Gain modeling method for recommendation system in presence of unobserved intermixture

Abstract

The invention relates to a gain modeling method for a recommendation system in the presence of non-observation mixing, belongs to the technical field of information recommendation, and solves the problem that the gain cannot be accurately estimated in the presence of non-observation mixing in the prior art. The method comprises the following steps: constructing a first training set and a second training set; the samples in the first training set are samples obtained by observability research; samples in the second training set are samples obtained by a random control test; each sample comprises a user article pair feature, a sample processing scheme and a real outcome of whether the user purchases the article or not; constructing a pre-training model, wherein the pre-training model is used for predicting potential outcomes of the sample under different processing schemes; training a pre-training model based on the first training set to obtain a trained pre-training model; and constructing a fine tuning model based on the trained pre-training model, and training the fine tuning model based on the second training set to obtain a gain prediction model for eliminating the unobserved hybrid influence. And accurate unbiased gain estimation is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and in particular to a gain modeling method for a recommendation system in the presence of unobserved confounding. Background Art

[0002] CATE (conditional average treatment effect) refers to the average causal effect of an intervention on the outcome of interest given a feature, and plays an important role in many fields such as e-commerce, healthcare, and economics. In a recommendation system, the platform hopes to predict how recommending a specific product or advertisement to a specific user affects the probability of purchase, thereby affecting the total profit. It is worth noting that predicting the conditional average causal effect is very different from traditional machine learning prediction problems. For example, for a specific user, if the user is pushed an advertisement, the probability of the user buying the product is 0.9. The traditional prediction problem will only focus on predicting the probability of the user buying the product to be 0.9, and will not pay attention to the probability of the user buying the product if the advertisement is not pushed to the user. If the probability of the user buying the product is also 0.9 when the advertisement is not pushed to the user, the causal effect estimated by the causal method will be 0.9-0.9=0, so the advertisement will not be recommended to the user. However, the traditional prediction problem will consider the probability of the user buying the product to be 0.9, which is a very large probability, and thus recommend advertisements to the user, resulting in a waste of resources. By estimating or predicting the conditional average causal effect, i.e., gain, of users with and without intervention, intervention measures (such as push advertising and coupons) can be implemented on the target population more efficiently and accurately.

[0003] A basic problem in estimating CATE is that we can only observe one outcome for a specific individual. For example, for a specific user, if the e-commerce company pushes ads, we can only observe the purchase outcome of the user being pushed ads. Conversely, if the e-commerce company does not push ads, we can only observe the purchase outcome of the user not being pushed ads. Since we can only observe one outcome, we need to use causal inference methods to estimate CATE, rather than just using traditional machine learning prediction methods.

[0004] With the significant progress of reinforcement learning and deep learning, in order to improve the accuracy of CATE estimation, representation-based learning methods have received increasing attention due to their excellent performance. These methods focus on generating feature representations, with the goal of mitigating confounding bias by minimizing the difference in feature representation distribution between the intervention group and the control group. In order to obtain this representation, previous methods have developed a lot of theories and carried out extensive practice. For example, some of them use IPM (integral probability metric) for regularization, some methods emphasize local similarity preservation, and some methods use techniques such as target learning and optimal transmission to estimate CATE. However, the above methods ignore unobserved confounding factors, which are very common in real-world scenarios. For example, in recommendation systems, users' financial income status is sensitive and usually cannot be collected. These unobserved variables may affect treatment options and results, resulting in confounding bias in causal effect estimates. Summary of the invention

[0005] In view of the above analysis, an embodiment of the present invention aims to provide a gain modeling method for a recommendation system in the presence of unobserved confounding, so as to solve the problem that the gain cannot be accurately estimated in the presence of unobserved confounding.

