WO2018014610A1 - 基于c4.5决策树算法的特定用户挖掘系统及其方法 - Google Patents
基于c4.5决策树算法的特定用户挖掘系统及其方法 Download PDFInfo
- Publication number
- WO2018014610A1 WO2018014610A1 PCT/CN2017/080782 CN2017080782W WO2018014610A1 WO 2018014610 A1 WO2018014610 A1 WO 2018014610A1 CN 2017080782 W CN2017080782 W CN 2017080782W WO 2018014610 A1 WO2018014610 A1 WO 2018014610A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- attribute
- algorithm
- classification
- decision tree
- split
- 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
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
Definitions
- the invention relates to the field of data mining of live website, and particularly relates to a specific user mining system and method thereof based on C4.5 decision tree algorithm.
- the traditional user interest degree mining is mostly manual extraction, and the user behavior is manually screened according to individual experience combined with effective behavior characteristics.
- the above methods of manual screening often have a large degree of subjectivity.
- the data tends to have many dimensions and a large amount of data. It is difficult to comprehensively calculate the relevant indicators of user interest by manual.
- the object of the present invention is to provide a specific user mining system based on the C4.5 decision tree algorithm, which can comprehensively count user interest degree related indicators based on the decision tree mining algorithm and mine specific users.
- a specific user mining system based on the C4.5 decision tree algorithm including:
- a sample selection module configured to select user samples, the user samples being classified into a paid user sample and an unpaid user sample according to the class label;
- a behavior attribute statistics module for counting attributes of a classification attribute of the user sample value
- a sample processing module for normalizing the attribute values of the classification attribute as training sample data
- An algorithm platform for receiving the training sample data the algorithm platform including the C4.5 decision tree algorithm, and providing an algorithm interface for the C4.5 decision tree algorithm, and the algorithm platform is based on the training
- the sample data and C4.5 decision tree algorithm train the C4.5 decision tree model.
- the classification attribute includes a viewing time, a viewing time, a number of hair curtains, a number of gift gifts, a number of gifts received, a number of attention rooms, and a number of attention zones.
- the algorithm platform is a Spark platform, and the Spark platform includes an algorithm component Spark MLlib, where the Spark MLlib includes an algorithm library, and the C4.5 decision tree algorithm is provided in the algorithm library. And the Spark MLlib provides an algorithm interface for the C4.5 decision tree algorithm.
- the algorithm platform is based on the C4.5 decision tree algorithm to find the classification attribute with the largest information gain rate in the classification attribute, and uses it as the classification attribute to be split, and the algorithm platform is also used to find Determining an optimal split position of the classification attribute to be split, and splitting the training sample data in an optimal split position of the classification attribute to be split, for each type of data obtained after splitting,
- the algorithm platform re-determines the classification attributes to be split in the remaining classification attributes based on the C4.5 decision tree algorithm, and then determines the optimal split position of each type of data, and finally optimizes each type of data in each.
- the split position is split according to the classification attribute that is determined again to be split until all the classification attributes are used up.
- the present invention also provides a method for mining a specific user based on a decision tree mining algorithm that can comprehensively calculate relevant indicators of user interest.
- the technical solution adopted by the present invention is: a method for mining a specific user by using the above specific user mining system, comprising the following steps:
- the sample selection module selects a user sample, and the user sample is divided into a paid user sample and an unpaid user sample according to the class label;
- the behavior attribute statistics module calculates the attribute value of the classification attribute of the user sample
- the sample processing module normalizes the attribute values of the classification attributes to obtain training sample data D;
- the algorithm platform receives the training sample data D and trains the C4.5 decision tree model
- the attribute value of the new user's classification attribute is normalized in step S3 and then input into the trained C4.5 decision tree model, and the process ends.
- the training C4.5 decision tree model includes the following steps:
- A is an attribute set of all classification attributes
- the classification attribute A j is an element in the attribute set A
- the classification attribute A j divides the training sample data D into v disjoint subsets ⁇ D 1 , D 2 . .., D v ⁇ , calculating the information entropy weighted sum of the subsets divided by the classification attribute A j , Wherein the value of v is equal to the number of class labels, and D j represents the training sample data corresponding to the jth class after dividing D into v classes;
- the attribute values of the classification attributes to be split in the training sample data D are sorted in an ascending order to obtain a data set, and the data set is divided into N+1 different two sub data sets, corresponding to N+1 dividing points, for N-1 division points between the first division point and the last division point, determine the position of the adjacent attribute values of N-1 by calculating the average value of the pairwise attribute values, and ensure all attribute values of the classification attribute to be split Located between the first dividing point and the last dividing point, according to N+1 different two sub-data sets, calculating the information gain of all the dividing points, using the dividing point with the largest information gain as the optimal splitting position, and then training
- the sample data D is split into categories equal to the number of class labels at the optimal splitting position according to the classification attribute to be split;
- the behavior attribute statistics module counts the user's viewing time, the number of views, the number of gift gifts, the number of gifts received, the number of rooms concerned, and the number of partitions.
- the algorithm platform is a Spark platform, and the The Spark platform includes an algorithm component Spark MLlib, the Spark MLlib includes an algorithm library, the C4.5 decision tree algorithm is provided in the algorithm library, and the Spark MLlib provides an algorithm interface for the C4.5 decision tree algorithm and Train the C4.5 decision tree model.
- the value of m in the step S41 is 2.
- the value of v in the step S42 is 2.
- the specific user mining system based on C4.5 decision tree algorithm of the invention comprises Spark platform, Spark platform includes algorithm component Spark MLlib, Spark MLlib provides algorithm interface for C4.5 decision tree algorithm, based on C4.5 decision tree algorithm, can When the data dimension is large and the data volume is large, the user interest degree related indicators are comprehensively counted, which is convenient for mining specific users. In addition, based on Spark's memory calculation, the calculation speed is faster, which greatly shortens the calculation cycle.
- FIG. 1 is a schematic structural diagram of a specific user mining system based on a C4.5 decision tree algorithm in the present invention
- FIG. 3 is a flow chart of training a C4.5 decision tree model in the present invention.
- the present invention provides a specific user mining system based on a C4.5 decision tree algorithm, which includes a sample selection module, a behavior attribute statistics module, a sample processing module, and an algorithm platform.
