WO2018196798A1 - 用户客群分类方法和装置 - Google Patents

用户客群分类方法和装置 Download PDF

Info

Publication number
WO2018196798A1
WO2018196798A1 PCT/CN2018/084565 CN2018084565W WO2018196798A1 WO 2018196798 A1 WO2018196798 A1 WO 2018196798A1 CN 2018084565 W CN2018084565 W CN 2018084565W WO 2018196798 A1 WO2018196798 A1 WO 2018196798A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
data samples
cluster
sample set
samples
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2018/084565
Other languages
English (en)
French (fr)
Inventor
陈玲
户保田
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tencent Technology Shenzhen Co Ltd
Original Assignee
Tencent Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tencent Technology Shenzhen Co Ltd filed Critical Tencent Technology Shenzhen Co Ltd
Publication of WO2018196798A1 publication Critical patent/WO2018196798A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/28Databases characterised by their database models, e.g. relational or object models
    • G06F16/284Relational databases
    • G06F16/285Clustering or classification

Definitions

  • the present application relates to the field of communications technologies, and in particular, to a user group classification method and apparatus.
  • the embodiment of the present application provides a user customer group classification method and device, which can not only greatly accelerate the convergence of the classification model, but also improve the accuracy and generalization ability of the classification model, and is beneficial to improving the efficiency of the classification of the user group.
  • the embodiment of the present application provides a user customer group classification method, including:
  • Collecting user data of a plurality of users the user data including data of a plurality of user behavior categories
  • the classification model training is performed based on the sample set to be trained, and the customer classification is performed according to the classification model after the training.
  • an embodiment of the present application further provides a user group classification device, including: a processor and a memory, where the computer stores computer readable instructions, where the computer readable instructions are executed by the processor to complete the following operating:
  • Collecting user data of a plurality of users the user data including data of a plurality of user behavior categories
  • the classification model training is performed based on the sample set to be trained, and the customer classification is performed according to the classification model after the training.
  • the embodiment of the present application further provides a user customer group classification method for a server, where the server includes: a processor and a memory, and the method includes:
  • Collecting user data of a plurality of users the user data including data of a plurality of user behavior categories
  • the classification model training is performed based on the sample set to be trained, and the customer classification is performed according to the classification model after the training.
  • the embodiment of the present application further provides a computer readable storage medium, wherein the computer readable storage medium stores instructions that, when run on a computer, cause the computer to perform the following operations:
  • Collecting user data of a plurality of users the user data including data of a plurality of user behavior categories
  • the classification model training is performed based on the sample set to be trained, and the customer classification is performed according to the classification model after the training.
  • 1A is a schematic diagram of a scenario of a method for classifying user groups provided by an embodiment of the present application
  • FIG. 1B is a flowchart of a user group classification method provided by an embodiment of the present application.
  • 2A is a diagram showing an example of a framework of a user group classification method provided by an embodiment of the present application
  • 2B is another flowchart of a method for classifying user groups provided by an embodiment of the present application.
  • 3A is a schematic structural diagram of a user group classification device provided by an embodiment of the present application.
  • FIG. 3B is another schematic structural diagram of a user group classification device provided by an embodiment of the present application.
  • FIG. 4 is a schematic structural diagram of a server provided by an embodiment of the present application.
  • sample labeling methods for data classification applications such as randomly selecting samples for labeling or selecting information entropy or larger confidence samples for labeling.
  • randomly selecting samples for labeling may make the sample distribution inconsistent with the expected sample, resulting in over-fitting or low accuracy of the classification model.
  • the method of selecting the information entropy or the larger confidence value for labeling may also have similar similarity between the selected samples. Therefore, the sample is too concentrated, which is not conducive to the convergence of the classification model. , greatly affecting the accuracy and generalization ability of the classification model, resulting in low efficiency of data classification.
  • the embodiment of the present application provides a user group classification method and device.
  • the user group classification device may be specifically integrated into a server, such as a service server.
  • the server can collect user data of multiple users, select data samples from the user data according to a preset policy, obtain a data sample set, and then concentrate the data sample.
  • the data samples are clustered to obtain multiple clusters, and the data samples in each cluster are respectively subjected to de-redundancy operations to obtain a sample set to be trained.
  • multiple clusters are respectively classified into class A and B.
  • class N, N class, etc. at this time, the data samples in clusters such as class A, class B, ... class N can be deduplicated separately, and multiple clusters after deduplication are obtained.
  • the server may perform classification model training based on the sample set to be trained.
  • the data samples in the sample set to be trained may be labeled, and then the initial classification model is trained according to the labeled data samples, and classified according to the training. Treat type classification data customer base classification.
  • the embodiment of the present application will be described from the perspective of a user group classification device, and the user group classification device may be specifically integrated into a server, such as a service server.
  • a user customer group classification method includes: collecting user data of a plurality of users, selecting data samples from user data of the plurality of users according to a preset policy, obtaining a data sample set, and collecting data samples in the data sample set.
  • Class operation obtain multiple clusters, perform de-redundancy operation on the data samples in each cluster, obtain the sample set to be trained, perform classification model training based on the sample set to be trained, and classify according to the post-training classification model
  • the data is classified into customer groups.
  • the specific process of the user group classification method may be as follows:
  • the user group classification device may specifically collect user data of various types of users from the network, or may collect other user data of various types of users by using other collection devices, and then when the user data needs to be used, The user customer group classification device acquires from the collection device, and the like.
  • the user data may include data of multiple user behavior categories, where the user behavior category may be classified according to user behavior characteristics.
  • social related data may be classified into social information
  • payment related data may be divided into
  • the payment behavior data, the entertainment-related data may be classified into entertainment and leisure behavior data, and the like; that is, the user data may include the user's account information, social information, payment behavior data, entertainment and leisure behavior data, educational behavior data, and/or Or other network behavior data, etc., for example, can be as follows:
  • User account information such as gender, region, age, and/or general settings.
  • User's social information such as chat, email, voice call, information published in Weibo space, community comments, and/or reading of public article articles.
  • User's payment behavior data such as online shopping, game card payment, and/or music traffic package.
  • Entertainment and leisure behavior data such as video on demand, music playback, K song, and / or news reading.
  • Educational behavior data such as online reading, open class learning, vocational exam exercises, skills training, and/or translation software use.
  • Its network behavior data such as terminal application (App, Application) download, and / or search.
  • the preset policy can be set according to the requirements of the actual application, for example, it can be set to “clean the user data”. That is, the step of selecting a data sample from the user data according to a preset policy to obtain a data sample set may specifically include:
  • the collected user data is cleaned to obtain a data sample set. That is, the collected user data can be cleaned at this time, and then the cleaned user data is added as a data sample to the data sample set.
  • the cleaning method can be various, such as “removing singular points”, “filling missing values”, and / or “removing data collected repeatedly”, etc.; for example, it can be as follows:
  • the singular points in the user data are culled, and the remaining user data is added as a data sample to the data sample set.
  • the preset value is assigned to the data that is not collected (that is, the default value is assigned), and then the collected user data and the assigned data are used as data samples, and then added. To the data sample set.
  • the repeatedly collected data in the user data is culled, and the remaining user data is added as a data sample to the data sample set.
  • the cleaned data that is, the data samples in the data sample set
  • the cleaned data can be classified and written into a preset database for use in subsequent steps.
  • the clustering algorithm may also be determined according to the requirements of the actual application. For example, a proximity algorithm (kNN, k-NearestNeighbor), or a k-means algorithm (a hard clustering algorithm) may be used.
  • a proximity algorithm kNN, k-NearestNeighbor
  • a k-means algorithm a hard clustering algorithm
  • the data samples belonging to the social information may be grouped into one cluster, and the data samples belonging to the entertainment and leisure information may be grouped into one cluster. And so on, and so on.
  • the similarity of the data samples in each cluster may be separately calculated, and the redundant samples in each cluster are eliminated according to the similarity to obtain a sample set to be trained, which may be as follows:
  • the maximum amount of information refers to the maximum amount of information that the data sample can provide, and the sum of the maximum amount of information and the amount of information of the same data sample is 1. For example, if the maximum amount of information for a data sample is 0.8, then the amount of information is 0.2, and so on, and so on.
  • the amount of information refers to the information metric or content required to select an event from M equal possible events, that is, the question "yes or no" is required in the process of identifying a specific event in M events. The minimum number of times.
  • the amount of information is related to the probability of occurrence of the event, and the maximum amount of information is also related to the probability of occurrence of the event. Therefore, the maximum amount of information can be calculated by calculating the probability of occurrence of the event, thereby determining the magnitude of the uncertainty of the event.
  • the probability of occurrence of an event mainly refers to a probability value that the classification model predicts that the data sample is a certain type. Therefore, in the embodiment of the present application, the probability value may be obtained by calculating the probability value.
  • the maximum amount of information that is, optionally, the step of "maximizing the maximum amount of information of the data samples in each cluster by using the maximum information amount algorithm" may include:
  • the customer classification of the data samples in each cluster is predicted by using a preset initial classification model, and the maximum information amount of the corresponding data samples is calculated according to the prediction result.
  • the maximum amount of information of the corresponding data sample may be directly calculated according to the prediction result (such as the probability that the user belongs to a certain classification), or the confidence of the prediction result may be calculated, and then the corresponding data sample is calculated according to the confidence level.
  • the confidence here the most uncertain
  • the calculation of the most uncertain can be calculated by the following formula 1 (not limited to this formula), that is, the lower the confidence of the most uncertain sample, the higher the converse.
  • the initial classification model may be preset by a maintenance personnel, or may be obtained by training by the user group classification device.
  • the user group classification method may further include:
  • a predetermined number of data samples are randomly selected from each cluster, and the customer group classification to which they belong is manually labeled, and the preset original model is trained according to the manually labeled data samples to obtain an initial classification model.
  • the step “predicting the customer group classification of the data samples in each cluster by using the preset initial classification model” may include: using the obtained initial classification model, and not performing other artificial in each cluster. The customer group classification of the annotated data samples is predicted.
  • the original model used in the training can be set according to the requirements of the actual application, and the predetermined quantity can also be set according to the requirements of the actual application, for example, taking 1000 initial training samples as an example, if there are 5 in total Clustering, each cluster can randomly select 200 data samples as the initial training samples of the corresponding clusters, and so on.
  • the corresponding maximum information amount calculation formula can be as follows:
  • x is the data sample
  • L is the initial training sample set (ie, the manually labeled data sample)
  • C is one or a set of classification models (also called classifiers)
  • Pr[x 1
  • L] is at the initial
  • the classification model C trained on the training sample set L predicts the probability value of the data sample as "Class 1" (where "1" represents "Class 1")
  • argmin is made The function that takes the smallest data sample, U is the unlabeled sample set (ie, the unlabeled data sample in the cluster).
  • a cluster A includes data samples x 1 , x 2 , and x 3 , and the classification model C calculates the probability of belonging to the classification 1 to be 0.49, 0.1, and 0.8, respectively.
  • the maximum amount of information calculated is 0.01, 0.4, and 0.3, that is, the uncertainty of x 1 is the largest, the uncertainty of x 3 is the second, and the uncertainty of x 2 is the smallest.
  • three samples are selected.
  • the priority of the sample set to be labeled is: x 1 , x 3 , x 2 . And so on, will not repeat them here.
  • K is a positive integer
  • the specific value can be set according to the requirements of the actual application. For example, taking K as 2 as an example, if in step S1, the priority of the three samples selected into the sample set to be labeled is: x 1 . x 3 , x 2 , then x 1 and x 3 can be added to the sample set to be labeled.
  • the sample set to be labeled includes data samples x 1 , x 3 , x 5 , and x 6 .
  • the similarity calculation may first quantize the user information of the data sample and convert it into a vector space, and then calculate the similarity by using the standardized Euclidean distance, the cosine of the included angle, and the correlation coefficient.
  • a threshold may be set according to the obtained similarity, and above the threshold, the two data samples are considered to be highly similar.
  • the similarity greater than the preset threshold is referred to as a height similarity.
  • the threshold can be set according to the requirements of the actual application, and details are not described herein again.
  • a data sample whose similarity is greater than a preset threshold may be obtained, and a candidate set is obtained, and a data sample whose maximum information amount satisfies a preset condition is selected from the candidate set to obtain a sample set to be trained.
