Potential user screening method and apparatus

Abstract

The application relates to the technical field of user screening, and provides a potential user screening method and device. The method comprises the following steps: inputting target data of positive samples and target data of unmarked samples into a potential user screening model; if the number of the positive samples is not equal to the number of reliable negative samples, and the number of model training times reaches a preset training number, then a historical weighted average probability is obtained according to historical output probabilities and historical weights corresponding to the remaining unmarked samples; the remaining unmarked samples are sequentially divided into reliable negative samples in ascending order of the historical weighted average probability until the number of the positive samples is equal to the number of the current reliable negative samples; the target data of the positive samples and the target data of the current reliable negative samples are input into the potential user screening model; and potential users meeting a preset bandwidth condition are screened out. The potential user screening method can improve the generalization ability and expression ability of the model, and the accuracy of the screened potential users is high.

Description

Technical Field

[0001] This application relates to the field of user screening technology, specifically to a method and apparatus for screening potential users. Background Technology

[0002] In recent years, with the gradual promotion of data mining technology in the industrial field, machine learning algorithms have also gradually developed, and traditional supervised learning is the mainstream algorithm in the field of data mining. In the field of data mining, operators mostly conduct supervised modeling based on labeled positive samples. However, for modeling problems such as potential user identification, the labeled method will artificially introduce subjective noise, thereby affecting the stability and generalization ability of the model.

[0003] For unlabeled data, semi-supervised learning is now the mainstream approach. Traditional semi-supervised learning selects reliable negative samples from unlabeled samples and then identifies potential users from these reliable negative samples. However, during model training, an imbalance between the number of positive samples and reliable negative samples can occur, leading to underfitting, decreased generalization ability, insufficient model expressive power, and reduced recognition accuracy. Summary of the Invention

[0004] This application provides a potential user screening method and apparatus to solve the technical problem that the imbalance between the number of positive samples and reliable negative samples leads to underfitting of the model, decreased generalization ability, and consequently, insufficient model expressive power and decreased recognition accuracy.

[0005] In a first aspect, embodiments of this application provide a method for screening potential users, including:

[0006] The target data of positive samples and the target data of unlabeled samples are input into the potential user screening model to train the potential user screening model; the target data of positive samples is the behavioral feature data of the first target user related to bandwidth information, and the target data of unlabeled samples is the behavioral feature data of the second target user related to bandwidth information. The first target user is a user who meets the preset bandwidth conditions, and the second target user is a user who does not meet the preset bandwidth conditions. The potential user screening model is any classification model or a combination of multiple classification models.

[0007] Determine whether the number of positive samples is equal to the number of reliable negative samples; the reliable negative samples are the samples obtained after classifying the unlabeled samples as negative samples.

[0008] If the number of positive samples is not equal to the number of reliable negative samples, and the model training times reach the preset number of training times, then the historical weighted average probability corresponding to the remaining unlabeled samples is obtained according to the historical output probability and historical weight corresponding to the remaining unlabeled samples; the remaining unlabeled samples are the samples among the unlabeled samples that have not yet been classified as reliable negative samples.

[0009] The remaining unlabeled samples are divided into reliable negative samples according to their historical weighted average probability from smallest to largest, until the number of positive samples is equal to the number of current reliable negative samples;

[0010] The target data of the positive samples and the target data of the current reliable negative samples are input into the potential user screening model to screen out potential users who meet the preset bandwidth conditions.

[0011] In one embodiment, inputting the target data of the positive samples and the target data of the current reliable negative samples into the potential user screening model to screen out potential users who meet the preset bandwidth conditions includes:

[0012] Input the target data of the positive sample and the target data of the current reliable negative sample into the potential user screening model to obtain the current output probability that the second target user in the current reliable negative sample meets the preset bandwidth condition;

[0013] Based on the current output probability, potential users who meet the preset bandwidth conditions are selected from the second target users within the current reliable negative sample.

[0014] In one embodiment, after determining whether the number of positive samples is equal to the number of reliable negative samples, the process includes:

[0015] If the number of positive samples is not equal to the number of reliable negative samples, and the number of training iterations of the model has not reached the preset number of training iterations, then the negative samples in the remaining unlabeled samples are oversampled to obtain an oversampled negative sample with the same number of positive samples.

[0016] The target data of the positive samples and the target data of the oversampled negative samples are input into the potential user screening model to train the potential user screening model;

[0017] The oversampled negative samples are classified to obtain new reliable negative samples. The process then returns to the step of determining whether the number of positive samples is equal to the number of reliable negative samples, until the number of positive samples is equal to the current number of reliable negative samples, or the number of training iterations of the model reaches the preset number of training iterations.

[0018] In one embodiment, before inputting the target data of positive samples and the target data of unlabeled samples into the potential user screening model, the following steps are included:

[0019] The first data to be processed is processed to remove missing data, resulting in the second data to be processed; the first data to be processed consists of the behavioral feature data of the first target user and the behavioral feature data of the second target user.

[0020] The second data to be processed is normalized to obtain the third data to be processed.

[0021] The third set of data to be processed is filtered to obtain the target data of the positive samples and the target data of the unlabeled samples.

[0022] In one embodiment, the missing data processing for the first data to be processed includes:

[0023] Delete the data corresponding to the duplicate behavior features of the same user in the first data to be processed to obtain non-overlapping behavior feature data;

[0024] If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data exceeds the first missing threshold, then the data corresponding to any behavioral feature will be deleted.

[0025] If the missing data corresponding to any behavior feature in the non-overlapping behavior feature data exceeds the second missing threshold but does not exceed the first missing threshold, then new labeled data is added to mark whether the data corresponding to the behavior feature is missing.

[0026] If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data does not exceed the second missing threshold, and the data corresponding to any behavioral feature is a continuous variable, then the average value of the data corresponding to any behavioral feature is filled into the missing position.

