Information recommendation method and device

Abstract

The embodiments of this specification provide an information recommendation method and apparatus. The information recommendation method includes: obtaining the object information of a target object under a target task and each piece of information to be recommended; extracting first target object features that are not affected by the environment from the object information, and for any piece of information to be recommended, predicting a prediction result of the piece of information to be recommended affecting the target object to perform the target task according to the first target object features and the piece of information to be recommended; determining a target recommended information among the pieces of information to be recommended based on the prediction results, and recommending the target recommended information to the target object. This solution can improve the accuracy of the relationship analysis result.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of computer technology, and particularly to an information recommendation method. Background Art

[0002] In information recommendation, relationship analysis can be performed on user behavior and each piece of information to be recommended, that is, causal effect prediction: predicting the probability distribution of a user performing a specified behavior after each piece of information to be recommended is respectively recommended to the user. Furthermore, according to the causal effect value, the target recommended information to be recommended to the user is determined, so that the target recommended information is more likely to be processed as expected by the user, such as viewed. In this way, the effectiveness of information recommendation can be ensured.

[0003] However, user behavior is easily affected by the environment, such as holidays, and the model used for causal effect prediction is trained using sample user behavior. Therefore, once the environment during prediction is inconsistent with the environment of the sample user behavior, the prediction result will be inaccurate, which also leads to the problem of inaccurate information recommendation. Therefore, a more accurate solution is needed. Summary of the Invention

[0004] In view of this, the embodiments of this specification provide an information recommendation method. One or more embodiments of this specification also relate to an information recommendation device, a computing device, a computer-readable storage medium, and a computer program, so as to solve the technical defects existing in the prior art.

[0005] According to the first aspect of the embodiments of this specification, an information recommendation method is provided, including:

[0006] Obtaining the object information of the target object under the target task, and each piece of information to be recommended;

[0007] Extracting the first target object feature that is not affected by the environment from the object information, and for any piece of information to be recommended, predicting the prediction result of this piece of information to be recommended affecting the target object to perform the target task according to the first target object feature and this piece of information to be recommended;

[0008] Based on each prediction result, determining the target recommended information among the pieces of information to be recommended, and recommending the target recommended information to the target object.

[0009] According to the second aspect of the embodiments of this specification, an information recommendation device is provided, including:

[0010] An information acquisition module, configured to obtain the object information of the target object under the target task, and each piece of information to be recommended;

[0011] A prediction module, configured to extract first target object features that are not affected by the environment from the object information, and for any piece of information to be recommended, predict a prediction result of the information to be recommended affecting the target object to perform the target task according to the first target object features and the information to be recommended;

[0012] An information recommendation module, configured to determine a target recommended information among the pieces of information to be recommended based on the analysis results of each prediction relationship, and recommend the target recommended information to the target object.

[0013] According to a third aspect of the embodiments of the present specification, a computing device is provided, including:

[0014] A memory and a processor;

[0015] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the above information recommendation method are implemented.

[0016] According to a fourth aspect of the embodiments of the present specification, a computer-readable storage medium is provided, which stores computer-executable instructions. When the instructions are executed by a processor, the steps of the above information recommendation method are implemented.

[0017] According to a fifth aspect of the embodiments of the present specification, a computer program is provided. When the computer program is executed on a computer, the computer is made to execute the steps of the above information recommendation method.

[0018] An embodiment of the present specification realizes obtaining object information of a target object under a target task and each piece of information to be recommended; extracting first target object features that are not affected by the environment from the object information, and for any piece of information to be recommended, predicting a prediction result of the information to be recommended affecting the target object to perform the target task according to the first target object features and the information to be recommended; determining a target recommended information among the pieces of information to be recommended based on each prediction result, and recommending the target recommended information to the target object.

[0019] In this way, it can be ensured that the first target object feature that is not affected by the environmental variables can be split from the object information of the target object under the target task, and the causal effect prediction of the user feature and the information to be recommended can be performed based on the first target object feature that is not affected by the environment, so as to achieve the effect of cutting off the influence of the object feature affected by the environmental variables on the prediction result. Therefore, the influence of the environmental variables on the prediction result can be reduced, and the accuracy of the prediction result can be improved. On this basis, based on each prediction result, the target recommended information in each piece of information to be recommended is determined, and the target recommended information is recommended to the target object, which can achieve a more accurate information recommendation with a reduced influence of the environmental variables. Therefore, the present solution can achieve a more accurate and applicable effect in more environments. Description of the Drawings

[0020] Figure 1 is a flowchart of a method for training a relationship analysis model provided by an embodiment of this specification;

[0021] Figure 2 is an example diagram of an application scenario of a method for training a relationship analysis model provided by an embodiment of this specification;

[0022] Figure 3 is a schematic diagram of a causal diagram of a relationship analysis model provided by an embodiment of this specification;

[0023] Figure 4 is a schematic diagram of the structure of a decoder in a method for training a relationship analysis model provided by an embodiment of this specification;

[0024] Figure 5 is a schematic diagram of the structure of an encoder in a method for training a relationship analysis model provided by an embodiment of this specification;

[0025] Figure 6 is a flowchart of an information recommendation method provided by an embodiment of this specification;

[0026] Figure 7 is an example diagram of an application scenario of an information recommendation method provided by an embodiment of this specification;

[0027] Figure 8 is a schematic diagram of the structure of an information recommendation device provided by an embodiment of this specification;

[0028] Figure 9 is a block diagram of the structure of a computing device provided by an embodiment of this specification. Detailed Embodiments

[0029] In the following description, numerous specific details are set forth in order to provide a thorough understanding of this specification. However, this specification can be implemented in many other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the spirit of this specification. Therefore, this specification is not limited by the specific implementations disclosed below.

[0030] The terms used in one or more embodiments of this specification are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of this specification. The singular forms “a,” “the,” and “said” used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly dictates otherwise. It should also be understood that the term “and / or” used in one or more embodiments of this specification refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0031] It should be understood that although the terms first, second, etc. may be used in one or more embodiments of this specification to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word “if” as used herein may be interpreted as “when” or “while” or “in response to determining.”

[0032] First, the noun terms related to one or more embodiments of this specification are explained.

[0033] Causal effect estimation: Causal effect estimation is a classic problem in causal counterfactual inference. Suppose t represents whether an intervention is applied to a certain user, y0 represents the potential outcome of this user when no intervention is applied (the user belongs to the control group), and y1 represents the potential outcome of this user when the intervention is applied (the user belongs to the experimental group). Since a user cannot belong to both the control group and the experimental group simultaneously, the observed value for this user is only y0 or y1. Causal effect estimation refers to the estimation of the incremental potential outcome τ = y1 - y0 when an intervention is applied to a user compared to when no intervention is applied, given the user characteristics.