- a sample selection module for selecting user samples, the user samples are classified into a paid user sample and an unpaid user sample according to the class label, wherein the paid user sample is marked as 1, unpaid The fee user sample is marked as 0.
- the behavior attribute statistics module is used for counting the attribute values of the classification attributes of the user samples.
- the statistical classification attributes in the present invention mainly include the viewing time, the number of views, the number of hair curtains, the number of gift gifts, the number of gift gifts, and the attention room. Number, attention to the number of partitions.
- Y (X-MinValue(X))/(MaxValue(X)-MinValue(X) Normalization, where X is the corresponding attribute value in a classification attribute, MinValue(X) represents the minimum value of the attribute value in the classification attribute, and MaxValue(X) is the maximum value of the attribute value in the classification attribute.
- the attribute values of the classified attributes after normalization are concentrated between (0, 1).
- An algorithm platform for receiving training sample data includes the C4.5 decision tree algorithm, and provides an algorithm interface for the C4.5 decision tree algorithm, and training the C4.5 decision tree model, and the algorithm platform is based on the training sample Data and C4.5 decision tree algorithms train the C4.5 decision tree model.
- the algorithm platform in the present invention is the Spark platform, and Spark is a general-purpose parallel framework of Hadoop MapReduce which is open sourced by UC Berkeley AMP lab, and has the advantages of Hadoop MapReduce. But unlike MapReduce, Job intermediate output can be saved in memory, so you no longer need to read and write HDFS. Based on Spark's memory calculation, it is faster in calculation speed and greatly shortens the calculation cycle. Therefore, Spark can be better applied to data mining and machine learning and other iterative MapReduce algorithms.
- the Spark platform includes the algorithm component Spark MLlib (Machine Learning Library), Spark MLlib includes an algorithm library, the algorithm library has a C4.5 decision tree algorithm, and Spark MLlib provides an algorithm interface for the C4.5 decision tree algorithm. Train the C4.5 decision tree model.
- the C4.5 decision tree algorithm is a series of algorithms used in the classification of machine learning and data mining. Its goal is to supervise learning: Given a data set, each of these tuples can be described by a set of attribute values, each of which belongs to a class in a mutually exclusive category. The goal of C4.5 is to find a mapping from attribute values to categories by learning, and this mapping can be used to classify entities with unknown new categories.
- the algorithm platform in the invention is based on the C4.5 decision tree algorithm to find the classification attribute with the largest information gain rate in the classification attribute, and uses it as the classification attribute to be split.
- the algorithm platform is also used to find the optimal splitting position of the classification attribute to be split, and split the training sample data in the best split position of the classification attribute to be split, and for each type of data obtained after splitting, the algorithm platform Based on the C4.5 decision tree algorithm, the classification attributes to be split are respectively determined in the remaining classification attributes, and then the optimal split positions of each type of data are determined respectively. Finally, each type of data is in its optimal split position. Split according to the classification attribute that is determined again to be split until all the classification attributes are used up.
- the present invention further provides a method for mining a specific user by a specific user mining system based on a C4.5 decision tree algorithm, including the following steps:
- the sample selection module selects a user sample, and the user sample is divided into a paid user sample and an unpaid user sample according to the class label;
- the behavior attribute statistics module calculates the attribute value of the classification attribute of the user sample
- the classification attributes of the users counted in the present invention include the viewing time, the number of views, the number of haircuts, the number of gift gifts, the number of gifts received, the number of rooms of interest, and the number of partitions of interest.
- the sample processing module normalizes the attribute values of the classification attributes to obtain training sample data D;
- each attribute value in each classification attribute is followed.
- the minimum value in the attribute value, MaxValue(X) is the maximum value of the attribute value in this classification attribute, and the attribute value of the classification attribute after normalization is concentrated between (0, 1).
- the algorithm platform receives the training sample data D and trains the C4.5 decision tree model
- the algorithm platform in the present invention is a Spark platform, the Spark platform includes an algorithm component Spark MLlib, the Spark MLlib includes an algorithm library, the C4.5 decision tree algorithm is provided in the algorithm library, and the Spark MLlib is the C4.5 decision tree algorithm. Provide an algorithm interface and train the C4.5 decision tree model.
- the algorithm platform training C4.5 decision tree model in the present invention includes the following steps:
- A is an attribute set of all classification attributes
- the classification attribute A j is an element in the attribute set A
- the classification attribute A j divides the training sample data D into v disjoint subsets ⁇ D 1 , D 2 . .., D v ⁇ , calculating the information entropy weighted sum of the subsets divided by the classification attribute A j
- the value of v is equal to the number of class labels
- D j represents the training sample data corresponding to the jth class after dividing D into v classes; in the present invention, the value of v is 2.
- the attribute values of the classification attributes to be split in the training sample data D are sorted in an ascending order to obtain a data set, and the data set is divided into N+1 different two sub data sets, corresponding to N+1 dividing points, for N-1 dividing points between the first dividing point and the last dividing point, determining the position of the pairwise adjacent attribute values by calculating N-1, and ensuring the classification of the training sample data D to be split All attribute values of an attribute are between the first and last dividing points.
- N+1 dividing points divide the data set into N+1 different two sub-data sets, and then calculate the information gain of all the dividing points by using the method described in steps S41-S43, and the dividing point with the largest information gain is the most
- the training sample data D is split into the categories equal to the number of class labels at the optimal splitting position according to the classification attribute to be split;
- the present invention ensures that all attribute values of the classification attribute to be split in the training sample data D are located between the first division point and the last division point, and the attribute value of the first record is decremented by 1 as the position of the first division point. , increments the attribute value of the last record by 1 as the position of the last division point.
- step S47 Repeat step S41-S46 for each type of data in which the training sample data D is split according to the classification attribute to be split at the optimal splitting position, and again determine the classification attribute to be split in the remaining classification attribute of the attribute set A. Then, determine the optimal split position of each type of data, and finally split each type of data in the respective optimal split position according to the classification attribute to be determined again until all the classification attributes in the attribute set A are used up. .
- step S41 the calculation is after splitting.
- the information entropy of each type of data not the information entropy of the training sample data D.
