Model training methods, population expansion methods, media, devices and computing equipment

Abstract

This disclosure provides a model training method, a crowd expansion method, a medium, an apparatus, and a computing device, relating to the field of computer technology. The model training method includes: acquiring seed users corresponding to media resources; determining a saliency index corresponding to preset media resource features based on the seed users and preset media resource features; identifying preset media resource features whose saliency index is greater than or equal to an index threshold as target features; and iteratively training a crowd expansion model based on the seed users, a first target group index corresponding to the target features of the seed users, and the target features, to obtain a trained crowd expansion model. This disclosure can obtain a more accurate crowd expansion model, and thus, when using the crowd expansion model for crowd expansion, it can more accurately obtain the expanded audience.

Description

Technical Field

[0001] The embodiments of this disclosure relate to the field of computer technology, and more specifically, the embodiments of this disclosure relate to a model training method, a population expansion method, a medium, an apparatus, and a computing device. Background Technology

[0002] This section is intended to provide background or context for the embodiments of this disclosure as set forth in the claims. The description herein is not intended to be a prior art simply because it is included in this section.

[0003] Audience expansion is often used for advertising and other media resource placement or marketing activities. For example, when placing an advertisement, expanding the audience based on the seed audience provided by the advertiser can effectively improve the click-through rate or purchase conversion rate of the advertisement.

[0004] Currently, population expansion is typically based on supervised learning algorithms. Specifically, a classification model is trained using a specified sample (i.e., the seed population) as positive samples and randomly selected samples from the remaining samples as negative samples. The trained classification model is then used to filter the candidate population to obtain the expanded population. However, this method cannot accurately obtain the expanded population. Summary of the Invention

[0005] This disclosure provides a model training method, a population expansion method, a medium, an apparatus, and a computing device to solve the problem that current methods cannot accurately obtain expanded population data.

[0006] In a first aspect of this disclosure, a model training method is provided, comprising:

[0007] Acquire seed users corresponding to media resources;

[0008] Based on the seed users and the characteristics of the preset media resources, determine the significance index corresponding to the characteristics of the preset media resources;

[0009] The target features are determined by identifying media resource features whose significance index is greater than or equal to the index threshold.

[0010] Based on seed users, the first target group index corresponding to the target features of seed users, and target features, the audience expansion model is iteratively trained to obtain a trained audience expansion model, which is used to expand the audience based on seed users.

[0011] Secondly, embodiments of this disclosure provide a method for expanding a user base, including:

[0012] Obtain candidate users corresponding to media resources;

[0013] Based on the candidate users and the seed users and target features corresponding to the audience expansion model, the target group index corresponding to the target features of the candidate users is determined. The audience expansion model is trained using the model training method described in the first aspect of this disclosure.

[0014] The candidate user and the target group index of the target feature corresponding to the candidate user are input into the population expansion model to obtain the predicted probability value of the candidate user. The predicted probability value is used to determine whether the candidate user is an expandable user.

[0015] Thirdly, embodiments of this disclosure provide a model training apparatus, comprising:

[0016] The acquisition module is used to acquire seed users corresponding to media resources;

[0017] The first determining module is used to determine the significance index corresponding to the preset media resource characteristics based on the seed users and preset media resource characteristics.

[0018] The second determining module is used to determine the preset media resource features whose significance index is greater than or equal to the index threshold as target features;

[0019] The training module is used to iteratively train the audience expansion model based on seed users, the first target group index corresponding to the seed users, and the target features to obtain a trained audience expansion model. The audience expansion model is used to expand the audience based on the seed users.

[0020] Fourthly, embodiments of this disclosure provide a crowd expansion device, comprising:

[0021] The acquisition module is used to acquire candidate users corresponding to media resources;

[0022] The determination module is used to determine the target group index of the target features corresponding to the candidate users based on the candidate users and the seed users and target features corresponding to the audience expansion model. The audience expansion model is trained using the model training method described in the first aspect of this disclosure.

[0023] The processing module is used to input the candidate user and the target group index of the target feature corresponding to the candidate user into the population expansion model to obtain the predicted probability value of the candidate user.

[0024] Fifthly, embodiments of this disclosure provide a computing device, including: a processor, and a memory communicatively connected to the processor;

[0025] The memory stores the instructions that the computer executes;

[0026] The processor executes computer execution instructions stored in memory to implement the model training method as described in the first aspect of this disclosure or the population expansion method as described in the second aspect.

[0027] Sixthly, embodiments of this disclosure provide a storage medium storing computer program instructions, which, when executed, implement the model training method as described in the first aspect or the population expansion method as described in the second aspect of this disclosure.

[0028] In a seventh aspect, embodiments of this disclosure provide a computer program product, including a computer program that, when executed by a processor, implements the model training method as described in the first aspect of this disclosure or the population expansion method as described in the second aspect.

[0029] The model training method, audience expansion method, medium, apparatus, and computing device provided in this disclosure acquire seed users corresponding to media resources, determine the salience index corresponding to the preset media resource features based on the seed users and preset media resource features, determine the preset media resource features whose salience index is greater than or equal to an index threshold as target features, and iteratively train the audience expansion model based on the seed users, the first target group index corresponding to the target features of the seed users, and the target features to obtain a trained audience expansion model. Because this disclosure determines the target features based on the salience index corresponding to the preset media resource features determined by the seed users and preset media resource features, it can obtain the target features more accurately. Furthermore, by training based on the target features, seed users, and the first target group index corresponding to the target features of the seed users to obtain a trained audience expansion model, a more accurate audience expansion model can be obtained. Therefore, when using the audience expansion model for audience expansion, it can more accurately obtain the expanded audience. Attached Figure Description

[0030] The above and other objects, features, and advantages of this disclosure will become readily apparent from the following detailed description of exemplary embodiments, taken in conjunction with the accompanying drawings. Several embodiments of this disclosure are illustrated in the drawings by way of example and not limitation, in which:

[0031] Figure 1 This is a schematic diagram of an application scenario provided by an embodiment of the present disclosure;

[0032] Figure 2 A flowchart of a model training method provided in an embodiment of this disclosure;

[0033] Figure 3 A flowchart of a model training method provided in another embodiment of this disclosure;

[0034] Figure 4 This is a schematic diagram illustrating iterative training of a population expansion model according to an embodiment of this disclosure;

[0035] Figure 5A flowchart illustrating a population expansion method provided in an embodiment of this disclosure;

[0036] Figure 6 This is a schematic diagram of the structure of a model training device provided in an embodiment of the present disclosure;

[0037] Figure 7 This is a schematic diagram of the structure of a crowd expansion device provided in an embodiment of the present disclosure;

[0038] Figure 8 A schematic diagram of a storage medium provided in an embodiment of this disclosure;

[0039] Figure 9 This is a schematic diagram of the structure of a computing device provided in an embodiment of the present disclosure.

[0040] In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts. Detailed Implementation

[0041] The principles and spirit of this disclosure will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are given merely to enable those skilled in the art to better understand and implement this disclosure, and are not intended to limit the scope of this disclosure in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.