  • the preset condition may be determined according to the needs of the actual application, for example, by counting highly similar counts, and rejecting samples with a high degree of similarity count, and the like. That is, the step of “selecting the data sample that meets the preset condition from the candidate set to obtain the sample set to be trained” may include:
  • N is a positive integer, and the value of N may be determined according to the needs of the actual application.
  • the preset initial classification model may be obtained, and the data samples in the sample set to be trained are marked, and then the initial classification model is trained according to the labeled data samples to obtain a post-training classification model.
  • the data to be classified may be obtained, and the data to be classified is imported into the post-training classification model to obtain an output result, and the customer group classification of the data to be classified is determined according to the output result.
  • the customer group refers to the customer group, for example, it can be divided into “student”, “white collar” and “golden collar”, for example, it can also be divided into “children”, “youth”, “middle age” and “old age”. , etc., the specific division requirements of the customer group can be determined according to the needs of the actual application, and will not be described here.
  • the embodiment of the present application may perform clustering operations on the data samples, and respectively perform de-redundancy operations on the data samples in each cluster to obtain a to-be-redundant operation.
  • the user group classification device is specifically integrated into the server as an example for description.
  • a user customer group classification method may be as follows:
  • the server collects user data of multiple users.
  • the server may collect user data of various types of users from the network, or may collect user data of various types of users by other collecting devices, and then collect the user data when the user data is needed. Get in the device, and so on.
  • the user data may include data of various user behavior categories, for example, may include user account information, social information, payment behavior data, entertainment and leisure behavior data, educational behavior data, and/or other network behavior data.
  • the server cleans the collected user data to obtain a data sample set.
  • the cleaning method can be various, such as “removing singular points”, “filling missing values”, and / or “removing data collected repeatedly”, etc.; that is, the step “the server will clean the collected user data, Get the data sample set” can be as follows:
  • the server may specifically cull the singular points in the user data, and add the remaining user data as a data sample to the data sample set;
  • the server may determine, according to the user data, that the data should be collected but not collected, and assign a preset value to the data that is not collected (ie, assign a default value), and then collect the collected user data and the assigned data. As a data sample, added to the data sample set;
  • the server may also cull the repeatedly collected data in the user data, add the remaining user data as a data sample, to the data sample set, and the like.
  • data samples of the data sample set are all data samples of the customer group classification that are not labeled.
  • the server obtains a user behavior category of the data sample, and performs clustering operation on the data sample in the data sample set according to the user behavior category to obtain multiple clusters.
  • the server may perform a clustering operation on the data samples in the data sample set by using a clustering algorithm such as the kNN or the k-means algorithm to obtain a plurality of clusters.
  • a clustering algorithm such as the kNN or the k-means algorithm
  • the server randomly selects a predetermined number of data samples from each cluster, manually labels the customer group classifications, and trains the preset original model according to the manually labeled data samples to obtain an initial classification model.
  • the original model used in the training can be set according to the requirements of the actual application, and the predetermined quantity can also be set according to the requirements of the actual application.
  • cluster A For example, as shown in FIG. 2A, if a total of 1000 data samples need to be selected, at this time, it is possible to separately select 5 clusters, such as cluster A, cluster B, cluster C, cluster D, and cluster E. Randomly select a certain number, such as about 200 unlabeled data samples, manually mark the customer group classification to form the initial training samples, and then use these initial training samples to train the preset original model to obtain the initial classification. model.
  • the set of the initial training samples may be referred to as an initial training sample set, and the initial training sample set may have multiple initial training sample subsets, and each initial training sample subset corresponds to one cluster.
  • the initial training sample subset corresponding to the cluster A the initial training sample subset corresponding to the cluster B
  • the initial training sample subset corresponding to the cluster C the initial training sample subset corresponding to the cluster D
  • the initial training sample subset corresponding to cluster E and so on.
  • the initial training sample subset corresponding to the cluster A may include the labeled data samples corresponding to the cluster A
  • the initial training sample subset corresponding to the cluster B may include the labeled data samples corresponding to the cluster B
  • clustering The initial training sample subset corresponding to C may include the labeled data samples corresponding to the cluster C
  • the initial training sample subset corresponding to the cluster D may include the labeled data samples corresponding to the cluster D
  • the cluster E corresponds to
  • the initial training sample subset may include the labeled data samples corresponding to the cluster E, and so on.
  • the server uses the initial classification model obtained in step 204 to predict the customer group classification of the data samples in each cluster, and calculates the maximum information amount of the corresponding data samples according to the prediction result.
  • the server may use the initial classification model to cluster respectively.
  • A, cluster B, cluster C, cluster D, and the customer group classification of the data samples in the cluster E are predicted, and then the maximum information amount of the corresponding data sample is calculated according to the prediction result.
  • the server may use the initial classification model to predict the customer group classification of the data samples x 1 , x 2 and x 3 , if the prediction data samples x 1 , x 2 and x 3 may belong to “student”, and the belongs to “ Students with a probability of 0.49, 0.1, and 0.8, respectively, can pass the following formula
  • the maximum amount of information for x 1 , x 2 and x 3 is calculated separately: 0.01, 0.4 and 0.3.
  • L is the initial training sample set (ie, the manually labeled data sample)
  • C is one or a group of classification models
  • Pr[x 1
  • L] is the classification model C pair trained on the initial training sample set L.
  • the probability that the data sample is predicted as "student”, To make The data sample of the minimum value is obtained, and U is an unlabeled sample set (ie, an unlabeled data sample in the cluster).
  • the maximum information amount of the data samples in the clusters B, C, D, and E can also be calculated in the above manner, and will not be described herein.
  • the server filters each cluster to obtain a sample set to be labeled.
  • the sample set to be labeled may include a plurality of sample subsets to be labeled, and each sample subset to be labeled corresponds to one cluster, for example, may include a subset of samples to be labeled corresponding to cluster A, and a corresponding group of cluster B Labeling the subset of samples, the subset of samples to be labeled corresponding to cluster C, the subset of samples to be labeled corresponding to cluster D, and the subset of samples to be labeled corresponding to cluster E, and so on.
  • the subset of samples to be labeled corresponding to A is called cluster A'
  • the subset of samples to be labeled corresponding to B is called cluster B'
  • the subset of samples to be labeled corresponding to C is called cluster C'
  • D corresponds to
  • the subset of samples to be labeled is called cluster D'
  • the subset of samples to be labeled corresponding to E is called cluster E'.
  • the screening method may be various.
  • the server may select the first K data samples with the largest uncertainty from each cluster according to the maximum information amount, and obtain the sample set to be labeled.
  • K is a positive integer, and the specific value can be set according to the requirements of the actual application.
  • the order is sorted according to the uncertainty from large to small.
  • the order of the data samples is: x 1 , x 3 , x 2 , since K is 2, therefore, the first two data samples with the most certain deterministic, ie, x 1 and x 3 , can be added to the sample set to be labeled.
  • the cluster A' ie, the subset of samples to be labeled corresponding to cluster A.
  • the first K data samples with the largest maximum amount of information can be selected from the clusters B, C, D and E in the same way, and added to the clusters B', C', D' and the clusters to be labeled respectively. E'.
  • the server calculates a similarity between the two data samples in the sample set to be labeled, and obtains the similarity of the data samples in each cluster.
  • the sample set to be labeled (ie, the subset of samples to be labeled corresponding to cluster A of the sample set to be labeled) includes data samples x 1 , x 3 , x 5 , x 6 and x 7 , then, at this time, it is necessary to calculate the similarity of x 1 and x 3 , the similarity of x 1 and x 5 , the similarity of x 1 and x 6 , the similarity of x 1 and x 7 , x 3 and x 5
  • the similarity, the similarity of x 3 and x 6 , the similarity of x 3 and x 7 , the similarity of x 5 and x 6 , x 5 and x 7 , and x 6 and x 7 ie, at this time:
  • the similarity of the data samples x 1 in this cluster is: the similarity of x 1 and x 3 , the similarity of x 1 and x 5 , the similarity of x 1 and x 6 , and the similarity of x 1 and x 7 ;
  • the similarity of the data samples x 3 in the cluster A is: the similarity of x 1 and x 3 , the similarity of x 3 and x 5 , the similarity of x 3 and x 6 , and the similarity of x 3 and x 7 ;
  • the similarity of the data samples x 5 in the cluster A is: the similarity of x 1 and x 5 , the similarity of x 3 and x 5 , the similarity of x 5 and x 6 , and the similarity of x 5 and x 7 ;
  • the similarity of the data samples x 6 in the cluster A is: the similarity of x 1 and x 6 , the similarity of x 3 and x 6 , the similarity of x 5 and x 6 , and the similarity of x 5 and x 7 ;
  • the similarity of the data samples x 7 in cluster A is: the similarity of x 1 and x 7 , the similarity of x 3 and x 7 , the similarity of x 5 and x 7 , and the similarity of x 6 and x 7 .
  • the user information of the data sample may be first quantized and converted into a vector space, and then the similarity is calculated by using the standardized Euclidean distance, the cosine of the included angle, and the correlation coefficient.
  • a threshold may be set according to the obtained similarity, and above the threshold, the two data samples are considered to be highly similar, which is called high similarity.
  • the threshold may be set according to the requirements of the actual application, for example, the threshold may be set to 0.6, and the like.
  • the server culls the redundant samples in each cluster according to the similarity (ie, performs a de-redundancy operation) to obtain a sample set to be trained.
  • the sample set to be trained may include a plurality of sample subsets to be trained, and each subset of samples to be trained corresponds to one cluster, for example, may include a subset of samples to be trained corresponding to cluster A, and a corresponding group of cluster B
  • the subset of samples to be trained corresponding to cluster A is called cluster A′′
  • the subset of samples to be trained corresponding to cluster B is called cluster B′′
  • the subset of samples to be trained corresponding to cluster C is called clustering.
  • C′′ the subset of samples to be trained corresponding to cluster D is called cluster D′′
  • the subset of samples to be trained corresponding to cluster E is called cluster E′′.
  • the manner of de-redundancy may be multiple.
  • the server may specifically obtain data samples whose similarity is greater than a preset threshold, obtain a candidate set, and then count the number of high similarities of each data sample in the candidate set. A highly similar count is obtained, and the top N data samples with the highest height similarity count are eliminated from the candidate set to obtain a sample set to be trained.
  • N is a positive integer, and the value of N may be determined according to the needs of the actual application.
  • the height similarity count of the data sample x 1 is 1, the height similarity count of the data sample x 3 is 3, and the data sample x 5
  • the height similarity count is 1
  • the data sample x 6 has a height similarity count of 2
  • the data sample x 7 has a height similarity count of 1, which is visible, wherein the data sample x 3 has the highest height similarity count, and if N is 1, then,
  • the server performs classification model training based on the sample set to be trained.
  • the server may obtain a preset initial classification model, label the data samples in the sample set to be trained, and then train the initial classification model according to the labeled data samples to obtain a post-training classification model. .
  • the labeling may be manually marked, or may be marked by the server according to a preset policy, and details are not described herein.
  • the original model, the initial classification model, and the post-training classification model are the expressions of the same “classification model” during different training periods. Therefore, for the convenience of description, in FIG. 2A, the original model and the initial classification model are used. And the post-training classification model is expressed as a “classification model”. It should be understood that the expression pattern of the classification model may be different at different times.
  • the server classifies the customer according to the post-training classification model.
  • the server may obtain the data to be classified, import the data to be classified into the post-training classification model to perform the customer group classification prediction, and then determine the customer group classification of the data to be classified according to the output of the post-training classification model to obtain the classification.
  • the data please refer to FIG. 2A.
  • the data to be classified such as the account information of the user to be classified, social information, payment behavior data, entertainment and leisure behavior data
  • Information such as educational behavior data and/or other network behavior data is imported into the post-training classification model, and then, according to the output of the post-training classification model, whether the data to be classified is data corresponding to “student” or data corresponding to “white-collar” is determined.
  • the corresponding users are thus classified into corresponding customer groups, such as "student” or "white collar”.
  • the embodiment of the present application may perform clustering operations on the data samples, and respectively perform de-redundancy operations on the data samples in each cluster to obtain a to-be-redundant operation.
  • the embodiment of the present application further provides a user customer group classification device, which may be integrated into a server, such as a service server or the like.
  • the user customer group classification device may include a processor and a memory, the computer readable instructions stored thereon, the computer readable instructions being executed by the processor to perform the following operations: collecting user data of a plurality of users
  • the user data includes data of a plurality of user behavior categories; selecting data samples from the user data of the plurality of users according to a preset policy to obtain a data sample set; performing clustering operations on the data samples in the data sample set Obtaining a plurality of clusters; respectively performing de-redundancy operations on the data samples in each cluster to obtain a sample set to be trained; performing classification model training based on the sample set to be trained, and classifying the data according to the post-training classification model Conduct customer group classification.
