Behavior prediction model training method and device, behavior prediction method and device, equipment, medium and product

Abstract

The embodiment of the invention discloses a behavior prediction model training method and device, a behavior prediction model prediction method and device, equipment, a medium and a product. According to one specific embodiment, the method comprises the steps that an object behavior information set is obtained, and object behavior information in the object behavior information set comprises a response tag, an intervention tag and an object feature information set; training a response label classification model and an intervention label classification model according to the object behavior information set; determining a screening causal feature set according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model; and training a response behavior prediction model according to the screening causal feature set and the object behavior information set. The embodiment is related to causal inference, existing real data can be utilized, assumed conditions are not needed, the complexity of an object response behavior prediction model and the risk of over-fitting are reduced, and the accuracy of causal inference is improved.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of computer technologies, and particularly to training of behavior prediction models, prediction methods, devices, equipment, media, and products. Background Art

[0002] Causal inference can provide more scientific and accurate data support for refined decision-making, enabling more precise decision-making and optimization. However, in causal inference, sometimes the causal relationship between two variables or multiple variables is relatively weak. Currently, to address the weak causal effect in causal inference, the commonly adopted methods are as follows: increasing the sample size or using counterfactual learning methods before modeling, adopting certain data augmentation, data expansion, or striving for the traffic of the experimental group and the control group, and lengthening the time window for observation to obtain more modeling data; or introducing stronger data processing methods during modeling and introducing semi-supervised learning methods to enhance the performance of the model using unlabeled data; or using advanced statistical models and introducing deconfounding means.

[0003] However, the inventors have found that when adopting the above methods, the following technical problems often exist: the enhanced or augmented data is not real, resulting in a large error in the causal inference result, while the acquisition of real control experiment data is difficult, and only increasing the sample size will lead to an increase in the complexity of modeling; introducing stronger data processing methods results in a higher model complexity and a greater risk of overfitting, and using semi-supervised or weak supervision leads to more cases of label noise and data imbalance, resulting in a large error in the causal inference result; when using advanced statistical models and deconfounding means, when there are many confounding factors and the influence of confounding factors is relatively complex, the accuracy of causal inference is poor, and the methods for dealing with confounding factors also need to rely on some assumptions, such as linear relationships and homogeneity, etc. When the assumptions do not hold, the accuracy of causal inference is poor.

[0004] The above information disclosed in this background art section is only used to enhance the understanding of the background of the inventive concept, and thus, it may include information that does not form the prior art known to ordinary skilled in the art in this country. Summary of the Invention

[0005] This content part of the present disclosure is used to briefly introduce the concepts, which will be described in detail in the following detailed implementation part. This content part of the present disclosure is not intended to identify the key features or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0006] Some embodiments of the present disclosure propose a response behavior prediction model training method, a response behavior prediction method, a device, an electronic device, a computer-readable medium, and a computer program product to solve one or more of the technical problems mentioned in the above background art section.

[0007] In a first aspect, some embodiments of the present disclosure provide a method for training a response behavior prediction model. The method includes: obtaining a set of object behavior information, where the object behavior information in the set of object behavior information includes a response label, an intervention label, and a set of object feature information; training a response label classification model and an intervention label classification model according to the set of object behavior information, where both the response label classification model and the intervention label classification model correspond to an object feature sequence; determining a screened causal feature set according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model; and training a response behavior prediction model according to the screened causal feature set and the set of object behavior information.

[0008] Optionally, before training the response label classification model and the intervention label classification model according to the set of object behavior information, the method further includes: dividing the set of object behavior information into a training set and a test set.

[0009] Optionally, training the response label classification model and the intervention label classification model according to the set of object behavior information includes: training the response label classification model according to the training set corresponding to the set of object behavior information and each response label included in the training set; and training the intervention label classification model according to the training set corresponding to the set of object behavior information and each intervention label included in the training set.

[0010] Optionally, determining the screened causal feature set according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model includes: determining each object feature that satisfies a preset important feature condition in the object feature sequence corresponding to the response label classification model as a first object feature set; determining each object feature that satisfies a preset important feature condition in the object feature sequence corresponding to the intervention label classification model as a second object feature set; and determining the intersection of the first object feature set and the second object feature set as the screened causal feature set.

[0011] Optionally, training the response behavior prediction model according to the screened causal feature set and the set of object behavior information includes: adding intervention features to the screened causal feature set to update the screened causal feature set; generating a training sample set according to the training set corresponding to the set of object behavior information and the updated screened causal feature set, where each training sample in the training sample set includes a response label and each object feature information corresponding to the screened causal feature set, and the object feature information corresponding to the intervention feature is an intervention label; and performing the following training steps based on the training sample set:

[0012] Input each training sample in at least one training sample in the training sample set into the embedding layer included in the initial response behavior prediction model to obtain word vectors corresponding to each training sample in the at least one training sample for the above-mentioned screened causal feature set; generate prediction response labels corresponding to each training sample in the at least one training sample according to the word vectors corresponding to each training sample in the at least one training sample and the initial response behavior prediction model; compare the prediction response labels corresponding to each training sample in the at least one training sample with the response labels to obtain a comparison result; in response to determining that the comparison result meets the optimization condition, determine the initial response behavior prediction model as the trained response behavior prediction model.

[0013] Optionally, the above training steps further include: generating a first loss value according to the prediction response labels and response labels corresponding to each training sample in the at least one training sample; generating a second loss value according to the word vectors corresponding to each training sample in the at least one training sample for the above-mentioned screened causal feature set; generating a model loss value according to the first loss value and the second loss value; in response to determining that the comparison result does not meet the optimization condition, adjust the network parameters of the initial response behavior prediction model according to the model loss value, form a training sample set with the unused training samples, use the adjusted initial response behavior prediction model as the initial response behavior prediction model, and execute the above training steps again.

[0014] Optionally, the generating, according to the word vectors corresponding to each training sample in the at least one training sample and the initial response behavior prediction model, prediction response labels corresponding to each training sample in the at least one training sample includes: inputting the word vectors corresponding to each training sample in the obtained at least one training sample for the above-mentioned screened causal feature set into the first feature interaction layer included in the initial response behavior prediction model to obtain concatenated vectors corresponding to each training sample in the at least one training sample; inputting the concatenated word vectors corresponding to each training sample in the obtained at least one training sample into the first prediction layer included in the initial response behavior prediction model to obtain prediction response labels corresponding to each training sample in the at least one training sample.

[0015] Optionally, each screened causal feature in the screened causal feature set corresponds to a feature domain; and the generating of the second loss value according to the word vectors corresponding to each training sample in at least one training sample for each screened causal feature in the screened causal feature set includes: for each training sample in the at least one training sample, performing the following steps: for each feature domain corresponding to the screened causal feature set, generating a distance in the domain of two word vectors corresponding to the two screened causal features corresponding to the feature domain in the screened causal feature set for the training sample; determining the sum of the generated distances in the domain as the total distance in the sample domain; for each feature domain corresponding to the screened causal feature set, performing the following steps according to each screened causal feature corresponding to the feature domain in the screened causal feature set: determining the word vectors corresponding to the screened causal feature set for the training sample as a feature matrix; removing the word vectors corresponding to the feature domain from the feature matrix to obtain a feature matrix after removal; determining the similarity between the word vector corresponding to the screened causal feature and each word vector in the feature matrix after removal; determining the sum of the determined similarities as the distance between sample domains; determining the sum of the determined distances between sample domains as the total distance between sample domains; determining the determined total distances in the sample domain as the total distance in the domain; determining the determined total distances between sample domains as the total distance between domains; determining the sum of the total distance in the domain and the total distance between domains as the second loss value.

[0016] Optionally, the generating of the model loss value according to the first loss value and the second loss value includes: for each training sample in the at least one training sample, performing the following steps: performing a masking process on the feature matrix composed of the word vectors corresponding to the screened causal feature set for the training sample to obtain a first masked feature matrix and a second masked feature matrix; inputting the first masked feature matrix into a second feature interaction layer included in the initial response behavior prediction model to obtain a first concatenated vector; inputting the second masked feature matrix into the second feature interaction layer to obtain a second concatenated vector; inputting the first concatenated vector into a mapping layer included in the initial response behavior prediction model to obtain a first hidden layer vector; inputting the second concatenated vector into the mapping layer to obtain a second hidden layer vector; determining the distance between the first hidden layer vector and the second hidden layer vector as a contrast distance; determining the sum of the obtained contrast distances as the total contrast distance; determining the sum of the first loss value, the second loss value, and the total contrast distance as the model loss value.

[0017] Optionally, the method further includes: generating model evaluation information of the response behavior prediction model according to the test set.

