WO2017173929A1 - 无监督的特征选择方法、装置 - Google Patents

无监督的特征选择方法、装置 Download PDF

Info

Publication number
WO2017173929A1
WO2017173929A1 PCT/CN2017/077997 CN2017077997W WO2017173929A1 WO 2017173929 A1 WO2017173929 A1 WO 2017173929A1 CN 2017077997 W CN2017077997 W CN 2017077997W WO 2017173929 A1 WO2017173929 A1 WO 2017173929A1
Authority
WO
WIPO (PCT)
Prior art keywords
node
feature
label
nodes
similarity
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/CN2017/077997
Other languages
English (en)
French (fr)
Inventor
黄光远
兰金鹤
麦港林
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.)
Alibaba Group Holding Ltd
Original Assignee
Alibaba Group Holding 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 Alibaba Group Holding Ltd filed Critical Alibaba Group Holding Ltd
Publication of WO2017173929A1 publication Critical patent/WO2017173929A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/28—Databases characterised by their database models, e.g. relational or object models
    • G06F16/284—Relational databases
    • G06F16/285—Clustering or classification
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
    • G06F16/2465—Query processing support for facilitating data mining operations in structured databases

Definitions

  • the present application relates to the field of computer data processing technologies, and in particular, to an unsupervised feature selection method and apparatus.
  • sample data has multiple features, but not every feature is useful for data analysis. Therefore, it is necessary to select effective features from a large number of features to reduce the number of features in the feature set, thereby reducing the specific application.
  • the complexity of the calculation process based on the feature set improves the prediction accuracy.
  • sample data does not have a classification indicator, making it difficult to select features by existing feature selection methods.
  • applying the existing feature selection method may result in a phenomenon in which the information represented by the selected two or more features is similar, that is, the information set finally obtained has information redundancy.
  • the present application provides an unsupervised feature selection method and apparatus.
  • an unsupervised feature selection method including:
  • a target feature set is determined according to the connected graph.
  • Feature topology map including:
  • the feature topology map is segmented to obtain one or more connected graphs, including:
  • the nodes connected by the edges are divided into the same connected graph to obtain one or more connected graphs.
  • the feature topology map is segmented to obtain one or more connected graphs, including:
  • the nodes with the same label are divided into the same connected graph to obtain one or more connected graphs.
  • determining the target feature set according to the connectivity graph including:
  • the core nodes in each connected graph are respectively determined, and the features corresponding to the core nodes are recorded in the target feature set.
  • the core nodes in each connected graph are respectively determined, including:
  • the core nodes in each connected graph are determined according to any one of the node degree value, the median value, and the Kcore coefficient of each node.
  • an unsupervised feature selection apparatus including:
  • a similarity calculation unit for calculating a similarity between all features of the sample data
  • a topology map construction unit configured to construct a feature topology map according to the similarity between the nodes by using the feature as a node
  • a connectivity graph mining unit configured to segment the feature topology map to obtain one or more connected graphs
  • the target feature selection unit is configured to determine the target feature set according to the connected graph.
  • the topology map construction unit is specifically configured to:
  • the connectivity graph mining unit includes: a first mining module
  • the first mining module is configured to:
  • the nodes connected by the edges are divided into the same connected graph to obtain one or more connected graphs.
  • the connectivity graph mining unit includes: a second mining module
  • the second mining module is configured to:
  • the nodes with the same label are divided into the same connected graph to obtain one or more connected graphs.
  • the target feature selection unit is specifically configured to:
  • the core nodes in each connected graph are respectively determined, and the features corresponding to the core nodes are recorded in the target feature set.
  • the target feature selection unit is specifically configured to: according to the node degree value, the median value, and the Kcore coefficient of each node Any one of them determines the core node in each connected graph.
  • the embodiment of the present application completes the analysis of the sample data based on the knowledge of the graph theory.
  • the feature topological map is constructed according to the similarity between the features, and then the feature topological map is segmented to make the feature with higher similarity.
  • Nodes are divided into the same connected graph to implement feature clustering of sample data, so that it cannot be predetermined
  • the characteristics of the sample data may also be selected in the scene of the classification of the sample data; after the feature clustering is completed through the connected graph, a node is further selected from each connected graph, and the corresponding feature of the node is representative.
  • the target feature set is recorded, thereby obtaining a comprehensive and non-repeating representative feature corresponding to the entire sample data.
  • the embodiment does not need to rely on the classification indication of the sample data, can realize unsupervised feature selection, and ensures that two or more similar features do not appear in the target feature set, so that the target feature set can be more intuitively described.
  • Sample data to avoid information redundancy.
  • FIG. 2 is a schematic diagram of a principle of mining a connected graph in an unsupervised feature selection method provided by an embodiment of the present application