- the classification attribute of the present invention includes the viewing time, the number of views, the number of haircuts, the number of gift gifts, the number of gifts received, the number of rooms of interest, and the number of partitions. If the step S45 determines that the viewing duration is the classification attribute to be split, then the optimal split position is determined by step S46, and then the training sample data D is split into two types of data according to the classification attribute to be split at the optimal split position (because The class tag of the present invention is 2), and then steps S41-S46 are repeatedly performed for each of the two types of data.
- the classification attributes of the two types of data to be split are determined again, and then Then determine the best split position of each of the two types of data, and then complete the split, the split continues until all the classification attributes are used up.
- the attribute value of the new user's classification attribute is normalized in step S3 and input into the trained C4.5 decision tree model, and the process ends.
- the new user data is input into the trained C4.5 decision tree model in the format of the training sample data, and the predicted new user category can be obtained, and the user interest degree related index can be obtained to determine whether it is a specific user.
Landscapes
- Business, Economics & Management (AREA)
- Strategic Management (AREA)
- Engineering & Computer Science (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Finance (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Economics (AREA)
- Marketing (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
一种基于C4.5决策树算法的特定用户挖掘系统及其方法,涉及直播网站数据挖掘领域,所述方法包括:选择用户样本(S1),用户样本根据类标签分为付费用户样本和未付费用户样本;统计用户样本的分类属性的属性值(S2);将分类属性的属性值归一化处理作为训练样本数据(S3);算法平台接收训练样本数据,并训练C4.5决策树模型(S4);C4.5决策树模型训练完毕后,将新的用户的分类属性的属性值归一化处理后输入训练好的C4.5决策树模型(S5)。基于C4.5决策树算法的特定用户挖掘系统能够在数据维度多、数据量大的情况下全面统计用户兴趣度相关指标,便于挖掘特定用户。
Description
本发明涉及直播网站数据挖掘领域,具体涉及一种基于C4.5决策树算法的特定用户挖掘系统及其方法。
近些年来,在直播行业飞速发展的同时,直播网站的用户也呈爆炸式增长。如何快速有效的将具有潜力的用户从全站用户中筛选出来,方便运营人员针对特定用户做进一步的精细化营销方案,提高用户的付费转化率,是每一个直播网站不可逃避的问题。
目前,传统的用户兴趣度挖掘多为人工提取、根据个人经验组合有效行为特征来人工筛选用户兴趣度。然而上述人工筛选的方法往往带有较大程度的主观性,另外,在海量数据的场景下,数据往往维度多、数据量大,靠人工是很难将用户兴趣度相关指标统计全面的。
发明内容
针对现有技术中存在的缺陷,本发明的目的在于提供一种基于C4.5决策树算法的特定用户挖掘系统,其基于决策树挖掘算法能全面的统计用户兴趣度相关指标,挖掘特定用户。
为达到以上目的,本发明采取的技术方案是:一种基于C4.5决策树算法的特定用户挖掘系统,包括:
样本选择模块,其用于选择用户样本,所述用户样本根据类标签分为付费用户样本和未付费用户样本;
行为属性统计模块,其用于统计所述用户样本的分类属性的属性
值;
样本处理模块,其用于将所述分类属性的属性值归一化处理作为训练样本数据;以及
算法平台,其用于接收所述训练样本数据,所述算法平台包括所述C4.5决策树算法,并为所述C4.5决策树算法提供算法接口,且所述算法平台基于所述训练样本数据和C4.5决策树算法训练C4.5决策树模型。
在上述技术方案的基础上,所述分类属性包括观看时长、观看次数、发弹幕数、送出虚礼礼物数、领取虚礼礼物数、关注房间数、关注分区数。
在上述技术方案的基础上,所述算法平台为Spark平台,所述Spark平台包括算法组件Spark MLlib,所述Spark MLlib包括算法库,所述算法库内设有所述C4.5决策树算法,且所述Spark MLlib为所述C4.5决策树算法提供算法接口。
在上述技术方案的基础上,所述算法平台基于C4.5决策树算法找出分类属性中信息增益率最大的分类属性,并将其作为待分裂的分类属性,所述算法平台还用于找出所述待分裂的分类属性的最佳分裂位置,并将所述训练样本数据在所述待分裂的分类属性的最佳分裂位置进行分裂,针对分裂后所得到的每一类数据,所述算法平台基于C4.5决策树算法在剩下的分类属性中各自再次确定待分裂的分类属性,再各自确定每一类数据各自的最佳分裂位置,最后对每一类数据在各自的最佳分裂位置按照再次确定待分裂的分类属性进行分裂,直到所有分类属性全部用完。