[0042] Those skilled in the art will recognize that embodiments of this disclosure can be implemented as a system, apparatus, device, method, or computer program product. Therefore, this disclosure can be specifically implemented in the following forms: entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0043] According to embodiments of this disclosure, a model training method, a population expansion method, a medium, an apparatus, and a computing device are proposed.

[0044] In this article, it is important to understand the following terms:

[0045] Target Group Index (TGI): TGI = [the proportion of people in the target group who have a certain characteristic / the proportion of people in the whole who have the same characteristic] * number of norms 100.

[0046] Furthermore, the number of any elements in the accompanying drawings is for illustrative purposes only and not for limitation, and any naming is for distinction only and has no limiting meaning.

[0047] In addition, the data involved in this disclosure may be data authorized by the user or fully authorized by all parties. The collection, dissemination and use of the data shall comply with the requirements of relevant national laws and regulations. The implementation methods / executives of this disclosure may be combined with each other.

[0048] The principles and spirit of this disclosure will be explained in detail below with reference to several representative embodiments. Invention Overview

[0050] The inventors have discovered that, in a related technology, when expanding the population, the population can be expanded based on a semi-supervised learning algorithm or a supervised learning algorithm. Taking the population expansion based on a supervised learning algorithm as an example, specifically, a specified sample (i.e., the seed population) is used as a positive sample, and samples are randomly selected from the remaining samples as negative samples to train a classification model. Then, the candidate group is screened through the trained classification model to obtain the expanded population, and the feature importance is output based on the trained classification model. However, the above method cannot accurately obtain the expanded population, and has the following disadvantages: (1) Taking advertising as an example, since users do not provide feedback on the advertising content, it does not mean that users are not interested in the advertising content. Therefore, the method of randomly selecting samples as negative samples cannot obtain true negative samples. The ratio of the number of samples to the number of specified samples will directly affect the training and prediction of the classification model. In real-world scenarios, this ratio is often unknowable; (2) In training the classification model, it is necessary to adjust the model parameters based on the test set, which requires a lot of computing time and manual intervention. It is impossible to quickly achieve similarity calculation of a large-scale population, that is, it is impossible to quickly expand the population.

[0051] In another related technology, taking advertising as an example of media resources, users are collected for a certain advertising campaign. By sampling users multiple times and performing cluster analysis, the features used in each cluster are recorded. The importance of the features is evaluated based on the number of times they are used, and then the audience is expanded based on the important features. This related technology has the following main drawbacks: (1) The clustering method itself cannot identify the importance of features. It only records the number of times the features are used, but it cannot record the degree of differentiation of a single feature in the cluster. The evaluation of the importance of features is biased, which leads to the inability to accurately obtain the expanded audience when expanding the audience based on important features; (2) It requires multiple clustering of samples, which is particularly resource-intensive in big data environments.

[0052] To address the aforementioned issues, this disclosure provides a model training method, a population expansion method, a medium, a device, and a computing device. By using an unsupervised algorithm, the saliency of features is determined, and new feature values ​​are assigned to the features. A population expansion model is constructed based on seed users, and the weight of each feature is determined. The predicted probability value corresponding to the seed user is obtained through weighted summation. This results in a more accurate population expansion model, which, when used for population expansion, can more accurately obtain the expanded population.

[0053] Application Scenarios Overview

[0054] First refer to Figure 1 Examples of application scenarios for the solutions provided in this disclosure are given. Figure 1 This is a schematic diagram of an application scenario provided by an embodiment of the present disclosure, such as... Figure 1 As shown, this application scenario includes a server cluster 11 and a terminal 12. The server cluster 11 comprises multiple servers 111 and a storage device 112. The terminal 12 can be a tablet, laptop, desktop computer, or smart home appliance, etc. The servers 111 are used to train the population expansion model, retrieving data from the storage device 112 during training and storing the generated data in the storage device 112. Furthermore, during training, communication with the terminal 12 occurs via a wireless network or wired network.

[0055] Furthermore, the embodiments disclosed herein can be applied to audience expansion scenarios. For example, when conducting advertising campaigns, audience expansion can be carried out based on the seed audience provided by the advertiser.

[0056] It should be noted that, Figure 1 This is merely a schematic diagram illustrating one application scenario provided by an embodiment of this disclosure; this embodiment does not necessarily represent an application scenario. Figure 1 The included equipment is not limited, nor is it restricted. Figure 1 The positional relationships between the devices are defined. The model training method provided in this embodiment can be applied to a server, which can be a standalone server or a service cluster, etc.

[0057] Exemplary methods

[0058] The following is combined with Figure 1 Application scenarios, refer to Figure 2 This document describes a model training method according to exemplary embodiments of the present disclosure. It should be noted that the above application scenarios are shown only to facilitate understanding of the spirit and principles of the present disclosure, and the embodiments of the present disclosure are not limited in any way. Rather, the embodiments of the present disclosure can be applied to any applicable scenario.

[0059] First, the model training method will be introduced through specific examples.

[0060] Figure 2 This is a flowchart illustrating a model training method provided in one embodiment of this disclosure. The method of this embodiment can be applied to a computing device, which may be a server or a server cluster, etc. Figure 2 As shown, the method in this embodiment includes:

[0061] S201. Obtain the seed users corresponding to the media resources.

[0062] In this embodiment of the disclosure, the seed user corresponding to the media resource can be input by the user to the electronic device executing this method embodiment, or it can be sent by another device to the electronic device executing this method embodiment. For example, the media resource may be an advertisement. After the advertiser uploads the seed user corresponding to the advertisement to the electronic device executing this method embodiment, the seed user corresponding to the advertisement can be obtained.

[0063] S202. Based on the seed users and the preset media resource characteristics, determine the significance index corresponding to the preset media resource characteristics.

[0064] In this step, a preset media resource characteristic is, for example, the number of days a song has been played in the past 30 days. After obtaining the seed users corresponding to the media resource, a significance index corresponding to the preset media resource characteristic can be determined based on the seed users and the preset media resource characteristic. The significance index is used to characterize the significance of the preset media resource characteristic. For details on how to determine the significance index corresponding to the preset media resource characteristic based on the seed users and the preset media resource characteristic, please refer to the subsequent embodiments, which will not be repeated here.

[0065] S203. Determine the preset media resource features whose significance index is greater than or equal to the index threshold as target features.

[0066] In this step, the index threshold can be determined as needed, and this disclosure does not limit it. After determining the significance index corresponding to the preset media resource features, the preset media resource features can be sorted in descending order of significance index, and the preset media resource features with a significance index greater than or equal to the index threshold can be determined as target features.

[0067] Optionally, determining the target feature may include: displaying preset media resource features in descending order of salience index; and determining the target feature in response to a selection operation targeting a preset media resource feature.

[0068] For example, after determining the saliency index corresponding to the preset media resource feature, the preset media resource features can be displayed in descending order of saliency index. After the user selects the preset media resource feature with a higher saliency index, the target feature can be determined as the preset media resource feature selected by the user.

