A training method and an interaction prediction method, device, storage medium, electronic equipment and program product of an interaction prediction model in a page display scenario

Abstract

The application discloses a kind of training methods of interactive prediction model in page display scene, interactive prediction method, device, storage medium, electronic equipment and program product. In which training method includes: obtaining sample data set, the characteristic information of sample object includes first characteristic information and second characteristic information;First characteristic information and second characteristic information of sample object are input to teacher network model, and first prediction interaction index is obtained;First characteristic information of sample object is input to student network model, and second prediction interaction index is obtained;Generation target loss function is adjusted to student network model parameter, until the student network model trained is obtained. Through the way of mixed distillation (at least two of structure distillation, feature distillation, task distillation) input data simple, the interactive prediction model of simple model structure is obtained, in the prediction process of interaction, the prediction accuracy is high, processing efficiency is high, and the consumption of computing resources is less.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, and in particular to a training method, interaction prediction method, device, storage medium, electronic device, and program product for an interaction prediction model in a page display scenario. Background Technology

[0002] With the rapid development of internet technology and the digital economy, online shopping has become an integral part of people's daily lives. Online shopping platforms showcase products for sale, and the interactions between the displayed products and the user are crucial factors influencing the products shown on the e-commerce platform's display page. Currently, deep learning models can be used to predict these interactions, providing a reference for deciding which products to display on the e-commerce platform's page.

[0003] In the process of realizing this invention, it was found that at least the following technical problems exist in the prior art: current deep learning models suffer from high computational overhead and poor prediction accuracy due to limitations in input features. Summary of the Invention

[0004] This invention provides a training method, an interaction prediction method, a device, a storage medium, an electronic device, and a program product for an interaction prediction model in a page display scenario, thereby reducing the computational overhead of the interaction indicator prediction process and improving the prediction accuracy of the interaction indicator.

[0005] According to one aspect of the present invention, a method for training an interaction prediction model in a page display scenario is provided, comprising:

[0006] Obtain a sample dataset, which includes multiple sample data. The sample data includes feature information and interaction tag information. The feature information includes first feature information obtained before the sample object is displayed on the display page and second feature information obtained after it is displayed on the display page.

[0007] The first feature information and the second feature information in the sample data are input into the teacher network model to obtain the first predictive interaction index; the first feature information in the sample data is input into the student network model to obtain the second predictive interaction index.

[0008] A target loss function is generated based on the interaction label information, the first predicted interaction index, and the second predicted interaction index in the sample data.

[0009] With the goal of reducing the target loss function as the training objective, the parameters of the student network model are adjusted until a well-trained student network model is obtained. The well-trained student network model is then identified as the interaction prediction model, which is used to predict the interaction metrics corresponding to the interactive operations set by the object to be displayed during the display process on the display page.

[0010] Optionally, the second feature information includes one or more of the display position information of the sample object and the interaction information on the display page;

[0011] The first feature information includes one or more of the object information of the sample object, user behavior feature information, user non-behavioral feature information, and user-object cross-combination information; wherein, the user non-behavioral feature information is the feature information other than the user behavior feature information in the user-related feature information, and the user-object cross-combination information is the feature information obtained by combining the object information of the sample object with the user non-behavioral feature information.

[0012] Optionally, generating the target loss function based on the interaction label information, the first predicted interaction index, and the second predicted interaction index in the sample data includes: generating a cross-entropy loss function based on the interaction label information and the second predicted interaction index in the sample data; generating a distillation loss function based on the first predicted interaction index and the second predicted interaction index; and determining the target loss function based on the cross-entropy loss function and the distillation loss function.

[0013] Optionally, the set interaction operation includes a transaction operation and a pre-transaction operation; the first predictive interaction indicator and the second predictive interaction indicator respectively include a first interaction indicator corresponding to the transaction operation and a second interaction indicator corresponding to the pre-transaction operation; the interaction tag information in the sample data includes a first tag corresponding to the transaction operation and a second tag corresponding to the pre-transaction operation.

[0014] The method further includes: determining sample data with label consistency based on the first label and the second label of the sample object; obtaining the second predictive interaction index corresponding to the sample data with label consistency, wherein the second predictive interaction index includes the first interaction index and the second interaction index; generating a consistency loss function based on the first interaction index and the second interaction index included in the second predictive interaction index; and determining the target loss function based on the cross-entropy loss function, the distillation loss function, and the consistency loss function.

[0015] Optionally, the student network model includes: a first feature extraction network block, a second feature extraction network block, and a multi-task processing network block, wherein the first feature extraction network block and the second feature extraction network block are respectively connected to the multi-task processing feature extraction network block.

[0016] Optionally, the step of inputting the first feature information from the sample data into the student network model to obtain the second predicted interaction index includes: inputting the first feature information into a first feature extraction network block for feature extraction to obtain user behavior sequence semantic features; inputting one or more of the object information of the sample object, the user non-behavioral feature information, and the user-object cross-combination information from the first feature information into a second feature extraction network block for feature extraction to obtain user-object implicit cross semantic features; and inputting the user behavior sequence semantic features and the user-object implicit cross semantic features into the multi-task processing network block for index prediction to obtain the second predicted interaction index.

[0017] Optionally, the teacher network model includes: a third feature extraction network block, a fourth feature extraction network block, a fifth feature extraction network block, a parameter personalization network block, and a multi-task processing network block, wherein the third feature extraction network block, the fourth feature extraction network block, and the fifth feature extraction network block are respectively connected to the parameter personalization network block, and the parameter personalization network block is connected to the multi-task processing network block.

[0018] Optionally, inputting the first feature information and the second feature information from the sample data into the teacher network model to obtain the first predicted interaction index includes: inputting the first feature information into the third feature extraction network block for feature extraction to obtain user behavior sequence semantic features; inputting one or more of the user non-behavioral feature information, the object information of the sample object, and the user-object cross-combination information from the first feature information into the fourth feature extraction network block for feature extraction to obtain user-object implicit cross semantic features; inputting the second feature information into the fifth feature extraction network block for feature extraction to obtain user-object explicit cross semantic features; inputting the user behavior sequence semantic features, the user-object implicit cross semantic features, the user-object explicit cross semantic features, and the second feature information into the parameter personalization network block to obtain personalized features; and inputting the personalized features into the multi-task processing network block for index prediction to obtain the first predicted interaction index.

[0019] According to another aspect of the present invention, an interactive prediction method is provided, comprising:

[0020] Obtain the feature information of the object to be displayed in the pre-display stage;

[0021] The feature information is input into a pre-trained interaction prediction model to obtain the predicted interaction indicators corresponding to the display process of the object to be displayed on the display page. The predicted interaction indicators include one or more of the first interaction indicator corresponding to the transaction operation and the second interaction indicator corresponding to the pre-transaction operation. The interaction prediction model is trained based on the training method of the interaction prediction model in the page display scenario provided in the embodiments of this disclosure.

