Gain model training method and device and method and device for predicting reach gain by using gain model

Abstract

The invention discloses a method and device for training a gain model and predicting reach gain by using the gain model, and relates to the technical field of artificial intelligence. A specific embodiment of the gain model training method comprises the steps of performing feature processing on experimental group user features and control group user features to obtain a first user representation based on facts, a second user representation based on anti-facts, a third user representation based on facts and a fourth user representation based on anti-facts; splicing the first user characterization and the third user characterization with the user touch features, and then performing model prediction to obtain experimental group prediction loss and control group prediction loss; performing comparative learning to obtain the characterization loss of the experimental group and the characterization loss of the control group; and performing model parameter adjustment based on the experimental group prediction loss, the control group prediction loss, the experimental group characterization loss and the control group characterization loss. According to the embodiment, the problem of unbalanced sample distribution is solved, the accuracy of the model prediction result is improved, and the intelligent marketing effect is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for training a gain model and using the gain model to predict the reach gain. Background Technology

[0002] With the development of artificial intelligence technology, intelligent marketing has gradually expanded into scenarios such as recommendation and advertising. The effectiveness of intelligent marketing is typically measured by gain models used to predict user engagement with promotional information (e.g., whether a user purchased a recommended product, or visited a store). In gain modeling, since it's impossible to simultaneously observe both engagement and non-engagement for the same sample, randomized controlled trials are usually conducted, requiring engagement to be independent of the user. However, in real-world scenarios, this is difficult to achieve. For example, when modeling the gain of coupons on conversion rates, even with random coupon distribution, the model may predict that users with high conversion rates are more likely to receive coupons. This means that engagement is not independent of the user. Furthermore, the magnitude of the difference can be significant; for instance, the number of samples receiving coupons may be far fewer than those not receiving them, leading to an imbalanced sample distribution and inaccurate model predictions, resulting in poor intelligent marketing effectiveness. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a method and apparatus for training a gain model and using the gain model to predict reach gain, which can solve the problem of unbalanced sample distribution, improve the accuracy of model prediction results, and enhance the effectiveness of intelligent marketing.

[0004] To achieve the above objectives, according to a first aspect of the present invention, a method for training a promotion information reach gain model is provided, comprising: Feature processing was performed on the user characteristics of the experimental group and the user characteristics of the control group respectively to obtain the first user representation based on facts and the second user representation based on counterfacts corresponding to the user characteristics of the experimental group, and the third user representation based on facts and the fourth user representation based on counterfacts corresponding to the user characteristics of the control group. The first user representation and the third user representation are concatenated with user reach features to obtain a first concatenated user representation and a third concatenated user representation; model prediction is performed based on the first concatenated user representation to obtain the experimental group prediction loss; model prediction is performed based on the third concatenated user representation to obtain the control group prediction loss. The experimental group representation loss is obtained by comparative learning based on the first user representation and the second user representation, and the control group representation loss is obtained by comparative learning based on the third user representation and the fourth user representation. The model parameters are adjusted based on the experimental group prediction loss, the control group prediction loss, the experimental group representation loss, and the control group representation loss to train the promotion information reach gain model.

[0005] Optionally, feature processing is performed on the user characteristics of the experimental group and the user characteristics of the control group respectively to obtain the first user representation and the second user representation based on facts corresponding to the user characteristics of the experimental group, and the third user representation and the fourth user representation based on facts corresponding to the user characteristics of the control group. This includes: inputting the user characteristics of the experimental group into the feature processing sub-model of the experimental group and the feature processing sub-model of the control group respectively to obtain the first user representation and the second user representation based on facts corresponding to the user characteristics of the experimental group; and inputting the user characteristics of the control group into the feature processing sub-model of the control group and the feature processing sub-model of the experimental group respectively to obtain the third user representation and the fourth user representation based on facts corresponding to the user characteristics of the control group.

[0006] Optionally, feature processing is performed on the user characteristics of the experimental group and the user characteristics of the control group, respectively, including: abstracting the user characteristics of the experimental group and the user characteristics of the control group into high-dimensional representations based on user behavior sequences for feature processing.

[0007] Optionally, the user reach features are generated based on the user behavior sequences of the experimental group and the user behavior sequences of the control group.

[0008] Optionally, the experimental group representation loss is obtained by comparative learning based on the first user representation and the second user representation, and the control group representation loss is obtained by comparative learning based on the third user representation and the fourth user representation. This includes: taking the first user representation and the second user representation belonging to the same sample as positive examples of the experimental group, and the first user representation and the second user representation not belonging to the same sample as negative examples of the experimental group, and obtaining the experimental group representation loss by comparative learning based on the positive and negative examples of the experimental group; taking the third user representation and the fourth user representation belonging to the same sample as positive examples of the control group, and the third user representation and the fourth user representation not belonging to the same sample as negative examples of the control group, and obtaining the control group representation loss by comparative learning based on the positive and negative examples of the control group.

[0009] Optionally, the experimental group representation loss is determined based on the similarity between positive examples and the similarity between negative examples in the experimental group; the control group representation loss is determined based on the similarity between positive examples and the similarity between negative examples in the control group.

[0010] Optionally, the characterization loss of the experimental group and the characterization loss of the control group are calculated using the following formula: ; ;in, To characterize the loss in the experimental group, The control group represents the loss. represents a positive example, Indicates a negative example. It's a hyperparameter. This indicates similarity calculation. This indicates the sample size in the experimental or control group, where k and i are the sample numbers. This represents the user representation output by the i-th sample after passing through the experimental group feature processing sub-model. This represents the user representation output by the i-th sample after passing through the feature processing sub-model of the control group. This represents the user representation output by the k-th sample after passing through the control group feature processing sub-model.

