Behavior prediction model training method and device, behavior prediction method and device and electronic equipment

Abstract

The invention discloses a behavior prediction model training method and device, a behavior prediction method and device and electronic equipment, and relates to the technical field of artificial intelligence. The training data set comprises sample object features of a plurality of sample objects corresponding to at least one business party and a preset behavior label of each sample object, and the sample object features are features of various object information; performing loop iteration training on the initial behavior prediction model based on the training data set to obtain a target behavior prediction model; and in a one-time loop iteration process, on the basis of performing undifferentiated depth feature extraction on the features of various object information in the sample object features of each current sample object, performing interactive joint learning on the features of the various object information in the sample object features of each current sample object. By using the technical scheme provided by the invention, the behavior prediction accuracy of the model can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a behavior prediction model training, a behavior prediction method, a device, and an electronic device. Background Art

[0002] With the development of artificial intelligence (AI), its application is becoming increasingly widespread. For example, in intelligent marketing scenarios such as ratings, activation after a loss, and policyholder conversions, AI is often incorporated to predict behavior, enabling more targeted services. Related technologies often directly predict user behavior, such as when visiting a store or making a purchase, by combining user-related features provided by multiple parties. This fails to effectively mine the corresponding user features, resulting in low prediction accuracy for the behavior prediction model and, consequently, inability to effectively provide relevant services. Summary of the Invention

[0003] The present application provides a behavior prediction model training, behavior prediction method, apparatus, equipment, storage medium and computer program product, which can effectively mine features associated with behavior prediction scenarios, improve the effectiveness of the mined features, and thereby improve the prediction accuracy of the behavior prediction model, and can provide users with corresponding services in a more targeted manner.

[0004] In one aspect, the present application provides a behavior prediction model training method, the method comprising: Obtaining a training data set corresponding to at least one business party, the training data set including sample object features of a plurality of sample objects corresponding to the at least one business party and a preset behavior label for each sample object, the preset behavior label being used to indicate a probability of each sample object having a preset behavior; the sample object features being features of a plurality of object information; The initial behavior prediction model is iteratively trained based on the training data set to obtain a target behavior prediction model; the initial behavior prediction model includes an adaptive learning layer to be trained and a behavior prediction layer to be trained corresponding to each business party, and in one iterative process, the following operations are performed: Acquire a current training data set from the training data set, where the multiple current sample objects corresponding to the current training data set include a sample object corresponding to each business party; Inputting the sample object features of each current sample object into the to-be-trained adaptive learning layer, so as to perform interactive joint learning on the features of the multiple object information in the sample object features of each current sample object based on indifferent deep feature extraction, to obtain a sample adaptive feature of each current sample object; Inputting the sample adaptive features corresponding to each business party into the to-be-trained behavior prediction layer corresponding to each business party to perform behavior prediction, thereby obtaining a predicted behavior label corresponding to each business party; Based on the predicted behavior label and the preset behavior label corresponding to each current sample object, the initial behavior prediction model is updated. On the other hand, a behavior prediction method is provided, comprising: Obtaining target object characteristics of a target object corresponding to a target business party; the target object characteristics are characteristics of multiple object information, and the target business party is any business party among at least one business party; The target object features are input into the target behavior prediction model trained according to any behavior prediction model training method provided in the embodiments of the present disclosure to perform behavior prediction processing to obtain a target behavior label corresponding to the target object, and the target behavior label is used to indicate the probability that the target object has a preset behavior.

[0005] Another aspect provides a behavior prediction model training device, the device comprising: a training data set acquisition module configured to acquire a training data set corresponding to at least one business party, the training data set including sample object features of a plurality of sample objects corresponding to the at least one business party and a preset behavior label of each sample object, the preset behavior label being used to indicate a probability of each sample object having a preset behavior; the sample object features being features of a plurality of object information; The cyclic iterative training module is configured to perform cyclic iterative training on the initial behavior prediction model based on the training data set to obtain a target behavior prediction model; the initial behavior prediction model includes an adaptive learning layer to be trained and a behavior prediction layer to be trained corresponding to each business party, and in one cyclic iteration process, a cyclic iterative operation is performed based on the following units in the cyclic iterative training module: a current training data set acquisition unit, configured to acquire a current training data set from the training data set, wherein the multiple current sample objects corresponding to the current training data set include a sample object corresponding to each business party; a first adaptive learning unit configured to input the sample object features of each current sample object into the to-be-trained adaptive learning layer, so as to perform interactive joint learning on the features of the multiple object information in the sample object features of each current sample object based on indifferent deep feature extraction, to obtain a sample adaptive feature of each current sample object; The first behavior prediction unit is configured to input the sample adaptive features corresponding to each business party into the to-be-trained behavior prediction layer corresponding to each business party to perform behavior prediction, and obtain a predicted behavior label corresponding to each business party; The model updating unit is configured to update the initial behavior prediction model based on the predicted behavior label and the preset behavior label corresponding to each current sample object.

[0006] In another aspect, a behavior prediction device is provided, comprising: A target object feature acquisition module is configured to acquire target object features of a target object corresponding to a target business party; the target object features are features of a plurality of object information, and the target business party is any one of the at least one business party; The second behavior prediction processing module is configured to execute behavior prediction processing by inputting the target object features into the target behavior prediction model trained according to any of the behavior prediction model training methods provided in the present disclosure, and obtain a target behavior label corresponding to the target object, and the target behavior label is used to indicate the probability that the target object has a preset behavior.

[0007] Another aspect provides an electronic device, comprising: a processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement any of the above-mentioned behavior prediction model training methods or behavior prediction methods.

[0008] On the other hand, a computer-readable storage medium is provided. When the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is capable of executing any of the above-mentioned behavior prediction model training methods or behavior prediction methods.

[0009] Another aspect provides a computer program product or computer program, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the behavior prediction model training method or behavior prediction method provided in the various optional implementations described above.

[0010] The behavior prediction model training, behavior prediction method, device, equipment, storage medium, and computer program product provided in this application have the following technical effects: In the process of training the behavior prediction model, the present application obtains a training data set corresponding to at least one business party, and the training data set includes sample object features of multiple sample objects corresponding to the at least one business party and a preset behavior label for each sample object, and the preset behavior label is used to indicate the probability of each sample object having a preset behavior; the sample object features are features of multiple object information; and in the process of iteratively training the initial behavior prediction model based on the training data set to obtain the target behavior prediction model, on the basis of indiscriminate deep feature extraction of features of multiple object information in the sample object features of each current sample object, interactive joint learning is performed on the features of multiple object information in the sample object features of each current sample object to obtain sample adaptive features of each current sample object, so as to achieve the learning of deep features of multiple object information while effectively mining features associated with behavior prediction scenarios, thereby improving the effectiveness of the mined features, and thereby improving the prediction accuracy of the behavior prediction model, and providing users with corresponding services in a more targeted manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0012] Figure 1 is a schematic diagram of an application environment provided by an embodiment of the present application; Figure 2 This is a flow chart of a behavior prediction model training method provided in an embodiment of the present application; Figure 3 A schematic diagram of a process for inputting the sample object features of each current sample object into a to-be-trained adaptive learning layer, based on indifferent deep feature extraction of features of multiple object information in the sample object features of each current sample object, and interactively and jointly learning the features of multiple object information in the sample object features of each current sample object to obtain a sample adaptive feature of each current sample object, provided by an embodiment of the present application; Figure 4 This is a schematic diagram of a feature fusion layer to be trained and behavior prediction layers to be trained corresponding to multiple business parties provided in an embodiment of the present application; Figure 5 This is a schematic diagram of a process for determining a feature fusion layer to be trained provided by an embodiment of the present application; Figure 6 This is a schematic diagram of another feature fusion layer to be trained and behavior prediction layers to be trained corresponding to multiple business parties provided in an embodiment of the present application; Figure 7 This is a schematic diagram of another feature fusion layer to be trained and behavior prediction layers to be trained corresponding to multiple business parties provided in an embodiment of the present application; Figure 8 is a schematic diagram of training an initial behavior prediction model according to an exemplary embodiment; Figure 9 This is a flow chart of a behavior prediction method provided in an embodiment of the present application; Figure 10 This is a structural diagram of a behavior prediction model training device provided in an embodiment of the present application; Figure 11 Schematic diagram of a behavior prediction device provided in an embodiment of the present application; Figure 12 This is a block diagram of an electronic device for behavior prediction model training or behavior prediction provided by an embodiment of the present application; Figure 13 This is a block diagram of another electronic device for behavior prediction model training or behavior prediction provided in an embodiment of the present application. DETAILED DESCRIPTION