[0034] Stable learning: General machine learning requires the training set and the test set to follow the independent and identically distributed assumption. Machine learning models usually mine correlations rather than causalities, which may lead to the model learning spurious associations and limited generalization ability when the data distributions of the training set and the test set are different. To improve the generalization ability of the model, stable learning decouples the training data features through a method of variable balancing by sample weighting to achieve the purpose of debiasing, and thus can improve the generalization ability of the model in an unknown environment.

[0035] Invariant learning: Invariant learning is an out-of-distribution generalization algorithm that assumes that some aspects of the data distribution vary in the training set, but the underlying causal mechanisms remain unchanged. This method learns representations or models with causally invariant properties from the data, enabling the model to be highly resistant to distribution shifts. A typical method is the Invariant Risk Minimization (IRM) algorithm.

[0036] Latent variable: In statistics, a latent variable is an unobservable random variable, and inferences about latent variables are typically made from samples of observable variables. Take the Gaussian mixture model as an example. In the Gaussian mixture model, the latent variable refers to the Gaussian component corresponding to each observation. Since the generation process is unobservable (or hidden), it is named the latent variable.

[0037] Prior distribution: Also translated as "prior distribution" or "a priori distribution", it is a type of probability distribution. Opposed to the "posterior distribution". It is independent of the experimental results or random sampling and reflects the distribution obtained based on other knowledge about the parameter θ before conducting a statistical experiment.

[0038] Posterior distribution: Given the distribution Pθ of the sample X and the prior distribution π(θ) of θ, using the method of finding the conditional probability distribution in probability theory, the conditional distribution π(θ|x) of θ can be calculated under the condition that X = x is known. Since this distribution is obtained after sampling, it is called the posterior distribution.

[0039] Variational Auto-Encoders (VAE): Auto-encoders are a class of artificial neural networks used in semi-supervised and unsupervised learning, whose function is to perform representation learning on the input information by taking the input information as the learning target. An auto-encoder consists of an encoder and a decoder. Variational Auto-Encoders are a subclass of auto-encoders that solve the problem of the irregularity of the latent space by making the encoder return a distribution in the latent space rather than a single point and adding a regularization term for the returned distribution to the loss function.

[0040] Exemplarily, take the scenario of the marketing pricing of MYbank Business Loan as an example. One of the core objectives of the marketing pricing of MYbank Business Loan is to maximize the balance or utilization of the overall users under the constraints of costs, return on investment (ROI), risks, etc. The pricing process mainly includes two stages: ① Estimating the causal effect of counterfactual inference for various types and intensities of benefits based on single-user characteristics; ② Recommending target benefits for users using methods such as operations research optimization under the constraints of costs, ROI, etc., so as to achieve project objectives such as maximizing the overall balance / utilization. Therefore, accurate causal effect estimation is the core and foundation for achieving project objectives. However, after analyzing and modeling the experimental data, it is found that when conducting multiple randomized experiments on the same group of people defined under the same conditions at different time periods and analyzing the causal effect of single characteristics of the group, there are situations where the causal relationship between some characteristics and the outcome variable is unstable, and even opposite in different-period experiments. Among them, the outcome variable refers to the intervention, such as each piece of information to be recommended. That is to say, user characteristics are easily affected by environmental variables, which in turn leads to inaccurate results of sensitivity modeling (causal effect estimation). The main reasons may include: ① Dynamic changes in the population: Although the same conditions are used for population screening, due to the influence of external activities, such as major promotions, etc., the overall population changes; ② External environmental factors: Due to external economic, policy, and pricing adjustments of competing enterprises during the activity period, the sensitivity of the population to benefits changes; ③ Demand-side factors: In the loan marketing scenario, user conversion is strongly driven by demand and the demand has a certain periodicity due to factors such as the capital turnover time, so the sensitivity varies at different times; ④ Random experiment fluctuations: Since the project objectives involve the user's utilization amount and account balance, there is a certain degree of randomness (individual sporadic capital demands), or due to the limited operation population, it is difficult to ensure accurate randomness, so there are fluctuations. Among them, the above four points can all be regarded as environmental variables.

[0041] Moreover, the main methods for causal effect estimation include: methods based on meta-learners, methods based on tree (forest) models, methods based on neural network discriminant models, methods based on generative models, methods based on VAE models, and so on. Among them, the core idea of the method based on meta-learners is to approximately obtain the gain of the counterfactual result, i.e., uplift, through the combination of a series of base-learners. The method based on tree (forest) models gradually increases the difference in the statistical indicators of the treatment group and the control group labels in the child nodes through splitting. The method based on neural network discriminant models, such as the Counter-Factual Regression (CFR) model. In the first half of the network, the sample representation is learned based on the sample features. Then, according to whether the sample group belongs to the treatment group or the control group, it enters two branches respectively. Two sub-networks are used to learn the result prediction of the treatment group and the control group respectively. In order to balance the sample representations learned by the treatment group and the control group, an Integral Probability Metric (IPM) is added to measure the difference in the representation distributions between the treatment group and the control group. The method based on generative models, for example, generates counterfactual results by establishing a counterfactual generator, and trains the generator to obtain causal effect estimation. The method based on VAE models, such as the Causal Effect Variational Auto-encoders (CEVAE) model and the Treatment Effect Estimation with Disentangled Latent Factors Variational Auto-encoders (TEDVAE) model, both use the architecture of the VAE model to simulate the data generation process and then perform confounding debiasing and causal effect estimation. The above-mentioned causal effect estimation methods do not solve the problem of causal effect changing with the environment. Therefore, when the causal relationship between variables in the training set and the test set changes, the generalization ability of the model may drop sharply, leading to inaccurate relationship analysis results.

[0042] Moreover, stable learning mainly achieves feature decoupling debiasing through the method of variable balancing by sample weighting, and improves the generalization ability of the model in unknown environments. The method of stable learning mainly solves the problem that the model learns spurious associations due to the biased distribution of training set samples. It can be understood that it is the internal cause of the data that leads to the decline in the generalization ability of the cross-environment model. However, the causal relationship between variables remains unchanged in different environments. Therefore, it is not applicable to scenarios where the causal relationship between variables changes due to external factors (environmental changes).

[0043] Moreover, invariant learning is a distributional appearance feature learning algorithm that achieves the ability to generalize in unknown environments by learning underlying causal mechanisms that remain invariant. However, the causal effect estimation method based on invariant learning has serious potential problems. For non-linear causal effects, it will fail unless the test data is distributed similarly enough to the training data. Therefore, for most scenarios with non-linear causal effects, invariant learning will lose the ability to generalize in unknown environments.

