of the training dataset using the initial item ranking model, the output of the initial item ranking model and the training dataset are jointly input into the creative selection model for processing. That is, the two models can perceive each other during training, thus improving the subsequent prediction performance of the model. In addition, after obtaining the outputs of the initial item ranking model and the initial creative selection model, joint training of the two models based on their respective outputs can reduce model training time and save more computing resources. Attached Figure Description

[0025] Figure 1A This is a schematic diagram of the structure of a recommendation system in related technologies;

[0026] Figure 1B This is a schematic diagram of the structure of another recommendation system in related technologies;

[0027] Figure 2 This is a flowchart illustrating a model training method in an embodiment of this application;

[0028] Figure 3 This is a schematic diagram of a structure for jointly training an initial item sorting model and an initial creative selection model in an embodiment of this application.

[0029] Figure 4 This is a flowchart illustrating an information recommendation method in an embodiment of this application;

[0030] Figure 5 This is a schematic diagram of the structure of a recommendation system according to an embodiment of this application;

[0031] Figure 6 This is a schematic diagram of a structure for predicting click-through rates using an item sorting model and a creative selection model, as described in an embodiment of this application.

[0032] Figure 7 This is a schematic diagram of the composition of the model training device according to an embodiment of this application;

[0033] Figure 8 This is a schematic diagram of the composition structure of the information recommendation device according to an embodiment of this application;

[0034] Figure 9 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0035] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments provided herein are merely illustrative of the present application and are not intended to limit the present application. Furthermore, the embodiments provided below are some embodiments for implementing the present application, and not all embodiments for implementing the present application. Unless otherwise specified, the technical solutions described in the embodiments of the present application can be implemented in any combination.

[0036] It should be noted that, in the embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a method or apparatus that includes a list of elements includes not only the elements expressly stated, but also other elements not expressly listed, or elements inherent to implementing the method or apparatus. Unless otherwise specified, an element defined by the phrase "comprising a..." does not exclude the presence of other related elements in the method or apparatus that includes that element (e.g., steps in the method or units in the apparatus; for example, a unit may be a portion of circuitry, a portion of a processor, a portion of a program or software, etc.).

[0037] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, I and / or J can represent three cases: I alone, I and J simultaneously, and J alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more of a plurality of elements. For example, including at least one of I, J, and R can mean including any one or more elements selected from the set consisting of I, J, and R.

[0038] For example, the model training method provided in the embodiments of this application includes a series of steps, but the model training method provided in the embodiments of this application is not limited to the steps described. Similarly, the model training device provided in the embodiments of this application includes a series of modules, but the model training device provided in the embodiments of this application is not limited to the modules explicitly described, but may also include modules that need to be set up for obtaining relevant time series data or processing based on time series data.

[0039] The embodiments of this application can be applied to computer systems composed of server components and can operate together with many other general-purpose or special-purpose computing system environments or configurations. Here, the server can be a distributed cloud computing technology environment including small computer systems, large computer systems, etc.

[0040] Server-side and other electronic devices can achieve corresponding functions through the execution of program modules. Typically, program modules can include routines, programs, object programs, components, logic, data structures, etc., which perform specific tasks or implement specific abstract data types. Computer systems / servers can be implemented in distributed cloud computing environments, where tasks are executed by remote processing devices connected via communication networks. In distributed cloud computing environments, program modules can reside on local or remote computing system storage media, including storage devices.

[0041] For recommendation systems, time constraints are extremely stringent; a single calculation typically needs to be completed within 0.5 seconds. If a method such as... Figure 1Aand Figure 1B The processing method shown also causes the ranking model and the creative model to compete for time and space, and greatly limits the model complexity and iteration space, which in turn reduces the recommendation effect.

[0042] To address the aforementioned technical problems, the following embodiments are proposed.

[0043] In some embodiments of this application, the model training method can be implemented using a processor in the model training device. The processor can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), controller, microcontroller, and microprocessor.

[0044] It should be noted that the collection, use, storage, sharing and transfer of user personal information involved in the technical solution of this application all comply with the provisions of relevant laws and regulations, and require notification to users and obtaining their consent or authorization. Where applicable, user personal information has been subjected to de-identification and / or anonymization and / or encryption technical processing.