[0006] On the one hand, an embodiment of the present invention provides a gain modeling method for a recommendation system in the presence of unobserved confounding, comprising the following steps:

[0007] Constructing a first training set and a second training set; the samples in the first training set are samples obtained from observational studies; the samples in the second training set are samples obtained from randomized controlled trials; each sample includes user-item pair features, sample processing schemes, and the actual outcome of whether the user purchases the item;

[0008] Constructing a pre-trained model, the pre-trained model is used to predict the potential outcome of the sample under different treatment schemes; training the pre-trained model based on the first training set to obtain a trained pre-trained model;

[0009] A fine-tuning model is constructed based on the trained pre-training model, and the fine-tuning model is trained based on the second training set to obtain a gain prediction model that eliminates the influence of unobserved confounding.

[0010] Based on the further improvement of the above method, the pre-training model includes a first representation module, a reconstruction module and a first prediction module;

[0011] The first characterization module is used to extract a first characterization of a sample based on the feature of the user item;

[0012] The reconstruction module is used to reconstruct the sample based on the first representation;

[0013] The first prediction module is used to predict the potential outcome of the sample under different treatment schemes based on the first representation.

[0014] Based on the further improvement of the above method, a fine-tuning model is constructed based on the trained pre-trained model, including:

[0015] Fix the parameters of the first representation module of the pre-trained model and use it as a module of the fine-tuned model;

[0016] Adding a second characterization module and a second prediction module based on the first characterization module;

[0017] The second characterization module is used to extract a second characterization of the sample based on the feature of the user-item pair;

[0018] The second prediction module is used to predict the potential outcomes of the sample under different treatment schemes based on the first representation and the second representation, and obtain a gain prediction result that eliminates the influence of unobserved confounding;

[0019] Initialize parameters of the second prediction module based on the first prediction module.

[0020] Based on the further improvement of the above method, the following formula is used to calculate the training loss of the fine-tuning model:

[0021]

[0022] in, represents the prediction loss of the second prediction module, represents the mutual information loss between the first representation and the second representation, represents the parameter deviation loss, and λ3 and λ4 represent the weights.

[0023] Based on the further improvement of the above method, the mutual information loss of the first representation and the second representation is calculated using the following formula:

[0024]

[0025]

[0026] Among them, q θ (·|·) represents the variational approximation of p(·|·), p(·|·) represents the conditional probability, Z represents the first representation variable of the current training batch samples, and Z i represents the first representation of the i-th sample in the current training batch, Z U Represents the second representation variable of the current training batch sample, Z U i represents the second representation of the i-th sample in the current training batch, and N represents the number of samples in the current training batch.

[0027] Based on the further improvement of the above method, the following formula is used to calculate the parameter deviation loss:

[0028]

[0029] in, represents the initial parameters of the first prediction head of the second prediction module, represents the initial parameters of the second prediction head of the second prediction module, represents the current parameters of the first prediction head of the second prediction module, represents the current parameters of the second prediction head of the second prediction module, and ||·||2 represents the l2 norm.

[0030] Based on the further improvement of the above method, the prediction loss of the second prediction module is calculated using the following formula:

[0031]

[0032] Where N is the number of samples in the current training batch, l(·,·) is the loss function, and Y i represents the true outcome of the i-th sample in the current training batch, represents the potential outcome of the i-th sample predicted by the second prediction module, w i Represents the weight of the i-th sample T i represents the processing scheme of the i-th sample, n represents the number of samples in the second training set, T j Represents the processing scheme for the jth sample in the second training set.

[0033] Based on the further improvement of the above method, the training loss of the pre-trained model is calculated using the following formula:

[0034]

[0035] in, represents the prediction loss of the first prediction module, represents the reconstruction loss, represents the feature imbalance loss, and λ1 and λ2 represent the weights.

[0036] Based on the further improvement of the above method, the prediction loss of the first prediction module is calculated using the following formula:

[0037]

[0038] Where N is the number of samples in the current training batch, l(·,·) is the loss function, and Y i represents the true outcome of the i-th sample in the current training batch, represents the potential outcome of the i-th sample predicted by the first prediction module, w i represents the weight of the i-th sample, represents the processing scheme of the i-th sample, m represents the number of samples in the first training set, T j Represents the processing scheme for the jth sample in the first training set.

[0039] Based on a further improvement of the above method, the feature imbalance loss is obtained by calculating the integral probability measure between the first representations of the samples with the intervention treatment scheme and the samples with the non-intervention treatment scheme in the current training batch.