[0044] To provide a more accurate solution, in this specification, an information recommendation method is provided. This specification also relates to an information recommendation device, a computing device, and a computer-readable storage medium, which will be described in detail one by one in the following embodiments.

[0045] See Figure 1 , Figure 1 shows a flowchart of a relationship analysis model training method provided by an embodiment of this specification, which specifically includes the following steps:

[0046] S102, obtain the object information, sample recommendation information, sample environmental variables of the sample object under the sample task, and the relationship annotation information indicating the influence of the sample recommendation information on the sample object to perform the sample task;

[0047] This step is to obtain the training set. Among them, the sample task can be tasks such as loan, viewing messages, playing music, purchasing goods, etc. The relationship annotation can be a binary label, such as executed / not executed. Or, the relationship annotation can be multi-valued, continuous-valued, linear-valued, etc. For example, the amount of the purchase price, the number of plays, etc. Exemplarily, as Figure 2 As shown in the application scenario example diagram of a relationship analysis model training method provided by an embodiment of this specification, obtaining the object information, sample recommendation information, sample environmental variables of the sample object under the sample task, and the relationship annotation information indicating the influence of the sample recommendation information on the sample object to perform the sample task is multi-environment data preparation: including user characteristics, the interventions imposed, and the final observation results. The multi-environment data provides the possibility for obtaining cross-environment causally invariant features later.

[0048] S104, perform an environmental impact analysis on the object information of the sample object based on the sample environmental variables, and extract the first sample object features that are not affected by the sample environmental variables from the object information of the sample object.

[0049] In specific applications, performing an environmental impact analysis on the object information of the sample object based on the sample environmental variables and extracting the first sample object features that are not affected by the sample environmental variables from the object information of the sample object can specifically be a binary classification model for classifying the object information of the sample object. Or, exemplarily, it can be extracted from the object information of the sample object by choosing conditions. And, asFigure 2 As shown in the figure, during the model training phase, the selection conditions can be specifically implemented through the gating mechanism and the decomposition of the latent variables of the relationship analysis model, that is, in the form of VAE latent variables. Thus, through the gating mechanism and the decomposition of the latent variables of the relationship analysis model, that is, VAE latent variables, the features can be effectively split into: 1. The environment-invariant component, that is, the first sample object feature that is not affected by the sample environment variables; 2. The environment-changing component, that is, the second sample object feature that is affected by the sample environment variables. On this basis, for greater accuracy, as Figure 2 shown in the figure, during the model training phase, a gating complementary sparse constraint can be added to the comparison result corresponding to the relationship analysis model, such as the loss value, to achieve the information amount constraint of the gating invariant branch. For the sake of easy understanding and reasonable layout, in the form of optional embodiments, the second example and the situation of adding the gating complementary sparse constraint will be specifically described later.

[0050] S106, input the first sample object feature and the sample recommendation information into the pre-trained relationship analysis model to obtain the predicted relationship analysis result of the sample recommendation information affecting the sample object to perform the sample task.

[0051] Exemplarily, as Figure 2 shown in the figure, the pre-trained relationship analysis model can achieve prediction based on the invariant latent variable: during the prediction process, only based on the stable latent variable, cut off the influence of the unstable component on the prediction result, and improve the generalization ability of the method in the unknown environment. Among them, for the specific processing logic of prediction based on the invariant latent variable, reference can be made to Figure 3 . Exemplarily, as Figure 3 shown in the schematic diagram of the causal graph of a relationship analysis model provided in an embodiment of this specification: among them, the posterior probability of the latent variable can be inferred by collecting samples, and then the estimated posterior probability can be used to cluster the data to achieve self-supervised training. e is the environment identifier (which can distinguish different environment variables in the training set and is unknown during the prediction process). X is the object information of the sample object, such as user features, which can specifically include user attribute features and user behavior features. The user attribute features can specifically include attribute features such as user identifier, user level, and user permissions. The object information X of the sample object can be split into a component X inv that does not change with the environment in the causal relationship with y and a component X var that changes with the environment. t is the intervention, that is, the type of the sample recommendation information. For the sake of easy understanding and description, it is assumed to be a binary intervention (with / without intervention) in the subsequent embodiments. In specific applications, it can be extended to multi-valued intervention and continuous-valued intervention. y is the result variable, that is, the predicted relationship analysis result. Z inv and Z var are the latent variables that do not change with the environment and change with the environment respectively. The above steps S104 and S106 are respectively used to identify Xinv and X var and cut off the unstable component X during the modeling process var and the latent variable Z that changes with the environment var and the relationship with the analysis result y of the prediction relationship, so as to improve the ability of the model to resist changes in causal relationships caused by environmental changes and be stable and effective across environments.

[0052] S108. Compare the analysis result of the prediction relationship with the relationship annotation information, adjust the model parameters of the pre-trained relationship analysis model based on the comparison result, and return the object information, sample recommendation information, sample environmental variables of the sample object under the task of obtaining samples, and the relationship annotation information indicating that the sample recommendation information affects the sample object to perform the sample task.

[0053] In specific applications, the above comparison of the analysis result of the prediction relationship and the relationship annotation information can be various. Exemplarily, the above comparison can be achieved by adding a target loss function with complementary sparse constraints. Or, exemplarily, when the above pre-trained relationship analysis model includes an autoencoder and a relationship analyzer, the above comparison can be achieved by fusing the loss values corresponding to the autoencoder and the relationship analyzer respectively. For the convenience of understanding and reasonable layout, step S108 will be specifically described in the form of optional embodiments later.

[0054] S110. When the preset training stop condition is reached, obtain the trained relationship analysis model.

[0055] In specific applications, the preset training stop condition can be various. Exemplarily, the preset training stop condition can include: if the comparison result is a loss value, the loss value reaches the loss threshold; or, the number of training rounds reaches the round threshold; or, an early stopping condition. For example, the preset number of training rounds is 8 rounds. In specific applications, when the training is less than 8 rounds, it has been fitted, and continuing the training will cause overfitting and reduce the accuracy of the model. Therefore, the training can be ended before reaching the number of training rounds, that is, early stopping. Early stopping is because the model training may...

[0056] An embodiment of this specification can ensure that the first target object feature that is not affected by the environmental variable can be split from the object information of the target object under the target task, and the causal effect prediction of the user feature and the information to be recommended can be performed based on the first target object feature that is not affected by the environment, so as to achieve the effect of cutting off the influence of the object feature affected by the environmental variable on the prediction result. Therefore, the influence of the environmental variable on the prediction result can be reduced, and the accuracy of the prediction result can be improved. On this basis, based on each prediction result, the target recommendation information in each piece of information to be recommended is determined, and the target recommendation information is recommended to the target object, so as to achieve a more accurate information recommendation with reduced influence of the environmental variable. Therefore, this solution can achieve a more accurate and more applicable effect in more environments.

