Index change prediction model training method, index change prediction method and device thereof

Abstract

The application discloses an index change prediction model training method, an index change prediction method and devices thereof. The index change prediction model training method can generate a social network according to an obtained sample data set. After the social network is divided into different sub-networks, lower weights are assigned to interference edges based on each edge weight matrix, so as to remove the interference in the graph structure, and higher weights are assigned to key edges to retain causal information. The complete first graph structure is formed by adding and combining each social sub-graph structure. Further, the second graph structure is obtained by performing symmetry, feature smoothing and the like on the first graph structure. Then, the index change initial model is trained based on the weighted denoised second graph structure and the feature information of the nodes, to obtain the index change prediction model. Thus, the robustness of the causal effect estimation method can be effectively improved, so that the method can still maintain good prediction effect when the network is disturbed by noise.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method for training an indicator change prediction model, an indicator change prediction method, and an apparatus thereof. Background Technology

[0002] Currently, there are many scenarios that require analysis of changes in indicator data. These scenarios are typically handled by humans who, based on experience, predict potential changes in the indicator data after a specified time point, and then process the possible outcomes. However, this method is heavily influenced by human subjectivity, potentially leading to lower accuracy in predicting changes in indicator data.

[0003] Therefore, improving the accuracy of analysis of indicator data has become an urgent problem to be solved. Summary of the Invention

[0004] This application provides a method for training an indicator change prediction model, an indicator change prediction method, and an apparatus thereof to solve problems such as inaccurate indicator change prediction values.

[0005] In a first aspect, embodiments of this application provide a method for training an indicator change prediction model, the method comprising:

[0006] Obtain a sample dataset; wherein the sample dataset includes: a target item, feature information of multiple samples, social association information of the multiple samples, and actual change values ​​of attention indicators of the multiple samples; the multiple samples include: a first sample and a second sample, wherein the first sample uses the target item, and the second sample does not use the target item; the actual change values ​​of the attention indicators include: the initial value of the attention indicator and the actual value of the attention indicator after a preset time.

[0007] Based on the social association information of the multiple samples, a social network corresponding to the multiple samples is generated;

[0008] Based on the social networks corresponding to the multiple samples and the feature information of each of the multiple samples, multiple social subgraph structures are generated;

[0009] Based on the multiple social subgraph structures, the first graph structure corresponding to the multiple samples is obtained;

[0010] The first graph structure corresponding to the plurality of samples is smoothed and symmetrically processed to obtain the second graph structure corresponding to the plurality of samples;

[0011] Based on the second graph structure corresponding to the multiple samples, the feature information of each of the multiple samples, the actual change value of the attention index of the first sample, and the actual change value of the attention index of the second sample, the initial model of the indicator change is trained to obtain the indicator change prediction model.

[0012] Secondly, embodiments of this application provide a method for predicting changes in indicators, the method comprising:

[0013] Obtain the feature information, social association information, and initial values ​​of the metrics of interest for the target to be tested;

[0014] The feature information, social association information, and initial values ​​of the attention indicators of the target to be tested are input into the indicator change prediction model to obtain the predicted values ​​of the attention indicators of the target to be tested after using the target item after a preset time.

[0015] The indicator change prediction model is obtained by training the indicator change prediction model training method as described in any one of claims 1 to 7.

[0016] Thirdly, embodiments of this application provide a training apparatus for an indicator change prediction model, the apparatus comprising:

[0017] A sample acquisition module is used to acquire a sample dataset; wherein, the sample dataset includes: a target item, feature information of multiple samples, social association information of the multiple samples, and actual change values ​​of attention indicators of the multiple samples; the multiple samples include: a first sample and a second sample, wherein the first sample uses the target item, and the second sample does not use the target item; the actual change values ​​of the attention indicators include: the initial value of the attention indicator and the actual value of the attention indicator after a preset time.

[0018] A social network generation module is used to generate social networks corresponding to the multiple samples based on the social association information of the multiple samples;

[0019] The social subgraph structure generation module is used to generate multiple social subgraph structures based on the social networks corresponding to the multiple samples and the feature information of the multiple samples respectively.

[0020] The first graph structure acquisition module is used to obtain the first graph structure corresponding to the multiple samples based on the multiple social subgraph structures;

[0021] The second graph structure acquisition module is used to perform smoothing and symmetry processing on the first graph structure corresponding to the multiple samples to obtain the second graph structure corresponding to the multiple samples.

[0022] The indicator change prediction model acquisition module is used to train the initial indicator change model based on the second graph structure corresponding to the multiple samples, the feature information of each of the multiple samples, the actual change value of the attention indicator of the first sample, and the actual change value of the attention indicator of the second sample, to obtain the indicator change prediction model.

[0023] Fourthly, embodiments of this application provide a training apparatus for an indicator change prediction model, the apparatus comprising:

[0024] The target information acquisition module is used to acquire the target's feature information, social association information, and initial values ​​of the attention indicators.

[0025] The prediction value acquisition module is used to input the feature information, social association information, and initial values ​​of the attention indicators of the target to be tested into the indicator change prediction model to obtain the predicted value of the attention indicators of the target to be tested after using the target item after a preset time.

[0026] The indicator change prediction model is obtained by training the indicator change prediction model training method as described in any one of claims 1 to 7.

[0027] Fifthly, embodiments of this application provide a computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the above-described method steps.

[0028] Sixthly, embodiments of this application provide an electronic device, which may include: a processor and a memory;

[0029] The memory stores a computer program adapted to be loaded by the processor and executed in accordance with the above-described method steps.

[0030] The beneficial effects of the technical solutions provided in some embodiments of this application include at least the following:

[0031] This embodiment of the application can generate a social network based on the obtained sample dataset. After dividing the social network into different sub-networks, lower weights are assigned to interfering edges based on the edge weight matrix to remove interference from the graph structure, while higher weights are assigned to key edges to retain causal information. The weighted social sub-graph structures are then combined to form a complete first graph structure. Further, the first graph structure is processed with symmetry, feature smoothing, etc., to obtain a second graph structure. Finally, the initial model for index changes is trained based on the weighted and denoised second graph structure and the feature information of the nodes to obtain an index change prediction model. Therefore, this embodiment of the application can effectively learn a clean graph structure from a disturbed network and retain the potential causal information in the graph structure, thereby improving the robustness of the causal effect estimation method and enabling it to maintain good prediction performance even when the network is disturbed by noise. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 A flowchart illustrating a method for training an indicator change prediction model provided in an embodiment of this application;

[0034] Figure 2 A schematic diagram of a social network in a method for training an indicator change prediction model provided in an embodiment of this application;

[0035] Figure 3 This is a schematic diagram of a portion of the training process of an indicator change prediction model in an embodiment of this application.

[0036] Figure 4 This is a schematic diagram of another part of the training process of the indicator change prediction model in the indicator change prediction model training method provided in the embodiments of this application;

[0037] Figure 5 A flowchart illustrating an indicator change prediction method provided in an embodiment of this application;

[0038] Figure 6 This application scenario diagram illustrates an indicator change prediction method provided in an embodiment of this application.