[0013] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0014] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0015] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or portion of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal. It can be implemented in whole or in part using software, hardware (such as processing circuits or memory), or a combination thereof. Similarly, a 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 part of an overall module or unit that includes the functionality of the module or unit.

[0016] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, autonomous driving, drones, digital twins, virtual humans, robots, artificial intelligence generated content (AIGC), conversational interaction, smart medical care, smart customer service, game AI, 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.

[0017] The solutions provided in the embodiments of this application involve technologies such as deep learning and federated learning of artificial intelligence, and are specifically described through the following embodiments: See also Figure 1 , Figure 1 This is a schematic diagram of an application environment provided by an embodiment of the present application. Optionally, taking the combination of horizontal federated learning for behavior prediction model training and behavior prediction as an example, the application environment may include at least a server 100 and a client 200.

[0018] In an optional embodiment, the server 100 may be the first participant in horizontal federated learning and may provide features of first object information. Optionally, the first object information may be object information on the first participant's side, such as gender, age, geographic location, and user behavior information on the first participant. The client 200 may be the second participant in horizontal federated learning and may provide features of second object information. The second object information may be object information on the second participant's side. The second participant may be a demander of behavior prediction. Optionally, the second object information may include information such as the user's intended products, intended stores, historical intended products, and historical stores. Specifically, the second participant may correspond to at least one business party, and different business parties may be stores of different brands.

[0019] In practical applications, the behavior prediction model can be trained on the second participant. Accordingly, the first participant side (client 200) can send an object intersection request to the second participant side (server 100), and the object intersection request includes the first object identification information of the first sample object corresponding to each business party local to the client 200; accordingly, the server 100 performs object intersection processing based on the first object identification information and the second object identification information of the second sample object local to the server 100, and obtains the third object identification information corresponding to the target intersection object; specifically, the target intersection object can be multiple sample objects corresponding to at least one business party, that is, it includes the sample object corresponding to each business party while including the second sample object; further, the server 100 can send intersection result feedback information to the client 200, and the intersection result feedback information includes the third object identification information. Further, the server 100 can obtain the target intersection object on the server 100 based on the third object identification information. The first object information of the local client 200 is obtained by extracting the features of the first object information to obtain the first object feature. Then, the first object feature is verified to obtain a first verification result. If the first verification result indicates that the verification is passed, the first object feature is encrypted and sent to the client 200. Accordingly, the client 200 can decrypt the encrypted first object feature. In addition, the client 200 can also obtain the second object information of the target intersection object on the client 200 and the preset behavior label corresponding to the target intersection object based on the third object identification information, and extract the features of the second object information to obtain the second object feature. Then, the second object feature and the preset behavior label are verified respectively to obtain a second verification result corresponding to the second object feature and a third verification result corresponding to the preset behavior label. If the second verification result indicates that the verification is passed and the third verification result indicates that the verification is passed, the decrypted first object feature, second object feature, and preset behavior label are used as a training data set. Then, the initial behavior prediction model can be iteratively trained based on the training data set to obtain a target behavior prediction model. Further, after obtaining the target behavior prediction model, the client 200 can perform behavior prediction processing for any object in combination with the target behavior prediction model.

[0020] In a specific embodiment, the verification items for verifying the first object feature and the second object feature can be set in combination with the actual application. For example, time consistency verification (i.e., taking the features of object information in the same time period), encryption and decryption consistency verification (the client side and the service side need to share the encryption and decryption algorithms and keys), feature format verification (whether it conforms to the preset format requirements, whether there are null values, etc.), feature value range verification (whether it is within the preset threshold range), feature distribution verification (specifically, the positive sample ratio can be statistically calculated, that is, the ratio of the object features corresponding to the sample objects with the preset behavior, and the negative sample ratio, that is, the ratio of the object features corresponding to the sample objects without the preset behavior, and control the positive sample ratio and the negative sample ratio within the preset range), feature quality verification. Specifically, the feature quality verification can be combined with the following formula:

[0021] Where, is the index data reflecting the quality of the object feature, n is the number of value types of the feature, is the positive sample quantity of the i-th object feature value in the object feature (the first object feature or the second object feature), is the negative sample quantity of the i-th object feature value of the object feature, is the total positive sample quantity in the object feature, is the total negative sample quantity in the object feature.

[0022] Furthermore, for object features with FV less than or equal to a certain threshold, the verification passes. For object features with a threshold <FV <= 1, after relevant personnel evaluate and pass, the features can be optimized by methods such as balancing the positive and negative sample ratios, long-tail truncation, and screening features for modeling, and then it is determined that the verification passes. For object features with FV greater than 1, the verification fails. Specifically, the threshold can be set in combination with the actual application requirements, such as 0.5.

[0023] In a specific embodiment, the verification items for verifying the preset behavior label can be set in combination with the actual application. For example, label value range verification (i.e., for classification task labels, the value range should be discrete category values, and for regression task labels, they should be continuous numerical values), verification of whether the label value range conforms to the business logic (such as in the behavior prediction scenario, the value is 0 or 1), balance verification, checking whether the proportion of each category in the dataset is relatively balanced (i.e., statistically calculating whether the proportion of positive and negative samples conforms to the preset requirements), stability verification, that is, verifying whether the label is stable under different data integrations). Specifically, the ks (Kolmogorov-Smirnov test) test method can be combined for stability verification).

[0024] In addition, before the server 100 and the client 200 interact, a connectivity check can be performed to ensure that the client side and the server side can communicate and exchange data normally. Specifically, the check can be performed in combination with the delay and packet loss rate.

[0025] In addition, the method for obtaining the test data set can refer to the method for obtaining the training data set mentioned above, and will not be repeated here.

[0026] In a specific embodiment, the server 100 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0027] In an optional embodiment, the client 200 may include, but is not limited to, electronic devices such as smartphones, desktop computers, tablet computers, laptop computers, smart speakers, digital assistants, augmented reality (AR) / virtual reality (VR) devices, smart wearable devices, vehicle-mounted terminals, and smart televisions. It may also be software running on the above electronic devices, such as applications and applets. The operating systems running on the electronic devices in the embodiments of the present application may include, but are not limited to, Android, iOS, Linux, and Windows.