与此同时,本发明还提供一种基于决策树挖掘算法能全面的统计用户兴趣度相关指标,挖掘特定用户的方法。
为达到以上目的,本发明采取的技术方案是:一种利用上述特定用户挖掘系统挖掘特定用户的方法,包括以下步骤:
S1.样本选择模块选择用户样本,用户样本根据类标签分为付费用户样本和未付费用户样本;
S2.行为属性统计模块统计用户样本的分类属性的属性值;
S3.样本处理模块将分类属性的属性值归一化处理得到训练样本数据D;
S4.算法平台接收训练样本数据D,并训练C4.5决策树模型;
S5.C4.5决策树模型训练完毕后,将新的用户的分类属性的属性值按步骤S3归一化处理后输入训练好的C4.5决策树模型,结束。
在上述技术方案的基础上,所述训练C4.5决策树模型包括以下步骤:
S42.记A为所有分类属性的属性集,分类属性Aj为属性集A中的一个元素,分类属性Aj将训练样本数据D分为v个不相交的子集{D1,D2...,Dv},计算按照分类属性Aj划分后的个子集的信息熵加权和,其中v的值与类标签的数量相等,Dj表示将D分为v类后第j类所对应的训练样本数据;
S45.找出属性集A中信息增益率最大的分类属性,并将其作为待分裂的分类属性;
S46.将训练样本数据D中待分裂的分类属性的属性值按照递增排序得到一数据集,将数据集划分为N+1种不同的两个子数据集,对应N+1个划分点,对于位于第一个划分点和最后一个划分点中间的N-1个划分点,通过计算N-1对两两相邻属性值的平均值来确定其位置,并保证待分裂的分类属性的所有属性值位于第一个划分点和最后一个划分点之间,根据N+1种不同的两个子数据集,计算所有划分点的信息增益,将信息增益最大的划分点作为最佳分裂位置,然后将训练样本数据D按照待分裂的分类属性在最佳分裂位置分裂成与类标签的数量相等的类别;
S47.对于训练样本数据D按照待分裂的分类属性在最佳分裂位置分裂成的每一类数据,重复执行步骤S41-S46,在属性集A剩下的分类属性中各自再次确定待分裂的分类属性,再各自确定每一类数据各自的最佳分裂位置,最后对每一类数据在各自的最佳分裂位置按照再次确定待分裂的分类属性进行分裂,直到属性集A中的所有分类属性全部用完。
在上述技术方案的基础上,所述行为属性统计模块统计用户的观看时长、观看次数、送出虚礼礼物数、领取虚礼礼物数、关注房间数、关注分区数。
在上述技术方案的基础上,所述算法平台为Spark平台,所述
Spark平台包括算法组件Spark MLlib,所述Spark MLlib包括算法库,所述算法库内设有所述C4.5决策树算法,且所述Spark MLlib为所述C4.5决策树算法提供算法接口并训练C4.5决策树模型。
在上述技术方案的基础上,所述步骤S41中m的取值为2。
在上述技术方案的基础上,所述步骤S42中v的取值为2。
与现有技术相比,本发明的优点在于:
本发明的基于C4.5决策树算法的特定用户挖掘系统包括Spark平台,Spark平台包括算法组件Spark MLlib,Spark MLlib为C4.5决策树算法提供算法接口,基于C4.5决策树算法,能够在数据维度多、数据量大的情况下全面统计用户兴趣度相关指标,便于挖掘特定用户。此外基于Spark的内存计算,在计算速度方面更快,大大缩短了计算周期。
图1为本发明中基于C4.5决策树算法的特定用户挖掘系统的结构示意图;
图2为本发明中挖掘特定用户的流程图;
图3为本发明中训练C4.5决策树模型的流程图。
以下结合附图对本发明作进一步详细说明。
参见图1所示,本发明提供一种基于C4.5决策树算法的特定用户挖掘系统,其包括样本选择模块、行为属性统计模块、样本处理模块和算法平台。
样本选择模块,其用于选择用户样本,用户样本根据类标签分为付费用户样本和未付费用户样本,其中付费用户样本标记为1,未付
费用户样本标记为0。
行为属性统计模块,其用于统计用户样本的分类属性的属性值,本发明中统计的分类属性主要包括观看时长、观看次数、发弹幕数、送出虚礼礼物数、领取虚礼礼物数、关注房间数、关注分区数。
样本处理模块,其用于将分类属性的属性值归一化处理作为训练样本数据。因为上述的分类属性的量纲不同。为了避免因分类属性量纲不同,对分类结果造成影响,故对每个分类属性中的每一个属性值按照公式Y=(X-MinValue(X))/(MaxValue(X)-MinValue(X))进行归一化,其中X为某一个分类属性中对应的属性值,MinValue(X)表示这个分类属性中属性值里面的最小值,MaxValue(X)是这个分类属性中属性值的最大值,归一化之后的分类属性的属性值都集中在(0,1]之间。
算法平台,其用于接收训练样本数据,算法平台包括所述C4.5决策树算法,且为C4.5决策树算法提供算法接口,并训练C4.5决策树模型,且算法平台基于训练样本数据和C4.5决策树算法训练C4.5决策树模型。
本发明中的算法平台为Spark平台,Spark是UC Berkeley AMP lab所开源的类Hadoop MapReduce的通用并行框架,其拥有Hadoop MapReduce所具有的优点。但不同于MapReduce的是Job中间输出结果可以保存在内存中,从而不再需要读写HDFS,基于Spark的内存计算,在计算速度方面更快,大大缩短了计算周期。因此Spark能更好地适用于数据挖掘与机器学习等需要迭代的MapReduce的算法。Spark平台包括算法组件Spark MLlib(Machine Learning Library,机器学习库),Spark MLlib包括算法库,算法库内设有C4.5决策树算法,且Spark MLlib为C4.5决策树算法提供算法接口,并训练C4.5决策树模型。
C4.5决策树算法是一系列用在机器学习和数据挖掘的分类问题中的算法。它的目标是监督学习:给定一个数据集,其中的每一个元组都能用一组属性值来描述,每一个元组属于一个互斥的类别中的某一类。C4.5的目标是通过学习,找到一个从属性值到类别的映射关系,并且这个映射能用于对新的类别未知的实体进行分类。
本发明中的算法平台基于C4.5决策树算法找出分类属性中信息增益率最大的分类属性,并将其作为待分裂的分类属性。算法平台还用于找出待分裂的分类属性的最佳分裂位置,并将训练样本数据在待分裂的分类属性的最佳分裂位置进行分裂,针对分裂后所得到的每一类数据,算法平台基于C4.5决策树算法在剩下的分类属性中各自再次确定待分裂的分类属性,再各自确定每一类数据各自的最佳分裂位置,最后对每一类数据在各自的最佳分裂位置按照再次确定待分裂的分类属性进行分裂,直到所有分类属性全部用完。
参见图2所示,本发明还提供一种基于C4.5决策树算法的特定用户挖掘系统挖掘特定用户的方法,包括以下步骤:
S1.样本选择模块选择用户样本,用户样本根据类标签分为付费用户样本和未付费用户样本;
本发明中的类标签有两个,即付费和未付费。
S2.行为属性统计模块统计用户样本的分类属性的属性值;
本发明中统计的用户的分类属性包括观看时长、观看次数、发弹幕数、送出虚礼礼物数、领取虚礼礼物数、关注房间数、关注分区数。
S3.样本处理模块将分类属性的属性值归一化处理得到训练样本数据D;
由于用户的分类属性的量纲不同。为了避免因分类属性量纲不同,对分类结果造成影响,故对每个分类属性中的每一个属性值按照