[0022] According to another aspect of the present invention, a training apparatus for an interaction prediction model in a page display scenario is provided, comprising:

[0023] The sample acquisition module is used to acquire a sample dataset, which includes multiple sample data. The sample data includes feature information and interaction tag information. The feature information includes first feature information acquired before the sample object is displayed on the display page and second feature information acquired after it is displayed on the display page.

[0024] The data processing module is used to input the first feature information and the second feature information in the sample data into the teacher network model to obtain the first predictive interaction index; and to input the first feature information in the sample data into the student network model to obtain the second predictive interaction index.

[0025] The loss function determination module is used to generate a target loss function based on the interaction label information, the first predicted interaction index, and the second predicted interaction index in the sample data.

[0026] The model tuning module is used to tune the parameters of the student network model with the goal of reducing the target loss function, until a well-trained student network model is obtained, and the trained student network model is determined as the interactive prediction model.

[0027] The interaction prediction model is used to predict the interaction metrics corresponding to the interactive operations set by the object to be displayed during the display process on the display page.

[0028] According to another aspect of the present invention, an interactive prediction apparatus is provided, comprising:

[0029] The feature acquisition module is used to acquire feature information of the object to be displayed in the pre-display stage;

[0030] The prediction module is used to input the feature information into a pre-trained interaction prediction model to obtain the interaction indicators corresponding to the display process of the object to be displayed on the display page. The interaction indicators include one or more of the first predicted interaction indicator corresponding to the transaction operation and the second predicted interaction indicator corresponding to the pre-transaction operation. The interaction prediction model is trained based on the training method of the interaction prediction model in the page display scenario provided in the embodiments of this disclosure.

[0031] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0032] At least one processor; and

[0033] A memory communicatively connected to the at least one processor; wherein,

[0034] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the training method of the interaction prediction model in the page display scenario according to any embodiment of the present invention, or the interaction prediction method.

[0035] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions, the computer instructions being configured to cause a processor to execute and implement, when executed, a training method for an interaction prediction model in a page display scenario as described in any embodiment of the present invention, or an interaction prediction method.

[0036] The technical solution of this invention obtains a lighter interactive prediction model compared to the teacher network model through knowledge distillation. In the knowledge distillation process, the lightweight student network model undergoes structural distillation based on the complex structure of the teacher network model. Furthermore, the input data for the teacher network model consists of the feature information of the sample object throughout all stages, while the input data for the student network model consists of the feature data from the previous stage. Feature distillation is performed on the student network model to achieve knowledge transfer. This knowledge distillation method yields an interactive prediction model with simpler input data, a simpler model structure, and higher prediction accuracy. This interactive prediction model exhibits higher prediction accuracy, higher processing efficiency, and lower computational resource consumption in predicting interactive indicators.

[0037] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of a training method for an interactive prediction model provided in Embodiment 1 of the present invention;

[0040] Figure 2 This is a schematic diagram of the student network model provided in an embodiment of the present invention;

[0041] Figure 3 This is a flowchart illustrating the processing of the first feature information by the student network model provided in an embodiment of the present invention.

[0042] Figure 4 This is a schematic diagram of the structure of the teacher network model provided in an embodiment of the present invention;

[0043] Figure 5 This is a schematic diagram of the structure of the third feature extraction network block provided in an embodiment of the present invention;

[0044] Figure 6 This is a schematic diagram of the structure of a parameter personalized feature extraction network block provided in an embodiment of the present invention;

[0045] Figure 7 The flowchart of the teacher network model for processing the first feature information and the second feature information provided in the embodiments of the present invention is shown.

[0046] Figure 8 This is a flowchart of an interactive prediction method provided in an embodiment of the present invention;

[0047] Figure 9 This is a schematic diagram of the structure of a training device for an interactive prediction model provided in an embodiment of the present invention;

[0048] Figure 10 This is a schematic diagram of the structure of an interactive prediction device provided in an embodiment of the present invention;

[0049] Figure 11 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0050] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0051] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0052] In this embodiment, the client displays multiple objects on a display page. For example, the display page may include multiple display slots, each display slot showcasing one object. For instance, the display slot may display an image and / or descriptive information of the object. In response to a trigger operation (e.g., a click operation) on any display slot on the display page, an information display page for the object displayed in the triggered slot is displayed. This information display page allows for user interaction with the object. Specifically, the user's set interaction operations on the information display page include, but are not limited to, transaction operations and pre-transaction operations. The interaction metrics of the set interaction operations of the objects displayed on the display page affect the conversion rate of those objects.

[0053] For the objects to be displayed, due to the large number of objects and the limited number of display slots on the display page, the objects are filtered before being displayed on the e-commerce platform's display page. Optionally, interaction metrics are predicted for each object to be displayed. These interaction metrics are predicted indicators corresponding to the set interactive operations performed during the display process. The objects to be displayed can be filtered based on their interaction metrics. These metrics may include a first interaction metric corresponding to a transaction operation and / or a second interaction metric corresponding to a pre-transaction operation. A transaction operation is an operation completed on the object's information display page, such as, but not limited to, checkout and order submission. A pre-transaction operation is an operation representing a transaction intent completed on the object's information display page, such as adding the object to the shopping cart or setting it as a favorite or collection. Correspondingly, the interaction metric corresponding to the set interactive operation can be understood as the probability value of completing the set interactive operation. Correspondingly, the first interaction indicator for a transaction operation can be understood as the probability that a user completes a transaction operation while the object to be displayed is being displayed on the display page; the second interaction indicator for a pre-transaction operation can be understood as the probability that a user completes a pre-transaction operation while the object to be displayed is being displayed on the display page.

[0054] Currently, when using deep learning models to predict interaction metrics for objects to be displayed, in order to improve the accuracy of interaction metric prediction, the complex structure and large parameter scale of deep learning models result in high computational overhead. It is understood that the influencing factors of the interaction metrics of the object to be displayed include the feature information that can be collected before the display and the feature information that can be collected after the display. In this embodiment, the interaction metric prediction stage is the pre-display stage, and the feature information that can be collected after the display cannot be collected, meaning the feature information used for interaction metric prediction is incomplete, leading to poor prediction accuracy.

[0055] To address the aforementioned technical problems, embodiments of the present invention provide a method for training an interaction prediction model in a page display scenario, see below. Figure 1 , Figure 1This is a flowchart of a training method for an interactive prediction model in a page display scenario provided in Embodiment 1 of the present invention. This embodiment is applicable to training a lightweight student network model through knowledge distillation to obtain a lightweight interactive prediction model. During knowledge distillation, a hybrid distillation method (e.g., a combination of at least two of structural distillation, feature distillation, and task distillation) is used to train a lightweight interactive prediction model with high prediction accuracy. This method can be executed by a training device for the interactive prediction model in a page display scenario. This training device can be implemented in hardware and / or software and can be configured in electronic devices such as mobile terminals, servers, and computers. Figure 1 As shown, the method includes:

[0056] S110. Obtain a sample dataset, which includes multiple sample data. The sample data includes feature information and interactive label information. The feature information of the sample object includes first feature information obtained in the pre-display stage and second feature information obtained in the post-display stage.