[0045] Figure 2 This is a flowchart illustrating a model training method in an embodiment of this application, as shown below. Figure 2 As shown, the method includes the following steps:

[0046] Step 100: Obtain the training dataset based on the historical behavior data of multiple users.

[0047] For example, the historical behavior data of multiple users can be the historical behavior data generated by multiple users when using the e-commerce platform within a first set time period; the historical behavior data can include user attribute data of each user and behavior data related to multiple items, wherein the behavior data includes, but is not limited to, the target user's browsing behavior data and click behavior data for each item; the items can be any type of item on the e-commerce platform; for example, it can be clothing items, food items, etc.

[0048] Here, the length of the first set time period can be determined according to the actual situation, and this application embodiment does not limit it; for example, it can be the last six months or the last three months.

[0049] For example, firstly, historical behavior data of multiple users is acquired. After acquiring the historical behavior data of multiple users, the historical behavior data can be cleaned to remove invalid and abnormal data. Then, based on the cleaned historical behavior data, browsing behavior data and click behavior data of each user for item samples are acquired. Then, based on the browsing behavior data and click behavior data of each user, the training dataset is determined. Here, the item samples include each item recalled by the target recall model based on the historical behavior data of each user.

[0050] In this embodiment of the application, the training dataset may include multiple item samples and label information for each item sample. For example, the label information is used to characterize whether an item sample is clicked after each user browses it; here, item samples clicked after each user browses them can be called positive samples, and item samples not clicked after each user browses them can be called negative samples. That is, the training dataset includes positive samples and negative samples; wherein, the label value corresponding to a positive sample is 1, and the label value corresponding to a negative sample is 0.

[0051] In one embodiment, in order to utilize behavioral data as much as possible and improve the accuracy of the model, the combination of positive and negative samples determined based on the historical behavioral data of multiple users can be used as the training dataset. In another embodiment, in order to improve the running efficiency of the model, after obtaining the positive and negative samples determined based on the historical behavioral data of multiple users, the positive and negative samples can be sampled according to a set ratio, and the combination of the sampled positive and negative samples can be used as the training dataset. Here, the value of the set ratio can be set according to the actual situation, and this embodiment does not limit it.

[0052] Understandably, to improve the recommendation efficiency of the model, the initial item ranking model and the initial creative selection model can be jointly trained offline using the training dataset in advance, and the trained item ranking model and creative selection model can be stored offline. Then, if information recommendation is needed for a user, these two models can be directly retrieved from the offline storage location for use.

[0053] Step 101: Process the training dataset using the initial item sorting model to obtain the first processing result; process the training dataset and the first processing result using the initial creative selection model to obtain the second processing result.

[0054] In some embodiments, the above method may further include: determining, based on the training dataset, the user sample feature vector, the item sample feature vector, and the creative sample feature vector for each user.

[0055] In this embodiment of the application, after obtaining the training dataset, features of different dimensions are first extracted from the training dataset to obtain user sample features, item sample features, and creative sample features for each user. Then, each feature is hashed using a hash function to obtain a vectorized representation of each feature, corresponding to the user sample feature vector, item sample feature vector, and creative sample feature vector for each user. Here, hashing includes hash mapping.

[0056] In some embodiments, processing the training dataset using an initial item ranking model to obtain a first processing result may include: processing user sample feature vectors and item sample feature vectors using the initial item ranking model to obtain a first processing result; processing the training dataset and the first processing result using an initial creative selection model to obtain a second processing result may include: processing user sample feature vectors, creative sample feature vectors, and the first processing result using the initial creative selection model to obtain a second processing result.

[0057] Furthermore, after obtaining the user sample feature vector, item sample feature vector, and creative sample feature vector for each user, refer to Figure 3 The user sample feature vector and the item sample feature vector can be input into the initial item ranking model to obtain the first processing result; here, the first processing result includes the estimated click-through rate (corresponding to) for each user on different item samples. Figure 3 (PCTR1); Next, the first processing result is used as an input feature, and together with the user sample feature vector and the creative sample feature vector, it is input into the initial creative selection model to obtain the second processing result; Here, the second processing result includes the estimated click-through rate (CTR) of each user for each creative item corresponding to different item samples (corresponding to PCTR1); Figure 3 (PCTR2).

