Data processing method and device, electronic equipment, storage medium and program product

Abstract

The invention discloses a data processing method and device, electronic equipment, a storage medium and a program product, and can be applied to the technical field of artificial intelligence such as machine learning. The method comprises the following steps: respectively predicting a first intervention response and a first contrast response of a to-be-predicted sample object through an intervention response prediction network and a contrast response prediction network so as to determine a first response gain of intervention processing on the to-be-predicted sample object; predicting a second intervention response and a second contrast response of the to-be-predicted sample object through the hybrid response prediction network to determine a second response gain of the intervention processing to the to-be-predicted sample object; combining the first response gain and the second response gain to obtain a trained first gain prediction model and a trained second gain prediction model; and predicting and intervening the prediction response gain of the to-be-predicted object through the trained gain prediction model. According to the application, the accuracy of predicting the response gain can be improved, so that the precision of intervention processing is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a data processing method and device, electronic equipment, storage medium and program product. BACKGROUND

[0002] In the interaction between the business subject and the business user, the intervention processing plays a key role on the business user. Through the intervention processing, the response of the business user to the business can be stimulated, and then the interaction possibility between the business user and the business subject is increased, so as to promote the continuous development of the business. In the prior art, an uplift model is often used to predict the influence of the intervention processing on the business user, so as to predict the response result of the business user to the intervention processing.

[0003] However, the existing uplift model is easily affected by factors such as data distribution, and fluctuates, resulting in low accuracy of predicting the response gain, so that the accuracy of the intervention processing is low. SUMMARY

[0004] The embodiments of the present application provide a data processing method and device, electronic equipment, storage medium and program product, which can improve the accuracy of predicting the response gain, determine a more accurate intervention processing strategy, and improve the accuracy of the intervention processing.

[0005] The embodiment of the present application provides a data processing method, comprising: obtaining a to-be-predicted sample object, a first gain prediction model and a second gain prediction model, the first gain prediction model comprising an intervention response prediction network trained by a first sample object and a control response prediction network trained by a second sample object, the second gain prediction model comprising a mixed response prediction network trained by the first sample object and the second sample object, the first sample object being a sample object subjected to an intervention treatment, and the second sample object being a sample object not subjected to the intervention treatment; predicting a first intervention response and a first control response of the to-be-predicted sample object by the intervention response prediction network and the control response prediction network respectively, so as to determine a first response gain of the intervention treatment on the to-be-predicted sample object from the first intervention response and the first control response; predicting a second intervention response and a second control response of the to-be-predicted sample object by the mixed response prediction network, so as to determine a second response gain of the intervention treatment on the to-be-predicted sample object from the second intervention response and the second control response; jointly training the first gain prediction model and the second gain prediction model by combining the first response gain and the second response gain, to obtain a trained first gain prediction model and a trained second gain prediction model; and predicting a prediction response gain of the intervention treatment on a to-be-predicted object by the trained first gain prediction model or the trained second gain prediction model, the prediction response gain being used to determine a processing strategy of the intervention treatment on the to-be-predicted object.

[0006] The embodiment of the present application further provides a data processing apparatus, comprising: an acquisition unit, configured to acquire a sample object to be predicted, a first gain prediction model and a second gain prediction model, wherein the first gain prediction model comprises an intervention response prediction network trained by a first sample object and a control response prediction network trained by a second sample object, the second gain prediction model comprises a mixed response prediction network trained by the first sample object and the second sample object, the first sample object is a sample object subjected to an intervention treatment, and the second sample object is a sample object not subjected to the intervention treatment; a first prediction unit, configured to predict a first intervention response and a first control response of the sample object to be predicted respectively by using the intervention response prediction network and the control response prediction network, and determine a first response gain of the intervention treatment on the sample object to be predicted according to the first intervention response and the first control response; a second prediction unit, configured to predict a second intervention response and a second control response of the sample object to be predicted by using the mixed response prediction network, and determine a second response gain of the intervention treatment on the sample object to be predicted according to the second intervention response and the second control response; a training unit, configured to jointly train the first gain prediction model and the second gain prediction model according to the first response gain and the second response gain, and obtain a trained first gain prediction model and a trained second gain prediction model; and a third prediction unit, configured to predict a predicted response gain of the intervention treatment on a sample object to be predicted by using the trained first gain prediction model or the trained second gain prediction model, and determine a treatment strategy of the intervention treatment on the sample object to be predicted according to the predicted response gain.

[0007] In some embodiments, the acquisition unit comprises a first acquisition subunit, a second acquisition subunit and a third acquisition subunit, comprising: the first acquisition subunit, configured to acquire a sample object set and a first initial gain prediction model, wherein the first initial gain prediction model comprises an initial intervention response prediction network and an initial control response prediction network, and the sample object set comprises the first sample object and the second sample object; the second acquisition subunit, configured to predict a first initial intervention response of the first sample object and a first initial control response of the second sample object respectively by using the initial intervention response prediction network and the initial control response prediction network; and the third acquisition subunit, configured to train the first initial gain prediction model according to the first initial intervention response and the first initial control response, and obtain a first gain prediction model.

[0008] In some embodiments, the sample object set includes a first object group, the first object group includes the first sample object and the second sample object, and the second obtaining sub-unit is specifically configured to: predict, by the initial intervention response prediction network, a first initial intervention response of the first sample object in the first object group; and predict, by the initial control response prediction network, a first initial control response of the second sample object in the first object group.

[0009] In some embodiments, the sample object set further includes a second object group, and a third obtaining sub-unit is specifically configured to: update the first initial gain prediction model based on the first initial intervention response and the first initial control response to obtain an intermediate gain prediction model, the intermediate gain prediction model including an intermediate intervention response prediction network updated by the initial intervention response prediction network and an intermediate control response prediction network updated by the initial control response prediction network; predict, by the intermediate intervention response prediction network and the intermediate control response prediction network, an intermediate intervention response and an intermediate control response of the sample object to be predicted in the second object group, respectively; determine an intermediate response gain of the intervention processing on the sample object to be predicted in the second object group according to the intermediate intervention response and the intermediate control response; and obtain the first gain prediction model from the intermediate gain prediction model based on the intermediate response gain.

[0010] In some embodiments, obtaining the first gain prediction model from the intermediate gain prediction model based on the intermediate response gain includes: determining a performance parameter of the intermediate gain prediction model based on the intermediate response gain; and training the intermediate gain prediction model to obtain the first gain prediction model by the sample object set when the performance parameter meets a preset performance condition.

[0011] In some embodiments, obtaining the first gain prediction model from the intermediate gain prediction model based on the intermediate response gain further includes: returning to execute the step of predicting, by the initial intervention response prediction network and the initial control response prediction network, the first initial intervention response of the first sample object and the first initial control response of the second sample object, respectively, and subsequent steps until the performance parameter meets the preset performance condition, when the performance parameter does not meet the preset performance condition.

[0012] In some embodiments, the obtaining unit comprises a fourth obtaining subunit, a fifth obtaining subunit and a sixth obtaining subunit, comprising: the fourth obtaining subunit is configured to obtain a sample object set and a second initial gain prediction model, the sample object set comprises the first sample object and the second sample object, and the second initial gain prediction model comprises an initial mixed response prediction network; the fifth obtaining subunit is configured to predict a second initial intervention response of the first sample object and a second initial control response of the second sample object through the initial mixed response prediction network; and the sixth obtaining subunit is configured to train the second initial gain prediction model based on the second initial intervention response and the second initial control response to obtain a second gain prediction model.

[0013] In some embodiments, the training unit comprises a first training subunit and a second training subunit, comprising: the first training subunit is configured to determine a first object set and a second object set from the sample object to be predicted, the first response gain corresponding to the first object set satisfies a gain threshold condition, and the second response gain corresponding to the second object set satisfies the gain threshold condition; and the second training subunit is configured to jointly train the first gain prediction model and the second gain prediction model based on the coincidence degree of the first object set and the second object set to obtain a trained first gain prediction model and a trained second gain prediction model.

[0014] In some embodiments, the second training subunit is specifically configured to: when the coincidence degree does not satisfy a preset coincidence degree condition, update the first gain prediction model and the second gain prediction model to obtain an updated first gain prediction model and an updated second gain prediction model; return to execute the step of obtaining the sample object, the first gain prediction model and the second gain prediction model and the subsequent steps with the updated first gain prediction model and the updated second gain prediction model as the first gain prediction model and the second gain prediction model respectively until the coincidence degree satisfies the preset coincidence degree condition, and take the updated first gain prediction model and the updated second gain prediction model as the trained first gain prediction model and the trained second gain prediction model respectively.

[0015] In some embodiments, the data processing apparatus further comprises a message reminding unit, and the message reminding unit is configured to perform the message reminding on the object to be predicted when the predicted response gain satisfies a specified gain type.

[0016] In some embodiments, the data processing apparatus further comprises a processing unit configured to: perform data preprocessing on the object information of the sample object to be predicted to obtain processed information; and perform vectorization processing on the processed information to obtain the object feature of the sample object to be predicted, the object feature being used to predict the intervention response and the control response of the sample object to be predicted.

[0017] An electronic device is also provided, which includes a processor and a memory. The memory stores a plurality of instructions. The processor loads the instructions from the memory to perform the steps of any of the data processing methods provided in the embodiments.

[0018] A computer readable storage medium is also provided, which stores a plurality of instructions. The instructions are adapted to be loaded by a processor to perform the steps of any of the data processing methods provided in the embodiments.

[0019] A computer program product is also provided, which includes a computer program or instructions. The computer program or instructions, when executed by a processor, implement the steps of any of the data processing methods provided in the embodiments.

[0020] The embodiment of the present application can obtain a to-be-predicted sample object, a first gain prediction model and a second gain prediction model, the first gain prediction model comprising an intervention response prediction network trained by a first sample object and a control response prediction network trained by a second sample object, the second gain prediction model comprising a mixed response prediction network trained by the first sample object and the second sample object, the first sample object being a sample object subjected to an intervention treatment, and the second sample object being a sample object not subjected to the intervention treatment; the first intervention response and the first control response of the to-be-predicted sample object are predicted by the intervention response prediction network and the control response prediction network respectively, so as to determine a first response gain of the intervention treatment on the to-be-predicted sample object from the first intervention response and the first control response; the second intervention response and the second control response of the to-be-predicted sample object are predicted by the mixed response prediction network, so as to determine a second response gain of the intervention treatment on the to-be-predicted sample object from the second intervention response and the second control response; the first gain prediction model and the second gain prediction model are jointly trained in combination with the first response gain and the second response gain, to obtain a trained first gain prediction model and a trained second gain prediction model; and the prediction response gain of the intervention treatment on a to-be-predicted object is predicted by the trained first gain prediction model or the trained second gain prediction model, and the prediction response gain is used to determine a processing strategy of the intervention treatment on the to-be-predicted object.