[0011] Optionally, adjusting the model parameters based on the experimental group prediction loss, the control group prediction loss, the experimental group characterization loss, and the control group characterization loss includes: performing a weighted summation of the experimental group prediction loss, the control group prediction loss, the experimental group characterization loss, and the control group characterization loss to obtain the model loss; and adjusting the model parameters based on the model loss.

[0012] According to a second aspect of the present invention, a method is provided for predicting the reach gain of a user to a promotional message using a promotional message reach gain model trained by the method described in the first aspect of the present invention, comprising: The user characteristics of the predicted user are input into the promotion information reach gain model to obtain the first predicted probability that the predicted user will reach the promotion information and the second predicted probability that the predicted user will not reach the promotion information. The reach gain of the promotional information to the predicted user is calculated based on the first prediction probability and the second prediction probability.

[0013] Optionally, the promotional information reach gain model includes an experimental group feature processing sub-model and a control group feature processing sub-model. Inputting the user features of the predicted user into the promotional information reach gain model to obtain a first predicted probability that the predicted user will reach the promotional information and a second predicted probability that the predicted user will not reach the promotional information includes: inputting the user features of the predicted user into the promotional information reach gain model; obtaining a first predicted user representation and a second predicted user representation corresponding to the user features of the predicted user through the experimental group feature processing sub-model and the control group feature processing sub-model of the promotional information reach gain model, respectively; concatenating the first predicted user representation with a preset user reach feature used to represent that the user has reached the promotional information, and then predicting the first predicted probability that the predicted user will reach the promotional information; concatenating the second predicted user representation with a preset user reach feature used to represent that the user has not reached the promotional information, and then predicting the second predicted probability that the predicted user will not reach the promotional information.

[0014] According to a third aspect of the present invention, a training apparatus for a promotion information reach gain model is provided, comprising: The user feature processing module is used to process the user features of the experimental group and the user features of the control group respectively, to obtain the first user representation based on facts and the second user representation based on counterfacts corresponding to the user features of the experimental group, and the third user representation based on facts and the fourth user representation based on counterfacts corresponding to the user features of the control group. The prediction loss determination module is used to concatenate the first user representation and the third user representation with user reach features to obtain a first concatenated user representation and a third concatenated user representation; perform model prediction based on the first concatenated user representation to obtain the experimental group prediction loss; and perform model prediction based on the third concatenated user representation to obtain the control group prediction loss. The representation loss determination module is used to obtain the experimental group representation loss by comparative learning based on the first user representation and the second user representation, and to obtain the control group representation loss by comparative learning based on the third user representation and the fourth user representation. The model parameter adjustment module is used to adjust the model parameters based on the experimental group prediction loss, the control group prediction loss, the experimental group representation loss, and the control group representation loss, so as to train the promotion information reach gain model.

[0015] According to a fourth aspect of the present invention, an apparatus is provided for predicting the reach gain of a user to a promotional message using a promotional message reach gain model trained using the apparatus described in the third aspect of the present invention, comprising: The prediction probability generation module is used to input the user characteristics of the predicted user into the promotion information reach gain model to obtain a first prediction probability that the predicted user will reach the promotion information and a second prediction probability that the predicted user will not reach the promotion information. The reach gain calculation module is used to calculate the reach gain of the predicted user to the promotional information based on the first prediction probability and the second prediction probability.

[0016] According to a fifth aspect of the present invention, an electronic device is provided, comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the training method for the promotional information reach gain model provided in the embodiments of the present invention.

[0017] According to a sixth aspect of the present invention, a computer-readable medium is provided having a computer program stored thereon, which, when executed by a processor, implements the training method for the promotional information reach gain model provided in the embodiments of the present invention.

[0018] According to a seventh aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the training method for the promotional information reach gain model provided in the embodiments of the present invention.

[0019] One embodiment of the above invention has the following advantages or beneficial effects: By performing feature processing on the user characteristics of the experimental group and the user characteristics of the control group respectively, a fact-based first user representation and a counterfactual second user representation corresponding to the user characteristics of the experimental group are obtained, as well as a fact-based third user representation and a counterfactual fourth user representation corresponding to the user characteristics of the control group; the first user representation and the third user representation are respectively concatenated with user reach features to obtain a first concatenated user representation and a third concatenated user representation; model prediction is performed based on the first concatenated user representation to obtain the experimental group prediction loss; model prediction is performed based on the third concatenated user representation to obtain the control group prediction loss; comparative learning is performed based on the first user representation and the second user representation. The experimental group representation loss is obtained, and the control group representation loss is obtained through comparative learning based on the third and fourth user representations. Model parameters are adjusted based on the experimental group prediction loss, control group prediction loss, experimental group representation loss, and control group representation loss to train the promotional information reach gain model. This model generates fact-based and counterfactual user representations for the experimental and control group user features, and compares and learns from them to train the gain model. This aligns the fact-based and counterfactual user representations of the same sample, thus equalizing the distribution of the experimental and control groups at the sample feature level, solving the problem of imbalanced sample distribution, improving the accuracy of model prediction results, and enhancing the effectiveness of intelligent marketing. Furthermore, the introduction of reach features into the generated features used for model prediction prevents reach features from being overwhelmed, further improving the accuracy of model prediction results.