[0028] In addition, it should be noted that Figure 1 What is shown is merely an application environment of a behavior prediction model training method. The embodiments of this specification are not limited to the above. For example, the training data set can also come from the same party, etc. Correspondingly, the behavior prediction model can also be trained on the server side, and the server can provide background support for the client so that the client can provide behavior prediction services to users.

[0029] In the embodiments of this specification, the server 100 and the client 200 may be connected directly or indirectly via wired or wireless communication, which is not limited in this application.

[0030] In the above-mentioned federated learning process of this application, one of the two participants in the federated learning provides part of the training data, and the other party provides the other part of the training data, and model training is performed on the side with preset behavior labels. No intermediary party is required to participate, and the training can be initiated by the party that trains the model without the need for other participants to initiate it again, which can greatly improve the security, efficiency and ease of use of the federated learning process.

[0031] The following describes a behavior prediction model training method of this application. Figure 2 This is a flow chart of a behavior prediction model training method provided by an embodiment of the present application, specifically as follows Figure 2 As shown, the method may include: S201: Obtain a training data set corresponding to at least one business party.

[0032] In a specific embodiment, the above-mentioned training data set may include sample object features of multiple sample objects corresponding to at least one business party and preset behavior labels for each sample object. The sample objects may be users; specifically, the preset behavior labels are used to indicate the probability that each sample object has a preset behavior. Optionally, if the sample object has a preset behavior, the preset behavior label may be 1; conversely, if the sample object does not have a preset behavior, the preset behavior label may be 0; specifically, the preset behavior may be set in combination with the needs of the business party, such as the behavior of purchasing a product, the behavior of going to a store, etc.

[0033] In an optional embodiment, the above-mentioned sample object features may come from a feature provider. Optionally, the above-mentioned sample object features include the first object features on the first participant side of the horizontal federated learning and the second object features on the second participant side of the horizontal federated learning, thereby better enriching the object features.

[0034] In a specific embodiment, the above-mentioned sample object features are features of multiple object information. Specifically, the multiple object information may include basic information of the object, such as gender, age, education level, etc., and may also include behavioral operation information of the object, such as behavioral operation information associated with preset behaviors (such as browsing information of business-related information), user intended products, intended stores, historical intended products, historical store visits and other behavioral operation information. Specifically, for the specific details of the steps for obtaining the training data set, please refer to the above-mentioned relevant description and will not be repeated here.

[0035] S203: Performing cyclic iterative training on the initial behavior prediction model based on the training data set to obtain a target behavior prediction model.

[0036] In a specific embodiment, the initial behavior prediction model may be a behavior prediction model to be trained; the initial behavior prediction model includes an adaptive learning layer to be trained and a behavior prediction layer to be trained corresponding to each business party, and Figure 2 As shown, during one loop iteration, the following operations are performed: S2031: Acquire a current training data set from the training data set, where the multiple current sample objects corresponding to the current training data set include a sample object corresponding to each business party; S2033: Inputting the sample object features of each current sample object into the to-be-trained adaptive learning layer, so as to perform interactive joint learning on the features of the multiple object information in the sample object features of each current sample object based on indifferent deep feature extraction, thereby obtaining a sample adaptive feature of each current sample object; S2035: Inputting the sample adaptive features corresponding to each business party into the to-be-trained behavior prediction layer corresponding to each business party to perform behavior prediction, thereby obtaining a predicted behavior label corresponding to each business party; S2037: Update the initial behavior prediction model based on the predicted behavior label and the preset behavior label corresponding to each current sample object.

[0037] In a specific embodiment, the current training data set may be a training data set selected for the current training iteration round, or a partial data set in the training data set. Specifically, the current training data set may include sample object features of multiple current sample objects and preset behavior labels of multiple current sample objects.

[0038] In an optional embodiment, the above-mentioned adaptive learning layer to be trained may include an interactive learning layer to be trained, a feature extraction layer to be trained, and a feature fusion layer to be trained; accordingly, Figure 3 As shown, the sample object features of each current sample object are input into the adaptive learning layer to be trained, so as to perform interactive joint learning on the features of multiple object information in the sample object features of each current sample object on the basis of performing indifferent deep feature extraction on the features of multiple object information in the sample object features of each current sample object, and the sample adaptive features of each current sample object may include: S301: Inputting the sample object features of each current sample object into the interactive learning layer to be trained, interactively and jointly learning the features of multiple object information in the sample object features of each current sample object to obtain sample interactive features; S303: Inputting the sample object features of each current sample object into the feature extraction layer to be trained, performing indifferent deep feature extraction on the features of multiple object information in the sample object features of each current sample object, and obtaining the sample object deep features; S305: Input the sample interaction features and the sample object depth features into the feature fusion layer to be trained for feature fusion processing to obtain sample adaptive features.

[0039] In a specific embodiment, the interactive learning layer to be trained may be a network layer constructed based on a deep neural network. The specific network structure may be set in combination with actual applications. Specifically, the interactive learning layer to be trained may be used to interactively and jointly learn features of multiple object information in the sample object features of each sample object. Optionally, the interactive learning layer to be trained may be combined with the following formula for interactive joint learning:

[0040] in, 、 、 、 、 are the network parameters (model parameters) in the interactive learning layer to be trained, i, j, k are the object serial numbers of the current sample object corresponding to the sample object features; n is the total number of object features in the sample object features; is the i-th object feature (feature of object information) among the sample object features of a current sample object; is the j-th object feature (feature of object information) among the sample object features of a current sample object; is the kth object feature (feature of object information) among the sample object features of a current sample object; It can be a sample adaptive feature of a current sample object.

[0041] In a specific embodiment, combined with the constant term in the interactive learning layer to be trained , a conventional offset can be introduced to better fit the data and extract features. It can be combined with the first term in the interactive learning layer to be trained Mining the independent feature information of each object information corresponding to each sample object feature, such as the store's historical store visit rate; combined with a second-order interaction term in the interactive learning layer to be trained Mining first-order interaction information, such as combining the gender and age features of each sample object, can mine features that reflect the life status of the user (object); combining the second-order interaction terms in the interactive learning layer to be trained Mining the secondary second-order interaction information, combining the information browsing operation characteristics and information browsing trend characteristics (consultation browsing status in two adjacent periods of time) in each sample object feature can mine the user's recent information browsing interest changes; combining the third-order interaction items in the interactive learning layer to be trained Third-order interaction information can be mined. For example, by combining the city characteristics and educational background characteristics of each sample object, the positioning of the user's target product can be mined.

[0042] In a specific embodiment, the sample interaction feature of each current sample object is an object feature adapted to the behavior prediction scenario, which is learned (mined) by interactively and jointly learning features of multiple object information in the sample object feature of the current sample object.

[0043] In a specific embodiment, the feature extraction layer to be trained can be a network layer constructed based on a deep neural network. The specific network structure can be set in combination with the actual application. Specifically, the interactive learning layer to be trained can be used to perform indiscriminate deep feature extraction on features of multiple object information in the sample object features of each sample object. Optionally, the feature extraction layer to be trained can be a fully connected neural network composed of multiple layers of neurons. The sample object deep features of each current sample object are object features mined by performing indiscriminate deep feature extraction on features of multiple object information in the sample object features of the current sample object.