[0018] Second aspect, some embodiments of the present disclosure provide a response behavior prediction model training device, which includes: an object behavior information set acquisition unit configured to acquire an object behavior information set, where the object behavior information in the object behavior information set includes a response label, an intervention label, and an object feature information set; a first training unit configured to train a response label classification model and an intervention label classification model according to the object behavior information set, where both the response label classification model and the intervention label classification model correspond to an object feature sequence; a determination unit configured to determine a screened causal feature set according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model; a second training unit configured to train a response behavior prediction model according to the screened causal feature set and the object behavior information set.

[0019] Optionally, before the first training unit, the prediction model training device further includes: a division unit configured to divide the object behavior information set into a training set and a test set.

[0020] Optionally, the first training unit is further configured to: train a response label classification model according to the training set corresponding to the object behavior information set and each response label included in the training set; train an intervention label classification model according to the training set corresponding to the object behavior information set and each intervention label included in the training set.

[0021] Optionally, the determination unit is further configured to: determine each object feature that satisfies a preset important feature condition in the object feature sequence corresponding to the response label classification model as a first object feature set; determine each object feature that satisfies a preset important feature condition in the object feature sequence corresponding to the intervention label classification model as a second object feature set; determine the intersection of the first object feature set and the second object feature set as the screened causal feature set.

[0022] Optionally, the second training unit is further configured to: add intervention features to the above-mentioned screened causal feature set to update the screened causal feature set; generate a training sample set according to the training set corresponding to the above-mentioned object behavior information set and the updated screened causal feature set, where each training sample in the above-mentioned training sample set includes a response label and each object feature information corresponding to the screened causal feature set, and the object feature information corresponding to the intervention feature is an intervention label; based on the training sample set, perform the following training steps: input each training sample in at least one training sample in the training sample set into the embedding layer included in the initial response behavior prediction model to obtain each word vector corresponding to the screened causal feature set for each training sample in at least one training sample; generate a predicted response label corresponding to each training sample in at least one training sample according to each word vector corresponding to each training sample in at least one training sample and the initial response behavior prediction model; compare the predicted response labels and response labels corresponding to each training sample in at least one training sample to obtain a comparison result; in response to determining that the comparison result meets the optimization condition, determine the initial response behavior prediction model as the trained response behavior prediction model.

[0023] Optionally, the above-mentioned training steps further include: generating a first loss value according to the predicted response labels and response labels corresponding to each training sample in at least one training sample; generating a second loss value according to each word vector corresponding to the screened causal feature set for each training sample in at least one training sample; generating a model loss value according to the first loss value and the second loss value; in response to determining that the comparison result does not meet the optimization condition, adjust the network parameters of the initial response behavior prediction model according to the model loss value, form a training sample set with the unused training samples, use the adjusted initial response behavior prediction model as the initial response behavior prediction model, and perform the above-mentioned training steps again.

[0024] Optionally, the second training unit is further configured to: input each word vector corresponding to the screened causal feature set for each training sample in at least one training sample obtained into the first feature interaction layer included in the initial response behavior prediction model to obtain a concatenated vector corresponding to each training sample in at least one training sample; input the concatenated word vector corresponding to each training sample in at least one training sample obtained into the first prediction layer included in the initial response behavior prediction model to obtain a predicted response label corresponding to each training sample in at least one training sample.

[0025] Optionally, the screened causal features in the screened causal feature set correspond to feature domains.

[0026] Optionally, the second training unit is further configured to, for each of the at least one training sample, perform the following steps: for each feature domain corresponding to the above-mentioned screened causal feature set, according to every two screened causal features corresponding to the above-mentioned feature domain in the above-mentioned screened causal feature set, generate the in-domain distance between the two word vectors of the above-mentioned training sample corresponding to the two screened causal features; determine the sum of the generated in-domain distances as the total in-domain distance of the sample; for each feature domain corresponding to the above-mentioned screened causal feature set, according to each screened causal feature corresponding to the above-mentioned feature domain in the above-mentioned screened causal feature set, perform the following steps: determine the respective word vectors of the above-mentioned training sample corresponding to the above-mentioned screened causal feature set as a feature matrix; remove the respective word vectors corresponding to the above-mentioned feature domain from the above-mentioned feature matrix to obtain a post-removal feature matrix; determine the similarity between the word vector corresponding to the above-mentioned screened causal feature and each word vector in the above-mentioned post-removal feature matrix; determine the sum of the determined similarities as the inter-domain distance of the sample; determine the sum of the determined inter-domain distances of the samples as the total inter-domain distance; determine the determined total in-domain distances of the samples as the total in-domain distance; determine the determined total inter-domain distances of the samples as the inter-domain distance; determine the sum of the above-mentioned total in-domain distance and the above-mentioned inter-domain distance as the second loss value.

[0027] Optionally, the second training unit is further configured to, for each of the at least one training sample, perform the following steps: perform a masking process on the feature matrix composed of the respective word vectors of the above-mentioned training sample corresponding to the above-mentioned screened causal feature set to obtain a first masked feature matrix and a second masked feature matrix; input the above-mentioned first masked feature matrix into the second feature interaction layer included in the initial response behavior prediction model to obtain a first concatenated vector; input the above-mentioned second masked feature matrix into the second feature interaction layer to obtain a second concatenated vector; input the above-mentioned first concatenated vector into the mapping layer included in the initial response behavior prediction model to obtain a first hidden layer vector; input the above-mentioned second concatenated vector into the mapping layer to obtain a second hidden layer vector; determine the distance between the above-mentioned first hidden layer vector and the above-mentioned second hidden layer vector as the contrast distance; determine the sum of the obtained contrast distances as the total contrast distance; determine the sum of the above-mentioned first loss value, the above-mentioned second loss value, and the above-mentioned total contrast distance as the model loss value.

[0028] Optionally, the prediction model training device further includes: a generation unit configured to generate model evaluation information of the above-mentioned response behavior prediction model according to the above-mentioned test set.

[0029] In a third aspect, some embodiments of the present disclosure provide a response behavior prediction method, which includes: for each target object in the target object set, obtaining the item transfer behavior information of the target item corresponding to the above-mentioned target object; inputting each obtained item transfer behavior information into a pre-trained item transfer behavior prediction model to obtain an item transfer prediction behavior information set corresponding to the above-mentioned target object set, where the above-mentioned item transfer behavior prediction model is a response behavior prediction model and is trained by the method described in any implementation manner of the above first aspect.

[0030] Optionally, the method further includes: obtaining the to-be-intervened information of the above-mentioned target item; for each target object in the above-mentioned target object set, in response to determining that the item transfer prediction behavior information corresponding to the above-mentioned target object meets the preset intervention condition, sending the above-mentioned to-be-intervened information to the terminal device corresponding to the above-mentioned target object; in response to determining that the number of the sent to-be-intervened information meets the preset quantity condition, controlling the associated item scheduling device to perform a replenishment operation on the above-mentioned target item.

[0031] In a fourth aspect, some embodiments of the present disclosure provide a response behavior prediction device, which includes: an item transfer behavior information acquisition unit configured to, for each target object in the target object set, obtain the item transfer behavior information of the target item corresponding to the above-mentioned target object; an input unit configured to input each obtained item transfer behavior information into a pre-trained item transfer behavior prediction model to obtain an item transfer prediction behavior information set corresponding to the above-mentioned target object set, where the above-mentioned item transfer behavior prediction model is a response behavior prediction model and is trained by the method described in any implementation manner of the above first aspect.

[0032] Optionally, the response behavior prediction device further includes: a to-be-intervened information acquisition unit, a sending unit, and a control unit. Among them, the to-be-intervened information acquisition unit is configured to obtain the to-be-intervened information of the above-mentioned target item. The sending unit is configured to, for each target object in the above-mentioned target object set, in response to determining that the item transfer prediction behavior information corresponding to the above-mentioned target object meets the preset intervention condition, send the above-mentioned to-be-intervened information to the terminal device corresponding to the above-mentioned target object. The control unit is configured to, in response to determining that the number of the sent to-be-intervened information meets the preset quantity condition, control the associated item scheduling device to perform a replenishment operation on the above-mentioned target item.

[0033] In a fifth aspect, some embodiments of the present disclosure provide an electronic device, which includes: one or more processors; a storage device on which one or more programs are stored, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation manner of the above first aspect or the third aspect.

[0034] In a sixth aspect, some embodiments of the present disclosure provide a computer-readable medium having a computer program stored thereon, wherein when the program is executed by a processor, the method described in any implementation manner of the above first aspect or third aspect is implemented.

[0035] In a seventh aspect, some embodiments of the present disclosure provide a computer program product including a computer program, and when the computer program is executed by a processor, the method described in any implementation manner of the above first aspect or third aspect is implemented.