[0027] If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data does not exceed the second missing threshold, and the data corresponding to any behavioral feature is a discrete variable, then the mode of the data corresponding to any behavioral feature is filled into the missing position.

[0028] In one embodiment, the filtering process for the third data to be processed includes:

[0029] The importance value of any behavioral feature in the third data to be processed is obtained by fitting the random forest model to the third data to be processed for the screening of potential users.

[0030] Behavioral features with importance values ​​greater than the importance threshold are identified as important behavioral features;

[0031] Calculate the correlation coefficient between any two important behavioral features, and randomly select one of the two important behavioral features whose correlation coefficient is greater than the coefficient threshold as the selected important behavioral feature;

[0032] The data corresponding to the key behavioral characteristics selected are identified as the selected data.

[0033] In one embodiment, the weights corresponding to the remaining unlabeled samples input into the potential user screening model obtained from the nth model training are obtained by the ratio of the screening accuracy of the potential user screening model obtained from the nth model training to the sum of the screening accuracies of the potential user screening models obtained from the Nth model training; where N is the preset number of training times and n is a positive integer less than or equal to N.

[0034] Secondly, embodiments of this application provide a potential user screening device, comprising:

[0035] The model training module is used to: input the target data of positive samples and the target data of unlabeled samples into the potential user screening model to train the potential user screening model; the target data of the positive samples is the behavioral feature data of the first target user related to bandwidth information, the target data of the unlabeled samples is the behavioral feature data of the second target user related to bandwidth information, the first target user is a user who meets the preset bandwidth conditions, the second user is a user who does not meet the preset bandwidth conditions, and the potential user screening model is any classification model or a combination of multiple classification models;

[0036] The sample quantity determination module is used to: determine whether the number of positive samples is equal to the number of reliable negative samples; the reliable negative samples are the samples obtained after classifying the unlabeled samples as negative samples;

[0037] The sample probability calculation module is used to: if the number of positive samples is not equal to the number of reliable negative samples, and the number of model training times reaches a preset number of training times, then obtain the historical weighted average probability corresponding to the remaining unlabeled samples based on the historical output probability and historical weight corresponding to the remaining unlabeled samples; the remaining unlabeled samples are the samples among the unlabeled samples that have not yet been classified as reliable negative samples;

[0038] The sample quantity balancing module is used to: divide the remaining unlabeled samples into reliable negative samples in ascending order of historical weighted average probability, until the number of positive samples is equal to the number of current reliable negative samples;

[0039] The potential user screening module is used to input the target data of the positive samples and the target data of the current reliable negative samples into the potential user screening model, and screen out potential users that meet the preset bandwidth conditions.

[0040] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the program to implement the steps of the potential user screening method described in the first aspect.

[0041] Fourthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the potential user screening method described in the first aspect.

[0042] The potential user screening method and apparatus provided in this application input the target data of positive samples and the target data of unlabeled samples into a potential user screening model to train the potential user screening model. It determines whether the number of positive samples is equal to the number of reliable negative samples. If the number of positive samples is not equal to the number of reliable negative samples, and the model training times reach the preset number of training times, it obtains the historical weighted average probability of the remaining unlabeled samples according to the historical output probability and historical weight of the remaining unlabeled samples. The remaining unlabeled samples are divided into reliable negative samples in ascending order of historical weighted average probability until the number of positive samples is equal to the number of current reliable negative samples. The target data of positive samples and the target data of current reliable negative samples are input into the potential user screening model to screen out potential users who meet the preset bandwidth conditions. In cases where the number of positive samples and reliable negative samples is unbalanced and the model has been trained to the preset number of training iterations, this application uses historical weighted average probability to select reliable negative samples from the remaining unlabeled samples, thereby balancing the number of positive samples and reliable negative samples. By inputting balanced positive samples and reliable negative samples into the model, overfitting is avoided, the model's generalization ability and stability are improved, and the model's expressive power is enhanced, resulting in high accuracy of the selected potential users. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is one of the flowcharts illustrating the potential user screening method provided in the embodiments of this application;

[0045] Figure 2 This is a second schematic flowchart of the potential user screening method provided in the embodiments of this application;

[0046] Figure 3 This is the third flowchart illustrating the potential user screening method provided in the embodiments of this application;

[0047] Figure 4 This is the fourth flowchart illustrating the potential user screening method provided in the embodiments of this application;

[0048] Figure 5 This is the fifth flowchart illustrating the potential user screening method provided in the embodiments of this application;

[0049] Figure 6 This is a schematic diagram of the potential user screening device provided in the embodiments of this application;

[0050] Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0052] Figure 1 This is one of the flowcharts illustrating the potential user screening method provided in this application embodiment. (Refer to...) Figure 1 This application provides a method for screening potential users, which may include:

[0053] 101. Input the target data of positive samples and the target data of unlabeled samples into the potential user screening model to train the potential user screening model;

[0054] The target data for positive samples is the behavioral feature data of the first target user related to bandwidth information, and the target data for unlabeled samples is the behavioral feature data of the second target user related to bandwidth information. The first target user is the user who meets the preset bandwidth conditions, and the second target user is the user who does not meet the preset bandwidth conditions. The potential user screening model is any classification model or a combination of multiple classification models.

[0055] 102. Determine whether the number of positive samples is equal to the number of reliable negative samples;

[0056] A reliable negative sample is a sample obtained by classifying unlabeled samples as negative samples.

[0057] 103. If the number of positive samples is not equal to the number of reliable negative samples, and the number of training times of the model reaches the preset number of training times, then the historical weighted average probability of the remaining unlabeled samples is obtained based on the historical output probability and historical weight of the remaining unlabeled samples.