[0057] S120. Input the first feature information and the second feature information in the sample data into the teacher network model to obtain the first predictive interaction index; input the first feature information in the sample data into the student network model to obtain the second predictive interaction index.

[0058] S130. Generate a target loss function based on the interaction label information in the sample data, the first predicted interaction index, and the second predicted interaction index.

[0059] S140. The parameters of the student network model are adjusted with the goal of reducing the target loss function. If the student network model does not meet the termination condition, step S120 is executed repeatedly. When the student network model meets the termination condition, a trained student network model is obtained, and this trained student network model is determined as the interactive prediction model.

[0060] The interaction prediction model is used to predict the interaction metrics corresponding to the interactive operations set by the object to be displayed during the display process on the display page.

[0061] In this embodiment, the sample dataset includes multiple sample data sets used for training the student network model and the teacher network model. The sample data includes feature information and interaction label information. Sample data can be understood as data generated during the historical display process of a sample object. A sample object can be an object that has already been displayed on the display page. The sample data corresponding to the sample object is obtained by collecting its feature information and the interaction label information set according to the set interactive operations during the display process. The type of sample object is not limited here. Optionally, the same sample object can correspond to multiple sample data sets. During the display process of a sample object on the display page, if any user enters the information display page of the sample object through a trigger operation on the display position of the sample object, the user's set interactive operation on the information display page of the sample object is detected, and a sample data set for that sample object is generated. It is understood that an authorization request is sent to the relevant client before obtaining the sample data. The sample data is obtained after authorization is granted. This sample data is only used for the training process of the interaction prediction model and is deleted after training is completed to avoid data leakage.

[0062] Optionally, a dataset is maintained, which includes sample data within a preset time range prior to the current time. For example, the preset time range could be one month or three months, etc., and is not limited here. Real-time acquired sample data that meets the preset time range is added to the dataset, while sample data exceeding the preset time range is removed. This dataset serves as the aforementioned sample dataset. This sample dataset is used to train the interactive prediction model, and the interactive prediction model is updated periodically to ensure its performance.

[0063] Optionally, the sample data may include positive sample data and negative sample data. Positive sample data may be generated after entering the information display page of the sample object and completing the set interactive operations, while negative sample data may be generated after entering the information display page of the sample object but without completing the set interactive operations. The ratio of positive to negative sample data in the sample dataset may meet a preset ratio, which may be set according to the model training requirements.

[0064] The interaction label information in the sample data can be marked using 0 and 1. For example, the interaction label information in sample data that has completed the set interaction operation can be 1, and the interaction label information in sample data that has not completed the set interaction operation can be 0. Optionally, the interaction label information in the sample data can include a first label corresponding to the transaction operation and / or a second label corresponding to the pre-transaction operation. For a sample object, if the transaction operation is completed, the first label of the sample object is 1, and if the transaction operation is not completed, the first label of the sample object is 0; if the pre-transaction operation is completed, the second label of the sample object is 1, and if the pre-transaction operation is not completed, the second label of the sample object is 0. Accordingly, the interaction label information in the sample data can include (0,0), (1,0), (0,1), and (1,1). Here, 0 and 1 are only one form of interaction label information and can be changed according to requirements, which is not limited here.

[0065] The feature information in the sample data consists of two parts, acquired in two separate stages. Specifically, the feature information includes the first feature information of the sample object acquired before display, and the second feature information acquired after display. The pre-display stage refers to the period before the sample object is displayed on the display page, and the post-display stage refers to the period after the sample object is displayed on the display page. It can be understood that "display" here refers to the current display.

[0066] The first set of characteristic information acquired before the display includes one or more of the following: sample object information, user behavior characteristic information, user non-behavioral characteristic information, and user-object cross-combination information. The sample object information can be understood as information used to introduce the sample object, including but not limited to the sample object's identification information, SKU (Stock Keeping Unit) information, category information, and description information. User behavior characteristic information refers to the behavioral characteristics of the users displayed to the sample object within a historical time period. This behavioral characteristic information includes, but is not limited to, characteristic information corresponding to interactive operations such as adding to cart, transaction, clicking, sharing, and favorites of historically displayed objects. User non-behavioral characteristic information refers to other characteristic information of the users displayed to the sample object besides the behavioral characteristic information. This may include basic user information, including but not limited to user location information and account information. User-related characteristic information is based on multiple characteristic information corresponding to the users displayed to the sample object. User-related characteristic information may include multiple preset types of characteristic information. Characteristic information other than user behavior characteristic information is defined as user non-behavioral characteristic information. User-object cross-combination information refers to feature information obtained by combining the object information of the sample object with the user's non-behavioral feature information, including but not limited to feature information composed of gender and category, and feature relationships composed of age and category. Optionally, for sample objects that have completed a transaction, the above-mentioned object information and user non-behavioral feature information of the sample objects that have completed the transaction are combined, the frequency of occurrence of each combination of feature information is determined, and the combination of feature information with a frequency less than a preset threshold is removed, and the remaining combination of feature information is used as user-object cross-combination information.

[0067] The second feature information acquired in the post-display phase includes one or more of the sample object's display position information and interaction information on the display page. The display page includes multiple display positions used to display relevant information about the sample object. Different display positions have different exposure effects on the sample object. The sample object's display position information can be the identifier or location of the display position where the sample object is located; it can be used to uniquely identify the display position. The interaction information on the display page is the information generated by the user's interactive operations on the information display page in response to the triggering operation of the sample object's display position on the display page. This includes, but is not limited to, interaction information generated by user interactions such as browsing time on the information display page, commenting on, favorited, sharing, commenting and querying, transaction operations, and pre-transaction operations related to the sample object.

[0068] The pre-built student and teacher network models are invoked, and knowledge distillation is performed on them using sample data from the aforementioned dataset to obtain a trained interaction prediction model. The student network model has a simpler structure and fewer parameters than the teacher network model. Through knowledge distillation, a lightweight student network model with a simple structure, fewer parameters, and higher prediction accuracy (compared to before knowledge distillation) is trained as the interaction prediction model. This improves the prediction accuracy of the interaction prediction model while reducing resource consumption during the interaction prediction process, thus increasing computational efficiency and reducing computational overhead.

[0069] In this embodiment, the prediction stage of the interaction index of the object to be displayed is located in the pre-display stage, and the second feature information of the object to be displayed cannot be obtained. In order to improve the prediction performance of the interaction prediction model, the second feature information is transferred through knowledge distillation so that the trained interaction prediction model can guarantee the prediction accuracy of the interaction without the need for the second feature information.

[0070] During training, sample data is extracted from the sample dataset. The first and second feature information from the sample data are used as input to the teacher network model, resulting in the first predicted interaction index output by the teacher network model. Similarly, the first feature information from the sample data is used as input to the student network model, resulting in the second predicted interaction index output by the student network model. The first predicted interaction index can be understood as the predicted probability obtained by the teacher network model using the first and second feature information from the sample data to predict the user's desired interaction during the display of the corresponding sample object on the page. The second predicted interaction index can be understood as the predicted probability obtained by the student network model using the first feature information from the sample data to predict the user's desired interaction during the display of the corresponding sample object on the page.