[0040] Compared with the prior art, the present invention constructs the first training set and the second training set by extracting samples from observational studies and self-controlled trials respectively, and uses a two-stage framework, that is, first training the pre-training model based on the samples of observational studies to obtain the basic representation of the features to estimate the counterfactual results, and then adjusting the deviation of the representation learned from the observational study data based on the samples of small-scale randomized controlled trials, so as to obtain an unbiased gain prediction model that eliminates the influence of unobserved confounding and avoids the overfitting problem. According to the trained gain prediction model, the intervention gain is accurately predicted, so that the intervention measures can be accurately implemented according to the predicted gain, improving efficiency and optimizing resource allocation.

[0041] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0043] Figure 1 A flow chart of a gain modeling method for a recommendation system when unobserved confounding exists in an embodiment of the present invention;

[0044] Figure 2 Schematic diagram of a two-stage CATE estimation framework according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0046] Unobserved confounding refers to the presence of unobserved variables in the study that affect the relationship between intervention and outcome, which may lead to bias in CATE estimation. Observational studies (OBS) are data collected naturally in real life. Due to the presence of unobserved confounding in these data, there will be bias in CATE estimation based solely on samples obtained from observational studies. Randomized controlled trial (RCT) data are considered the gold standard for causal effect estimation. Randomized controlled trial data are artificially randomly assigned treatment plans (i.e., individuals are in the intervention group or the control group). Data from randomized controlled trials (RCTs) are not affected by confounding. However, the collection of these data is costly and usually has a small sample size. Directly training a causal effect prediction model on RCT data alone faces the risk of overfitting due to insufficient RCT sample size.

[0047] In order to accurately estimate the gain, a specific embodiment of the present invention discloses a gain modeling method for a recommendation system in the presence of unobserved confounding, such as Figure 1 As shown, the following steps are included:

[0048] S1. Construct a first training set and a second training set; the samples in the first training set are samples obtained from observational studies; the samples in the second training set are samples obtained from randomized controlled trials; each sample includes user-item pair features, sample processing schemes, and the actual outcome of whether the user purchases the item;

[0049] S2. construct a pre-training model, wherein the pre-training model is used to predict the potential outcome of the sample under different treatment schemes; and train the pre-training model based on the first training set to obtain a trained pre-training model;

[0050] S3. Construct a fine-tuning model based on the trained pre-training model, and train the fine-tuning model based on the second training set to obtain a gain prediction model that eliminates the influence of unobserved confounding.

[0051] Compared with the prior art, the gain modeling method of the recommendation system in the presence of unobserved confounding provided in this embodiment constructs the first training set and the second training set by extracting samples from observational studies and self-controlled trials respectively, and uses a two-stage framework, that is, first training the pre-training model based on the samples of the observational study to obtain the basic representation of the features to estimate the counterfactual results, and then adjusting the deviation of the representation learned from the observational study data based on the samples of the small-scale randomized controlled trial, so as to obtain an unbiased gain prediction model that eliminates the influence of unobserved confounding and avoids the overfitting problem. The intervention gain is accurately predicted according to the trained gain prediction model, so that the intervention measures can be accurately implemented according to the predicted gain, improving efficiency and optimizing resource allocation.

[0052] In addition, the prior art often performs gain estimation under strong assumptions, for example, assuming that the statistical characteristics of all individuals without unobserved confounding are similar to those with unobserved confounding, independent assumptions, and linear additivity generation assumptions. These assumptions cannot be tested through data and are easily violated in practice. The present invention does not rely on these assumptions and does not restrict the data generation process (for example, linear or additive noise assumptions). Therefore, it is more in line with actual application conditions. The present invention is robust to unobserved confounding of different intensities and different proportions of RCT data.