公式Y=(X-MinValue(X))/(MaxValue(X)-MinValue(X))进行归一化,其中X为某一个分类属性中对应的属性值,MinValue(X)表示这个分类属性中属性值里面的最小值,MaxValue(X)是这个分类属性中属性值的最大值,归一化之后的分类属性的属性值都集中在(0,1]之间。
S4.算法平台接收训练样本数据D,并训练C4.5决策树模型;
本发明中的算法平台为Spark平台,Spark平台包括算法组件Spark MLlib,Spark MLlib包括算法库,算法库内设有所述C4.5决策树算法,且Spark MLlib为所述C4.5决策树算法提供算法接口,并训练C4.5决策树模型。参见图3所示,本发明中的算法平台训练C4.5决策树模型包括以下步骤:
S42.记A为所有分类属性的属性集,分类属性Aj为属性集A中的一个元素,分类属性Aj将训练样本数据D分为v个不相交的子集{D1,D2...,Dv},计算按照分类属性Aj划分后的个子集的信息熵加权和,其中v的值与类标签的数量相等,Dj表示将D分为v类后第j类所对应的训练样本数据;本发明中v的取值为2。
S45.找出属性集A中信息增益率最大的分类属性,并将其作为待分裂的分类属性;
S46.将训练样本数据D中待分裂的分类属性的属性值按照递增排序得到一数据集,将数据集划分为N+1种不同的两个子数据集,对应N+1个划分点,对于位于第一个划分点和最后一个划分点中间的N-1个划分点,通过计算N-1对两两相邻属性值的平均值来确定其位置,并保证训练样本数据D中待分裂的分类属性的所有属性值位于第一个划分点和最后一个划分点之间。N+1个划分点将数据集划分为N+1种不同的两个子数据集,然后利用步骤S41-S43中所述的方式计算所有划分点的信息增益,将信息增益最大的划分点作为最佳分裂位置,然后将训练样本数据D按照待分裂的分类属性在最佳分裂位置分裂成与类标签的数量相等的类别;
本发明为了保证训练样本数据D中待分裂的分类属性的所有属性值位于第一个划分点和最后一个划分点之间,将第一个记录的属性值减1作为第一个划分点的位置,将最后一个记录的属性值加1作为最后一个划分点的位置。
S47.对于训练样本数据D按照待分裂的分类属性在最佳分裂位置分裂成的每一类数据,重复执行步骤S41-S46,在属性集A剩下的分类属性中再次确定待分裂的分类属性,再确定每一类数据各自的最佳分裂位置,最后对每一类数据在各自的最佳分裂位置按照再次确定待分裂的分类属性进行分裂,直到属性集A中的所有分类属性全部用完。
需要指出的是,重复执行步骤S41的过程中,其计算的是分裂后
每一类数据的信息熵,而不是训练样本数据D的信息熵。
具体的,本发明的分类属性包括观看时长、观看次数、发弹幕数、送出虚礼礼物数、领取虚礼礼物数、关注房间数、关注分区数。若步骤S45确定了观看时长为待分裂的分类属性,则通过步骤S46确定最佳分裂位置,然后便将训练样本数据D按照待分裂的分类属性在最佳分裂位置分裂成的两类数据(因为本发明的类标签为2),然后针对这两类数据各自重复执行步骤S41-S46,此时在除观看时长外的分类属性中,再次确定这两类数据各自的待分裂的分类属性,然后再分别确定这两类数据各自的最佳分裂位置,然后完成分裂,分裂一直持续到所有分类属性全部用完为止。
S5.C4.5决策树模型训练完毕后,将新的用户的分类属性的属性值按步骤S3归一化后输入训练好的C4.5决策树模型,结束。
因为C4.5通过学习找到一个从属性值到类别的映射关系,并且这个映射能用于对新的类别未知的实体进行分类。故将新的用户数据按训练样本数据的格式输入训练好的C4.5决策树模型,就能得到预测的新用户的类别,得到用户兴趣度相关指标,从而判断其是否是特定用户。
本发明不局限于上述实施方式,对于本技术领域的普通技术人员来说,在不脱离本发明原理的前提下,还可以做出若干改进和润饰,这些改进和润饰也视为本发明的保护范围之内。本说明书中未作详细描述的内容属于本领域专业技术人员公知的现有技术。
Claims (10)
- 一种基于C4.5决策树算法的特定用户挖掘系统,其特征在于,包括:样本选择模块,其用于选择用户样本,所述用户样本根据类标签分为付费用户样本和未付费用户样本;行为属性统计模块,其用于统计所述用户样本的分类属性的属性值;样本处理模块,其用于将所述分类属性的属性值归一化处理作为训练样本数据;以及算法平台,其用于接收所述训练样本数据,所述算法平台包括所述C4.5决策树算法,并为所述C4.5决策树算法提供算法接口,且所述算法平台基于所述训练样本数据和C4.5决策树算法训练C4.5决策树模型。
- 如权利要求1所述的基于C4.5决策树算法的特定用户挖掘系统,其特征在于:所述分类属性包括观看时长、观看次数、发弹幕数、送出虚礼礼物数、领取虚礼礼物数、关注房间数、关注分区数。
- 如权利要求1所述的基于C4.5决策树算法的特定用户挖掘系统,其特征在于:所述算法平台为Spark平台,所述Spark平台包括算法组件Spark MLlib,所述Spark MLlib包括算法库,所述算法库内设有所述C4.5决策树算法,且所述Spark MLlib为所述C4.5决策树算法提供算法接口。
- 如权利要求1所述的基于C4.5决策树算法的特定用户挖掘系统,其特征在于:所述算法平台基于C4.5决策树算法找出分类属性中信息增益率最大的分类属性,并将其作为待分裂的分类属性,所述算法平台还用于找出所述待分裂的分类属性的最佳分裂位置,并将所 述训练样本数据在所述待分裂的分类属性的最佳分裂位置进行分裂,针对分裂后所得到的每一类数据,所述算法平台基于C4.5决策树算法在剩下的分类属性中各自再次确定待分裂的分类属性,再各自确定每一类数据各自的最佳分裂位置,最后对每一类数据在各自的最佳分裂位置按照再次确定待分裂的分类属性进行分裂,直到所有分类属性全部用完。
- 一种利用权利要求1所述的特定用户挖掘系统挖掘特定用户的方法,其特征在于,包括以下步骤:S1.样本选择模块选择用户样本,用户样本根据类标签分为付费用户样本和未付费用户样本;S2.行为属性统计模块统计用户样本的分类属性的属性值;S3.样本处理模块将分类属性的属性值归一化处理得到训练样本数据D;S4.算法平台接收训练样本数据D,并训练C4.5决策树模型;S5.C4.5决策树模型训练完毕后,将新的用户的分类属性的属性值按步骤S3归一化处理后输入训练好的C4.5决策树模型,结束。
- 如权利要求5所述的挖掘特定用户的方法,其特征在于,所述训练C4.5决策树模型包括以下步骤:S42.记A为所有分类属性的属性集,分类属性Aj为属性集A中的一个元素,分类属性Aj将训练样本数据D分为v个不相交的子集{D1,D2...,Dv},计算按照分类属性Aj划分后的个子集的信息熵加权和, 其中v的值与类标签的数量相等,Dj表示将D分为v类后第j类所对应的训练样本数据;S45.找出属性集A中信息增益率最大的分类属性,并将其作为待分裂的分类属性;S46.将训练样本数据D中待分裂的分类属性的属性值按照递增排序得到一数据集,将数据集划分为N+1种不同的两个子数据集,对应N+1个划分点,对于位于第一个划分点和最后一个划分点中间的N-1个划分点,通过计算N-1对两两相邻属性值的平均值来确定其位置,并保证待分裂的分类属性的所有属性值位于第一个划分点和最后一个划分点之间,根据N+1种不同的两个子数据集,计算所有划分点的信息增益,将信息增益最大的划分点作为最佳分裂位置,然后将训练样本数据D按照待分裂的分类属性在最佳分裂位置分裂成与类标签的数量相等的类别;S47.对于训练样本数据D按照待分裂的分类属性在最佳分裂位置分裂成的每一类数据,重复执行步骤S41-S46,在属性集A剩下的分类属性中各自再次确定待分裂的分类属性,再各自确定每一类数据各自的最佳分裂位置,最后对每一类数据在各自的最佳分裂位置按照 再次确定待分裂的分类属性进行分裂,直到属性集A中的所有分类属性全部用完。