[0036] The above-mentioned various embodiments of the present disclosure have the following beneficial effects: Through the response behavior prediction model training method of some embodiments of the present disclosure, existing real data can be utilized without assuming conditions, reducing the model complexity and the risk of overfitting, and improving the accuracy of causal inference. Specifically, the reasons for the relatively high model complexity, the relatively high risk of overfitting, and the relatively poor accuracy of causal inference are as follows: The enhanced or augmented data does not actually exist, resulting in a relatively large error in the causal inference result. Moreover, it is difficult to collect real control experiment data, and simply increasing the sample size will only increase the complexity of modeling; Introducing a stronger data processing method results in a relatively high model complexity and a relatively high risk of overfitting. Moreover, using semi-supervised or weakly supervised methods leads to more cases of label noise and data imbalance, bringing a relatively large error to the causal inference result; When using advanced statistical models and deconfounding means, when there are many confounding factors and the influence of confounding factors is relatively complex, the accuracy of causal inference is relatively poor, and the methods for dealing with confounding factors also need to rely on some assumptions, such as linear relationships and homogeneity. When the assumptions do not hold, the accuracy of causal inference is relatively poor. Based on this, for the response behavior prediction model training method of some embodiments of the present disclosure, first, an object behavior information set is obtained. Among them, the object behavior information in the above object behavior information set includes a response label, an intervention label, and an object feature information set. Thus, the obtained object behavior information set can be used as a real data set for training the response behavior prediction model. Then, according to the above object behavior information set, a response label classification model and an intervention label classification model are trained. Among them, the above response label classification model and the above intervention label classification model both correspond to an object feature sequence. Thus, the classification models of the response label and the intervention label can be trained using the response label and the intervention label respectively, and the object feature sequence can represent each feature adopted by the classification model. Next, according to the object feature sequence corresponding to the above response label classification model and the object feature sequence corresponding to the above intervention label classification model, a screened causal feature set is determined. Thus, the screened causal feature set can represent the important object features screened according to the object feature sequences corresponding to the response label classification model and the intervention label classification model, so that the range of subsequent representation learning to strengthen feature representation can be pre-selected. Finally, according to the above screened causal feature set and the above object behavior information set, a response behavior prediction model is trained. Thus, the screened causal feature set can make subsequent representation learning pay more attention to the screened feature expression during the modeling process, reducing the risk of overfitting caused by network computational complexity. Also, because existing real data sets are used instead of enhanced or augmented samples, the error brought to the causal inference result and the complexity of modeling are reduced, and there is no need to extend the collection time to collect real control experiment data. Also, because the object behavior information set contains response labels, there is no need to use semi-supervised or weakly supervised methods, reducing the cases of label noise and data imbalance caused by semi-supervised or weakly supervised methods.Moreover, since there is no need to obfuscate, assumptions can be dispensed with, and situations where there are many confounding factors and the influence of confounding factors is relatively complex do not need to be considered, thus improving the accuracy of causal inference. Therefore, existing real data can be utilized without making assumptions, reducing the model complexity and the risk of overfitting, and improving the accuracy of causal inference. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In conjunction with the accompanying drawings and with reference to the following detailed description, the above and other features, advantages, and aspects of the various embodiments of the present disclosure will become more apparent. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic and that elements and elements are not necessarily drawn to scale.

[0038] Figure 1 FIG. is a schematic diagram of an application scenario of a method for training a response behavior prediction model according to some embodiments of the present disclosure.

[0039] Figure 2 FIG. is a flowchart of some embodiments of a method for training a response behavior prediction model according to the present disclosure;

[0040] Figure 3 FIG. is a schematic structural diagram of a response behavior prediction model of some embodiments of a method for training a response behavior prediction model according to the present disclosure;

[0041] Figure 4 FIG. is a flowchart of some embodiments of a response behavior prediction method according to the present disclosure;

[0042] Figure 5 FIG. is a schematic structural diagram of some embodiments of a response behavior prediction model training apparatus according to the present disclosure;

[0043] Figure 6 FIG. is a schematic structural diagram of some embodiments of a response behavior prediction apparatus according to the present disclosure;

[0044] Figure 7 FIG. is a schematic structural diagram of an electronic device suitable for implementing some embodiments of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. On the contrary, these embodiments are provided to more thoroughly and completely understand the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0046] In addition, it should be noted that for the sake of convenience of description, only parts related to the relevant invention are shown in the drawings. Without conflict, the embodiments in the present disclosure and the features in the embodiments can be combined with each other.

[0047] It should be noted that the concepts such as "first" and "second" mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order of functions executed by these devices, modules or units or their interdependent relationships.

[0048] It should be noted that the modifications of "one" and "multiple" mentioned in the present disclosure are illustrative rather than restrictive. Those skilled in the art should understand that unless clearly specified otherwise in the context, it should be understood as "one or more".

[0049] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are only for illustrative purposes and are not used to limit the scope of these messages or information.

[0050] Regarding the operations of collecting, storing, using, etc. of the user's personal information (such as object behavior information) involved in the present disclosure, before performing the corresponding operations, relevant organizations or individuals shall fulfill obligations including conducting personal information security impact assessments, fulfilling the obligation of notification to the personal information subject, and obtaining the prior authorization and consent of the personal information subject.

[0051] The present disclosure will be described in detail below with reference to the drawings and in conjunction with the embodiments.

[0052] Figure 1 It is a schematic diagram of an application scenario of a response behavior prediction model training method according to some embodiments of the present disclosure.

[0053] In Figure 1In the application scenario, first, the computing device 101 can obtain the object behavior information set 102. Among them, the object behavior information in the object behavior information set 102 includes a response label, an intervention label, and a set of object feature information. Then, the computing device 101 can train a response label classification model 103 and an intervention label classification model 104 according to the object behavior information set 102. Among them, both the response label classification model 103 and the intervention label classification model 104 correspond to an object feature sequence. The response label classification model 103 corresponds to the object feature sequence 105. The intervention label classification model 104 corresponds to the object feature sequence 106. Next, the computing device 101 can determine a screened causal feature set 107 according to the object feature sequence 105 corresponding to the response label classification model 103 and the object feature sequence 106 corresponding to the intervention label classification model 104. Finally, the computing device 101 can train a response behavior prediction model 108 according to the screened causal feature set 107 and the object behavior information set 102.

[0054] It should be noted that the computing device 101 can be hardware or software. When the computing device is hardware, it can be implemented as a distributed cluster composed of multiple servers or terminal devices, or as a single server or a single terminal device. When the computing device is embodied as software, it can be installed in the above-listed hardware devices. It can be implemented as, for example, multiple software or software modules for providing distributed services, or as a single software or software module. No specific limitation is made here.

[0055] It should be understood that Figure 1 the number of computing devices in

[0056] Continuing to refer to Figure 2 , a process 200 of some embodiments of the response behavior prediction model training method according to the present disclosure is shown. The response behavior prediction model training method includes the following steps:

[0057] Step 201, obtain an object behavior information set.

[0058] In some embodiments, the execution subject of the response behavior prediction model training method (such as Figure 1The computing device 101 shown can obtain an object behavior information set from a database through a wired connection or a wireless connection. Among them, the above object behavior information set can be the object behavior information of each object. The object can be a user. The object behavior information can be information related to object attributes and object behaviors. The object behavior information in the above object behavior information set can include a response tag, an intervention tag, and an object feature information set. The above response tag can characterize the type of response made by the object to the intervention matter. The intervention tag can characterize the intervention matter (for example, in online post-loan operations, the intervention variable can be what kind of tool is used to remind the user to repay the loan (phone, IVR robot, text message, letter, etc.), or whether it is changed from IVR robot collection to manual collection on the Xth day, and the response variable can be whether the number of overdue days of this user will exceed Y days since the overdue record is generated. Another example is in the advertising recommendation scenario, the intervention variable can be whether to push an advertisement for the user's item, and the response variable can be whether this user will purchase this item). The object feature information set can be the respective attribute and / or behavior feature information of the object (for example, in the online post-loan operation scenario, the object feature information set can include but is not limited to age, gender, number of overdue days, and overdue amount. Another example is in the advertising recommendation scenario, the object feature information set can include but is not limited to age, gender, preferred item type, activity level, and number of item views).

[0059] It should be noted that the above wireless connection method can include but is not limited to 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other currently known or future-developed wireless connection methods.

[0060] Optionally, the above execution entity can divide the above object behavior information set into a training set and a test set. In practice, the above execution entity can divide the above object behavior information set into a training set and a test set according to a preset division ratio. For example, the preset division ratio can be 8:2.

[0061] Step 202: Train a response tag classification model and an intervention tag classification model according to the object behavior information set.

[0062] In some embodiments, the above-mentioned execution entity may train a response label classification model and an intervention label classification model according to the above-mentioned object behavior information set. Among them, both the above-mentioned response label classification model and the above-mentioned intervention label classification model correspond to an object feature sequence. Each object feature in the object feature sequence may be arranged in descending order of importance. In practice, first, the above-mentioned execution entity may use the response labels included in each object behavior information in the object behavior information set as sample labels, and use the intervention labels and the object feature information set included in each object behavior information in the object behavior information set as input features, and train a response label classification model according to the above-mentioned object behavior information set. For example, the response label classification model may be LightGBM. The model type of the response label classification model is not limited. Then, the object features used for judgment in the classification process of the response label classification model may be sequentially combined to obtain an object feature sequence corresponding to the above-mentioned response label classification model.