[0044] In an optional embodiment, when the total number of objects corresponding to the multiple sample objects is less than a first preset threshold, the feature fusion layer to be trained includes a first fusion layer to be trained; the sample interaction features and the sample object depth features are input into the feature fusion layer to be trained for feature fusion processing, and the sample adaptive features obtained include: The sample interaction features and the sample object depth features are input into the first fusion layer to be trained for feature fusion processing to obtain sample adaptive features.

[0045] In a specific embodiment, the first preset threshold can be set in combination with the actual application, such as 200,000; in actual applications, a small number of samples is prone to overfitting. Accordingly, when the total number of objects is less than the first preset threshold, in order to avoid overfitting, a first fusion layer to be trained with a relatively simple structure can be used as the feature fusion layer to be trained. Specifically, the first fusion layer to be trained can be a fully connected neural network composed of multiple layers of neurons.

[0046] In a specific embodiment, the first fusion layer to be trained is a feature fusion layer to be trained, and at least one business party is a plurality of business parties (assuming m), such as Figure 4 As shown, Figure 4 This is a schematic diagram of a feature fusion layer to be trained and a behavior prediction layer to be trained corresponding to multiple business parties provided in an embodiment of the present application; specifically, Figure 4 It can be seen that the sample interaction features and sample object deep features can be input into the feature fusion layer to be trained (the first fusion layer to be trained) for feature fusion processing to obtain sample adaptive features. Furthermore, the sample adaptive features of the current sample object corresponding to each business party can be input into the behavior prediction layer to be trained corresponding to the business party.

[0047] In the above embodiment, when the total number of objects corresponding to multiple sample objects during the training process is less than a first preset threshold, a fully connected neural network composed of multiple layers of neurons is selected as the feature fusion layer to be trained. This can effectively avoid overfitting by reducing the complexity of the network structure, thereby improving the effectiveness of the fused features for behavior prediction.

[0048] In an optional embodiment, when at least one business party is multiple business parties and the total number of objects corresponding to the multiple sample objects is greater than or equal to a first preset threshold, the above method may further include: determining a feature fusion layer to be trained, specifically, Figure 5 As shown, this may include: S501: Acquire a test data set corresponding to at least one business party, where the test data set includes sample object features of multiple test objects corresponding to the at least one business party and a preset behavior label for each test object; S503: Inputting the sample object features of the multiple test objects into the initial behavior prediction model corresponding to the current loop iteration round to perform behavior prediction processing to obtain predicted behavior labels corresponding to the multiple test objects; S505: Determine a first model performance indicator based on the preset behavior labels corresponding to the multiple test objects and the predicted behavior labels corresponding to the multiple test objects; S507: Obtaining a second model performance indicator; S509: Determine the difference between the second model performance index and the first model performance index; S511: Determine a feature fusion layer to be trained from the second fusion layer to be trained and the third fusion layer to be trained according to the difference.

[0049] In a specific embodiment, the sample object features of multiple test objects are input into the initial behavior prediction model corresponding to the current loop iteration round for behavior prediction processing to obtain the specific details of the predicted behavior labels corresponding to the multiple test objects. Please refer to the above-mentioned specific details of obtaining the predicted behavior labels corresponding to the sample objects, which will not be repeated here.

[0050] In a specific embodiment, the first model performance indicator may be a model performance indicator determined based on a test data set. Specifically, the model performance indicator may be an indicator that characterizes the performance of the model, such as AUC (Area Under Curve): used to calculate the number of times the predicted probability of a positive sample is greater than the predicted probability of a negative sample.

[0051] In a specific embodiment, the above-mentioned second model performance indicator is the largest model performance indicator among the model performance indicators corresponding to a preset number of loop iteration rounds. Specifically, the preset number of loop iteration rounds corresponds to the current loop iteration round and the preset number before the current loop iteration round minus one loop iteration round; specifically, the preset number can be set in combination with actual applications, for example 5, and the model performance indicator corresponding to each loop iteration round can be a model performance indicator determined based on the current training data set corresponding to the loop iteration round.

[0052] In practical applications, the second model performance metric for the training dataset is often greater than the first model performance metric for the test dataset. Accordingly, the difference can be calculated by subtracting the first model performance metric from the second model performance metric. Specifically, the larger the difference, the more severe the overfitting. Consequently, in cases of severe overfitting, the complexity of the model structure can be reduced.

[0053] In a specific embodiment, the second fusion layer to be trained may include a single gating layer, and the third fusion layer to be trained may include multiple gating layers, with the multiple gating layers corresponding to the multiple service parties. That is, the structure of the second fusion layer to be trained is simpler than that of the third fusion layer to be trained.

[0054] In an optional embodiment, determining the feature fusion layer to be trained from the second fusion layer to be trained and the third fusion layer to be trained based on the difference may include: When the difference is greater than or equal to the second preset threshold, the second fusion layer to be trained is used as the feature fusion layer to be trained; Accordingly, the sample interaction features and the sample object depth features are input into the feature fusion layer to be trained for feature fusion processing to obtain the sample adaptive features, which may include: The sample interaction features and sample object depth features are input into the second fusion layer to be trained for feature fusion processing to obtain sample adaptive features and shared feature weights.

[0055] In a specific embodiment, the second preset threshold can be set in combination with actual applications, such as 0.1, etc. The shared feature weight can represent the weight of the feature in the sample adaptive feature, that is, different business parties share the shared feature weight.

[0056] In an optional embodiment, the second fusion layer to be trained includes a single gating layer and a fourth fusion layer to be trained; inputting the sample interaction features and the sample object depth features into the second fusion layer to be trained for feature fusion processing to obtain the sample adaptive features and shared feature weights may include: Input the sample interaction features and the sample object depth features into the fourth fusion layer to be trained for feature fusion processing to obtain sample adaptive features; The sample interaction features and sample object depth features are input into a single gating layer for feature weight analysis to obtain shared feature weights.

[0057] In a specific embodiment, the single gating layer included in the second fusion layer to be trained can be used to control the weight of each feature in the sample adaptive features, thereby more specifically extracting features from the sample adaptive features for behavior prediction. Specifically, the fourth fusion layer to be trained can be a fully connected neural network composed of multiple layers of neurons. The specific network structure can be configured based on the actual application.

[0058] In a specific embodiment, the second fusion layer to be trained is the feature fusion layer to be trained, and the multiple business parties are m business parties, such as Figure 6 As shown, Figure 6 This is a schematic diagram of another feature fusion layer to be trained and a behavior prediction layer to be trained corresponding to multiple business parties provided in an embodiment of the present application; specifically, Figure 6 It can be seen that the sample interaction features and sample object depth features can be input into the fourth fusion layer to be trained for feature fusion processing to obtain sample adaptive features; and the sample interaction features and sample object depth features can be input into a single gating layer for feature weight analysis to obtain shared feature weights; further, while the sample adaptive features of the current sample object corresponding to each business party are input into the behavior prediction layer to be trained corresponding to the business party, the shared feature weights are input into the behavior prediction layer to be trained corresponding to the business party.