[0020] The further effects of the aforementioned unconventional alternative methods will be explained below in conjunction with specific implementation methods. Attached Figure Description

[0021] The accompanying drawings are provided to better understand the invention and are not intended to unduly limit the scope of the invention. Wherein: Figure 1 This is a schematic diagram of the main steps of the training method for the promotion information reach gain model according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the implementation principle of a training method for a promotional information reach gain model according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the main steps of a method for predicting the reach gain of promotional information according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the main modules of a training device for a promotional information reach gain model according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the main modules of a device for predicting the reach gain of promotional information according to an embodiment of the present invention; Figure 6 This is an exemplary system architecture diagram in which embodiments of the present invention can be applied; Figure 7 This is a schematic diagram of the structure of a computer system suitable for implementing terminal devices or servers of the present invention. Detailed Implementation

[0022] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0023] It should be noted that the technical solutions disclosed in this invention, regarding the collection, updating, analysis, processing, use, transmission, and storage of user personal information, all comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. Necessary measures are taken to prevent unauthorized access to user personal information data and to safeguard user personal information security, network security, and national security.

[0024] In the current training process of promotional information reach gain models, it is impossible to simultaneously observe the conversion status of the same sample when it is reached and when it is not reached. Therefore, it is necessary to infer a counterfactual result from observed instances. That is, if it is an experimental group sample (i.e., users who actually reached the promotional information), it is necessary to infer the situation when they did not reach the promotional information; if it is a control group sample (i.e., users who did not actually reach the promotional information), it is necessary to infer the situation when they reached the promotional information. However, actual reach is not independent of the user, resulting in an imbalance in the distribution of the experimental and control groups. To alleviate this imbalance, it is necessary to control the distribution of the training data for both groups. In one embodiment of this invention, each of the experimental and control groups has a sub-model that can generate fact-based user representations. Then, the other's sub-model is used to generate corresponding counterfactual user representations. The fact-based user representations and counterfactual user representations are then compared and learned to align the representation of the same sample in the two sub-models, thereby aligning the distribution of the experimental and control groups at the sample feature level. In this way, both models process the full training data, further solving the problem of sample imbalance. Furthermore, reach features are also incorporated into the final generated features to predict the corresponding conversion results and prevent reach features from being overwhelmed.

[0025] Figure 1 This is a schematic diagram illustrating the main steps of the training method for the promotional information reach gain model according to an embodiment of the present invention. Figure 1As shown, the training method of the promotion information reach gain model in this embodiment of the invention mainly includes the following steps S101 to S104.

[0026] Step S101: Perform feature processing on the user characteristics of the experimental group and the control group respectively to obtain the fact-based first user representation and counterfactual second user representation corresponding to the user characteristics of the experimental group, and the fact-based third user representation and counterfactual fourth user representation corresponding to the user characteristics of the control group. Before model training, all training data is first divided into experimental and control groups. The training data here mainly comes from real promotion data, that is, actual marketing promotion tests without special processing. The experimental group consists of users who have been reached by promotional information, and the control group consists of users who have not been reached by promotional information. Here, promotional information includes, for example, social advertisements, outbound call notifications, coupon pushes, etc. User characteristics are, for example, basic user characteristics such as gender and education level represented by numbers.

[0027] According to an embodiment of the present invention, when training the model, the training data corresponding to the experimental group and the training data corresponding to the control group can be divided into multiple batches, and one batch of training data can be used for model training each time.

[0028] According to an embodiment of the present invention, step S101 may include: inputting the experimental group user characteristics into the experimental group feature processing sub-model and the control group feature processing sub-model respectively to obtain a fact-based first user representation and a counterfactual second user representation corresponding to the experimental group user characteristics; inputting the control group user characteristics into the control group feature processing sub-model and the experimental group feature processing sub-model respectively to obtain a fact-based third user representation and a counterfactual fourth user representation corresponding to the control group user characteristics. The experimental group feature processing sub-model is a feature processing sub-model corresponding to the experimental group user characteristics. Since all users in the experimental group have been exposed to the promotional information, the feature processing of the experimental group user characteristics yields a fact-based first user representation. The control group feature processing sub-model is a feature processing sub-model corresponding to the control group user characteristics. Since all users in the control group have not been exposed to the promotional information, the feature processing of the control group user characteristics yields a fact-based third user representation. Correspondingly, after the feature processing sub-model of the control group processes the user features of the experimental group, the result is the second user representation of the counterfactual fact; after the feature processing sub-model of the experimental group processes the user features of the control group, the result is the fourth user representation of the counterfactual fact.

[0029] According to one embodiment of the present invention, feature processing is performed on the user features of the experimental group and the user features of the control group, respectively. Specifically, this may include: abstracting the user features of the experimental group and the user features of the control group into high-dimensional representations by combining user behavior sequences for feature processing. The feature processing sub-models for the experimental group and the control group are, for example, multi-layered DNNs (Deep Neural Networks) or RNNs (Recurrent Neural Networks) containing user behavior sequences. Their main function is to abstract the original user features into high-dimensional representations to represent the inherent characteristics or behavioral characteristics of the users.

[0030] Step S102: Concatenate the first user representation and the third user representation with the user reach features to obtain the first concatenated user representation and the third concatenated user representation; perform model prediction based on the first concatenated user representation to obtain the experimental group prediction loss; perform model prediction based on the third concatenated user representation to obtain the control group prediction loss.