[0058] The remaining unlabeled samples are those unlabeled samples that have not yet been classified as reliable negative samples;

[0059] 104. Divide the remaining unlabeled samples into reliable negative samples in ascending order of their historical weighted average probability, until the number of positive samples is equal to the number of current reliable negative samples;

[0060] 105. Input the target data of the positive samples and the target data of the current reliable negative samples into the potential user screening model to screen out potential users that meet the preset bandwidth conditions.

[0061] In step 101, the preset bandwidth condition can be set according to actual needs, and there is no limitation here. In this embodiment, the preset bandwidth condition can be set to "has subscribed to home broadband service and bandwidth ≥ 500M", then the first target user can be defined as a gigabit user, and the first target user is used as a positive sample.

[0062] Users who do not meet the preset bandwidth requirement may be unable to reach it because their place of residence does not offer broadband services of 500 Mbps or higher, or they may not have been aware of the availability of such services and therefore did not apply for them. In these cases, it is not that the user subjectively does not want to meet the preset bandwidth requirement, and therefore these are spurious negative samples. Since it is difficult to ascertain the user's subjective intentions, obtaining truly negative samples is extremely challenging. Therefore, not meeting the preset bandwidth requirement can be defined as "having subscribed to home broadband service with a bandwidth ≥ 500 Mbps," and these second target users who do not meet the preset bandwidth requirement can be treated as unlabeled samples to avoid inaccurate labeling caused by pre-labeling negative samples.

[0063] The statistical definitions for positive and unlabeled samples are shown in the table below:

[0064] Table 1. Statistical Scope of the Sample

[0065] Serial Number Sample categories rule 1 Positive samples The first target users are those who have subscribed to home broadband service and have a bandwidth of ≥500M. 2 unlabeled samples Second target users: those who have subscribed to home broadband service and whose bandwidth is <500M

[0066] Considering that potential users of gigabit bandwidth generally have the following user characteristics:

[0067] 1. High data usage;

[0068] 2. Frequent data usage exceeding limits;

[0069] 3. Home gateways have many connected terminals;

[0070] 4. Games and video applications are used frequently;

[0071] 5. Others.

[0072] Therefore, relevant behavioral features can be selected based on the above user characteristics, and behavioral feature data of the first and second target users over the past month can be collected. The specific collected behavioral features are shown in the table below:

[0073] Table 2 User Behavior Characteristics

[0074] Serial Number Behavioral characteristics Serial Number Behavioral characteristics 1 User ID 17 Days spent using short video apps 2 Gateway SN 18 Number of times using short video apps 3 bandwidth 19 Duration of time spent using video apps 4 Broadband expiration time 20 Number of times lag occurred when using gaming apps 5 LAN port maximum negotiation speed 21 Lag duration when using gaming apps 6 Maximum negotiated receive rate of WLAN port 22 Number of times using game apps 7 Maximum negotiated transmit rate of WLAN port 23 Duration of time using gaming apps 8 Is the gateway dual-band? 24 Number of times buffering occurred when using live streaming apps 9 WIFI version 25 Lag duration when using live streaming apps 10 Number of downstream devices 26 Number of days using live streaming apps 11 Number of times buffering occurs when using video apps 27 Number of times using live streaming apps 12 Days of using video apps 28 Duration of use of live streaming apps 13 Number of times video apps are used 29 WAN port downlink rate saturation 14 Duration of time spent using video apps 30 WAN port uplink rate saturation 15 Number of times buffering occurs when using short video apps 31 LAN port downlink rate saturation 16 Lag duration when using short video apps 32 LAN port uplink rate saturation

[0075] Potential user screening models can be a combination of one or more classification models such as logistic regression, decision tree, random forest, gradient boosting decision tree, and deep learning.

[0076] In step 102, unlabeled samples can be used as negative samples. A probability threshold is set to classify unlabeled samples as negative samples as reliable negative samples. Based on this probability threshold, unlabeled samples as negative samples are classified to obtain reliable negative samples.

[0077] Since the sample data is input into the potential user screening model, the output probability of the sample meeting the preset bandwidth condition is obtained. The probability threshold is set according to the output probability. Therefore, the lower the probability threshold, the smaller the probability that the unlabeled sample that is a negative sample is actually a positive sample, that is, the greater the probability that it is actually a true negative sample. In this embodiment, the probability threshold can be set to 0. Based on the probability threshold, the unlabeled sample that is an initial output probability of 0 after being input into the potential user screening model is marked as a reliable negative sample. Since the initial output probability is 0, theoretically, the unlabeled sample that is a negative sample cannot be a positive sample. Therefore, classifying it as a reliable negative sample can be infinitely close to a true negative sample.

[0078] At this point, the sample types include a small number of positive samples, reliable negative samples, and a large number of unlabeled samples containing the initial output probability.

[0079] It should be noted that if one sample corresponds to the target data of one user, then the number of positive samples is the number of the first target users, the number of unlabeled samples is the number of the second target users, and the number of reliable negative samples is the corresponding number of the second target users.

[0080] In step 103, since the potential user screening model is initially trained using positive samples and unlabeled samples (which serve as negative samples), the number of positive and negative samples is imbalanced. This generally results in low accuracy and generalization ability of the trained potential user screening model. Therefore, it is necessary to further train the potential user screening model using reliable negative and positive samples obtained from the split to improve the model's generalization ability. Meanwhile, the probability threshold set in step 102 is quite strict to obtain reliable negative samples. If the number of reliable negative samples is still less than the number of positive samples when the model has been trained a preset number of times, the problem of sample imbalance persists, thus affecting the model's generalization ability.

[0081] It should be noted that if, after the judgment in step 102, the number of positive samples is equal to the number of reliable negative samples, then you can directly jump to step 105.