[0069] S204. Based on seed users, the first target group index corresponding to the target features of seed users, and target features, iteratively train the population expansion model to obtain a trained population expansion model.

[0070] Among them, the audience expansion model is used to expand the audience based on seed users.

[0071] In this step, the first target group index corresponding to the target features of the seed users can be obtained based on the seed users and preset media resource features, as detailed in subsequent embodiments. For example, taking advertising as an example, given the specific nature of advertising scenarios, only one batch of positive feedback user samples (i.e., seed users) can be collected after each advertising campaign. In this case, it is necessary to use a purely unsupervised learning algorithm to perform profile analysis on the seed users already acquired, and to find similar groups of people through core profile features (i.e., target features) for the next advertising campaign, thereby improving the effectiveness of the campaign. In advertising business scenarios, analyzing user profiles is only the first step in analyzing the effectiveness of advertising campaigns. More importantly, it is necessary to adjust the advertising strategy based on the current advertising effectiveness and then select the precise target audience, that is, to expand the audience through a lookalike method using the known advertising audience. This embodiment of the disclosure constructs an audience expansion model by obtaining target features with high significance indices from the seed population. The default label value for the specified sample (i.e., seed user) is 1, meaning that the probability of this batch of samples being positive samples is 100%. In the construction and iteration of the audience expansion model, the goal is to train the weights of the target features, calculate the probability that a sample is a positive sample through weighted summation of the target features, and minimize the difference between this probability and the actual probability (100%). In this step, after determining the target features, the audience expansion model can be iteratively trained based on seed users, the first target group index corresponding to the seed users' target features, and the target features themselves, to obtain a well-trained audience expansion model. For details on how to obtain the well-trained audience expansion model, please refer to subsequent embodiments; they will not be elaborated upon here.

[0072] The model training method provided in this disclosure involves acquiring seed users corresponding to media resources, determining a saliency index corresponding to the preset media resource features based on the seed users and preset media resource features, identifying preset media resource features whose saliency index is greater than or equal to an index threshold as target features, and iteratively training the audience expansion model based on the seed users, the first target group index corresponding to the target features of the seed users, and the target features to obtain a trained audience expansion model. Because this disclosure determines the target features based on the saliency index corresponding to the preset media resource features determined by the seed users and preset media resource features, it can obtain the target features more accurately. Furthermore, by training based on the target features, seed users, and the first target group index corresponding to the target features of the seed users, a more accurate audience expansion model can be obtained. Consequently, when using the audience expansion model for audience expansion, the expanded audience can be obtained more accurately.

[0073] Figure 3 A flowchart illustrating a model training method provided in another embodiment of this disclosure. Based on the above embodiments, this disclosure further describes the model training method. Figure 3 As shown, the method in this embodiment of the disclosure may include:

[0074] S301. Obtain the seed users corresponding to the media resources.

[0075] For a detailed description of this step, please refer to [link / reference]. Figure 2 The relevant description of S201 in the illustrated embodiment will not be repeated here.

[0076] In this embodiment of the disclosure, Figure 2 Step S202 can further include the following five steps: S302 to S306:

[0077] S302. Obtain all users within a preset time range.

[0078] For example, the preset time range is such as one month, one quarter, or one week, and the total number of users within the preset time range is such as the total number of monthly active users.

[0079] S303. Based on the preset boxing rules, determine the proportion of the first user in each box corresponding to the preset media resource characteristics of all users and the proportion of the second user in each box corresponding to the preset media resource characteristics of seed users.

[0080] For example, after obtaining the full set of monthly active users, the preset binning rules are as follows: based on the full set of monthly active users, preset media resource features are binned according to business scenarios. Alternatively, equal-frequency binning can be selected to binnify the preset media resource features. Theoretically, the user proportion in each bin should not be less than 5%, and the number of bins for each preset media resource feature is usually less than or equal to 20 to ensure the stability of the calculation results. After binning the preset media resource features, the first user proportion (e.g., represented by ratio_all) of each bin corresponding to the preset media resource features can be determined, and the second user proportion (e.g., represented by ratio_sample) of each bin corresponding to the seed users can be determined.

[0081] S304. Based on the proportion of the first user and the proportion of the second user, determine the second target group index corresponding to each sub-box of the preset media resource characteristics, and the first target group index corresponding to the seed user of the preset media resource characteristics.

[0082] For example, after obtaining the first user percentage and the second user percentage, the second target group index (TGI) corresponding to each sub-box of the preset media resource characteristics can be determined according to the following formula:

[0083]

[0084] It is understandable that after determining the second target group index corresponding to each sub-box according to Formula 1 above, since each seed user belongs to the corresponding sub-box, the first target group index corresponding to the preset media resource characteristics of the seed user can be determined according to the second target group index corresponding to each sub-box.

[0085] S305. Determine the aggregated value of the second user percentage corresponding to the second target group index that is greater than the average target group index under the preset media resource characteristics.

[0086] The mean of the target group index is obtained based on the second target group index.

[0087] In this step, the mean target group index can be obtained based on the second target group index corresponding to each sub-box of all preset media resource features. Optionally, the mean target group index is obtained by at least one of the following methods: determining the mean target group index as the average of the second target group index; weighting the second target group index and the second user proportion corresponding to the second target group index to obtain the mean target group index; sorting the second target group index and determining the median of the sorted second target group index as the mean target group index; sorting the number of all users corresponding to the second target group index and determining the second target group index corresponding to the largest number as the mean target group index.

[0088] For example, the average value of the second target group index corresponding to each sub-box of all preset media resource features can be used as the target group index mean; the product of the second target group index and the proportion of the second user corresponding to the second target group index can be summed and the average value can be used as the target group index mean; the second target group index can be sorted from high to low and the median of the sorted second target group index can be determined as the target group index mean; the number of all users corresponding to the second target group index can be sorted and the second target group index corresponding to the largest number can be determined as the target group index mean.

[0089] After determining the mean index of the target group (e.g., represented by avg_TGI), this step can determine the aggregated value of the second user proportion corresponding to the second target group index that is greater than the mean index of the target group under the preset media resource characteristics, according to the following formula two. This aggregated value can also be called the significant proportion:

[0090] ratio_important = ∑(ratio_TGI), if TGI > avg_TGI (Formula 2)

[0091] Here, ratio_important represents the significant proportion; ratio_TGI represents the proportion of the second user corresponding to the second target group index that is greater than the average index of the target group.

[0092] For example, assuming the preset media resource characteristic is the number of days songs have been played in the past 30 days, this preset media resource characteristic corresponds to six bins: 1 day, 2-7 days, 7-14 days, 14-21 days, 21-28 days, and 30 days. The TGI (i.e., the second target group index corresponding to each bin) and significance percentages determined after binning are shown in Table 1.

[0093] Table 1

[0094]

[0095]

[0096] Based on Table 1 above, assuming the average target group index is 150, the total value (i.e. significant percentage) of the second user proportion corresponding to the second target group index that is greater than the average target group index of 150 under the preset media resource characteristics can be determined to be 30%.

[0097] S306. Based on the aggregated value and the second target group index corresponding to each sub-box of the preset media resource characteristics, determine the significance index corresponding to the preset media resource characteristics.