  • the performing clustering operation on the data samples in the data sample set to obtain a plurality of clusters includes: acquiring a user behavior category of the data samples; and performing, according to the user behavior category, the data samples in the data sample set Clustering operations to obtain multiple clusters.
  • the performing a de-redundancy operation on the data samples in each cluster to obtain a sample set to be trained includes: calculating a similarity of the data samples in each cluster separately; and performing each of the similarities according to the similarity
  • the redundant samples in the cluster are culled to obtain a sample set to be trained.
  • the calculating the similarity of the data samples in each cluster separately includes: calculating a maximum information amount of the data samples in each cluster by using a maximum information amount algorithm; and obtaining each cluster according to the maximum information amount Select the top K data samples with the largest uncertainty, and obtain the sample set to be labeled, where K is a positive integer; calculate the similarity between the two data samples in the sample set to be labeled, and obtain the data samples in each cluster. Similarity.
  • the method for maximizing the maximum amount of data samples in each cluster by using a maximum information amount algorithm includes: predicting a customer group classification of data samples in each cluster by using a preset initial classification model; The maximum amount of information of the corresponding data sample is calculated based on the predicted result.
  • the method further includes: randomly selecting a predetermined number of data samples from each cluster for manual labeling;
  • the labeled data sample trains the preset original model to obtain an initial classification model;
  • the preset initial classification model predicts the customer group classification of the data samples in each cluster, including: using the obtained initial A classification model that predicts the population classification of other data samples that are not manually labeled in each cluster.
  • the culling the redundant samples in each type of data samples according to the similarity to obtain a sample set to be trained includes: acquiring data samples with similarities greater than a preset threshold to obtain candidate sets; and screening the candidate sets to the largest The data sample that satisfies the preset condition is obtained, and the sample set to be trained is obtained.
  • the instruction is, for example, an instruction module in the form of software, and the specific description is as follows:
  • the user group classification device includes an acquisition unit 301, a selection unit 302, a clustering unit 303, a deduplication unit 304, and a classification unit 305, as follows:
  • the collecting unit 301 is configured to collect user data of multiple users.
  • the collecting unit 301 can collect user data of various types of users from the network, or can collect user data of various types of users by other collecting devices, and then collect the user data when the user data needs to be used.
  • Unit 301 retrieves from the collection device, and so on.
  • the user data may include data of various user behavior categories, for example, may include user account information, social information, payment behavior data, entertainment and leisure behavior data, educational behavior data, and/or other network behavior data.
  • the selecting unit 302 is configured to select a data sample from the user data according to a preset policy to obtain a data sample set.
  • the preset policy can be set according to the requirements of the actual application, for example, it can be set to “clean the user data”.
  • the cleaning method can be various, such as "removing singular points”, “filling missing values”, and/or “removing data collected repeatedly”, and the like.
  • the clustering unit 303 is configured to perform clustering operations on the data samples in the data sample set to obtain a plurality of clusters.
  • the clustering unit 303 may be configured to acquire a user behavior category of the data sample, and perform clustering operations on the data samples in the data sample set according to the user behavior category to obtain multiple clusters.
  • the clustering algorithm may also be determined according to the requirements of the actual application. For example, a kNN or k-means algorithm may be used.
  • the redundancy unit 304 is configured to perform a de-redundancy operation on the data samples in each cluster to obtain a sample set to be trained.
  • the deduplication unit 304 can include a calculation subunit and a filter subunit, as follows:
  • the calculation subunit is configured to calculate the similarity of the data samples in each cluster separately.
  • the screening subunit is configured to cull redundant samples in each cluster according to the similarity to obtain a sample set to be trained.
  • the calculation subunit may be specifically configured to calculate a maximum information amount of each data sample in each cluster by using a maximum information amount algorithm, and select the most uncertain before each cluster according to the maximum information amount.
  • K data samples obtain the sample set to be labeled, calculate the similarity between the two data samples in the sample set to be labeled, and obtain the similarity of the data samples in each cluster; wherein K is a positive integer.
  • the screening sub-unit may be configured to obtain a data sample with a similarity greater than a preset threshold, obtain a candidate set, and filter a data sample that meets a preset condition from a candidate set to obtain a sample set to be trained; for example, the statistics may be collected.
  • the number of high similarities of each data sample in the candidate set is highly similar, and the top N data samples with the highest height similarity count are eliminated from the candidate set, the sample set to be trained is obtained, and the like.
  • the calculating sub-unit may be specifically used to predict a customer group classification of data samples in each cluster by using a preset initial classification model, and calculate a maximum information amount of the corresponding data samples according to the prediction result; for example, directly according to the prediction
  • the result (such as the probability that the user belongs to a certain classification) calculates the maximum amount of information of the corresponding data sample, or can also calculate the confidence of the prediction result, and then, according to the confidence level, calculate the maximum information amount of the corresponding data sample, etc. .
  • the initial classification model may be preset by a maintenance personnel or may be obtained by training by the user group classification device. For example, a certain number of data samples may be randomly selected from each cluster for manual labeling to form an initial Training the sample, and then training the preset original model with the initial training samples to obtain an initial classification model; that is, optionally, as shown in FIG. 3B, the user group classification device may further include a model establishing unit 306. as follows:
  • the model establishing unit 306 can be configured to randomly select a predetermined number of data samples from each cluster for manual labeling, and train the preset original model according to the manually labeled data samples to obtain an initial classification model.
  • the calculation sub-unit can be specifically used for the initial classification model obtained by the model establishment unit, and the customer group classification of other data samples that are not manually labeled in each cluster is predicted.
  • the original model used in the training may be set according to the requirements of the actual application, and the predetermined number may also be set according to the requirements of the actual application, and details are not described herein again.
  • the classification unit 305 is configured to perform classification model training based on the to-be-trained sample set, and perform customer group classification according to the post-training classification model.
  • the classification unit 305 is specifically configured to acquire a preset initial classification model, label the data samples in the sample set to be trained, and train the initial classification model according to the labeled data samples to obtain a post-training classification model.
  • the data to be classified is obtained, and the data to be classified is imported into the post-training classification model to obtain an output result, and the customer group classification of the data to be classified is determined according to the output result.
  • the foregoing various units may be implemented as a separate entity, and may be implemented in any combination, and may be implemented as the same or a plurality of entities.
  • the foregoing embodiments, and details are not described herein again.
  • the clustering unit 303 can perform clustering operations on the data samples, and the deduplication unit 304 respectively Performing a de-redundancy operation on the data samples in each cluster to obtain a sample set to be trained, and then the classification unit 305 performs the classification model training based on the sample set to be trained, and performs the guest classification data according to the post-training classification model.
  • Group classification because the scheme clusters the data samples before training the classification model, and then selects the data samples from each cluster for training, so that each cluster has data.
  • the sample is selected into the training samples of the classification model, so that the classification model can learn from each cluster to the greatest extent, covering the expected distribution of the sample, greatly improving the accuracy and generalization ability of the classification model; De-redundancy operations can be performed on data samples in each cluster to eliminate redundant data samples with high similarity, so that Minimize the similarity of data samples, so it can greatly accelerate the convergence of the classification model; in short, the program can not only greatly accelerate the convergence of the classification model, but also improve the accuracy and generalization ability of the classification model, which is conducive to improving users. The efficiency of group classification.
  • the embodiment of the present application further provides a server, as shown in FIG. 4, which shows a schematic structural diagram of a server involved in the embodiment of the present application, specifically:
  • the server may include one or more processor cores 401, one or more memories 402, a power source 403, and an input unit 404.
  • Memory 402 is a computer readable storage medium.
  • the server structure shown in FIG. 4 does not constitute a limitation to the server, and may include more or less components than those illustrated, or some components may be combined, or different component arrangements. among them:
  • the processor 401 is the control center of the server, connecting various portions of the entire server using various interfaces and lines, by running or executing software programs and/or modules stored in the memory 402, and recalling data stored in the memory 402, Execute the server's various functions and process data to monitor the server as a whole.
  • the processor 401 may include one or more processing cores; preferably, the processor 401 may integrate an application processor and a modem processor, where the application processor mainly processes an operating system, a user interface, an application, and the like.
  • the modem processor primarily handles wireless communications. It can be understood that the above modem processor may not be integrated into the processor 401.
  • the memory 402 can be used to store software programs and modules, and the processor 401 executes various functional applications and data processing by running software programs and modules stored in the memory 402.
  • the memory 402 may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application required for at least one function (such as a sound playing function, an image playing function, etc.), and the like; the storage data area may be stored according to Data created by the use of the server, etc.
  • memory 402 can include high speed random access memory, and can also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid state storage device. Accordingly, memory 402 can also include a memory controller to provide processor 401 access to memory 402.
  • the server also includes a power source 403 that supplies power to the various components.
  • the power source 403 can be logically coupled to the processor 401 via a power management system to enable management of charging, discharging, and power management functions through the power management system.
  • Power source 403 may also include any one or more of a DC or AC power source, a recharging system, a power failure detection circuit, a power converter or inverter, a power status indicator, and the like.
  • the server can also include an input unit 404 that can be used to receive input numeric or character information and to generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function controls.
  • an input unit 404 can be used to receive input numeric or character information and to generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function controls.
  • the server may further include a display unit or the like, and details are not described herein again.
  • the processor 401 in the server loads the executable file corresponding to the process of one or more applications into the memory 402 according to the following instruction, and is stored and stored by the processor 401.
  • Collecting user data of multiple users includes data of a plurality of user behavior categories, selecting data samples from the user data according to a preset policy, obtaining a data sample set, and performing clustering operations on the data samples in the data sample set.
  • Obtaining multiple clusters respectively performing de-redundancy operations on the data samples in each cluster, obtaining a sample set to be trained, performing classification model training based on the sample set to be trained, and performing classification data according to the post-training classification model Customer group classification.
  • the similarity of the data samples in each cluster may be separately calculated, and the redundant samples in each cluster are eliminated according to the similarity, and the samples to be trained are obtained. set.
  • the maximum information amount algorithm can be used to calculate the maximum information amount of each data sample in each cluster, and the top K data samples with the largest uncertainty are selected from each cluster according to the maximum information amount, and the sample to be labeled is obtained.
  • Set calculate the similarity between the two data samples in the sample set to be labeled, obtain the similarity of the data samples in each cluster, obtain the data samples whose similarity is greater than the preset threshold, obtain the candidate set, and count the candidate set.
  • the number of highly similarities of the data samples is highly similar, and the top N data samples with the highest height similarity count are eliminated from the candidate set, the sample set to be trained is obtained, and the like.
  • the server in the embodiment of the present application may perform clustering operations on the data samples, and perform deduplication operations on the data samples in each cluster respectively.
  • the sample set to be trained is obtained, and then the classification model is trained based on the sample set to be trained, and the customer classification is performed according to the classification model after the training; since the scheme performs the classification model training, the data sample is first performed. Clustering, and then selecting data samples from each cluster for training, so that it is ensured that each cluster has data samples selected into the training samples of the classification model, maximizing the classification model.
  • the scheme can learn from each cluster, cover the expected distribution of samples, greatly improve the accuracy and generalization ability of the classification model; in addition, because the scheme can perform de-redundancy operation on the data samples in each cluster to eliminate Redundant data samples with high similarity, so that data samples with similar similarity can be reduced as much as possible. In general, the scheme can not only greatly accelerate the convergence of the classification model, but also improve the accuracy and generalization ability of the classification model, which is beneficial to improve the efficiency of user group classification.
  • All or part of the steps of the above embodiments may be completed by a program to instruct related hardware, and the program may be stored in a computer readable storage medium, and the storage medium may include: a read only memory (ROM, Read) Only Memory), Random Access Memory (RAM), disk or CD.
  • ROM read only memory
  • RAM Random Access Memory