[0071] A target loss function is generated based on the interaction label information, the first predicted interaction metric, and the second predicted interaction metric in the sample data. This target loss function is then fed back into the student network model to adjust its parameters. Specifically, a cross-entropy loss function is generated based on the interaction label information and the second predicted interaction metric in the sample data; a distillation loss function is generated based on the first and second predicted interaction metric. The target loss function includes one or more of the cross-entropy loss function and the distillation loss function. The type of loss function used for distillation is not limited here.

[0072] For example, the cross-entropy loss function can be expressed as: Where y represents the interactive label information. This serves as the second predictive interaction indicator. For example, y might be the second label corresponding to a pre-trading operation. The second interactive indicator is the one corresponding to the pre-trading operation in the second predictive interactive indicator; or, y is the first label corresponding to the trading operation. The first interaction indicator is the one corresponding to the trading operation in the second predictive interaction indicator. In some embodiments, the cross-entropy loss function may include the sum of the first loss function corresponding to the trading operation and the second loss function corresponding to the pre-trading operation. The first loss function corresponding to the trading operation is generated based on the first interaction indicator in the first label and the second predictive interaction indicator; the second loss function corresponding to the pre-trading operation is generated based on the second label and the second interaction indicator in the second predictive interaction indicator.

[0073] The distillation loss function can be expressed as: Where, r t r is the first predicted interaction metric output by the teacher network model. s This is the second predicted interaction metric output by the student network model. For example, r t The second interaction indicator, r, corresponds to the pre-trading operation in the first predictive interaction indicator output by the teacher network model. s The second interaction indicator corresponding to the pre-trading operation in the second predictive interaction indicator output by the student network model; or, r t The first interaction indicator, r, is the first interaction indicator corresponding to the transaction operation in the first predictive interaction indicator output by the teacher network model. s The first interaction index corresponds to the transaction operation in the second predicted interaction index output by the student network model. τ is a hyperparameter in the knowledge distillation process, representing the temperature during knowledge distillation. σ represents the sigmoid function. In some embodiments, the distillation loss function can be the sum of the distillation loss function corresponding to the transaction operation and the distillation loss function corresponding to the pre-transaction operation, where the distillation loss function corresponding to the transaction operation and the distillation loss function corresponding to the pre-transaction operation can be calculated based on the formulas for the distillation loss functions described above.

[0074] Optionally, the sum of the cross-entropy loss function and the distillation loss function can be used as the target loss function. Alternatively, a weighted sum of the cross-entropy loss function and the distillation loss function can be used as the target loss function.

[0075] In this embodiment, during the interaction with the sample object, a pre-transaction operation identifies the sample object as an object with transaction intent. Here, the second label in the sample data is 1. The transaction operation realizes the transformation of the object to be displayed, and here the first label in the sample data is 1. If no transaction operation is performed, the first label in the sample data is 0. It can be seen that the sample data with the second label of 1 includes the sample data with the first label of 1 and the sample data with the first label of 0. It can be understood that the pre-transaction operation can serve as the basis for the transaction operation.

[0076] The second predictive interaction metric includes a first interaction metric corresponding to a transaction operation and a second interaction metric corresponding to a pre-transaction operation. The first and second interaction metrics are correlated, and the second interaction metric can provide prior information for the first interaction metric of a sample object. Optionally, the sample data in the sample dataset is denoised to remove noisy sample data lacking a first or second label, thus avoiding interference from noisy sample data during the training process.

[0077] Based on the first and second labels of the sample objects, label-consistent sample data is determined. A second predictive interaction index is obtained from the label-consistent sample data, where the second predictive interaction index includes both the first and second interaction indices. A consistency loss function is generated based on the first and second interaction indices included in the second predictive interaction index. A target loss function is determined based on the cross-entropy loss function, the distillation loss function, and the consistency loss function. For the first and second labels in the sample data, if the first and second labels are the same, the sample data is label-consistent. If the first and second labels are different, the sample data is not label-consistent.

[0078] For label-consistent sample data, the second predictive interaction index of the label-consistent sample data includes the first interaction index and the second interaction index. During the training process, the second interaction index, which is relatively dense among the samples, guides the learning of the first interaction index with relative sample coefficients. The loss function is the consistency loss function.

[0079] For example, the consistency loss function can be expressed as: Among them, A s C is the second interaction indicator in the second predictive interaction indicator. s This is the first interaction indicator in the second prediction interaction indicator. The accuracy of the transaction prediction task is further improved by performing task distillation on the transaction trend transformation prediction task in the student network model.

[0080] Accordingly, the target loss function includes one or more of the cross-entropy loss function, the distillation loss function, and the consistency loss function. Optionally, the sum of the cross-entropy loss function, the distillation loss function, and the consistency loss function is used as the target loss function; or, the weighted sum of the cross-entropy loss function, the distillation loss function, and the consistency loss function is used as the target loss function.

[0081] Optionally, the teacher network model can be pre-trained. Optionally, the teacher network model and the student network model are jointly trained. The training process of the teacher network model may include: generating a teacher loss function for the teacher network model based on the interaction label information and the first predicted interaction index of the sample objects in the sample data; and inputting this teacher loss function back into the teacher network model to adjust the model parameters. Optionally, the teacher loss function can be a cross-entropy function. The interaction label information in the sample data includes a first label corresponding to the transaction operation and a second label corresponding to the pre-transaction operation; the first predicted interaction index includes a first interaction index corresponding to the transaction operation and a second interaction index corresponding to the pre-transaction operation.

[0082] A first loss term is generated based on the first interaction indicator and the first label in the first predictive interaction indicator, and a second loss term is generated based on the second interaction indicator and the second label in the first predictive interaction indicator. The first loss term and the second loss term can form the teacher loss function.

[0083] The above training process is iteratively executed to train the student network model. By processing the first and second feature information through the teacher network model and processing the first feature information through the student network model, the student network model is trained through knowledge transfer. This improves the interactive prediction accuracy of the student network model while reducing the amount of data processed and increasing the processing efficiency.

[0084] Based on the above embodiments, optionally, the student network model includes: a first feature extraction network block, a second feature extraction network block, and a multi-task processing network block, wherein the first feature extraction network block and the second feature extraction network block are respectively connected to the multi-task processing network block; for example, see [link to example]. Figure 2 , Figure 2This is a schematic diagram of the student network model provided in an embodiment of the present invention. The first feature extraction network block and the second feature extraction network block are used for feature extraction, and their network structures can be different. A multi-task processing network block can be used to predict a second predictive interaction index. In some embodiments, the first feature extraction network block can be a transformer network block, the second feature extraction network block can be a DNN (Deep Neural Network) network block, and the multi-task processing network block can be an MMOE (Multi-Gated Mixture-of-Experts) network block. The first and second feature extraction network blocks run in parallel and are respectively connected to the multi-task processing network block. The first and second feature extraction network blocks can respectively extract features from a portion of the first feature information. The intermediate features obtained by the first and second feature extraction network blocks are concatenated and then input into the multi-task processing feature extraction network block.