[0058] It should be noted that before inputting the first processing result as an input feature into the initial creative selection model, the first processing result needs to be hashed. Since the value type of the first processing result is usually double and greater than 0 and less than 1, directly hashing it will result in an overly sparse feature vector after mapping. To solve this problem, the first processing result can be evenly divided into 1000 classes before hashing, for example, 0.001-0.002 in one bucket and 0.002-0.003 in another bucket, to improve data retrieval efficiency.

[0059] Understandably, by inputting the output of the initial item ranking model into the initial idea selection model, it can be ensured that the two models perceive and influence each other during training, thereby enhancing the correlation between candidate items and candidate ideas and improving the subsequent prediction accuracy of the model.

[0060] Step 102: Based on the first and second processing results, jointly train the initial item ranking model and the initial creative selection model to obtain the item ranking model and the creative selection model.

[0061] In some embodiments, the initial item ranking model and the initial creative selection model are jointly trained based on the first processing result and the second processing result to obtain the item ranking model and the creative selection model. This may include: determining a first loss based on the first processing result and label information; determining a second loss based on the second processing result and label information; determining a target loss based on the first loss and the second loss; and jointly training the initial item ranking model and the initial creative selection model based on the target loss until convergence to obtain the item ranking model and the creative selection model.

[0062] For example, the initial item ranking model and the initial creative selection model can output the estimated click-through rate of each user for different item samples and the estimated click-through rate of each user for each item creative corresponding to different item samples through activation functions.

[0063] Here, the type of activation function can be determined according to the actual situation. This application does not limit this. For example, it can be a sigmoid function or other types of activation functions. The sigmoid function is used for binary classification problems, and the corresponding output range is between 0 and 1.

[0064] In this embodiment of the application, after obtaining the first processing result, the first loss can be determined based on the first processing result and the tag information (i.e., the actual click results of each item sample); similarly, after obtaining the second processing result, the second loss can be determined based on the second processing result and the tag information (i.e., the actual click results of each item creative corresponding to each item sample).

[0065] For example, when the activation function is the sigmoid function, the first loss and the second loss can be determined according to the cross-entropy loss function. In this embodiment of the application, after obtaining the first loss and the second loss, the first loss and the second loss can be fused to obtain the target loss. Then, the initial item ranking model and the initial creative selection model are jointly trained according to the target loss to obtain the item ranking model and the creative selection model.

[0066] The target loss is a composite loss formed by combining the first loss and the second loss; loss fusion is the process of converting multiple losses into a single loss through one or more preset logical operations. These preset logical operations include, but are not limited to, mixed arithmetic operations, weighted summation, and machine learning algorithms.

[0067] For example, when the preset logical operation is weighted summation, a weighting coefficient can also be introduced. The weighting coefficient can be a value set based on experience or experimentation, which will not be elaborated here.

[0068] For example, after obtaining the target loss, the parameters of the initial item ranking model and the initial creative selection model can be iteratively updated through backpropagation of the target loss until convergence, i.e., reaching the preset training stopping condition. Each time the target loss is backpropagated, it first updates the parameters of the initial creative selection model, then updates the parameters of the initial item ranking model, and also updates all the underlying feature vectors shared by the two models.

[0069] Here, the training stopping condition could be that the target loss reaches a preset minimum value, or that the model performance of the intermediate model does not significantly improve after a preset number of iterations. At this point, the trained item ranking model and creative selection model can be obtained.

[0070] As can be seen, in this embodiment of the application, by jointly training the initial item sorting model and the initial creative selection model offline, the training efficiency of the model can be improved. In addition, since the input features of the initial creative selection model include the output of the initial item sorting model, it can be ensured that the two models are aware of each other during the training process, thus improving the subsequent prediction effect of the model.

[0071] This application proposes a model training method, which includes: obtaining a training dataset based on the historical behavior data of multiple users; the training dataset includes multiple item samples and label information for each item sample, the label information being used to characterize whether an item sample is clicked after each user browses it; processing the training dataset using an initial item ranking model to obtain a first processing result; processing the training dataset and the first processing result using an initial creative selection model to obtain a second processing result; and jointly training the initial item ranking model and the initial creative selection model based on the first and second processing results to obtain an item ranking model and a creative selection model. It can be seen that, after obtaining the training dataset corresponding to multiple users, this application, during the process of processing the training dataset using the initial item ranking model, inputs the output result of the initial item ranking model and the training dataset into the creative selection model for processing. That is, the two models can perceive each other during training, thus improving the subsequent prediction performance of the model. Furthermore, after obtaining the output results of the initial item ranking model and the initial creative selection model respectively, jointly training the two models based on their respective output results can reduce model training time and save more computing resources.