[0082] The potential user screening method provided in this embodiment inputs the target data of positive samples and the target data of unlabeled samples into the potential user screening model to train the potential user screening model. It determines whether the number of positive samples is equal to the number of reliable negative samples. If the number of positive samples is not equal to the number of reliable negative samples, and the model training times reach the preset number of training times, the historical weighted average probability of the remaining unlabeled samples is obtained according to the historical output probability and historical weight of the remaining unlabeled samples. The remaining unlabeled samples are divided into reliable negative samples in ascending order of historical weighted average probability until the number of positive samples is equal to the number of current reliable negative samples. The target data of positive samples and the target data of current reliable negative samples are input into the potential user screening model to screen out potential users who meet the preset bandwidth conditions. In cases where the number of positive samples and reliable negative samples is unbalanced and the model has been trained to the preset number of training iterations, this application uses historical weighted average probability to select reliable negative samples from the remaining unlabeled samples, thereby balancing the number of positive samples and reliable negative samples. By inputting balanced positive samples and reliable negative samples into the model, overfitting is avoided, the model's generalization ability and stability are improved, and the model's expressive power is enhanced, resulting in high accuracy of the selected potential users.

[0083] In one embodiment, inputting the target data of positive samples and the target data of current reliable negative samples into a potential user screening model to screen potential users who meet preset bandwidth conditions may include:

[0084] Input the target data of the positive sample and the target data of the current reliable negative sample into the potential user screening model to obtain the current output probability that the second target user in the current reliable negative sample meets the preset bandwidth condition. Then, based on the current output probability, screen the potential users who meet the preset bandwidth condition from the second target users in the current reliable negative sample.

[0085] A filtering threshold can be set for the output probability. When the current output probability is greater than the filtering threshold, the second target user in the current reliable negative sample will be filtered out as a potential user that meets the preset bandwidth condition.

[0086] The final model output structure can be represented by the following table:

[0087] Table 3 Model Output Results

[0088] Serial Number user Are you a potential gigabit bandwidth user? The probability of being a potential gigabit bandwidth user 1 135******** 0 or 1 <![CDATA[p i ∈[0,1]]]>

[0089] As shown in the table above, user i is identified by their mobile phone number. 0 indicates that user i is not a potential gigabit bandwidth user, and 1 indicates that user i is a potential gigabit bandwidth user. i This represents the probability that user i is a potential gigabit bandwidth user.

[0090] This embodiment inputs balanced positive sample data and current reliable negative sample data into the potential user screening model, and filters potential users based on the current output probability. Under the condition of balanced samples, it can improve the generalization ability and expressive ability of the model, resulting in a high accuracy of the screened potential users.

[0091] Figure 2 This is the second flowchart illustrating the potential user screening method provided in this application embodiment. (Refer to...) Figure 2 In one embodiment, after determining whether the number of positive samples is equal to the number of reliable negative samples, the process may include:

[0092] 201. If the number of positive samples is not equal to the number of reliable negative samples, and the number of training iterations of the model has not reached the preset number of training iterations, then oversample the negative samples in the remaining unlabeled samples to obtain an oversampled negative sample with the same number of positive samples.

[0093] 202. Input the target data of positive samples and the target data of oversampled negative samples into the potential user screening model, and train the potential user screening model.

[0094] 203. Classify the oversampled negative samples to obtain new reliable negative samples. Return to the step of determining whether the number of positive samples is equal to the number of reliable negative samples, until the number of positive samples is equal to the current number of reliable negative samples, or the number of training iterations of the model reaches the preset number of training iterations.

[0095] In step 201, any oversampling method can be used to oversample the negative samples in the remaining unlabeled samples. There is no limitation here. In this embodiment, synthetic minority class oversampling technique can be used for oversampling.

[0096] In step 202, the model is trained using balanced positive samples and oversampled negative samples, which can improve the problem of sample imbalance and enhance the model's generalization and expressive abilities.

[0097] In step 203, since it cannot be guaranteed that the oversampled negative samples are reliable negative samples, the improvement effect of step 202 on the model is limited. It is still necessary to classify the oversampled negative samples to obtain new reliable negative samples, and add the new reliable negative samples to the existing reliable negative samples to obtain the current reliable negative samples.

[0098] The method for classifying oversampled negative samples is the same as the method for classifying unlabeled samples that are negative samples, and will not be described again here.

[0099] In this embodiment, when the number of positive samples and reliable negative samples is unbalanced and the model training times have not reached the preset number of training times, negative samples with the same number as positive samples are obtained through oversampling. The model is then trained using an equal number of positive and negative samples to improve its generalization and expressive abilities. The negative samples are then classified to obtain new reliable negative samples, which can further increase the number of reliable negative samples. Since reliable negative samples have a better model improvement effect than negative samples, this helps to balance the number of positive samples and reliable negative samples, thereby enabling the model to be trained again in the future to further improve its generalization and expressive abilities.

[0100] Figure 3 This is the third flowchart illustrating the potential user screening method provided in this application's embodiments. (Refer to...) Figure 3 In one embodiment, before inputting the target data of positive samples and the target data of unlabeled samples into the potential user screening model, the following may be included:

[0101] 301. Perform missing data processing on the first data to be processed to obtain the second data to be processed;

[0102] The first set of data to be processed consists of behavioral characteristic data of the first target user and behavioral characteristic data of the second target user.

[0103] 302. Normalize the second set of data to be processed to obtain the third set of data to be processed;

[0104] 303. The third set of data to be processed is filtered to obtain the target data of positive samples and the target data of unlabeled samples.

[0105] In step 302, any normalization method can be used to normalize the second data to be processed; no limitation is made here. In this embodiment, a linear normalization method can be used to normalize the second data to be processed. The specific normalization formula is as follows:

[0106]

[0107] Where, x * Let x be the normalized third data to be processed from any second data to be processed. min x is the minimum value of the behavioral feature data that has the same behavioral feature as x in the second set of data to be processed. max It is the maximum value of the behavioral feature data in the second set of data to be processed that has the same behavioral feature as x.