[0031] According to one embodiment of the present invention, user reach features are generated based on the user behavior sequences of the experimental group and the user behavior sequences of the control group. Here, the reach features can be single-dimensional features indicating whether or not reach has occurred, or they can be multiple features, such as a combination of reach methods, reach times, etc. If no reach has occurred, the feature values ​​corresponding to reach methods, reach times, etc., can be, for example, 0. It should be noted that if there are multiple features, the training data itself needs to be sufficiently rich in features. That is, during actual marketing promotion testing, there can be multiple reach methods, multiple reach times, etc., and these data can be collected as reach features.

[0032] The promotional information reach gain model of this invention is used to predict whether a user will convert in the future, such as by visiting a store or placing an order. Specifically, the prediction result is, for example, the user's predicted conversion probability.

[0033] According to one embodiment of the present invention, the experimental group prediction loss and the control group prediction loss are calculated, for example, using the cross-entropy loss function. Taking the experimental group prediction loss as an example, ,in Indicates the predicted conversion probability. This represents the actual conversion result. It measures the difference between the distribution of the predicted conversion probability predicted by the promotional information reach gain model and the probability distribution of the actual conversion result (i.e., the value of the reach feature). The smaller the value, the better the model's prediction effect. For this invention, the prediction problem is a binary classification problem.

[0034] Step S103: Obtain the experimental group representation loss by performing contrastive learning based on the first user representation and the second user representation, and obtain the control group representation loss by performing contrastive learning based on the third user representation and the fourth user representation. In one embodiment of the present invention, contrastive learning can be performed, for example, through a contrastive learning sub-model.

[0035] According to an embodiment of the present invention, step S103 may include: taking the first user representation and the second user representation belonging to the same sample as positive examples of the experimental group, and the first user representation and the second user representation not belonging to the same sample as negative examples of the experimental group, and performing comparative learning based on the positive examples and negative examples of the experimental group to obtain the representation loss of the experimental group; taking the third user representation and the fourth user representation belonging to the same sample as positive examples of the control group, and the third user representation and the fourth user representation not belonging to the same sample as negative examples of the control group, and performing comparative learning based on the positive examples and negative examples of the control group to obtain the representation loss of the control group.

[0036] According to one embodiment of the present invention, the characterization loss of the experimental group is determined based on the similarity between positive examples and the similarity between negative examples in the experimental group; the characterization loss of the control group is determined based on the similarity between positive examples and the similarity between negative examples in the control group.

[0037] According to one embodiment of the present invention, the characterization loss of the experimental group and the characterization loss of the control group are calculated, for example, by the following formula: ; ;in, To characterize the loss in the experimental group, The control group represents the loss. represents a positive example, Indicates a negative example. It's a hyperparameter. This indicates similarity calculation. This indicates the sample size in the experimental or control group, where k and i are the sample numbers. This represents the user representation output by the i-th sample after passing through the experimental group feature processing sub-model. This represents the user representation output by the i-th sample after passing through the feature processing sub-model of the control group. This represents the user representation output by the k-th sample after passing through the control group feature processing sub-model. Calculating the representation loss of the experimental group and the control group using the above formulas can make the user representations of different samples closer together and the user representations of different samples farther apart.

[0038] Step S104: Adjust the model parameters based on the experimental group prediction loss, the control group prediction loss, the experimental group representation loss, and the control group representation loss to train the promotion information reach gain model.

[0039] According to one embodiment of the present invention, step S104 may include: weighted summing of the experimental group prediction loss, the control group prediction loss, the experimental group characterization loss, and the control group characterization loss to obtain the model loss; and adjusting the model parameters based on the model loss. Assume the experimental group prediction loss is... The predicted loss for the control group was... The characterization loss of the experimental group was The control group showed the following characteristics: loss Then the model loss L is: , in, , , These are all weighted hyperparameters.

[0040] When adjusting model parameters based on model loss, one could set a loss threshold and adjust the model parameters accordingly until the model loss does not exceed the loss threshold.

[0041] Figure 2 This is a schematic diagram illustrating the implementation principle of a training method for a promotional information reach gain model according to an embodiment of the present invention. For example... Figure 2 As shown, for the training data corresponding to the experimental group, the user features of the experimental group (t1, t2, ... t) are... n Input experimental group feature processing sub-model We obtain a high-dimensional, fact-based first user representation, and then combine the user features of the experimental group (t1, t2, ... t) n Input control group feature processing sub-model A high-dimensional counterfactual second user representation is obtained. Then, the first user representation is concatenated with the user reach feature T to obtain a first concatenated user representation. Based on this first concatenated user representation, model prediction is performed to obtain the predicted conversion result. And based on the predicted conversion results and actual conversion results The experimental group's predicted loss is obtained; simultaneously, the experimental group's representation loss is obtained through comparative learning based on the first user's representation and the second user's representation.

[0042] For the training data corresponding to the control group, the user features of the control group (c1, c2, ... c) n Input experimental group feature processing sub-model We obtain a high-dimensional counterfactual fourth user representation, and then combine the user characteristics of the experimental group (c1, c2, ... c)n Input control group feature processing sub-model A high-dimensional fact-based third-party user representation is obtained. Then, this third-party user representation is concatenated with the user reach feature T to obtain a concatenated third-party user representation. Based on this concatenated third-party user representation, model prediction is performed to obtain the predicted conversion result. And based on the predicted conversion results and actual conversion results The predicted loss of the control group is obtained; at the same time, the control group representation loss is obtained by comparative learning based on the representations of the third user and the fourth user.