[0021] In the present application, different first gain prediction models and second gain prediction models are used to predict the response gain of the intervention treatment on the sample object, so as to combine the prediction results of the two gain prediction models for joint training, and guide the training process of the two gain prediction models. The joint training can utilize the correlation between the prediction tasks of the two gain prediction models to improve the volatility of the output of a single gain prediction model, so that the trained gain model can output stable and consistent prediction results, improve the accuracy of the prediction response gain, determine a more accurate intervention treatment strategy based on the predicted response gain, and improve the accuracy of the intervention treatment. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0023] Figure 1a is a scene schematic diagram of the data processing method provided by the embodiment of the present application;

[0024] Figure 1b is a flowchart of a data processing method provided by an embodiment of the present application;

[0025] Figure 1c is a schematic diagram of obtaining a sample object provided by an embodiment of the present application;

[0026] Figure 1d is a schematic diagram of a first gain prediction model provided by an embodiment of the present application;

[0027] Figure 1e is a schematic diagram of a second gain prediction model provided by an embodiment of the present application;

[0028] Figure 1f is a schematic diagram of an independent training flow of the first gain prediction model provided by an embodiment of the present application;

[0029] Figure 2a is a flowchart of a data processing method provided by another embodiment of the present application;

[0030] Figure 2b is a flowchart of another data processing method provided by an embodiment of the present application;

[0031] Figure 2c is a schematic diagram of an independent training flow of the first gain prediction model provided by another embodiment of the present application;

[0032] Figure 2d is a schematic diagram of an independent training flow of the second initial gain prediction model provided by an embodiment of the present application;

[0033] Figure 2e is a schematic diagram of a joint training flow of the first gain prediction model and the second gain prediction model provided by an embodiment of the present application;

[0034] Figure 3 is a structural schematic diagram of a data processing apparatus provided by an embodiment of the present application;

[0035] Figure 4 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0037] In the following description, the terms "first", "second", "third", "fourth" are merely used to distinguish similar objects, and do not represent a specific order or sequence of the objects. Understandably, the "first", "second", "third", "fourth" can be interchanged with a specific order or sequence as allowed, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein.

[0038] The embodiments of the present application provide a data processing method and device, an electronic device, a storage medium and a program product.

[0039] The data processing device can be integrated in an electronic device, which can be a terminal, a server, etc. The terminal can be a mobile phone, a tablet computer, a smart Bluetooth device, a notebook computer, or a personal computer (PC), etc. The server can be a single server or a server cluster composed of multiple servers.

[0040] In some embodiments, the data processing device can also be integrated in multiple electronic devices, for example, the data processing device can be integrated in multiple servers, and the multiple servers can implement the data processing method of the present application.

[0041] In some embodiments, the server can also be implemented in the form of a terminal.

[0042] For example, refer to Figure 1aThe data processing method is applied to a message reminding system, and the message reminding system comprises a terminal and a server. The server is configured to obtain a to-be-predicted sample object, a first gain prediction model, and a second gain prediction model. The first gain prediction model comprises an intervention response prediction network trained by a first sample object and a control response prediction network trained by a second sample object. The second gain prediction model comprises a mixed response prediction network trained by the first sample object and the second sample object. The first sample object is a sample object that has been intervened, and the second sample object is a sample object that has not been intervened. The first intervention response and the first control response of the to-be-predicted sample object are predicted by the intervention response prediction network and the control response prediction network, respectively, so as to determine a first response gain of the intervention on the to-be-predicted sample object according to the first intervention response and the first control response. The second intervention response and the second control response of the to-be-predicted sample object are predicted by the mixed response prediction network, so as to determine a second response gain of the intervention on the to-be-predicted sample object according to the second intervention response and the second control response. The first gain prediction model and the second gain prediction model are jointly trained according to the first response gain and the second response gain, so as to obtain a trained first gain prediction model and a trained second gain prediction model. The prediction response gain of the intervention on the to-be-predicted object is predicted by the trained first gain prediction model or the trained second gain prediction model, and the prediction response gain is used to determine a processing strategy of the intervention on the to-be-predicted object. The terminal is configured to obtain the to-be-predicted object and receive the prediction response result sent by the server.

[0043] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined target, and can be implemented entirely or partially by using software, hardware (such as a processing circuit or a memory) or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an overall module or unit that includes the functions of the module or unit.

[0044] The following are described in detail. It should be noted that the order of the following embodiments is not limited as the preferred order of the embodiments. It can be understood that in the detailed description of the present application, the data related to the user such as the sample object, the to-be-predicted object, the object information, the object feature, the user, and the customer needs to be obtained with the permission or consent of the user, and the collection, use and processing of the related data need to comply with the relevant laws, regulations and standards of the country and region.

[0045] Artificial Intelligence (AI) is a technology that uses digital computers to simulate human perception of the environment, acquire knowledge, and use knowledge, which can enable machines to have functions similar to human perception, reasoning, and decision-making. Artificial intelligence basic technologies generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, automatic driving, intelligent transportation, etc.

[0046] Machine Learning (ML) is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is a subject that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

[0047] With the research and progress of artificial intelligence technology, artificial intelligence technology is researched and applied in many fields, such as common smart home, smart wearable device, virtual assistant, smart speaker, smart marketing, unmanned driving, autonomous driving, unmanned aerial vehicle, robot, intelligent medical treatment, intelligent customer service, Internet of Vehicles, intelligent transportation, etc. It is believed that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0048] In this embodiment, a data processing method related to artificial intelligence is provided, as shown in Figure 1b The specific process of the data processing method can be as follows:

[0049] 110, obtaining a to-be-predicted sample object, a first gain prediction model, and a second gain prediction model.

[0050] In the embodiment of the application, the first gain prediction model includes an intervention response prediction network trained by a first sample object and a control response prediction network trained by a second sample object, and the second gain prediction model includes a mixed response prediction network trained by the first sample object and the second sample object.

[0051] The sample object refers to an object used for training the gain prediction model, which can be an entity or an individual. In the embodiments of the present application, the object can be a user interacting with the business subject (i.e., a business user). For example, the business subject can be a virtual platform such as an application program, and the object can be a user of the application program, or the business subject can be an entity such as an enterprise, and the object can be a customer of the enterprise.

[0052] The first sample object is a sample object that has been intervened, and the second sample object is a sample object that has not been intervened. In the embodiments of the present application, the first sample object and the second sample object are respectively taken as an experimental group and a control group, so as to train the gain prediction model by comparing the two groups of samples, so that the trained gain prediction model (i.e., the first gain prediction model and the second gain prediction model) can be used to evaluate the effect and influence of the intervention on the sample object. The to-be-predicted sample object refers to a sample object used for jointly training the first gain prediction model and the second gain prediction model, and the to-be-predicted sample object can include one or more of the first sample object, the second sample object, or other sample objects. The other sample objects can be sample objects that have been intervened or have not been intervened, in addition to the first sample object and the second sample object.

[0053] The intervention refers to a means of giving intervention to the object. It can be understood that the intervention aims to stimulate the response of the business user to the business, and thus promote the interaction between the business user and the business subject. Therefore, the intervention can include, but is not limited to, one or more of important notifications, personalized recommendations, business reminders, and activity invitations. In the embodiments of the present application, the intervention can be performed in one or more ways such as application messages, short messages, telephone, email, and mail. In different application scenarios, the specific implementation of the intervention can be different. For example, for users of a virtual platform, the intervention can be performed by sending a pop-up reminder in an application program, pushing a notification, or sending an email. For customers of an entity, the intervention can be performed by sending a short message, making a phone call, or mailing a letter.

[0054] In this context, the gain prediction model refers to a model used to predict response gain. Response gain refers to the gain of the intervention response relative to the control response; that is, response gain can be used to evaluate the predictive effect of an intervention on a sample object (i.e., gain). For example, the gain prediction model can be various machine learning models or neural network models. Response gain can represent the predictive effect of an intervention on an object (i.e., gain). Specifically, response gain can be used to evaluate the impact of an intervention on an object to characterize the gain of the object's response. The higher the response gain, the greater the impact of the intervention on the object, that is, the greater the improvement in the object's response. Response refers to the likelihood of an object making a positive response; the higher the response, the higher the likelihood of interaction between the object and the business entity.

[0055] In this context, the response prediction network refers to the network structure used in the gain prediction model to predict the response. For example, the response prediction network can include, but is not limited to, tree classifiers such as decision trees, random forests, and gradient boosting trees, or logistic regression, or one or more combinations of other neural network structures that can be used to predict responses. Thus, the response prediction network can use information about the object to model and predict the object's response to an intervention. It should be noted that the intervention response prediction network can be used to predict the response of a sample object when an intervention is applied, the control response prediction network can be used to predict the response of a sample object when no intervention is applied, and the hybrid response prediction network can be used to predict both the response of a sample object when an intervention is applied and the response of a sample object when no intervention is applied. The intervention response prediction network, control response prediction network, and hybrid response prediction network can have the same network structure or different network structures.

[0056] For example, such as Figure 1c The diagram illustrating sample object acquisition shows that, with the individual consent or permission of the business users, information on multiple business users (i.e., sample objects) can be obtained from the business system and from business users interacting with the business entity. Sample objects may include business users who have undergone intervention (i.e., first sample objects) and multiple business users who have not undergone intervention (i.e., second sample objects), and these one or more sample objects can be used as the sample objects to be predicted. Furthermore, information such as... Figure 1d The first gain prediction model shown and as follows Figure 1e The second gain prediction model shown is different from the first gain prediction model, which includes two independent response prediction networks (i.e., the intervention response prediction model and the control response prediction model), while the second gain prediction model includes only one response prediction network (i.e., the dual response prediction model).

[0057] In some embodiments, the object of the sample object can be pre-processed and vectorized to improve the subsequent model prediction accuracy and effect. Specifically, the data processing method further comprises:

[0058] performing data preprocessing on the object information of the sample object to be predicted to obtain processed information;

[0059] performing vectorization processing on the processed information to obtain object features of the sample object to be predicted, the object features being used to predict the intervention response and the control response of the sample object to be predicted.

[0060] The object information refers to information related to the object. In the embodiments of the present application, the object information can be feature information of the object, which can include but is not limited to one or more of attributes, states, historical behaviors, interests, social relationships, etc.

[0061] The data preprocessing refers to a processing process of identifying, correcting or removing data with errors, incompleteness, inconsistency or abnormal values, etc. before model processing. Data preprocessing can ensure the quality and reliability of the data to improve the subsequent model prediction accuracy and effect. In the embodiments of the present application, data preprocessing can include but is not limited to one or more combinations of missing value filling, abnormal value replacement and other processing methods.

[0062] The vectorization processing refers to a processing process of converting data into a vector form. The object information of the sample object is represented as a vector through vectorization processing, so as to facilitate subsequent model processing and understanding, thereby improving the subsequent model prediction accuracy and effect. In the embodiments of the present application, the vectorization processing method can include but is not limited to one or more combinations of encoder encoding, embedding processing or other encoding methods.

[0063] For example, the feature information of a plurality of sample objects can be obtained from business users interacting with business subjects, and the obtained feature information can be pre-processed by missing value filling and abnormal value replacement. The processed information of each sample object obtained by preprocessing is encoded to obtain the object features of each sample object. When the sample object is obtained, the object features of the sample object are directly obtained for subsequent model processing.

[0064] In some embodiments, the response prediction network of the first gain prediction model can be used to predict the responses of the sample objects with intervention processing and the sample objects without intervention processing, respectively, to train the first gain prediction model based on the predicted responses. During the training process, the response prediction network of the model can learn the features of the two types of sample objects in advance independently through the predicted responses of different sample objects, so as to accurately predict the response of the object based on these features. Specifically, the first gain prediction model is obtained by the following steps:

[0065] obtain a sample object set and a first initial gain prediction model, the first initial gain prediction model comprising an initial intervention response prediction network and an initial control response prediction network, the sample object set comprising a first sample object and a second sample object;

[0066] predict a first initial intervention response of the first sample object and a first initial control response of the second sample object by the initial intervention response prediction network and the initial control response prediction network, respectively;

[0067] train the first initial gain prediction model based on the first initial intervention response and the first initial control response to obtain a first gain prediction model.

[0068] The first initial gain prediction model, the initial intervention response prediction network, and the initial control response prediction network refer to the initial state of the gain prediction model, the intervention response prediction network, and the control response prediction network to be trained, respectively.