[0072] Based on the foregoing embodiments, this application provides an information recommendation method, such as... Figure 4 As shown, the method includes the following steps:

[0073] Step 200: Based on the target user's historical behavior data, determine the target user's user characteristics, the item characteristics of multiple candidate items, and the creative characteristics of each candidate idea corresponding to the multiple candidate items.

[0074] In this embodiment of the application, the information recommendation method can be applied to an e-commerce platform. The target user, also known as the user to be recommended, refers to any user on the e-commerce platform who has historical behavioral data.

[0075] For example, the target user's historical behavior data can be the historical behavior data generated when the target user uses the e-commerce platform during a second set time period. This historical behavior data may include the target user's user attribute data and behavior data related to multiple candidate items. The target user can be one of the aforementioned multiple users.

[0076] For example, the second set time period refers to the time period before the current time. Here, the length of the second set time period is not limited. For example, it can be the last month or the last two weeks.

[0077] For example, the first and second time periods mentioned above both refer to historical time periods, and the second time period is after the first time period. Since the historical behavior data generated by users recently can more accurately reflect the user's preference information than the data generated earlier, the historical time period that is closer to the current time can be selected as the first and second time periods. For example, the first time period can be set to the last six months and the second time period can be set to the last two weeks after the last six months.

[0078] In some embodiments, historical behavior data of the target user is first obtained. After obtaining the historical behavior data of the target user, the user feature vector of the target user, the item feature vectors of multiple candidate items, and the creative feature vectors of each candidate idea corresponding to the multiple candidate items are determined based on the historical behavior data of the target user. This may include: extracting features from the historical behavior data of the target user to obtain the initial user features of the target user; determining the item information of multiple candidate items based on the historical behavior data; extracting features from the item information of multiple candidate items to obtain the initial item features of multiple candidate items and the initial creative features of each candidate idea corresponding to the multiple candidate items; and performing hash processing on the initial user features, initial item features, and initial creative features respectively to obtain the user feature vector of the target user, the item feature vectors of multiple candidate items, and the creative feature vectors of each candidate idea corresponding to the multiple candidate items.

[0079] For example, after obtaining the historical behavior data of the target user, the historical behavior data can be cleaned first to remove invalid and abnormal data. Then, the cleaned behavior data is first subjected to user-dimensional feature extraction to obtain the initial user features of the target user. Then, based on the cleaned behavior data, the item information of multiple candidate items is determined, and the item information of multiple candidate items is subjected to item-dimensional feature extraction to obtain the initial item features of multiple candidate items and the initial creative features of each candidate creative corresponding to multiple candidate items.

[0080] For example, candidate items represent items that the target user may be interested in; it should be noted that multiple candidate items are obtained by recalling items in advance through a pre-trained item recall model. Specifically, the item recall model recalls items from a massive pool of items on the e-commerce platform based on the target user's historical behavior data; here, the type of recall model can be set according to the actual situation, and this application does not limit it.

[0081] It should be noted that each candidate item can correspond to at least two candidate creatives; each candidate creative can include, but is not limited to, an item image and an item title, and may also include reasons for recommending the item; there are differences between the different candidate creatives corresponding to each candidate item; for example, if a candidate item corresponds to 10 candidate creatives, the content included in these 10 candidate creatives will be somewhat different.

[0082] For example, initial user characteristics include, but are not limited to, age, gender, education level, and activity level; initial item characteristics include, but are not limited to, item identification document (ID), category, and brand; initial creative characteristics include, but are not limited to, item images, item titles, and reasons for recommending the item.

[0083] Furthermore, after obtaining the initial user features, initial item features, and initial creative features, each feature can be hashed using a hash function to obtain the vectorized representations of the initial user features, initial item features, and initial creative features, which correspond to user feature vectors, item feature vectors, and creative feature vectors, respectively.

[0084] Step 201: Process the user feature vector and the item feature vector using the item ranking model to obtain the first prediction result; process the user feature vector and the creative feature vector using the creative selection model to obtain the second prediction result.