[0039] Figure 7 A schematic diagram of the structure of a training device for an indicator change prediction model provided in an embodiment of this application;

[0040] Figure 8 This is a schematic diagram of the structure of an indicator change prediction device provided in an embodiment of this application;

[0041] Figure 9 A schematic diagram of the structure of an electronic device including a training program for a predictive model of index changes, provided in an embodiment of this application;

[0042] Figure 10 A schematic diagram of the structure of an electronic device including an index change prediction program provided in an embodiment of this application. Detailed Implementation

[0043] In the following description, when referring to the accompanying drawings, the same numbers in different drawings denote the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0044] In the description of this application, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances. Furthermore, in the description of this application, unless otherwise stated, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0045] In recent years, research on causal relationships in deep learning has received widespread attention. Among these studies, causal effect estimation is a crucial area, aiming to identify confounding variables and discover the causal relationship between the applied influence and the outcome. For example, in summer, ice cream shops experience higher electricity costs and better ice cream sales. If predictions are based solely on this data, the model might incorrectly link electricity costs to sales. However, the common cause of both increased electricity costs and ice cream sales is rising temperature, which is considered a confounding variable.

[0046] Research on causal effect estimation is of great significance to the development of many fields, including economics, education, and healthcare. For example, in advertising, causal effect estimation can predict the probability of users making a purchase after an advertisement is placed, thereby increasing the return on advertising. In the medical field, it can predict changes in patients' attention indicators after medication, thus helping to select the best treatment plan.

[0047] Randomized controlled trials are an effective method for estimating causal effects. Keeping all other conditions the same, by setting up experimental and control groups, the differences in outcomes between the two groups after applying different effects are observed. Observational data typically contains numerous entities and rich characteristic information. For example, suppose we are studying the impact of education policies on the test scores of a group of students. We can randomly assign some students to an experimental group and the other to a control group. The experimental group receives an intervention under a certain education policy, while the control group receives no intervention. Observations show that the average score of the experimental group is 80 points, and the average score of the control group is 75 points. However, we cannot conclude a causal effect of the education policy on student performance solely from these observations, because the results are influenced by multiple factors. Instead, we can define a causal effect by comparing potential outcomes. Assuming that without intervention, the potential outcome for the experimental group is 85 points, and the potential outcome for the control group is 80 points, we can conclude that the causal effect of the education policy on student performance is 5 points.

[0048] Next, based on the causal effect estimation introduced above, we will introduce a training method for an indicator change prediction model provided in the embodiments of this application.

[0049] In one embodiment, Figure 1 As shown, this application provides a flowchart illustrating a method for training an indicator change prediction model. Figure 1 As shown, the training method for this indicator change prediction model may include the following steps:

[0050] S101, Obtain the sample dataset.

[0051] The sample dataset in this embodiment may include: a target item, feature information of multiple samples, social association information of multiple samples, and actual changes in attention metrics for each of the multiple samples. The multiple samples may include: a first sample and a second sample. The first sample uses the target item, while the second sample does not. The actual changes in attention metrics include: the initial value of the attention metrics and the actual value of the attention metrics after a preset time.

[0052] Possibly, the target item in this application embodiment can be medicine, reagent, health product, etc. The sample can be a person participating in model training. The characteristic information of the sample can include: the person's name, gender, age, occupation, past medical history, etc. The social association information of the samples represents the social relationship between the samples. For example, sample person 1 and sample person 2 are friends in social communication software A, can communicate with each other, and have a social relationship; sample person 1 and sample person 3 are not friends in social communication software A, cannot communicate with each other, and have no social relationship.

[0053] In this study, the first sample represents the participants in the experimental group, and the second sample represents the participants in the control group. The first sample using the target item indicates participants who used the target item, for example, participant 1 took medication a. The second sample not using the target item indicates participants who did not use the target item, for example, participant 2 did not take medication a. The actual change value of the indicator of interest can include: the initial value of the indicator of interest and the actual value of the indicator of interest after a preset time. For example, the initial value of the indicator of interest could be the blood glucose level of participant 1 before taking medication a, and the actual value of the indicator of interest could be the blood glucose level of participant 1 half an hour after taking medication a. Alternatively, the initial value of the indicator of interest could be the blood glucose level of participant 2 who did not take medication a, and the actual value of the indicator of interest could be the blood glucose level of participant 2 half an hour later.

[0054] S102, Based on the social association information of multiple samples, generate social networks corresponding to multiple samples.

[0055] Understandably, in real-world applications, entities may have certain relationships, meaning they are interconnected, and these connections can be represented by networks. For example, in social networks, users have friend relationships, and in document repositories, connections are established through citation relationships. Typically, these networks contain valuable causal information, such as hidden variables like social status and interpersonal relationships. These are generally not directly observable from an individual's external characteristics, health data, or daily life, but can be inferred through relational networks.

[0056] For example, the obtained sample dataset may include 100 employees randomly selected from Company A. The employees may or may not know each other. According to the embodiments of this application, a social network of relationships among the 100 employees can be established based on the relationship list of each employee.

[0057] S103 generates multiple social subgraph structures based on the social networks corresponding to multiple samples and the feature information of each sample.

[0058] Possibly, embodiments of this application can, based on the social association information of each sample, split the social networks corresponding to multiple samples to obtain a first sub-network corresponding to the first sample, a second sub-network corresponding to the second sample, and a third sub-network where there is a social association between the first sample and the second sample; based on the feature information of the first sample in the first sub-network, obtain a first edge weight matrix; based on the feature information of the second sample in the second sub-network, obtain a second edge weight matrix; based on the feature information of the first sample and the feature information of the second sample in the third sub-network, obtain a third edge weight matrix; based on the first edge weight matrix and the first sub-network, obtain a first social subgraph structure; based on the second edge weight matrix and the second sub-network, obtain a second social subgraph structure; based on the third edge weight matrix and the third sub-network, obtain a third social subgraph structure.

[0059] It is understood that, based on the social association information between the sample personnel, the social network can be split into two groups. The experimental group corresponding to the sample personnel who applied the item is divided into one group, namely the first sample. Then, based on the association relationship between the sample personnel in the experimental group, a first sub-network is obtained. The sample personnel in the control group corresponding to the sample personnel who did not apply the item are divided into one group, namely the second sample. Then, based on the association relationship between the sample personnel in the control group, a second sub-network is obtained. In addition, the embodiments of this application further divide the sample personnel in the experimental group and the control group who have an association relationship into another group, and based on the association relationship between the sample personnel in the experimental group and the sample personnel in the control group, a third sub-network is obtained.

[0060] Possibly, embodiments of this application may employ the Edge Reweighting of Multi-Subgraph (ERMS) method to calculate the weights of different types of edges in the network, assigning more weights to key edges and less weights to interfering edges. This effectively learns a clean graph structure from the perturbed network while preserving the potential causal information in the graph structure, thereby improving the robustness of the causal effect estimation method and enabling it to maintain good performance even when the network is subject to noise interference.