[0057] In an alternative embodiment, for the above environmental impact analysis of object information based on sample environmental variables, extracting first sample object features that are not affected by the sample environmental variables from the object information may specifically include the following steps:

[0058] Obtain the influence result of the sample environmental variables on the object information of the sample object under the sample task;

[0059] Extract second sample object features affected by the sample environmental variables from the influence result, and determine a third selection condition for extracting the second sample object features from the object information of the sample object;

[0060] Use a fourth selection condition complementary to the third selection condition to extract the first sample object features from the object information of the sample object.

[0061] Among them, the fourth selection condition can be in the form of a gating mechanism to achieve feature classification. In this way, the feature extraction is divided into two branches. The upper branch only inputs the first sample object features regulated by the gating mechanism, and the lower branch adds the input of the environmental variables so that the unstable components in the object information of the sample object, that is, the second sample object features affected by the sample environmental variables, can interact with the environmental variables. Exemplarily, the calculation method divided into two branches by the gating mechanism is shown in the following formula (1):

[0062]

[0063] Among them, g(x) is the fourth selection condition, (1 - g(x)) is the third selection condition, W g is the weight of the activation function, x inv is the first sample object feature, x var is the second sample object feature, e is the natural constant, and the symbol is the multiplication symbol. That is to say, the influence result of the sample environmental variables on the object information of the sample object can be obtained through the Concat() method. Among them, the Concat() method is used to concatenate two or more arrays.

[0064] In an alternative embodiment, for the above comparison of the prediction relationship analysis result and the relationship annotation information, it may specifically include the following steps:

[0065] Obtain an objective loss function containing sparse constraints, where the sparse constraints are set for the first selection condition and the second selection condition;

[0066] Input the prediction relationship analysis result and the relationship annotation information into the objective loss function to obtain a comparison result, and adjust the parameters of the third selection condition based on the comparison result.

[0067] In a specific application, there can be various target loss functions that include sparse constraints. Exemplarily, the target loss function including the above-mentioned sparse constraint can be directly read, that is, a sparse constraint is added to the loss function from the beginning of model training. Or, exemplarily, a sparse constraint can be added to a preset loss function under certain conditions. For the sake of easy understanding and reasonable layout, specific descriptions will be given in the form of optional embodiments hereinafter. In this embodiment, the sparse constraints set for the first selection condition and the second selection condition can be regarded as imposing complementary sparse constraints on the dual branches of the gated output, so that features can be classified according to the environment-dependent invariance. That is, the parameters of the third selection condition, such as W g , are adjusted based on the comparison result, so as to achieve a more accurate effect of feature extraction.

[0068] In an optional implementation manner, the above-mentioned obtaining the target loss function including the sparse constraint may specifically include the following steps:

[0069] Determine the influence degree of the sample environment variable on the object information of the sample object;

[0070] When the influence degree reaches a preset degree threshold, add a sparse constraint to the preset loss function to obtain the target loss function.

[0071] Among them, the influence degree of the sample environment variable on the object information of the sample object can be the training using the first N batches of epoch training samples, the duration of the influence, the difference between the influence result and the object information of the sample object, etc. In this embodiment, by adding a sparse constraint to the preset loss function when the influence degree reaches the preset degree threshold to obtain the target loss function, it is possible to allow features to selectively interact with the environment sufficiently in the early stage to ensure the accuracy of feature extraction.

[0072] Exemplarily, the above-mentioned sparse constraint can be set to the form of. In this way, when the influence degree does not reach the preset degree threshold, no sparse constraint is imposed on the gate, and the sparsity of the gate sparse constraint in the loss, that is, the sparsity of the sparse constraint in this embodiment, is αg = 0. For example, αg = 0 is set in the training using the first N batches of epoch training samples. Correspondingly, when the influence degree reaches the preset degree threshold, adding a sparse constraint to the preset loss function can be setting αg ≠ 0.

[0073] In an optional implementation manner, the above-mentioned sparse constraint includes a regulation coefficient;

[0074] Correspondingly, after obtaining the target loss function including the sparse constraint, the relationship analysis model training method provided in the embodiments of this specification may further include the following steps:

[0075] Increase the adjustment coefficient according to the preset adjustment frequency.

[0076] In this embodiment, by gradually increasing the gated sparsity constraint, for example, increasing the adjustment coefficient according to the preset adjustment frequency, it can ensure that the stability of model training gradually increases after N epochs (for example, first linear and then fixed) αg. Among them, linear means a proportional and straight-line relationship between quantities, which can be understood mathematically as a function with a constant first derivative.

[0077] In an optional embodiment, before inputting the first sample object feature and the sample recommendation information into the pre-trained relationship analysis model, the relationship analysis model training method provided by the embodiments of this specification may further include the following steps:

[0078] Conduct an environmental impact analysis on the object information of the sample object based on the sample environmental variable, extract the second sample object feature affected by the sample environmental variable from the object information of the sample object, and extract the original sample object feature from the object information of the sample object;

[0079] Correspondingly, inputting the first sample object feature and the sample recommendation information into the pre-trained relationship analysis model to obtain the predicted relationship analysis result of the sample recommendation information affecting the sample object to perform the sample task may specifically include the following steps:

[0080] Input the first sample object feature and the sample recommendation information into the pre-trained relationship analysis model to obtain the first predicted relationship analysis result;

[0081] Input the first sample object feature, the second sample object feature, the original sample object feature, and the sample recommendation information into the pre-trained relationship analysis model to obtain the second predicted relationship analysis result;

[0082] Use the first predicted relationship analysis result and the second predicted relationship analysis result as the predicted relationship analysis result of the sample recommendation information affecting the sample object to perform the sample task.