[0053] During implementation, each sample contains the user-item pair feature X, the treatment plan T (T∈{0,1}, T=1 represents the intervention group, where intervention measures are implemented on the user, T=0 represents the control group, where no intervention measures are implemented on the user), and the true outcome Y∈{0,1} (Y=1 represents that the user purchased the item, and Y=0 represents that the user did not purchase the item). In addition, G is used to represent the set to which the sample belongs. Indicates that the sample belongs to the OBS set. Indicates that the sample belongs to the RCT set, that is, the second training set, and has no confounding bias. The i-th sample is represented by D i =(X i ,T i ,Y i ,G i ). According to the Neyman-Rubin potential outcome framework, Y(1) and Y(0) are defined as potential outcomes, where Y(0) represents the potential outcome without the implementation of the intervention measures, and Y(1) represents the potential outcome with the implementation of the intervention measures. For a specific individual, only one outcome can be observed, so the true outcome Y of the sample is one of the potential outcomes Y(1) or Y(0).

[0054] During implementation, the user-item pair feature may be a feature obtained based on user attributes and item attributes.

[0055] The propensity score, that is, the probability of the sample implementing the intervention, is e(X,G)=P(T=1||(X,G)), and the conditional mean outcome is μ(X,G)=E[Y|X,G]. CATE is defined as the conditional expected difference of potential outcomes, expressed as follows:

[0056] τ(x)=E[Y(1)-Y(0)∣X=x].

[0057] That is, the mean difference between the potential outcomes of implementing the intervention and the potential outcomes of not implementing the intervention when the sample characteristics are x.

[0058] In order to identify CATE from observed data, in addition to SUTVA (stable unit treatment value assumption), the following two assumptions are required:

[0059] Assumption 1: Ignorability (Y(1), Y(0)) ⊥ T | X

[0060] This means that, given characteristic X, the potential outcomes are independent of the intervention assignment.

[0061] Assumption 2: Positivity

[0062] This means that for all features X, in both RCT and OBS data, the probability of implementing an intervention e(X,G) must be between 0 and 1, ensuring that each individual has the possibility of receiving a different treatment.

[0063] Assumption 1 is also known as no unobserved confounding, which is assumed to be true in RCTs due to random treatment assignment. We can identify and calculate CATE based on RCT data as follows:

[0064]

[0065] Compared with RCT data, there may be unobserved confounding in OBS data. Therefore, for observational data, we do not assume no confounding, that is, the following relationship does not hold:

[0066]

[0067] This means that we may not be able to identify τ(x) based on the OBS data alone. We express the difference in conditional mean results in the observed data as:

[0068]

[0069] Due to unobserved confounding, ω(x)≠τ(x) holds for any x. The difference between these two quantities is the confounding effect, which we call the residual function and express as:

[0070] n(x)=τ(x)-(x).

[0071] The residual function η(x) captures the deviation between CATE and the conditional mean result in the observed data due to unobserved confounding.

[0072] Therefore, the residual function is defined as follows:

[0073]

[0074]

[0075] Here, η t (x) represents the residual between the OBS data and the RCT data in the intervention group, and η c (x) represents the residual between OBS data and RCT data in the control group. These two residual functions reflect the impact of unobserved confounding on different treatment groups.

[0076] Based on the above analysis, the present invention adopts a two-stage CATE estimation framework of pre-training-fine-tuning paradigm, such as Figure 2 As shown, in the first stage, large-scale OBS data is used to train the basic representation of features, and in the second stage, relatively small-scale unbiased RCT data is used to calibrate the bias in the representation learned from the OBS data to train an unbiased prediction model.

[0077] During implementation, the first stage trains a pre-trained model based on the first training set. Specifically, the pre-trained model includes a first representation module, a reconstruction module and a first prediction module;

[0078] The first characterization module is used to extract a first characterization of a sample based on the feature of the user item;

[0079] The reconstruction module is used to reconstruct the sample based on the first representation;

[0080] The first prediction module is used to predict the potential outcome of the sample under different treatment schemes based on the first representation.