[0098] In this step, after determining the aggregate value, the significance index corresponding to the preset media resource feature can be determined based on the aggregate value and the second target group index corresponding to each sub-box of the preset media resource feature.

[0099] Further, optionally, determining the significance index corresponding to the preset media resource characteristics based on the aggregated value and the second target group index corresponding to each sub-box of the preset media resource characteristics may include: determining the standard deviation of the second target group index under the preset media resource characteristics based on the mean of the second target group index under the preset media resource characteristics and the second target group index; determining the significance index corresponding to the preset media resource characteristics as the product of the aggregated value and the standard deviation.

[0100] For example, the mean of the second target group index under the preset media resource characteristics can be determined based on the second target group index corresponding to different sub-boxes under the preset media resource characteristics (e.g., represented by avg_tgi). Then, the standard deviation of the second target group index under the preset media resource characteristics can be determined according to the following formula three:

[0101]

[0102] Where tgi represents the second target group index corresponding to different sub-boxes under the preset media resource characteristics, and n is the total number of sub-boxes.

[0103] Based on the example in Table 1 above, the mean value of the second target group index under the preset media resource characteristics can be determined to be 107, and then std = 94 can be determined according to Formula 3 above.

[0104] According to Formula 4 below, the significance index (e.g., represented by score_important) corresponding to the preset media resource characteristics can be determined as the product of the summary value and the standard deviation:

[0105] score_importantratio_important×std formula four

[0106] Based on the example in Table 1 above, after determining the summary value to be 30% and std = 94, the significance index can be determined as: score_important = 94 * 0.3 = 28.2.

[0107] S307. Determine the preset media resource features whose significance index is greater than or equal to the index threshold as target features.

[0108] For a detailed description of this step, please refer to [link / reference]. Figure 2 The relevant description of S203 in the illustrated embodiment will not be repeated here.

[0109] In this embodiment of the disclosure, Figure 2Step S204 can further include the following three steps: S308 to S310:

[0110] S308. Normalize the first target group index corresponding to the target features of the seed users to obtain the normalized first target group index.

[0111] The normalization process includes maximum value normalization or normal distribution normalization.

[0112] It is understandable that before constructing a suitable population expansion model, it is necessary to assign values ​​to the target features. In this embodiment of the disclosure, an assignment scheme with low computational cost and close to the positive sample probability (i.e., 100%) is constructed through an unsupervised calculation method. Based on the above formula 1 and the probability formula, the following formula derivation can be obtained as shown in formula 5:

[0113] five

[0114] Where M represents the full sample; M i Indicates a specified sample; M ij Indicates that in the specified sample M i Samples with feature j in M; j represents the feature; j P(Y=1|F) represents a sample in the full sample M that possesses feature j; j ) represents the probability that Y (i.e., the sample label) is equal to 1 given that feature (F) is j; C represents a constant.

[0115] The derivation of Formula 5 above proves that, with the specified sample (i.e., seed user) remaining unchanged, the TGI of a feature can be used as an approximation of the probability that a sample becomes a positive sample under the specified bin of that feature.

[0116] Since the TGI values ​​for different features vary significantly, excessively large values ​​can lead to very small parameter estimates during subsequent iterative training of the population expansion model. Therefore, normalization is necessary. Considering that the probability of a sample being positive is equal to the product of the TGI and a constant, where the constant is the ratio of the total number of samples to the specified number of samples, and this ratio is greater than or equal to 1, it indicates that the sample probability is directly proportional to the TGI index. In this embodiment, the first target group index can be normalized, i.e., globally standardized, to obtain a normalized first target group index. This normalization includes maximum value normalization or normal distribution normalization. The normalized first target group index can be assigned to different bins corresponding to the target features of the seed users, serving as target features for model training and prediction.

[0117] S309. Traverse the target features in descending order of significance index. For each target feature traversed, perform the following operations: Based on the normalized first target group index, the first residual obtained from the previous training of the population expansion model, and the loss function, determine the loss function value, the target weight, and the second residual corresponding to the current training; wherein, the target weight is determined based on the minimum obtained loss function value; in the first training of the population expansion model, the corresponding weight is the preset weight, and the preset initial probability is used as the first residual.

[0118] For example, in this embodiment of the disclosure, the initial probability is preset to 100%; the framework of the population expansion model adopts the boosting algorithm framework to construct a regression tree model based on conditional probability. Here, the linear relationship between feature values ​​and target variables is verified through the transformation of conditional probability; the multicollinearity problem in linear models is avoided by applying the regression tree algorithm; and the sample probability is calculated through the boosting algorithm framework. Based on the example of step S308, the currently assigned target feature value approximates the conditional probability of the sample, where the condition is the corresponding target feature. Therefore, the linear combination of multiple target features ultimately produced by the population expansion model is the probability that the sample is a positive sample. Thus, each target feature is considered a separate sub-model, and each sub-model contains the feature value of the target feature, the target weight corresponding to the target feature, and the sample prediction probability of that sub-model. In the boosting algorithm framework, the optimization objective of each iteration is the residual of the previous iteration, achieving the final prediction probability as close as possible to the true probability. Finally, the final prediction probability of the sample is obtained based on the linear fusion of multiple sub-models. Based on the principle of binary classification models, samples with a predicted probability greater than 0.5 are labeled as 1, and samples with a predicted probability less than or equal to 0.5 are labeled as 0. This allows us to determine the accuracy of the overall population expansion model.

[0119] Since each sample does not have a fixed positive or negative label, in this embodiment, all specified samples (i.e., seed users) are assumed to be positive samples, and the initial probability of each positive sample is 100% (corresponding to an original label of 1). In the initial training of the audience expansion model, the corresponding weights are preset weights (e.g., 1), and the preset initial probability of 100% is used as the first residual. Each iteration uses only one target feature. According to the boosting algorithm framework, the goal of each iteration is the residual of the previous iteration. Therefore, the iteration goal of this embodiment is:

[0120]

[0121] in, f(x i) represents the first target group index corresponding to the target characteristics of seed users; w i This represents the target weight corresponding to the target feature; the value of i ranges from 1 to N, where N represents the number of seed users; y represents the iteration target, y t Let represent the iteration objective obtained in the t-th iteration.

[0122] In this step, the target features can be traversed in descending order of significance index. For the traversed target features, the following operations are performed: Based on the normalized first target group index, the first residual obtained from the previous training of the population expansion model, and the loss function, the loss function value corresponding to the current training, the target weight corresponding to the target feature, and the second residual are determined; wherein, the target weight is determined based on the minimum obtained loss function value.

[0123] Optionally, the loss function value is determined as follows: the mean squared error is determined based on the first target group index and the first residual corresponding to the target feature of the seed user; the loss function value is determined based on the mean squared error, the first regularization, and the second regularization, where the first regularization is used to represent the sum of the magnitudes of the target weights corresponding to the target feature, and the second regularization is used to represent the sum of the squares of the target weights corresponding to the target feature.