[0083] Among them, the prior distribution assumptions of each latent variable, that is, the first sample object feature, the second sample object feature, and the original sample object feature, follow the Gaussian distribution shown in the following formulas (2-1) to (2-3):

[0084]

[0085]

[0086]

[0087] Among them, the outputs of formulas (2-1) to (2-3) are the prior distributions of the first sample object feature, the second sample object feature, and the original sample object feature respectively. The parameters in formulas (2-1) to (2-3) are parameters under the Gaussian distribution. The specific meanings of the parameters are as follows: If the random variable X follows a normal distribution with a mathematical expectation of μ and a variance of σ2, it is denoted as N(μ, σ2). The probability density function of the random variable X is that the expected value μ of the normal distribution determines the position of the random variable X, and the standard deviation σ of the random variable X determines the amplitude of the distribution. The normal distribution when μ = 0 and σ = 1 is the standard normal distribution. D is the total number of random variables, and j is the serial number of the random variable. On this basis, for the first sample object feature, the second sample object feature, and the original sample object feature, the specific values of the mathematical expectation μ and the specific values of the standard deviation σ are all estimated values. The subscripts of the parameters in formulas (2-1) to (2-3) represent the sample object features corresponding to the specific parameters. For example, the subscript of the parameter corresponding to the first sample object feature is inv, the subscript of the parameter corresponding to the second sample object feature is var, and the parameter corresponding to the original sample object feature has no subscript.

[0088] Exemplarily, such as Figure 4 In a method for training a relationship analysis model provided by an embodiment of this specification, as shown in the structural example diagram of the decoder, the original sample object feature p(Z0) and the sample recommendation information p(t|Z0, Z τ inv , Z τ var ) are input into the pre-trained relationship analysis model to obtain the first predicted relationship analysis result p(y|t = 0, Z0); the first sample object feature p(Z τ inv ) and the sample recommendation information p(t|Z0, Z τ inv , Z τ var ) are input into the pre-trained relationship analysis model to obtain the second predicted relationship analysis result p(y|t = 1, Z0, Z τ inv ); the first sample object feature p(Z τ inv ), the second sample object feature p(Z τ var ), the original sample object feature p(Z0) and the sample recommendation information p(t|Z0, Z τ inv , Z τ var ) are input into the pre-trained relationship analysis model to obtain the third predicted relationship analysis result p(y|t = 1, Z0, Z τ inv , Zτ var );The analysis results of the second prediction relationship and the third prediction relationship are used as the sample recommendation information to affect the prediction relationship analysis result of the sample object performing the sample task. In this way, in this embodiment, both the relationship analysis result based on the first target object feature and the relationship analysis result based on the first target object feature, the second target object feature, and the original object feature are output, which is more comprehensive and accurate.

[0089] Moreover, the decoder mainly restores the user feature X, that is, the sample object feature, the intervention t, and the result y based on the latent variable. See Figure 4 , where it is assumed that the causal effect increment τ is additive, τ = τ inv +τ var . Since the input of the decoder is the sampling of the encoder latent variable result distribution, it is the result of noisy perturbation and has better noise resistance. Combining the user feature X, the intervention t, and the reconstruction calculation method of the result variable y is shown in the following formulas (3-1) to (3-3):

[0090]

[0091]

[0092]

[0093] Among them, the parameters in formulas (3-1) to (3-3) are combined with Figure 4 shown as follows, and the specific meanings are: is the user feature, specifically including: the feature for the original variable z0, the feature for the variable that is not affected by the sample environment variable feature, and the feature for the variable affected by the sample environment variable feature, and each feature follows a Gaussian distribution. p(t|Z0, Z τ inv , Z τ var ) is the sample recommendation information, which is obtained through the function Bern() referring to the Bernoulli distribution. The Bernoulli distribution, also known as the "zero-one distribution" and "two-point distribution". The Bernoulli distribution means that for the random variable X with parameter p (0 < p < 1), it is assumed that the random variable X takes the values 1 and 0 with probabilities p and 1 - p respectively. EX = p, DX = p(1 - p). The number of successful Bernoulli trials follows the Bernoulli distribution, and the parameter p is the probability of success in the trial. The Bernoulli distribution is a discrete probability distribution and a special case of the binomial distribution when N = 1. δ is the activation function, and f0 to f8 all represent neural networks. In order to obtain the causal effect estimate not affected by environmental factors, combined with τ = τ inv +τvar Adding a branch of latent variables based on stable latent variables, i.e., the first sample features, for causal effect estimation modeling can obtain variables that combine features not affected by sample environmental variables and a causal effect estimation model p(τ|Z of features of variables affected by sample environmental variables τ inv , Z τ var ), a causal effect estimation model p(τ inv |Z τ inv ) only for variables not affected by sample environmental variables, and a causal effect estimation model p(τ var |Z τ var ) only for variables affected by sample environmental variables. In this way, since the decoder fully simulates the data generation process, the problem of confounding debiasing in the case of biased data is solved. In addition, for the meanings of the same parameters in formulas (3-1) to (3-3) and formulas (2-1) to (2-3), reference can be made to the symbol explanations of formulas (2-1) to (2-3), which will not be elaborated here.

[0094] In an alternative embodiment, inputting the first sample object features and the sample recommendation information into the pre-trained relationship analysis model to obtain the predicted relationship analysis result of the influence of the sample recommendation information on the sample object to perform the sample task may specifically include the following steps:

[0095] Extracting the recommended features of the sample recommendation information for at least one of the second sample object features and the original sample object features, and the annotation features of the relationship annotation information of the first sample object features;

[0096] Inputting at least one of the second sample object features and the original sample object features, the first sample object features, and the recommended features of the sample recommendation information into the pre-trained relationship analysis model to obtain the predicted relationship analysis result of the influence of the sample recommendation information on the sample object to perform the sample task.

[0097] In a specific application, the extraction of each object feature can be implemented by an encoder. Exemplarily, as Figure 5 shown in the structural example diagram of the encoder in a method for training a relationship analysis model provided in an embodiment of this specification: To help the latent variables effectively learn, an auxiliary branch (represented by a dashed line) is added for predicting the intervention t and the result y at the encoding end, and the calculation method is the same as that of the decoder. And, refer to Figure 5, the Gate Layer is the above-mentioned gating mechanism. Since the model based on the variational autoencoder framework performs relatively better in the field of causal effects, this method is implemented based on the variational autoencoder framework. Figure 5 The encoder is shown, and the white nodes represent the neural network layers.

[0098] Among them, the gating layer is the Gate Layer, p(X) and p(e) respectively correspond to the distributions of the user features and environmental variables in the training set, q(Z0|X) is the distribution of the latent variables without intervention, and q(Z τ inv |X), q(Z τ inv |X,e) respectively correspond to the distributions of the latent variables corresponding to the causal effect increment τ for the case of unchanged environment and changed environment, q(t|Z0, Z τ inv , Z τ var ) is the distribution of the latent variables of the user features after applying the environmental variable, that is, the intervention t, and q(y|Z0, Z τ inv , Z τ var ) is the distribution of the latent variables of the sample recommendation information for the latent variables of the user features.

[0099] This embodiment realizes increasing the prediction of the intervention t and the result y by the encoder, and improving the extraction accuracy of the latent variables, that is, the features of each object.