[0069] The intervention response refers to the response predicted by the sample object to the intervention processing of the sample object. The control response refers to the response predicted by the sample object to the non-intervention processing of the sample object. The first initial intervention response and the first initial control response refer to the responses obtained by processing the initial model or network.

[0070] For example, the object features of the first sample object can be input into the initial intervention response prediction network for response prediction to obtain the first initial intervention response of the sample object. The object features of the second sample object can be input into the initial control response prediction network for response prediction to obtain the first initial control response of the sample object. The difference (such as loss) between the first initial intervention response and the first initial control response and the predicted target (i.e., sample label) of the sample object can be calculated. The model parameters of the first initial gain prediction model are adjusted by minimizing the loss until the difference between the first initial response gain and the predicted target is less than a preset threshold. The gain prediction model obtained by the last adjustment is the first gain prediction model. The preset threshold can be set according to the actual situation.

[0071] In some embodiments, the sample objects in the sample object set can be used as the sample objects to be predicted to increase the consistency of the process of training the gain prediction model individually and jointly, and to increase the reliability and stability of the gain prediction model after training.

[0072] In some embodiments, the response of the sample object with intervention and the sample object without intervention can be predicted by the response prediction network of the second gain prediction model, so as to train the first gain prediction model based on the predicted response. During the training process, the response prediction network of the model can learn the characteristics of the two types of sample objects in advance and independently, so as to accurately predict the response of the object based on the characteristics. Specifically, the second gain prediction model is obtained by the following steps:

[0073] obtaining a sample object set and a second initial gain prediction model, the sample object set comprising a first sample object and a second sample object, and the second initial gain prediction model comprising an initial mixed response prediction network;

[0074] predicting, by the initial mixed response prediction network, a second initial intervention response of the first sample object and a second initial control response of the second sample object;

[0075] training the second initial gain prediction model based on the second initial intervention response and the second initial control response, to obtain the second gain prediction model.

[0076] wherein the second initial gain prediction model and the initial mixed response prediction network refer to the initial state of the gain prediction model and the mixed response prediction network to be trained, and the second initial intervention response and the second initial control response refer to the responses obtained by the initial model or network.

[0077] For example, the object characteristics of the first sample object and the object characteristics of the second sample object can be input into the initial mixed response prediction network for response prediction, to obtain the second initial intervention response of the first sample object and the second initial control response of the second sample object. The difference (such as loss) between the second initial intervention response and the second initial control response and the predicted target (i.e. sample label) of the sample object can be calculated, the model parameters of the second initial gain prediction model are adjusted by minimizing the loss, until the difference between the second initial response gain and the predicted target is less than a preset threshold, and the gain prediction model obtained by the last adjustment is the second gain prediction model. The preset threshold can be set according to the actual situation.

[0078] In some embodiments, the sample objects can be grouped for training the gain prediction model, so as to improve the performance and prediction accuracy of the model. Specifically, the sample object set comprises a first object group, the first object group comprising a first sample object and a second sample object, and the first initial intervention response of the first sample object and the first initial control response of the second sample object are predicted by the initial intervention response prediction network and the initial control response prediction network, comprising:

[0079] Predict the first initial intervention response of the first sample object in the first object group using the initial intervention response prediction network;

[0080] The initial control response prediction network is used to predict the first initial control response of the second sample object in the first object group.

[0081] For example, such as Figure 1f The independent training process of the first gain prediction model shown can be as follows: the object features of the first sample object in the sample object set such as the first object group in the subset D1 of the dataset D can be input into the initial intervention response prediction network to predict the response and obtain the response of the first sample object (i.e., the first initial intervention response); and the object features of the second sample object in the subset D1 can be input into the initial control response prediction network to predict the response and obtain the response of the second sample object (i.e., the first initial control response).

[0082] In some implementations, sample objects can be grouped for training the first gain prediction model. One group of objects trains the model based on the predicted response, and another group trains the model based on the predicted response gain. This allows for multi-dimensional evaluation of the model's performance across different datasets, improving both model performance and prediction accuracy. Specifically, the sample object set also includes a second group of objects. The first initial gain prediction model is trained based on the first initial response gain to obtain the first gain prediction model, which includes:

[0083] Based on the first initial intervention response and the first initial control response, the first initial gain prediction model is updated to obtain the intermediate gain prediction model. The intermediate gain prediction model includes the intermediate intervention response prediction network updated by the initial intervention response prediction network and the intermediate control response prediction network updated by the initial control response prediction network.

[0084] The intermediate intervention response prediction network and the intermediate control response prediction network are used to predict the intermediate intervention response and the intermediate control response of the sample subjects in the second subject group, respectively.

[0085] Based on the intermediate intervention response and the intermediate control response, determine the intermediate response gain of the intervention treatment on the sample objects to be predicted in the second object group;

[0086] Based on the intermediate response gain, the first gain prediction model is obtained from the intermediate gain prediction model.

[0087] The intermediate gain prediction model, the intermediate intervention response prediction network, and the intermediate control response prediction network refer to the gain prediction model, the intervention response prediction network, and the control response prediction network used in the intermediate step or process, respectively. The intermediate intervention response, the intermediate control response, and the intermediate response gain refer to the response and the response gain processed by the model or network used in the intermediate step. It should be noted that in the embodiments of the present application, the second object group can include one or more of the first sample object and the second sample object.

[0088] For example, as shown in the independent training process of the first gain prediction model, Figure 1f The obtained sample object set, such as the data set D, can be randomly or according to a preset rule divided into a sub-data set D1 (i.e., the first object group) and a sub-data set D2 (i.e., the second object group). The sub-data set D1 and the sub-data set D2 each contain a plurality of first sample objects and a plurality of second sample objects. The difference (such as loss) between the first initial intervention response and the first initial control response predicted by the first initial gain prediction model for the sub-data set D1 and the prediction target (i.e., the sample label) of the sample object can be calculated. By minimizing the loss, the model parameters of the first initial gain prediction model are adjusted to obtain the intermediate gain prediction model. The object features of each sample object in the sub-data set D2 can be input into the intermediate intervention response prediction network for response prediction to obtain the response (i.e., the intermediate intervention response) of each sample object, and the object features of each sample object in the sub-data set D2 can be input into the intermediate control response prediction network for response prediction to obtain the response (i.e., the intermediate control response) of each sample object. For any sample object in the sub-data set D2, the intermediate response gain of the sample object can be calculated by comparing the intermediate intervention response and the intermediate control response of the sample object using difference calculation, ratio calculation, and percentage difference analysis methods. The performance of the intermediate gain prediction model can be evaluated based on the intermediate response gain, so that the intermediate gain prediction model with expected performance is used as the first gain prediction model.

[0089] In some embodiments, during the training process of the first gain prediction model, the sample objects can be grouped for training and validating the gain prediction model to evaluate the performance of the model while training the gain prediction model, thereby improving the performance and prediction accuracy of the gain model. Meanwhile, when the performance of the gain prediction model reaches the expectation (i.e., meets the preset performance condition), all sample objects can be used to train the gain prediction model to further train the model using more and more comprehensive sample data, so that the gain prediction model learns the overall characteristics and distribution of more data and improves the generalization ability of the model, thereby improving the performance and prediction accuracy of the model. Specifically, as shown in the response prediction process, Figure 1f The first gain prediction model obtained from the intermediate gain prediction model based on the intermediate response gain includes:

[0090] Based on the intermediate response gain, determine the performance parameters of the intermediate gain prediction model;

[0091] When the performance parameters meet the preset performance conditions, the intermediate gain prediction model is trained using the sample object set to obtain the first gain prediction model.

[0092] Performance parameters refer to parameters used to evaluate the performance of the gain prediction model. Performance parameters can be used to measure the model's prediction accuracy, generalization ability, and other related performance. For example, in the embodiments of this application, performance parameters may include, but are not limited to, one or more combinations of accuracy, precision, recall, F1 score, area under the curve (AUUC), mean squared error (MSE), mean absolute error (MAE), and coefficient of variation (CV).

[0093] The preset performance conditions refer to the conditions set in advance to determine whether the model's predictive performance meets expectations. These preset performance conditions can be determined based on specific application scenarios and requirements. For example, they may include one or more of the following: minimum requirements for performance parameters, threshold settings, or other constraints. For instance, prediction performance conditions could be achieving a certain accuracy, recall, or AUC value, or a mean squared error or mean absolute error within a specific error range. When the model's performance parameters reach or exceed the preset performance conditions, the prediction performance conditions are satisfied.

[0094] For example, such as Figure 1f The independent training process of the first gain prediction model shown can be achieved by determining the performance parameters of the intermediate gain prediction model, such as mean, variance, area under the curve (AUUC), and coefficient of variation (CV), based on the intermediate response gains of the sample objects in the subset D2. It is then determined whether these performance parameters meet the prediction performance conditions. If they do, the intermediate intervention response prediction network and the initial control response prediction network can be used to predict the target intervention response of the first sample object in the sample object set and the target control response of the second sample object in the sample object set, respectively. Based on the target intervention response and the target control response, the intermediate gain prediction model is trained to obtain the first gain prediction model.

[0095] Specifically, the object features of each first sample object in the data set D are input into the intermediate intervention response prediction network for response prediction to obtain the response of each first sample object (i.e., the target intervention response), and the object features of each second sample object in the data set D are input into the intermediate control response prediction network for response prediction to obtain the response of the second sample object (i.e., the target control response). The difference (such as loss) between the target intervention response and the target control response and the predicted target (i.e., the sample label) of the sample object can be calculated, and the model parameters of the intermediate gain prediction model are adjusted by minimizing the loss until the difference between the target response gain and the predicted target is less than a preset threshold. The gain prediction model obtained by the last adjustment is the first gain prediction model. The preset threshold can be set according to actual conditions.

[0096] In some embodiments, when the model performance of the first gain prediction model does not meet the expectation (i.e., does not satisfy the preset performance condition), the first gain prediction model can be repeatedly trained again until the expectation is reached to improve the performance and prediction accuracy of the model. Specifically, based on the intermediate response gain, the first gain prediction model is obtained from the intermediate gain prediction model, further comprising:

[0097] When the performance parameter does not satisfy the preset performance condition, the intermediate gain prediction model is taken as the first initial gain prediction model, and the steps of predicting the first initial intervention response of the first sample object and the first initial control response of the second sample object by the initial intervention response prediction network and the initial control response prediction network, respectively, and the subsequent steps are executed until the performance parameter satisfies the preset performance condition.

[0098] For example, when the performance parameter of the intermediate gain prediction model does not satisfy the prediction performance condition, the intermediate gain prediction model can be taken as the first initial gain prediction model, and the steps of updating the model parameters of the first initial gain prediction model based on the first initial response gain to obtain the intermediate gain prediction model, and determining the performance parameter of the intermediate gain prediction model based on the intermediate response gain are repeatedly executed until the performance parameter satisfies the prediction performance condition.

[0099] In some embodiments, the difference between the mean value of the responses of the first sample objects and the mean value of the responses of the second sample objects can be calculated to obtain the average response gain of the plurality of sample objects, so as to comprehensively consider the overall influence degree of the intervention processing on the response gain of the sample objects, and more comprehensively and accurately evaluate the prediction ability of the first gain prediction model. Specifically, the first initial gain prediction model is trained based on the first initial intervention response and the first initial control response to obtain the first gain prediction model, comprising:

[0100] The mean of the first initial intervention response was calculated to obtain the first mean, and the mean of the first initial control response was calculated to obtain the second mean;

[0101] The difference between the first mean and the second mean is calculated to obtain the first initial response gain of the first initial intervention response relative to the first initial control response;

[0102] Based on the first initial response gain, a first initial gain prediction model is trained to obtain the first gain prediction model.