[0059] In an optional embodiment, determining the feature fusion layer to be trained from the second fusion layer to be trained and the third fusion layer to be trained based on the difference may include: When the difference is less than the second preset threshold, the third fusion layer to be trained is used as the feature fusion layer to be trained; Accordingly, the sample interaction features and the sample object depth features are input into the feature fusion layer to be trained for feature fusion processing to obtain the sample adaptive features, which may include: The sample interaction features and the sample object depth features are input into the third fusion layer to be trained for feature fusion processing to obtain the sample adaptive features and the feature weights corresponding to the multiple business parties.

[0060] In a specific embodiment, the feature weight corresponding to any of the above business parties may represent the weight of the feature in the sample adaptive feature corresponding to any of the business parties, that is, different business parties have different feature weights.

[0061] In an optional embodiment, the second fusion layer to be trained includes multiple gating layers and a fifth fusion layer to be trained; the sample interaction features and the sample object depth features are input into the third fusion layer to be trained for feature fusion processing, and the sample adaptive features and the feature weights corresponding to the multiple business parties are obtained, including: Input the sample interaction features and the sample object depth features into the fifth fusion layer to be trained for feature fusion processing to obtain sample adaptive features; The sample interaction features and sample object depth features are input into the gating layer corresponding to each business party for feature weight analysis to obtain the feature weight corresponding to each business party.

[0062] In a specific embodiment, each of the multiple gating layers can be used to control the weight of each feature in the sample adaptive features corresponding to the corresponding business entity, thereby more specifically extracting features from the sample adaptive features for behavior prediction for that business entity. Specifically, the fifth to-be-trained fusion layer can be a fully connected neural network composed of multiple layers of neurons. The specific network structure can be configured based on the actual application.

[0063] In a specific embodiment, the third fusion layer to be trained is the feature fusion layer to be trained, and the multiple business parties are m business parties, such as Figure 7 As shown, Figure 7 This is a schematic diagram of another feature fusion layer to be trained and a behavior prediction layer to be trained corresponding to multiple business parties provided in an embodiment of the present application; specifically, Figure 7 It can be seen that the sample interaction features and sample object depth features can be input into the fifth fusion layer to be trained for feature fusion processing to obtain sample adaptive features; and the sample interaction features and sample object depth features can be input into each gating layer for feature weight analysis to obtain the feature weight corresponding to each business party; further, while the sample adaptive features of the current sample object corresponding to each business party are input into the behavior prediction layer to be trained corresponding to the business party, the feature weight corresponding to each business party is input into the behavior prediction layer to be trained corresponding to the business party.

[0064] In the above embodiment, when the total number of objects corresponding to multiple sample objects during the training process is greater than or equal to the first preset threshold, the doorway layer is integrated into the feature fusion layer to be trained, so that features for behavior prediction can be extracted from the sample adaptive features in a more targeted manner. In order to effectively avoid overfitting, overfitting judgment can be made in combination with the model performance indicators in the training process and the model performance indicators in the testing process, thereby effectively alleviating overfitting and improving the effectiveness of the extracted features for behavior prediction.

[0065] In a specific embodiment, the network structure of the behavior prediction layer to be trained can be set in combination with actual applications, such as a sigmoid network layer. Specifically, the predicted behavior label corresponding to each current sample object can be the probability that the sample object predicted by the current initial behavior prediction model has a preset behavior; in a specific embodiment, based on the predicted behavior label and the preset behavior label corresponding to each current sample object, updating the initial behavior prediction model may include: substituting the predicted behavior label corresponding to each current sample object and the preset behavior label into a preset loss function to determine the behavior prediction loss; and combining the gradient descent method and the behavior prediction loss to update the model parameters of the initial behavior prediction model, and performing the next round of cyclic iteration operations based on the updated initial behavior prediction model until the preset convergence condition is met, and using the initial behavior prediction model corresponding to the preset convergence condition as the target behavior prediction model.

[0066] In a specific embodiment, the behavior prediction loss can represent the behavior prediction performance of the current initial behavior prediction model. Specifically, the preset loss function can be set in combination with the actual application, such as cross entropy loss, L1 regularization loss function, etc.

[0067] In a specific embodiment, the preset convergence conditions can be set in combination with actual applications, such as the number of executions of the loop iteration operation reaches a preset number, the behavior prediction loss is less than a specified threshold, etc., which can be set specifically in combination with the training speed and model accuracy requirements.

[0068] In a specific embodiment, Figure 8 As shown, Figure 8It is a schematic diagram of a training initial behavior prediction model provided according to an exemplary embodiment. Optionally, taking the training data set coming from two participants as an example, the sample object feature can control the first object feature and the second object feature, wherein the first object feature is the feature representation model on the server (one participant), obtained by performing feature representation on the first object information; the second object feature is the feature representation model on the client (another participant), obtained by performing feature representation on the second object information; specifically, the sample object features of each current sample object can be input into the interactive learning layer to be trained and the feature extraction layer to be trained in the adaptive learning layer to be trained, and the features of multiple object information in the sample object features of each current sample object are interactively and jointly learned in the interactive learning layer to be trained to obtain the sample interactive features. In the feature extraction layer to be trained, the sample object features of each current sample object are extracted and analyzed. The features of multiple object information are extracted indiscriminately to obtain sample object deep features; further, the sample interaction features and the sample object deep features can be input into the feature fusion layer to be trained for feature fusion processing to obtain sample adaptive features; further, the sample adaptive features corresponding to each business party are input into the behavior prediction layer to be trained corresponding to each business party to perform behavior prediction to obtain the predicted behavior label corresponding to each business party; further, based on the predicted behavior label and preset behavior label corresponding to the current sample object of each business party, the initial behavior prediction model is updated, and the next round of cyclic iteration operation is performed based on the updated initial behavior prediction model until the preset convergence condition is met, and the initial behavior prediction model corresponding to the preset convergence condition is used as the target behavior prediction model.

[0069] It can be seen from the technical solutions provided in the above embodiments of this specification that, in the process of training the behavior prediction model, this specification obtains a training data set corresponding to at least one business party, and the training data set includes sample object features of multiple sample objects corresponding to at least one business party and preset behavior labels for each sample object, and the preset behavior labels are used to indicate the probability that each sample object has a preset behavior; the sample object features are features of multiple object information; and in the process of iteratively training the initial behavior prediction model based on the training data set to obtain the target behavior prediction model, on the basis of indiscriminate deep feature extraction of the features of multiple object information in the sample object features of each current sample object, the features of multiple object information in the sample object features of each current sample object are interactively and jointly learned to obtain sample adaptive features of each current sample object, so as to achieve the learning of deep features of multiple object information while effectively mining features associated with behavior prediction scenarios, thereby improving the effectiveness of the mined features, and thereby improving the prediction accuracy of the behavior prediction model, and providing users with corresponding services in a more targeted manner.

[0070] Based on the above target behavior prediction model, the following describes a behavior prediction method of the present application, such as Figure 9 As shown, the method may include: S910: Obtain target object characteristics of the target object corresponding to the target business party.

[0071] In a specific embodiment, the target object characteristics may be characteristics of multiple types of object information, the target business party may be any one of the at least one business party, and the target object may be a customer (user) corresponding to the target business party. For details on the process of acquiring target object characteristics, please refer to the process of acquiring sample object characteristics, and will not be further elaborated here.

[0072] S920: Input the target object features into the target behavior prediction model to perform behavior prediction processing to obtain a target behavior label corresponding to the target object.