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本申请实施例公开了一种用户客群分类方法和装置;本申请实施例可以采集多个用户的用户数据,按照预设策略从这些用户数据中选择数据样本,然后,对这些数据样本进行聚类运算,并分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,再基于该待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。

Description

用户客群分类方法和装置
本申请要求于2017年04月28日提交中国专利局、申请号为201710293870.6、发明名称为“一种用户客群分类方法和装置”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请涉及通信技术领域,具体涉及一种用户客群分类方法和装置。
背景技术
在许多数据分类应用中,如文本分类、图像分类、以及特殊客群的挖掘等,都需要大量的标注样本来进行分类模型训练,而标注样本通常难以自动获取,一般都需要人工进行标注。如何合理地从数据中选择样本进行标注为业界研究的一个问题。
发明内容
本申请实施例提供一种用户客群分类方法和装置,不仅可以大大加快分类模型的收敛,而且可以提高分类模型的准确率和泛化能力,有利于提高用户客群分类的效率。
本申请实施例提供一种用户客群分类方法,包括:
采集多个用户的用户数据,所述用户数据包括多种用户行为类别的数据;
按照预设策略从所述多个用户的用户数据中选择数据样本,得到数据样本集;
对所述数据样本集中的数据样本进行聚类运算,得到多个聚类;
分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集;
基于所述待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。
相应地,本申请实施例还提供一种用户客群分类装置,包括:处理器和存储器,所述存储器上存储有计算机可读指令,所述计算机可读指令由所述处理器执行以完成以下操作:
采集多个用户的用户数据,所述用户数据包括多种用户行为类别的数据;
按照预设策略从所述多个用户的用户数据中选择数据样本,得到数据样本集;
对所述数据样本集中的数据样本进行聚类运算,得到多个聚类;
分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集;
基于所述待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。
本申请实施例还提供一种用户客群分类方法,用于服务器,所述服务器包括:处理器和存储器,所述方法包括:
采集多个用户的用户数据,所述用户数据包括多种用户行为类别的数据;
按照预设策略从所述多个用户的用户数据中选择数据样本,得到数据样本集;
对所述数据样本集中的数据样本进行聚类运算,得到多个聚类;
分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集;
基于所述待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。
本申请实施例还提供一种计算机可读存储介质,其中,所述计算机可读存储介质中存储有指令,当其在计算机上运行时,使得计算机执行以下操作:
采集多个用户的用户数据,所述用户数据包括多种用户行为类别的数据;
按照预设策略从所述多个用户的用户数据中选择数据样本,得到数据样本集;
对所述数据样本集中的数据样本进行聚类运算,得到多个聚类;
分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集;
基于所述待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。
附图简要说明
为了更清楚地说明本申请实施例中的技术方案,下面将对实施例描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本申请 的一些实施例,对于本领域技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的附图。
图1A是本申请实施例提供的用户客群分类方法的场景示意图;
图1B是本申请实施例提供的用户客群分类方法的流程图;
图2A是本申请实施例提供的用户客群分类方法的框架示例图;
图2B是本申请实施例提供的用户客群分类方法的另一流程图;
图3A是本申请实施例提供的用户客群分类装置的结构示意图;
图3B是本申请实施例提供的用户客群分类装置的另一结构示意图;
图4是本申请实施例提供的服务器的结构示意图。
实施本发明的方式
下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本申请一部分实施例,而不是全部的实施例。基于本申请中的实施例,本领域技术人员在没有作出创造性劳动前提下所获得的所有其他实施例,都属于本申请保护的范围。
目前已有一些用于数据分类应用的样本标注方法,例如随机选择样本进行标注或者将信息熵或置信度较大的样本挑选出来进行标注。但随机选择样本进行标注可能会使样本分布与预期样本不一致,导致分类模型过拟合或者准确率不高。而使用将信息熵或者置信度较大的样本挑选出来进行标注的方法,所挑选的样本之间的相似性也可能较大,因此,往往会出现样本过于集中的情况,不利于分类模型的收敛,大大影响分类模型的准确率和泛化能力,导致数据分类的效率较低。
本申请实施例提供一种用户客群分类方法和装置。
其中,该用户客群分类装置具体可以集成在服务器,如业务服务器等设备中。例如,以集成在服务器中为例,参见图1A,该服务器可以采集多个用户的用户数据,按照预设策略从该用户数据中选择数据样本,得到数据样本集,然后,对该数据样本集中的数据样本进行聚类运算,得到多个聚类,并分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,比如,以多个聚类分别为A类、B类、……N类等为例,则此时,可以分别对A类、B类、……N类等聚类中的数据样本进行去冗余操作,得到去冗余后的多个聚类,如A″类、B″ 类、……N″类等,其中,为了描述方便,在本申请实施例中,将这去冗余后的多个聚类统称为待训练样本集;此后,服务器便可以基于该待训练样本集进行分类模型训练,比如,可以对该待训练样本集中的数据样本进行标注,然后,根据标注后的数据样本对初始分类模型进行训练,并根据训练后分类模型对待分类数据进行客群分类。
以下分别进行详细说明。
本申请实施例将从用户客群分类装置的角度进行描述,该用户客群分类装置具体可以集成在服务器,比如业务服务器等设备中。
一种用户客群分类方法,包括:采集多个用户的用户数据,按照预设策略从该多个用户的用户数据中选择数据样本,得到数据样本集,对该数据样本集中的数据样本进行聚类运算,得到多个聚类,分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,基于该待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。
如图1B所示,该用户客群分类方法的具体流程可以如下:
101、采集多个用户的用户数据。
例如,用户客群分类装置具体可以从网络中采集各种类型用户的用户数据,或者,也可以由其他的采集设备对各种类型用户的用户数据进行采集,然后在需要使用该用户数据时,用户客群分类装置从该采集设备中进行获取,等等。
其中,该用户数据可以包括多种用户行为类别的数据,其中,用户行为类别可以根据用户行为特性来进行划分,比如,与社交相关的数据可以划分为社交信息,与支付相关的数据可以划分为支付行为数据,与娱乐休闲相关的数据可以划分为娱乐休闲行为数据,等等;即该用户数据可以包括用户的帐号信息、社交信息、支付行为数据、娱乐休闲行为数据、教育行为数据、和/或其他网络行为数据,等等,例如,可以如下:
1)用户的帐号信息,比如性别、区域、年龄、和/或通用设置等。
2)用户的社交信息,比如聊天、电子邮件、语音通话、微博空间发表的信息、社区评论、和/或公众号文章阅读等。
3)用户的支付行为数据,比如网上购物、游戏点卡支付、和/或音乐流量包等。
4)娱乐休闲行为数据:如视频点播、音乐播放、K歌、和/或新闻阅读等。
5)教育行为数据:如线上读书、公开课学习、职业考试练习、技能培训、和/或翻译软件使用等.
6)其网络行为数据:如终端应用(App,Application)下载、和/或搜索等。
102、按照预设策略从该多个用户的用户数据中选择数据样本,得到数据样本集。
其中,该预设策略可以根据实际应用的需求进行设置,例如,可以设置为“对用户数据进行清洗”。即步骤“按照预设策略从该用户数据中选择数据样本,得到数据样本集”具体可以包括:
将采集到的用户数据进行清洗,得到数据样本集。即此时可以将采集到的用户数据进行清洗,然后,将清洗后的用户数据作为数据样本,添加至数据样本集中。
其中,清洗的方式可以多种,比如“去除奇异点”、“补齐缺失值”、和/或“去掉重复采集的数据”,等等;例如,可以如下:
(1)去除奇异点;
将该用户数据中的奇异点剔除,并将剩余的用户数据作为数据样本,添加至数据样本集中。
其中,该奇异点指的是“由于数据采集过程中的错误或者其他行为造成的数据可能性取值外的数据”。
(2)补齐缺失值;
根据该用户数据确定本应采集但没有采集到数据,对没有采集到的数据赋予预设值(即赋默认值),然后,将采集到的用户数据和该赋值后的数据作为数据样本,添加至数据样本集中。
(3)去掉重复采集的数据;
将该用户数据中重复采集的数据进行剔除,并将剩余的用户数据作为数据样本,添加至数据样本集中。
在将用户数据进行清洗后,便可以将清洗后的数据,即数据样本集中的数据样本进行归类,并写入预设数据库中,以供后续步骤使用。
103、对该数据样本集中的数据样本进行聚类运算,得到多个聚类;例如,具体可以如下:
获取数据样本的用户行为类别,根据该用户行为类别对该数据样本集中的数据样本进行聚类运算,得到多个聚类。
其中,聚类算法也可以根据实际应用的需求而定,比如,可以采用邻近算法(kNN,k-NearestNeighbor)、或k-means算法(一种硬聚类算法)等。
比如,以用户行为类别包括社交信息和娱乐休闲信息为例,则经过聚类运算后,可以将属于社交信息的数据样本组成一个聚类,以及将属于娱乐休闲信息的数据样本组成一个聚类,以此类推,等等。
104、分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集。
例如,具体可以分别计算每个聚类中的数据样本的相似度,根据该相似度对每个聚类中的冗余样本进行剔除,得到待训练样本集,具体可以如下:
(1)分别计算每个聚类中的数据样本的相似度,比如,具体可以包括步骤S1、S2和S3,如下:
S1、利用最大化信息量算法分别计算每个聚类中的数据样本的最大信息量。
其中,最大信息量,指的是数据样本所能够提供的信息的最大含量,同一数据样本的最大信息量和信息量的和为1。比如,若数据样本的最大信息量为0.8,则其信息量为0.2,以此类推,等等。其中,信息量指的是从M个相等可能事件中选出一个事件所需要的信息度量或含量,也就是在辩识M个事件中特定的一个事件的过程中所需要提问“是或否”的最少次数。一个事件的可能性愈小(即出现的概率小),不确定性越多,确定其所需获取的信息越多,即信息量愈大;而事件的可能性愈大(即出现的概率大),不确定性越少,则确定其所需获取的信息愈少,即信息量越小。可见,该信息量的多少是与事件发生的概率大小有关,而最大信息量的多少同样与事件发生的概率大小有关。所以,可以通过计算事件发生的概率来计算该最大信息量,进而确定事件的不确定性的大小。其中,最大信息量越小(即信息量越大),不确定性越大,最大信息 量越大(即信息量越小),不确定性越小。比如,在本申请实施例中,事件发生的概率主要指的是“分类模型预测数据样本为某一类”的概率值,因此,在本申请实施例中,可以通过计算该概率值来得到该最大信息量。即,可选地,步骤“利用最大化信息量算法分别计算每个聚类中的数据样本的最大信息量”可以包括:
采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测,根据预测结果计算相应数据样本的最大信息量。比如,可以直接根据预测结果(如该用户属于某一分类的概率)计算相应数据样本的最大信息量,或者,也可以计算预测结果的置信度,然后,根据该置信度来计算相应数据样本的最大信息量,等等。这里的置信度=最不确定性,最不确定性的计算可以由下文的公式1来计算(不仅限于此公式),即最不确定的样本置信度越低,反之越高。其中,该初始分类模型可以由维护人员进行预设,也可以由该用户客群分类装置通过训练来得到,比如,可以从每个聚类中随机选择一定数量的数据样本,对其所属的客群分类进行人工标注,构成初始训练样本(这些初始训练样本构成初始训练样本集),然后,利用这些初始训练样本对预设的原始模型进行训练,得到初始分类模型;即,可选地,在步骤“采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测”之前,该用户客群分类方法还可以包括:
从每个聚类中随机选择预定数量的数据样本,对其所属的客群分类进行人工标注,根据人工标注后的数据样本对预设的原始模型进行训练,得到初始分类模型。
则此时,步骤“采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测”可以包括:采用得到的初始分类模型,对每个聚类中的其他未进行人工标注的数据样本的客群分类进行预测。
其中,训练时所采用的原始模型可以根据实际应用的需求进行设置,而该预定数量也可以根据实际应用的需求进行设置,比如,以需要选择1000个初始训练样本为例,若总共有5个聚类,则每个聚类可以随机挑选200个数据样本作为相应聚类的初始训练样本,以此类推,等等。
例如,以二元分类(即存在两种分类:分类1和分类2)为例,则其相应的最大信息量计算公式可以如下:
Figure PCTCN2018084565-appb-000001
其中,x为数据样本,L为初始训练样本集(即人工标注后的数据样本),C为一个或一组分类模型(也称为分类器),Pr[x=1|L]为在初始训练样本集L上训练出来的分类模型C对数据样本预测为“分类1”(其中的“1”表示“分类1”)的概率值,
Figure PCTCN2018084565-appb-000002
为使得
Figure PCTCN2018084565-appb-000003
取得最小值的数据样本,argmin为求使得
Figure PCTCN2018084565-appb-000004
取得最小值的数据样本的函数,U为未标注样本集(即该聚类中未标注的数据样本)。
从公式
Figure PCTCN2018084565-appb-000005
可以看出,概率值越接近
Figure PCTCN2018084565-appb-000006
的样本越容易被选中,因为越接近
Figure PCTCN2018084565-appb-000007