[0103] For example, an initial intervention response prediction network can predict the first initial intervention response as {f1, f2, ..., fm} for all first sample objects (if there are m objects) in the sample, where f represents the first initial intervention response for any first sample object. The average of these m first responses can be calculated to obtain the first mean. Similarly, an initial control response prediction network can predict the first initial control response as {f1, f2, ..., fm} for all second sample objects in the sample. ’ 1,f ’ 2,…,f ’ n}, where f ’ This represents the second response of any second sample object. The average of the n first initial control responses can be calculated to obtain the second mean. Subtracting the second mean from the first mean yields the first initial response gain, which represents the average response gain of multiple sample objects predicted by the first initial gain prediction model. The model parameters of the first initial gain prediction model can be adjusted based on the difference between the first initial response gain and the prediction target (i.e., the sample label) of the sample object until the difference between the first initial response gain and the prediction target is less than a preset threshold. The gain prediction model obtained after this final adjustment is the first gain prediction model. The preset threshold can be set according to actual conditions.

[0104] In some implementations, the difference between the mean response of the first sample object and the mean response of the second sample object can be calculated to obtain the average response gain representing multiple sample objects. This allows for a comprehensive consideration of the overall impact of the intervention on the response gain of the sample objects, and a more comprehensive and accurate evaluation of the predictive ability of the second gain prediction model. Specifically, based on the second initial intervention response and the second initial control response, a second initial gain prediction model is trained to obtain the second gain prediction model, including:

[0105] The mean of the second initial intervention response was calculated to obtain the third mean, and the mean of the second initial control response was calculated to obtain the fourth mean;

[0106] a difference between the third mean value and the fourth mean value is calculated to obtain a second initial response gain of the second initial intervention response relative to the second initial control response;

[0107] Based on the second initial response gain, a second initial gain prediction model is trained to obtain a second gain prediction model.

[0108] For example, the second initial intervention response of all first sample objects (such as m) in the sample object can be predicted by the hybrid response prediction network as {F1, F2, …, Fm}, where F represents the second initial intervention response of any first sample object. The average value of the m first responses can be calculated to obtain a third mean value. Similarly, the second initial control response of all second sample objects (such as n) in the sample object can be predicted by the hybrid response prediction network as {F ’ 1, F ’ 2, …, F ’ n}, where F ’ represents the second initial control response of any second sample object. The average value of the n second initial control responses can be calculated to obtain a fourth mean value. The third mean value can be subtracted from the fourth mean value to obtain a second initial response gain, which represents the average response gain of the plurality of sample objects predicted by the second initial gain prediction model. The model parameters of the second initial gain prediction model can be adjusted according to the difference between the second initial response gain and the predicted target (i.e., the sample label) of the sample object, until the difference between the second initial response gain and the predicted target is less than a preset threshold. The gain prediction model obtained by the last adjustment is the second gain prediction model. The preset threshold can be set according to actual conditions.

[0109] 120. The first response gain of the to-be-predicted sample object by the intervention is determined by the first intervention response and the first control response predicted by the intervention response prediction network and the control response prediction network.

[0110] The first intervention response is the intervention response of the sample object predicted by the intervention response prediction network, the second intervention response is the control response of the sample object predicted by the control response prediction network, and the first response gain is the gain of the first intervention response relative to the first control response.

[0111] For example, as Figure 1dThe first gain prediction model shown can input the object features of the sample object into the intervention response prediction network to predict the response and obtain the first intervention response of the sample object. Similarly, it can input the object features of the sample object into the control response prediction network to predict the response and obtain the first control response of the sample object. For any sample object, the first response gain of the sample object can be calculated using comparative analysis methods such as difference calculation, ratio calculation, and percentage difference, based on both the first intervention response and the first control response, thus obtaining the first response gain for each sample object.

[0112] 130. Using a hybrid response prediction network, predict the second intervention response and the second control response of the sample to be predicted, so as to determine the second response gain of the intervention treatment on the sample to be predicted from the second intervention response and the second control response.

[0113] Wherein, the second intervention response is the intervention response of the sample object predicted by the hybrid response prediction network, the second control response is the control response of the sample object predicted by the control response prediction network, and the second response gain is the gain of the second intervention response relative to the second control response.

[0114] For example, such as Figure 1e The second gain prediction model shown can input the object characteristics of the sample object into a hybrid response prediction network to predict the response and obtain the second intervention response and the second control response of the sample object. For any sample object, the second response gain of the sample object can be calculated using comparative analysis methods such as difference calculation, ratio calculation, and percentage difference, thereby obtaining the second response gain of each sample object.

[0115] In some implementations, the difference between the intervention response and the control response can be used as the response gain. For example, the difference between the first intervention response and the first control response of each subject to be predicted can be calculated as the first response gain of each subject to be predicted. Similarly, the difference between the second intervention response and the second control response of each subject to be predicted can be calculated as the second response gain of each subject to be predicted.

[0116] In some embodiments, the object features of the plurality of sample objects can be input into the first gain prediction model and the second gain prediction model in a sequence for processing. In order to enable the response prediction network (especially the hybrid response prediction network) in the gain prediction model to distinguish between the intervention response or the control response predicted for the sample object, an intervention label can be added to the object features of the sample object, so that the gain prediction model distinguishes between the object features of the sample object for predicting the intervention response or the control response according to the label. The intervention label refers to a label for marking the object features of the sample object for predicting the intervention response, or a label for marking the object features of the sample object for predicting the control response. The intervention label can be a specific symbol, number or other label form, which can be determined according to the application scenario or actual demand.

[0117] For example, when the object features of the sample object are input into the hybrid response prediction network for response prediction, a first label value (e.g., 1) of the intervention label can be added to the sequence of the object features of the plurality of sample objects to obtain a first sequence, and the first sequence is input into the hybrid response prediction network for response prediction to obtain a second intervention response of the sample object. A second label value (e.g., 0) of the intervention label can be added to the sequence of the object features of the plurality of sample objects to obtain a second sequence, and the second sequence is input into the hybrid response prediction network for response prediction to obtain a second control response of the sample object. In this way, the intervention response and the control response of the sample object can be predicted simultaneously by using one hybrid response prediction network.

[0118] 140、The first gain prediction model and the second gain prediction model are jointly trained in combination with the first response gain and the second response gain to obtain a trained first gain prediction model and a trained second gain prediction model.

[0119] The joint training refers to a method of training the first gain prediction model and the second gain prediction model simultaneously using the first response gain and the second response gain.

[0120] For example, the first response gain and the second response gain can be used to jointly evaluate the prediction effect of the first gain prediction model and the second gain prediction model to guide the training process of the two gain prediction models. In the embodiments of the present application, the two gain prediction models are jointly trained by combining the prediction results of the two gain prediction models, so as to utilize the correlation between the prediction tasks of the two gain prediction models to improve the volatility of the output of a single gain prediction model, so that the gain model can output stable and consistent prediction results to improve the accuracy of the message prompt.

[0121] In some embodiments, the coincidence of the prediction results of the two gain prediction models can be evaluated in combination, and the higher the coincidence is, the better the stability of the prediction results of the models is, that is, the higher the accuracy of the prediction results is. In this way, by evaluating the coincidence of the prediction results of the two gain prediction models, the volatility of the output of a single gain prediction model can be improved, so that the gain model can output stable and consistent prediction results, so as to improve the prediction accuracy and reliability of the trained model. Specifically, the first gain prediction model and the second gain prediction model are trained in combination based on the first response gain and the second response gain, to obtain the trained first gain prediction model and the trained second gain prediction model, including:

[0122] determining a first object set and a second object set from the sample objects to be predicted, the first response gain corresponding to the first object set satisfying the gain threshold condition, and the second response gain corresponding to the second object set satisfying the gain threshold condition;

[0123] training the first gain prediction model and the second gain prediction model in combination based on the coincidence of the first object set and the second object set, to obtain the trained first gain prediction model and the trained second gain prediction model.

[0124] Wherein, the object set refers to a set of sample objects in the plurality of sample objects that satisfy the gain threshold condition. The coincidence refers to the similarity of the two object sets, which can measure how many common sample objects are between the two object sets. For example, the coincidence can be in the form of a numerical value, a percentage, etc. The coincidence can be represented as the number of common sample objects between the two object sets, the proportion of common sample objects between the two object sets in the total amount of sample objects, etc.

[0125] Wherein, the gain threshold condition refers to a condition for screening sample objects based on a preset gain threshold value. The preset gain threshold value can be set according to actual needs or application scenarios. The preset gain threshold value can be in the form of a numerical value or a percentage, etc. For example, the gain threshold condition can be to add sample objects with a first response gain greater than a preset gain threshold value to the first object set, and to add sample objects with a second response gain greater than a preset gain threshold value to the second object set. Or, for example, the gain threshold condition can be to add the first preset number (such as the top 10%) of sample objects in the plurality of sample objects sorted according to the first response gain to the first object set, and to add the first preset number (such as the top 10%) of sample objects in the plurality of sample objects sorted according to the second response gain to the second object set.

[0126] For example, the sample objects in the data set D can be sorted according to the first response gain and the second response gain, respectively, and the top 10% of the sample objects in the sorted results can be taken as the first object set and the second object set, respectively. Then, the proportion of the sample objects in the data set D that are included in the first object set and the second object set (i.e., the coincidence degree) is calculated. If the coincidence degree meets the expectation (e.g., is greater than a preset proportion), the first gain prediction model and the second gain prediction model are taken as the trained first gain prediction model and the trained second gain prediction model, respectively.

[0127] In some embodiments, the gain threshold condition can be a condition for screening sample objects based on a plurality of preset gain threshold values. In this way, a plurality of first object sets and a plurality of second object sets corresponding to different preset gain threshold values can be determined according to the gain threshold condition, so as to more comprehensively evaluate the accuracy of the prediction results of the gain prediction models under different gain threshold values, thereby improving the prediction accuracy and reliability of the models. For example, the preset gain threshold values can include 10 threshold values in the range of 10%-20%. In this way, 10 first object sets and 10 second object sets corresponding to different threshold values can be obtained. For any threshold value (e.g., 15%), the proportion of the sample objects in the data set D that are included in the first object set and the second object set (i.e., the coincidence degree) can be calculated. Similarly, 10 coincidence degrees corresponding to different threshold values can be calculated. When all the coincidence degrees are greater than a preset proportion, it is considered that the coincidence degree meets the expectation.

[0128] In some embodiments, when the coincidence degree of the prediction results of the two gain prediction models does not meet the expectation, the model parameters of the gain prediction models can be iteratively updated, so that the coincidence degree of the prediction results of the trained models is high, the prediction error of the models is reduced, and the prediction accuracy and reliability of the trained models are improved. Specifically, based on the coincidence degree of the first object set and the second object set, the first gain prediction model and the second gain prediction model are jointly trained to obtain a trained first gain prediction model and a trained second gain prediction model, including:

[0129] When the coincidence degree does not meet the preset coincidence degree condition, the first gain prediction model and the second gain prediction model are updated to obtain an updated first gain prediction model and an updated second gain prediction model.

[0130] The updated first gain prediction model and the updated second gain prediction model are returned to the step of obtaining the sample objects, the first gain prediction model and the second gain prediction model, and the subsequent steps until the coincidence degree meets the preset coincidence degree condition, and the updated first gain prediction model and the updated second gain prediction model are taken as the trained first gain prediction model and the trained second gain prediction model respectively.