[0085] The process of inputting the first feature information from the sample data into the student network model to obtain the second predicted interaction index includes: inputting the first feature information into a first feature extraction network block for feature extraction to obtain user behavior sequence semantic features; inputting one or more of the object information of the sample object, the user non-behavioral feature information, and the user-object cross-combination information from the first feature information into a second feature extraction network block for feature extraction to obtain user-object implicit cross semantic features; and inputting the user behavior sequence semantic features and the user-object implicit cross semantic features into the multi-task processing network block for index prediction to obtain the second predicted interaction index.

[0086] In this embodiment, before performing interaction prediction, the first and second feature information are discretized to obtain discretized features. After discretization, the first and second feature information are hash-mapped to different feature spaces. Taking the interaction information on the display page in the second feature information as an example, the interaction information on the display page can include OA, OB, OC, etc., where OA can represent the dwell time on the display page, OB can represent the operation of viewing comments on the sample object, and OC can represent the sharing operation on the sample object. The interaction information on the display page after hash mapping can be represented as X1 = hash(OA, OB, OC, ..., OF), where X1 can be a feature vector. Correspondingly, X2 can represent the feature vector obtained by hash mapping the display position information of the sample object. Similarly, the various feature information in the first feature information are hash-mapped to convert them into feature vectors.

[0087] The feature vector corresponding to the first feature information is input into the first feature extraction network block for feature extraction to obtain the user behavior sequence semantic features. The user behavior sequence semantic features can be understood as the sequence semantic features extracted from the user's historical interaction operations. Here, the user's historical interaction operations in the first feature information are each set with a timestamp, which can provide a temporal relationship for the user behavior sequence semantic features. In some embodiments, the feature vector corresponding to the user behavior feature information in the first feature information can be input into the first feature extraction network block for feature extraction to obtain the user behavior sequence semantic features.

[0088] By inputting object information of the sample object, user non-behavioral feature information, and user-object cross-combination information into the second feature extraction network block for feature extraction, implicit user-object cross-semantic features are obtained. It is understood that different user groups may interact differently with objects, meaning there is a hidden correlation between user information and object information. The second feature extraction network block extracts implicit user-object cross-semantic features to represent this hidden correlation. Here, user non-behavioral feature information refers to feature information other than user behavioral feature information, and user-object cross-combination information can include at least one combination of user non-behavioral feature information and object information of the sample object, providing comprehensive cross-feature information for implicit user-object cross-semantic features.

[0089] The second predicted interaction metric is obtained by predicting the semantic features of user behavior sequences and the implicit cross-semantic features between users and objects through a multi-task processing network block. Semantic features of different dimensions are extracted through the first and second feature extraction network blocks, providing multi-dimensional features for the prediction of the interaction metric, improving the comprehensiveness of features, and enhancing the accuracy of interaction metric extraction.

[0090] For example, see Figure 3 , Figure 3 This is a flowchart illustrating the processing of first feature information by the student network model provided in this embodiment of the invention. In the student network model, the first feature extraction network block is a transformer network block, which extracts features from the first feature information to obtain semantic features of the user's historical behavior sequence. The second feature extraction network block is a DNN feature extraction network block, which extracts features from local features in the first feature information. These local features are related to the user and the object, namely, one or more of the object information of the sample object, the user's non-behavioral feature information, and user-object cross-combination information. The outputs of the first and second feature extraction network blocks are concatenated (e.g., concatenated along the channel dimension) and used as input to a multi-task processing feature extraction network block, which is an MMOE feature extraction network block. The multi-task processing feature extraction network block outputs a first interaction index and a second interaction index.

[0091] Based on the above embodiments, optionally, the teacher network model includes: a third feature extraction network block, a fourth feature extraction network block, a fifth feature extraction network block, a parameter personalization network block, and the multi-task processing network block, wherein the third feature extraction network block, the fourth feature extraction network block, and the fifth feature extraction network block are respectively connected to the parameter personalization network block, and the parameter personalization network block is connected to the multi-task processing network block. For example, see [link to example]. Figure 4 , Figure 4 This is a schematic diagram of the structure of the teacher network model provided in an embodiment of the present invention.

[0092] The network structures of the third, fourth, and fifth feature extraction network blocks can be different. These three network blocks are set up in parallel and connected to the parameter personalization network block, which in turn is connected to the multi-task processing network block.

[0093] In some embodiments, the third feature extraction network block can have the same structure as the first feature extraction network block, for example, both can be transformer network blocks. Similarly, the second and fourth feature extraction network blocks can have the same structure, for example, both can be DNN network blocks. The fifth feature extraction network block can be a DCN (Deep Cross Neural Network) network block. The parameter personalization network block can be a PPNet (Parameter Personalized Net) network block, and the multi-task processing network block can be an MMOE network block.

[0094] In some embodiments, the fifth feature extraction network block may include a cross network, a deep network, and a neural network (NN) processing unit, such as a multilayer perceptron. For example, see [link to example]. Figure 5 , Figure 5 This is a schematic diagram of the structure of the fifth feature extraction network block provided in this embodiment of the invention. Cross-network units and deep network units are arranged in parallel and connected to the neural network processing unit, respectively. The cross-network units, deep network units, and neural network processing unit in the third feature extraction network block may each include multiple network layers. The processing procedure of each network layer in the fifth feature extraction network block for the input data can be represented as follows:

[0095] F dcn ={L i =L i-1 L i-1 T w c,i-1 +b c,i-1 +L i-1} i＝1,2,3..n

[0096] Where L represents the input of each layer, n represents the number of layers in the fifth feature extraction network block, and w and b are the model parameters in the fifth feature extraction network block, respectively.

[0097] In some embodiments, the parameter personalization network block may include gate network units, deep network units, and neural network processing units (NN networks). For example, see [link to example]. Figure 6 , Figure 6 This is a schematic diagram of the structure of a parameter personalization network block provided in an embodiment of the present invention.

[0098] The process of the parameter personalization network block processing the input data can be represented as follows:

[0099] Fppnet ={L i =(L i-1 w c,i-1 +b c,i-1 Gate i (L 0)} i＝1,2,3..n

[0100] Here, L represents the input of each layer, and the input of the next network layer is the product of the output of the neural network processing unit in the previous network layer and the feature of the gated network unit. The feature of the gated network unit is called the Gate. i = sigmoid(relu(L0)).

[0101] In this embodiment, a third feature extraction network block (e.g., DCN) is introduced into the teacher network model to enhance the explicit high-order feature interactions between various features. A parameter personalization network block is introduced to enhance the implicit feature interactions and feature selection between advantageous and existing features, thereby strengthening the teacher network model's perception and modeling effectiveness for these advantageous features. Knowledge distillation is performed on the lightweight student network model using the complex teacher network model, thus reducing the model structure of the student network model while maintaining the accuracy of interaction metric predictions.