[0061] Specifically, embodiments of this application assume that the edges between different groups and within each group have different properties. See also Figure 2The diagram illustrates a social network. It is understandable that closely connected users in a social network tend to exhibit similar behaviors, making them more likely to make the same choices. Correspondingly, this leads to a higher edge density within the experimental group T (control group C) than the edge density between the experimental and control groups, but noise or interfering edges within each group are often randomly generated. Therefore, embodiments of this application can learn the weights of different types of edges separately and identify interfering edges that violate these characteristics based on the features of that type of edge, removing noise by assigning lower weights to the interfering edges.

[0062] Specifically, embodiments of this application can learn attention weights for different types of edges using a subgraph edge reweighting method, assigning higher weights to important edges and lower weights to interfering edges. For the causal effect estimation task, the adjacency matrix A of the graph G(A, X) corresponding to the social network can be divided into three subnetworks according to the edge type, denoted as the first subnetwork A. t Second subnetwork A c Third subnetwork A ct These three subnetworks are respectively composed of edges inside the Treatment group, edges inside the Control group, and edges inside both the Treatment group and the Control group.

[0063] Possibly, embodiments of this application can establish three learnable edge weight matrices in the following manner, namely: the first edge weight matrix W t The second side weight matrix W c The third side weight matrix W ct .

[0064] Understandably, analyzing attention weights helps in understanding the model's focus on the input. Attention weights can be calculated using the sigmoid function, which maps raw weight values ​​to between 0 and 1, representing the activation level of the input. By analyzing the distribution and changes in attention weights, we can discover the model's sensitivity to different features. For example, in image recognition tasks, if attention weights are concentrated on image edges or texture features, the model pays more attention to these features and ignores other irrelevant information. This analysis can optimize model design and parameter settings, improving model performance.

[0065] Therefore, in this embodiment of the application, the attention weights of the edges in each sub-network are learned to obtain the edge weight matrix, wherein the elements W in the edge weight matrix are... ij for:

[0066] W ij =Sigmoid(sim(x) i ,x j )),

[0067] Where, x i Let x represent the feature vector of the i-th sample person. j Let represent the feature vector of the j-th sample person. sim(·) is a function that measures vector similarity; specifically, in this embodiment, the dot product is used as the function of sim(·).

[0068] Furthermore, the weighted social subgraph structure can be represented as:

[0069]

[0070] Among them, * could be t, c, or ct.

[0071] When * is t, the first social subgraph structure It can be represented as Among them, A t For the first sub-network, W t This is the weight matrix of the first side;

[0072] When * is c, the second social subgraph structure It can be represented as Among them, A c For the second sub-network, W c This is the weight matrix for the second side;

[0073] When * is ct, the third social subgraph structure It can be represented as Among them, A ct For the third sub-network, W ct This is the weight matrix of the third side.

[0074] Possibly, in embodiments of this application, the first social subgraph structure, the second social subgraph structure, and the third social subgraph structure can be input into the edge density loss function to obtain the edge density loss value.

[0075] Specifically, the loss function (edge ​​density loss function) controlling the edge density distribution can be defined as follows:

[0076]

[0077] Where τ is a hyperparameter used to control the difference in edge density within and between groups. The larger τ is, the greater the difference in density within and between groups. d(·) represents the density function, which is used to measure the density of edges in the graph structure.

[0078] Specifically, the density function of the first social subgraph structure As shown below:

[0079]

[0080] Specifically, the density function of the second social subgraph structure:

[0081]

[0082] Specifically, the density function of the third social subgraph structure:

[0083]

[0084] Where, n t and n c These represent the sample sizes of the experimental group and the control group, respectively. This represents the element in the i-th row and j-th column of the adjacency matrix corresponding to the first social subgraph structure. This represents the element in the i-th row and j-th column of the adjacency matrix corresponding to the second social subgraph structure. This represents the element in the i-th row and j-th column of the adjacency matrix corresponding to the third social subgraph structure. Let represent the element in the i-th row and j-th column of the adjacency matrix corresponding to the first sub-network. This represents the element in the i-th row and j-th column of the adjacency matrix corresponding to the second sub-network. W represents the element in the i-th row and j-th column of the adjacency matrix corresponding to the third sub-network. t ij W represents the weight matrix of the first side. t The element in the i-th row and j-th column, W t ij yes The corresponding edge weights, W represents the weight matrix of the second side. c The element in the i-th row and j-th column, yes The corresponding edge weights, W represents the weight matrix of the second side. ct The element in the i-th row and j-th column, yes The corresponding edge weights.

[0085] S104, based on multiple social subgraph structures, obtains the first graph structure corresponding to multiple samples.

[0086] Specifically, embodiments of this application can utilize the first social subgraph structure learned in the above embodiments. Second social subgraph structure Third social subgraph structure The three social subgraph structures are merged to obtain the complete first graph structure.

[0087] Understandably, the main challenge in estimating causal effects using observed data is the presence of confounding factors—factors that cannot be directly observed in the data and thus affect the outcome. Many related techniques for causal effect estimation assume the absence of unobservable confounding factors in the data. However, in reality, the assumption of strong ignorability is impractical because hidden confounding factors often exist in the observed data and cannot be directly observed. Therefore, this application's embodiment combines social network structure with data for causal effect estimation, which can help further improve model performance.

[0088] Current causal effect estimation methods based on network structure are highly susceptible to network structure interference, and the data collected in real-world scenarios often contains a certain degree of noise. Therefore, the edge weight matrix used in this application can effectively remove noise information from sub-networks in the network data while simultaneously recovering the potential causal information contained therein. After calculating the weights of different types of edges in the network, more weights are assigned to key edges and less weights to interfering edges. This effectively learns a clean social subgraph structure from the perturbed network while preserving the potential causal information in the sub-network, thereby improving the robustness of the causal effect estimation method and enabling it to maintain good performance even when the network structure is disturbed by noise.

[0089] S105, perform smoothing and symmetry processing on the first graph structure corresponding to multiple samples to obtain the second graph structure corresponding to multiple samples.

[0090] Possibly, embodiments of this application may employ a symmetric loss function to adjust the first graph structure to obtain the minimum value of the symmetric loss; employ a feature smoothing loss function to adjust the first graph structure to obtain the minimum value of the feature smoothing loss; determine the first graph structure when both the symmetric loss and the feature smoothing loss are at their minimum values; and obtain the second graph structure corresponding to multiple samples based on the first graph structure when both the symmetric loss and the feature smoothing loss are at their minimum values.

[0091] It is understandable that by preserving the characteristic smoothness and symmetry of the graph structure, the impact of perturbations in the graph structure can be further reduced. Characteristic smoothness means making the characteristics of connected nodes in the graph structure more similar, and symmetry means making the adjacency matrix of the undirected graph structure a symmetric matrix.