[0131] The preset coincidence degree condition refers to a condition preset for judging whether the coincidence degree meets the expectation. The preset coincidence degree condition can be determined according to specific application scenarios and requirements. For example, the preset coincidence degree condition can include one or more of a minimum value requirement for the coincidence degree, a threshold setting or other constraint conditions. For example, the coincidence degree condition can be to reach or be greater than a coincidence degree threshold. When the coincidence degree meets or exceeds the preset performance condition, the preset coincidence degree condition is met.

[0132] For example, if the coincidence degree does not meet the expectation (e.g., is less than or equal to a preset proportion), the model parameters of the first gain prediction model and the second gain prediction model are iteratively updated. In each iteration process, the first response gain and the second response gain of each sample object are predicted according to the first gain prediction model and the second gain prediction model of the iteration, to calculate the coincidence degree of the object set in the iteration, until the coincidence degree calculated in any iteration meets the expectation (e.g., is greater than the preset proportion), and the first gain prediction model and the second gain prediction model updated in the last iteration are taken as the trained first gain prediction model and the trained second gain prediction model respectively.

[0133] In some embodiments, in the joint iterative training of the first gain prediction model and the second gain prediction model based on the coincidence degree of the prediction result, the first gain prediction model and the second gain prediction model can be trained separately based on the response gain, so as to optimize the training result of the gain prediction model through a two-layer iterative training process, to improve the prediction accuracy and reliability of the trained model. The outer-layer iterative training process is the joint iterative training of the first gain prediction model and the second gain prediction model, which can combine the information of the two models to jointly evaluate the accuracy of the prediction result of the gain prediction model. The inner-layer iterative training process is to train the first gain prediction model and the second gain prediction model separately based on the response gain in each outer-layer iteration, which can adjust the parameters of each model to minimize the prediction error associated with it. Specifically, when the coincidence degree does not meet the preset coincidence degree condition, the first gain prediction model and the second gain prediction model are updated to obtain an updated first gain prediction model and an updated second gain prediction model, including:

[0134] When the coincidence degree does not satisfy the preset coincidence degree condition, the first gain prediction model and the second gain prediction model are returned as a first initial gain prediction model and a second initial gain prediction model respectively, and the first initial intervention response prediction network and the initial control response prediction network are executed to predict a first initial intervention response of the first sample object and a first initial control response of the second sample object respectively, and the first initial intervention response and the first initial control response are used to train the first initial gain prediction model to obtain the first gain prediction model.

[0135] The trained first gain prediction model and the trained second gain prediction model are used as an updated first gain prediction model and an updated second gain prediction model respectively.

[0136] For example, if the coincidence degree does not satisfy the expectation (e.g., is less than or equal to the preset proportion), the model parameters of the first gain prediction model and the second gain prediction model are updated through outer iteration. In each round of outer iteration, the first gain prediction model and the second gain prediction model are used as a first initial gain prediction model and a second initial gain prediction model respectively, and the first initial gain prediction model and the second initial gain prediction model are used to update the first gain prediction model and the second gain prediction model through multiple rounds of inner iteration. In each round of outer iteration, the first response gain and the second response gain of each sample object are predicted according to the first gain prediction model and the second gain prediction model of the round, and the coincidence degree of the object set in the round is calculated until the coincidence degree calculated in any round satisfies the expectation (e.g., is greater than the preset proportion), and the first gain prediction model and the second gain prediction model updated in the last outer iteration are used as a trained first gain prediction model and a trained second gain prediction model respectively.

[0137] 150. Predicting the predicted response gain of the intervention processing on the to-be-predicted object by using the trained first gain prediction model or the trained second gain prediction model.

[0138] The to-be-predicted object refers to an object whose response gain is to be predicted. The predicted response gain refers to the response gain predicted for the to-be-predicted object, and the predicted response gain can be used to evaluate the intervention effect (i.e., gain) of predicting the intervention processing on the to-be-predicted object. In the embodiments of the present application, the predicted response gain is used to determine the processing strategy of the intervention processing on the to-be-predicted object. The processing strategy refers to the strategy of performing intervention processing or not performing intervention processing.

[0139] For example, when obtaining separate consent or permission of a business user, the business user interacting with the business subject can be taken as a to-be-predicted object to obtain object features of the to-be-predicted object. The object features of the to-be-predicted object can be input into the trained first gain prediction model or the trained second gain prediction model to obtain a predicted response gain of intervention processing on the to-be-predicted object. Whether the intervention processing is to be performed on the to-be-predicted object is determined according to the predicted response gain. In the embodiments of the present application, the response gain of the intervention processing on the user is predicted by using the gain prediction model, so as to quantitatively evaluate the acceptance degree of the user to the intervention processing, to identify the user who is more sensitive to the specific message prompt, and to further improve the accuracy of the predicted response gain, so as to determine a more accurate intervention processing strategy based on the predicted response gain, and to improve the accuracy of the intervention processing.

[0140] In some embodiments, whether the intervention processing is performed can be evaluated by specifying a gain type, so that the decision-making process of the intervention processing is more standardized, and the decision-making efficiency is higher. Specifically, the intervention processing includes message prompting, and after the predicted response gain of the intervention processing on the to-be-predicted object is predicted by using the trained first gain prediction model or the trained second gain prediction model, the intervention processing further includes:

[0141] When the predicted response gain meets the specified gain type, the to-be-predicted object is prompted by a message.

[0142] The specified gain type refers to a type of response gain that is pre-set and can be used for message prompting. The specified gain type can be determined according to specific application scenarios and requirements, for example, the specified gain type can be a prompt-sensitive type.

[0143] For example, the object features of the to-be-predicted object can be input into the trained intervention response prediction network to perform response prediction, to obtain a trained intervention response of the to-be-predicted object. The object features of the to-be-predicted object can be input into the trained control response prediction network to perform response prediction, to obtain a trained control response of the to-be-predicted object. The predicted response gain of the to-be-predicted object can be calculated by using difference calculation, ratio calculation, and percentage difference comparison analysis method on the trained intervention response and the trained control response of the to-be-predicted object. If the predicted response gain matches the value of the response gain corresponding to the specified gain type, the to-be-predicted object is prompted by a message.

[0144] For another example, the object features of the to-be-predicted object can be input into the trained hybrid response prediction network to perform response prediction, to obtain a trained intervention response and a trained control response of the to-be-predicted object. The predicted response gain of the to-be-predicted object can be calculated by using difference calculation, ratio calculation, and percentage difference comparison analysis method on the trained intervention response and the trained control response of the to-be-predicted object. If the predicted response gain is greater than a preset prediction threshold, the to-be-predicted object is prompted by a message.

[0145] In some embodiments, the specified gain type can be represented as a preset gain threshold. For example, a preset gain threshold can be set, and when the predicted response gain of the to-be-predicted object is greater than the preset gain threshold, the predicted response gain belongs to the type of alert sensitivity, so as to send a message alert to the to-be-predicted object. The preset gain threshold can be in the form of a numerical value, a percentage, etc., and can be set according to actual needs or application scenarios.

[0146] In some embodiments, the mean of the response gains of a plurality of to-be-predicted objects with similar characteristics (i.e., a group with similar characteristics) can be predicted, so as to determine whether to intervene in these to-be-predicted objects according to the mean of the response gains, so as to achieve the same intervention on the group with similar characteristics and reduce the cost of intervention. In addition, group prediction can smooth out the abnormal values of individuals to some extent, reducing the risk of intervention decision errors due to individual differences. Specifically, there are a plurality of to-be-predicted objects, and the predicted response gain of the to-be-predicted object to the intervention is predicted by using the trained first gain prediction model or the trained second gain prediction model, including:

[0147] The object response gain of each to-be-predicted object to the intervention is predicted by using the trained first gain prediction model or the trained second gain prediction model.

[0148] The object response gain is subjected to mean processing to obtain the predicted response gain of the to-be-predicted object to the intervention.

[0149] The object response gain refers to the response gain of each sample object predicted by the trained first gain prediction model or the trained second gain prediction model. The mean processing refers to a method of calculating the average value of a group of data.

[0150] For example, after obtaining the post-training intervention response and the post-training control response of each to-be-predicted object, the post-training intervention response and the post-training control response of each to-be-predicted object can be subtracted to obtain the object response gain of each to-be-predicted object. The average value of the object response gain of all to-be-predicted objects is calculated to obtain the predicted response gain.

[0151] The data processing scheme provided in the embodiments of the present application can be applied in various intervention processing scenarios. For example, taking message reminding as an example, a to-be-predicted sample object, a first gain prediction model and a second gain prediction model are obtained, the first gain prediction model includes an intervention response prediction network trained by a first sample object and a control response prediction network trained by a second sample object, and the second gain prediction model includes a mixed response prediction network trained by the first sample object and the second sample object, the first sample object is a sample object that has undergone intervention processing, and the second sample object is a sample object that has not undergone intervention processing; the first intervention response and the first control response of the to-be-predicted sample object are predicted by the intervention response prediction network and the control response prediction network respectively, so as to determine a first response gain of the intervention processing on the to-be-predicted sample object from the first intervention response and the first control response; the second intervention response and the second control response of the to-be-predicted sample object are predicted by the mixed response prediction network, so as to determine a second response gain of the intervention processing on the to-be-predicted sample object from the second intervention response and the second control response; the first gain prediction model and the second gain prediction model are jointly trained in combination with the first response gain and the second response gain, to obtain a trained first gain prediction model and a trained second gain prediction model; and the prediction response gain of the intervention processing on the to-be-predicted object is predicted by the trained first gain prediction model or the trained second gain prediction model, and the prediction response gain is used to determine a processing strategy of the intervention processing on the to-be-predicted object.

[0152] As can be seen from the above, the embodiments of the present application use different first gain prediction models and second gain prediction models to predict the response gain of the intervention processing on the sample object, to combine the prediction results of the two gain prediction models for joint training, and to guide the training process of the two gain prediction models. The joint training can utilize the correlation between the prediction tasks of the two gain prediction models to improve the volatility of the output of a single gain prediction model, so that the trained gain model can output stable and consistent prediction results, improve the accuracy of the prediction response gain, determine a more accurate intervention processing strategy based on the predicted response gain, and improve the accuracy of the intervention processing.

[0153] The method described in the above embodiments will be further described in detail below.

[0154] In this embodiment, the method of the present application will be described in detail by taking message reminding for customers of an entity as an example.

[0155] As shown in Figure 2a and Figure 2b , a specific process of a data processing method is as follows:

[0156] 210, a sample object set, a first initial gain prediction model and a second initial gain prediction model are obtained.

[0157] For example, as shown in the data processing method shown in Figure 2b , the data processing method can include a training phase (steps 210-260) and an inference phase (step 270). Among them, the training phase can be divided into two layers of training, the outer layer is the joint training of two gain prediction models, and the inner layer is the independent training of each gain prediction model. In the embodiment of the application, the sample object can be a customer of an entity.

[0158] Specifically, the object information of a plurality of (such as d) sample objects can be obtained for the training phase, which can come from the feature data of the historical period of the historical customer group. For example, the object information of the d sample objects can be preprocessed such as missing value filling, abnormal value replacement, etc., and vectorized to obtain a feature (covariate) set of the object features (x1, x2, x3, …, x d ) of the d sample objects. The object features of the d sample objects can form a sample data set D (i.e. a sample object set, the sample objects in the sample object set are also the to-be-predicted sample objects in the foregoing embodiments), such as (y, t, x1, x2, …, x d ), wherein y is the target variable (i.e. the sample label), t is whether to remind the user (i.e. the intervention mark), and r is the proportion of the customer group that needs to be reminded. Generally, if intervention processing is performed, t = 1, and if intervention processing is not performed, t = 0. In the embodiment of the application, the sample data set D includes the object features of the first sample objects (i.e. the intervention group) and the object features of the second sample objects (i.e. the non-intervention group), the sample label of the first sample object is 1, and the sample label of the second sample object is 0. In the embodiment of the application, the intervention processing can be that the entity sends a specific message reminder to the customer, for example, in the financial business scenario, the specific message reminder can be a repayment reminder.