[0102] The process of inputting the first feature information and the second feature information from the sample data into the teacher network model to obtain the first predicted interaction index includes: inputting the first feature information into the third feature extraction network block for feature extraction to obtain user behavior sequence semantic features; inputting one or more of the user non-behavioral feature information, the object information of the sample object, and the user-object cross-combination information from the first feature information into the fourth feature extraction network block for feature extraction to obtain user-object implicit cross semantic features; inputting the second feature information into the fifth feature extraction network block for feature extraction to obtain user-object explicit cross semantic features; inputting the user behavior sequence semantic features, the user-object implicit cross semantic features, the user-object explicit cross semantic features, and the second feature information into the parameter personalization network block to obtain personalized features; and inputting the personalized features into the multi-task processing network block for index prediction to obtain the first predicted interaction index.

[0103] In this embodiment, the teacher network model includes a third feature extraction network block for extracting semantic features of user behavior sequences and a fourth feature extraction network block for extracting implicit user-object cross-semantic features, as well as a fifth feature extraction network block for extracting explicit user-object cross-semantic features. The second feature information in the sample data is the feature information obtained after the sample object is displayed on the page, including the user's interaction information on the sample object's information display page. The fifth feature extraction network block extracts features from the second feature information input to obtain explicit user-object cross-semantic features, which can characterize the association information between the user and the object during the user's interaction on the sample object's information display page. Optionally, the second feature information can be input into a cross-network unit and a deep network unit respectively. The output features of the cross-network unit and the deep network unit are input into a neural network processing unit, which outputs the explicit user-object cross-semantic features.

[0104] By processing user behavior sequence semantic features, implicit user-object cross semantic features, explicit user-object cross semantic features, and second feature information through a parameter-personalized network block, multi-dimensional feature processing is achieved, improving feature comprehensiveness. Personalized features extracted through the parameter-personalized network block can reflect the personalization and differentiation of different sample objects for different users. Optionally, the second feature information can be input into a gating network unit, while the user behavior sequence semantic features, implicit user-object cross semantic features, and explicit user-object cross semantic features can be input into a deep network unit and a neural network processing unit. The fused features of the output features of the gating network unit and the output features of the deep network unit are then input into the neural network processing unit to obtain the personalized features output by the neural network processing unit.

[0105] The personalized features are predicted by a multi-task processing network block to obtain the first prediction interaction index.

[0106] For example, see Figure 7 , Figure 7This is a flowchart illustrating the processing of the first and second feature information by the teacher network model provided in this embodiment of the invention. In the teacher network model, the third feature extraction network block, a transformer network block, extracts features from the first feature information to obtain user behavior sequence semantic features. The fourth feature extraction network block, a DNN network block, extracts features from local features in the first feature information to obtain user-object implicit cross-semantic features. These local features are related information about the user and the object, namely, one or more of the object information of the sample object, user non-behavioral feature information, and user-object cross-combination information. The fifth feature extraction network block, a DCN network block, extracts features from the second feature information (i.e., X1 and X2 mentioned above) to obtain user-object explicit cross-semantic features. The user behavior sequence semantic features, user-object implicit cross-semantic features, user-object explicit cross-semantic features, and the second feature information are input to a parameter personalization network block (which can be a PPNET network block), which outputs personalized features. These personalized features are then input to a multi-task processing network block, which outputs a first predicted interaction index.

[0107] The technical solution of this embodiment obtains a lightweight interactive prediction model through knowledge distillation. During knowledge distillation, a structural distillation is performed on a lightweight student network model based on a complex teacher network model. Furthermore, the input data for the teacher network model consists of the feature information of the sample object throughout all stages, while the output data for the student network model consists of the feature data from the previous stage. Feature distillation is performed on the student network model to achieve knowledge transfer. Additionally, task distillation is performed on the transaction interaction prediction task through the transaction trend interaction prediction task in the student network, improving the prediction accuracy of the transaction accuracy information prediction task. This knowledge distillation method yields an interactive prediction model with simple input data, a simple model structure, and higher prediction accuracy. This interactive prediction model exhibits high prediction accuracy, high processing efficiency, and low computational resource consumption during the interaction prediction process.

[0108] Figure 8 This is a flowchart of an interaction prediction method provided by an embodiment of the present invention. This embodiment is applicable to situations where a lightweight interaction prediction model trained through knowledge distillation is used to predict interaction metrics of an object to be displayed. This method can be executed by an interaction prediction device, which can be implemented in hardware and / or software. The interaction prediction device can be configured in electronic devices such as mobile terminals, servers, and computers. Figure 8 As shown, the method includes:

[0109] S210. Obtain the feature information of the object to be displayed in the pre-display stage.

[0110] S220. Input the feature information into a pre-trained interaction prediction model to obtain the predicted interaction indicators corresponding to the display process of the object to be displayed on the display page, wherein the predicted interaction indicators include one or more of the first interaction indicator corresponding to the transaction operation and the second interaction indicator corresponding to the pre-transaction operation.

[0111] In this embodiment, the interaction prediction model is trained based on the training method of the interaction prediction model in the page display scenario.

[0112] Optionally, upon detecting a preset trigger operation on the client side, the feature information of the object to be displayed in the pre-display stage is obtained, and the predicted interaction metrics of the object to be displayed are determined. The preset trigger operation can be an application startup operation, where the application displays the object through a display page. Alternatively, the preset trigger operation can be a page refresh operation.

[0113] For each object to be displayed, the predicted interaction metrics of the object are determined to identify the target object to be displayed on the display page, and the target object is then displayed on the display page. The number of target objects can be determined based on the number of display slots on the display page. For example, multiple objects to be displayed can be sorted based on the predicted interaction metrics, and the target object can be determined based on this sorting. Optionally, multiple objects to be displayed can be sorted based on a first interaction metric among the predicted interaction metrics, or based on a weighted average of the first and second interaction metrics among the predicted interaction metrics.

[0114] The technical solution of this embodiment acquires feature information before the object to be displayed is displayed, and processes it through a pre-trained lightweight interactive prediction model. This results in high prediction accuracy, high processing efficiency, and low computational resource consumption.

[0115] Figure 9 This is a schematic diagram of the structure of a training device for an interactive prediction model provided in an embodiment of the present invention. Figure 9 As shown, the device includes:

[0116] The sample acquisition module 310 is used to include feature information of sample objects and interactive tag information in the sample data in the sample dataset. The feature information includes first feature information obtained before the sample object is displayed on the display page and second feature information obtained after the sample object is displayed on the display page.

[0117] The data processing module 320 is used to input the first feature information and the second feature information in the sample data into the teacher network model to obtain the first predictive interaction index; and to input the first feature information in the sample data into the student network model to obtain the second predictive interaction index.