[0081] The goal of the pre-training module is to accurately estimate the potential outcomes of the OBS data. Therefore, the deep features of the sample, namely the first representation, are first extracted through the first representation module, and then the first prediction module is used to predict the potential outcomes based on the deep features. When implemented, the first representation module can use a multi-layer feedforward neural network, represented as φ. The first representation module obtains the first representation Z of the feature X of the intervention group and the control group. In other words, for a single sample Regardless of T i =0 or T i =1, representing Z i =φ(X i ) are the same. To better estimate the causal effect, we adopt the idea of ​​feature balancing in the representation space proposed by Shalit et al. Specifically, the representation of the intervention group is considered to be distributed The control group is similarly considered to be an independent and identically distributed sample randomly drawn from We expect the distributions of representations for the intervention and control groups to be similar.

[0082] In implementation, the first prediction module includes two prediction heads, denoted as h0 and h1, which respectively predict the potential outcomes of the sample under treatment scheme T = 0 and treatment scheme T = 1. The prediction head adopts a multi-layer perceptron.

[0083] In order to ensure that the extracted first representation retains as much information about the original features as possible, a reconstruction module is used to reconstruct samples based on the first representation. In implementation, the reconstruction module uses a decoder network ψ to reconstruct the original features

[0084] By adding a reconstruction module, the learned first representation can contain information in the features, rather than just the information required to fit the predicted outcome Y, thereby enhancing the generalization ability of the first representation module.

[0085] Specifically, the training loss of the pre-trained model is calculated using the following formula:

[0086]

[0087] in, represents the prediction loss of the first prediction module, represents the reconstruction loss, represents the feature imbalance loss, and λ1 and λ2 represent the weights.

[0088] Specifically, the prediction loss of the first prediction module is calculated using the following formula:

[0089]

[0090] Where N is the number of samples in the current training batch, l(·,·) is the loss function, and Y i represents the true outcome of the i-th sample in the current training batch, represents the potential outcome of the i-th sample predicted by the first prediction module, w i represents the weight of the i-th sample. When implemented, the loss function l(·,·) can use cross entropy loss.

[0091] It should be noted that the two prediction heads of the first prediction module predict the potential outcomes of the sample in the intervention group and the control group respectively, and the actual outcome of the sample is one of them. Therefore, if the sample is actually in the intervention group, then Take the prediction result of prediction head h1. If the sample is actually in the control group, then Take the prediction result of prediction head h0.

[0092] In implementation, the weight of the i-th sample is calculated using the following formula:

[0093]

[0094] Where m represents the number of samples in the first training set, T j represents the processing scheme of the jth sample in the first training set, T i Represents the processing scheme for the i-th sample.

[0095] In order to balance the features between the intervention group and the control group, the feature imbalance loss is obtained by calculating the integral probability measure between the first representations of the samples with intervention and the samples with no intervention in the current training batch.

[0096] After the pre-trained model is trained based on the first training set to obtain a trained pre-trained model, a fine-tuning model is constructed in the second stage to obtain a gain prediction model that eliminates the influence of unobserved confounding.

[0097] Specifically, a fine-tuning model is constructed based on the trained pre-trained model, including:

[0098] Fix the parameters of the first representation module of the pre-trained model and use it as a module of the fine-tuned model;

[0099] Adding a second characterization module and a second prediction module based on the first characterization module;

[0100] The second characterization module is used to extract a second characterization of the sample based on the feature of the user-item pair;

[0101] The second prediction module is used to predict the potential outcomes of the sample under different treatment schemes based on the first representation and the second representation, and obtain a gain prediction result that eliminates the influence of unobserved confounding;

[0102] Initialize parameters of the second prediction module based on the first prediction module.

[0103] In the second stage, the fine-tuning model keeps the bias representation Z generated by φ unchanged, but regards it as part of the representation, while introducing another representation module (the second representation module) φ U Generated additional representation (second representation) Z U , the final representation is adjusted using the second representation to make it unbiased.

[0104] When implemented, the second representation module can use a feedforward neural network, φ U The width and depth of φ can be adjusted according to the scale of RCT data. If the scale of RCT data is comparable to that of OBS data, we can use the same architecture as φ. In reality, RCT data is relatively scarce compared to OBS data, so φ U The scale should be smaller.