[0159] 220, by the sample object set, training the first initial gain prediction model to obtain the first gain prediction model.

[0160] For example, as shown in the data processing method shown in Figure 2b , in the inner layer of the training phase, the sample data set D can be used to independently train the first initial gain prediction model to obtain the first gain prediction model, which can be a T-learner (double model). The T-learner is a method in the Uplift model, which predicts the response of the intervention group and the non-intervention group by establishing two models, and then calculates the difference between the two models to evaluate the effect of the strategy. The Uplift model is a model for evaluating marketing strategies by analyzing intervention effects. It compares the behavior differences of the intervention group and the non-intervention group under the marketing strategy by dividing the customers into the intervention group and the non-intervention group, so as to evaluate the effectiveness of the marketing strategy.

[0161] Specifically, the independent training process (i.e., step 220) of the first gain prediction model as shown in FIG. 2 can include the following steps 221-228. Figure 2c

[0162] 221. Predicting, by the initial intervention response prediction network, a first initial intervention response of a first sample object in the first object group.

[0163] 222. Predicting, by the initial control response prediction network, a first initial control response of a second sample object in the first object group.

[0164] 223. Updating, based on the first initial intervention response and the first initial control response, the first initial gain prediction model to obtain an intermediate gain prediction model.

[0165] The intermediate gain prediction model includes an intermediate intervention response prediction network updated by the initial intervention response prediction network, and an intermediate control response prediction network updated by the initial control response prediction network.

[0166] 224. Predicting, by the intermediate intervention response prediction network and the intermediate control response prediction network, an intermediate intervention response and an intermediate control response of a sample object in the second object group, respectively.

[0167] 225. Determining, according to the intermediate intervention response and the intermediate control response, an intermediate response gain of the intervention treatment on the sample object in the second object group.

[0168] 226. Determining, based on the intermediate response gain, a performance parameter of the intermediate gain prediction model.

[0169] 227. When the performance parameter does not satisfy a preset performance condition, taking the intermediate gain prediction model as the first initial gain prediction model, returning to execute steps 221-227 until the performance parameter satisfies the preset performance condition.

[0170] 228. When the performance parameter satisfies the preset performance condition, training the intermediate gain prediction model by the sample object set to obtain the first gain prediction model.

[0171] For example, a sample data set D can be taken as an input of the first initial gain prediction model. The sample data set D includes a feature set X of object features (i.e., intervention group) of the first sample object and object features (non-intervention group) of the second sample object, and a label set Y. The output of the model is an intervention effect prediction model f(x), where x is a feature vector of the object features.

[0172] ​The sample data set D can be randomly divided into two parts, i.e., a sub-data set D1 and a sub-data set D2, and the object features of the first sample object and the object features of the second sample object are included in the sub-data set D1 and the sub-data set D2. The intervention group and the non-intervention group in the sub-data set D1 can be used to train two classifiers f1(x) (i.e., an intervention response prediction network) and f0(x) (i.e., a control response prediction network) respectively, wherein f1(x) and f0(x) are used to predict the responses of the intervention group and the non-intervention group respectively.

[0173] The responses of the intervention group and the non-intervention group in the sub-data set D1 are predicted by f1(x) and f0(x) of the first initial gain prediction model respectively to obtain predicted labels y1 (i.e., first initial intervention responses) and y0 (first initial control responses). The model parameters of the first initial gain prediction model can be adjusted according to the difference between the predicted target (i.e., sample label) of the first initial intervention response and the first initial control response, until the difference between the first initial response gain and the predicted target is less than a preset threshold, and the gain prediction model obtained by the last adjustment is the intermediate gain prediction model. The preset threshold can be set according to actual conditions.

[0174] The responses of the intervention group and the non-intervention group in the sub-data set D1 can be predicted by f1(x) and f0(x) of the intermediate gain prediction model to obtain predicted labels y1 (i.e., intermediate intervention responses) and predicted labels y0 (intermediate control responses) of each sample object in the intervention group and the non-intervention group. The predicted value τ(x) (i.e., intermediate response gain) of the treatment effect of each sample object can be obtained according to the difference between the intermediate intervention response and the intermediate control response of each sample object. The performance parameters of the intermediate gain prediction model can be calculated based on the intermediate response gain, such as a combination of one or more of the following: area under the uplift region (AUUC), mean square error (MSE), mean absolute error (MAE), coefficient of variation (CV), etc. When the performance parameters do not meet the preset performance conditions, the intermediate gain prediction model is used as the first initial gain prediction model, and steps 221-227 are repeated until the performance parameters meet the preset performance conditions. The f1(x) and f0(x) of the intermediate gain prediction model updated last time are trained using the entire data set D to obtain the final intervention effect prediction model (i.e., the first gain prediction model).

[0175] Specifically, the T-learner model predicts the responses of the intervention group and the non-intervention group by training two classifiers f1(x) and f0(x) and calculates the difference between them, i.e., the treatment effect. Among them, the training of f1(x) and f0(x) is respectively for the intervention group and the non-intervention group. For the intervention group, f1(x) is trained using the intervention group samples in the data set D1. Let the output of f1(x) be y1(x). For the non-intervention group, f0(x) is trained using the non-intervention group samples in the data set D1. Let the output of f0(x) be y0(x). Then, for the intervention group and the non-intervention group in the data set D2, f1(x) and f0(x) are used for prediction to obtain the predicted labels y1 and y0. Finally, the difference between y1 and y0 is calculated to obtain the predicted value τ(x) of the treatment effect. Then f1(x) and f0(x) are trained using the entire data set D to obtain the final intervention effect prediction model.

[0176] 230. Training a second initial gain prediction model through the sample object set to obtain a second gain prediction model.

[0177] For example, as shown in the data processing method, Figure 2b the second initial gain prediction model can be independently trained in the inner layer of the training stage using the sample data set to obtain a second gain prediction model. The gain prediction model can be an S-learner (double model). The S-learner is a method in the Uplift model, which is a single-model-based method that trains a classifier to predict the response of all samples to obtain the predicted value of the intervention effect.

[0178] Specifically, as shown in the independent training process of the second initial gain prediction model (i.e., step 230), Figure 2d the independent training process includes the following steps 231-232:

[0179] 231. Predicting a second initial intervention response of the first sample object and a second initial control response of the second sample object through an initial mixed response prediction network.

[0180] 232. Training a second initial gain prediction model based on the second initial intervention response and the second initial control response to obtain a second gain prediction model.

[0181] For example, as shown in the data processing method, Figure 2b the sample data set D can be used as the input of the second initial gain prediction model. The sample data set D includes a feature set X containing the object features of the first sample object (i.e., the intervention group) and the object features of the second sample object (the non-intervention group), and a label set Y. The output of the model is an intervention effect prediction model f(x), where x is the feature vector of the object features.

[0182] The second initial gain prediction model can be trained by all sample objects in the sample data set D, and a trained classifier f(x) (i.e., a mixed response prediction network) is obtained. During the training process, the sample data set D can be divided into an intervention group and a non-intervention group to respectively predict the response of the intervention group (i.e., a second initial intervention response) and the response of the non-intervention group (i.e., a second initial control response). The model parameters of the second initial gain prediction model can be adjusted according to the difference between the second initial intervention response and the second initial control response and the predicted target (i.e., a sample label) of the sample object, until the difference between the second initial response gain and the predicted target is less than a preset threshold. The gain prediction model obtained by the last adjustment is the second gain prediction model. The preset threshold can be set according to actual conditions.

[0183] 240. The first intervention response and the first control response of the sample object are respectively predicted by the intervention response prediction network and the control response prediction network, so as to determine the first response gain of the sample object by the first intervention response and the first control response.

[0184] 250. The second intervention response and the second control response of the sample object are predicted by the mixed response prediction network, so as to determine the second response gain of the sample object by the second intervention response and the second control response.

[0185] For example, as shown in the data processing method shown in Figure 2b the outer layer of the training stage, the sample data set D can be input into the first gain prediction model and the second gain prediction model trained to respectively predict the first intervention response and the first control response and the second intervention response and the second control response of each sample object, so as to obtain the first response gain by the difference between the first intervention response and the first control response of each sample object, and obtain the second response gain by the difference between the second intervention response and the second control response of each sample object.

[0186] 260. The first gain prediction model and the second gain prediction model are jointly trained in combination with the first response gain and the second response gain, and a trained first gain prediction model and a trained second gain prediction model are obtained.

[0187] For example, as shown in the data processing method shown in Figure 2b the outer layer of the training stage, the first gain prediction model and the second gain prediction model can be jointly trained in combination with the first response gain and the second response gain.

[0188] Specifically, the joint training process (i.e., step 260) of the first gain prediction model and the second gain prediction model as shown in Figure 2e includes the following steps 261-263:

[0189] 261、determining a first object set and a second object set from the sample object set, the first object set corresponding to a first response gain satisfying a gain threshold condition, and the second object set corresponding to a second response gain satisfying the gain threshold condition.

[0190] 262、updating the first gain prediction model and the second gain prediction model when the coincidence degree does not satisfy the preset coincidence degree condition, to obtain an updated first gain prediction model and an updated second gain prediction model.

[0191] 263、returning to execute steps 210-260 with the first gain prediction model and the second gain prediction model as a first initial gain prediction model and a second initial gain prediction model respectively, until the coincidence degree satisfies the preset coincidence degree condition, and taking the updated first gain prediction model and the updated second gain prediction model as a trained first gain prediction model and a trained second gain prediction model respectively.

[0192] For example, the sample objects in the sample data set D can be sorted according to the first response gain and the second response gain respectively, and the top 10% of the sample objects in each sorting result (i.e., the head customer group) can be taken as the first object set and the second object set respectively. Then, the proportion of the sample objects (i.e., the coincidence degree) that coincide in the first object set and the second object set (i.e., the intersection) in the sample objects in the sample data set D is calculated. If the coincidence degree meets the expectation (e.g., greater than a preset proportion), the updated first gain prediction model and the updated second gain prediction model are taken as the trained first gain prediction model and the trained second gain prediction model respectively. If the coincidence degree does not meet the expectation (e.g., less than or equal to the preset proportion), the model parameters of the first gain prediction model and the second gain prediction model are iteratively updated (i.e., returning to execute steps 210-260).

[0193] 270、predicting the predicted response gain of the intervention processing on the to-be-predicted object by the trained first gain prediction model or the trained second gain prediction model.

[0194] 280、when the predicted response gain satisfies a specified gain type, sending a message reminder to the to-be-predicted object.

[0195] For example, the object features of the to-be-predicted object can be input into the trained intervention response prediction network for response prediction to obtain the trained intervention response of the to-be-predicted object. The object features of the to-be-predicted object can also be input into the trained contrast response prediction network for response prediction to obtain the trained contrast response of the to-be-predicted object. The difference between the trained intervention response and the trained contrast response of the to-be-predicted object can be taken as the predicted response gain of the to-be-predicted object. If the predicted response gain satisfies the specified gain type, a specific message reminder can be sent to the to-be-predicted object.