[0043] Finally, the model parameters can be adjusted based on the experimental group prediction loss, the control group prediction loss, the experimental group representation loss, and the control group representation loss to train the promotion information reach gain model.

[0044] Figure 3 This is a schematic diagram illustrating the main steps of a method for predicting the reach gain of promotional information according to an embodiment of the present invention. Figure 3 As shown, the method for predicting the reach gain of promotional information in this embodiment of the invention mainly includes the following steps S301 and S302.

[0045] Step S301: Input the predicted user's user characteristics into the promotional information reach gain model to obtain a first predicted probability that the predicted user will receive the promotional information and a second predicted probability that the predicted user will not receive the promotional information. The promotional information reach gain model is trained using the training method described in the previous embodiment. The predicted user is the user to be predicted whether to receive promotional information, and the predicted user's user characteristics are, for example, basic user characteristics such as gender and education level represented by numbers.

[0046] According to an embodiment of the present invention, step S301 may include: inputting the user characteristics of the predicted user into the promotional information reach gain model; obtaining a first predicted user representation and a second predicted user representation corresponding to the user characteristics of the predicted user through the experimental group feature processing sub-model and the control group feature processing sub-model of the promotional information reach gain model, respectively; concatenating the first predicted user representation with a preset user reach feature used to represent that the user has reached the promotional information, and predicting a first predicted probability that the predicted user has reached the promotional information; concatenating the second predicted user representation with a preset user reach feature used to represent that the user has not reached the promotional information, and predicting a second predicted probability that the predicted user has not reached the promotional information. The experimental group feature processing sub-model is a feature processing sub-model corresponding to the experimental group user characteristics, and the control group feature processing sub-model is a feature processing sub-model corresponding to the control group user characteristics. By inputting the user characteristics of the predicted user into the promotional information reach gain model, the first predicted user representation and the second predicted user representation corresponding to the user characteristics of the predicted user can be obtained through the experimental group feature processing sub-model and the control group feature processing sub-model.

[0047] In specific implementation, preset user reach characteristics are used to represent that a user has reached the promotional information, such as the user reaching the promotional information, or the reach characteristics such as the time and method of reaching the user reaching the promotional information; preset user reach characteristics are used to represent that a user has not reached the promotional information, such as the reach characteristics of a user not reaching the promotional information, etc.

[0048] Step S302: Calculate the reach gain of the promotional information for the predicted users based on the first and second predicted probabilities. In practice, the difference between the first and second predicted probabilities can be used as the reach gain for the predicted users. During actual marketing campaigns, multiple predicted users can be ranked according to their reach gain, and a certain percentage of the top predicted users can be selected for information promotion. This can improve the conversion rate of the information promotion.

[0049] Figure 4 This is a schematic diagram of the main modules of a training device for a promotional information reach gain model according to an embodiment of the present invention. Figure 4 As shown, the training device 400 for the promotion information reach gain model in this embodiment of the invention mainly includes a user feature processing module 401, a prediction loss determination module 402, a representation loss determination module 403, and a model parameter adjustment module 404.

[0050] User feature processing module 401 is used to process the user features of the experimental group and the user features of the control group respectively to obtain the first user representation based on facts and the second user representation based on counterfacts corresponding to the user features of the experimental group, and the third user representation based on facts and the fourth user representation based on counterfacts corresponding to the user features of the control group. The prediction loss determination module 402 is used to concatenate the first user representation and the third user representation with the user reach features to obtain the first concatenated user representation and the third concatenated user representation, respectively; perform model prediction based on the first concatenated user representation to obtain the experimental group prediction loss; and perform model prediction based on the third concatenated user representation to obtain the control group prediction loss. The representation loss determination module 403 is used to obtain the representation loss of the experimental group by comparative learning based on the first user representation and the second user representation, and to obtain the representation loss of the control group by comparative learning based on the third user representation and the fourth user representation. The model parameter adjustment module 404 is used to adjust the model parameters based on the experimental group prediction loss, the control group prediction loss, the experimental group representation loss, and the control group representation loss in order to train the promotion information reach gain model.

[0051] According to an embodiment of the present invention, the user feature processing module 401 can be specifically used to: input the experimental group user features into the experimental group feature processing sub-model and the control group feature processing sub-model respectively to obtain the first user representation based on facts and the second user representation based on counterfacts corresponding to the experimental group user features; input the control group user features into the control group feature processing sub-model and the experimental group feature processing sub-model respectively to obtain the third user representation based on facts and the fourth user representation based on counterfacts corresponding to the control group user features.

[0052] According to one embodiment of the present invention, the user feature processing module 401 can be specifically used to: perform feature abstraction and generate high-dimensional representations for the experimental group user features and the control group user features by combining user behavior sequences, respectively, for feature processing.

[0053] According to one embodiment of the present invention, the user reach features are generated based on the user behavior sequences of the experimental group and the user behavior sequences of the control group.

[0054] According to one embodiment of the present invention, the representation loss determination module 403 can be specifically used to: take the first user representation and the second user representation belonging to the same sample as positive examples of the experimental group, and the first user representation and the second user representation not belonging to the same sample as negative examples of the experimental group, and perform comparative learning based on the positive examples and negative examples of the experimental group to obtain the representation loss of the experimental group; take the third user representation and the fourth user representation belonging to the same sample as positive examples of the control group, and the third user representation and the fourth user representation not belonging to the same sample as negative examples of the control group, and perform comparative learning based on the positive examples and negative examples of the control group to obtain the representation loss of the control group.