[0124] For example, assuming the crowd expansion model uses a boosting algorithm framework and constructs a regression tree model based on conditional probability, the initial loss function can be the mean squared error shown in Formula 6 below:

[0125]

[0126] Where MSE represents the initial loss function; yi represents the first residual obtained by expanding the model from the previous training population.

[0127] Meanwhile, considering that an increase in the number of target features would lead to an increase in the complexity of the population expansion model, in order to prevent overfitting, a first regularization and a second regularization are added for restriction. The final loss function is as follows: Formula 7:

[0128]

[0129] Where Loss represents the final loss function; L1 represents the first regularization; L2 represents the second regularization; α represents the weight of L1; and β represents the weight of L2.

[0130] The loss function value can be determined based on the loss function defined in Formula 7 above.

[0131] S310. When the decrease rate of the loss function value is less than the decrease threshold, obtain the trained population expansion model and the predicted probability value corresponding to the seed user.

[0132] For example, the decrease threshold could be 0.001. The loss function value can be determined based on the loss function shown in Formula 7 above. After several iterations of training, training stops when the decrease ratio of the loss function value is less than the decrease threshold, thus obtaining a trained crowd expansion model. It can be understood that the trained crowd expansion model determines the target weights corresponding to the target features.

[0133] For example, Figure 4 This is a schematic diagram illustrating iterative training of a population expansion model according to an embodiment of this disclosure, as shown below. Figure 4 As shown, seed users are used as the specified samples, with an original label of 1 (i.e., a preset initial probability of 1). The population expansion model is trained in iteration 1, using the preset initial probability as the first residual (i.e., the iteration target), and the preset weight is set to 1, serving as the target weight for target feature 1 in iteration 1. Taking the first specified sample as an example, the first target group index corresponding to target feature 1 for the first specified sample is 0.2. Therefore, the product of the first target group index and the target weight corresponding to target feature 1 for the first specified sample is 0.2, and the second residual corresponding to the first specified sample in iteration 1 (i.e., the initial probability of 1) can be determined. Figure 4 The residual (1) shown is 0.8. When iteration 1 training is complete, the loss function value corresponding to iteration 1 is 2.0. For iteration 2 training of the population expansion model, the iteration target for the first specified sample is the second residual of 0.8 determined in iteration 1. The target weight corresponding to target feature 2 is 0.5, and the first target population index corresponding to target feature 2 for the first specified sample is 0.6. Therefore, the product of the first target population index and the target weight corresponding to target feature 2 for the first specified sample is 0.3, and the second residual corresponding to iteration 2 (i.e., Figure 4 The residual (2) shown is 0.5. When iteration 2 training is complete, the loss function value corresponding to iteration 2 is 1.3. From the loss function value of 2.0 corresponding to iteration 1 and the loss function value of 1.3 corresponding to iteration 2, the decrease ratio of the loss function value can be determined to be 0.35. Similarly, the iteration target, the second residual, and the loss function value corresponding to each iteration for each specified sample in each iteration can be obtained. After T consecutive iterations, when the decrease ratio of the loss function value is determined to be less than the decrease threshold of 0.001, the iteration stops, and the trained population expansion model is obtained. Simultaneously with obtaining the trained population expansion model, the predicted probability value corresponding to the seed user can be obtained.

[0134] Based on the above embodiments, assuming the audience expansion model is based on conditional probability, the final output of the audience expansion model can be approximately equal to the predicted probability value of the positive sample (i.e., seed user). The predicted probability value output by the audience expansion model can be processed as follows:

[0135]

[0136] Where, r i r represents the predicted probability value for each seed user output by the audience expansion model. i =∑f(x) i )×w i That is, the product of the first target group index and the target weight obtained for each seed user in each iteration is summed to obtain the predicted probability value for each seed user.

[0137] Therefore, the overall prediction accuracy of the population expansion model (e.g., expressed as rate) can be determined as:

[0138] rate=∑R i / N*100%

[0139] The higher the accuracy of the prediction, the better the effect of the population expansion model, and the more precise the target population selected.

[0140] The model training method provided in this disclosure is based on preset binning rules. It determines the significance index corresponding to the preset media resource features based on seed users and preset media resource features. The target features are then determined based on the significance index, enabling more accurate acquisition of target features. The trained audience expansion model is then obtained based on the target features, seed users, and the first target group index corresponding to the target features of the seed users. In each iteration of the audience expansion model, the optimization objective is the residual from the previous iteration, maximizing the approximation of the true probability in the final prediction. The audience expansion model has a clear and simple structure; the target features at each layer approximate the sample probability. Therefore, each iteration only needs to record the final prediction probability value, reducing computational complexity and enabling large-scale lookalike calculations. Furthermore, the loss function of the audience expansion model incorporates first and second regularization constraints, significantly reducing its complexity and accelerating the calculation of sample prediction probabilities. Thus, a more accurate and efficient audience expansion model is obtained, enabling more accurate and faster acquisition of the expanded audience when used for audience expansion.

[0141] Based on the model training method provided in this disclosure, taking advertising as an example of media resources, it can output the significance ranking of the features corresponding to the audience expansion model for any advertising audience sample under unsupervised conditions, and can quantify the magnitude of the significance differences between different features. Furthermore, since the number of samples that can usually be collected in advertising campaigns is relatively small, conventional models cannot achieve good training and prediction. However, the model training method provided in this disclosure is also applicable to supervised learning. For supervised situations with insufficient sample size or fluctuations in positive and negative samples over time, the algorithm framework and theory provided in this disclosure can be used to construct the model. Finally, the model can be further validated and optimized using labels.

[0142] Figure 5 This is a flowchart illustrating a method for expanding a user base according to an embodiment of this disclosure. The method described in this embodiment can be applied to a computing device, which may be a server or a server cluster, etc. Figure 5 As shown, the method in this embodiment includes:

[0143] S501, Obtain candidate users corresponding to media resources.

[0144] In this step, the candidate user corresponding to the media resource can be input by a user (such as an advertiser) to the electronic device executing this method embodiment, or it can be sent by other devices to the electronic device executing this method embodiment. It can be understood that the candidate user corresponding to the media resource can be any media resource audience user.

[0145] S502. Based on the candidate users and the seed users and target characteristics corresponding to the audience expansion model, determine the target group index corresponding to the target characteristics of the candidate users.

[0146] The population expansion model is trained using the model training method described in any of the above method embodiments.

[0147] In this step, the audience expansion model is trained using the model training method described in any of the above method embodiments. That is, the seed users and target features corresponding to the audience expansion model have been determined. Therefore, the target group index corresponding to the target features of the candidate users can be determined based on the candidate users and the seed users and target features corresponding to the audience expansion model.

[0148] Further, optionally, the target group index of the target features corresponding to the candidate users and the seed users and target features corresponding to the audience expansion model is determined. This may include: determining the proportion of the third user in each bin of the target features corresponding to the candidate users and the proportion of the fourth user in each bin of the target features corresponding to the seed users based on the preset binning rules; and determining the target group index of the target features corresponding to the candidate users based on the proportion of the third user and the proportion of the fourth user.