[0073] In a specific embodiment, the target behavior tag is used to indicate the probability of the target object engaging in a preset behavior. Furthermore, the target business entity can combine the target behavior tag to convert each user's probability of engaging in the preset behavior (intention) into a rating result, namely, a rating score for each user. This conversion process typically follows business needs, such as mapping a probability value of 0-1 to a score of 1-5 at a certain ratio, with 1 being the lowest intention and 5 being the highest. Furthermore, corresponding services can be provided in conjunction with the rating.

[0074] In a specific embodiment, the target behavior prediction model includes an adaptive learning layer and a behavior prediction layer corresponding to the target business party; optionally, the target behavior prediction model may also include a behavior prediction layer for any business party other than the target business party in at least one business party.

[0075] Furthermore, the target object features are input into the target behavior prediction model for behavior prediction processing to obtain the target behavior label corresponding to the target object, including: Inputting the target object features into the adaptive learning layer, so as to interactively and jointly learn the features of the multiple object information in the target object features on the basis of indiscriminate deep feature extraction of the features of the multiple object information in the target object features, and obtain the target adaptive features of the target object; The target adaptive features are input into the behavior prediction layer corresponding to the target business party to perform behavior prediction and obtain the target behavior label.

[0076] In a specific embodiment, the target object features are input into the adaptive learning layer, so as to perform interactive joint learning on the features of multiple object information in the target object features on the basis of indifferent deep feature extraction of the features of multiple object information in the target object features, so as to obtain the specific refinement of the target adaptive features of the target object. Please refer to the above-mentioned input of the sample object features of each current sample object into the adaptive learning layer to be trained, so as to perform interactive joint learning on the features of multiple object information in the sample object features of each current sample object on the basis of indifferent deep feature extraction of the features of multiple object information in the sample object features of each current sample object, so as to obtain the specific refinement of the sample adaptive features of each current sample object. No further details will be given here.

[0077] It can be seen from the technical solutions provided in the above embodiments of this specification that, in the process of behavior prediction, this specification obtains the target object features corresponding to the target object, which contain multiple object information features, and inputs the target object features into the target behavior prediction model. On the basis of indiscriminate deep feature extraction of the features of multiple object information in the target object features, the features of multiple object information in the target object features are interactively and jointly learned, so as to realize the learning of the deep features of multiple object information while effectively mining the features associated with the behavior prediction scenario, thereby improving the effectiveness of the mined features, and then improving the accuracy of behavior prediction, so as to provide users with corresponding services in a more targeted manner.

[0078] The present application also provides a behavior prediction model training device, such as Figure 10 As shown, the above device includes: The training data set acquisition module 1010 is configured to acquire a training data set corresponding to at least one business party, wherein the training data set includes sample object features of a plurality of sample objects corresponding to the at least one business party and a preset behavior label of each sample object, wherein the preset behavior label indicates a probability of each sample object having a preset behavior; the sample object features are features of a plurality of object information; The cyclic iterative training module 1020 is configured to perform cyclic iterative training on the initial behavior prediction model based on the training data set to obtain a target behavior prediction model; the initial behavior prediction model includes a to-be-trained adaptive learning layer and a to-be-trained behavior prediction layer corresponding to each business party, and in one cyclic iteration process, cyclic iteration operations are performed based on the following units in the cyclic iterative training module: a current training data set acquisition unit, configured to acquire a current training data set from the training data set, wherein the multiple current sample objects corresponding to the current training data set include a sample object corresponding to each business party; a first adaptive learning unit configured to input the sample object features of each current sample object into the to-be-trained adaptive learning layer, so as to perform interactive joint learning on the features of the multiple object information in the sample object features of each current sample object based on indifferent deep feature extraction, to obtain a sample adaptive feature of each current sample object; The first behavior prediction unit is configured to input the sample adaptive features corresponding to each business party into the to-be-trained behavior prediction layer corresponding to each business party to perform behavior prediction, and obtain a predicted behavior label corresponding to each business party; The model updating unit is configured to update the initial behavior prediction model based on the predicted behavior label and the preset behavior label corresponding to each current sample object.

[0079] In an optional embodiment, the adaptive learning layer to be trained includes an interactive learning layer to be trained, a feature extraction layer to be trained, and a feature fusion layer to be trained; the first adaptive learning unit 1140 includes: an interactive joint learning unit configured to input the sample object features of each current sample object into the interactive learning layer to be trained, and perform interactive joint learning on features of multiple object information in the sample object features of each current sample object to obtain a sample interactive feature; a depth feature extraction unit configured to input the sample object features of each current sample object into the feature extraction layer to be trained, perform indifferent depth feature extraction on features of multiple object information in the sample object features of each current sample object, and obtain a sample object depth feature; The feature fusion processing unit is configured to input the sample interaction feature and the sample object depth feature into the feature fusion layer to be trained to perform feature fusion processing to obtain the sample adaptive feature.

[0080] In an optional embodiment, when the total number of objects corresponding to the multiple sample objects is less than a first preset threshold, the feature fusion layer to be trained includes a first fusion layer to be trained; and the feature fusion processing unit includes: The first feature fusion processing subunit is configured to input the sample interaction feature and the sample object depth feature into the first fusion layer to be trained for feature fusion processing to obtain the sample adaptive feature.

[0081] In an optional embodiment, when the at least one business party is a plurality of business parties, and the total number of objects corresponding to the plurality of sample objects is greater than or equal to a first preset threshold, the apparatus further includes: a test data set acquisition module configured to acquire a test data set corresponding to the at least one business party, the test data set including sample object features of a plurality of test objects corresponding to the at least one business party and a preset behavior label for each test object; A first behavior prediction processing module is configured to input the sample object features of the multiple test objects into the initial behavior prediction model corresponding to the current loop iteration round to perform behavior prediction processing to obtain predicted behavior labels corresponding to the multiple test objects; a model performance indicator determination module, configured to determine a first model performance indicator based on preset behavior labels corresponding to the plurality of test objects and predicted behavior labels corresponding to the plurality of test objects; a model performance indicator acquisition module configured to execute acquisition of a second model performance indicator, where the second model performance indicator is a maximum model performance indicator among model performance indicators corresponding to a preset number of loop iterations, where the preset number of loop iterations corresponds to a current loop iteration and a preset number of loop iterations before the current loop iteration minus one; a performance indicator difference determination module, configured to determine a difference between the second model performance indicator and the first model performance indicator; The feature fusion layer determination module is configured to determine the feature fusion layer to be trained from the second fusion layer to be trained and the third fusion layer to be trained based on the difference; the second fusion layer to be trained is a fusion layer to be trained including a single gating layer, and the third fusion layer to be trained is a fusion layer to be trained including multiple gating layers, and the multiple gating layers correspond one to one to the multiple business parties.

[0082] In an optional embodiment, the feature fusion layer determination module includes: A first feature fusion layer determining unit is configured to, when the difference is greater than or equal to a second preset threshold, use the second to-be-trained fusion layer as the to-be-trained feature fusion layer; The feature fusion processing unit includes: The second feature fusion processing sub-unit is configured to perform feature fusion processing by inputting the sample interaction feature and the sample object depth feature into the second fusion layer to be trained to obtain the sample adaptive feature and the shared feature weight, wherein the shared feature weight represents the weight of the feature in the sample adaptive feature.