[0055] According to one embodiment of the present invention, the characterization loss of the experimental group is determined based on the similarity between positive examples and the similarity between negative examples in the experimental group; the characterization loss of the control group is determined based on the similarity between positive examples and the similarity between negative examples in the control group.

[0056] According to one embodiment of the present invention, the characterization loss of the experimental group and the characterization loss of the control group are calculated by the following formula: ; ;in, To characterize the loss in the experimental group, The control group represents the loss. represents a positive example, Indicates a negative example. It's a hyperparameter. This indicates similarity calculation. This indicates the sample size in the experimental or control group, where k and i are the sample numbers. This represents the user representation output by the i-th sample after passing through the experimental group feature processing sub-model. This represents the user representation output by the i-th sample after passing through the feature processing sub-model of the control group. This represents the user representation output by the k-th sample after passing through the control group feature processing sub-model.

[0057] According to one embodiment of the present invention, the model parameter adjustment module 404 can be specifically used to: perform a weighted summation of the experimental group prediction loss, the control group prediction loss, the experimental group characterization loss, and the control group characterization loss to obtain the model loss; and adjust the model parameters based on the model loss.

[0058] Figure 5 This is a schematic diagram of the main modules of an apparatus for predicting the reach gain of promotional information according to an embodiment of the present invention. Figure 5 As shown, the device 500 for predicting the reach gain of promotional information in this embodiment of the invention mainly includes a prediction probability generation module 501 and a reach gain calculation module 502.

[0059] The prediction probability generation module 501 is used to input the user characteristics of the predicted user into the promotion information reach gain model to obtain the first prediction probability that the predicted user will reach the promotion information and the second prediction probability that the predicted user will not reach the promotion information. The reach gain calculation module 502 is used to calculate the reach gain of the predicted user to the promotional information based on the first prediction probability and the second prediction probability.

[0060] According to an embodiment of the present invention, the promotional information reach gain model includes an experimental group feature processing sub-model and a control group feature processing sub-model; the prediction probability generation module 501 is specifically used to: input the user features of the predicted user into the promotional information reach gain model; obtain a first predicted user representation and a second predicted user representation corresponding to the user features of the predicted user through the experimental group feature processing sub-model and the control group feature processing sub-model of the promotional information reach gain model, respectively; concatenate the first predicted user representation with a preset user reach feature used to represent that the user has reached the promotional information, and predict a first prediction probability that the predicted user has reached the promotional information; concatenate the second predicted user representation with a preset user reach feature used to represent that the user has not reached the promotional information, and predict a second prediction probability that the predicted user has not reached the promotional information.

[0061] According to the technical solution of the present invention, feature processing is performed on the user characteristics of the experimental group and the user characteristics of the control group respectively to obtain a fact-based first user representation and a counterfactual second user representation corresponding to the user characteristics of the experimental group, and a fact-based third user representation and a counterfactual fourth user representation corresponding to the user characteristics of the control group; the first user representation and the third user representation are respectively concatenated with user reach features to obtain a first concatenated user representation and a third concatenated user representation; model prediction is performed based on the first concatenated user representation to obtain the experimental group prediction loss; model prediction is performed based on the third concatenated user representation to obtain the control group prediction loss; comparative learning is performed based on the first user representation and the second user representation to obtain the experimental group prediction loss. The model employs a group representation loss method, which generates a control group representation loss through comparative learning based on the third and fourth user representations. Model parameters are then adjusted based on the experimental group prediction loss, control group prediction loss, experimental group representation loss, and control group representation loss to train the promotional information reach gain model. This method generates fact-based and counterfactual user representations for the experimental and control group user features, allowing for comparative learning and training of the gain model. This aligns the fact-based and counterfactual user representations of the same sample, thus equalizing the distribution of experimental and control groups at the sample feature level. This addresses the problem of imbalanced sample distribution, improves the accuracy of model predictions, and enhances the effectiveness of intelligent marketing. Furthermore, the introduction of reach features into the generated features used for model prediction prevents reach features from being overwhelmed, further improving the accuracy of model predictions.

[0062] Figure 6 An exemplary system architecture 600 is shown, which can be applied to the training method of the promotional information reach gain model, the method of predicting the reach gain of user on promotional information, or the training device of the promotional information reach gain model and the device of predicting the reach gain of user on promotional information.

[0063] like Figure 6 As shown, system architecture 600 may include terminal devices 601, 602, and 603, a network 604, and a server 605. Network 604 serves as the medium for providing communication links between terminal devices 601, 602, and 603 and server 605. Network 604 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0064] Users can use terminal devices 601, 602, and 603 to interact with server 605 via network 604 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 601, 602, and 603, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0065] Terminal devices 601, 602, and 603 can be various electronic devices with displays and web browsing capabilities, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0066] Server 605 can be a server providing various services, such as a backend management server supporting shopping websites browsed by users using terminal devices 601, 602, and 603 (for example only). The backend management server can perform feature processing, feature concatenation, comparative learning, and model parameter adjustment on the received data such as the characteristics of users in the experimental group and the control group, and feed back the processing results (such as the trained promotional information reach gain model - for example only) to the terminal devices.

[0067] It should be noted that the training method of the promotional information reach gain model or the method of predicting the user's reach gain of the promotional information provided in the embodiments of the present invention are generally executed by the server 605. Correspondingly, the training device of the promotional information reach gain model or the method of predicting the user's reach gain of the promotional information are generally set in the server 605.