[0085] Here, the item ranking model can be any type of click-through rate prediction model, such as a multilayer perceptron (MLP) model or other types of ranking models; similarly, the creative selection model can also be any type of click-through rate prediction model, such as a multilayer perceptron (MLP) model or other types of ranking models.

[0086] For example, the MLP model is a common feedforward neural network model; it consists of multiple neuron layers, each connected to both the previous and next layers, with full connections between layers. In the MLP model, each neuron receives the outputs of all neurons in the previous layer, sums them in a weighted manner, and then processes the sum through a non-linear activation function to obtain the output of that layer. This output then serves as the input to the next layer, and so on, until the output layer obtains the final output.

[0087] For example, by using an item ranking model to process the user feature vector and the item feature vector, a first prediction result can be obtained; wherein, the first prediction result may include the first estimated click-through rate of multiple candidate items, that is, the estimated click-through rate (PCTR) of each candidate item.

[0088] For example, a creative selection model is used to process user feature vectors and creative features to obtain a second prediction result; wherein, the second prediction result may include the second estimated click-through rate of each candidate creative corresponding to each candidate item, that is, the estimated click-through rate of each candidate creative.

[0089] It should be noted that user feature vectors are input into both the item ranking model and the creative selection model; that is, these two models share a portion of the underlying feature vectors. The click-through rate prediction processes based on the input feature vectors of the item ranking model and the creative selection model are performed in parallel. (Refer to...) Figure 5 First, an item recall model is used to recall a massive pool of items on an e-commerce platform. Then, an item ranking model and a creative selection model are used to predict the click-through rate of each recalled candidate item and each candidate creative corresponding to each candidate item in parallel online, obtaining a first prediction result and a second prediction result. Subsequently, recommendation information for target users is determined based on these two prediction results.

[0090] In some embodiments, the above method may further include: determining the average click-through rate (CTR) for multiple users based on the training dataset; and accordingly, processing the user feature vector and the creative feature vector using a creative selection model to obtain a second prediction result, which may include: processing the user feature vector, the creative feature vector, and the average CTR using a creative selection model to obtain a second prediction result.

[0091] For example, as can be seen from the above, the training dataset may include positive samples and negative samples; refer to Figure 6 Both the item ranking model and the creative selection model are MLP models. The input features of the item ranking model include user feature vectors and item feature vectors, while the input features of the creative selection model include user feature vectors, creative feature vectors, and average click-through rate.

[0092] For example, the average click-through rate (CTR), also known as the fixed CTR value, is related to the number of positive and negative samples in the training dataset. In one embodiment, the ratio of the number of positive samples to the total number of samples can be determined as the average CTR for multiple users. For example, assuming the training dataset contains a total of 10,000 samples, of which 2,000 are positive samples and 8,000 are negative samples, the ratio of 2,000 to 10,000, 0.2, can be determined as the average CTR for multiple users.

[0093] As can be seen, in this embodiment of the application, when using the creative selection model to predict the click-through rate, in addition to inputting the user feature vector and creative features, the average click-through rate is also input. Since the average click-through rate is related to the items clicked by each user in the training dataset, inputting it into the creative selection model helps it perceive item-related information and improves the prediction effect of the creative selection model.

[0094] Step 202: Determine the recommended information based on the first and second prediction results.

[0095] Here, the recommendation information may include at least one target item from multiple candidate items and a target creative corresponding to each target item. The target creative is the creative that can be combined with the target item from among the multiple candidate creatives corresponding to each target item.

[0096] For example, as described above, the first prediction result may include the first estimated click-through rate of multiple candidate items, and the second prediction result may include the second estimated click-through rate of each candidate creative corresponding to each candidate item. In some embodiments, determining recommendation information based on the first and second prediction results may include: sorting multiple candidate items based on the first estimated click-through rate to obtain a first sorting result; determining at least one target item from the multiple candidate items based on the first sorting result; sorting each candidate creative corresponding to each target item based on the second estimated click-through rate to obtain at least one second sorting result; determining each target creative combined with the at least one target item based on the at least one second sorting result; and combining the at least one target item and each target creative to obtain recommendation information.

[0097] For example, after obtaining the first estimated click-through rate of each candidate item among multiple candidate items, the multiple candidate items can be sorted in descending order to obtain a first sorting result; then, based on the first sorting result, at least one target item with the highest ranking is determined from the multiple candidate items; here, at least one target item is the items to be recommended to the target user.