[0083] In an optional embodiment, the second fusion layer to be trained includes the single gating layer and the fourth fusion layer to be trained; and the second feature fusion processing subunit includes: A third feature fusion processing subunit is configured to input the sample interaction feature and the sample object depth feature into the fourth fusion layer to be trained to perform feature fusion processing to obtain the sample adaptive feature; The first feature weight analysis unit is configured to perform feature weight analysis on the sample interaction feature and the sample object depth feature by inputting them into the single gating layer to obtain the shared feature weight.

[0084] In an optional embodiment, the feature fusion layer determination module includes: A second feature fusion layer determining unit is configured to, when the difference is less than a second preset threshold, use the third to-be-trained fusion layer as the to-be-trained feature fusion layer; The feature fusion processing unit includes: The fourth feature fusion processing sub-unit is configured to execute feature fusion processing by inputting the sample interaction feature and the sample object depth feature into the third fusion layer to be trained, and obtain the sample adaptive feature and the feature weights corresponding to each of the multiple business parties. The feature weight corresponding to any business party represents the weight of the feature in the sample adaptive feature corresponding to any business party.

[0085] In an optional embodiment, the second to-be-trained fusion layer includes the multiple gating layers and the fifth to-be-trained fusion layer; and the fourth feature fusion processing subunit includes: A fifth feature fusion processing subunit is configured to input the sample interaction feature and the sample object depth feature into the fifth to-be-trained fusion layer for feature fusion processing to obtain the sample adaptive feature; The second feature weight analysis unit is configured to perform feature weight analysis by inputting the sample interaction feature and the sample object depth feature into the gating layer corresponding to each business party to obtain the feature weight corresponding to each business party.

[0086] In an optional embodiment, the sample object features include first object features on the first participant side of the horizontal federated learning and second object features on the second participant side of the horizontal federated learning.

[0087] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0088] The present application also provides a behavior prediction device, such as Figure 11 As shown, the above device includes: The target object feature acquisition module 1110 is configured to acquire target object features of a target object corresponding to a target business party; the target object features are features of a variety of object information, and the target business party is any one of the at least one business party; The second behavior prediction processing module 1120 is configured to execute behavior prediction processing by inputting the target object features into the target behavior prediction model trained by the behavior prediction model training method, and obtain the target behavior label corresponding to the target object, and the target behavior label is used to indicate the probability that the target object has a preset behavior.

[0089] In an optional embodiment, the target behavior prediction model includes an adaptive learning layer and a behavior prediction layer corresponding to the target business party; the second behavior prediction processing module 1120 includes: a second adaptive learning unit configured to input the target object feature into the adaptive learning layer, and perform interactive joint learning on the features of the multiple object information in the target object feature based on indifferent deep feature extraction of the features of the multiple object information in the target object feature, to obtain a target adaptive feature of the target object; The second behavior prediction unit is configured to input the target adaptive feature into the behavior prediction layer corresponding to the target business party to perform behavior prediction and obtain the target behavior label.

[0090] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0091] Figure 12 This is a block diagram of an electronic device for behavior prediction model training or behavior prediction provided by an embodiment of the present application. The electronic device may be a terminal, and its internal structure diagram may be as shown in FIG. Figure 12 As shown. The electronic device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the electronic device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a behavior prediction model training method or a behavior prediction method is implemented. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the electronic device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the electronic device, or an external keyboard, touchpad or mouse, etc. Figure 13 This is a block diagram of another electronic device for behavior prediction model training or behavior prediction provided by an embodiment of the present application. The electronic device may be a server, and its internal structure diagram may be as shown in FIG. Figure 13 As shown. The electronic device includes a processor, a memory and a network interface connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the electronic device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a behavior prediction model training method or a behavior prediction method is implemented. Those skilled in the art will understand that Figure 12 or Figure 13 The structure shown in the figure is only a block diagram of a part of the structure related to the scheme of the present disclosure, and does not constitute a limitation on the electronic device to which the scheme of the present disclosure is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement. In an exemplary embodiment, an electronic device is also provided, comprising: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement a behavior prediction model training method or a behavior prediction method as in the embodiments of the present disclosure.

[0092] In an exemplary embodiment, a computer-readable storage medium is also provided. When the instructions in the storage medium are executed by the processor of an electronic device, the electronic device can execute the behavior prediction model training method or behavior prediction method in the embodiment of the present disclosure. In an exemplary embodiment, a computer program product or computer program is also provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the behavior prediction model training method or behavior prediction method provided in the various optional implementations described above.

[0093] It is understandable that in the specific implementation of this application, user-related data is involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0094] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, which can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0095] Other embodiments of the present disclosure will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the claims.

[0096] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A behavior prediction model training method, characterized in that: The method comprises: Obtaining a training data set corresponding to at least one business party, the training data set including sample object features of a plurality of sample objects corresponding to the at least one business party and a preset behavior label for each sample object, the preset behavior label being used to indicate a probability of each sample object having a preset behavior; the sample object features being features of a plurality of object information; The initial behavior prediction model is iteratively trained based on the training data set to obtain a target behavior prediction model; the initial behavior prediction model includes an adaptive learning layer to be trained and a behavior prediction layer to be trained corresponding to each business party, and in one iterative process, the following operations are performed: Acquire a current training data set from the training data set, where the multiple current sample objects corresponding to the current training data set include a sample object corresponding to each business party; Inputting the sample object features of each current sample object into the to-be-trained adaptive learning layer, so as to perform interactive joint learning on the features of the multiple object information in the sample object features of each current sample object based on indifferent deep feature extraction, to obtain a sample adaptive feature of each current sample object; Inputting the sample adaptive features corresponding to each business party into the to-be-trained behavior prediction layer corresponding to each business party to perform behavior prediction, thereby obtaining a predicted behavior label corresponding to each business party; The initial behavior prediction model is updated based on the predicted behavior label and the preset behavior label corresponding to each current sample object.

2. The method according to claim 1, characterized in that The adaptive learning layer to be trained includes an interactive learning layer to be trained, a feature extraction layer to be trained, and a feature fusion layer to be trained; the sample object features of each current sample object are input into the adaptive learning layer to be trained, so as to perform interactive joint learning on the features of multiple object information in the sample object features of each current sample object on the basis of performing indifferent deep feature extraction on the features of multiple object information in the sample object features of each current sample object, and obtain the sample adaptive features of each current sample object, including: Inputting the sample object features of each current sample object into the interactive learning layer to be trained, and performing interactive joint learning on features of multiple object information in the sample object features of each current sample object to obtain sample interactive features; Inputting the sample object features of each current sample object into the feature extraction layer to be trained, performing indifferent depth feature extraction on features of multiple object information in the sample object features of each current sample object to obtain sample object depth features; The sample interaction feature and the sample object depth feature are input into the feature fusion layer to be trained for feature fusion processing to obtain the sample adaptive feature.

3. The method according to claim 2, characterized in that When the total number of objects corresponding to the multiple sample objects is less than a first preset threshold, the feature fusion layer to be trained includes a first fusion layer to be trained; and inputting the sample interaction feature and the sample object depth feature into the feature fusion layer to be trained for feature fusion processing to obtain the sample adaptive feature includes: The sample interaction feature and the sample object depth feature are input into the first fusion layer to be trained for feature fusion processing to obtain the sample adaptive feature.