[0149] For example, referring to the example of step S303 in the above embodiments, the proportion of the third user in each bin corresponding to the target feature of the candidate user and the proportion of the fourth user in each bin corresponding to the target feature of the seed user can be determined based on the preset binning rules. Referring to Formula 1 above, the target group index corresponding to the target feature of the candidate user can be determined according to the proportion of the third user and the proportion of the fourth user.

[0150] S503. Input the candidate user and the target group index of the target feature corresponding to the candidate user into the population expansion model to obtain the predicted probability value of the candidate user.

[0151] The predicted probability value is used to determine whether a candidate user is a scalable user.

[0152] In this step, after determining the target group index of the target feature corresponding to the candidate user, the candidate user and the target group index of the target feature corresponding to the candidate user can be input into the population expansion model to obtain the predicted probability value of the candidate user.

[0153] Further, optionally, inputting the candidate user and the target group index of the target feature corresponding to the candidate user into the population expansion model to obtain the predicted probability value of the candidate user may include: inputting the candidate user and the target group index of the target feature corresponding to the candidate user into the population expansion model, and summing the product of the target group index of the target feature corresponding to the candidate user and the target weight corresponding to the target feature to obtain the predicted probability value of the candidate user.

[0154] For example, the target group index of the candidate user and the target feature corresponding to the candidate user is input into the population expansion model. The target weight of the target feature corresponding to the population expansion model has been determined through training. Therefore, the product of the target group index of the target feature corresponding to the candidate user and the target weight corresponding to the target feature can be obtained. By summing all the products, the predicted probability value corresponding to the candidate user can be obtained.

[0155] After obtaining the predicted probability values ​​for candidate users, it is possible to determine whether a candidate user is a potential expandable user based on these predicted probability values. For example, candidate users whose predicted probability values ​​are greater than a probability threshold can be identified as potential expandable users.

[0156] The audience expansion method provided in this disclosure acquires candidate users corresponding to media resources, determines the target group index of the target features corresponding to the candidate users based on the candidate users, seed users, and target features corresponding to the audience expansion model, and inputs the candidate users and the target group index of the target features corresponding to the candidate users into the audience expansion model to obtain the predicted probability value of the candidate users. The predicted probability value is used to determine whether the candidate users are expandable users. Since the audience expansion model of this disclosure has higher accuracy and better generalization ability, it can more accurately determine expandable users.

[0157] Based on the above embodiments, taking advertising as an example of media resources, automated display of audience profiles for advertising can be achieved. Specifically, based on feature saliency ranking, the most significant features of a specified audience compared to the general audience can be output in real time for any advertising audience, and secondary drill-down operation of TGI for specific feature bins can be implemented. Since the sample distribution has already been calculated when determining the TGI, while displaying significant features, the distribution map corresponding to the features can be automatically generated, and the feature bins with the largest distribution differences can be marked. Real-time lookalike calculation of the advertising audience can be achieved. Specifically, based on the audience expansion model trained by the model training method provided in this embodiment, the number of expandable audiences can be quickly calculated, and secondary precise screening of the audience can be achieved based on the predicted probability values.

[0158] Exemplary device

[0159] Having introduced the medium of exemplary embodiments of this disclosure, the following references are made to... Figure 6 The model training apparatus according to exemplary embodiments of this disclosure will be described. The apparatus of the exemplary embodiments of this disclosure can implement the various processes in the foregoing model training method embodiments and achieve the same functions and effects.

[0160] Figure 6 This is a schematic diagram of the structure of a model training device provided in an embodiment of the present disclosure, as shown below. Figure 6 As shown, the model training apparatus 600 of this embodiment includes: an acquisition module 601, a first determination module 602, a second determination module 603, and a training module 604. Wherein:

[0161] The acquisition module 601 is used to acquire the seed users corresponding to the media resources.

[0162] The first determining module 602 is used to determine the significance index corresponding to the preset media resource characteristics based on the seed users and preset media resource characteristics.

[0163] The second determining module 603 is used to determine the preset media resource features whose significance index is greater than or equal to the index threshold as target features.

[0164] Training module 604 is used to iteratively train the audience expansion model based on seed users, the first target group index corresponding to the seed users' target features, and target features, to obtain a trained audience expansion model. The audience expansion model is used to expand the audience based on seed users.

[0165] In one possible implementation, the first determining module 602 may be specifically used to: acquire all users within a preset time range; determine, based on preset binning rules, the first user percentage of each bin corresponding to the preset media resource characteristics and the second user percentage of each bin corresponding to the preset media resource characteristics; determine, based on the first user percentage and the second user percentage, the second target group index corresponding to each bin of the preset media resource characteristics and the first target group index corresponding to the preset media resource characteristics; determine, based on the first user percentage and the second user percentage, the second target group index corresponding to each bin of the preset media resource characteristics and the first target group index corresponding to the preset media resource characteristics; determine, based on the sum of the second user percentages corresponding to the second target group index under the preset media resource characteristics that are greater than the average target group index, wherein the average target group index is obtained based on the second target group index; and determine, based on the sum of the sum and the second target group index corresponding to each bin of the preset media resource characteristics, the significance index corresponding to the preset media resource characteristics.

[0166] In one possible implementation, when the first determining module 602 determines the significance index corresponding to the preset media resource feature based on the summary value and the second target group index corresponding to each sub-box of the preset media resource feature, it can specifically be used to: determine the standard deviation of the second target group index under the preset media resource feature based on the mean of the second target group index under the preset media resource feature and the second target group index; and determine that the significance index corresponding to the preset media resource feature is the product of the summary value and the standard deviation.

[0167] In one possible implementation, the first determining module 602 may obtain the target group index mean in at least one of the following ways: determining the target group index mean as the average of the second target group index; weighting the second target group index and the second user proportion corresponding to the second target group index to obtain the target group index mean; sorting the second target group index and determining the median of the sorted second target group index as the target group index mean; sorting the number of all users corresponding to the second target group index and determining the second target group index corresponding to the largest number as the target group index mean.

[0168] In one possible implementation, the training module 604 may be specifically used to: traverse the target features in descending order of significance index, and for the traversed target features, perform the following operations: based on the first target group index corresponding to the target feature of the seed user, the first residual obtained from the previous training of the population expansion model, and the loss function, determine the loss function value corresponding to the current training, the target weight corresponding to the target feature, and the second residual; wherein, the target weight is determined based on the minimum obtained loss function value; in the first training of the population expansion model, the corresponding weight is a preset weight, and the preset initial probability is used as the first residual; when it is determined that the decrease ratio of the loss function value is less than the decrease threshold, obtain the trained population expansion model and the predicted probability value corresponding to the seed user.

[0169] In one possible implementation, the training module 604 can determine the loss function value by: determining the mean squared error based on the first target group index and the first residual corresponding to the target feature of the seed user; and determining the loss function value based on the mean squared error, the first regularization, and the second regularization, wherein the first regularization is used to represent the sum of the magnitudes of the target weights corresponding to the target feature, and the second regularization is used to represent the sum of the squares of the target weights corresponding to the target feature.