[0068] It should be understood that Figure 6 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0069] The following is for reference. Figure 7 It shows a schematic diagram of the structure of a computer system 700 suitable for implementing terminal devices or servers of the present invention. Figure 7 The terminal device or server shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0070] like Figure 7 As shown, the computer system 700 includes a central processing unit (CPU) 701, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 702 or programs loaded from storage section 708 into random access memory (RAM) 703. The RAM 703 also stores various programs and data required for the operation of the system 700. The CPU 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0071] The following components are connected to the I / O interface 705: an input section 706 including a keyboard, mouse, etc.; an output section 707 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a LAN card, modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to the I / O interface 705 as needed. A removable medium 711, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 710 as needed so that computer programs read from it can be installed into the storage section 708 as needed.

[0072] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 709, and / or installed from removable medium 711. When the computer program is executed by central processing unit (CPU) 701, it performs the functions defined above in the system of this invention.

[0073] It should be noted that the computer-readable medium shown in this invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a 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, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0074] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0075] The units or modules described in the embodiments of the present invention can be implemented in software or hardware. The described units or modules can also be housed in a processor; for example, a processor can be described as including a user feature processing module, a prediction loss determination module, a representation loss determination module, and a model parameter adjustment module. The names of these units or modules do not necessarily limit the specific unit or module itself. For example, the model parameter adjustment module can also be described as "a module for adjusting model parameters based on the experimental group prediction loss, the control group prediction loss, the experimental group representation loss, and the control group representation loss to train a promotion information reach gain model."

[0076] In another aspect, the present invention also provides a computer-readable medium, which may be included in the device described in the above embodiments; or it may exist independently and not assembled into the device. The computer-readable medium carries one or more programs, which, when executed by the device, cause the device to include: performing feature processing on experimental group user features and control group user features respectively to obtain a fact-based first user representation and a counterfactual second user representation corresponding to the experimental group user features, and a fact-based third user representation and a counterfactual fourth user representation corresponding to the control group user features; concatenating the first user representation and the third user representation with user reach features respectively to obtain a first concatenated user representation and a third concatenated user representation; performing model prediction based on the first concatenated user representation to obtain an experimental group prediction loss; performing model prediction based on the third concatenated user representation to obtain a control group prediction loss; performing comparative learning based on the first user representation and the second user representation to obtain an experimental group representation loss, and performing comparative learning based on the third user representation and the fourth user representation to obtain a control group representation loss; and adjusting model parameters based on the experimental group prediction loss, the control group prediction loss, the experimental group representation loss, and the control group representation loss to train a promotional information reach gain model.

[0077] According to the technical solution of the present invention, feature processing is performed on the user characteristics of the experimental group and the user characteristics of the control group respectively to obtain a fact-based first user representation and a counterfactual second user representation corresponding to the user characteristics of the experimental group, and a fact-based third user representation and a counterfactual fourth user representation corresponding to the user characteristics of the control group; the first user representation and the third user representation are respectively concatenated with user reach features to obtain a first concatenated user representation and a third concatenated user representation; model prediction is performed based on the first concatenated user representation to obtain the experimental group prediction loss; model prediction is performed based on the third concatenated user representation to obtain the control group prediction loss; comparative learning is performed based on the first user representation and the second user representation to obtain the experimental group prediction loss. The model employs a group representation loss method, which generates a control group representation loss through comparative learning based on the third and fourth user representations. Model parameters are then adjusted based on the experimental group prediction loss, control group prediction loss, experimental group representation loss, and control group representation loss to train the promotional information reach gain model. This method generates fact-based and counterfactual user representations for the experimental and control group user features, allowing for comparative learning and training of the gain model. This aligns the fact-based and counterfactual user representations of the same sample, thus equalizing the distribution of experimental and control groups at the sample feature level. This addresses the problem of imbalanced sample distribution, improves the accuracy of model predictions, and enhances the effectiveness of intelligent marketing. Furthermore, the introduction of reach features into the generated features used for model prediction prevents reach features from being overwhelmed, further improving the accuracy of model predictions.

[0078] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A training method for a promotional information reach gain model, characterized in that, include: Feature processing was performed on the user characteristics of the experimental group and the user characteristics of the control group respectively to obtain the first user representation based on facts and the second user representation based on counterfacts corresponding to the user characteristics of the experimental group, and the third user representation based on facts and the fourth user representation based on counterfacts corresponding to the user characteristics of the control group. The first user representation and the third user representation are concatenated with user reach features to obtain a first concatenated user representation and a third concatenated user representation; model prediction is performed based on the first concatenated user representation to obtain the experimental group prediction loss; model prediction is performed based on the third concatenated user representation to obtain the control group prediction loss. The experimental group representation loss is obtained by comparative learning based on the first user representation and the second user representation, and the control group representation loss is obtained by comparative learning based on the third user representation and the fourth user representation. The model parameters are adjusted based on the experimental group prediction loss, the control group prediction loss, the experimental group representation loss, and the control group representation loss to train the promotion information reach gain model.

2. The method according to claim 1, characterized in that, Feature processing was performed on the user characteristics of the experimental group and the user characteristics of the control group respectively to obtain the fact-based first user representation and counterfactual second user representation corresponding to the user characteristics of the experimental group, and the fact-based third user representation and counterfactual fourth user representation corresponding to the user characteristics of the control group, including: The user characteristics of the experimental group are input into the feature processing sub-model of the experimental group and the feature processing sub-model of the control group, respectively, to obtain the first user representation based on facts and the second user representation based on counterfactual facts corresponding to the user characteristics of the experimental group. By inputting the user characteristics of the control group into the feature processing sub-model of the control group and the feature processing sub-model of the experimental group respectively, the fact-based third user representation and the counterfactual fourth user representation corresponding to the user characteristics of the control group are obtained.