[0196] In this embodiment, the objects to be predicted can be further categorized into four types based on the predicted response gain: reminder-sensitive, natural recall, churn, and reminder-reverse. For example, if the predicted response gain belongs to the reminder-sensitive type, a specific service reminder is sent to the object to be predicted. Alternatively, a preset gain threshold can be used; if the predicted response gain of the object to be predicted is greater than the preset gain threshold, then the predicted response gain belongs to the reminder-sensitive type, and a specific service reminder is sent to the object to be predicted.

[0197] As can be seen from the above, in this embodiment, different first and second gain prediction models are used to predict the response gain of the intervention treatment on the sample object. The prediction results of these two gain prediction models are then combined for joint training to guide the training process of these two gain prediction models. Joint training can utilize the correlation between the prediction tasks of these two gain prediction models to improve the volatility of the output of a single gain prediction model, enabling the trained gain model to output stable and consistent prediction results, improving the accuracy of predicting response gain, and determining a more precise intervention strategy based on the predicted response gain, thereby improving the accuracy of the intervention treatment.

[0198] To better implement the above methods, this application also provides a data processing device, which can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, personal computer, etc.; the server can be a single server or a server cluster composed of multiple servers.

[0199] For example, in this embodiment, the method of this application embodiment will be described in detail by taking the data processing device specifically integrated into the server as an example.

[0200] For example, such as Figure 3 As shown, the data processing device may include an acquisition unit 310, a first prediction unit 320, a second prediction unit 330, a training unit 340, and a third prediction unit 350, as follows:

[0201] (I) Acquisition Unit 310

[0202] The method is used to obtain the sample object to be predicted, the first gain prediction model, and the second gain prediction model. The first gain prediction model includes an intervention response prediction network trained on the first sample object and a control response prediction network trained on the second sample object. The second gain prediction model includes a hybrid response prediction network trained on the first sample object and the second sample object. The first sample object is the sample object that has undergone intervention treatment, and the second sample object is the sample object that has not undergone intervention treatment.

[0203] In some embodiments, the acquisition unit comprises a first acquisition subunit, a second acquisition subunit, and a third acquisition subunit, and comprises: the first acquisition subunit is configured to acquire a sample object set and a first initial gain prediction model, the first initial gain prediction model comprises an initial intervention response prediction network and an initial control response prediction network, and the sample object set comprises a first sample object and a second sample object; the second acquisition subunit is configured to predict a first initial intervention response of the first sample object and a first initial control response of the second sample object through the initial intervention response prediction network and the initial control response prediction network, respectively; and the third acquisition subunit is configured to train the first initial gain prediction model based on the first initial intervention response and the first initial control response to obtain a first gain prediction model.

[0204] In some embodiments, the sample object set comprises a first object group, the first object group comprises the first sample object and the second sample object, and the second acquisition subunit is specifically configured to: predict the first initial intervention response of the first sample object in the first object group through the initial intervention response prediction network; and predict the first initial control response of the second sample object in the first object group through the initial control response prediction network.

[0205] In some embodiments, the sample object set further comprises a second object group, and the third acquisition subunit is specifically configured to: update the first initial gain prediction model based on the first initial intervention response and the first initial control response to obtain an intermediate gain prediction model, the intermediate gain prediction model comprises an intermediate intervention response prediction network updated from the initial intervention response prediction network and an intermediate control response prediction network updated from the initial control response prediction network; predict an intermediate intervention response and an intermediate control response of a sample object to be predicted in the second object group through the intermediate intervention response prediction network and the intermediate control response prediction network, respectively; determine an intermediate response gain of the intervention processing on the sample object to be predicted in the second object group according to the intermediate intervention response and the intermediate control response; and obtain the first gain prediction model from the intermediate gain prediction model based on the intermediate response gain.

[0206] In some embodiments, obtaining the first gain prediction model from the intermediate gain prediction model based on the intermediate response gain comprises: determining a performance parameter of the intermediate gain prediction model based on the intermediate response gain; and when the performance parameter meets a preset performance condition, training the intermediate gain prediction model to obtain the first gain prediction model through the sample object set.

[0207] In some embodiments, the first gain prediction model is obtained from the intermediate gain prediction model based on the intermediate response gain, and further comprising: when the performance parameter does not satisfy the preset performance condition, taking the intermediate gain prediction model as a first initial gain prediction model, and returning to execute the step of predicting a first initial intervention response of the first sample object and a first initial control response of the second sample object by the initial intervention response prediction network and the initial control response prediction network respectively, and the subsequent steps until the performance parameter satisfies the preset performance condition.

[0208] In some embodiments, the obtaining unit comprises a fourth obtaining subunit, a fifth obtaining subunit and a sixth obtaining subunit, comprising: the fourth obtaining subunit is configured to obtain a sample object set and a second initial gain prediction model, the second initial gain prediction model comprising an initial mixed response prediction network, and the sample object set comprising the first sample object and the second sample object; the fifth obtaining subunit is configured to predict a second initial intervention response of the first sample object and a second initial control response of the second sample object by the initial mixed response prediction network; and the sixth obtaining subunit is configured to train the second initial gain prediction model based on the second initial intervention response and the second initial control response to obtain a second gain prediction model.

[0209] (ii) the first prediction unit 320

[0210] for predicting a first intervention response and a first control response of the to-be-predicted sample object by the intervention response prediction network and the control response prediction network respectively, to determine a first response gain of the intervention processing on the to-be-predicted sample object from the first intervention response and the first control response.

[0211] (iii) the second prediction unit 330

[0212] for predicting a second intervention response and a second control response of the to-be-predicted sample object by the mixed response prediction network, to determine a second response gain of the intervention processing on the to-be-predicted sample object from the second intervention response and the second control response.

[0213] (iv) the training unit 340

[0214] for jointly training the first gain prediction model and the second gain prediction model in combination with the first response gain and the second response gain to obtain a trained first gain prediction model and a trained second gain prediction model.

[0215] In some embodiments, the training unit comprises a first training subunit and a second training subunit, the first training subunit is configured to determine a first object set and a second object set from the sample object to be predicted, the first object set corresponds to a first response gain satisfying a gain threshold condition, and the second object set corresponds to a second response gain satisfying the gain threshold condition; and the second training subunit is configured to jointly train the first gain prediction model and the second gain prediction model based on a coincidence degree of the first object set and the second object set, to obtain the trained first gain prediction model and the trained second gain prediction model.

[0216] In some embodiments, the second training subunit is specifically configured to: when the coincidence degree does not satisfy a preset coincidence degree condition, update the first gain prediction model and the second gain prediction model to obtain an updated first gain prediction model and an updated second gain prediction model; and return to execute the step of obtaining the sample object, the first gain prediction model and the second gain prediction model, and the subsequent steps, with the updated first gain prediction model and the updated second gain prediction model as the first gain prediction model and the second gain prediction model respectively, until the coincidence degree satisfies the preset coincidence degree condition, and the updated first gain prediction model and the updated second gain prediction model are taken as the trained first gain prediction model and the trained second gain prediction model respectively.

[0217] (Five) The third prediction unit 350

[0218] The trained first gain prediction model or the trained second gain prediction model is used to predict a predicted response gain of the intervention processing on the object to be predicted, and the predicted response gain is used to determine a processing strategy of the intervention processing on the object to be predicted.

[0219] In some embodiments, the data processing apparatus further comprises a message reminding unit, the message reminding unit is configured to: when the predicted response gain satisfies a specified gain type, perform message reminding on the object to be predicted.

[0220] In some embodiments, the data processing apparatus further comprises a processing unit, the processing unit is configured to: perform data preprocessing on object information of the sample object to be predicted to obtain processed information; and perform vectorization processing on the processed information to obtain object features of the sample object to be predicted, the object features being used to predict an intervention response and a control response of the sample object to be predicted.

[0221] In implementation, each of the above units can be implemented as an independent entity, or can be combined as the same or several entities. The implementation of each of the above units can be referred to the method embodiments above, and will not be described here.

[0222] As can be seen, the embodiments of the present application use different first gain prediction models and second gain prediction models to predict the response gain of the sample object to the intervention processing, combine the prediction results of the two gain prediction models to perform joint training, and guide the training process of the two gain prediction models. The joint training can utilize the correlation between the prediction tasks of the two gain prediction models to improve the volatility of the output of a single gain prediction model, enable the trained gain model to output stable and consistent prediction results, improve the accuracy of predicting the response gain, determine a more accurate intervention processing strategy based on the predicted response gain, and improve the accuracy of the intervention processing.

[0223] The embodiments of the present application also provide an electronic device, which can be a terminal, a server, etc. The terminal can be a mobile phone, a tablet computer, a smart Bluetooth device, a notebook computer, a personal computer, etc. The server can be a single server or a server cluster composed of multiple servers, etc.

[0224] In some embodiments, the data processing apparatus can also be integrated in multiple electronic devices, for example, the data processing apparatus can be integrated in multiple servers to implement the data processing method of the present application by the multiple servers.

[0225] In the present embodiment, the electronic device of the present embodiment is taken as an example to be described in detail, for example, as shown in Figure 4 The server structure related to the embodiments of the present application is shown, and specifically:

[0226] The server can include a processor 410 with one or more processing cores, a memory 420 with one or more computer readable storage media, a power supply 430, an input module 440, a communication module 450, etc. Those skilled in the art can understand that the server structure shown in Figure 4 does not limit the server, and the server can include more or fewer components than those shown in the figure, or combine certain components, or different component arrangements. Among them:

[0227] The processor 410 is the control center of the server, which connects all parts of the server through various interfaces and lines, executes various functions of the server and processes data by running or executing software programs and / or modules stored in the memory 420 and calling data stored in the memory 420. In some embodiments, the processor 410 can include one or more processing cores; in some embodiments, the processor 410 can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application program, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 410.

[0228] The memory 420 can be used to store software programs and modules, and the processor 410 can execute various function applications and data processing by running the software programs and modules stored in the memory 420. The memory 420 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, application programs required by at least one function (such as a sound playing function, an image playing function, etc.), and the like; and the data storage area can store data created according to the use of the server, etc. In addition, the memory 420 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. Accordingly, the memory 420 can also include a memory controller to provide the processor 410 with access to the memory 420.

[0229] The server further includes a power supply 430 for supplying power to various components, and in some embodiments, the power supply 430 can be logically connected to the processor 410 through a power management system, so that the power management system can realize functions such as management of charging, discharging, and power consumption management. The power supply 430 can also include one or more direct current or alternating current power supplies, a recharging system, a power failure detection circuit, a power converter or inverter, a power state indicator, and the like.

[0230] The server can further include an input module 440, which can be used to receive input digital or character information, and to generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function controls.

[0231] The server can further include a communication module 450, which in some embodiments can include a wireless module, and the server can perform short-range wireless transmission through the wireless module of the communication module 450, thereby providing the user with wireless broadband Internet access. For example, the communication module 450 can be used to help the user send and receive emails, browse web pages, and access streaming media, etc.

[0232] Although not shown, the server can also include a display unit, etc., which will not be described here. Specifically, in the present embodiment, the processor 410 in the server will load the executable file corresponding to the process of one or more application programs into the memory 420 according to the following instructions, and the processor 410 will run the application programs stored in the memory 420 to realize various functions, as follows:

[0233] The first gain prediction model includes an intervention response prediction network trained by a first sample object and a control response prediction network trained by a second sample object, and the second gain prediction model includes a mixed response prediction network trained by the first sample object and the second sample object. The first sample object is a sample object that has been subjected to intervention processing, and the second sample object is a sample object that has not been subjected to intervention processing. The first intervention response and the first control response of the sample object to be predicted are predicted by the intervention response prediction network and the control response prediction network, respectively, to determine a first response gain of the sample object to be predicted by the intervention processing. The second intervention response and the second control response of the sample object to be predicted are predicted by the mixed response prediction network, to determine a second response gain of the sample object to be predicted by the intervention processing. The first gain prediction model and the second gain prediction model are jointly trained by combining the first response gain and the second response gain, to obtain a trained first gain prediction model and a trained second gain prediction model. The prediction response gain of the sample object to be predicted by the intervention processing is predicted by the trained first gain prediction model or the trained second gain prediction model, and the prediction response gain is used to determine a processing strategy for the sample object to be predicted by the intervention processing.