[0063] Secondly, the above-mentioned execution entity may use the intervention labels included in each object behavior information in the object behavior information set as sample labels, and use the response labels and the object feature information set included in each object behavior information in the object behavior information set as input features, and train an intervention label classification model according to the above-mentioned object behavior information set. For example, the intervention label classification model may be LightGBM. The model type of the intervention label classification model is not limited. Then, the object features used for judgment in the classification process of the intervention label classification model may be sequentially combined to obtain an object feature sequence corresponding to the above-mentioned intervention label classification model.

[0064] In some optional implementation manners of some embodiments, the above-mentioned execution entity may train a response label classification model and an intervention label classification model according to the above-mentioned object behavior information set through the following steps:

[0065] In the first step, a response label classification model is trained according to the training set corresponding to the above-mentioned object behavior information set and each response label included in the training set. In practice, the above-mentioned execution entity may use the response labels included in each object behavior information in the training set as sample labels, and use the intervention labels and the object feature information set included in each object behavior information in the training set as input features, and train a response label classification model according to the above-mentioned training set. Then, the object features used for judgment in the classification process of the response label classification model may be sequentially combined to obtain an object feature sequence corresponding to the above-mentioned response label classification model.

[0066] Step 2: Train an intervention label classification model based on the training set corresponding to the above object behavior information set and each intervention label included in the above training set. In practice, the above execution entity can use the intervention labels included in each object behavior information in the training set as sample labels, and use the response labels and the object feature information set included in each object behavior information in the training set as input features, and train an intervention label classification model according to the above training set. Then, each object feature used for judgment in the classification process of the intervention label classification model can be combined in sequence to obtain an object feature sequence corresponding to the above intervention label classification model. Thus, a response label classification model and an intervention label classification model can be trained respectively according to the training set.

[0067] Step 203: Determine a screened causal feature set according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model.

[0068] In some embodiments, the above execution entity can determine a screened causal feature set according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model. In practice, the above execution entity can determine the intersection of the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model as the screened causal feature set.

[0069] In some optional implementation manners of some embodiments, the above execution entity can determine a screened causal feature set according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model through the following steps.

[0070] First step: Determine each object feature in the object feature sequence corresponding to the above response label classification model that satisfies a preset important feature condition as a first object feature set. Wherein, the above preset important feature condition can be that the object feature is one of the top N object features.

[0071] Second step: Determine each object feature in the object feature sequence corresponding to the above intervention label classification model that satisfies a preset important feature condition as a second object feature set.

[0072] Third step: Determine the intersection of the above first object feature set and the above second object feature set as the screened causal feature set. Thus, relatively important object features can be further screened.

[0073] Step 204: Train a response behavior prediction model according to the screened causal feature set and the object behavior information set.

[0074] In some embodiments, the execution subject may train the response behavior prediction model based on the screened causal feature set and the object behavior information set. In practice, the execution subject may use each response label included in the object behavior information set as each sample label, and each intervention label included in the object behavior information set and each object behavior feature information corresponding to the screened causal feature set as input features to train the initial neural network model, and obtain the trained initial neural network model as the response behavior prediction model. For example, the initial neural network model may be a DCN (Deep & Cross Network) or a transformer. Here, there is no limitation on the type of the initial neural network model.

[0075] In some optional implementations of some embodiments, the execution subject may train the response behavior prediction model according to the above-mentioned screening causal feature set and the above-mentioned object behavior information set through the following steps:

[0076] In the first step, the intervention feature is added to the above-mentioned screening causal feature set to update the screening causal feature set. The intervention feature may be a feature corresponding to the intervention label.

[0077] The second step is to generate a training sample set based on the training set corresponding to the above object behavior information set and the updated screening causal feature set. Each training sample in the above training sample set includes a response label and each object feature information of the corresponding screening causal feature set. The object feature information corresponding to the intervention feature is the intervention label. In practice, for each object behavior information in the training set, the above execution entity can remove each object feature information in the above object behavior information except for each object feature information of the corresponding updated screening causal feature set to obtain a training sample. In this way, a training sample set can be obtained.

[0078] The third step is to perform the following training steps based on the training sample set:

[0079] In the first sub-step, each training sample of at least one training sample in the training sample set is input into the embedding layer included in the initial response behavior prediction model to obtain each word vector corresponding to the above-mentioned screening causal feature set for each training sample of at least one training sample. Wherein, when the training step is executed for the first time, the initial response behavior prediction model can be an initial neural network model that has not been adjusted. When the training step is executed subsequently, the initial response behavior prediction model can be the initial response behavior prediction model adjusted when the training step was executed last time. The above-mentioned embedding layer can be used to generate a word vector (Embedding) of the object feature information. One screening causal feature can correspond to one word vector.

[0080] The second sub-step is to generate a predicted response label corresponding to each training sample in at least one training sample according to each word vector corresponding to each training sample in at least one training sample and the initial response behavior prediction model. In practice, the above-mentioned execution entity can input the above-mentioned word vectors into the model structure layer after the embedding layer in the initial response behavior prediction model to obtain the predicted response label corresponding to the above-mentioned training sample. Here, there is no limitation on the model structure layer after the embedding layer in the initial response behavior prediction model.

[0081] The third sub-step is to compare the predicted response labels and response labels corresponding to each training sample in at least one training sample to obtain a comparison result. In practice, the above-mentioned execution entity can determine whether the predicted response labels and response labels corresponding to each training sample are the same, and use the accuracy of each predicted response label corresponding to the above-mentioned training sample as the comparison result.

[0082] The fourth sub-step is to determine the initial response behavior prediction model as the trained response behavior prediction model in response to determining that the comparison result meets the optimization condition. Among them, the above-mentioned optimization condition can be that the accuracy is greater than a preset value.

[0083] Optionally, the above training steps may further include:

[0084] The fifth sub-step is to generate a first loss value according to the predicted response labels and response labels corresponding to each training sample in at least one training sample. In practice, for each training sample in at least one training sample, the above-mentioned execution entity can generate a sample loss value based on the predicted response label and response label corresponding to the above-mentioned training sample. Here, the sample loss value can be determined by a cross-entropy loss function. Then, the sum of the generated sample loss values can be determined as the first loss value.

[0085] The sixth sub-step is to generate a second loss value according to each word vector corresponding to the above-mentioned screened causal feature set for each training sample in at least one training sample. Thus, the second loss value can be used as a loss value considering the distance between features.

[0086] The seventh sub-step is to generate a model loss value according to the first loss value and the second loss value.

[0087] The eighth sub-step is to, in response to determining that the comparison result does not meet the optimization condition, adjust the network parameters of the initial response behavior prediction model according to the model loss value, form a training sample set with the unused training samples, use the adjusted initial response behavior prediction model as the initial response behavior prediction model, and execute the above training steps again. In practice, the above-mentioned execution entity can use the gradient descent method to adjust the network parameters of the initial response behavior prediction model according to the model loss value. Thus, iterative training of the initial response behavior prediction model can be realized.

[0088] In some alternative implementations of some embodiments, the above-mentioned execution subject may generate a predicted response label corresponding to each training sample in at least one training sample according to each word vector corresponding to each training sample in at least one training sample and an initial response behavior prediction model through the following steps:

[0089] First, input each word vector corresponding to each training sample in the obtained at least one training sample to the first feature interaction layer included in the initial response behavior prediction model to obtain a concatenated vector corresponding to each training sample in at least one training sample. For example, the above-mentioned first feature interaction layer may be an encoder of DCN or a transformer.

[0090] Second, input the concatenated word vector corresponding to each training sample in the obtained at least one training sample to the first prediction layer included in the initial response behavior prediction model to obtain a predicted response label corresponding to each training sample in at least one training sample. For example, the above-mentioned prediction layer may be an output layer or a decoder.

[0091] Optionally, the screened causal features in the screened causal feature set correspond to feature domains. The feature domain may be the category to which the object feature belongs. One feature domain may include at least one object feature (for example, in the post-loan operation scenario, the feature domain "overdue information" may include screened causal features such as overdue days and overdue amount). Here, the feature domains of object features may be pre-divided.

[0092] In some alternative implementations of some embodiments, the above-mentioned execution subject may generate a second loss value according to each word vector corresponding to each training sample in at least one training sample through the following steps, including:

[0093] First, for each training sample in the above-mentioned at least one training sample, perform the following steps:

[0094] The first sub-step: For each feature domain corresponding to the above-mentioned screened causal feature set, generate the in-domain distance between two word vectors corresponding to the above-mentioned two screened causal features of the above-mentioned training sample. In practice, the above-mentioned execution subject may determine the L2 distance between the above-mentioned two word vectors as the in-domain distance. Thus, the in-domain distance between two word vectors of a training sample may represent the distance between any two features in the same feature domain.