[0092] Specifically, symmetric loss functions can be used respectively. and feature smoothing loss function To ensure the symmetry of the learned graph structure and the smoothness of its features, the following definitions are used:

[0093]

[0094]

[0095] in, The smaller the value, the more symmetrical the learned graph structure. X represents the feature matrix corresponding to the sample dataset. In the expression yes The graph Laplace matrix, D is The degree matrix can be obtained by using the Laplace regularization matrix. This allows feature smoothness to be independent of node degree. It can also be converted to the following form:

[0096]

[0097] Where, d i The degree of node i, also known as the associativity, refers to the number of edges associated with that node. It can measure the similarity of features between connected nodes; the smaller the value, the more similar the features between connected nodes.

[0098] Therefore, embodiments of this application can be modified by adjusting the structure of the first figure to achieve the desired effect. functions and The function output value is minimized, thus obtaining a second graph structure with smooth features and symmetry.

[0099] S106. Based on the second graph structure corresponding to multiple samples, the feature information of each sample, the actual change value of the focus indicator of the first sample, and the actual change value of the focus indicator of the second sample, the initial model of indicator change is trained to obtain the indicator change prediction model.

[0100] Possibly, embodiments of this application may obtain the representation vector of the first sample based on the second graph structure corresponding to multiple samples and the feature information of the first sample; obtain the representation vector of the second sample based on the second graph structure corresponding to multiple samples and the feature information of the second sample; obtain the predicted change value of the attention indicator of the first sample after a preset time based on the representation vector of the first sample; obtain the predicted value of the attention indicator of the second sample after a preset time based on the representation vector of the second sample; and train the initial model of indicator change based on the predicted value of the attention indicator of the first sample after a preset time, the predicted value of the attention indicator of the second sample, the actual value of the attention indicator of the first sample, and the actual value of the attention indicator of the second sample to obtain the indicator change prediction model.

[0101] Specifically, the embodiments of this application can be implemented through the structure shown in the second figure. We learn the feature representation of the dataset and perform causal effect estimation using the feature matrix X corresponding to the sample dataset.

[0102] Specifically, in this embodiment of the application, the representation vector matrix corresponding to the sample dataset can be obtained through a graph convolutional neural network, which can be represented in the following form:

[0103]

[0104] Where θ is the parameter of the graph convolutional neural network function g(·), r n These are the elements in the representation vector matrix corresponding to the sample dataset;

[0105] The distributions of the representation vector P of the first sample and the representation vector Q of the second sample can be expressed in the following form:

[0106]

[0107]

[0108] Where X1 represents the feature matrix corresponding to the first sample, p n X1 represents the element in the representation vector matrix corresponding to the first sample, and X2 represents the feature matrix corresponding to the second sample.

[0109] Where X1 represents the feature matrix corresponding to the first sample, p n X1 represents the element in the representation vector matrix corresponding to the first sample, X2 represents the feature matrix corresponding to the second sample, and q represents the element in the representation vector matrix corresponding to the first sample. n These are the elements in the representation vector matrix corresponding to the second sample.

[0110] Furthermore, in this embodiment, the representation vectors corresponding to the sample dataset can be used to estimate the output results (i.e., the predicted values ​​corresponding to each sample person) through a feedforward neural network f, as shown below:

[0111]

[0112] Where t=0 represents the second sample without intervention (i.e., the control group), and t=1 represents the first sample with intervention (i.e., the experimental group). This represents the predicted value for each sample of individuals.

[0113] Possibly, embodiments of this application can obtain an error loss value based on an error loss function, the predicted values ​​of the indicators of interest for multiple samples, and the actual values ​​of the indicators of interest for multiple samples; obtain a distribution difference loss value based on a distribution difference function, the representation vector of the first sample, and the representation vector of the second sample; and train an initial model for indicator change based on the error loss value, the distribution difference loss value, the symmetry loss value, the feature smoothing loss value, and the edge density loss value to obtain an indicator change prediction model.

[0114] Specifically, the expression for the error loss function in this application embodiment is as follows:

[0115]

[0116] Among them, y i Here, n represents the actual value corresponding to each sample person, and n is the number of samples in the sample dataset. This represents the error loss value.

[0117] In addition to minimizing the error of the prediction results during the model training process, in order to ensure the balance between the experimental group and the control group, this application also introduces the Integral Probability Metric (IPM) to measure the imbalance of data distribution between the experimental group and the control group.

[0118] Specifically, the expression for the distribution difference loss function of IPM in this embodiment is as follows:

[0119]

[0120] Where ρ(·,·) is the distribution difference loss function that measures the difference between distributions. This represents the loss value due to distributional differences.

[0121] Possibly, the distribution difference loss function in the embodiments of this application can be implemented using the Wasserstein distance.

[0122] It is understood that the embodiments of this application can adjust the loss value of the initial model for index change by adjusting the loss value until the loss value can no longer be reduced, indicating that the training is complete, that is, the index change prediction model is obtained.

[0123] Specifically, embodiments of this application can be based on error loss values. Distribution difference loss value Symmetric loss value Feature smoothing loss value and edge density loss value Determine the loss value of the initial model for changes in indicators.

[0124] Furthermore, in this embodiment of the application, the loss value of the initial model due to changes in the index can be calculated by setting the objective function of ERMS. The objective function is as follows:

[0125]

[0126] Where α = 10 -3 β=5, γ=10 3λ = 10 -1 objective function Clean graph structures can be learned in conjunction with causal effect estimation tasks, making the learned graph structures more suitable for downstream causal effect estimation tasks. For example, based on the changes in indicators after medication, the subsequent treatment plan for the sample can be determined.

[0127] In a specific example, a sample dataset of 100 individuals is obtained. This dataset includes: the blood glucose-lowering drug b, the age, occupation, gender, and other characteristics of each individual, their Weibo (e.g., mutual following relationships) information, and the actual changes in the following indicators for each individual. The first group of participants took blood glucose-lowering drug b (experimental group), while the second group did not (control group). The actual changes in the following indicators include the initial blood glucose level and the actual blood glucose level one hour later.

[0128] For a more convenient example than the one above, please refer to... Figure 3 The diagram illustrates a portion of the training process of the indicator change prediction model in this embodiment of the application. First, a social network corresponding to the sample individuals can be generated based on their Weibo association information. This social network is then split into sub-network 1 (experimental group), sub-network 2 (control group), and sub-network 3 (Weibo users in the experimental and control groups who follow each other). A first social subgraph structure is obtained based on sub-network 1 and edge weight matrix 1; a second social subgraph structure is obtained based on sub-network 2 and edge weight matrix 2; and a third social subgraph structure is obtained based on sub-network 3 and edge weight matrix 3. Merging the first, second, and third social subgraph structures yields the first graph structure corresponding to 100 sample individuals.