[0118] The loss function determination module 330 is used to generate a target loss function based on the interaction label information, the first predicted interaction index and the second predicted interaction index in the sample data.

[0119] The model tuning module 340 is used to tune the parameters of the student network model with the goal of reducing the target loss function, until a well-trained student network model is obtained, and the trained student network model is determined as the interactive prediction model.

[0120] The interaction prediction model is used to predict the interaction metrics corresponding to the interactive operations set by the object to be displayed during the display process on the display page.

[0121] Based on the above embodiments, optionally, the second feature information includes one or more of the display position information of the sample object and the interaction information on the display page;

[0122] The first feature information includes one or more of the object information of the sample object, user behavior feature information, user non-behavioral feature information, and user-object cross-combination information; wherein, the user non-behavioral feature information is the feature information other than the user behavior feature information in the user-related feature information, and the user-object cross-combination information is the feature information obtained by combining the object information of the sample object with the user non-behavioral feature information.

[0123] Based on the above embodiments, optionally, the loss function determination module 330 is used to: generate a cross-entropy loss function based on the interaction label information in the sample data and the second predicted interaction index; generate a distillation loss function based on the first predicted interaction index and the second predicted interaction index; and determine the target loss function based on the cross-entropy loss function and the distillation loss function.

[0124] Optionally, the setting interaction operation includes transaction operation and pre-transaction operation;

[0125] The first predictive interaction indicator and the second predictive interaction indicator respectively include the first interaction indicator corresponding to the transaction operation and the second interaction indicator corresponding to the pre-transaction operation; the interaction tag information in the sample data includes the first tag corresponding to the transaction operation and the second tag corresponding to the pre-transaction operation.

[0126] The loss function determination module 330 is further configured to: determine sample data with label consistency based on the first label and the second label of the sample object; obtain the second predictive interaction index of the sample data with label consistency, wherein the second predictive interaction index includes the first interaction index and the second interaction index; generate a consistency loss function based on the first interaction index and the second interaction index included in the second predictive interaction index; and determine the target loss function based on the cross-entropy loss function, the distillation loss function, and the consistency loss function.

[0127] Based on the above embodiments, optionally, the student network model includes: a first feature extraction network block, a second feature extraction network block, and a multi-task processing network block, wherein the first feature extraction network block and the second feature extraction network block are respectively connected to the multi-task processing feature extraction network block;

[0128] The data processing module 320 is used to: input the first feature information into a first feature extraction network block for feature extraction to obtain user behavior sequence semantic features; input one or more of the object information of the sample object, the user non-behavioral feature information, and the user-object cross-combination information in the first feature information into a second feature extraction network block for feature extraction to obtain user-object implicit cross-semantic features; and input the user behavior sequence semantic features and the user-object implicit cross-semantic features into the multi-task processing network block for index prediction to obtain the second predicted interaction index.

[0129] Optionally, the teacher network model includes: a third feature extraction network block, a fourth feature extraction network block, a fifth feature extraction network block, a parameter personalization network block, and a multi-task processing network block, wherein the third feature extraction network block, the fourth feature extraction network block, and the fifth feature extraction network block are respectively connected to the parameter personalization network block, and the parameter personalization network block is connected to the multi-task processing network block.

[0130] Optionally, the data processing module 320 is further configured to: input the first feature information into the third feature extraction network block for feature extraction to obtain user behavior sequence semantic features; input one or more of the user non-behavioral feature information, the object information of the sample object, and the user-object cross-combination information in the first feature information into the fourth feature extraction network block for feature extraction to obtain user-object implicit cross semantic features; and input the second feature information into the fifth feature extraction network block for feature extraction to obtain user-object explicit cross semantic features.

[0131] The user behavior sequence semantic features, the user-object implicit cross semantic features, the user-object explicit cross semantic features, and the second feature information are input into the parameter personalization network block to obtain personalized features; the personalized features are input into the multi-task processing network block for index prediction to obtain the first predicted interaction index.

[0132] The training device for the interaction prediction model in the page display scenario provided in the embodiments of the present invention can execute the training method for the interaction prediction model in the page display scenario provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0133] Figure 10 This is a schematic diagram of the structure of an interactive prediction device provided in an embodiment of the present invention. Figure 10 As shown, the device includes:

[0134] Feature acquisition module 410 is used to acquire feature information of the object to be displayed in the pre-display stage;

[0135] The prediction module 420 is used to input the feature information into a pre-trained interaction prediction model to obtain the interaction indicators corresponding to the display process of the object to be displayed on the display page. The interaction indicators include one or more of the first predicted interaction indicator corresponding to the transaction operation and the second predicted interaction indicator corresponding to the pre-transaction operation. The interaction prediction model is trained based on the training method of the interaction prediction model in the page display scenario provided in any of the above embodiments.

[0136] The interactive prediction device provided in the embodiments of the present invention can execute the interactive prediction method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0137] Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0138] like Figure 11As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 may also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0139] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0140] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as training methods for interactive prediction models in a page display scene, or interactive prediction methods.

[0141] In some embodiments, the method for training an interaction prediction model in a page display scenario, or the interaction prediction method, may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, the method for training an interaction prediction model in a page display scenario, or one or more steps of the interaction prediction method, described above, may be executed. Alternatively, in other embodiments, processor 11 may be configured by any other suitable means (e.g., by means of firmware) to execute the method for training an interaction prediction model in a page display scenario, or the interaction prediction method.

[0142] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0143] The training method for the interaction prediction model in the page display scenario for implementing the present invention, or the computer program for the interaction prediction method, can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer program causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer program can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0144] This invention also provides a computer-readable storage medium storing computer instructions for causing a processor to execute a training method for an interaction prediction model in a page display scenario, the method comprising:

[0145] A sample dataset is obtained, comprising feature information of sample objects and interaction label information. The feature information includes first feature information acquired before the sample object is displayed on the display page and second feature information acquired after the object is displayed on the display page. The first and second feature information from the sample data are input into a teacher network model to obtain a first predicted interaction index. The first feature information from the sample data is input into a student network model to obtain a second predicted interaction index. A target loss function is generated based on the interaction label information, the first predicted interaction index, and the second predicted interaction index from the sample data. The parameters of the student network model are adjusted with the goal of reducing the target loss function until a well-trained student network model is obtained. The well-trained student network model is determined as the interaction prediction model, which is used to predict the interaction index corresponding to the set interaction operation of the object to be displayed on the display page.

[0146] Alternatively, a computer-readable storage medium stores computer instructions for causing a processor to execute an interactive prediction method, the method comprising:

[0147] Obtain feature information of the object to be displayed in the pre-display stage; input the feature information into a pre-trained interaction prediction model to obtain the predicted interaction index corresponding to the display process of the object to be displayed on the display page, wherein the predicted interaction index includes one or more of the first interaction index corresponding to the transaction operation and the second interaction index corresponding to the pre-transaction operation, and the interaction prediction model is trained based on the training method of the interaction prediction model in the page display scenario provided in any of the above embodiments.