[0095] The second sub-step is to determine the sum of the distances in each generated domain as the total distance in the sample domain. Thus, the sum of the distances between any two word vectors in each group under each feature domain in a training sample can be determined. Consequently, the total distance in the sample domain can be used as the feature alignment loss of a training sample. The feature alignment constraint aims to minimize the distances between features under the same feature domain, enabling the features within the same feature domain to be more closely distributed in the low-dimensional space.

[0096] The third sub-step is to perform the following steps for each feature domain corresponding to the above-mentioned screened causal feature set, according to each screened causal feature corresponding to the above-mentioned feature domain in the above-mentioned screened causal feature set:

[0097] First, determine the word vectors of the above-mentioned training sample corresponding to the above-mentioned screened causal feature set as the feature matrix. In practice, the above-mentioned execution entity can use the above-mentioned word vectors as the row vectors to form the feature matrix. That is, each row vector in the feature matrix corresponds to a screened causal feature.

[0098] Then, remove the word vectors corresponding to the above-mentioned feature domain from the above-mentioned feature matrix to obtain the feature matrix after removal.

[0099] Next, determine the similarity between the word vectors corresponding to the above-mentioned screened causal feature and each word vector in the above-mentioned feature matrix after removal. Here, the similarity can be the cosine similarity.

[0100] After that, determine the sum of the determined similarities as the distance between sample domains.

[0101] The fourth sub-step is to determine the sum of the determined distances between sample domains as the total distance between sample domains. Thus, the total distance between sample domains can be used as the feature uniformity loss of a training sample. The uniformity constraint aims to maximize the distances between features between different feature domains, enabling the features between different feature domains to be as far apart as possible in the low-dimensional space.

[0102] The second step is to determine the total distance in each sample domain as the total distance in the domain.

[0103] The third step is to determine the total distance between each sample domain as the total distance between domains.

[0104] The fourth step is to determine the sum of the above-mentioned total distance in the domain and the above-mentioned total distance between domains as the second loss value. Thus, by introducing the feature alignment constraint and the uniformity constraint, the quality and generalization ability of the feature representation in the model can be effectively optimized, and the accuracy of the causal inference task can be improved.

[0105] In some optional implementation manners of some embodiments, the above-mentioned execution entity can generate the model loss value according to the first loss value and the second loss value through the following steps:

[0106] In the first step, for each of the at least one training sample described above, perform the following steps:

[0107] In the first sub-step, perform masking on the feature matrix composed of the word vectors corresponding to the above training sample for each of the above screening causal feature sets to obtain a first masked feature matrix and a second masked feature matrix. In practice, the above execution entity can randomly select two masking methods to perform masking on the feature matrix composed of the word vectors corresponding to the above training sample for each of the above screening causal feature sets to obtain a first masked feature matrix and a second masked feature matrix. The masking methods can include, but are not limited to: random mask method, dimension mask method, feature mask method. The random mask method can be multiplying a randomly initialized matrix on the basis of the feature matrix. The randomly initialized matrix has the same number of rows and columns as the feature matrix. All elements in the randomly initialized matrix are randomly assigned negative infinity, and the rest of the elements are assigned 1, aiming to randomly mask all elements in the feature matrix. The dimension mask method can be similar to the random mask method, where the matrix in the randomly initialized matrix is transformed into all elements in randomly selected several columns being assigned negative infinity, and the rest of the elements are assigned 1. The purpose is to mask the elements in randomly selected several columns of the feature matrix. The dimension mask method is a vertical mask. While the feature mask method is a horizontal mask, where all elements in randomly selected several rows in the randomly initialized matrix are assigned negative infinity, and the rest of the elements are assigned 1, aiming to mask the elements in randomly selected several rows of the feature matrix.

[0108] In the second sub-step, input the above first masked feature matrix into the second feature interaction layer included in the initial response behavior prediction model to obtain a first concatenated vector. The type of the second feature interaction layer can be the same as that of the first feature interaction layer.

[0109] In the third sub-step, input the above second masked feature matrix into the second feature interaction layer to obtain a second concatenated vector.

[0110] In the fourth sub-step, input the above first concatenated vector into the mapping layer included in the initial response behavior prediction model to obtain a first hidden layer vector. The mapping layer can be used to compress the dimension of the vector. Thus, the training process can be made more stable.

[0111] In the fifth sub-step, input the above second concatenated vector into the mapping layer to obtain a second hidden layer vector.

[0112] Sixth sub-step, determine the distance between the above first hidden layer vector and the above second hidden layer vector as the contrast distance. In practice, the above execution entity may determine the L2 distance between the above first hidden layer vector and the above second hidden layer vector as the contrast distance. Thus, the contrast distance can be used as the contrast loss of a training sample. The core idea of contrastive learning is to make the neural network learn that the distances of positive sample pairs are as close as possible.

[0113] Second step, determine the sum of the obtained contrast distances as the total contrast distance.

[0114] Third step, determine the sum of the above first loss value, the above second loss value and the above total contrast distance as the model loss value. Thus, the second loss value and the total contrast distance can be used as two additional losses introduced by representation learning, which can strengthen the screening causal feature representations in the screening causal feature set, thereby achieving the effect of enhancing the feature representation of weak causal features in the causal inference model.

[0115] As an example, the model structure of the response behavior prediction model can refer to Figure 3 . Figure 3 In, the response behavior prediction model includes an embedding layer, a first feature interaction layer, a second feature interaction layer, a prediction layer and a mapping layer. The loss of the response behavior prediction model includes feature alignment & feature uniformity loss (i.e., the second loss value), cross-entropy loss (i.e., the first loss value) and contrast loss (i.e., the total contrast distance). E represents the feature matrix. Feature alignment & feature uniformity loss represents the loss after introducing feature alignment constraints and uniformity constraints. E1 represents the first masked feature matrix. E2 represents the second masked feature matrix. h represents the concatenated vector. h1 represents the first concatenated vector. h2 represents the second concatenated vector. Mask represents masking processing. Figure 3 The two Masks in represent two masking processing methods. For example, the two Masks can be any two of the following, but not limited to: random mask method, dimension mask method, feature mask method.

[0116] Optionally, the above execution entity can also generate model evaluation information of the above response behavior prediction model according to the above test set. In practice, the above execution entity can input the test set into the response behavior prediction model to obtain each predicted response label. Then, based on each predicted response label and each response label included in the test set, model evaluation information can be generated. The model evaluation information can include at least one of the following, but not limited to: uplift gain (gain value), AUUC (Area Under the Uplift Curve). Thus, the response behavior prediction model can be evaluated through the model evaluation information.

[0117] The above-mentioned various embodiments of the present disclosure have the following beneficial effects: Through the response behavior prediction model training method of some embodiments of the present disclosure, existing real data can be utilized without assuming conditions, reducing the model complexity and the risk of overfitting, and improving the accuracy of causal inference. Specifically, the reasons for the relatively high model complexity, the relatively high risk of overfitting, and the relatively poor accuracy of causal inference are as follows: The enhanced or augmented data does not actually exist, resulting in a large error in the causal inference result. Moreover, it is difficult to collect real control experiment data, and simply increasing the sample size will only increase the complexity of model building. Introducing a stronger data processing method results in a relatively high model complexity and a relatively high risk of overfitting, and using semi-supervised or weakly supervised methods leads to more cases of label noise and data imbalance, bringing a large error to the causal inference result. When using advanced statistical models and deconfounding means, when there are many confounding factors and the influence of confounding factors is relatively complex, the accuracy of causal inference is relatively poor, and the methods for dealing with confounding factors also need to rely on some assumptions, such as linear relationships and homogeneity. When the assumptions do not hold, the accuracy of causal inference is relatively poor. Based on this, for the response behavior prediction model training method of some embodiments of the present disclosure, first, an object behavior information set is obtained. Among them, the object behavior information in the above object behavior information set includes a response label, an intervention label, and an object feature information set. Thus, the obtained object behavior information set can be used as a real data set for training the response behavior prediction model. Then, according to the above object behavior information set, a response label classification model and an intervention label classification model are trained. Among them, the above response label classification model and the above intervention label classification model both correspond to an object feature sequence. Thus, the classification models of the response label and the intervention label can be trained using the response label and the intervention label respectively, and the object feature sequence can represent each feature adopted by the classification model. Next, according to the object feature sequence corresponding to the above response label classification model and the object feature sequence corresponding to the above intervention label classification model, a screened causal feature set is determined. Thus, the screened causal feature set can represent the important object features screened according to the object feature sequences corresponding to the response label classification model and the intervention label classification model, so that the range of subsequent representation learning to strengthen the feature representation can be pre-selected. Finally, according to the above screened causal feature set and the above object behavior information set, a response behavior prediction model is trained. Thus, the screened causal feature set can make subsequent representation learning pay more attention to the screened feature expression in the modeling process, reducing the risk of overfitting caused by network computational complexity. Also, because instead of using enhanced or augmented samples, the existing real data set is used, the error brought to the causal inference result and the complexity of model building are reduced, and there is no need to extend the acquisition time to collect real control experiment data. Also, because the response label is included in the object behavior information set, there is no need to use semi-supervised or weakly supervised methods, reducing the cases of label noise and data imbalance caused by semi-supervised or weakly supervised methods.Moreover, since there is no need to deconfound, no assumptions need to be made, and situations where there are many confounding factors and the influence of confounding factors is relatively complex do not need to be considered, thus improving the accuracy of causal inference. Therefore, existing real data can be utilized without making assumptions, reducing the model complexity and the risk of overfitting, and improving the accuracy of causal inference.