[0100] In an alternative embodiment, the pre-trained relationship analysis model includes: a variational autoencoder and a relationship analyzer, where the variational autoencoder is used to extract latent variables from the first sample object features, sample recommendation information, and relationship annotation information respectively;

[0101] Correspondingly, comparing the above-mentioned predicted relationship analysis result and relationship annotation information, and adjusting the model parameters of the pre-trained relationship analysis model based on the comparison result, specifically, it may include the following steps:

[0102] Compare the first sample object features, sample recommendation information, and relationship annotation information with the corresponding latent variables respectively to obtain a first comparison result, and adjust the model parameters of the variational autoencoder based on the first comparison result;

[0103] Compare the predicted relationship analysis result and relationship annotation information to obtain a second comparison result, and adjust the model parameters of the relationship analyzer based on the second comparison result and the first comparison result.

[0104] Exemplarily, the loss of the variational autoencoder, that is, the first comparison result, is specifically shown in the following formula (4):

[0105]

[0106] Among them, That is, the results of separately extracting latent variables from the first sample object feature, sample recommendation information, and relationship annotation information, D KL is the KL divergence (Kullback-Leibler divergence) of the distribution. The KL divergence is a method for measuring the matching degree between two distributions (such as two lines). The smaller the KL divergence, the better the matching between the true distribution and the approximate distribution.

[0107] The overall loss of the model, that is, the second comparison result, is specifically shown in the following formula (5):

[0108]

[0109] Wherein:

[0110]

[0111]

[0112] and are respectively the calculation of the comparison result, such as the loss value, for each prediction relationship analysis result. Specifically, is the prediction relationship analysis result of the first sample object feature, the original sample object feature, and the sample recommendation information with the causal effect increment added; is the prediction relationship analysis result of the original sample object feature, the first sample object feature with the causal effect increment added, the second sample object feature with the causal effect increment added, the original sample object feature, and the sample environment variable; is the prediction relationship analysis result of the original sample object feature, the first sample object feature with the causal effect increment added, the second sample object feature with the causal effect increment added, and the sample recommendation information. are respectively the loss function parameters for each of the above-mentioned prediction relationship analyses. In this way, feature extraction and relationship analysis can be effectively combined. In addition, in formulas (4) and (5), for the same parameters as in formulas (2-1) to (2-3), (3-1) to (3-3), the symbol explanations in formulas (2-1) to (2-3), (3-1) to (3-3) can be referred to and will not be elaborated here.

[0113] See Figure 6 , Figure 6 shows the flowchart of an information recommendation method provided by an embodiment of this specification, specifically including the following steps:

[0114] S602. Obtain the object information of the target object under the target task and each piece of information to be recommended.

[0115] The target task is similar to the sample task in the above Figure 1 embodiment, the object information of the target object is similar to the object information of the sample object in the above Figure 1 embodiment, and the information to be recommended is similar to the sample recommended information, with the difference being adaptive adjustment according to the specific application scenario. For the same parts, they will not be elaborated here. For details, see the corresponding descriptions in the above Figure 1 embodiment and Figure 1 the corresponding descriptions of the optional embodiments.

[0116] S604. Extract the first target object feature that is not affected by the environment from the object information, and for any piece of information to be recommended, based on the first target object feature and the piece of information to be recommended, predict the prediction result of the piece of information to be recommended affecting the target object to perform the target task.

[0117] The specific implementation manner of this step can be various. For the convenience of understanding and reasonable layout, it will be given in the form of optional embodiments later.

[0118] S606. Based on each prediction result, determine the target recommended information in each piece of information to be recommended, and recommend the target recommended information to the target object.

[0119] Exemplarily, based on each prediction result, determining the target recommended information in each piece of information to be recommended may include: if the prediction result is that the piece of information to be recommended affects whether the target object performs the target task, then if it is to perform, determine the piece of information to be recommended as the target recommended information; or, if the prediction result is the probability that the piece of information to be recommended affects the target object to perform the target task, determine the piece of information to be recommended with the highest probability among each prediction result as the target recommended information. Any manner that can determine the target recommended information in each piece of information to be recommended based on each prediction result can be used in this specification, and this embodiment does not limit this.

[0120] An embodiment of this specification can ensure that the first target object feature that is not affected by the environment can be split from the object information of the target object under the target task, and based on the first target object feature that is not affected by the environment, perform causal effect prediction between the user feature and the information to be recommended, so as to achieve the effect of cutting off the influence of the object feature affected by the environmental variable on the prediction result. Therefore, the influence of the environmental variable on the prediction result can be reduced, and the accuracy of the prediction result can be improved. On this basis, based on each prediction result, determine the target recommended information in each piece of information to be recommended, and recommend the target recommended information to the target object, which can achieve more accurate information recommendation with reduced influence of the environmental variable. Therefore, this solution can achieve more accurate and applicable to more environments.

[0121] In an optional implementation manner, the above-mentioned method extracts the first target object feature that is not affected by the environment from the object information, and for any piece of information to be recommended, based on the first target object feature and the information to be recommended, predicts the prediction result of the information to be recommended affecting the target object to perform the target task. Specifically, it may include the following steps:

[0122] For any piece of information to be recommended, input the object information and the information to be recommended into a pre-trained relationship analysis model. The relationship analysis model extracts the first target object feature that is not affected by the environment, and based on the first target object feature and the information to be recommended, predicts the prediction result of the information to be recommended affecting the target object to perform the target task.

[0123] [[ID= six]]The relationship analysis model in this embodiment is the same as the above-mentioned Figure 1 trained relationship analysis model in the embodiment. Therefore, the process of obtaining the relationship analysis model in this embodiment, that is, the training method of the relationship analysis model, is the same as that of the above-mentioned Figure 1 embodiment and Figure 1 the steps of the relationship analysis model training method provided in the optional embodiment are the same. The same parts will not be elaborated here. For details, please refer to the description of the above-mentioned Figure 1 embodiment and Figure 1 optional embodiment.

[0124] In an optional implementation manner, the above-mentioned method extracts the first target object feature that is not affected by the environment from the object information. Specifically, it may include the following steps:

[0125] Obtain the influence result of the preset environmental variable on the object information under the target task;

[0126] Extract the second target object feature affected by the preset environmental variable from the influence result, and determine the first selection condition for extracting the second target object feature from the object information;

[0127] Use the second selection condition complementary to the first selection condition to extract the first target object feature from the object information.