[0170] In one possible implementation, the training module 604 may be specifically used to: normalize the first target group index to obtain a normalized first target group index, wherein the normalization process includes maximum value normalization or normal distribution normalization; and iteratively train the expansion model based on the seed users, the normalized first target group index, and the target features to obtain a trained population expansion model.

[0171] In one possible implementation, the second determining module 603, when determining the target feature, may specifically be used to: display preset media resource features in descending order of salience index; and determine the target feature in response to a selection operation targeting the preset media resource features.

[0172] The apparatus of this disclosure embodiment can be used to execute the model training method in any of the above method embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0173] Figure 7 This is a schematic diagram of the structure of a crowd expansion device provided in an embodiment of the present disclosure, as shown below. Figure 7 As shown, the crowd expansion device 700 of this embodiment includes: an acquisition module 701, a determination module 702, and a processing module 703. Wherein:

[0174] The acquisition module 701 is used to acquire candidate users corresponding to media resources.

[0175] The determination module 702 is used to determine the target group index of the target features corresponding to the candidate users based on the candidate users and the seed users and target features corresponding to the audience expansion model. The audience expansion model is trained using the model training method in any of the above method embodiments.

[0176] The processing module 703 is used to input the candidate user and the target group index of the target feature corresponding to the candidate user into the population expansion model to obtain the predicted probability value of the candidate user.

[0177] In one possible implementation, the processing module 703 may be specifically used to: input the candidate user and the target group index of the target feature corresponding to the candidate user into the population expansion model, and sum the product of the target group index of the target feature corresponding to the candidate user and the target weight corresponding to the target feature to obtain the predicted probability value corresponding to the candidate user.

[0178] In one possible implementation, the determining module 702 may be specifically used to: determine the proportion of the third user in each bin corresponding to the target feature of the candidate user and the proportion of the fourth user in each bin corresponding to the target feature of the seed user based on a preset binning rule; and determine the target group index of the target feature corresponding to the candidate user based on the proportion of the third user and the proportion of the fourth user.

[0179] The apparatus of this disclosure embodiment can be used to execute the population expansion method in any of the above method embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0180] Exemplary media

[0181] After introducing the methods of exemplary embodiments of this disclosure, the following references are made. Figure 8 The storage medium of the exemplary embodiments of this disclosure will be described.

[0182] Figure 8 This is a schematic diagram of a storage medium provided according to an embodiment of the present disclosure. (See reference...) Figure 8 As shown, the storage medium 800 stores a program product for implementing the above-described method according to embodiments of the present disclosure. This program product may be a portable compact disc read-only memory (CD-ROM) and includes program code, and can run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto.

[0183] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0184] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying readable program code. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium.

[0185] Program code for performing the operations disclosed herein can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing devices can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN).

[0186] Exemplary computing device

[0187] Having described the methods, media, and apparatus of exemplary embodiments of this disclosure, the following references... Figure 9 A computing device according to an exemplary embodiment of the present disclosure will be described.

[0188] Figure 9 The computing device 900 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments disclosed herein.

[0189] Figure 9 This is a schematic diagram of the structure of a computing device provided in an embodiment of the present disclosure, as shown below. Figure 9As shown, the computing device 900 is presented in the form of a general-purpose computing device. The components of the computing device 900 may include, but are not limited to: at least one processing unit 901, at least one storage unit 902, and a bus 903 connecting different system components (including the processing unit 901 and the storage unit 902). For example, the processing unit 901 may specifically be a processor, and the storage unit 902 stores computer execution instructions. The processing unit 901 executes the computer execution instructions stored in the storage unit 902 to implement the aforementioned model training method and population expansion method.

[0190] The 903 bus includes a data bus, a control bus, and an address bus.

[0191] Storage unit 902 may include readable media in the form of volatile memory, such as random access memory (RAM) 9021 and / or cache memory 9022, and may further include readable media in the form of non-volatile memory, such as read-only memory (ROM) 9023.

[0192] Storage unit 902 may also include a program / utility 9025 having a set (at least one) of program modules 9024, such program modules 9024 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0193] The computing device 900 can also communicate with one or more external devices 904 (e.g., keyboard, pointing device, etc.). This communication can be performed via input / output (I / O) interface 905. Furthermore, the computing device 900 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 906. Figure 9 As shown, network adapter 906 communicates with other modules of computing device 900 via bus 903. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with computing device 900, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0194] It should be noted that although several units / modules or sub-units / modules of the model training device or population expansion device are mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0195] Furthermore, although the operations of the methods disclosed herein are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0196] While the spirit and principles of this disclosure have been described with reference to several specific embodiments, it should be understood that this disclosure is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for convenience of expression. This disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A model training method comprising: obtaining seed users corresponding to a media resource; obtaining all users in a preset time range; determining, based on a preset binning rule, a first user proportion of the all users corresponding to each bin of a preset media resource feature and a second user proportion of the seed users corresponding to each bin of the preset media resource feature; determining, according to the first user proportion and the second user proportion, a second target group index corresponding to each bin of the preset media resource feature and a first target group index of the seed users corresponding to the preset media resource feature; determining a summary value of a second user proportion corresponding to a second target group index greater than a target group index mean value of the preset media resource feature, the target group index mean value being obtained according to the second target group index; determining a standard deviation of the second target group index of the preset media resource feature according to the mean value of the second target group index of the preset media resource feature and the second target group index; determining a significance index corresponding to the preset media resource feature as a product of the summary value and the standard deviation; determining a preset media resource feature with a significance index greater than or equal to an index threshold as a target feature; iteratively training a people expansion model based on the seed users, the first target group index of the seed users corresponding to the target feature, and the target feature to obtain a trained people expansion model, the people expansion model being used to expand people based on the seed users; the iteratively training a people expansion model based on the seed users, the first target group index of the seed users corresponding to the target feature, and the target feature to obtain a trained people expansion model comprises: traversing the target features in order of significance index from high to low, and for a traversed target feature, performing the following operations: determining, based on the first target group index of the seed users corresponding to the target feature, a first residual obtained by a previous training of the people expansion model, and a loss function, a loss function value corresponding to a current training, a target weight corresponding to the target feature, and a second residual; wherein the target weight is determined based on the obtained loss function value being the smallest; in the first training of the people expansion model, the corresponding weight is a preset weight, and a preset initial probability is taken as the first residual; when it is determined that a decline ratio of the loss function value is less than a decline threshold, obtaining the trained people expansion model and a predicted probability value corresponding to the seed users; the loss function value is determined by: determining a mean square error according to the first target group index of the seed users corresponding to the target feature and the first residual; determining the loss function value according to the mean square error, a first regular, and a second regular, the first regular being used to represent a sum of a modulus of the target weight corresponding to the target feature, and the second regular being used to represent a sum of a square of the target weight corresponding to the target feature.