[0098] Furthermore, after obtaining the second estimated click-through rate of at least one target item and each candidate creative corresponding to each target item, the candidate creatives corresponding to each target item are sorted in descending order to obtain at least one second sorting result. From each second sorting result, the top-ranked target creative is determined to obtain at least one target creative. Then, the at least one target item is combined with the at least one target creative to obtain at least one combination result, and this at least one combination result is determined as the recommendation information to be recommended to the target user. Finally, the at least one combination result is displayed to the target user to attract the target user's attention, improve the target user's click-through rate, and increase the traffic value of the e-commerce platform.

[0099] As can be seen, the information recommendation method provided in this application, after obtaining the feature vectors corresponding to the target user, can simultaneously use the item ranking model and the creative selection model to perform parallel prediction processing on the feature vectors. That is, the item ranking model and the creative selection model can share the time consumption. In this way, the problem of excessive time consumption caused by the need for sequential execution of these two models in related technologies can be solved, thereby improving recommendation efficiency. In addition, when the time required for one information recommendation is fixed, since the item ranking model and the creative selection model can save a certain amount of time, if this saved time is used to improve the complexity and iteration space of the model, the accuracy of information recommendation can be ensured.

[0100] Figure 7 This is a schematic diagram of the composition structure of the model training device according to an embodiment of this application, as shown below. Figure 7 As shown, the device includes: an acquisition module 300, a processing module 301, and a training module 302, wherein:

[0101] The acquisition module 300 is used to acquire a training dataset based on the historical behavior data of multiple users; the training dataset includes multiple item samples and label information for each item sample, the label information being used to characterize whether the item sample was clicked after each user browsed it;

[0102] Processing module 301 is used to process the training dataset using an initial item sorting model to obtain a first processing result; and to process the training dataset and the first processing result using an initial creative selection model to obtain a second processing result.

[0103] Training module 302 is used to jointly train the initial item sorting model and the initial creative selection model based on the first processing result and the second processing result to obtain the item sorting model and the creative selection model.

[0104] In some embodiments, the processing module 301 is further configured to:

[0105] Based on the training dataset, determine the user sample feature vector, item sample feature vector, and creative sample feature vector for each user;

[0106] The user sample feature vector and the item sample feature vector are processed using the initial item sorting model to obtain the first processing result;

[0107] The user sample feature vector, the creative sample feature vector, and the first processing result are processed using the initial creative selection model to obtain the second processing result.

[0108] In some embodiments, the training module 302 is further configured to: jointly train the initial item ranking model and the initial creative selection model based on the first processing result and the second processing result to obtain the item ranking model and the creative selection model, including:

[0109] Based on the first processing result and the tag information, the first loss is determined;

[0110] Based on the second processing result and the tag information, the second loss is determined;

[0111] Based on the first loss and the second loss, a target loss is determined, and the initial item ranking model and the initial creative selection model are jointly trained based on the target loss until convergence, thereby obtaining the item ranking model and the creative selection model.

[0112] Figure 8 This is a schematic diagram of the composition structure of the information recommendation device according to an embodiment of this application, such as... Figure 8 As shown, the device includes: a first determining module 400, a predicting module 401, and a second determining module 402, wherein:

[0113] The first determining module 400 is used to determine the user feature vector of the target user, the item feature vectors of multiple candidate items, and the creative feature vectors of each candidate idea corresponding to the multiple candidate items based on the target user's historical behavior data.

[0114] The prediction module 401 is used to process the user feature vector and the item feature vector using an item sorting model to obtain a first prediction result; and to process the user feature vector and the creative feature vector using a creative selection model to obtain a second prediction result.

[0115] The second determining module 402 is used to determine recommendation information based on the first prediction result and the second prediction result; the recommendation information includes at least one target item among the plurality of candidate items and a target idea corresponding to each target item; wherein the item ranking model and the idea selection model are obtained according to the model training method provided by one or more of the aforementioned technical solutions.