[0108] In step 303, the filtered behavioral feature data includes the behavioral feature data of the first target user and the behavioral feature data of the second target. Therefore, the filtered behavioral feature data of the first target user is the target data of the positive sample, and the filtered behavioral feature data of the second target user is the target data of the unlabeled sample.

[0109] This embodiment performs missing data processing, normalization processing, and filtering on the behavioral feature data of the first target user and the behavioral feature data of the second target user, which enables the obtained target data to have better feature representation capabilities, thus facilitating the classification and filtering of subsequent models.

[0110] Figure 4 This is the fourth flowchart illustrating the potential user screening method provided in this application's embodiments. (Refer to...) Figure 4 In one embodiment, handling missing data in the first data to be processed may include:

[0111] 401. Delete the data corresponding to the duplicate behavioral features of the same user in the first data to be processed to obtain non-overlapping behavioral feature data;

[0112] 402. If the number of missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data exceeds the first missing threshold, then the data corresponding to that behavioral feature shall be deleted.

[0113] 403. If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data exceeds the second missing threshold but does not exceed the first missing threshold, then add new labeled data to mark whether the data corresponding to the behavioral feature is missing.

[0114] 404. If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data does not exceed the second missing threshold, and the data corresponding to the behavioral feature is a continuous variable, then the average value of the data corresponding to the behavioral feature is used to fill the missing parts.

[0115] 405. If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data does not exceed the second missing threshold, and the data corresponding to the behavioral feature is a discrete variable, then the mode of the data corresponding to the behavioral feature is filled into the missing position.

[0116] The first missing threshold and the second missing threshold can be set according to the actual situation, and are not limited here. In this embodiment, the first missing threshold can be set to 90% and the second missing threshold can be set to 30%.

[0117] In step 402, assuming that in the non-overlapping behavioral feature data, more than 90 of the 100 data corresponding to behavioral feature A are missing, that is, in the data corresponding to behavioral feature A of 100 users, more than 90 users have missing data corresponding to behavioral feature A, then behavioral feature A is deleted, that is, the 10 data that actually exist for behavioral feature A are also deleted.

[0118] In step 403, assuming that among the non-overlapping behavioral feature data, 31-90 of the 100 data corresponding to behavioral feature A are missing, that is, among the data corresponding to behavioral feature A of 100 users, 31-90 of the data corresponding to behavioral feature A are missing, then new labeled data is added to mark whether all data corresponding to behavioral feature A are missing.

[0119] In step 404, assuming that in the non-overlapping behavioral feature data, no more than 30 of the 100 data corresponding to behavioral feature A are missing, that is, in the data corresponding to behavioral feature A of 100 users, no more than 30 users have missing data corresponding to behavioral feature A, and the non-missing data corresponding to behavioral feature A is a continuous variable, then the average value of all non-missing data is used to fill the missing parts.

[0120] In step 405, assuming that in the non-overlapping behavioral feature data, no more than 30 of the 100 data corresponding to behavioral feature A are missing, that is, in the data corresponding to behavioral feature A of 100 users, no more than 30 users have missing data corresponding to behavioral feature A, and the non-missing data corresponding to behavioral feature A are discrete variables, then the mode of all non-missing data is filled into the missing positions.

[0121] It should be noted that for any behavioral feature, if the data is an unreasonable outlier, then the data will be considered missing.

[0122] In practical applications, there is no strict timing relationship between steps 402 and 405; that is, they can be executed simultaneously, or any one step can be executed first, depending on the actual needs, and no restrictions are made here.

[0123] This embodiment sets a missing threshold and performs targeted missing processing on the data corresponding to behavioral features that meet different missing threshold conditions, so that the final behavioral feature data can maintain a reasonable quantity and high quality to the greatest extent.

[0124] Figure 5This is the fifth flowchart illustrating the potential user screening method provided in this application's embodiments. (Refer to...) Figure 5 In one embodiment, filtering the third data to be processed may include:

[0125] 501. Fit the third set of data to be processed using a random forest model to obtain the importance value of any behavioral feature in the third set of data to be processed for the screening of potential users.

[0126] 502. Behavioral features with importance values ​​greater than the importance threshold are identified as important behavioral features;

[0127] 503. Calculate the correlation coefficient between any two important behavioral features, and randomly select one of the two important behavioral features whose correlation coefficient is greater than the coefficient threshold as the selected important behavioral feature;

[0128] 504. The data corresponding to the key behavioral characteristics selected are identified as the selected data.

[0129] In step 503, any correlation coefficient calculation method can be used to calculate the correlation coefficient between two important behavioral features. There is no limitation here. In this embodiment, the Spearman correlation coefficient calculation method can be used to calculate the correlation coefficient between two important behavioral features.

[0130] The threshold value of this coefficient can be set according to the actual situation, and there is no limitation here. In this embodiment, the threshold value of this coefficient can be set to 0.8.

[0131] This embodiment obtains the importance value of any behavioral feature for potential user screening through a random forest model, then filters out important behavioral features based on a threshold, selects important behavioral features with high correlation coefficients, and finally filters out the data corresponding to the important behavioral features. This data is not only highly important for potential user screening but also highly correlated with other data, which can improve the efficiency and accuracy of subsequent model training, sample screening, and potential user screening.

[0132] In one embodiment, the historical weights can be obtained as follows:

[0133] The weights corresponding to the remaining unlabeled samples input into the potential user screening model obtained from the nth model training are obtained by the ratio of the screening accuracy of the potential user screening model obtained from the nth model training to the sum of the screening accuracies of the potential user screening models obtained from the Nth model training.

[0134] Where N is the preset number of training iterations, and n is a positive integer less than or equal to N.