[0129] See also Figure 4 The diagram illustrates another part of the training process of the indicator change prediction model in this embodiment of the application. In this embodiment, the first graph structure can be smoothed and symmetrically processed to obtain a second graph structure. Then, the second graph structure and the feature information of the sample personnel in the sample dataset are input into the graph convolutional layer structure to obtain the representation vector corresponding to the sample dataset. Furthermore, the initial indicator change model can be trained by minimizing the loss value output by the objective function to obtain the indicator change prediction model. Wherein, This represents the predicted blood glucose level of the i-th sample in the control group one hour later. This represents the predicted blood glucose level of the i-th sample in the experimental group one hour later.

[0130] It is understandable that, assuming that the initial blood glucose value of sample person 1 is 100, and one hour after taking hypoglycemic drug b, the actual blood glucose value of sample person 1 is 85, and the predicted blood glucose value of sample person 1 is between 83 and 88, and the initial blood glucose value of sample person 2 is 95, and one hour later, the actual blood glucose value of sample person 2 is 91, and the predicted blood glucose value of sample person 2 is between 88 and 93, then it shows that the indicator change prediction model trained in the embodiments of this application meets expectations.

[0131] Therefore, this embodiment first employs a subgraph edge reweighting method. This involves dividing the social network into different sub-networks based on different groups, then learning the weight matrices for each sub-network. By learning the weights of edges in different groups, lower weights are assigned to interfering edges to remove interference from the graph structure, while higher weights are assigned to key edges to preserve causal information. Finally, the weighted social subgraph structures are combined to form a complete first graph structure. Then, using graph property exploration, constraints such as symmetry and feature smoothness are applied to ensure the weighted first graph structure conforms to the basic properties of a graph structure, resulting in a second graph structure. Finally, through representation learning and causal effect estimation, the weighted and denoised second graph structure obtained in the first two parts, along with the node feature information, is input into a graph convolutional neural network to learn representations of confounding factors and estimate causal effects. Furthermore, this embodiment also minimizes the Individual Product Model (IPM) to balance the imbalance in data distribution, thereby improving the accuracy of the model training output.

[0132] In some implementations... Figure 5 An exemplary flowchart of an indicator change prediction method provided in an embodiment of this application is shown. Figure 5 As shown, the method for predicting changes in indicators can include at least the following steps:

[0133] S501: Obtain the feature information, social association information, and initial values ​​of the indicators of interest for the target to be tested.

[0134] S502, input the feature information, social association information, and initial values ​​of the attention indicators of the target to be tested into the indicator change prediction model to obtain the predicted values ​​of the attention indicators of the target to be tested after using the target item after a preset time.

[0135] The indicator change prediction model is trained using the indicator change prediction model training method described in the above embodiments of this application.

[0136] In a specific example, doctors often need to make advance judgments about the effects of a certain drug on patients. However, different patients may have different reactions to the drug after taking it, meaning that the degree of change in the indicators that need to be monitored may be different. Doctors cannot accurately estimate the different reactions of each patient after taking the drug, and therefore cannot accurately determine whether the drug is suitable for that patient.

[0137] See Figure 6 The diagram illustrates an application scenario for the indicator change prediction method. The model inputs the age, gender, occupation, social connections, and initial blood glucose level of the person being tested, and outputs the predicted blood glucose level for that person.

[0138] Therefore, the embodiments of this application can adopt an indicator change prediction model, based on the feature information and social association information of the target to be tested, to obtain the predicted value of the indicator that the target to be tested will pay attention to after using the target item after a preset time, thereby solving the problem that the prediction method in related technologies is greatly affected by subjective consciousness factors, resulting in inaccurate prediction results of indicator data changes.

[0139] Figure 7 This is a schematic diagram of the structure of an indicator change prediction model training device provided in an exemplary embodiment of this application. This indicator change prediction model training device can be installed in a server or other device to execute the indicator change prediction model training method of any of the above embodiments of this application. Figure 7 As shown, the training device for the indicator change prediction model may include:

[0140] The sample acquisition module 71 is used to acquire a sample dataset; wherein, the sample dataset includes: a target item, feature information of multiple samples, social association information of the multiple samples, and actual change values ​​of attention indicators of the multiple samples; the multiple samples include: a first sample and a second sample, wherein the first sample uses the target item, and the second sample does not use the target item; the actual change values ​​of the attention indicators include: the initial value of the attention indicator and the actual value of the attention indicator after a preset time;

[0141] The social network generation module 72 is used to generate a social network corresponding to the multiple samples based on the social association information of the multiple samples;

[0142] The social subgraph structure generation module 73 is used to generate multiple social subgraph structures based on the social networks corresponding to the multiple samples and the feature information of the multiple samples respectively;

[0143] The first graph structure acquisition module 74 is used to obtain the first graph structure corresponding to the multiple samples based on the multiple social subgraph structures;

[0144] The second graph structure acquisition module 75 is used to perform smoothing and symmetry processing on the first graph structure corresponding to the plurality of samples to obtain the second graph structure corresponding to the plurality of samples.

[0145] The indicator change prediction model acquisition module 76 is used to train the initial indicator change model based on the second graph structure corresponding to the multiple samples, the feature information of each of the multiple samples, the actual change value of the attention indicator of the first sample, and the actual change value of the attention indicator of the second sample, to obtain the indicator change prediction model.

[0146] In some embodiments, the social subgraph structure generation module 73 includes:

[0147] The sub-network obtaining unit is used to split the social network corresponding to the multiple samples based on the social association information of each sample, and obtain the first sub-network corresponding to the first sample, the second sub-network corresponding to the second sample, and the third sub-network where there is a social association between the first sample and the second sample.

[0148] The edge weight matrix obtaining unit is used to obtain a first edge weight matrix based on the feature information of the first sample in the first sub-network; to obtain a second edge weight matrix based on the feature information of the second sample in the second sub-network; and to obtain a third edge weight matrix based on the feature information of the first sample and the feature information of the second sample in the third sub-network.

[0149] A social subgraph structure obtaining unit is used to obtain a first social subgraph structure based on the first edge weight matrix and the first subnetwork; to obtain a second social subgraph structure based on the second edge weight matrix and the second subnetwork; and to obtain a third social subgraph structure based on the third edge weight matrix and the third subnetwork.

[0150] In some embodiments, the apparatus further includes:

[0151] The edge density loss module is used to input the first social subgraph structure, the second social subgraph structure, and the third social subgraph structure into the edge density loss function to obtain the edge density loss value.

[0152] In some embodiments, the module 75 for obtaining the second graph structure includes:

[0153] The first adjustment unit is used to adjust the first graph structure using a symmetric loss function to obtain the minimum value of the symmetric loss; and to adjust the first graph structure using a feature smoothing loss function to obtain the minimum value of the feature smoothing loss.

[0154] The first graph structure determination unit is used to determine the first graph structure when both the symmetric loss value and the feature smoothing loss value are at their minimum values.