[0116] In some embodiments, the first prediction result includes a first estimated click-through rate of the plurality of candidate items, the second prediction result includes a second estimated click-through rate of each candidate creative corresponding to each candidate item, and the second determining module 402 is further configured to:

[0117] The candidate items are sorted based on the first estimated click-through rate to obtain a first sorting result;

[0118] Based on the first sorting result, at least one target item is determined from the plurality of candidate items;

[0119] Based on the second estimated click-through rate, the candidate creatives corresponding to each of the at least one target item are sorted to obtain at least one second sorting result;

[0120] Based on the at least one second sorting result, each target idea that is combined with the at least one target item is determined;

[0121] The recommendation information is obtained by combining the at least one target item and each target idea.

[0122] In some embodiments, the first determining module 400 is further configured to: determine the average click-through rate of the plurality of users for each item sample based on the training dataset;

[0123] Accordingly, the prediction module 401 is further configured to: process the user feature vector, the creative feature vector and the average click-through rate using a creative selection model to obtain the second prediction result.

[0124] In some embodiments, the first determining module 400 is further configured to:

[0125] Feature extraction is performed on the historical behavior data of the target user to obtain the initial user features of the target user; based on the historical behavior data, the item information of the multiple candidate items is determined, and feature extraction is performed on the item information of the multiple candidate items to obtain the initial item features of the multiple candidate items and the initial creative features of each candidate creative corresponding to the multiple candidate items.

[0126] The initial user features, the initial item features, and the initial creative features are hashed respectively to obtain the user feature vector of the target user, the item feature vectors of multiple candidate items, and the creative feature vectors of each candidate creative corresponding to the multiple candidate items.

[0127] In practical applications, the acquisition module 300, processing module 301, training module 302, first determination module 400, prediction module 401, second determination module 402, and training module can all be implemented by a processor located in an electronic device. The processor can be at least one of ASIC, DSP, DSPD, PLD, FPGA, CPU, controller, microcontroller, and microprocessor.

[0128] Furthermore, in this embodiment, the functional modules can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional module.

[0129] If the integrated unit is implemented as a software functional module and is not sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the method of this embodiment. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0130] Specifically, the computer program instructions corresponding to a model training method or information recommendation method in this embodiment can be stored on storage media such as optical discs, hard disks, and USB flash drives. When the computer program instructions corresponding to a model training method or information recommendation method in the storage media are read or executed by an electronic device, any of the model training methods or information recommendation methods in the aforementioned embodiments are implemented.

[0131] Based on the same technical concept as the foregoing embodiments, see Figure 9 It illustrates the electronic device 500 provided in this application, which may include: a memory 501 and a processor 502; wherein,

[0132] Memory 501 is used to store computer programs and data;

[0133] The processor 502 is used to execute a computer program stored in the memory to implement any of the model training methods or information recommendation methods in the foregoing embodiments.

[0134] In practical applications, the aforementioned memory 501 can be volatile memory, such as RAM; or non-volatile memory, such as 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 502.

[0135] The processor 502 described above can be at least one of ASIC, DSP, DSPD, PLD, FPGA, CPU, controller, microcontroller, and microprocessor. It is understood that for different object attribute determination devices, the electronic device used to implement the above processor function can also be other types, and this application embodiment does not specifically limit the specific types.

[0136] In some embodiments, the functions or modules of the apparatus provided in this application can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0137] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities can be referred to each other. For the sake of brevity, they will not be repeated here.

[0138] The methods disclosed in the various method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict.

[0139] The features disclosed in the various product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0140] The features disclosed in the various method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0141] 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.

[0142] 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, create means for implementing the functions specified in one or more flowchart illustrations and / or one or more block diagrams.

[0143] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0144] The above are merely preferred embodiments of this application and are not intended to limit the scope of protection of this application.

Claims

1. A model training method, characterized in that, The method further includes: A training dataset is obtained based on the historical behavior data of multiple users; the training dataset includes multiple item samples and label information for each item sample, the label information being used to characterize whether the item sample was clicked after each user browsed it; The training dataset is processed using an initial item ranking model to obtain a first processing result; the training dataset and the first processing result are processed using an initial creative selection model to obtain a second processing result; the first processing result includes the estimated click-through rate for each user on different item samples; the second processing result includes the estimated click-through rate for each user on each creative corresponding to different item samples. Based on the first processing result and the second processing result, the initial item sorting model and the initial creative selection model are jointly trained to obtain the item sorting model and the creative selection model.