[0234] The specific implementation of each operation can refer to the foregoing embodiments, which will not be described here.

[0235] As can be seen from the above, the embodiments of the present application use different first gain prediction models and second gain prediction models to predict the response gain of the sample object by the intervention processing, to combine the prediction results of the two gain prediction models for joint training, and guide the training process of the two gain prediction models. The joint training can utilize the correlation between the prediction tasks of the two gain prediction models to improve the volatility of the output of a single gain prediction model, so that the trained gain model can output stable and consistent prediction results, improve the accuracy of the prediction response gain, determine a more accurate intervention processing strategy based on the predicted response gain, and improve the accuracy of the intervention processing.

[0236] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by related hardware controlled by the instructions, which can be stored in a computer readable storage medium and loaded and executed by a processor.

[0237] Therefore, the embodiments of the present application provide a computer readable storage medium, which stores a plurality of instructions. The instructions can be loaded by a processor to execute the steps in any data processing method provided by the embodiments of the present application. For example, the instructions can execute the following steps:

[0238] obtaining a sample object to be predicted, a first gain prediction model and a second gain prediction model, the first gain prediction model comprising an intervention response prediction network trained by a first sample object and a control response prediction network trained by a second sample object, the second gain prediction model comprising a mixed response prediction network trained by the first sample object and the second sample object, the first sample object being a sample object subjected to intervention processing, and the second sample object being a sample object not subjected to intervention processing; predicting a first intervention response and a first control response of the sample object to be predicted by the intervention response prediction network and the control response prediction network respectively, to determine a first response gain of the intervention processing on the sample object to be predicted from the first intervention response and the first control response; predicting a second intervention response and a second control response of the sample object to be predicted by the mixed response prediction network, to determine a second response gain of the intervention processing on the sample object to be predicted from the second intervention response and the second control response; training the first gain prediction model and the second gain prediction model jointly based on the first response gain and the second response gain, to obtain a trained first gain prediction model and a trained second gain prediction model; and predicting a prediction response gain of the intervention processing on the sample object to be predicted by the trained first gain prediction model or the trained second gain prediction model, the prediction response gain being used to determine a processing strategy of the intervention processing on the sample object to be predicted.

[0239] The storage medium can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disk, or the like.

[0240] According to an aspect of the present disclosure, a computer program product or a computer program is provided, including computer programs or instructions, which, when executed by a processor, implement the steps in the methods provided in the various optional implementation manners provided in the above embodiments. The computer programs / instructions are stored in a computer readable storage medium. The processor of the electronic device reads the computer programs / instructions from the computer readable storage medium, and the processor executes the computer programs / instructions, so that the electronic device executes the methods provided in the various optional implementation manners provided in the above embodiments.

[0241] Since the instructions stored in the storage medium can execute the steps in any of the data processing methods provided in the embodiments of the present disclosure, the beneficial effects that can be achieved by any of the data processing methods provided in the embodiments of the present disclosure can be achieved. Details are described in the above embodiments, which will not be described here.

[0242] The above describes in detail the data processing method, device, electronic device, storage medium and program product provided by the embodiments of the application. The principles and implementation manners of the application are described by applying specific examples. The above description of the embodiments is only used to help understand the method of the application and its core idea. Meanwhile, for those skilled in the art, according to the idea of the application, the specific implementation manners and application ranges can be changed. In conclusion, the content of the specification should not be understood as a limitation of the application.

Claims

1. A data processing method, characterized in that, include: The sample object to be predicted, the first gain prediction model, and the second gain prediction model are obtained. The first gain prediction model includes an intervention response prediction network trained by the first sample object and a control response prediction network trained by the second sample object. The second gain prediction model includes a hybrid response prediction network trained by the first sample object and the second sample object. The first sample object is a sample object that has undergone intervention treatment, and the second sample object is a sample object that has not undergone the intervention treatment. The intervention response prediction network and the control response prediction network are used to predict the first intervention response and the first control response of the sample object to be predicted, respectively, so as to determine the first response gain of the intervention treatment on the sample object to be predicted by the first intervention response and the first control response. The hybrid response prediction network is used to predict the second intervention response and the second control response of the sample object to be predicted, so as to determine the second response gain of the intervention treatment on the sample object to be predicted by the second intervention response and the second control response; By combining the first response gain and the second response gain, the first gain prediction model and the second gain prediction model are jointly trained to obtain the trained first gain prediction model and the trained second gain prediction model. The predicted response gain of the intervention treatment on the object to be predicted is predicted by the first gain prediction model or the second gain prediction model after training, and the predicted response gain is used to determine the treatment strategy for the intervention treatment on the object to be predicted.

2. The data processing method as described in claim 1, characterized in that, The first gain prediction model is obtained through the following steps: Obtain a sample object set and a first initial gain prediction model, the first initial gain prediction model including an initial intervention response prediction network and an initial control response prediction network, the sample object set including the first sample object and the second sample object; The first initial intervention response of the first sample object and the first initial control response of the second sample object are predicted by the initial intervention response prediction network and the initial control response prediction network, respectively. Based on the first initial intervention response and the first initial control response, the first initial gain prediction model is trained to obtain the first gain prediction model.

3. The data processing method as described in claim 2, characterized in that, The sample object set includes a first object group, which includes the first sample object and the second sample object. The step of predicting the first initial intervention response of the first sample object and the first initial control response of the second sample object using the initial intervention response prediction network and the initial control response prediction network, respectively, includes: The initial intervention response prediction network is used to predict the first initial intervention response of the first sample object in the first object group. The initial control response prediction network is used to predict the first initial control response of the second sample object in the first object group.

4. The data processing method as described in claim 3, characterized in that, The sample object set also includes a second object group. The process of training the first initial gain prediction model based on the first initial intervention response and the first initial control response to obtain the first gain prediction model includes: Based on the first initial intervention response and the first initial control response, the first initial gain prediction model is updated to obtain an intermediate gain prediction model. The intermediate gain prediction model includes an intermediate intervention response prediction network updated by the initial intervention response prediction network and an intermediate control response prediction network updated by the initial control response prediction network. The intermediate intervention response prediction network and the intermediate control response prediction network are used to predict the intermediate intervention response and intermediate control response of the sample objects in the second object group, respectively. Based on the intermediate intervention response and the intermediate control response, determine the intermediate response gain of the intervention treatment on the sample objects in the second object group; Based on the intermediate response gain, the first gain prediction model is obtained from the intermediate gain prediction model.

5. The data processing method as described in claim 4, characterized in that, The step of obtaining the first gain prediction model based on the intermediate response gain and the intermediate gain prediction model includes: Based on the intermediate response gain, determine the performance parameters of the intermediate gain prediction model; When the performance parameters meet the preset performance conditions, the intermediate gain prediction model is trained using the sample object set to obtain the first gain prediction model.

6. The data processing method as described in claim 5, characterized in that, The step of obtaining the first gain prediction model based on the intermediate response gain and the intermediate gain prediction model further includes: When the performance parameters do not meet the preset performance conditions, the intermediate gain prediction model is used as the first initial gain prediction model, and the execution steps are returned to predict the first initial intervention response of the first sample object and the first initial control response of the second sample object through the initial intervention response prediction network and the initial control response prediction network, respectively, and subsequent steps, until the performance parameters meet the preset performance conditions.

7. The data processing method as described in claim 1, characterized in that, The second gain prediction model is obtained through the following steps: Obtain a sample object set and a second initial gain prediction model, the second initial gain prediction model including an initial mixed response prediction network, the sample object set including the first sample object and the second sample object; The initial mixed response prediction network is used to predict the second initial intervention response of the first sample object and the second initial control response of the second sample object. Based on the second initial intervention response and the second initial control response, the second initial gain prediction model is trained to obtain the second gain prediction model.

8. The data processing method as described in claim 1, characterized in that, The step of combining the first response gain and the second response gain to jointly train the first gain prediction model and the second gain prediction model to obtain the trained first gain prediction model and the trained second gain prediction model includes: A first object set and a second object set are determined from the sample objects to be predicted. The first response gain corresponding to the first object set satisfies the gain threshold condition, and the second response gain corresponding to the second object set satisfies the gain threshold condition. Based on the overlap between the first object set and the second object set, the first gain prediction model and the second gain prediction model are jointly trained to obtain the trained first gain prediction model and the trained second gain prediction model.

9. The data processing method as described in claim 8, characterized in that, The step of jointly training the first gain prediction model and the second gain prediction model based on the overlap between the first object set and the second object set to obtain the trained first gain prediction model and the trained second gain prediction model includes: When the overlap does not meet the preset overlap condition, the first gain prediction model and the second gain prediction model are updated to obtain the updated first gain prediction model and the updated second gain prediction model. The updated first gain prediction model and the updated second gain prediction model are used as the first gain prediction model and the second gain prediction model, respectively. The execution steps are returned to obtain the sample object, the first gain prediction model and the second gain prediction model and subsequent steps, until the overlap meets the preset overlap condition. The updated first gain prediction model and the updated second gain prediction model are used as the trained first gain prediction model and the trained second gain prediction model, respectively.

10. The data processing method as described in claim 1, characterized in that, The method further includes: The object information of the sample object to be predicted is preprocessed to obtain the processed information; The processed information is vectorized to obtain the object features of the sample object to be predicted. The object features are used to predict the intervention response and control response of the sample object to be predicted.

11. The data processing method according to any one of claims 1 to 10, characterized in that, The intervention process includes message notification. After predicting the predicted response gain of the object to be predicted by the intervention process using the trained first gain prediction model or the trained second gain prediction model, it further includes: When the predicted response gain meets the specified gain type, the message reminder is sent to the object to be predicted.

12. A data processing apparatus, characterized in that, include: The acquisition unit is used to acquire the sample object to be predicted, the first gain prediction model and the second gain prediction model. The first gain prediction model includes an intervention response prediction network trained by the first sample object and a control response prediction network trained by the second sample object. The second gain prediction model includes a hybrid response prediction network trained by the first sample object and the second sample object. The first sample object is a sample object that has undergone intervention treatment and the second sample object is a sample object that has not undergone the intervention treatment. The first prediction unit is configured to predict the first intervention response and the first control response of the sample object to be predicted through the intervention response prediction network and the control response prediction network, respectively, so as to determine the first response gain of the intervention treatment on the sample object to be predicted by the first intervention response and the first control response. The second prediction unit is used to predict the second intervention response and the second control response of the sample object to be predicted through the hybrid response prediction network, so as to determine the second response gain of the intervention treatment on the sample object to be predicted by the second intervention response and the second control response. The training unit is used to combine the first response gain and the second response gain to jointly train the first gain prediction model and the second gain prediction model to obtain the trained first gain prediction model and the trained second gain prediction model. The third prediction unit is used to predict the predicted response gain of the intervention treatment on the object to be predicted through the trained first gain prediction model or the trained second gain prediction model. The predicted response gain is used to determine the treatment strategy for the intervention treatment on the object to be predicted.

13. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing multiple instructions; the processor loads instructions from the memory to perform the steps of the data processing method as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the steps of the data processing method according to any one of claims 1 to 11.

15. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the steps of the data processing method according to any one of claims 1 to 11.

Images