表示分类模型对数据样本属于哪一类的预测越不能确定。例如聚类后某一聚类A中包括数据样本x 1、x 2和x 3,用分类模型C算出属于分类1的概率分别为0.49、0.1和0.8,则通过公式
Figure PCTCN2018084565-appb-000008
算出的最大信息量分别为0.01、0.4和0.3,即x 1的不确定性最大,x 3的不确定性次之,x 2的不确定性最小,按照最不确定性原则,三个样本入选待标注样本集的优先级是:x 1,x 3,x 2。依此类推,在此不再赘述。
S2、根据最大信息量从每个聚类中选取不确定性最大的前K个数据样本,得到待标注样本集。
其中,K为正整数,具体取值可以根据实际应用的需求进行设置,比如,以K为2为例,若在步骤S1中,三个样本入选待标注样本集的优先级是:x 1,x 3,x 2,则此时,可以将x 1和x 3添加至待标注样本集中。
S3、计算待标注样本集中两两数据样本之间的相似度,得到每个聚类中的数据样本的相似度。
例如,若该待标注样本集中包括数据样本x 1、x 3、x 5和x 6,则此时,需要计算x 1和x 3的相似度、x 1和x 5的相似度、x 1和x 6的相似度、x 3和x 5的相似度、x 3和x 6 的相似度、以及x 5和x 6的相似度,等等。
需说明的是,相似度的计算可以先将数据样本的用户信息进行量化,转换为向量空间,然后,用标准化欧氏距离、夹角余弦、以及相关系数等计算出相似度。
此外,还需说明的是,在计算出相似度之后,可以根据得到的相似度设置一个阈值,高于该阈值的则认为两个数据样本高度相似,为了描述方便,在申请实施例中,将该大于该预设阈值的相似度称为高度相似度。其中,该阈值可以根据实际应用的需求进行设置,在此不再赘述。
(2)根据该相似度对每个聚类中的冗余样本进行剔除,得到待训练样本集。
例如,具体可以获取相似度大于预设阈值的数据样本,得到候选集,从候选集中筛选最大信息量满足预设条件的数据样本,得到待训练样本集。
其中,该预设条件可以根据实际应用的需求而定,比如,可以通过统计高度相似计数,并将高度相似计数较多的样本进行剔除,等等。即可选地,步骤“从候选集中筛选最大信息量满足预设条件的数据样本,得到待训练样本集”可以包括:
统计该候选集中每个数据样本所具有的高度相似度(大于预设阈值的相似度)的数量,得到高度相似计数,从该候选集中剔除高度相似计数最高的前N个数据样本,得到待训练样本集。
其中,N为正整数,N的取值可以根据实际应用的需求而定。
105、基于该待训练样本集进行分类模型训练。
例如,具体可以获取预设的初始分类模型,对该待训练样本集中的数据样本进行标注,然后,根据标注后的数据样本对该初始分类模型进行训练,得到训练后分类模型。
106、根据训练后分类模型对待分类数据进行客群分类。
例如,可以获取待分类数据,将该待分类数据导入训练后分类模型中,得到输出结果,根据该输出结果确定该待分类数据的客群分类。
其中,客群指的是客户群体,比如可以划分为“学生”、“白领”和“金领”等,又比如,还可以划分为“儿童”、“青年”、“中年”和“老年”,等等,客群的具体划分需求可以根据实际应用的需求而定,在此不再赘述。
由上可知,本申请实施例在从采集到的用户数据中选择数据样本后,可以对这些数据样本进行聚类运算,并分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,然后,才基于该待训练样本集对分类模型进行训练,并根据训练后分类模型对待分类数据进行客群分类;由于该方案在进行分类模型训练之前,先对数据样本进行了聚类,然后,再分别从每个聚类中选择数据样本进行训练,所以,可以保证每个聚类中都有数据样本被选中进入到分类模型的训练样本中,最大程度上使分类模型能从每个聚类中去学习,覆盖样本预期分布,大大提高分类模型的准确率和泛化能力;此外,由于该方案可以对每个聚类中的数据样本进行去冗余操作,以剔除相似度高的冗余数据样本,使得可以尽可能减少标注相似的数据样本,因此,可以大大加快分类模型的收敛;总而言之,该方案不仅可以大大加快分类模型的收敛,而且可以提高分类模型的准确率和泛化能力,有利于提高用户客群分类的效率。
根据前述实施例所描述的方法,以下将举例作进一步详细说明。
在本申请实施例中,将以该用户客群分类装置具体集成在服务器中为例进行说明。
如图2A和图2B所示,一种用户客群分类方法,具体流程可以如下:
201、服务器采集多个用户的用户数据。
例如,服务器具体可以从网络中采集各种类型用户的用户数据,或者,也可以由其他的采集设备对各种类型用户的用户数据进行采集,然后在需要使用该用户数据时,服务器从该采集设备中进行获取,等等。
其中,该用户数据可以包括多种用户行为类别的数据,比如,可以包括用户的帐号信息、社交信息、支付行为数据、娱乐休闲行为数据、教育行为数据、和/或其他网络行为数据等信息。
202、服务器将采集到的用户数据进行清洗,得到数据样本集。
其中,清洗的方式可以多种,比如“去除奇异点”、“补齐缺失值”、和/或“去掉重复采集的数据”,等等;即步骤“服务器将采集到的用户数据进行清洗,得到数据样本集”具体可以如下:
服务器具体可以将该用户数据中的奇异点剔除,并将剩余的用户数据作为数据样本,添加至数据样本集中;
或者,服务器也可以根据该用户数据确定本应采集但没有采集到数据,对没有采集到的数据赋予预设值(即赋默认值),然后,将采集到的用户数据和该赋值后的数据作为数据样本,添加至数据样本集中;
又或者,服务器还可以将该用户数据中重复采集的数据进行剔除,并将剩余的用户数据作为数据样本,添加至数据样本集中,等等。
需说明的是,该数据样本集的数据样本均为未标注所属的客群分类的数据样本。
203、服务器获取数据样本的用户行为类别,根据该用户行为类别对该数据样本集中的数据样本进行聚类运算,得到多个聚类。
例如,服务器具体可以根据该用户行为类别,采用kNN或k-means算法等聚类算法对该数据样本集中的数据样本进行聚类运算,得到多个聚类;为了描述方便,在本申请实施例中,将均以最终聚成5类:聚类A、聚类B、聚类C、聚类D、以及聚类E为例进行说明,比如,具体可参见图2A。
204、服务器从每个聚类中随机选择预定数量的数据样本,对其所属的客群分类进行人工标注,根据人工标注后的数据样本对预设的原始模型进行训练,得到初始分类模型。
其中,训练时所采用的原始模型可以根据实际应用的需求进行设置,而该预定数量也可以根据实际应用的需求进行设置。
例如,如图2A所示,如果总共需要挑选1000个数据样本,则此时,可以分别从聚类A、聚类B、聚类C、聚类D、以及聚类E等5个聚类中随机选择一定数量,如大约200个左右的未标注的数据样本,对其客群分类进行人工标注,构成初始训练样本,然后,利用这些初始训练样本对预设的原始模型进行训练,得到初始分类模型。
需说明的是,为了描述方便,可以将这些初始训练样本所在的集合称为初始训练样本集,该初始训练样本集中可以多个初始训练样本子集,每个初始训练样本子集对应一个聚类,比如,可以包括聚类A对应的初始训练样本子集、聚类B对应的初始训练样本子集、聚类C对应的初始训练样本子集、聚类D对应的初始训练样本子集、以及聚类E对应的初始训练样本子集,等等。其中,聚类A对应的初始训练样本子集可以包括聚类A对应的已标注的数据样本,聚类B对应的初始训练样本子集可以包括聚类B对应的已标注的数据样本,聚类C对应 的初始训练样本子集可以包括聚类C对应的已标注的数据样本,聚类D对应的初始训练样本子集可以包括聚类D对应的已标注的数据样本,而聚类E对应的初始训练样本子集可以包括聚类E对应的已标注的数据样本,等等,具体可参见图2A,在此不再赘述。
205、服务器采用步骤204中得到的初始分类模型对每个聚类中的数据样本的客群分类进行预测,根据预测结果计算相应数据样本的最大信息量。
例如,参见图2A,还是以最终聚成聚类A、聚类B、聚类C、聚类D、以及聚类E为例进行说明,则此时,服务器可以采用初始分类模型分别对聚类A、聚类B、聚类C、聚类D、以及聚类E中的数据样本的客群分类进行预测,然后,根据预测结果分别计算相应数据样本的最大信息量。
比如,以计算聚类A中的数据样本的最大信息量,且用户存在两种分类“学生”和“白领”为例,若聚类A中包括数据样本x 1、x 2和x 3,则此时,服务器可以采用初始分类模型对数据样本x 1、x 2和x 3的客群分类进行预测,若预测数据样本x 1、x 2和x 3均可能属于“学生”,且该属于“学生”的概率分别为0.49、0.1和0.8,则可以通过如下公式
Figure PCTCN2018084565-appb-000009
分别算出x 1、x 2和x 3的最大信息量:0.01、0.4和0.3。
其中,L为初始训练样本集(即人工标注后的数据样本),C为一个或一组分类模型,Pr[x=1|L]为在初始训练样本集L上训练出来的分类模型C对数据样本预测为“学生”的概率值,
Figure PCTCN2018084565-appb-000010
为使得
Figure PCTCN2018084565-appb-000011
取得最小值的数据样本,U为未标注样本集(即该聚类中未标注的数据样本)。
同理,对于聚类B、C、D和E中的数据样本的最大信息量,也可以采用上述方式进行计算,在此不再赘述。
206、服务器对每个聚类进行筛选,得到待标注样本集。
其中,该待标注样本集可以包括多个待标注样本子集,每个待标注样本子集对应一个聚类,比如,可以包括聚类A对应的待标注样本子集、聚类B对应的待标注样本子集、聚类C对应的待标注样本子集、聚类D对应的待标注样本子集、以及聚类E对应的待标注样本子集,等等。其中,A对应的待标注样本子集 称为聚类A',B对应的待标注样本子集称为聚类B',C对应的待标注样本子集称为聚类C',D对应的待标注样本子集称为聚类D',E对应的待标注样本子集称为聚类E',具体可参见图2A。
其中,筛选的方式可以有多种,例如,服务器具体可以根据最大信息量从每个聚类中选取不确定性最大的前K个数据样本,得到待标注样本集。
其中,K为正整数,具体取值可以根据实际应用的需求进行设置。
比如,以K为2为例,若在步骤205中,聚类A中有三个数据样本的最大信息量分别为:0.01、0.4和0.3,则按照不确定性从大到小进行排序,这三个数据样本的次序依次为:x 1,x 3,x 2,由于K为2,因此,此时可以将确定性最大的前两个数据样本,即x 1和x 3添加至待标注样本集中的聚类A'中(即聚类A对应的待标注样本子集中)。
同理,可以采用同样的方式从聚类B、C、D和E中选取最大信息量最大的前K个数据样本,分别添加至待标注样本集中的聚类B'、C'、D'和E'中。
207、服务器计算待标注样本集中两两数据样本之间的相似度,得到每个聚类中的数据样本的相似度。
例如,还是以聚类A为例,若该待标注样本集中(即该待标注样本集的聚类A对应的待标注样本子集中)包括数据样本x 1、x 3、x 5、x 6和x 7,则此时,需要计算x 1和x 3的相似度、x 1和x 5的相似度、x 1和x 6的相似度、x 1和x 7的相似度、x 3和x 5的相似度、x 3和x 6的相似度、x 3和x 7的相似度、x 5和x 6、x 5和x 7、以及x 6和x 7的相似度,即此时:
这个聚类中数据样本x 1的相似度为:x 1和x 3的相似度、x 1和x 5的相似度、x 1和x 6的相似度、以及x 1和x 7的相似度;
聚类A中数据样本x 3的相似度为:x 1和x 3的相似度、x 3和x 5的相似度、x 3和x 6的相似度、以及x 3和x 7的相似度;
聚类A中数据样本x 5的相似度为:x 1和x 5的相似度、x 3和x 5的相似度、x 5和x 6的相似度、以及x 5和x 7的相似度;
聚类A中数据样本x 6的相似度为:x 1和x 6的相似度、x 3和x 6的相似度、x 5和x 6的相似度、以及x 5和x 7的相似度;
聚类A中数据样本x 7的相似度为:x 1和x 7的相似度、x 3和x 7的相似度、x 5和x 7的相似度、以及x 6和x 7的相似度。
需说明的是,在计算相似度时,可以先将数据样本的用户信息进行量化,转换为向量空间,然后,再用标准化欧氏距离、夹角余弦、以及相关系数等计算出相似度。
比如,参见表一,每个聚类中的数据样本的相似度可以如下:
表一:
数据样本 x 1 x 3 x 5 x 6 x 7
x 1 1 0.8 0.2 0.3 0.4
x 3 0.8 1 0.6 0.7 0.8
x 5 0.2 0.6 1 0.9 0.5
x 6 0.3 0.7 0.9 1 0.3
x 7 0.4 0.8 0.5 0.3 1
需说明的是,在计算出相似度之后,可以根据得到的相似度设置一个阈值,高于该阈值的则认为两个数据样本高度相似,称为高度相似度。其中,该阈值可以根据实际应用的需求进行设置,比如,该阈值可以设置为0.6,等等。
208、服务器根据该相似度对每个聚类中的冗余样本进行剔除(即进行去冗余操作),得到待训练样本集。
其中,该待训练样本集可以包括多个待训练样本子集,每个待训练样本子集对应一个聚类,比如,可以包括聚类A对应的待训练样本子集、聚类B对应的待训练样本子集、聚类C对应的待训练样本子集、聚类D对应的待训练样本子集、以及聚类E对应的待训练样本子集,等等。其中,聚类A对应的待训练样本子集称为聚类A″,聚类B对应的待训练样本子集称为聚类B″,聚类C对应的待训练样本子集称为聚类C″,聚类D对应的待训练样本子集称为聚类D″,聚类E对应的待训练样本子集称为聚类E″,具体可参见图2A。
其中,去冗余的方式可以由多种,例如,服务器具体可以获取相似度大于预设阈值的数据样本,得到候选集,然后,统计该候选集中每个数据样本所具有的高度相似度的数量,得到高度相似计数,并从该候选集中剔除高度相似计数最高的前N个数据样本,得到待训练样本集。
其中,N为正整数,N的取值可以根据实际应用的需求而定。
比如,以值大于0.6的相似度称为高度相似度为例,则根据表一可知,数据样本x 1的高度相似计数为1,数据样本x 3的高度相似计数为3,数据样本x 5的高度相似计数为1,数据样本x 6的高度相似计数为2,数据样本x 7的高度相似计数为1,可见,其中数据样本x 3的高度相似计数最高,若N为1,则此时,需要从该候选集中剔除数据样本x 3,以得到待训练样本集,即此时待训练样本集可以包括数据样本x 1、x 5、x 6和x 7
209、服务器基于该待训练样本集进行分类模型训练。
例如,参见图2A,服务器具体可以获取预设的初始分类模型,对该待训练样本集中的数据样本进行标注,然后,根据标注后的数据样本对该初始分类模型进行训练,得到训练后分类模型。
其中,标注时可以采用人工标注,也可以由该服务器按照预设策略进行标注,在此不再赘述。
需说明的是,原始模型、初始分类模型、以及训练后分类模型分别为同一“分类模型”在不同训练期间的表现形式,因此,为了描述方便,在图2A中,将原始模型、初始分类模型、以及训练后分类模型均表示为“分类模型”,应当理解的是,在不同的时间,该分类模型的表现形式可以有所不同。
210、服务器根据训练后分类模型对待分类数据进行客群分类。
例如,服务器可以获取待分类数据,将该待分类数据导入训练后分类模型中,以进行客群分类预测,然后,根据训练后分类模型输出的结果确定该待分类数据的客群分类,得到分类后数据,具体可参见图2A。