[0105] Similarly, the fine-tuning model predicts the potential outcomes of the samples under different treatment schemes based on the first representation and the second representation through the second prediction module. The second prediction module includes a first prediction head g0 and a second prediction head g1. The first prediction head g0 is used to predict the potential outcomes of the samples in the control group, and the second prediction head g1 is used to predict the potential outcomes of the samples in the intervention group.

[0106] It should be noted that g0,g1 and h0,h1 have the same depth l p , but the width of each layer of g0,g1 is greater than h0,h1.

[0107] When implemented, the fine-tuning model architecture of the second phase is different from that of the first phase. Therefore, the initialization of the model parameters is crucial to the effectiveness of the fine-tuning phase. The goal of initialization is to ensure that the model produces the same predictions as the pre-trained model trained in the first phase. Therefore, the parameters of the second prediction module are initialized based on the first prediction module. That is, the prediction head g0 of the second prediction module is initialized based on the prediction head h0 of the first prediction module, and the prediction head g1 of the second prediction module is initialized based on the prediction head h1 of the first prediction module.

[0108] To simplify the representation, g is used t Represents the prediction head of the second prediction module. t represents the prediction head of the first prediction module, t∈{0,1}.

[0109] When implementing, first set g t The parameters are divided into blocks:

[0110]

[0111] in, Indicates g t The weight parameters from layer l-1 to layer l are: Yes t The bias parameter of the lth layer, The shape and same. Indicates h t The weight parameters from layer l-1 to layer l are: Yes t Bias parameter of the lth layer. l=1,2,…,l p .

[0112] Based on the partitioning of parameters, the following initialization strategy is proposed:

[0113]

[0114] Upcoming parameters Initialize to Parameters, parameters Initialize to The other parameters are initialized to 0.

[0115] In implementation, the rectified linear unit (ReLU) is used as the activation function so that t Deactivate and h t Compared to the extra neurons.

[0116] The fine-tuned model is then trained based on the second training set.

[0117] Specifically, the following formula is used to calculate the training loss of the fine-tuning model:

[0118]

[0119] in, represents the prediction loss of the second prediction module, represents the mutual information loss between the first representation and the second representation, represents the parameter deviation loss, and λ3 and λ4 represent the weights.

[0120] Specifically, the prediction loss of the second prediction module is calculated using the following formula:

[0121]

[0122] Where N is the number of samples in the current training batch, l(·,·) is the loss function, and Y i represents the true outcome of the i-th sample in the current training batch, represents the potential outcome of the i-th sample predicted by the second prediction module, w i Represents the weight of the i-th sample n represents the number of samples in the second training set, T i represents the processing scheme of the i-th sample, T j Represents the processing scheme for the jth sample in the second training set.

[0123] Similarly, if the sample is actually in the intervention group, then Take the prediction result of the second prediction head g1. If the sample is actually in the control group, then Take the prediction result of the first prediction head g0.

[0124] To ensure that Z U Capture the features different from Z and use mutual information to control the overlap between the two, that is, add the mutual information loss of the first representation and the second representation to the loss function. Specifically, the following formula is used to calculate the mutual information loss of the first representation and the second representation:

[0125]

[0126]

[0127] Among them, q θ (·|·) represents the variational approximation of p(·|·), p(·|·) represents the conditional probability, Z represents the first representation variable of the current training batch samples, and Z i represents the first representation of the i-th sample in the current training batch, Z U Represents the second representation variable of the current training batch sample, Z U i represents the second representation of the i-th sample in the current training batch, and N represents the number of samples in the current training batch.

[0128] It is assumed that unobserved confounding factors may cause the distribution of OBS data to be slightly offset from that of RCT data. Therefore, the fine-tuned model in the second stage should not deviate significantly from the pre-trained model in the first stage. The initial parameters of the fine-tuned model are denoted as In order to constrain the deviation from the initial value, the parameter deviation loss is added to the loss function. Specifically, the parameter deviation loss is calculated using the following formula:

[0129]

[0130] in, represents the initial parameters of the first prediction head of the second prediction module, represents the initial parameters of the second prediction head of the second prediction module, represents the current parameters of the first prediction head of the second prediction module, represents the current parameters of the second prediction head of the second prediction module, and ||·||2 represents the l2 norm.