[0155] The second graph structure obtaining unit is used to obtain the second graph structure corresponding to the multiple samples based on the first graph structure when both the symmetric loss value and the feature smoothing loss value are at their minimum values.

[0156] In some implementations, the indicator change prediction model obtaining module 76 includes:

[0157] The representation vector obtaining unit is used to obtain the representation vector of the first sample based on the second graph structure corresponding to the plurality of samples and the feature information of the first sample; and to obtain the representation vector of the second sample based on the second graph structure corresponding to the plurality of samples and the feature information of the second sample.

[0158] The prediction value acquisition unit is used to obtain the predicted change value of the attention index of the first sample after a preset time based on the representation vector of the first sample; and to obtain the predicted value of the attention index of the second sample after a preset time based on the representation vector of the second sample.

[0159] The indicator change prediction model obtaining unit is used to train the initial indicator change model based on the predicted value of the indicator of concern of the first sample after a preset time, the predicted value of the indicator of concern of the second sample, the actual value of the indicator of concern of the first sample, and the actual value of the indicator of concern of the second sample, to obtain the indicator change prediction model.

[0160] In some implementations, the indicator change prediction model obtains a unit, including:

[0161] The error loss value sub-unit is used to obtain the error loss value based on the error loss function, the predicted values ​​of the attention indicators of the multiple samples, and the actual values ​​of the attention indicators of the multiple samples.

[0162] The distribution difference loss value is obtained by sub-units, which are used to obtain the distribution difference loss value based on the distribution difference function, the representation vector of the first sample, and the representation vector of the second sample.

[0163] The indicator change prediction model obtains a sub-unit, which is used to train the initial indicator change model based on the error loss value, the distribution difference loss value, the symmetry loss value, the feature smoothing loss value, and the edge density loss value, to obtain the indicator change prediction model.

[0164] In some implementations, the indicator change prediction model obtains a sub-unit, specifically used for: determining the loss value of the initial indicator change model based on the error loss value, the distribution difference loss value, the symmetry loss value, the feature smoothing loss value, and the edge density loss value; adjusting the graph structures in the initial indicator change model until the loss value of the initial indicator change model no longer decreases, thereby obtaining the indicator change prediction model.

[0165] It should be noted that the above embodiments of the indicator change prediction model training device are only illustrated by the division of the above functional modules when executing the indicator change prediction model training method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the indicator change prediction model training device and the indicator change prediction model training method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0166] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0167] Figure 8 This is a schematic diagram of the structure of an indicator change prediction device provided in an exemplary embodiment of this application. This indicator change prediction device can be installed in a terminal or other device to execute the indicator change prediction method of any of the embodiments described above. Figure 8 As shown, the indicator change prediction device may include:

[0168] The target information acquisition module is used to acquire the target's feature information, social association information, and initial values ​​of the attention indicators.

[0169] The prediction value acquisition module is used to input the feature information, social association information, and initial values ​​of the attention indicators of the target to be tested into the indicator change prediction model to obtain the predicted value of the attention indicators of the target to be tested after using the target item after a preset time.

[0170] The indicator change prediction model is obtained by training the indicator change prediction model training method described in this application.

[0171] It should be noted that the indicator change prediction device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the indicator change prediction method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the indicator change prediction device and the indicator change prediction method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0172] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0173] Please see Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 9 As shown, the electronic device 90 may include: at least one processor 91, at least one network interface 94, a user interface 93, a memory 95, and at least one communication bus 92.

[0174] The communication bus 92 is used to enable communication between these components.

[0175] The user interface 93 may include a display screen and a camera. Optionally, the user interface 93 may also include a standard wired interface and a wireless interface.

[0176] Among them, the network interface 94 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0177] The processor 91 may include one or more processing cores. The processor 91 connects to various parts within the electronic device 90 using various interfaces and lines, and performs various functions and processes data of the electronic device 90 by running or executing instructions, programs, code sets, or instruction sets stored in the memory 95, and by calling data stored in the memory 95. Optionally, the processor 91 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 91 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display on the screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 91 and may be implemented as a separate chip.

[0178] The memory 95 may include random access memory (RAM) or read-only memory. Optionally, the memory 95 may include a non-transitory computer-readable storage medium. The memory 95 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 95 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 95 may also be at least one storage device located remotely from the aforementioned processor 91. Figure 9 As shown, the memory 95, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an information push application.

[0179] exist Figure 9In the electronic device 90 shown, the user interface 93 is mainly used to provide an input interface for the user and to obtain the user's input data; while the processor 91 can be used to call the indicator change prediction model training application stored in the memory 95, and specifically perform the following operations:

[0180] Obtain a sample dataset; wherein the sample dataset includes: a target item, feature information of multiple samples, social association information of the multiple samples, and actual change values ​​of attention indicators of the multiple samples; the multiple samples include: a first sample and a second sample, wherein the first sample uses the target item, and the second sample does not use the target item; the actual change values ​​of the attention indicators include: the initial value of the attention indicator and the actual value of the attention indicator after a preset time.

[0181] Based on the social association information of the multiple samples, a social network corresponding to the multiple samples is generated;

[0182] Based on the social networks corresponding to the multiple samples and the feature information of each of the multiple samples, multiple social subgraph structures are generated;

[0183] Based on the multiple social subgraph structures, the first graph structure corresponding to the multiple samples is obtained;

[0184] The first graph structure corresponding to the plurality of samples is smoothed and symmetrically processed to obtain the second graph structure corresponding to the plurality of samples;

[0185] Based on the second graph structure corresponding to the multiple samples, the feature information of each of the multiple samples, the actual change value of the attention index of the first sample, and the actual change value of the attention index of the second sample, the initial model of the indicator change is trained to obtain the indicator change prediction model.

[0186] In some implementations, when the processor 91 generates multiple social subgraph structures based on the social networks corresponding to the multiple samples and the feature information of each of the multiple samples, it specifically performs the following:

[0187] Based on the social association information of each sample, the social networks corresponding to the multiple samples are split to obtain a first sub-network corresponding to the first sample, a second sub-network corresponding to the second sample, and a third sub-network in which there is a social association between the first sample and the second sample.

[0188] Based on the feature information of the first sample in the first sub-network, a first side weight matrix is ​​obtained; based on the feature information of the second sample in the second sub-network, a second side weight matrix is ​​obtained; based on the feature information of the first sample and the feature information of the second sample in the third sub-network, a third side weight matrix is ​​obtained.

[0189] Based on the first edge weight matrix and the first sub-network, a first social subgraph structure is obtained; based on the second edge weight matrix and the second sub-network, a second social subgraph structure is obtained; based on the third edge weight matrix and the third sub-network, a third social subgraph structure is obtained.

[0190] In some implementations, the processor 91 also performs:

[0191] The first social subgraph structure, the second social subgraph structure, and the third social subgraph structure are input into the edge density loss function to obtain the edge density loss value.