2. The method according to claim 1, characterized in that, The method further includes: Based on the training dataset, determine the user sample feature vector, item sample feature vector, and creative sample feature vector for each user; Accordingly, the step of processing the training dataset using the initial item sorting model to obtain a first processing result includes: The user sample feature vector and the item sample feature vector are processed using the initial item sorting model to obtain the first processing result; Accordingly, the step of processing the training dataset and the first processing result using the initial creative selection model to obtain the second processing result includes: The user sample feature vector, the creative sample feature vector, and the first processing result are processed using the initial creative selection model to obtain the second processing result.

3. The method according to claim 2, characterized in that, The step of jointly training the initial item ranking model and the initial creative selection model based on the first processing result and the second processing result to obtain the item ranking model and the creative selection model includes: Based on the first processing result and the tag information, the first loss is determined; Based on the second processing result and the tag information, the second loss is determined; Based on the first loss and the second loss, a target loss is determined, and the initial item ranking model and the initial creative selection model are jointly trained based on the target loss until convergence, thereby obtaining the item ranking model and the creative selection model.

4. An information recommendation method, characterized in that, The method includes: Based on the target user's historical behavior data, determine the target user's user feature vector, the item feature vectors of multiple candidate items, and the creative feature vectors of each candidate idea corresponding to the multiple candidate items; The user feature vector and the item feature vector are processed using an item ranking model to obtain a first prediction result; the user feature vector and the creative feature vector are processed using a creative selection model to obtain a second prediction result. Based on the first prediction result and the second prediction result, recommendation information is determined; the recommendation information includes at least one target item among the plurality of candidate items and a target idea corresponding to each target item; wherein, the item ranking model and the idea selection model are obtained by the method according to any one of claims 1 to 3.

5. The method according to claim 4, characterized in that, The first prediction result includes the first estimated click-through rate of the plurality of candidate items, and the second prediction result includes the second estimated click-through rate of each candidate creative corresponding to each candidate item. Determining the recommendation information based on the first prediction result and the second prediction result includes: The candidate items are sorted based on the first estimated click-through rate to obtain a first sorting result; Based on the first sorting result, at least one target item is determined from the plurality of candidate items; Based on the second estimated click-through rate, the candidate creatives corresponding to each of the at least one target item are sorted to obtain at least one second sorting result; Based on the at least one second sorting result, each target idea that is combined with the at least one target item is determined; The recommendation information is obtained by combining the at least one target item and each target idea.

6. The method according to claim 4, characterized in that, The method further includes: Based on the training dataset, determine the average click-through rate of the multiple users for each item sample; Accordingly, the step of processing the user feature vector and the creative feature vector using the creative selection model to obtain the second prediction result includes: The second prediction result is obtained by processing the user feature vector, the creative feature vector, and the average click-through rate using a creative selection model.

7. The method according to claim 4, characterized in that, The step of determining the target user's user feature vector, multiple candidate item feature vectors, and the creative feature vectors of each candidate creative corresponding to the multiple candidate items based on the target user's historical behavior data includes: Feature extraction is performed on the historical behavior data of the target user to obtain the initial user features of the target user; based on the historical behavior data, the item information of the multiple candidate items is determined, and feature extraction is performed on the item information of the multiple candidate items to obtain the initial item features of the multiple candidate items and the initial creative features of each candidate creative corresponding to the multiple candidate items. The initial user features, the initial item features, and the initial creative features are hashed respectively to obtain the user feature vector of the target user, the item feature vectors of multiple candidate items, and the creative feature vectors of each candidate creative corresponding to the multiple candidate items.

8. A model training device, characterized in that, The device includes: The acquisition module is used to acquire a training dataset based on the historical behavior data of multiple users; the training dataset includes multiple item samples and label information for each item sample, the label information being used to characterize whether the item sample was clicked after each user browsed it; The processing module is used to process the training dataset using an initial item ranking model to obtain a first processing result; and to process the training dataset and the first processing result using an initial creative selection model to obtain a second processing result; the first processing result includes the estimated click-through rate of each user for different item samples; the second processing result includes the estimated click-through rate of each user for each item creative corresponding to different item samples. The training module is used to jointly train the initial item sorting model and the initial creative selection model based on the first processing result and the second processing result, so as to obtain the item sorting model and the creative selection model.

9. An electronic device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method according to any one of claims 1 to 7.

10. A computer storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method described in any one of claims 1 to 7.

Images