[0118] Further reference is made to Figure 4 , which shows the flow 400 of some embodiments of the response behavior prediction method. The flow 400 of the response behavior prediction method includes the following steps:

[0119] Step 401, for each target object in the target object set, obtain the item transfer behavior information of the target object corresponding to the target item.

[0120] In some embodiments, the execution entity (such as a computing device) of the response behavior prediction method can, for each target object in the target object set, obtain the item transfer behavior information of the above-mentioned target object corresponding to the target item from the database. Here, the above-mentioned execution entity can be the same as the execution entity of the response behavior prediction model training method, or different from the execution entity of the response behavior prediction model training method. The above-mentioned target object set can be a set of users whose response behaviors are to be predicted. Here, the predicted response behavior can be the transfer behavior of the target item (for example, the behavior of whether to purchase the target item). The target item can be any item. The item transfer behavior information can be information related to the historical operations of the above-mentioned target object for the target item and object attribute-related information. The above-mentioned item transfer behavior information can include, but is not limited to: object information, the number of times the detail page of the above-mentioned target item is viewed, the historical transfer times, the preferred item type, and intervention information. The intervention information can represent the recommendation information (such as an advertisement) of the above-mentioned target item pushed to the target object.

[0121] Step 402, input each piece of item transfer behavior information obtained into a pre-trained item transfer behavior prediction model to obtain an item transfer prediction behavior information set corresponding to the target object set.

[0122] In some embodiments, the above-mentioned execution entity can input each piece of item transfer behavior information obtained into a pre-trained item transfer behavior prediction model to obtain an item transfer prediction behavior information set corresponding to the above-mentioned target object set. Among them, the above-mentioned item transfer behavior prediction model is a response behavior prediction model, which is obtained through Figure 2It is obtained by training the steps in the corresponding embodiments. It can be understood that when training the item transfer behavior prediction model, the training data used is the item transfer behavior information set. Each item transfer prediction behavior information in the item transfer prediction behavior information set can represent the item transfer behavior of the predicted target object for the target item. For example, the item transfer prediction behavior information can be "User 001, 1". 1 can represent that after pushing the intervention information of the target item to User 001, User 001 will transfer the target item.

[0123] Optionally, the above-mentioned execution entity can also perform the following steps:

[0124] First step, obtain the intervention information to be processed for the above-mentioned target item. The intervention information to be processed can be the intervention information of the above-mentioned target item to be pushed to the user.

[0125] Second step, for each target object in the above-mentioned target object set, in response to determining that the item transfer prediction behavior information corresponding to the target object meets the preset intervention condition, send the above-mentioned intervention information to the terminal device corresponding to the target object. Among them, the above-mentioned preset intervention condition can be that the item transfer prediction behavior information represents that the target object will transfer the target item. The terminal device of the target object can be the device logged in with the account of the target object.

[0126] Third step, in response to determining that the number of the sent intervention information meets the preset quantity condition, control the associated item scheduling device to perform a replenishment operation for the above-mentioned target item. The preset quantity condition can be that the number of the sent intervention information is greater than the preset quantity. Here, the specific setting of the preset quantity is not limited. In practice, the above-mentioned execution entity can control the associated item scheduling device to schedule a predetermined quantity of the above-mentioned target item to the warehouse corresponding to the above-mentioned target item. Thus, according to the predicted transfer behavior of the item, replenishment can be automatically carried out in advance.

[0127] The above-mentioned various embodiments of the present disclosure have the following beneficial effects: Through the response behavior prediction method of some embodiments of the present disclosure, the accuracy of predicting the item transfer behavior is improved. Specifically, the reasons for the poor accuracy of predicting the item transfer behavior are as follows: When training the item transfer behavior prediction model, the enhanced or augmented data does not actually exist, resulting in a large error in the causal inference result. Moreover, it is difficult to collect real control experiment data, and only increasing the sample size will increase the complexity of modeling; Introducing a stronger data processing method results in a higher model complexity, a greater risk of overfitting, and using semi-supervised or weak supervision, resulting in more label noise and data imbalance situations, bringing a large error to the causal inference result; When using advanced statistical models and deconfounding means, when there are many confounding factors and the influence of confounding factors is relatively complex, the accuracy of causal inference is poor, and the method of dealing with confounding factors also needs to rely on some assumptions, such as linear relationship and homogeneity, etc. When the assumptions do not hold, the accuracy of causal inference is poor, that is, the accuracy of predicting the item transfer behavior is poor. Based on this, in the response behavior prediction method of some embodiments of the present disclosure, first, for each target object in the target object set, obtain the item transfer behavior information of the target item corresponding to the above target object. Then, input each obtained item transfer behavior information into a pre-trained item transfer behavior prediction model to obtain an item transfer prediction behavior information set corresponding to the above target object set. Among them, the above item transfer behavior prediction model is a response behavior prediction model, which is obtained through Figure 2 the steps in the corresponding embodiments. Also because the item transfer behavior prediction model is obtained through Figure 2 the steps in the corresponding embodiments, the accuracy of causal inference is improved, that is, the accuracy of predicting the item transfer behavior is improved.

[0128] Further referring to Figure 5 , as an implementation of the methods shown in the above figures, the present disclosure provides some embodiments of a response behavior prediction model training device. These device embodiments correspond to Figure 2 the method embodiments shown, and the device can be specifically applied to various electronic devices.

[0129] As Figure 5As shown, the response behavior prediction model training device 500 of some embodiments includes: an object behavior information set acquisition unit 501, a first training unit 502, a determination unit 503, and a second training unit 504. Among them, the object behavior information set acquisition unit 501 is configured to acquire an object behavior information set, where the object behavior information in the object behavior information set includes a response label, an intervention label, and an object feature information set; the first training unit 502 is configured to train a response label classification model and an intervention label classification model according to the object behavior information set, where both the response label classification model and the intervention label classification model correspond to an object feature sequence; the determination unit 503 is configured to determine a screened causal feature set according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model; the second training unit 504 is configured to train a response behavior prediction model according to the screened causal feature set and the object behavior information set.

[0130] Optionally, before the first training unit 502, the prediction model training device 500 may further include: a division unit (not shown in the figure), configured to divide the object behavior information set into a training set and a test set.

[0131] Optionally, the first training unit 502 may be further configured to: train a response label classification model according to the training set corresponding to the object behavior information set and each response label included in the training set; train an intervention label classification model according to the training set corresponding to the object behavior information set and each intervention label included in the training set.

[0132] Optionally, the determination unit 503 may be further configured to: determine each object feature that satisfies a preset important feature condition in the object feature sequence corresponding to the response label classification model as a first object feature set; determine each object feature that satisfies a preset important feature condition in the object feature sequence corresponding to the intervention label classification model as a second object feature set; determine the intersection of the first object feature set and the second object feature set as the screened causal feature set.

[0133] Optionally, the second training unit 504 may be further configured to: add intervention features to the above-mentioned screened causal feature set to update the screened causal feature set; generate a training sample set according to the training set corresponding to the above-mentioned object behavior information set and the updated screened causal feature set, where each training sample in the above-mentioned training sample set includes a response label and respective object feature information corresponding to the screened causal feature set, and the object feature information corresponding to the intervention feature is an intervention label; based on the training sample set, perform the following training steps: input each training sample in at least one training sample in the training sample set into the embedding layer included in the initial response behavior prediction model to obtain respective word vectors corresponding to each training sample in at least one training sample for the above-mentioned screened causal feature set; generate a predicted response label corresponding to each training sample in at least one training sample according to the respective word vectors corresponding to each training sample in at least one training sample and the initial response behavior prediction model; compare the predicted response labels and the response labels corresponding to each training sample in at least one training sample to obtain a comparison result; in response to determining that the comparison result meets the optimization condition, determine the initial response behavior prediction model as the trained response behavior prediction model.