4. The method according to claim 2, characterized in that In a case where the at least one business party is a plurality of business parties, and the total number of objects corresponding to the plurality of sample objects is greater than or equal to a first preset threshold, the method further includes: Acquire a test data set corresponding to the at least one business party, the test data set including sample object features of a plurality of test objects corresponding to the at least one business party and a preset behavior label for each test object; Inputting the sample object features of the multiple test objects into the initial behavior prediction model corresponding to the current loop iteration round to perform behavior prediction processing to obtain predicted behavior labels corresponding to the multiple test objects; determining a first model performance indicator according to the preset behavior labels corresponding to the multiple test objects and the predicted behavior labels corresponding to the multiple test objects; Obtaining a second model performance indicator, where the second model performance indicator is a maximum model performance indicator among model performance indicators corresponding to a preset number of loop iterations, where the preset number of loop iterations corresponds to a current loop iteration and a preset number of loop iterations before the current loop iteration minus one; determining a difference between the second model performance indicator and the first model performance indicator; According to the difference, the feature fusion layer to be trained is determined from the second fusion layer to be trained and the third fusion layer to be trained; the second fusion layer to be trained is a fusion layer to be trained including a single gating layer, and the third fusion layer to be trained is a fusion layer to be trained including multiple gating layers, and the multiple gating layers correspond one to one to the multiple business parties.

5. The method according to claim 4, characterized in that Determining the feature fusion layer to be trained from the second fusion layer to be trained and the third fusion layer to be trained according to the difference includes: When the difference is greater than or equal to a second preset threshold, the second fusion layer to be trained is used as the feature fusion layer to be trained; Inputting the sample interaction feature and the sample object depth feature into the feature fusion layer to be trained for feature fusion processing to obtain the sample adaptive feature includes: The sample interaction feature and the sample object depth feature are input into the second fusion layer to be trained for feature fusion processing to obtain the sample adaptive feature and shared feature weight, where the shared feature weight represents the weight of the feature in the sample adaptive feature.

6. The method according to claim 5, characterized in that The second fusion layer to be trained includes the single gating layer and the fourth fusion layer to be trained; the inputting the sample interaction feature and the sample object depth feature into the second fusion layer to be trained for feature fusion processing to obtain the sample adaptive feature and shared feature weight includes: Inputting the sample interaction feature and the sample object depth feature into the fourth fusion layer to be trained for feature fusion processing to obtain the sample adaptive feature; The sample interaction features and the sample object depth features are input into the single gating layer to perform feature weight analysis to obtain the shared feature weight.

7. The method according to claim 4, characterized in that Determining the feature fusion layer to be trained from the second fusion layer to be trained and the third fusion layer to be trained according to the difference includes: When the difference is less than a second preset threshold, using the third fusion layer to be trained as the feature fusion layer to be trained; Inputting the sample interaction feature and the sample object depth feature into the feature fusion layer to be trained for feature fusion processing to obtain the sample adaptive feature includes: The sample interaction features and the sample object depth features are input into the third fusion layer to be trained for feature fusion processing to obtain the sample adaptive features and the feature weights corresponding to each of the multiple business parties. The feature weight corresponding to any business party represents the weight of the features in the sample adaptive features corresponding to any business party.

8. The method according to claim 7, characterized in that The second fusion layer to be trained includes the multiple gating layers and the fifth fusion layer to be trained; the inputting the sample interaction feature and the sample object depth feature into the third fusion layer to be trained for feature fusion processing to obtain the sample adaptive feature and the feature weights corresponding to each of the multiple business parties includes: Inputting the sample interaction feature and the sample object depth feature into the fifth fusion layer to be trained for feature fusion processing to obtain the sample adaptive feature; The sample interaction features and the sample object depth features are input into the gating layer corresponding to each business party to perform feature weight analysis to obtain the feature weight corresponding to each business party.

9. The method according to any one of claims 1 to 8, characterized in that: The sample object features include a first object feature on a first participant side of horizontal federated learning and a second object feature on a second participant side of horizontal federated learning.

10. A behavior prediction method, characterized in that: The method comprises: Obtaining target object characteristics of a target object corresponding to a target business party; the target object characteristics are characteristics of multiple object information, and the target business party is any business party among at least one business party; The target object features are input into the target behavior prediction model trained according to the behavior prediction model training method according to any one of claims 1 to 9 for behavior prediction processing to obtain a target behavior label corresponding to the target object, and the target behavior label is used to indicate the probability that the target object has a preset behavior.

11. The method according to claim 10, characterized in that The target behavior prediction model includes an adaptive learning layer and a behavior prediction layer corresponding to the target business party; the inputting of the target object features into the target behavior prediction model trained by the behavior prediction model training method according to any one of claims 1 to 9 for behavior prediction processing to obtain a target behavior label corresponding to the target object includes: Inputting the target object features into the adaptive learning layer, so as to interactively and jointly learn the features of the multiple object information in the target object features based on indifferent deep feature extraction of the features of the multiple object information in the target object features, and obtain the target adaptive features of the target object; The target adaptive feature is input into the behavior prediction layer corresponding to the target business party to perform behavior prediction to obtain the target behavior label.

12. A behavior prediction model training device, characterized in that: The device comprises: a training data set acquisition module configured to acquire a training data set corresponding to at least one business party, the training data set including sample object features of a plurality of sample objects corresponding to the at least one business party and a preset behavior label of each sample object, the preset behavior label being used to indicate a probability of each sample object having a preset behavior; the sample object features being features of a plurality of object information; The cyclic iterative training module is configured to perform cyclic iterative training on the initial behavior prediction model based on the training data set to obtain a target behavior prediction model; the initial behavior prediction model includes an adaptive learning layer to be trained and a behavior prediction layer to be trained corresponding to each business party, and in one cyclic iteration process, a cyclic iterative operation is performed based on the following units in the cyclic iterative training module: a current training data set acquisition unit, configured to acquire a current training data set from the training data set, wherein the multiple current sample objects corresponding to the current training data set include a sample object corresponding to each business party; a first adaptive learning unit configured to input the sample object features of each current sample object into the to-be-trained adaptive learning layer, so as to perform interactive joint learning on the features of the multiple object information in the sample object features of each current sample object based on indifferent deep feature extraction, to obtain a sample adaptive feature of each current sample object; The first behavior prediction unit is configured to input the sample adaptive features corresponding to each business party into the to-be-trained behavior prediction layer corresponding to each business party to perform behavior prediction, and obtain a predicted behavior label corresponding to each business party; The model updating unit is configured to update the initial behavior prediction model based on the predicted behavior label and the preset behavior label corresponding to each current sample object.

13. A behavior prediction device, characterized in that: The device comprises: A target object feature acquisition module is configured to acquire target object features of a target object corresponding to a target business party; the target object features are features of a plurality of object information, and the target business party is any one of the at least one business party; The second behavior prediction processing module is configured to perform behavior prediction processing by inputting the target object features into the target behavior prediction model trained according to the behavior prediction model training method according to any one of claims 1 to 9 to obtain a target behavior label corresponding to the target object, and the target behavior label is used to indicate the probability that the target object has a preset behavior.

14. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the behavior prediction model training method according to any one of claims 1 to 9 or the behavior prediction method according to any one of claims 10 to 11.

15. A computer-readable storage medium, characterized in that When the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the behavior prediction model training method as described in any one of claims 1 to 8 or the behavior prediction method as described in any one of claims 10 to 11.

Images