[0135] This embodiment uses the proportion of the screening accuracy of the potential user screening model after historical training as a weight, which can take into account the screening accuracy of the model when calculating the historical weighted average probability of the remaining unlabeled samples. Based on this, the reliable negative samples are also more accurate.

[0136] The potential user screening device provided in the embodiments of this application is described below. The potential user screening device described below can be referred to in correspondence with the potential user screening method described above.

[0137] Figure 6 A schematic diagram of the potential user screening device provided in an embodiment of this application. (Refer to...) Figure 6 This application provides a potential user screening device, which may include:

[0138] The model training module 601 is used to: input the target data of positive samples and the target data of unlabeled samples into the potential user screening model to train the potential user screening model; the target data of the positive samples is the behavioral feature data of the first target user related to bandwidth information, the target data of the unlabeled samples is the behavioral feature data of the second target user related to bandwidth information, the first target user is a user who meets the preset bandwidth conditions, the second user is a user who does not meet the preset bandwidth conditions, and the potential user screening model is any classification model or a combination of multiple classification models;

[0139] The sample quantity determination module 602 is used to: determine whether the number of positive samples is equal to the number of reliable negative samples; the reliable negative samples are the samples obtained after classifying the unlabeled samples as negative samples;

[0140] The sample probability calculation module 603 is used to: if the number of positive samples is not equal to the number of reliable negative samples, and the number of model training times reaches a preset number of training times, then obtain the historical weighted average probability corresponding to the remaining unlabeled samples based on the historical output probability and historical weight corresponding to the remaining unlabeled samples; the remaining unlabeled samples are the samples among the unlabeled samples that have not yet been classified as reliable negative samples;

[0141] The sample quantity balancing module 604 is used to: divide the remaining unlabeled samples into reliable negative samples in ascending order of historical weighted average probability, until the number of positive samples is equal to the number of current reliable negative samples;

[0142] The potential user screening module 605 is used to: input the target data of the positive sample and the target data of the current reliable negative sample into the potential user screening model, and screen out potential users that meet the preset bandwidth conditions.

[0143] The potential user screening device provided in this embodiment inputs the target data of positive samples and the target data of unlabeled samples into the potential user screening model to train the potential user screening model. It determines whether the number of positive samples is equal to the number of reliable negative samples. If the number of positive samples is not equal to the number of reliable negative samples, and the model training times reach the preset number of training times, it obtains the historical weighted average probability of the remaining unlabeled samples according to the historical output probability and historical weight of the remaining unlabeled samples. The remaining unlabeled samples are divided into reliable negative samples in ascending order of historical weighted average probability until the number of positive samples is equal to the number of current reliable negative samples. The target data of positive samples and the target data of current reliable negative samples are input into the potential user screening model to screen out potential users who meet the preset bandwidth conditions. In cases where the number of positive samples and reliable negative samples is unbalanced and the model has been trained to the preset number of training iterations, this application uses historical weighted average probability to select reliable negative samples from the remaining unlabeled samples, thereby balancing the number of positive samples and reliable negative samples. By inputting balanced positive samples and reliable negative samples into the model, overfitting is avoided, the model's generalization ability and stability are improved, and the model's expressive power is enhanced, resulting in high accuracy of the selected potential users.

[0144] In one embodiment, the potential user screening module 605 is specifically used for:

[0145] Input the target data of the positive sample and the target data of the current reliable negative sample into the potential user screening model to obtain the current output probability that the second target user in the current reliable negative sample meets the preset bandwidth condition;

[0146] Based on the current output probability, potential users who meet the preset bandwidth conditions are selected from the second target users within the current reliable negative sample.

[0147] In one embodiment, an oversampling module (not shown) is further included, for:

[0148] If the number of positive samples is not equal to the number of reliable negative samples, and the number of training iterations of the model has not reached the preset number of training iterations, then the negative samples in the remaining unlabeled samples are oversampled to obtain an oversampled negative sample with the same number of positive samples.

[0149] The target data of the positive samples and the target data of the oversampled negative samples are input into the potential user screening model to train the potential user screening model;

[0150] The oversampled negative samples are classified to obtain new reliable negative samples. The process then returns to the step of determining whether the number of positive samples is equal to the number of reliable negative samples, until the number of positive samples is equal to the current number of reliable negative samples, or the number of training iterations of the model reaches the preset number of training iterations.

[0151] In one embodiment, a target data acquisition module (not shown in the figure) is further included, for:

[0152] The first data to be processed is processed to remove missing data, resulting in the second data to be processed; the first data to be processed consists of the behavioral feature data of the first target user and the behavioral feature data of the second target user.

[0153] The second data to be processed is normalized to obtain the third data to be processed.

[0154] The third set of data to be processed is filtered to obtain the target data of the positive samples and the target data of the unlabeled samples.

[0155] In one embodiment, the target data acquisition module is specifically used for:

[0156] Delete the data corresponding to the duplicate behavior features of the same user in the first data to be processed to obtain non-overlapping behavior feature data;

[0157] If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data exceeds the first missing threshold, then the data corresponding to any behavioral feature will be deleted.

[0158] If the missing data corresponding to any behavior feature in the non-overlapping behavior feature data exceeds the second missing threshold but does not exceed the first missing threshold, then new labeled data is added to mark whether the data corresponding to the behavior feature is missing.

[0159] If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data does not exceed the second missing threshold, and the data corresponding to any behavioral feature is a continuous variable, then the average value of the data corresponding to any behavioral feature is filled into the missing position.

[0160] If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data does not exceed the second missing threshold, and the data corresponding to any behavioral feature is a discrete variable, then the mode of the data corresponding to any behavioral feature is filled into the missing position.

[0161] In one embodiment, the target data acquisition module is specifically used for:

[0162] The importance value of any behavioral feature in the third data to be processed is obtained by fitting the random forest model to the third data to be processed for the screening of potential users.