- 如权利要求5所述的挖掘特定用户的方法,其特征在于:所述行为属性统计模块统计用户的观看时长、观看次数、送出虚礼礼物数、领取虚礼礼物数、关注房间数、关注分区数的数据。
- 如权利要求5所述的挖掘特定用户的方法,其特征在于:所述算法平台为Spark平台,所述Spark平台包括算法组件Spark MLlib,所述Spark MLlib包括算法库,所述算法库内设有所述C4.5决策树算法,且所述Spark MLlib为所述C4.5决策树算法提供算法接口并训练C4.5决策树模型。
- 如权利要求6所述的挖掘特定用户的方法,其特征在于:所述步骤S41中m的取值为2。
- 如权利要求6所述的挖掘特定用户的方法,其特征在于:所述步骤S42中v的取值为2。
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610573315.4 | 2016-07-20 | ||
| CN201610573315.4A CN106228398A (zh) | 2016-07-20 | 2016-07-20 | 基于c4.5决策树算法的特定用户挖掘系统及其方法 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2018014610A1 true WO2018014610A1 (zh) | 2018-01-25 |
Family
ID=57531576
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2017/080782 Ceased WO2018014610A1 (zh) | 2016-07-20 | 2017-04-17 | 基于c4.5决策树算法的特定用户挖掘系统及其方法 |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN106228398A (zh) |
| WO (1) | WO2018014610A1 (zh) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197187A (zh) * | 2018-02-24 | 2019-09-03 | 腾讯科技(深圳)有限公司 | 对用户流失进行预测的方法、设备、存储介质以及处理器 |
| CN110674468A (zh) * | 2019-09-25 | 2020-01-10 | 郑州轻工业学院 | 一种基于改进粗集算法的细纱断纱因素定量分析方法 |
| CN111259933A (zh) * | 2020-01-09 | 2020-06-09 | 中国科学院计算技术研究所 | 基于分布式并行决策树的高维特征数据分类方法及系统 |
| KR20200103098A (ko) * | 2018-01-31 | 2020-09-01 | 에이에스엠엘 네델란즈 비.브이. | 공정 매개변수를 기반으로 기판을 라벨링하는 방법 |
| CN111860946A (zh) * | 2020-06-09 | 2020-10-30 | 广东电网有限责任公司 | 基于c4.5算法的防鸟装置有效性评价方法 |
| CN112183604A (zh) * | 2020-09-22 | 2021-01-05 | 国网江苏省电力有限公司营销服务中心 | 一种基于决策树的电能计量装置选型方法和系统 |
| CN112183606A (zh) * | 2020-09-23 | 2021-01-05 | 内蒙古电力(集团)有限责任公司乌海电业局 | 基于c4.5算法的电力系统故障识别分类方法及系统 |
| CN112364743A (zh) * | 2020-11-02 | 2021-02-12 | 北京工商大学 | 一种基于半监督学习和弹幕分析的视频分类方法 |
| CN112714080A (zh) * | 2020-12-23 | 2021-04-27 | 上海观安信息技术股份有限公司 | 一种基于spark图算法的互连关系分类方法及系统 |
| CN114254697A (zh) * | 2021-12-07 | 2022-03-29 | 中信银行股份有限公司 | 一种基于决策树和二分k均值聚类的客户分群方法及系统 |
| CN114398757A (zh) * | 2021-12-10 | 2022-04-26 | 国网浙江省电力有限公司温州供电公司 | 基于大数据分析的变电检修效能优化方法 |
| CN114662412A (zh) * | 2022-05-23 | 2022-06-24 | 深圳市远湖科技有限公司 | 基于深度学习的服装设计方法、装置、设备及存储介质 |
| CN114925833A (zh) * | 2022-04-20 | 2022-08-19 | 中国人民解放军91977部队 | 一种基于能力数据底图的目标状态规律知识挖掘方法 |
| CN115600513A (zh) * | 2022-12-16 | 2023-01-13 | 山东省地质矿产勘查开发局八〇一水文地质工程地质大队(山东省地矿工程勘察院)(Cn) | 一种岩溶塌陷监测预警及防控一体化信息化模拟研判系统 |
| CN115914141A (zh) * | 2022-09-23 | 2023-04-04 | 暨南大学 | 一种基于p4硬件交换机的网络数据流分类预测方法 |
| CN116010688A (zh) * | 2022-12-12 | 2023-04-25 | 和元达信息科技有限公司 | 一种用户行为标签识别方法、系统、设备及存储介质 |
| CN116017546A (zh) * | 2022-12-27 | 2023-04-25 | 天翼物联科技有限公司 | 一种拨测系统 |
| CN119272086A (zh) * | 2024-08-30 | 2025-01-07 | 深圳市正针金属科技有限公司 | 产业中试平台化工具中的数据采集与处理方法及系统 |
| CN120354059A (zh) * | 2025-06-24 | 2025-07-22 | 国网甘肃省电力公司电力科学研究院 | 一种虚拟电厂数据处理方法及系统 |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106228398A (zh) * | 2016-07-20 | 2016-12-14 | 武汉斗鱼网络科技有限公司 | 基于c4.5决策树算法的特定用户挖掘系统及其方法 |
| CN106845706A (zh) * | 2017-01-19 | 2017-06-13 | 浙江工商大学 | 在线社交网络用户关系强度预测方法 |