[0131] Since there is no unobserved confounding in RCT data, the unobserved confounding factors are adjusted according to the RCT data in the second stage to calibrate the representation learned from the OBS data, eliminate the influence of unobserved confounding, and obtain a gain prediction model that eliminates the influence of unobserved confounding. In order to avoid overfitting caused by small-scale RCT data in the second stage, the network parameters are partially initialized in the second stage, using the parameters of the pre-trained network in the first stage instead of training an independent network. Compared with the existing technology, our method has the advantage of not restricting the data generation process (for example, linear or additive noise assumptions) and can effectively alleviate the overfitting problem.

[0132] For the user-item pair to be processed, its features are input into the gain prediction model. The second prediction module predicts its potential outcomes in the prediction group and the control group. The difference in potential outcomes between the prediction group and the control group is the intervention gain. Based on the intervention gain, it can be determined whether to intervene in the user. For example, if the intervention gain exceeds the predicted threshold, intervention measures are implemented for the user, otherwise no intervention measures are implemented.

[0133] The present invention significantly improves the estimation accuracy of the conditional average causal effect (CATE) by combining large-scale observational study data with small-scale randomized controlled trial (RCT) data. It can effectively reduce the bias caused by unmeasured confounding factors and provide policymakers with a more reliable basis for causal inference, thereby optimizing resource allocation and intervention measures.

[0134] During implementation, the Jobs dataset is used to verify the method of the present invention. The Jobs dataset is a commonly used benchmark dataset in the field of causal gain modeling, which includes observational data (OBS data) and randomized controlled trial data (RCT data). The binary treatment variable T in the Jobs dataset i Indicates whether the i-th user receives job training (T i =1 means receiving training, T i =0 means no training), the outcome variable Y i is their employment status after training (Y i =1 means employment, Y i =0 means unemployed). Covariates Including age, education level, previous income, etc. During implementation, OBS data and a small amount of RCT data are used as training sets to obtain a gain prediction model according to the method of the present invention, and other RCT data are used as validation sets for effect verification.

[0135] During implementation, we use the common evaluation index on the Jobs dataset, namely, strategy risk R pol , to evaluate the effects of the algorithm of the present invention and the existing algorithms. Specifically, R pol is defined as:

[0136]

[0137] Express expectations, represents the probability, Y(1) represents the actual result of the sample implementing the interference measures, represents the potential outcome of the predicted sample when the intervention is implemented, It indicates the potential outcome of the predicted sample when no intervention measures are implemented. T = 1 indicates that the sample is in the intervention group, and T = 0 indicates that the sample is in the control group. It is worth noting that R Pol It is measured on randomized controlled data, so it is an unbiased evaluation criterion. The specific experimental results are shown in Table 1 below, where the out-of-sample indicates the evaluation results on the unbiased data of the test set, and the in-sample indicates the evaluation results on the unbiased data of the training set.

[0138] Table 1 Comparison of the effects of different methods

[0139]

[0140] We consider three types of baseline methods: the first type is based on meta-learning, including T-learner methods, the second type is based on tree models, including Causal Forest, and the third type is based on representation learning, including TarNet, DragonNet and ESCFR. It can be found that the risk of our proposed method in both sample and out-of-sample is statistically significantly lower than that of the baseline method. This is because we can use RCT data more reasonably, and the flexible model structure enables our method to avoid the risk of overfitting.

[0141] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0142] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A gain modeling method for a recommendation system in the presence of unobserved confounding, characterized in that: The following steps are involved: Constructing a first training set and a second training set; the samples in the first training set are samples obtained from observational studies; the samples in the second training set are samples obtained from randomized controlled trials; each sample includes user-item pair features, sample processing schemes, and the actual outcome of whether the user purchases the item; Constructing a pre-trained model, the pre-trained model is used to predict the potential outcome of the sample under different treatment schemes; training the pre-trained model based on the first training set to obtain a trained pre-trained model; A fine-tuning model is constructed based on the trained pre-training model, and the fine-tuning model is trained based on the second training set to obtain a gain prediction model that eliminates the influence of unobserved confounding.