[0163] Behavioral features with importance values ​​greater than the importance threshold are identified as important behavioral features;

[0164] Calculate the correlation coefficient between any two important behavioral features, and randomly select one of the two important behavioral features whose correlation coefficient is greater than the coefficient threshold as the selected important behavioral feature;

[0165] The data corresponding to the key behavioral characteristics selected are identified as the selected data.

[0166] In one embodiment, a weight acquisition module (not shown in the figure) is further included, for:

[0167] The weights corresponding to the remaining unlabeled samples input into the potential user screening model obtained from the nth model training are obtained by the ratio of the screening accuracy of the potential user screening model obtained from the nth model training to the sum of the screening accuracies of the potential user screening models obtained from the Nth model training; where N is the preset number of training times and n is a positive integer less than or equal to N.

[0168] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other via the communication bus 740. The processor 710 can call a computer program in the memory 730 to execute steps of a potential user screening method, such as:

[0169] The target data of positive samples and the target data of unlabeled samples are input into the potential user screening model to train the potential user screening model; the target data of positive samples is the behavioral feature data of the first target user related to bandwidth information, and the target data of unlabeled samples is the behavioral feature data of the second target user related to bandwidth information. The first target user is a user who meets the preset bandwidth conditions, and the second target user is a user who does not meet the preset bandwidth conditions. The potential user screening model is any classification model or a combination of multiple classification models.

[0170] Determine whether the number of positive samples is equal to the number of reliable negative samples; the reliable negative samples are the samples obtained after classifying the unlabeled samples as negative samples.

[0171] If the number of positive samples is not equal to the number of reliable negative samples, and the model training times reach the preset number of training times, then the historical weighted average probability corresponding to the remaining unlabeled samples is obtained according to the historical output probability and historical weight corresponding to the remaining unlabeled samples; the remaining unlabeled samples are the samples among the unlabeled samples that have not yet been classified as reliable negative samples.

[0172] The remaining unlabeled samples are divided into reliable negative samples according to their historical weighted average probability from smallest to largest, until the number of positive samples is equal to the number of current reliable negative samples;

[0173] The target data of the positive samples and the target data of the current reliable negative samples are input into the potential user screening model to screen out potential users who meet the preset bandwidth conditions.

[0174] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0175] On the other hand, embodiments of this application also provide a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the steps of the potential user screening methods provided in the above embodiments, such as including:

[0176] The target data of positive samples and the target data of unlabeled samples are input into the potential user screening model to train the potential user screening model; the target data of positive samples is the behavioral feature data of the first target user related to bandwidth information, and the target data of unlabeled samples is the behavioral feature data of the second target user related to bandwidth information. The first target user is a user who meets the preset bandwidth conditions, and the second target user is a user who does not meet the preset bandwidth conditions. The potential user screening model is any classification model or a combination of multiple classification models.

[0177] Determine whether the number of positive samples is equal to the number of reliable negative samples; the reliable negative samples are the samples obtained after classifying the unlabeled samples as negative samples.

[0178] If the number of positive samples is not equal to the number of reliable negative samples, and the model training times reach the preset number of training times, then the historical weighted average probability corresponding to the remaining unlabeled samples is obtained according to the historical output probability and historical weight corresponding to the remaining unlabeled samples; the remaining unlabeled samples are the samples among the unlabeled samples that have not yet been classified as reliable negative samples.

[0179] The remaining unlabeled samples are divided into reliable negative samples according to their historical weighted average probability from smallest to largest, until the number of positive samples is equal to the number of current reliable negative samples;

[0180] The target data of the positive samples and the target data of the current reliable negative samples are input into the potential user screening model to screen out potential users who meet the preset bandwidth conditions.

[0181] On the other hand, embodiments of this application also provide a processor-readable storage medium storing a computer program for causing a processor to perform the steps of the methods provided in the above embodiments, such as including:

[0182] The target data of positive samples and the target data of unlabeled samples are input into the potential user screening model to train the potential user screening model; the target data of positive samples is the behavioral feature data of the first target user related to bandwidth information, and the target data of unlabeled samples is the behavioral feature data of the second target user related to bandwidth information. The first target user is a user who meets the preset bandwidth conditions, and the second target user is a user who does not meet the preset bandwidth conditions. The potential user screening model is any classification model or a combination of multiple classification models.

[0183] Determine whether the number of positive samples is equal to the number of reliable negative samples; the reliable negative samples are the samples obtained after classifying the unlabeled samples as negative samples.

[0184] If the number of positive samples is not equal to the number of reliable negative samples, and the model training times reach the preset number of training times, then the historical weighted average probability corresponding to the remaining unlabeled samples is obtained according to the historical output probability and historical weight corresponding to the remaining unlabeled samples; the remaining unlabeled samples are the samples among the unlabeled samples that have not yet been classified as reliable negative samples.

[0185] The remaining unlabeled samples are divided into reliable negative samples according to their historical weighted average probability from smallest to largest, until the number of positive samples is equal to the number of current reliable negative samples;

[0186] The target data of the positive samples and the target data of the current reliable negative samples are input into the potential user screening model to screen out potential users who meet the preset bandwidth conditions.

[0187] The processor-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).