2. The model training method of claim 1, wherein the target group index mean value is obtained according to at least one of the following ways: determining the target group index mean value as an average of the second target group index; weighting the second target group index and a second user proportion corresponding to the second target group index to obtain the target group index mean value; sorting the second target group index to determine a median of the sorted second target group index as the target group index mean value; sorting a number of total users corresponding to the second target group index to determine a second target group index corresponding to a maximum number as the target group index mean value. 3.The model training method of claim 1 or 2, wherein the iterative training of the crowd expansion model based on the seed user, the first target group index of the seed user corresponding to the target feature, and the target feature, to obtain the trained crowd expansion model, comprises: normalizing the first target group index to obtain a normalized first target group index, the normalization comprising maximum value normalization or normal distribution normalization; and iteratively training the expansion model based on the seed user, the normalized first target group index, and the target feature, to obtain the trained crowd expansion model. 4.The model training method of claim 1 or 2, wherein the target feature is determined by: displaying the preset media resource features in descending order of the saliency index; and determining the target feature in response to a selection operation on the preset media resource features. 5.A crowd expansion method, comprising: obtaining a candidate user corresponding to a media resource; determining a target group index of the candidate user corresponding to a target feature according to the candidate user, a seed user, and the target feature corresponding to a crowd expansion model, the crowd expansion model being trained by the model training method of any one of claims 1 to 4; and inputting the candidate user and the target group index of the candidate user corresponding to the target feature into the crowd expansion model to obtain a predicted probability value of the candidate user, the predicted probability value being used to determine whether the candidate user is an expandable user. 6.The crowd expansion method of claim 5, wherein the inputting of the candidate user and the target group index of the candidate user corresponding to the target feature into the crowd expansion model to obtain the predicted probability value of the candidate user comprises: inputting the candidate user and the target group index of the candidate user corresponding to the target feature into the crowd expansion model, and performing sum processing on a product of the target group index of the candidate user corresponding to the target feature and a target weight corresponding to the target feature to obtain the predicted probability value of the candidate user. 7.The crowd expansion method of claim 5 or 6, wherein the determining of the target group index of the candidate user corresponding to the target feature according to the candidate user, the seed user, and the target feature corresponding to the crowd expansion model comprises: ​ ​ ​ ​ ​ ​ ​ ​ determine, based on the third user proportion and the fourth user proportion, a target group index of the candidate user corresponding to the target feature.

8. A model training apparatus, comprising: an acquisition module configured to acquire a seed user corresponding to a media resource; a first determination module configured to determine a saliency index corresponding to a preset media resource feature based on the seed user and the preset media resource feature; a second determination module configured to determine the preset media resource feature with the saliency index greater than or equal to an index threshold as a target feature; a training module configured to iteratively train a crowd expansion model based on the seed user, a first target group index of the seed user corresponding to the target feature, and the target feature, to obtain a trained crowd expansion model, the crowd expansion model being configured to expand a crowd based on the seed user; the first determination module is specifically configured to: acquire all users in a preset time range; determine, based on a preset binning rule, a first user proportion of the all users corresponding to each bin of the preset media resource feature and a second user proportion of the seed user corresponding to each bin of the preset media resource feature; determine, based on the first user proportion and the second user proportion, a second target group index corresponding to each bin of the preset media resource feature and a first target group index of the seed user corresponding to the preset media resource feature; determine a summary value of the second user proportion corresponding to the second target group index greater than a target group index mean value of the preset media resource feature, the target group index mean value being obtained based on the second target group index; determine a standard deviation of the second target group index of the preset media resource feature based on the mean value of the second target group index of the preset media resource feature and the second target group index; determine the saliency index corresponding to the preset media resource feature as a product of the summary value and the standard deviation; the training module is specifically configured to: traverse the target features in descending order of saliency index, and perform the following operations on the traversed target feature: determine a loss function value corresponding to the current training, a target weight corresponding to the target feature, and a second residual error based on the first target group index of the seed user corresponding to the target feature, a first residual error obtained in the last training of the crowd expansion model, and a loss function; wherein the target weight is determined based on the obtained minimum loss function value; in the first training of the crowd expansion model, the corresponding weight is a preset weight, and a preset initial probability is taken as the first residual error; obtain the trained crowd expansion model and a predicted probability value corresponding to the seed user when it is determined that a decline ratio of the loss function value is less than a decline threshold; the training module determines the loss function value by the following way: ​ determine a mean square error according to the first target group index of the seed user corresponding to the target feature and the first residual; determine the loss function value according to the mean square error, a first regularization and a second regularization, the first regularization being used to represent a sum of norms of target weights corresponding to the target feature, and the second regularization being used to represent a sum of squares of the target weights corresponding to the target feature.

9. The model training apparatus of claim 8, wherein the first determination module obtains the target group index mean value according to at least one of the following manners: determining the target group index mean value as an average value of the second target group indexes; performing weighted processing on the second target group indexes and second user proportions corresponding to the second target group indexes to obtain the target group index mean value; sorting the second target group indexes and determining a median of the sorted second target group indexes as the target group index mean value; sorting numbers of total users corresponding to the second target group indexes and determining a second target group index corresponding to a maximum number as the target group index mean value.

10. The model training apparatus of claim 8 or 9, wherein the training module is specifically configured to: perform normalization processing on the first target group indexes to obtain normalized first target group indexes, the normalization processing including maximum value normalization processing or normal distribution normalization processing; perform iterative training on an expansion model based on the seed user, the normalized first target group indexes and the target feature to obtain a trained people expansion model.

11. The model training apparatus of claim 8 or 9, wherein the second determination module is specifically configured to, when determining the target feature: display the preset media resource features in a descending order of the significance indexes; and determine the target feature in response to a selection operation on the preset media resource features.

12. A people expansion apparatus, comprising: an acquisition module configured to acquire candidate users corresponding to media resources; a determination module configured to determine target group indexes of the candidate users corresponding to a target feature according to the candidate users, seed users and the target feature corresponding to a people expansion model, the people expansion model being trained by using the model training method of any one of claims 1 to 4; a processing module configured to input the candidate users and the target group indexes of the candidate users corresponding to the target feature into the people expansion model to obtain predicted probability values of the candidate users.

13. The people expansion apparatus of claim 12, wherein the processing module is specifically configured to: input the candidate users and the target group indexes of the candidate users corresponding to the target feature into the people expansion model, perform sum processing on products of the target group indexes of the candidate users corresponding to the target feature and target weights corresponding to the target feature to obtain the predicted probability values of the candidate users.

14. The people expansion apparatus of claim 12 or 13, wherein the determination module is specifically configured to: determine, based on a preset binning rule, a third user proportion of the candidate user corresponding to each bin of the target feature and a fourth user proportion of the seed user corresponding to each bin of the target feature; determine, according to the third user proportion and the fourth user proportion, a target group index of the candidate user corresponding to the target feature.

15. A computing device comprising: a processor, and a memory connected to the processor in communication; the memory stores computer-executed instructions; the processor executes the computer-executed instructions stored in the memory to implement the method according to any one of claims 1 to 7.

16. A storage medium, the storage medium storing computer program instructions, the computer program instructions being executed to implement the method according to any one of claims 1 to 7.

17. A computer program product, comprising a computer program, the computer program being executed to implement the method according to any one of claims 1 to 7.

Images