2. The gain modeling method for a recommendation system in the presence of unobserved confounding according to claim 1, characterized in that: The pre-training model includes a first representation module, a reconstruction module and a first prediction module; The first characterization module is used to extract a first characterization of a sample based on the feature of the user item; The reconstruction module is used to reconstruct the sample based on the first representation; The first prediction module is used to predict the potential outcome of the sample under different treatment schemes based on the first representation.

3. The gain modeling method for a recommendation system in the presence of unobserved confounding according to claim 2, characterized in that: Build a fine-tuning model based on the trained pre-trained model, including: Fix the parameters of the first representation module of the pre-trained model and use it as a module of the fine-tuned model; Adding a second characterization module and a second prediction module based on the first characterization module; The second characterization module is used to extract a second characterization of the sample based on the feature of the user-item pair; The second prediction module is used to predict the potential outcomes of the sample under different treatment schemes based on the first representation and the second representation, and obtain a gain prediction result that eliminates the influence of unobserved confounding; Initialize parameters of the second prediction module based on the first prediction module.

4. The gain modeling method for a recommendation system in the presence of unobserved confounding according to claim 3, characterized in that: The training loss of the fine-tuned model is calculated using the following formula: in, represents the prediction loss of the second prediction module, represents the mutual information loss between the first representation and the second representation, represents the parameter deviation loss, and λ3 and λ4 represent the weights.

5. The gain modeling method for a recommendation system in the presence of unobserved confounding according to claim 4, characterized in that: The mutual information loss between the first representation and the second representation is calculated using the following formula: Among them, q θ (·|·) represents the variational approximation of p(·|·), p(·|·) represents the conditional probability, Z represents the first representation variable of the current training batch samples, and Z i represents the first representation of the i-th sample in the current training batch, Z U Represents the second representation variable of the current training batch sample, Z U i represents the second representation of the i-th sample in the current training batch, and N represents the number of samples in the current training batch.

6. The gain modeling method for a recommendation system in the presence of unobserved confounding according to claim 4, characterized in that: The parameter deviation loss is calculated using the following formula: in, represents the initial parameters of the first prediction head of the second prediction module, represents the initial parameters of the second prediction head of the second prediction module, represents the current parameters of the first prediction head of the second prediction module, represents the current parameters of the second prediction head of the second prediction module, and ||·||2 represents the l2 norm.

7. The gain modeling method for a recommendation system in the presence of unobserved confounding according to claim 4, characterized in that: The prediction loss of the second prediction module is calculated using the following formula: Where N is the number of samples in the current training batch, l(·,·) is the loss function, and Y i represents the true outcome of the i-th sample in the current training batch, represents the potential outcome of the i-th sample predicted by the second prediction module, w i represents the weight of the i-th sample, T i represents the processing scheme of the i-th sample, n represents the number of samples in the second training set, T j Represents the processing scheme for the jth sample in the second training set.

8. The gain modeling method for a recommendation system in the presence of unobserved confounding according to claim 2, characterized in that: The training loss of the pre-trained model is calculated using the following formula: in, represents the prediction loss of the first prediction module, represents the reconstruction loss, represents the feature imbalance loss, and λ1 and λ2 represent the weights.

9. The gain modeling method for a recommendation system in the presence of unobserved confounding according to claim 8, characterized in that: The prediction loss of the first prediction module is calculated using the following formula: Where N is the number of samples in the current training batch, l(·,·) is the loss function, and Y i represents the true outcome of the i-th sample in the current training batch, represents the potential outcome of the i-th sample predicted by the first prediction module, w i represents the weight of the i-th sample, T i represents the processing scheme of the i-th sample, m represents the number of samples in the first training set, T j Represents the processing scheme for the jth sample in the first training set.

10. The gain modeling method for a recommendation system in the presence of unobserved confounding according to claim 8, characterized in that: The feature imbalance loss is obtained by calculating the integral probability metric between the first representations of the samples with intervention and the samples with no intervention in the current training batch.

Images