| CN107122390B (zh) * | 2017-03-04 | 2020-09-25 | 华数传媒网络有限公司 | 基于用户群组的推荐系统构建方法 |
| CN107666615B (zh) * | 2017-09-04 | 2020-05-19 | 广州虎牙信息科技有限公司 | 潜力主播用户的挖掘方法、装置及服务器 |
| CN111199243B (zh) * | 2019-12-19 | 2022-10-14 | 中国人民解放军战略支援部队信息工程大学 | 一种基于改进决策树的空中目标识别方法及系统 |
| CN113760484B (zh) * | 2020-06-29 | 2024-10-22 | 北京沃东天骏信息技术有限公司 | 数据处理的方法和装置 |
| CN111752995A (zh) * | 2020-06-30 | 2020-10-09 | 成都信息工程大学 | 一种学生数据挖掘系统及方法 |
| CN112258338A (zh) * | 2020-09-14 | 2021-01-22 | 陕西讯格信息科技有限公司 | 一种基于最近邻算法的基站费用自动化稽核方法 |
| CN112348583B (zh) * | 2020-11-04 | 2022-12-06 | 贝壳技术有限公司 | 用户偏好生成方法与生成系统 |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102214213A (zh) * | 2011-05-31 | 2011-10-12 | 中国科学院计算技术研究所 | 一种采用决策树的数据分类方法和系统 |
| CN102567661A (zh) * | 2010-12-31 | 2012-07-11 | 北京奇虎科技有限公司 | 基于机器学习的程序识别方法及装置 |
| CN102968447A (zh) * | 2012-10-24 | 2013-03-13 | 西安工程大学 | 基于决策树算法的seo关键词竞争程度计算方法 |
| CN103853979A (zh) * | 2010-12-31 | 2014-06-11 | 北京奇虎科技有限公司 | 基于机器学习的程序识别方法及装置 |
| CN104794195A (zh) * | 2015-04-17 | 2015-07-22 | 南京大学 | 一种用于电信潜在换机用户发现的数据挖掘方法 |
| CN106228398A (zh) * | 2016-07-20 | 2016-12-14 | 武汉斗鱼网络科技有限公司 | 基于c4.5决策树算法的特定用户挖掘系统及其方法 |
-
2016
- 2016-07-20 CN CN201610573315.4A patent/CN106228398A/zh active Pending
-
2017
- 2017-04-17 WO PCT/CN2017/080782 patent/WO2018014610A1/zh not_active Ceased
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102567661A (zh) * | 2010-12-31 | 2012-07-11 | 北京奇虎科技有限公司 | 基于机器学习的程序识别方法及装置 |
| CN103853979A (zh) * | 2010-12-31 | 2014-06-11 | 北京奇虎科技有限公司 | 基于机器学习的程序识别方法及装置 |
| CN102214213A (zh) * | 2011-05-31 | 2011-10-12 | 中国科学院计算技术研究所 | 一种采用决策树的数据分类方法和系统 |
| CN102968447A (zh) * | 2012-10-24 | 2013-03-13 | 西安工程大学 | 基于决策树算法的seo关键词竞争程度计算方法 |
| CN104794195A (zh) * | 2015-04-17 | 2015-07-22 | 南京大学 | 一种用于电信潜在换机用户发现的数据挖掘方法 |
| CN106228398A (zh) * | 2016-07-20 | 2016-12-14 | 武汉斗鱼网络科技有限公司 | 基于c4.5决策树算法的特定用户挖掘系统及其方法 |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20200103098A (ko) * | 2018-01-31 | 2020-09-01 | 에이에스엠엘 네델란즈 비.브이. | 공정 매개변수를 기반으로 기판을 라벨링하는 방법 |
| US12045555B2 (en) | 2018-01-31 | 2024-07-23 | Asml Netherlands B.V. | Method to label substrates based on process parameters |
| KR102439008B1 (ko) * | 2018-01-31 | 2022-08-31 | 에이에스엠엘 네델란즈 비.브이. | 공정 매개변수를 기반으로 기판을 라벨링하는 방법 |
| CN110197187A (zh) * | 2018-02-24 | 2019-09-03 | 腾讯科技(深圳)有限公司 | 对用户流失进行预测的方法、设备、存储介质以及处理器 |
| CN110674468A (zh) * | 2019-09-25 | 2020-01-10 | 郑州轻工业学院 | 一种基于改进粗集算法的细纱断纱因素定量分析方法 |
| CN110674468B (zh) * | 2019-09-25 | 2022-09-27 | 郑州轻工业学院 | 一种基于改进粗集算法的细纱断纱因素定量分析方法 |
| CN111259933A (zh) * | 2020-01-09 | 2020-06-09 | 中国科学院计算技术研究所 | 基于分布式并行决策树的高维特征数据分类方法及系统 |
| CN111259933B (zh) * | 2020-01-09 | 2023-06-13 | 中国科学院计算技术研究所 | 基于分布式并行决策树的高维特征数据分类方法及系统 |
| CN111860946B (zh) * | 2020-06-09 | 2023-08-04 | 广东电网有限责任公司 | 基于c4.5算法的防鸟装置有效性评价方法 |
| CN111860946A (zh) * | 2020-06-09 | 2020-10-30 | 广东电网有限责任公司 | 基于c4.5算法的防鸟装置有效性评价方法 |
| CN112183604B (zh) * | 2020-09-22 | 2024-05-28 | 国网江苏省电力有限公司营销服务中心 | 一种基于决策树的电能计量装置选型方法和系统 |
| CN112183604A (zh) * | 2020-09-22 | 2021-01-05 | 国网江苏省电力有限公司营销服务中心 | 一种基于决策树的电能计量装置选型方法和系统 |