[0192] In some embodiments, when the processor 91 performs smoothing and symmetry processing on the first graph structure corresponding to the plurality of samples to obtain the second graph structure corresponding to the plurality of samples, it specifically performs the following:

[0193] The first graph structure is adjusted using a symmetric loss function to obtain the minimum value of the symmetric loss; the first graph structure is also adjusted using a feature smoothing loss function to obtain the minimum value of the feature smoothing loss.

[0194] The first graph structure is determined when both the symmetric loss value and the feature smoothing loss value are minimized.

[0195] Based on the first graph structure where both the symmetric loss value and the feature smoothing loss value are at their minimum values, the second graph structure corresponding to the multiple samples is obtained.

[0196] In some implementations, when the processor 91 trains the initial model for indicator change based on the second graph structure corresponding to the plurality of samples, the feature information of each of the plurality of samples, the actual change value of the attention indicator of the first sample, and the actual change value of the attention indicator of the second sample to obtain the indicator change prediction model, it specifically performs the following:

[0197] Based on the second graph structure corresponding to the multiple samples and the feature information of the first sample, the representation vector of the first sample is obtained; based on the second graph structure corresponding to the multiple samples and the feature information of the second sample, the representation vector of the second sample is obtained.

[0198] Based on the representation vector of the first sample, the predicted change value of the attention index of the first sample after a preset time is obtained; based on the representation vector of the second sample, the predicted value of the attention index of the second sample after a preset time is obtained.

[0199] Based on the predicted values ​​of the attention indicators of the first sample, the predicted values ​​of the attention indicators of the second sample, the actual values ​​of the attention indicators of the first sample, and the actual values ​​of the attention indicators of the second sample after a preset time, the initial model of the indicator change is trained to obtain the indicator change prediction model.

[0200] In some implementations, when the processor 91 trains the initial model for indicator change based on the predicted values ​​of the attention indicators of the first sample, the predicted values ​​of the attention indicators of the second sample, the actual values ​​of the attention indicators of the first sample, and the actual values ​​of the attention indicators of the second sample after a preset time, to obtain the indicator change prediction model, it specifically performs the following:

[0201] The error loss value is obtained based on the error loss function, the predicted values ​​of the attention indicators of the multiple samples, and the actual values ​​of the attention indicators of the multiple samples.

[0202] Based on the distribution difference function, the representation vector of the first sample, and the representation vector of the second sample, the distribution difference loss value is obtained.

[0203] Based on the error loss value, the distribution difference loss value, the symmetry loss value, the feature smoothing loss value, and the edge density loss value, the initial model for index change is trained to obtain the index change prediction model.

[0204] In some implementations, when the processor 91 trains the initial model for index change based on the error loss value, the distribution difference loss value, the symmetry loss value, the feature smoothing loss value, and the edge density loss value to obtain the index change prediction model, it specifically performs the following:

[0205] Based on the error loss value, the distribution difference loss value, the symmetry loss value, the feature smoothing loss value, and the edge density loss value, the loss value of the initial model for the index change is determined;

[0206] The structure of each graph in the initial model of the indicator change is adjusted until the loss value of the initial model of the indicator change no longer decreases, thus obtaining the indicator change prediction model.

[0207] Please see Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 10 As shown, the electronic device 10 may include: at least one processor 11, at least one network interface 14, user interface 13, memory 15, and at least one communication bus 12.

[0208] The communication bus 12 is used to enable communication between these components.

[0209] The user interface 13 may include a display screen and a camera. Optionally, the user interface 13 may also include a standard wired interface and a wireless interface.

[0210] The network interface 14 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0211] The processor 11 may include one or more processing cores. The processor 11 connects to various parts within the electronic device 10 using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 15, and by calling data stored in the memory 15. Optionally, the processor 11 may be implemented using at least one hardware form of digital signal processing, field-programmable gate array, or programmable logic array. The processor 11 may integrate one or more of the following: a central processing unit (CPU), a graphics processing unit (GPU), and a modem. Specifically, the CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 11 and may be implemented as a separate chip.

[0212] The memory 15 may include random access memory (RAM) or read-only memory (ROM). Optionally, the memory 15 may include a non-transitory computer-readable medium. The memory 15 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 15 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), instructions for implementing the various method embodiments described above, etc.; the data storage area may store data involved in the various method embodiments described above, etc. Optionally, the memory 15 may also be at least one storage device located remotely from the aforementioned processor 11. Figure 10 As shown, the memory 15, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an information push application.

[0213] exist Figure 10 In the electronic device 10 shown, the user interface 13 is mainly used to provide an input interface for the user and to obtain the user's input data; while the processor 11 can be used to call the indicator change prediction application stored in the memory 15 and specifically perform the following operations:

[0214] Obtain the feature information, social association information, and initial values ​​of the metrics of interest for the target to be tested;

[0215] The feature information, social association information, and initial values ​​of the attention indicators of the target to be tested are input into the indicator change prediction model to obtain the predicted values ​​of the attention indicators of the target to be tested after using the target item after a preset time.

[0216] The indicator change prediction model is trained using the indicator change prediction model training method described above in this application.

[0217] This application also provides a computer-readable storage medium storing instructions that, when executed on a computer or processor, cause the computer or processor to perform the above-described instructions. Figure 2 and Figure 5 One or more steps in the illustrated embodiment. If the constituent modules of the above-described information push device are implemented as software functional units and sold or used as independent products, they can be stored in the computer-readable storage medium.

[0218] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital versatile discs (DVDs)), or semiconductor media (e.g., solid-state drives (SSDs)).

[0219] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium includes various media capable of storing program code, such as Read Only Memory (ROM), Random Access Memory (RAM), magnetic disks, or optical disks. Unless otherwise specified, the technical features of this embodiment and its implementation schemes can be combined arbitrarily.

[0220] The embodiments described above are merely preferred embodiments of this application and are not intended to limit the scope of this application. Any modifications and improvements made by those skilled in the art to the technical solutions of this application without departing from the spirit of this application should fall within the protection scope defined by the claims of this application.

Claims

1. A method for training an indicator change prediction model, characterized in that, The method includes: Obtain a sample dataset; wherein the sample dataset includes a target item, feature information of multiple samples, social association information of the multiple samples, and actual change values ​​of attention indicators of the multiple samples. The target item includes hypoglycemic drugs. The feature information of the samples includes at least the age, occupation, and gender of the samples. The social association information of the multiple samples includes at least Weibo association information. The multiple samples include a first sample and a second sample. The first sample uses the target item, and the second sample does not use the target item. The first sample includes people who take the hypoglycemic drugs, and the second sample includes people who do not take the hypoglycemic drugs. The actual change values ​​of attention indicators include the initial value of the attention indicator and the actual value of the attention indicator after a preset time. Based on the social association information of the multiple samples, a social network corresponding to the multiple samples is generated; Based on the social networks corresponding to the multiple samples and the feature information of each of the multiple samples, multiple social subgraph structures are generated; Based on the multiple social subgraph structures, the first graph structure corresponding to the multiple samples is obtained; The first graph structure corresponding to the plurality of samples is smoothed and symmetrically processed to obtain the second graph structure corresponding to the plurality of samples; Based on the second graph structure corresponding to the multiple samples, the feature information of each of the multiple samples, the actual change value of the attention index of the first sample, and the actual change value of the attention index of the second sample, the initial model of the indicator change is trained to obtain the indicator change prediction model.