比如,还是以用户存在两种分类“学生”和“白领”为例,则此时,可以将该待分类数据,比如待分类用户的帐号信息、社交信息、支付行为数据、娱乐休闲行为数据、教育行为数据、和/或其他网络行为数据等信息导入训练后分类模型中,然后,根据训练后分类模型输出的结果确定该待分类数据是“学生”对应的数据还是“白领”对应的数据,从而将相应的用户归类到相应的客群,比如“学生”或“白领”中。
由上可知,本申请实施例在从采集到的用户数据中选择数据样本后,可以对这些数据样本进行聚类运算,并分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,然后,才基于该待训练样本集对分类模型进行训练,并根据训练后分类模型对待分类数据进行客群分类;由于该方案在进行分类模 型训练之前,先对数据样本进行了聚类,然后,再分别从每个聚类中选择数据样本进行训练,所以,可以保证每个聚类中都有数据样本被选中进入到分类模型的训练样本中,最大程度上使分类模型能从每个聚类中去学习,覆盖样本预期分布,大大提高分类模型的准确率和泛化能力;此外,由于该方案可以对每个聚类中的数据样本进行去冗余操作,以剔除相似度高的冗余数据样本,使得可以尽可能减少标注相似的数据样本,因此,可以大大加快分类模型的收敛;总而言之,该方案不仅可以大大加快分类模型的收敛,而且可以提高分类模型的准确率和泛化能力,有利于提高用户客群分类的效率。
为了更好地实施以上方法,本申请实施例还提供一种用户客群分类装置,该用户客群分类装置具体可以集成在服务器,比如业务服务器等设备中。
所述用户客群分类装置可以包括:处理器和存储器,所述存储器上存储有计算机可读指令,所述计算机可读指令由所述处理器执行以完成以下操作:采集多个用户的用户数据,所述用户数据包括多种用户行为类别的数据;按照预设策略从所述多个用户的用户数据中选择数据样本,得到数据样本集;对所述数据样本集中的数据样本进行聚类运算,得到多个聚类;分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集;基于所述待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。
其中,所述对所述数据样本集中的数据样本进行聚类运算,得到多个聚类,包括:获取数据样本的用户行为类别;根据所述用户行为类别对所述数据样本集中的数据样本进行聚类运算,得到多个聚类。
其中,所述分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,包括:分别计算每个聚类中的数据样本的相似度;根据所述相似度对每个聚类中的冗余样本进行剔除,得到待训练样本集。
其中,所述分别计算每个聚类中的数据样本的相似度,包括:利用最大化信息量算法分别计算每个聚类中的数据样本的最大信息量;根据最大信息量从每个聚类中选取不确定性最大的前K个数据样本,得到待标注样本集,其中,K为正整数;计算待标注样本集中两两数据样本之间的相似度,得到每个聚类中的数据样本的相似度。
其中,所述利用最大化信息量算法分别计算每个聚类中的数据样本的最大信息量,包括:采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测;根据预测结果计算相应数据样本的最大信息量。
其中,所述采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测之前,还包括:从每个聚类中随机选择预定数量的数据样本进行人工标注;根据人工标注后的数据样本对预设的原始模型进行训练,得到初始分类模型;所述采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测,包括:采用得到的初始分类模型,对每个聚类中的其他未进行人工标注的数据样本的客群分类进行预测。
其中,所述根据所述相似度对每类数据样本中的冗余样本进行剔除,得到待训练样本集,包括:获取相似度大于预设阈值的数据样本,得到候选集;从候选集中筛选最大信息量满足预设条件的数据样本,得到待训练样本集。
所述指令例如为软件形式的指令模块,具体描述如下:
如图3A所示,该用户客群分类装置包括采集单元301、选择单元302、聚类单元303、去冗余单元304、以及分类单元305,如下:
(1)采集单元301;
采集单元301,用于采集多个用户的用户数据。
例如,该采集单元301可以从网络中采集各种类型用户的用户数据,或者,也可以由其他的采集设备对各种类型用户的用户数据进行采集,然后在需要使用该用户数据时,由采集单元301从该采集设备中进行获取,等等。
其中,该用户数据可以包括多种用户行为类别的数据,比如,可以包括用户的帐号信息、社交信息、支付行为数据、娱乐休闲行为数据、教育行为数据、和/或其他网络行为数据等信息。
(2)选择单元302;
选择单元302,用于按照预设策略从该用户数据中选择数据样本,得到数据样本集。
其中,该预设策略可以根据实际应用的需求进行设置,例如,可以设置为“对用户数据进行清洗”。其中,清洗的方式可以多种,比如“去除奇异点”、“补齐缺失值”、和/或“去掉重复采集的数据”,等等。
(3)聚类单元303;
聚类单元303,用于对该数据样本集中的数据样本进行聚类运算,得到多个聚类。
例如,该聚类单元303,具体可以用于获取数据样本的用户行为类别,根据该用户行为类别对该数据样本集中的数据样本进行聚类运算,得到多个聚类。
其中,该聚类算法也可以根据实际应用的需求而定,比如,可以采用kNN或k-means算法等。
(4)去冗余单元304;
去冗余单元304,用于分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集。
例如,该去冗余单元304可以包括计算子单元和筛选子单元,如下:
该计算子单元,用于分别计算每个聚类中的数据样本的相似度。
该筛选子单元,用于根据该相似度对每个聚类中的冗余样本进行剔除,得到待训练样本集。
比如,该计算子单元,具体可以用于利用最大化信息量算法分别计算每个聚类中的数据样本的最大信息量,根据该最大信息量从每个聚类中选取不确定性最大的前K个数据样本,得到待标注样本集,计算待标注样本集中两两数据样本之间的相似度,得到每个聚类中的数据样本的相似度;其中,K为正整数。
该筛选子单元,具体可以用于获取相似度大于预设阈值的数据样本,得到候选集,从候选集中筛选最大信息量满足预设条件的数据样本,得到待训练样本集;比如,可以统计该候选集中每个数据样本所具有的高度相似度的数量,得到高度相似计数,从该候选集中剔除高度相似计数最高的前N个数据样本,得到待训练样本集,等等。
其中,计算该最大信息量的方式可以有多种,比如,可以如下:
该计算子单元,具体可以用于采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测,根据预测结果计算相应数据样本的最大信息量;比如,可以直接根据预测结果(如该用户属于某一分类的概率)计算相应数据样本的最大信息量,或者,也可以计算预测结果的置信度,然后,根据该置信度来计算相应数据样本的最大信息量,等等。
其中,该初始分类模型可以由维护人员进行预设,也可以由该用户客群分类装置通过训练来得到,比如,可以从每个聚类中随机选择一定数量的数据样本进行人工标注,构成初始训练样本,然后,利用这些初始训练样本对预设的原始模型进行训练,得到初始分类模型;即,可选地,如图3B所示,该用户客群分类装置还可以包括模型建立单元306,如下:
模型建立单元306,可以用于从每个聚类中随机选择预定数量的数据样本进行人工标注,根据人工标注后的数据样本对预设的原始模型进行训练,得到初始分类模型。
则此时,计算子单元,具体可以用于采用模型建立单元得到的初始分类模型,对每个聚类中的其他未进行人工标注的数据样本的客群分类进行预测。
其中,训练时所采用的原始模型可以根据实际应用的需求进行设置,而该预定数量也可以根据实际应用的需求进行设置,在此不再赘述。
(5)分类单元305;
分类单元305,用于基于该待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。
例如,该分类单元305,具体用于获取预设的初始分类模型,对该待训练样本集中的数据样本进行标注,根据标注后的数据样本对该初始分类模型进行训练,得到训练后分类模型,获取待分类数据,将该待分类数据导入训练后分类模型中,得到输出结果,根据该输出结果确定该待分类数据的客群分类。
具体实施时,以上各个单元可以作为独立的实体来实现,也可以进行任意组合,作为同一或若干个实体来实现,以上各个单元的具体实施可参见前面的实施例,在此不再赘述。
由上可知,本申请实施例的用户客群分类装置在从采集到的用户数据中选择数据样本后,可以由聚类单元303对这些数据样本进行聚类运算,并由去冗余单元304分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,然后,才由分类单元305基于该待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类;由于该方案在进行分类模型训练之前,先对数据样本进行了聚类,然后,再分别从每个聚类中选择数据样本进行训练,所以,可以保证每个聚类中都有数据样本被选中进入到分类模型的训练样本中,最大程度上使分类模型能从每个聚类中去学习,覆盖样本预期分 布,大大提高分类模型的准确率和泛化能力;此外,由于该方案可以对每个聚类中的数据样本进行去冗余操作,以剔除相似度高的冗余数据样本,使得可以尽可能减少标注相似的数据样本,因此,可以大大加快分类模型的收敛;总而言之,该方案不仅可以大大加快分类模型的收敛,而且可以提高分类模型的准确率和泛化能力,有利于提高用户客群分类的效率。
本申请实施例还提供一种服务器,如图4所示,其示出了本申请实施例所涉及的服务器的结构示意图,具体来讲:
该服务器可以包括一个或者一个以上处理核心的处理器401、一个或一个以上存储器402、电源403和输入单元404等部件。存储器402为计算机可读存储介质。图4中示出的服务器结构并不构成对服务器的限定,可以包括比图示更多或更少的部件,或者组合某些部件,或者不同的部件布置。其中:
处理器401是该服务器的控制中心,利用各种接口和线路连接整个服务器的各个部分,通过运行或执行存储在存储器402内的软件程序和/或模块,以及调用存储在存储器402内的数据,执行服务器的各种功能和处理数据,从而对服务器进行整体监控。可选地,处理器401可包括一个或多个处理核心;优选的,处理器401可集成应用处理器和调制解调处理器,其中,应用处理器主要处理操作系统、用户界面和应用程序等,调制解调处理器主要处理无线通信。可以理解的是,上述调制解调处理器也可以不集成到处理器401中。
存储器402可用于存储软件程序以及模块,处理器401通过运行存储在存储器402的软件程序以及模块,从而执行各种功能应用以及数据处理。存储器402可主要包括存储程序区和存储数据区,其中,存储程序区可存储操作系统、至少一个功能所需的应用程序(比如声音播放功能、图像播放功能等)等;存储数据区可存储根据服务器的使用所创建的数据等。此外,存储器402可以包括高速随机存取存储器,还可以包括非易失性存储器,例如至少一个磁盘存储器件、闪存器件、或其他易失性固态存储器件。相应地,存储器402还可以包括存储器控制器,以提供处理器401对存储器402的访问。
服务器还包括给各个部件供电的电源403,优选的,电源403可以通过电源管理系统与处理器401逻辑相连,从而通过电源管理系统实现管理充电、放电、以及功耗管理等功能。电源403还可以包括一个或一个以上的直流或交流 电源、再充电系统、电源故障检测电路、电源转换器或者逆变器、电源状态指示器等任意组件。
该服务器还可包括输入单元404,该输入单元404可用于接收输入的数字或字符信息,以及产生与用户设置以及功能控制有关的键盘、鼠标、操作杆、光学或者轨迹球信号输入。
尽管未示出,服务器还可以包括显示单元等,在此不再赘述。具体在本申请实施例中,服务器中的处理器401会按照如下的指令,将一个或一个以上的应用程序的进程对应的可执行文件加载到存储器402中,并由处理器401来运行存储在存储器402中的应用程序,从而实现各种功能,如下:
采集多个用户的用户数据,该用户数据包括多种用户行为类别的数据,按照预设策略从该用户数据中选择数据样本,得到数据样本集,对该数据样本集中的数据样本进行聚类运算,得到多个聚类,分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,基于该待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。
其中,去冗余的方式可以有多种,例如,可以分别计算每个聚类中的数据样本的相似度,根据该相似度对每个聚类中的冗余样本进行剔除,得到待训练样本集。
比如,可以利用最大化信息量算法分别计算每个聚类中的数据样本的最大信息量,根据最大信息量从每个聚类中选取不确定性最大的前K个数据样本,得到待标注样本集,计算待标注样本集中两两数据样本之间的相似度,得到每个聚类中的数据样本的相似度,获取相似度大于预设阈值的数据样本,得到候选集,统计该候选集中每个数据样本所具有的高度相似度的数量,得到高度相似计数,从该候选集中剔除高度相似计数最高的前N个数据样本,得到待训练样本集,等等。
以上各个操作的具体实施可参见前面的实施例,在此不再赘述。
由上可知,本申请实施例的服务器在从采集到的用户数据中选择数据样本后,可以对这些数据样本进行聚类运算,并分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,然后,才基于该待训练样本集对分类模型进行训练,并根据训练后分类模型对待分类数据进行客群分类;由于该方案在进行分类模型训练之前,先对数据样本进行了聚类,然后,再分别从每个聚类中 选择数据样本进行训练,所以,可以保证每个聚类中都有数据样本被选中进入到分类模型的训练样本中,最大程度上使分类模型能从每个聚类中去学习,覆盖样本预期分布,大大提高分类模型的准确率和泛化能力;此外,由于该方案可以对每个聚类中的数据样本进行去冗余操作,以剔除相似度高的冗余数据样本,使得可以尽可能减少标注相似的数据样本,因此,可以大大加快分类模型的收敛;总而言之,该方案不仅可以大大加快分类模型的收敛,而且可以提高分类模型的准确率和泛化能力,有利于提高用户客群分类的效率。
上述实施例的各种方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成,该程序可以存储于一计算机可读存储介质中,存储介质可以包括:只读存储器(ROM,Read Only Memory)、随机存取记忆体(RAM,Random Access Memory)、磁盘或光盘等。
以上对本申请实施例所提供的一种用户客群分类方法和装置进行了详细介绍,本文中应用了具体个例对本申请的原理及实施方式进行了阐述,以上实施例的说明只是用于帮助理解本申请的方法及其核心思想;同时,对于本领域的技术人员,依据本申请的思想,在具体实施方式及应用范围上均会有改变之处,综上所述,本说明书内容不应理解为对本申请的限制。