[0134] Optionally, the above-mentioned training steps may further include: generating a first loss value according to the predicted response labels and the response labels corresponding to each training sample in at least one training sample; generating a second loss value according to the respective word vectors corresponding to each training sample in at least one training sample for the above-mentioned screened causal feature set; generating a model loss value according to the first loss value and the second loss value; in response to determining that the comparison result does not meet the optimization condition, adjust the network parameters of the initial response behavior prediction model according to the model loss value, form a training sample set with the unused training samples, use the adjusted initial response behavior prediction model as the initial response behavior prediction model, and perform the above-mentioned training steps again.

[0135] Optionally, the second training unit 504 may be further configured to: input the respective word vectors corresponding to each training sample in at least one training sample obtained for the above-mentioned screened causal feature set into the first feature interaction layer included in the initial response behavior prediction model to obtain a concatenated vector corresponding to each training sample in at least one training sample; input the respective concatenated word vectors corresponding to each training sample in at least one training sample obtained into the first prediction layer included in the initial response behavior prediction model to obtain a predicted response label corresponding to each training sample in at least one training sample.

[0136] Optionally, the screened causal features in the screened causal feature set correspond to feature domains.

[0137] Optionally, the second training unit 504 may be further configured to, for each of the at least one training sample, perform the following steps: for each feature domain corresponding to the above-mentioned screened causal feature set, according to every two screened causal features corresponding to the above-mentioned feature domain in the above-mentioned screened causal feature set, generate the in-domain distance between the two word vectors corresponding to the above-mentioned two screened causal features of the above-mentioned training sample; determine the sum of the generated in-domain distances as the total in-domain distance of the sample; for each feature domain corresponding to the above-mentioned screened causal feature set, according to each screened causal feature corresponding to the above-mentioned feature domain in the above-mentioned screened causal feature set, perform the following steps: determine the respective word vectors corresponding to the above-mentioned screened causal feature set of the above-mentioned training sample as a feature matrix; remove the respective word vectors corresponding to the above-mentioned feature domain from the above-mentioned feature matrix to obtain a post-removal feature matrix; determine the similarity between the word vector corresponding to the above-mentioned screened causal feature and each word vector in the above-mentioned post-removal feature matrix; determine the sum of the determined similarities as the inter-domain distance of the sample; determine the sum of the determined inter-domain distances of the samples as the total inter-domain distance; determine the determined total in-domain distances of the samples as the total in-domain distance; determine the determined total inter-domain distances of the samples as the inter-domain distance; determine the sum of the above-mentioned total in-domain distance and the above-mentioned inter-domain distance as the second loss value.

[0138] Optionally, the second training unit 504 may be further configured to, for each of the at least one training sample, perform the following steps: perform a masking process on the feature matrix composed of the respective word vectors corresponding to the above-mentioned screened causal feature set of the above-mentioned training sample to obtain a first masked feature matrix and a second masked feature matrix; input the above-mentioned first masked feature matrix into the second feature interaction layer included in the initial response behavior prediction model to obtain a first concatenated vector; input the above-mentioned second masked feature matrix into the second feature interaction layer to obtain a second concatenated vector; input the above-mentioned first concatenated vector into the mapping layer included in the initial response behavior prediction model to obtain a first hidden layer vector; input the above-mentioned second concatenated vector into the mapping layer to obtain a second hidden layer vector; determine the distance between the above-mentioned first hidden layer vector and the above-mentioned second hidden layer vector as the contrast distance; determine the sum of the obtained contrast distances as the total contrast distance; determine the sum of the above-mentioned first loss value, the above-mentioned second loss value, and the above-mentioned total contrast distance as the model loss value.

[0139] Optionally, the prediction model training apparatus 500 may further include: a generation unit (not shown in the figure), configured to generate model evaluation information of the above-mentioned response behavior prediction model according to the above-mentioned test set.

[0140] It can be understood that the various units described in the apparatus 500 and the reference Figure 2corresponds to each step in the described method. Thus, the operations, features, and beneficial effects described above for the method also apply to the apparatus 500 and the units included therein, and will not be elaborated here.

[0141] Further referring to Figure 6 , as an implementation of the methods shown in the above figures, the present disclosure provides some embodiments of a response behavior prediction apparatus, and these apparatus embodiments correspond to Figure 4 the method embodiments shown, and the apparatus can be specifically applied to various electronic devices.

[0142] As shown in Figure 6 , some embodiments of the response behavior prediction apparatus 600 include: an item transfer behavior information acquisition unit 601 and an input unit 602. Among them, the item transfer behavior information acquisition unit 601 is configured to, for each target object in the target object set, acquire the item transfer behavior information of the target item corresponding to the above target object; the input unit 602 is configured to input each acquired item transfer behavior information into a pre-trained item transfer behavior prediction model to obtain an item transfer prediction behavior information set corresponding to the above target object set, where the above item transfer behavior prediction model is a response behavior prediction model and is trained through Figure 2 the steps in the corresponding embodiments.

[0143] Optionally, the response behavior prediction apparatus 600 may further include: an information to be intervened acquisition unit, a sending unit, and a control unit (not shown in the figure). Among them, the information to be intervened acquisition unit is configured to acquire the information to be intervened of the above target item. The sending unit is configured to, for each target object in the above target object set, in response to determining that the item transfer prediction behavior information corresponding to the above target object meets a preset intervention condition, send the above information to be intervened to the terminal device corresponding to the above target object. The control unit is configured to, in response to determining that the number of the sent information to be intervened meets a preset number condition, control the associated item scheduling device to perform a replenishment operation on the above target item.

[0144] It can be understood that the various units described in the apparatus 600 correspond to Figure 4 each step in the method described with reference to

[0145] Below referring to Figure 7 , which shows a schematic structural diagram of an electronic device 700 (such as Figure 1 the computing device 101 in Figure 7The electronic device shown is merely an example and should not impose any limitations on the functions and scope of use of the embodiments of the present disclosure.

[0146] As Figure 7 shown, the electronic device 700 may include a processing device 701 (such as a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage device 708 into a random access memory (RAM) 703. In the RAM 703, various programs and data required for the operation of the electronic device 700 are also stored. The processing device 701, the ROM 702, and the RAM 703 are connected to each other through a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0147] Generally, the following devices may be connected to the I / O interface 705: an input device 706 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 707 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 708 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 709. The communication device 709 can allow the electronic device 700 to communicate with other devices wirelessly or wirelessly to exchange data. Although Figure 7 the electronic device 700 with various devices is shown, it should be understood that it is not required to implement or have all the shown devices. Instead, more or fewer devices may be implemented or had. Figure 7 Each block shown in

[0148] particular, according to some embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of the present disclosure include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program codes for executing the methods shown in the flowcharts. In such some embodiments, the computer program can be downloaded and installed from the network through the communication device 709, or installed from the storage device 708, or installed from the ROM 702. When the computer program is executed by the processing device 701, the above-mentioned functions defined in the methods of some embodiments of the present disclosure are executed.

[0149] It should be noted that the computer-readable media described in some embodiments of the present disclosure may be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of the computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In some embodiments of the present disclosure, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of the present disclosure, the computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries computer-readable program code. Such a propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer-readable signal medium may also be any computer-readable medium other than the computer-readable storage medium, and the computer-readable signal medium may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0150] In some embodiments, the client and the server can communicate using any currently known or future-developed network protocol such as HTTP (HyperText Transfer Protocol), and can be interconnected with digital data communication in any form or medium (e.g., a communication network). Examples of communication networks include local area networks ("LAN"), wide area networks ("WAN"), the Internet (e.g., the Internet), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or future-developed network.

[0151] The above computer-readable medium may be included in the above electronic device; or it may exist separately and not be assembled into the electronic device. The above computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device is caused to: obtain a set of object behavior information, wherein the object behavior information in the set of object behavior information includes response labels, intervention labels, and a set of object feature information; train a response label classification model and an intervention label classification model according to the set of object behavior information, wherein both the response label classification model and the intervention label classification model correspond to an object feature sequence; determine a set of screened causal features according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model; and train a response behavior prediction model according to the set of screened causal features and the set of object behavior information.

[0152] Or cause the electronic device to: for each target object in the target object set, obtain the item transfer behavior information of the target item corresponding to the target object; input each obtained item transfer behavior information into a pre-trained item transfer behavior prediction model to obtain a set of item transfer prediction behavior information corresponding to the target object set, wherein the item transfer behavior prediction model is a response behavior prediction model and is trained through Figure 2 the steps in the corresponding embodiments.

[0153] Computer program code for performing the operations of some embodiments of the present disclosure may be written in one or more programming languages or combinations thereof. The programming languages include object-oriented programming languages such as Java, Smalltalk, C++, and also include conventional procedural programming languages such as the "C" language or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, executed as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (for example, by using an Internet service provider to connect through the Internet).