2. The method as described in claim 1, characterized in that, The generation of multiple social subgraph structures based on the social networks corresponding to the multiple samples and the feature information of each of the multiple samples includes: Based on the social association information of each sample, the social networks corresponding to the multiple samples are split to obtain a first sub-network corresponding to the first sample, a second sub-network corresponding to the second sample, and a third sub-network in which there is a social association between the first sample and the second sample. Based on the feature information of the first sample in the first sub-network, a first side weight matrix is ​​obtained; based on the feature information of the second sample in the second sub-network, a second side weight matrix is ​​obtained; based on the feature information of the first sample and the feature information of the second sample in the third sub-network, a third side weight matrix is ​​obtained. Based on the first edge weight matrix and the first sub-network, a first social subgraph structure is obtained; based on the second edge weight matrix and the second sub-network, a second social subgraph structure is obtained; based on the third edge weight matrix and the third sub-network, a third social subgraph structure is obtained.

3. The method as described in claim 2, characterized in that, The method further includes: The first social subgraph structure, the second social subgraph structure, and the third social subgraph structure are input into the edge density loss function to obtain the edge density loss value.

4. The method as described in claim 3, characterized in that, The step of smoothing and symmetrically processing the first graph structure corresponding to the plurality of samples to obtain the second graph structure corresponding to the plurality of samples includes: The first graph structure is adjusted using a symmetric loss function to obtain the minimum value of the symmetric loss; the first graph structure is also adjusted using a feature smoothing loss function to obtain the minimum value of the feature smoothing loss. The first graph structure is determined when both the symmetric loss value and the feature smoothing loss value are minimized. Based on the first graph structure where both the symmetric loss value and the feature smoothing loss value are at their minimum values, the second graph structure corresponding to the multiple samples is obtained.

5. The method as described in claim 4, characterized in that, The method of training the initial model of indicator change based on the second graph structure corresponding to the multiple samples, the feature information of each of the multiple samples, the actual change value of the attention indicator of the first sample, and the actual change value of the attention indicator of the second sample to obtain the indicator change prediction model includes: Based on the second graph structure corresponding to the multiple samples and the feature information of the first sample, the representation vector of the first sample is obtained; based on the second graph structure corresponding to the multiple samples and the feature information of the second sample, the representation vector of the second sample is obtained. Based on the representation vector of the first sample, the predicted change value of the attention index of the first sample after a preset time is obtained; based on the representation vector of the second sample, the predicted value of the attention index of the second sample after a preset time is obtained. Based on the predicted values ​​of the attention indicators of the first sample, the predicted values ​​of the attention indicators of the second sample, the actual values ​​of the attention indicators of the first sample, and the actual values ​​of the attention indicators of the second sample after a preset time, the initial model of the indicator change is trained to obtain the indicator change prediction model.

6. The method according to claim 5, characterized in that, The indicator change prediction model is trained based on the predicted values ​​of the attention indicators of the first sample after a preset time, the predicted values ​​of the attention indicators of the second sample, the actual values ​​of the attention indicators of the first sample, and the actual values ​​of the attention indicators of the second sample, to obtain the indicator change prediction model, including: The error loss value is obtained based on the error loss function, the predicted values ​​of the attention indicators of the multiple samples, and the actual values ​​of the attention indicators of the multiple samples. Based on the distribution difference function, the representation vector of the first sample, and the representation vector of the second sample, the distribution difference loss value is obtained. Based on the error loss value, the distribution difference loss value, the symmetry loss value, the feature smoothing loss value, and the edge density loss value, the initial model for index change is trained to obtain the index change prediction model.

7. The method according to claim 6, characterized in that, The error loss value, The distribution difference loss value, the symmetry loss value, the feature smoothing loss value, and the edge density loss value are used to train the initial model of index change to obtain the index change prediction model, including: Based on the error loss value, the distribution difference loss value, the symmetry loss value, the feature smoothing loss value, and the edge density loss value, the loss value of the initial model for the index change is determined; The structure of each graph in the initial model of the indicator change is adjusted until the loss value of the initial model of the indicator change no longer decreases, thus obtaining the indicator change prediction model.

8. A method for predicting changes in an indicator, characterized in that, The method includes: Obtain the feature information, social association information, and initial values ​​of the metrics of interest for the target to be tested; The feature information, social association information, and initial values ​​of the attention indicators of the target to be tested are input into the indicator change prediction model to obtain the predicted values ​​of the attention indicators of the target to be tested after using the target item after a preset time. The indicator change prediction model is trained using the indicator change prediction model training method as described in any one of claims 1 to 7.

9. A training device for a predictive model of index changes, characterized in that, The device includes: A sample acquisition module is used to acquire a sample dataset. The sample dataset includes a target item, feature information of multiple samples, social association information of the multiple samples, and actual change values ​​of attention indicators for each of the multiple samples. The target item includes a hypoglycemic drug. The feature information of the samples includes at least the sample's age, occupation, and gender. The social association information of the multiple samples includes at least Weibo association information. The multiple samples include a first sample and a second sample. The first sample uses the target item, while the second sample does not use the target item. The first sample includes individuals taking the hypoglycemic drug, and the second sample includes individuals not taking the hypoglycemic drug. The actual change values ​​of the attention indicators include the initial value of the attention indicator and the actual value of the attention indicator after a preset time. A social network generation module is used to generate a social network corresponding to the multiple samples based on the social association information of the multiple samples; The social subgraph structure generation module is used to generate multiple social subgraph structures based on the social networks corresponding to the multiple samples and the feature information of the multiple samples respectively. The first graph structure acquisition module is used to obtain the first graph structure corresponding to the multiple samples based on the multiple social subgraph structures; The second graph structure acquisition module is used to perform smoothing and symmetry processing on the first graph structure corresponding to the multiple samples to obtain the second graph structure corresponding to the multiple samples. The indicator change prediction model acquisition module is used to train the initial indicator change model based on the second graph structure corresponding to the multiple samples, the feature information of each of the multiple samples, the actual change value of the attention indicator of the first sample, and the actual change value of the attention indicator of the second sample, to obtain the indicator change prediction model.

10. A device for predicting changes in an indicator, characterized in that, The device includes: The target information acquisition module is used to acquire the target's feature information, social association information, and initial values ​​of the attention indicators. The prediction value acquisition module is used to input the feature information, social association information, and initial values ​​of the attention indicators of the target to be tested into the indicator change prediction model to obtain the predicted value of the attention indicators of the target to be tested after using the target item after a preset time. The indicator change prediction model is trained using the indicator change prediction model training method as described in any one of claims 1 to 7.

Images