| CN112183606A (zh) * | 2020-09-23 | 2021-01-05 | 内蒙古电力(集团)有限责任公司乌海电业局 | 基于c4.5算法的电力系统故障识别分类方法及系统 |
| CN112364743A (zh) * | 2020-11-02 | 2021-02-12 | 北京工商大学 | 一种基于半监督学习和弹幕分析的视频分类方法 |
| CN112714080A (zh) * | 2020-12-23 | 2021-04-27 | 上海观安信息技术股份有限公司 | 一种基于spark图算法的互连关系分类方法及系统 |
| CN112714080B (zh) * | 2020-12-23 | 2023-10-17 | 上海观安信息技术股份有限公司 | 一种基于spark图算法的互连关系分类方法及系统 |
| CN114254697A (zh) * | 2021-12-07 | 2022-03-29 | 中信银行股份有限公司 | 一种基于决策树和二分k均值聚类的客户分群方法及系统 |
| CN114398757A (zh) * | 2021-12-10 | 2022-04-26 | 国网浙江省电力有限公司温州供电公司 | 基于大数据分析的变电检修效能优化方法 |
| CN114925833A (zh) * | 2022-04-20 | 2022-08-19 | 中国人民解放军91977部队 | 一种基于能力数据底图的目标状态规律知识挖掘方法 |
| CN114925833B (zh) * | 2022-04-20 | 2023-07-21 | 中国人民解放军91977部队 | 一种基于能力数据底图的目标状态规律知识挖掘方法 |
| CN114662412A (zh) * | 2022-05-23 | 2022-06-24 | 深圳市远湖科技有限公司 | 基于深度学习的服装设计方法、装置、设备及存储介质 |
| CN114662412B (zh) * | 2022-05-23 | 2022-10-11 | 深圳市远湖科技有限公司 | 基于深度学习的服装设计方法、装置、设备及存储介质 |
| CN115914141A (zh) * | 2022-09-23 | 2023-04-04 | 暨南大学 | 一种基于p4硬件交换机的网络数据流分类预测方法 |
| CN116010688A (zh) * | 2022-12-12 | 2023-04-25 | 和元达信息科技有限公司 | 一种用户行为标签识别方法、系统、设备及存储介质 |
| CN115600513B (zh) * | 2022-12-16 | 2023-08-04 | 山东省地质矿产勘查开发局八〇一水文地质工程地质大队(山东省地矿工程勘察院) | 一种岩溶塌陷监测预警及防控一体化信息化模拟研判系统 |
| CN115600513A (zh) * | 2022-12-16 | 2023-01-13 | 山东省地质矿产勘查开发局八〇一水文地质工程地质大队(山东省地矿工程勘察院)(Cn) | 一种岩溶塌陷监测预警及防控一体化信息化模拟研判系统 |
| CN116017546A (zh) * | 2022-12-27 | 2023-04-25 | 天翼物联科技有限公司 | 一种拨测系统 |
| CN119272086A (zh) * | 2024-08-30 | 2025-01-07 | 深圳市正针金属科技有限公司 | 产业中试平台化工具中的数据采集与处理方法及系统 |
| CN120354059A (zh) * | 2025-06-24 | 2025-07-22 | 国网甘肃省电力公司电力科学研究院 | 一种虚拟电厂数据处理方法及系统 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106228398A (zh) | 2016-12-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11238310B2 (en) | Training data acquisition method and device, server and storage medium | |
| CN107545038B (zh) | 一种文本分类方法与设备 | |
| CN102411563B (zh) | 一种识别目标词的方法、装置及系统 | |
| CN110990718B (zh) | 一种公司形象提升系统的社会网络模型构建模块 | |
| CN108446741B (zh) | 机器学习超参数重要性评估方法、系统及存储介质 | |
| CN104965784B (zh) | 自动测试方法及装置 | |
| CN106228398A (zh) | 基于c4.5决策树算法的特定用户挖掘系统及其方法 | |
| CN105069470A (zh) | 分类模型训练方法及装置 | |
| CN109934704A (zh) | 信息推荐方法、装置、设备和存储介质 | |
| CN110689368B (zh) | 一种移动应用内广告点击率预测系统设计方法 | |
| CN104077723B (zh) | 一种社交网络推荐系统及方法 | |
| CN106156809A (zh) | 用于更新分类模型的方法及装置 | |
| US10387805B2 (en) | System and method for ranking news feeds | |
| CN103617435A (zh) | 一种主动学习图像分类方法和系统 | |
| CN107862022A (zh) | 文化资源推荐系统 | |
| CN114692978A (zh) | 一种基于大数据的社交媒体用户行为预测方法及系统 | |
| WO2020024444A1 (zh) | 人群绩效等级识别方法、装置、存储介质及计算机设备 | |
| CN112463964A (zh) | 文本分类及模型训练方法、装置、设备及存储介质 | |
| CN110610373A (zh) | 一种潜在客户挖掘处理方法及装置 | |
| CN106911591A (zh) | 网络流量的分类方法和系统 | |
| CN106168976A (zh) | 一种基于朴素贝叶斯算法的特定用户挖掘方法及系统 | |
| CN104850868A (zh) | 一种基于k-means和神经网络聚类的客户细分方法 | |
| CN109995611B (zh) | 流量分类模型建立及流量分类方法、装置、设备和服务器 | |
| CN109685122B (zh) | 一种基于密度峰和引力影响的半监督游客画像数据聚类方法 | |
| CN108549857A (zh) | 事件检测模型训练方法、装置及事件检测方法 |
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: 17830248 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: 17830248 Country of ref document: EP Kind code of ref document: A1 |