3. The method according to claim 1 or 2, characterized in that, Feature processing was performed on the user characteristics of the experimental group and the user characteristics of the control group, including: For the user characteristics of the experimental group and the user characteristics of the control group, high-dimensional representations are generated by combining user behavior sequences to perform feature abstraction for feature processing.

4. The method according to claim 1, characterized in that, The user reach features were generated based on the user behavior sequences of the experimental group and the user behavior sequences of the control group.

5. The method according to claim 1, characterized in that, The experimental group representation loss is obtained through comparative learning based on the first user representation and the second user representation, and the control group representation loss is obtained through comparative learning based on the third user representation and the fourth user representation, including: The first user representation and the second user representation that belong to the same sample are taken as positive examples of the experimental group, and the first user representation and the second user representation that do not belong to the same sample are taken as negative examples of the experimental group. The experimental group representation loss is obtained by comparative learning based on the positive examples and negative examples of the experimental group. The third and fourth user representations belonging to the same sample are used as positive examples in the control group, and the third and fourth user representations not belonging to the same sample are used as negative examples in the control group. The control group representation loss is obtained by comparative learning based on the positive and negative examples in the control group.

6. The method according to claim 5, characterized in that, The experimental group representation loss is determined based on the similarity between positive examples in the experimental group and the similarity between negative examples in the experimental group; The control group representation loss is determined based on the similarity between positive examples and the similarity between negative examples in the control group.

7. The method according to claim 6, characterized in that, The characterization losses of the experimental group and the control group were calculated using the following formula: ； ； in, To characterize the loss in the experimental group, The control group represents the loss. represents a positive example, Indicates a negative example. It's a hyperparameter. This indicates similarity calculation. This indicates the sample size in the experimental or control group, where k and i are the sample numbers. This represents the user representation output by the i-th sample after passing through the experimental group feature processing sub-model. This represents the user representation output by the i-th sample after passing through the feature processing sub-model of the control group. This represents the user representation output by the k-th sample after passing through the control group feature processing sub-model.

8. The method according to claim 1, characterized in that, Model parameter adjustments are made based on the experimental group prediction loss, the control group prediction loss, the experimental group representation loss, and the control group representation loss, including: The model loss is obtained by weighted summing of the prediction loss of the experimental group, the prediction loss of the control group, the characterization loss of the experimental group, and the characterization loss of the control group. The model parameters are adjusted based on the model loss.

9. A method for predicting user reach gain for promotional information using a promotional information reach gain model trained according to any one of claims 1-8, characterized in that, include: The user characteristics of the predicted user are input into the promotion information reach gain model to obtain the first predicted probability that the predicted user will reach the promotion information and the second predicted probability that the predicted user will not reach the promotion information. The reach gain of the promotional information to the predicted user is calculated based on the first prediction probability and the second prediction probability.

10. The method according to claim 9, characterized in that, The promotional information reach gain model includes an experimental group feature processing sub-model and a control group feature processing sub-model; The user characteristics of the predicted user are input into the promotional information reach gain model to obtain a first predicted probability that the predicted user will reach the promotional information and a second predicted probability that the predicted user will not reach the promotional information, including: Input the predicted user characteristics into the promotional information reach gain model; The first and second predicted user representations corresponding to the user characteristics of the predicted user are obtained by using the experimental group feature processing sub-model and the control group feature processing sub-model of the promotion information reach gain model, respectively. After concatenating the first predicted user representation with the preset user reach features used to represent that the user has reached the promotional information, the first predicted probability of the predicted user reaching the promotional information is predicted. After concatenating the second predicted user representation with a preset user reach feature used to represent that the user does not reach the promotional information, a second predicted probability of the predicted user not reaching the promotional information is predicted.

11. A training device for a promotion information reach gain model, characterized in that, include: The user feature processing module is used to process the user features of the experimental group and the user features of the control group respectively, to obtain the first user representation based on facts and the second user representation based on counterfacts corresponding to the user features of the experimental group, and the third user representation based on facts and the fourth user representation based on counterfacts corresponding to the user features of the control group. The prediction loss determination module is used to concatenate the first user representation and the third user representation with user reach features to obtain a first concatenated user representation and a third concatenated user representation; perform model prediction based on the first concatenated user representation to obtain the experimental group prediction loss; and perform model prediction based on the third concatenated user representation to obtain the control group prediction loss. The representation loss determination module is used to obtain the experimental group representation loss by comparative learning based on the first user representation and the second user representation, and to obtain the control group representation loss by comparative learning based on the third user representation and the fourth user representation. The model parameter adjustment module is used to adjust the model parameters based on the experimental group prediction loss, the control group prediction loss, the experimental group representation loss, and the control group representation loss, so as to train the promotion information reach gain model.

12. An apparatus for predicting user reach gain for promotional information using a promotional information reach gain model trained using the apparatus as described in claim 11, characterized in that, include: The prediction probability generation module is used to input the user characteristics of the predicted user into the promotion information reach gain model to obtain a first prediction probability that the predicted user will reach the promotion information and a second prediction probability that the predicted user will not reach the promotion information. The reach gain calculation module is used to calculate the reach gain of the predicted user to the promotional information based on the first prediction probability and the second prediction probability.

13. An electronic device, characterized in that, include: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-10.

14. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-10.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-10.

Images