[0188] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0189] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for screening potential users, characterized in that, include: The target data of positive samples and the target data of unlabeled samples are input into the potential user screening model to train the potential user screening model; the target data of positive samples is the behavioral feature data of the first target user related to bandwidth information, and the target data of unlabeled samples is the behavioral feature data of the second target user related to bandwidth information. The first target user is a user who meets the preset bandwidth conditions, and the second target user is a user who does not meet the preset bandwidth conditions. The potential user screening model is any classification model or a combination of multiple classification models. Determine whether the number of positive samples is equal to the number of reliable negative samples; the reliable negative samples are the samples obtained after classifying the unlabeled samples as negative samples. If the number of positive samples is not equal to the number of reliable negative samples, and the model training times reach the preset number of training times, then the historical weighted average probability corresponding to the remaining unlabeled samples is obtained according to the historical output probability and historical weight corresponding to the remaining unlabeled samples; the remaining unlabeled samples are the samples among the unlabeled samples that have not yet been classified as reliable negative samples. The remaining unlabeled samples are divided into reliable negative samples according to their historical weighted average probability from smallest to largest, until the number of positive samples is equal to the number of current reliable negative samples; The historical weights are obtained as follows: the weights corresponding to the remaining unlabeled samples input into the potential user screening model obtained from the nth model training are obtained by the ratio of the screening accuracy of the potential user screening model obtained from the nth model training to the sum of the screening accuracies of the potential user screening models obtained from the Nth model training; where N is the preset number of training times, and n is a positive integer less than or equal to N. The target data of the positive samples and the target data of the current reliable negative samples are input into the potential user screening model to screen out potential users who meet the preset bandwidth conditions.

2. The potential user screening method according to claim 1, characterized in that, The step of inputting the target data of the positive samples and the target data of the current reliable negative samples into the potential user screening model to screen out potential users who meet the preset bandwidth conditions includes: Input the target data of the positive sample and the target data of the current reliable negative sample into the potential user screening model to obtain the current output probability that the second target user in the current reliable negative sample meets the preset bandwidth condition; Based on the current output probability, potential users who meet the preset bandwidth conditions are selected from the second target users within the current reliable negative sample.

3. The potential user screening method according to claim 1, characterized in that, After determining whether the number of positive samples is equal to the number of reliable negative samples, the process includes: If the number of positive samples is not equal to the number of reliable negative samples, and the number of training iterations of the model has not reached the preset number of training iterations, then the negative samples in the remaining unlabeled samples are oversampled to obtain an oversampled negative sample with the same number of positive samples. The target data of the positive samples and the target data of the oversampled negative samples are input into the potential user screening model to train the potential user screening model; The oversampled negative samples are classified to obtain new reliable negative samples. The process then returns to the step of determining whether the number of positive samples is equal to the number of reliable negative samples, until the number of positive samples is equal to the current number of reliable negative samples, or the number of training iterations of the model reaches the preset number of training iterations.

4. The potential user screening method according to claim 1, characterized in that, Before inputting the target data of positive samples and the target data of unlabeled samples into the potential user screening model, the following steps are included: The first data to be processed is processed to remove missing data, resulting in the second data to be processed; the first data to be processed consists of the behavioral feature data of the first target user and the behavioral feature data of the second target user. The second data to be processed is normalized to obtain the third data to be processed. The third set of data to be processed is filtered to obtain the target data of the positive samples and the target data of the unlabeled samples.

5. The potential user screening method according to claim 4, characterized in that, The missing data processing for the first data to be processed includes: Delete the data corresponding to the duplicate behavior features of the same user in the first data to be processed to obtain non-overlapping behavior feature data; If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data exceeds the first missing threshold, then the data corresponding to any behavioral feature will be deleted. If the missing data corresponding to any behavior feature in the non-overlapping behavior feature data exceeds the second missing threshold but does not exceed the first missing threshold, then new labeled data is added to mark whether the data corresponding to the behavior feature is missing. If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data does not exceed the second missing threshold, and the data corresponding to any behavioral feature is a continuous variable, then the average value of the data corresponding to any behavioral feature is filled into the missing position. If the missing data corresponding to any behavioral feature in the non-overlapping behavioral feature data does not exceed the second missing threshold, and the data corresponding to any behavioral feature is a discrete variable, then the mode of the data corresponding to any behavioral feature is filled into the missing position.

6. The potential user screening method according to claim 4, characterized in that, The filtering process for the third data to be processed includes: The importance value of any behavioral feature in the third data to be processed is obtained by fitting the random forest model to the third data to be processed for the screening of potential users. Behavioral features with importance values ​​greater than the importance threshold are identified as important behavioral features; Calculate the correlation coefficient between any two important behavioral features, and randomly select one of the two important behavioral features whose correlation coefficient is greater than the coefficient threshold as the selected important behavioral feature; The data corresponding to the key behavioral characteristics selected are identified as the selected data.

7. A potential user screening device, characterized in that, For performing the potential user screening method of claim 1, comprising: The model training module is used to: input the target data of positive samples and the target data of unlabeled samples into the potential user screening model to train the potential user screening model; the target data of the positive samples is the behavioral feature data of the first target user related to bandwidth information, the target data of the unlabeled samples is the behavioral feature data of the second target user related to bandwidth information, the first target user is a user who meets the preset bandwidth conditions, the second user is a user who does not meet the preset bandwidth conditions, and the potential user screening model is any classification model or a combination of multiple classification models; The sample quantity determination module is used to: determine whether the number of positive samples is equal to the number of reliable negative samples; the reliable negative samples are the samples obtained after classifying the unlabeled samples as negative samples; The sample probability calculation module is used to: if the number of positive samples is not equal to the number of reliable negative samples, and the number of model training times reaches a preset number of training times, then obtain the historical weighted average probability corresponding to the remaining unlabeled samples based on the historical output probability and historical weight corresponding to the remaining unlabeled samples; the remaining unlabeled samples are the samples among the unlabeled samples that have not yet been classified as reliable negative samples; The sample quantity balancing module is used to: divide the remaining unlabeled samples into reliable negative samples in ascending order of historical weighted average probability, until the number of positive samples is equal to the number of current reliable negative samples; The potential user screening module is used to input the target data of the positive samples and the target data of the current reliable negative samples into the potential user screening model, and screen out potential users that meet the preset bandwidth conditions.

8. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the potential user screening method according to any one of claims 1 to 6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the potential user screening method according to any one of claims 1 to 6.

Images