Claims (26)

  1. 一种用户客群分类方法,包括:
    采集多个用户的用户数据,所述用户数据包括多种用户行为类别的数据;
    按照预设策略从所述多个用户的用户数据中选择数据样本,得到数据样本集;
    对所述数据样本集中的数据样本进行聚类运算,得到多个聚类;
    分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集;
    基于所述待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。
  2. 根据权利要求1所述的方法,其中,所述对所述数据样本集中的数据样本进行聚类运算,得到多个聚类,包括:
    获取数据样本的用户行为类别;
    根据所述用户行为类别对所述数据样本集中的数据样本进行聚类运算,得到多个聚类。
  3. 根据权利要求1所述的方法,其中,所述分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,包括:
    分别计算每个聚类中的数据样本的相似度;
    根据所述相似度对每个聚类中的冗余样本进行剔除,得到待训练样本集。
  4. 根据权利要求3所述的方法,其中,所述分别计算每个聚类中的数据样本的相似度,包括:
    利用最大化信息量算法分别计算每个聚类中的数据样本的最大信息量;
    根据最大信息量从每个聚类中选取不确定性最大的前K个数据样本,得到待标注样本集,其中,K为正整数;
    计算待标注样本集中两两数据样本之间的相似度,得到每个聚类中的数据样本的相似度。
  5. 根据权利要求4所述的方法,其中,所述利用最大化信息量算法分别计算每个聚类中的数据样本的最大信息量,包括:
    采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测;
    根据预测结果计算相应数据样本的最大信息量。
  6. 根据权利要求5所述的方法,其中,所述采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测之前,还包括:
    从每个聚类中随机选择预定数量的数据样本进行人工标注;
    根据人工标注后的数据样本对预设的原始模型进行训练,得到初始分类模型;
    所述采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测,包括:采用得到的初始分类模型,对每个聚类中的其他未进行人工标注的数据样本的客群分类进行预测。
  7. 根据权利要求3所述的方法,其中,所述根据所述相似度对每类数据样本中的冗余样本进行剔除,得到待训练样本集,包括:
    获取相似度大于预设阈值的数据样本,得到候选集;
    从候选集中筛选最大信息量满足预设条件的数据样本,得到待训练样本集。
  8. 根据权利要求7所述的方法,其中,所述从候选集中筛选最大信息量满足预设条件的数据样本,得到待训练样本集,包括:
    统计所述候选集中每个数据样本所具有的高度相似度的数量,得到高度相似计数,所述高度相似度为大于预设阈值的相似度;
    从所述候选集中剔除高度相似计数最高的前N个数据样本,得到待训练样本集,其中,N为正整数。
  9. 根据权利要求1至8任一项所述的方法,其中,所述基于所述待训练样本集进行分类模型训练,包括:
    获取预设的初始分类模型;
    对所述待训练样本集中的数据样本进行标注;
    根据标注后的数据样本对所述初始分类模型进行训练,得到训练后分类模型。
  10. 根据权利要求1至8任一项所述的方法,其中,所述根据训练后分类模型对待分类数据进行客群分类,包括:
    获取待分类数据;
    将所述待分类数据导入训练后分类模型中,得到输出结果;
    根据所述输出结果确定所述待分类数据的客群分类。
  11. 一种用户客群分类装置,包括:处理器和存储器,所述存储器上存储有计算机可读指令,所述计算机可读指令由所述处理器执行以完成以下操作:
    采集多个用户的用户数据,所述用户数据包括多种用户行为类别的数据;
    按照预设策略从所述多个用户的用户数据中选择数据样本,得到数据样本集;
    对所述数据样本集中的数据样本进行聚类运算,得到多个聚类;
    分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集;
    基于所述待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。
  12. 根据权利要求11所述的装置,其中,所述对所述数据样本集中的数据样本进行聚类运算,得到多个聚类,包括:
    获取数据样本的用户行为类别;
    根据所述用户行为类别对所述数据样本集中的数据样本进行聚类运算,得到多个聚类。
  13. 根据权利要求11所述的装置,其中,所述分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,包括:
    分别计算每个聚类中的数据样本的相似度;
    根据所述相似度对每个聚类中的冗余样本进行剔除,得到待训练样本集。
  14. 根据权利要求13所述的装置,其中,所述分别计算每个聚类中的数据样本的相似度,包括:
    利用最大化信息量算法分别计算每个聚类中的数据样本的最大信息量;
    根据最大信息量从每个聚类中选取不确定性最大的前K个数据样本,得到待标注样本集,其中,K为正整数;
    计算待标注样本集中两两数据样本之间的相似度,得到每个聚类中的数据样本的相似度。
  15. 根据权利要求13所述的装置,其中,所述根据所述相似度对每类数据样本中的冗余样本进行剔除,得到待训练样本集,包括:
    获取相似度大于预设阈值的数据样本,得到候选集;
    从候选集中筛选最大信息量满足预设条件的数据样本,得到待训练样本集。
  16. 一种用户客群分类方法,用于服务器,所述服务器包括:处理器和存储器,所述方法包括:
    采集多个用户的用户数据,所述用户数据包括多种用户行为类别的数据;
    按照预设策略从所述多个用户的用户数据中选择数据样本,得到数据样本集;
    对所述数据样本集中的数据样本进行聚类运算,得到多个聚类;
    分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集;
    基于所述待训练样本集进行分类模型训练,并根据训练后分类模型对待分类数据进行客群分类。
  17. 根据权利要求16所述的方法,其中,所述对所述数据样本集中的数据样本进行聚类运算,得到多个聚类,包括:
    获取数据样本的用户行为类别;
    根据所述用户行为类别对所述数据样本集中的数据样本进行聚类运算,得到多个聚类。
  18. 根据权利要求16所述的方法,其中,所述分别对每个聚类中的数据样本进行去冗余操作,得到待训练样本集,包括:
    分别计算每个聚类中的数据样本的相似度;
    根据所述相似度对每个聚类中的冗余样本进行剔除,得到待训练样本集。
  19. 根据权利要求18所述的方法,其中,所述分别计算每个聚类中的数据样本的相似度,包括:
    利用最大化信息量算法分别计算每个聚类中的数据样本的最大信息量;
    根据最大信息量从每个聚类中选取不确定性最大的前K个数据样本,得到待标注样本集,其中,K为正整数;
    计算待标注样本集中两两数据样本之间的相似度,得到每个聚类中的数据样本的相似度。
  20. 根据权利要求19所述的方法,其中,所述利用最大化信息量算法分别计算每个聚类中的数据样本的最大信息量,包括:
    采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测;
    根据预测结果计算相应数据样本的最大信息量。
  21. 根据权利要求20所述的方法,其中,所述采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测之前,还包括:
    从每个聚类中随机选择预定数量的数据样本进行人工标注;
    根据人工标注后的数据样本对预设的原始模型进行训练,得到初始分类模型;
    所述采用预设的初始分类模型对每个聚类中的数据样本的客群分类进行预测,包括:采用得到的初始分类模型,对每个聚类中的其他未进行人工标注的数据样本的客群分类进行预测。
  22. 根据权利要求18所述的方法,其中,所述根据所述相似度对每类数据样本中的冗余样本进行剔除,得到待训练样本集,包括:
    获取相似度大于预设阈值的数据样本,得到候选集;
    从候选集中筛选最大信息量满足预设条件的数据样本,得到待训练样本集。
  23. 根据权利要求22所述的方法,其中,所述从候选集中筛选最大信息量满足预设条件的数据样本,得到待训练样本集,包括:
    统计所述候选集中每个数据样本所具有的高度相似度的数量,得到高度相似计数,所述高度相似度为大于预设阈值的相似度;
    从所述候选集中剔除高度相似计数最高的前N个数据样本,得到待训练样本集,其中,N为正整数。
  24. 根据权利要求16至23任一项所述的方法,其中,所述基于所述待训练样本集进行分类模型训练,包括:
    获取预设的初始分类模型;
    对所述待训练样本集中的数据样本进行标注;
    根据标注后的数据样本对所述初始分类模型进行训练,得到训练后分类模型。
  25. 根据权利要求16至23任一项所述的方法,其中,所述根据训练后分类模型对待分类数据进行客群分类,包括:
    获取待分类数据;
    将所述待分类数据导入训练后分类模型中,得到输出结果;
    根据所述输出结果确定所述待分类数据的客群分类。
  26. 一种计算机可读存储介质,其中,所述计算机可读存储介质中存储有指令,当其在计算机上运行时,使得计算机执行上述权利要求1-10中任一项用户客群分类方法。
PCT/CN2018/084565 2017-04-28 2018-04-26 用户客群分类方法和装置 Ceased WO2018196798A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710293870.6A CN108304427B (zh) 2017-04-28 2017-04-28 一种用户客群分类方法和装置
CN201710293870.6 2017-04-28