[0154] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a segment of a program, or a portion of code that contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions noted in the blocks may occur in a different order than that noted in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or by a combination of dedicated hardware and computer instructions.

[0155] The units described in some embodiments of the present disclosure can be implemented in software or in hardware. The described units can also be provided in a processor. For example, it can be described as: a processor includes an object behavior information set acquisition unit, a first training unit, a determination unit, and a second training unit. Among them, the names of these units do not constitute a limitation on the unit itself in some cases. For example, the object behavior information set acquisition unit can also be described as "the unit for acquiring the object behavior information set".

[0156] The functions described above herein can be performed, at least in part, by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that can be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on a chip (SOCs), complex programmable logic devices (CPLDs), and so on.

[0157] Some embodiments of the present disclosure also provide a computer program product, including a computer program that, when executed by a processor, implements any of the above response behavior prediction model training methods or response behavior prediction methods.

[0158] The above description is only some preferred embodiments of the present disclosure and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, technical solutions formed by mutually replacing the above features with (but not limited to) technical features having similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A method for training a response behavior prediction model, comprising: Obtain an object behavior information set, where the object behavior information in the object behavior information set includes a response label, an intervention label, and an object feature information set; Train a response label classification model and an intervention label classification model according to the object behavior information set, where both the response label classification model and the intervention label classification model correspond to an object feature sequence; Determine a screened causal feature set according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model; Train a response behavior prediction model according to the screened causal feature set and the object behavior information set.

2. The method according to claim 1, wherein, Before training the response label classification model and the intervention label classification model according to the object behavior information set, the method further includes: Divide the object behavior information set into a training set and a test set.

3. The method according to claim 2, wherein, Training the response label classification model and the intervention label classification model according to the object behavior information set includes: Train a response label classification model according to the training set corresponding to the object behavior information set and each response label included in the training set; Train an intervention label classification model according to the training set corresponding to the object behavior information set and each intervention label included in the training set.

4. The method according to claim 1, wherein, Determining the screened causal feature set according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model includes: Determine each object feature that satisfies a preset important feature condition in the object feature sequence corresponding to the response label classification model as a first object feature set; Determine each object feature that satisfies a preset important feature condition in the object feature sequence corresponding to the intervention label classification model as a second object feature set; Determine the intersection of the first object feature set and the second object feature set as the screened causal feature set.

5. The method according to claim 2, wherein, Training the response behavior prediction model according to the screened causal feature set and the object behavior information set includes: Add intervention features to the screened causal feature set to update the screened causal feature set; Generate a training sample set according to the training set corresponding to the object behavior information set and the updated screened causal feature set, where each training sample in the training sample set includes a response label and each object feature information corresponding to the screened causal feature set, and the object feature information corresponding to the intervention feature is an intervention label; Based on the training sample set, perform the following training steps: Input each training sample in at least one training sample in the training sample set into the embedding layer included in the initial response behavior prediction model to obtain each word vector corresponding to the screened causal feature set for each training sample in at least one training sample; Generate a predicted response label corresponding to each training sample in at least one training sample according to each word vector corresponding to each training sample in at least one training sample and the initial response behavior prediction model; Compare the predicted response labels corresponding to each training sample in at least one training sample with the response labels to obtain a comparison result; In response to determining that the comparison result reaches the optimization condition, the initial response behavior prediction model is determined as the trained response behavior prediction model.

6. The method according to claim 5, wherein, The training step further includes: generating a first loss value according to the predicted response labels and response labels corresponding to each training sample in at least one training sample; generating a second loss value according to each word vector corresponding to the screened causal feature set in each training sample of at least one training sample; generating a model loss value according to the first loss value and the second loss value; In response to determining that the comparison result does not reach the optimization condition, adjusting the network parameters of the initial response behavior prediction model according to the model loss value, forming a training sample set with the unused training samples, using the adjusted initial response behavior prediction model as the initial response behavior prediction model, and executing the training step again.

7. The method according to claim 5, wherein, The generating, according to each word vector corresponding to each training sample in at least one training sample and the initial response behavior prediction model, a predicted response label corresponding to each training sample in at least one training sample includes: inputting each word vector corresponding to the screened causal feature set in each training sample obtained into a first feature interaction layer included in the initial response behavior prediction model to obtain a concatenated vector corresponding to each training sample in at least one training sample; inputting each concatenated word vector corresponding to each training sample in at least one training sample obtained into a first prediction layer included in the initial response behavior prediction model to obtain a predicted response label corresponding to each training sample in at least one training sample.

8. The method according to claim 5, wherein, The screened causal features in the screened causal feature set correspond to feature domains; and The generating, according to each word vector corresponding to the screened causal feature set in each training sample of at least one training sample, a second loss value includes: For each training sample in the at least one training sample, perform the following steps: For each feature domain corresponding to the screened causal feature set, generate an in-domain distance between two word vectors corresponding to the two screened causal features in the feature domain for the training sample; determine the sum of the generated in-domain distances as the total in-domain distance of the sample; For each feature domain corresponding to the screened causal feature set, according to each screened causal feature corresponding to the feature domain, perform the following steps: determine each word vector corresponding to the screened causal feature set for the training sample as a feature matrix; exclude each word vector corresponding to the feature domain from the feature matrix to obtain an excluded feature matrix; determine the similarity between the word vector corresponding to the screened causal feature and each word vector in the excluded feature matrix; determine the sum of the determined similarities as the inter-domain distance of the sample; determine the sum of the determined inter-domain distances of each sample as the total inter-domain distance; determine the total in-domain distance of each sample determined as the in-domain total distance; determine the sum of the determined total inter-domain distances of each sample as the inter-domain total distance; determine the sum of the in-domain total distance and the inter-domain total distance as the second loss value.

9. The method according to claim 5, wherein, Generating a model loss value based on the first loss value and the second loss value includes: For each of the at least one training sample, perform the following steps: Perform masking processing on the feature matrix composed of word vectors corresponding to the training sample in the screened causal feature set to obtain a first masked feature matrix and a second masked feature matrix; Input the first masked feature matrix into the second feature interaction layer included in the initial response behavior prediction model to obtain a first concatenated vector; Input the second masked feature matrix into the second feature interaction layer to obtain a second concatenated vector; Input the first concatenated vector into the mapping layer included in the initial response behavior prediction model to obtain a first hidden layer vector; Input the second concatenated vector into the mapping layer to obtain a second hidden layer vector; Determine the distance between the first hidden layer vector and the second hidden layer vector as the contrast distance; Determine the sum of the obtained contrast distances as the total contrast distance; Determine the sum of the first loss value, the second loss value, and the total contrast distance as the model loss value.

10. The method according to any one of claims 2 - 3, 5 - 9, wherein, The method further includes: Generating model evaluation information of the response behavior prediction model according to the test set.

11. A response behavior prediction method, comprising: For each target object in the target object set, obtain the item transfer behavior information of the target object corresponding to the target item; Input each obtained item transfer behavior information into a pre-trained item transfer behavior prediction model to obtain an item transfer prediction behavior information set corresponding to the target object set, where the item transfer behavior prediction model is a response behavior prediction model and is trained by the method according to any one of claims 1-10.

12. The method according to claim 11, wherein, The method further includes: Obtain the information to be intervened for the target item; For each target object in the target object set, in response to determining that the item transfer prediction behavior information corresponding to the target object meets a preset intervention condition, send the information to be intervened to the terminal device corresponding to the target object; In response to determining that the number of the sent information to be intervened meets a preset quantity condition, control the associated item scheduling device to perform a replenishment operation for the target item.

13. A response behavior prediction model training device, comprising: An object behavior information set acquisition unit is configured to acquire an object behavior information set, where the object behavior information in the object behavior information set includes a response label, an intervention label, and an object feature information set; A first training unit is configured to train a response label classification model and an intervention label classification model according to the object behavior information set, where both the response label classification model and the intervention label classification model correspond to an object feature sequence; A determination unit is configured to determine a screened causal feature set according to the object feature sequence corresponding to the response label classification model and the object feature sequence corresponding to the intervention label classification model; A second training unit is configured to train a response behavior prediction model according to the screened causal feature set and the object behavior information set.

14. A response behavior prediction device, comprising: An item transfer behavior information acquisition unit is configured to, for each target object in the target object set, obtain the item transfer behavior information of the target object corresponding to the target item; An input unit configured to input the obtained item transfer behavior information for each item into a pre-trained item transfer behavior prediction model to obtain an item transfer prediction behavior information set corresponding to the target object set, wherein the item transfer behavior prediction model is a response behavior prediction model and is trained by the method according to any one of claims 1-10.

15. An electronic device, comprising: One or more processors; A storage device having stored thereon one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the method according to any one of claims 1-10 or 11-12.

16. A computer-readable medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, the method according to any one of claims 1-10 or 11-12 is implemented.

17. A computer program product, comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1-10 or 11-12.

Images