[0128] Each step of this embodiment is similar to each step of the above-mentioned optional embodiment regarding the extraction of the first sample object feature, except that the data processed in each step is adaptively adjusted for actual applications. The same parts will not be elaborated here. For details, please refer to the description of the above-mentioned optional embodiment regarding the extraction of the first sample object feature. Figure 1 Figure 1

[0129]

[0129] The following combines the attached Figure 7 , taking the application of the information recommendation method provided in this specification in the data service platform as an example, to further illustrate the information recommendation method. Among them, Figure 7FIG. 0 shows an example application scenario diagram of an information recommendation method provided by an embodiment of this specification, which is applied to a data service platform and specifically includes the following steps: receiving a request sent by a terminal logged in by a target object, where the request is a request for information to be recommended; invoking a prediction module to extract features of the information to be recommended and the object information of the target object, performing relationship prediction based on the extraction results, obtaining a prediction result, and feeding it back to the terminal. Among them, the relationship prediction can be implemented by a relationship prediction model pre-trained in the prediction model. The relationship prediction model is obtained by the data service platform invoking the prediction module to train using a training set in the database. For the specific process of this training, reference can be made to the above Figure 1 embodiment and Figure 1 the description of the alternative embodiments, which will not be elaborated here.

[0130] This method introduces an "environment" variable to construct a causal graph, and the designed solution can cope with environmental changes to obtain a cross-environment effective causal effect estimation method INV, that is, a relationship analysis model (INV, with Invariant Feature Extraction), which can be applied to scenarios such as the pricing of new and old customers in the online business loan market, competitor pricing, and industry pricing.

[0131] Corresponding to the above method embodiment, this specification also provides an embodiment of an information recommendation device, Figure 8 showing a schematic structural diagram of an information recommendation device provided by an embodiment of this specification. As Figure 8 shown, the device includes:

[0132] An information acquisition module 802, configured to acquire the object information of the target object under the target task and each piece of information to be recommended;

[0133] A prediction module 804, configured to extract the first target object features that are not affected by the environment from the object information, and for any piece of information to be recommended, predict the prediction result of the information to be recommended affecting the target object to perform the target task according to the first target object features and the information to be recommended;

[0134] An information recommendation module 806, configured to determine the target recommended information among each piece of information to be recommended based on each prediction relationship analysis result, and recommend the target recommended information to the target object.

[0135] In an alternative implementation manner, the above prediction module 804 is further configured to:

[0136] For any piece of information to be recommended, the object information and the piece of information to be recommended are input into a pre-trained relationship analysis model. The relationship analysis model extracts first target object features that are not affected by the environment, and based on the first target object features and the piece of information to be recommended, a prediction result of the piece of information to be recommended affecting the target object to perform a target task is predicted.

[0137] In an alternative embodiment, the above prediction module 804 is further configured to:

[0138] Obtain the influence result of a preset environmental variable on the object information under the target task;

[0139] Extract second target object features affected by the preset environmental variable from the influence result, and determine a first selection condition for extracting the second target object features from the object information;

[0140] Extract the first target object features from the object information using a second selection condition that is complementary to the first selection condition.

[0141] The above is a schematic solution of an information recommendation device according to this embodiment. It should be noted that the technical solution of this information recommendation device and the technical solution of the above information recommendation method belong to the same concept. For the details not described in the technical solution of the information recommendation device, reference can be made to the description of the technical solution of the above information recommendation method.

[0142] Figure 9 The structural block diagram of a computing device according to an embodiment of this specification is shown. The components of the computing device 900 include but are not limited to a memory 910 and a processor 920. The processor 920 is connected to the memory 910 through a bus 930, and a database 950 is used to store data.

[0143] The computing device 900 also includes an access device 940, which enables the computing device 900 to communicate via one or more networks 960. Examples of such networks include the Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 940 may include one or more of any type of wired or wireless network interface (e.g., Network Interface Controller (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, Worldwide Interoperability for Microwave Access (Wi-MAX) interface, Ethernet interface, Universal Serial Bus (USB) interface, cellular network interface, Bluetooth interface, Near Field Communication (NFC) interface, and so on.

[0144] In one embodiment of the present specification, the above components of the computing device 900 and Figure 9 other components not shown may also be connected to each other, for example, via a bus. It should be understood that Figure 9 the block diagram of the computing device shown is for illustrative purposes only and is not a limitation on the scope of the present specification. Those skilled in the art can add or replace other components as needed.

[0145] The computing device 900 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device 900 can also be a mobile or stationary server.

[0146] Among them, the processor 920 is used to execute the following computer-executable instructions, and when the computer-executable instructions are executed by the processor, the steps of the above information recommendation method are implemented.

[0147] The above is a schematic solution of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the above information recommendation method belong to the same concept. For the details not described in detail in the technical solution of the computing device, reference can be made to the description of the technical solution of the above information recommendation method.

[0148] An embodiment of this specification also provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by a processor, the steps of the above information recommendation method are implemented.

[0149] The above is a schematic solution of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solution of the above information recommendation method belong to the same concept. For the details not described in detail in the technical solution of the storage medium, reference can be made to the description of the technical solution of the above information recommendation method.

[0150] An embodiment of this specification also provides a computer program. When the computer program is executed on a computer, the computer is made to execute the steps of the above information recommendation method.

[0151] The above is a schematic solution of a computer program according to this embodiment. It should be noted that the technical solution of this computer program and the technical solution of the above information recommendation method belong to the same concept. For the details not described in detail in the technical solution of the computer program, reference can be made to the description of the technical solution of the above information recommendation method.

[0152] The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve the desired result. Additionally, the processes depicted in the figures do not necessarily require the particular order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0153] The computer instructions include computer program code, which may be in source code form, object code form, executable file, or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, removable hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0154] It should be noted that, for the foregoing method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the embodiments of this specification are not limited by the described action sequence, because according to the embodiments of this specification, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments of this specification.

[0155] In the above embodiments, the descriptions of the respective embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0156] The preferred embodiments of this specification disclosed above are only used to help explain this specification. The alternative embodiments do not elaborate on all details and do not limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of the embodiments of this specification. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments of this specification, so that those skilled in the art can well understand and utilize this specification. This specification is only limited by the claims and their full scope and equivalents.

Claims

1. An information recommendation method, comprising: Obtaining object information of a target object under a target task, and each piece of information to be recommended; Extracting a first target object feature that is not affected by the environment from the object information, and for any piece of information to be recommended, predicting a prediction result of the piece of information to be recommended affecting the target object to perform the target task according to the first target object feature and the piece of information to be recommended, wherein the extracting a first target object feature that is not affected by the environment from the object information includes: obtaining an influence result of a preset environmental variable on the object information under the target task; extracting a second target object feature affected by the preset environmental variable from the influence result, and determining a first selection condition for extracting the second target object feature from the object information; using a second selection condition complementary to the first selection condition to extract the first target object feature from the object information; Based on each prediction result, determining a target recommended information among the pieces of information to be recommended, and recommending the target recommended information to the target object.