[0148] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0149] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0150] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0151] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0152] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements a training method for an interaction prediction model in a page display scenario, or an interaction prediction method, as described in any embodiment of the invention.

[0153] In implementing the computer program product, computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0154] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0155] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for training an interaction prediction model in a page display scenario, characterized in that, include: Obtain a sample dataset, which includes multiple sample data. The sample data includes feature information and interaction tag information. The feature information includes first feature information obtained before the sample object is displayed on the display page and second feature information obtained after it is displayed on the display page. The first feature information and the second feature information in the sample data are input into the teacher network model to obtain the first predictive interaction index; the first feature information in the sample data is input into the student network model to obtain the second predictive interaction index. A target loss function is generated based on the interaction label information, the first predicted interaction index, and the second predicted interaction index in the sample data. With the goal of reducing the target loss function as the training objective, the parameters of the student network model are adjusted until a well-trained student network model is obtained. The well-trained student network model is then identified as the interaction prediction model, which is used to predict the interaction metrics corresponding to the interactive operations set by the object to be displayed during the display process on the display page.

2. The method according to claim 1, characterized in that, The second feature information includes one or more of the display position information of the sample object and the interaction information on the display page; The first feature information includes one or more of the object information of the sample object, user behavior feature information, user non-behavioral feature information, and user-object cross-combination information; wherein, the user non-behavioral feature information is the feature information other than the user behavior feature information in the user-related feature information, and the user-object cross-combination information is the feature information obtained by combining the object information of the sample object with the user non-behavioral feature information.

3. The method according to claim 1, characterized in that, The step of generating a target loss function based on the interaction label information, the first predicted interaction index, and the second predicted interaction index in the sample data includes: A cross-entropy loss function is generated based on the interaction label information in the sample data and the second predicted interaction index. A distillation loss function is generated based on the first and second predictive interaction indices. The target loss function is determined based on the cross-entropy loss function and the distillation loss function.

4. The method according to claim 3, characterized in that, The defined interactive operations include transaction operations and pre-transaction operations; The first predictive interaction indicator and the second predictive interaction indicator respectively include the first interaction indicator corresponding to the transaction operation and the second interaction indicator corresponding to the pre-transaction operation. The interactive tag information in the sample data includes a first tag corresponding to the transaction operation and a second tag corresponding to the pre-transaction operation; The method further includes: Based on the first label and the second label of the sample data, determine the sample data with consistent labels; Obtain the sample data corresponding to the label consistency of the second predicted interaction index, which includes the first interaction index and the second interaction index; generate a consistency loss function based on the first interaction index and the second interaction index included in the second predicted interaction index; The target loss function is determined based on the cross-entropy loss function, the distillation loss function, and the consistency loss function.

5. The method according to claim 2, characterized in that, The student network model includes: a first feature extraction network block, a second feature extraction network block, and a multi-task processing network block, wherein the first feature extraction network block and the second feature extraction network block are respectively connected to the multi-task processing feature extraction network block; The step of inputting the first feature information from the sample data into the student network model to obtain the second predictive interaction index includes: The first feature information is input into the first feature extraction network block for feature extraction to obtain user behavior sequence semantic features; one or more of the object information of the sample object, the user non-behavioral feature information, and the user-object cross-combination information in the first feature information are input into the second feature extraction network block for feature extraction to obtain user-object implicit cross semantic features; the user behavior sequence semantic features and the user-object implicit cross semantic features are input into the multi-task processing network block for index prediction to obtain the second predicted interaction index.

6. The method according to claim 2, characterized in that, The teacher network model includes: a third feature extraction network block, a fourth feature extraction network block, a fifth feature extraction network block, a parameter personalization network block, and a multi-task processing network block, wherein the third feature extraction network block, the fourth feature extraction network block, and the fifth feature extraction network block are respectively connected to the parameter personalization network block, and the parameter personalization network block is connected to the multi-task processing network block; The first feature information and the second feature information in the sample data are input into the teacher network model to obtain the first predictive interaction index, including: The first feature information is input into the third feature extraction network block for feature extraction to obtain user behavior sequence semantic features; one or more of the user non-behavioral feature information, the object information of the sample object, and the user-object cross-combination information in the first feature information are input into the fourth feature extraction network block for feature extraction to obtain user-object implicit cross semantic features; the second feature information is input into the fifth feature extraction network block for feature extraction to obtain user-object explicit cross semantic features. The user behavior sequence semantic features, the user-object implicit cross semantic features, the user-object explicit cross semantic features, and the second feature information are input into the parameter personalization network block to obtain personalized features; the personalized features are input into the multi-task processing network block for index prediction to obtain the first predicted interaction index.

7. An interactive prediction method, characterized in that, include: Obtain the feature information of the object to be displayed in the pre-display stage; The feature information is input into a pre-trained interaction prediction model to obtain the predicted interaction indicators corresponding to the display process of the object to be displayed on the display page. The predicted interaction indicators include one or more of the first interaction indicator corresponding to the transaction operation and the second interaction indicator corresponding to the pre-transaction operation. The interaction prediction model is trained based on the training method of the interaction prediction model in the page display scenario described in any one of claims 1-6.

8. A training device for an interaction prediction model in a page display scenario, characterized in that, include: The sample acquisition module is used to acquire a sample dataset, which includes multiple sample data. The sample data includes feature information and interaction tag information. The feature information includes first feature information acquired before the sample object is displayed on the display page and second feature information acquired after it is displayed on the display page. The data processing module is used to input the first feature information and the second feature information in the sample data into the teacher network model to obtain the first predictive interaction index; and to input the first feature information in the sample data into the student network model to obtain the second predictive interaction index. The loss function determination module is used to generate a target loss function based on the interaction label information, the first predicted interaction index, and the second predicted interaction index in the sample data. The model tuning module is used to tune the parameters of the student network model with the goal of reducing the target loss function, until a well-trained student network model is obtained, and the trained student network model is determined as the interactive prediction model. The interaction prediction model is used to predict the interaction metrics corresponding to the interactive operations set by the object to be displayed during the display process on the display page.

9. An interactive prediction device, characterized in that, include: The feature acquisition module is used to acquire feature information of the object to be displayed in the pre-display stage; The prediction module is used to input the feature information into a pre-trained interaction prediction model to obtain the interaction indicators corresponding to the display process of the object to be displayed on the display page. The interaction indicators include one or more of the first predicted interaction indicator corresponding to the transaction operation and the second predicted interaction indicator corresponding to the pre-transaction operation. The interaction prediction model is trained based on the training method of the interaction prediction model in the page display scenario according to any one of claims 1-6.

10. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the training method of the interaction prediction model in the page display scenario of any one of claims 1-6, or the interaction prediction method of claim 7.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute and implement the training method for the interaction prediction model in the page display scenario according to any one of claims 1-6, or the interaction prediction method according to claim 7.

12. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the training method for the interaction prediction model in the page display scenario according to any one of claims 1-6, or the interaction prediction method according to claim 7.

Images