Publications (1)

Publication Number Publication Date
WO2018196798A1 true WO2018196798A1 (zh) 2018-11-01

Family

ID=62872172

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/084565 Ceased WO2018196798A1 (zh) 2017-04-28 2018-04-26 用户客群分类方法和装置

Country Status (2)

Country Link
CN (1) CN108304427B (zh)
WO (1) WO2018196798A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113743435A (zh) * 2020-06-22 2021-12-03 北京沃东天骏信息技术有限公司 业务数据分类模型的训练、业务数据的分类方法和装置
CN118447546A (zh) * 2024-04-17 2024-08-06 科大讯飞股份有限公司 一种认知评估方法、装置、设备及存储介质

Families Citing this family (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110858313B (zh) * 2018-08-24 2023-01-31 国信优易数据股份有限公司 一种人群分类方法以及人群分类系统
CN110880006B (zh) * 2018-09-05 2024-05-14 广州视源电子科技股份有限公司 用户分类方法、装置、计算机设备和存储介质
CN109409672A (zh) * 2018-09-25 2019-03-01 深圳市元征科技股份有限公司 一种汽车维修技师分类评级建模方法及装置
CN109213856A (zh) * 2018-10-22 2019-01-15 广东小天才科技有限公司 一种语义识别方法及系统
CN109816043B (zh) * 2019-02-02 2021-01-01 拉扎斯网络科技(上海)有限公司 用户识别模型的确定方法、装置、电子设备及存储介质
CN111767735B (zh) * 2019-03-26 2024-08-16 北京京东尚科信息技术有限公司 执行任务的方法、装置及计算机可读存储介质
CN110110080A (zh) * 2019-03-29 2019-08-09 平安科技(深圳)有限公司 文本分类模型训练方法、装置、计算机设备及存储介质
CN110188197B (zh) * 2019-05-13 2021-09-28 北京一览群智数据科技有限责任公司 一种用于标注平台的主动学习方法及装置
CN114039794B (zh) * 2019-12-11 2024-12-03 支付宝(杭州)信息技术有限公司 基于半监督学习的异常流量检测模型训练方法及装置
CN111597336B (zh) * 2020-05-14 2023-12-22 腾讯科技(深圳)有限公司 训练文本的处理方法、装置、电子设备及可读存储介质
US11494700B2 (en) * 2020-09-16 2022-11-08 International Business Machines Corporation Semantic learning in a federated learning system
CN111967541B (zh) * 2020-10-21 2021-01-05 上海冰鉴信息科技有限公司 基于多平台样本的数据分类方法及装置
CN112507093B (zh) * 2020-12-10 2023-08-11 亿企赢网络科技有限公司 一种语料混淆处理方法、装置及介质
US11609273B2 (en) * 2020-12-14 2023-03-21 Guangzhou Automobile Group Co., Ltd. Method and system for optimizing BMS model, storage medium and electric vehicle
CN112541595B (zh) * 2020-12-30 2024-12-06 中国建设银行股份有限公司 模型构建方法及装置、存储介质及电子设备
CN114741500B (zh) * 2021-01-07 2025-07-29 中国科学院软件研究所 一种基于深度主动学习的用户意图识别方法及系统
CN113176905B (zh) * 2021-04-26 2021-10-26 北京麟卓信息科技有限公司 一种基于机器学习的通用ide性能优化方法及系统
CN113127605B (zh) * 2021-06-17 2021-11-02 明品云(北京)数据科技有限公司 一种目标识别模型的建立方法、系统、电子设备及介质
CN113487320A (zh) * 2021-06-28 2021-10-08 深圳索信达数据技术有限公司 欺诈交易检测方法、装置、计算机设备和存储介质
CN114330578A (zh) * 2021-12-31 2022-04-12 北京字跳网络技术有限公司 对象分类方法、装置、可读介质及电子设备
CN117520836A (zh) * 2022-07-29 2024-02-06 上海智臻智能网络科技股份有限公司 训练样本的生成方法、装置、设备和存储介质
CN116610952A (zh) * 2023-05-16 2023-08-18 湖南视比特机器人有限公司 目标检测模型的优化方法及计算机存储介质和终端设备

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101211339A (zh) * 2006-12-29 2008-07-02 上海芯盛电子科技有限公司 基于用户行为的智能网页分类器
CN101980202A (zh) * 2010-11-04 2011-02-23 西安电子科技大学 不平衡数据的半监督分类方法
CN103617429A (zh) * 2013-12-16 2014-03-05 苏州大学 一种主动学习分类方法和系统
CN103885977A (zh) * 2012-12-21 2014-06-25 腾讯科技(深圳)有限公司 一种网页数据的分类方法、装置和系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101211339A (zh) * 2006-12-29 2008-07-02 上海芯盛电子科技有限公司 基于用户行为的智能网页分类器
CN101980202A (zh) * 2010-11-04 2011-02-23 西安电子科技大学 不平衡数据的半监督分类方法
CN103885977A (zh) * 2012-12-21 2014-06-25 腾讯科技(深圳)有限公司 一种网页数据的分类方法、装置和系统
CN103617429A (zh) * 2013-12-16 2014-03-05 苏州大学 一种主动学习分类方法和系统

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113743435A (zh) * 2020-06-22 2021-12-03 北京沃东天骏信息技术有限公司 业务数据分类模型的训练、业务数据的分类方法和装置
CN113743435B (zh) * 2020-06-22 2024-10-22 北京沃东天骏信息技术有限公司 业务数据分类模型的训练、业务数据的分类方法和装置
CN118447546A (zh) * 2024-04-17 2024-08-06 科大讯飞股份有限公司 一种认知评估方法、装置、设备及存储介质

Also Published As

Publication number Publication date
CN108304427B (zh) 2020-03-17
CN108304427A (zh) 2018-07-20

Similar Documents

Publication Publication Date Title
WO2018196798A1 (zh) 用户客群分类方法和装置
CN113569895B (zh) 图像处理模型训练方法、处理方法、装置、设备及介质
CN107766929B (zh) 模型分析方法及装置
CN111898675B (zh) 信贷风控模型生成方法、装置、评分卡生成方法、机器可读介质及设备
WO2022222942A1 (zh) 问答记录生成方法、装置、电子设备及存储介质
CN111754345A (zh) 一种基于改进随机森林的比特币地址分类方法
CN105069534A (zh) 客户流失预测方法以及装置
US12499374B2 (en) Extracting and classifying entities from digital content items
CN111931809A (zh) 数据的处理方法、装置、存储介质及电子设备
CN108319720A (zh) 基于人工智能的人机交互方法、装置及计算机设备
CN105225135B (zh) 潜力客户识别方法以及装置
CN110929169A (zh) 基于改进Canopy聚类协同过滤算法的职位推荐方法
CN108717459B (zh) 一种面向用户评论信息的移动应用缺陷定位方法
CN110807086A (zh) 文本数据标注方法及装置、存储介质、电子设备
CN112883990A (zh) 数据分类方法及装置、计算机存储介质、电子设备
CN113888265A (zh) 产品推荐方法、装置、设备及计算机可读存储介质
CN114398473A (zh) 企业画像生成方法、装置、服务器及存储介质
CN110390019A (zh) 一种试题的聚类方法、去重方法及系统
WO2022142019A1 (zh) 基于智能机器人的问题分发方法、装置、电子设备及存储介质
CN116645554A (zh) 多模态医学报告检索方法、装置、电子设备及存储介质
CN112084764A (zh) 数据检测方法、装置、存储介质及设备
EP4697196A1 (en) Intelligent question answering method and apparatus, computing device, program product, and storage medium
WO2020135054A1 (zh) 视频推荐方法、装置、设备及存储介质
CN110347821A (zh) 一种文本类别标注的方法、电子设备和可读存储介质
CN107943947A (zh) 一种基于Hadoop平台的改进并行KNN网络舆情分类算法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18790332

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18790332

Country of ref document: EP

Kind code of ref document: A1