2. The method according to claim 1, wherein the extracting a first target object feature that is not affected by the environment from the object information, and for any piece of information to be recommended, predicting a prediction result of the piece of information to be recommended affecting the target object to perform the target task according to the first target object feature and the piece of information to be recommended, includes: For any piece of information to be recommended, inputting the object information and the piece of information to be recommended into a pre-trained relationship analysis model, extracting a first target object feature that is not affected by the environment through the relationship analysis model, and predicting a prediction result of the piece of information to be recommended affecting the target object to perform the target task according to the first target object feature and the piece of information to be recommended.

3. The method according to claim 2, wherein the training method of the relationship analysis model includes: Obtaining object information of a sample object, sample recommended information, sample environmental variables under a sample task, and relationship annotation information of the sample recommended information affecting the sample object to perform the sample task; Performing environmental impact analysis on the object information of the sample object based on the sample environmental variables, and extracting a first sample object feature that is not affected by the sample environmental variables from the object information of the sample object; Inputting the first sample object feature and the sample recommended information into a pre-trained relationship analysis model to obtain a predicted relationship analysis result of the sample recommended information affecting the sample object to perform the sample task; Comparing the predicted relationship analysis result with the relationship annotation information, adjusting model parameters of the pre-trained relationship analysis model based on the comparison result, and returning to execute obtaining object information of a sample object, sample recommended information, sample environmental variables under a sample task, and relationship annotation information of the sample recommended information affecting the sample object to perform the sample task; Obtaining the trained relationship analysis model when a preset training stop condition is reached.

4. The method according to claim 3, wherein the environmental impact analysis of the object information based on the sample environmental variables and extracting first sample object features that are not affected by the sample environmental variables from the object information includes: Obtaining the influence result of the sample environmental variables on the object information of the sample object under the sample task; Extracting second sample object features affected by the sample environmental variables from the influence result, and determining a third selection condition for extracting the second sample object features from the object information of the sample object; Using a fourth selection condition complementary to the third selection condition to extract the first sample object features from the object information of the sample object.

5. The method according to claim 4, wherein the comparing the prediction relationship analysis result and the relationship annotation information includes: Obtaining a target loss function including a sparsity constraint, wherein the sparsity constraint is set for the third selection condition and the fourth selection condition; Inputting the prediction relationship analysis result and the relationship annotation information into the target loss function to obtain a comparison result, and adjusting parameters of the third selection condition based on the comparison result.

6. The method according to claim 5, wherein the obtaining a target loss function including a sparsity constraint includes: Determining the influence degree of the sample environmental variables on the object information of the sample object; Adding the sparsity constraint to a preset loss function when the influence degree reaches a preset degree threshold to obtain the target loss function.

7. The method according to claim 5 or 6, wherein the sparsity constraint includes a regularization coefficient; Accordingly, after obtaining the target loss function including the sparsity constraint, the method further includes: Increasing the regularization coefficient at a preset adjustment frequency.

8. The method according to any one of claims 4-6, before inputting the first sample object features and the sample recommendation information into a pre-trained relationship analysis model, the method further includes: Performing environmental impact analysis on the object information of the sample object based on the sample environmental variables, extracting second sample object features affected by the sample environmental variables from the object information of the sample object, and extracting original sample object features from the object information of the sample object; Accordingly, the inputting the first sample object features and the sample recommendation information into a pre-trained relationship analysis model to obtain a prediction relationship analysis result of the influence of the sample recommendation information on the sample object to perform the sample task includes: Inputting the first sample object features and the sample recommendation information into a pre-trained relationship analysis model to obtain a first prediction relationship analysis result; Inputting the first sample object features, the second sample object features, the original sample object features, and the sample recommendation information into a pre-trained relationship analysis model to obtain a second prediction relationship analysis result; Using the first prediction relationship analysis result and the second prediction relationship analysis result as the prediction relationship analysis result of the influence of the sample recommendation information on the sample object to perform the sample task.

9. The method according to claim 8, wherein the step of inputting the first sample object feature and the sample recommendation information into a pre-trained relationship analysis model to obtain a predicted relationship analysis result of the sample recommendation information affecting the sample object to perform the sample task includes: extracting recommended features of the sample recommendation information and labeled features of the relationship annotation information for at least one of the second sample object feature and the original sample object feature, and the first sample object feature; inputting at least one of the second sample object feature and the original sample object feature, the first sample object feature, and the recommended features of the sample recommendation information into a pre-trained relationship analysis model to obtain a predicted relationship analysis result of the sample recommendation information affecting the sample object to perform the sample task.

10. The method according to any one of claims 3-6, 9, wherein the pre-trained relationship analysis model comprises: a variational autoencoder and a relationship analyzer, wherein the variational autoencoder is configured to perform latent variable extraction on the first sample object feature, the sample recommendation information, and the relationship annotation information respectively; Accordingly, the step of comparing the predicted relationship analysis result and the relationship annotation information and adjusting the model parameters of the pre-trained relationship analysis model based on the comparison result includes: comparing the first sample object feature, the sample recommendation information, and the relationship annotation information with the corresponding latent variables respectively to obtain a first comparison result, and adjusting the model parameters of the variational autoencoder based on the first comparison result; comparing the predicted relationship analysis result and the relationship annotation information to obtain a second comparison result, and adjusting the model parameters of the relationship analyzer based on the second comparison result and the first comparison result.

11. An information recommendation device, comprising: an information acquisition module configured to acquire object information of a target object and each piece of information to be recommended under a target task; a prediction module configured to extract a first target object feature that is not affected by the environment from the object information, and for any piece of information to be recommended, predict a prediction result of the piece of information to be recommended affecting the target object to perform the target task according to the first target object feature and the piece of information to be recommended; The prediction module is further configured to acquire an influence result of a preset environmental variable on the object information under the target task; extract a second target object feature affected by the preset environmental variable from the influence result, and determine a first selection condition for extracting the second target object feature from the object information; use a second selection condition complementary to the first selection condition to extract the first target object feature from the object information; an information recommendation module configured to determine a target recommended information among the pieces of information to be recommended based on each predicted relationship analysis result, and recommend the target recommended information to the target object.

12. A computing device, comprising: a memory and a processor; the memory is used for storing computer-executable instructions, and the processor is used for executing the computer-executable instructions, and when the computer-executable instructions are executed by the processor, the steps of the information recommendation method according to any one of claims 1 to 10 are implemented.

13. A computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps of the information recommendation method according to any one of claims 1 to 10 are implemented.

Images