  • FIG. 3 is a schematic diagram of a method for segmenting a feature topology diagram in an unsupervised feature selection method provided by an embodiment of the present application
  • FIG. 4 is a structural block diagram of an unsupervised feature selection apparatus according to an embodiment of the present application.
  • FIG. 1 A flow chart of an unsupervised feature selection method is shown in FIG. As shown in Figure 1, the method includes the following steps.
  • the embodiment of the present application can calculate the similarity between two features according to any one of the Euclidean distance, the cosine distance, and the Jaccard similarity coefficient.
  • the Euclidean distance represents the absolute distance between two points in space. Assuming that the two features are X and Y respectively, it can be regarded as two points in the multidimensional space, and the coordinates are X(x 1 , x 2 ,%) and Y(y 1 , y 2 ,7), respectively. Then its Euclidean distance calculation formula is:
  • the calculation result Eu(X, Y) can be used as the similarity between the features X and Y; where x i represents the coordinate of the feature i in the i-dimensional dimension of the multidimensional coordinate system, and y i represents the feature i in the multidimensional coordinate system.
  • the coordinates of the dimension, i 1, 2, ....
  • the cosine distance also known as the cosine similarity, is a measure of the magnitude of the difference between two individuals using the cosine of the angle between the two vectors in the vector space.
  • the cosine distance of the features X and Y, ie the vector corresponding to X and Y with The cosine of the included angle is calculated as:
  • the result of calculation cos(X, Y) can be used as the similarity between the features X and Y.
  • the cosine distance pays more attention to the difference in direction between the two features.
  • the ratio of the Jaccard similarity coefficient, the number of intersection elements of the two sets to the number of union elements of the two sets, is called the Jaccard coefficient of the two sets; for the features X and Y, the corresponding Vector with For multidimensional vectors, the values of each dimension are treated as one element, and the values of each dimension of the multidimensional vector form a set. Therefore, the features X and Y can also be regarded as two sets, and then the Jaccard similarity coefficient can be calculated. for: The calculation result J(X, Y) can be used as the similarity between the features X and Y.
  • the embodiment of the present application may also use other methods to calculate the similarity between two features. In actual applications, different ones may be selected according to specific scenarios. Formula to calculate the similarity between features.
  • the feature is a node, and the feature topology map is constructed according to the similarity between the nodes.
  • any two nodes are connected by a path (both directly connected by one edge, and indirectly connected by at least one node and at least two edges), it is called a connected graph.
  • the connection relationship between the nodes in the same connected graph is closer to the nodes outside the connected graph, that is, the similarity between the features corresponding to the nodes in the same connected graph is more High, that is, step S13 actually aggregates the nodes corresponding to the similar features into the same connected graph, and implements feature clustering on the sample data, so that the classification label of the sample data does not need to be predicted.
  • step S14 of the embodiment may select a node from each connected graph, and the corresponding feature of the node is used as a representative feature of the feature set corresponding to the connected graph. , the target feature set is recorded, so that each type of feature in the target feature set contains only one, and the target feature set is guaranteed. The difference between features.
  • the embodiment of the present application completes the analysis of the sample data based on the knowledge of the graph theory. Firstly, the feature topological map is constructed according to the similarity between the features, and then the feature topological map is segmented to make the feature with higher similarity.
  • the nodes are divided into the same connected graph to realize feature clustering of the sample data, so that the characteristics of the sample data can also be selected in the scenario where the classification of the sample data cannot be determined in advance; after the feature clustering is completed through the connected graph Further, a node is selected from each connected graph, and the corresponding feature of the node is represented as a representative feature, and is recorded in the target feature set, thereby obtaining a comprehensive and non-repeating representative feature corresponding to the entire sample data. Therefore, the embodiment does not need to rely on the classification indication of the sample data, can realize unsupervised feature selection, and ensures that two or more similar features do not appear in the target feature set, so that the target feature set can be more intuitively described. Sample data to avoid information redundancy.
  • the feature topology map is constructed according to the similarity between the nodes, as described in step S12, and specifically includes:
  • the process of constructing the feature top map corresponding to the eight features is: after the similarity calculation, the similarity between the two features labeled 1 and 2 is 0.7, and the corresponding nodes are connected by one edge. The similarity between the two features labeled 3 and 7 is 0.2, then the corresponding nodes are connected by one edge... and so on, the feature topology diagram shown in Fig. 2 can be obtained.
  • the similarity between the features it can be set that the smaller the similarity, the larger the corresponding side length, so that the distance between the corresponding two nodes is larger (for example, in Figure 2, due to 0.7> 0.2, so the edge between the labels 1 and 2 is shorter than the edge between the labels 3 and 7, so that the degree of similarity of the corresponding features can be determined according to the degree of node density, that is, the more dense between the nodes, the corresponding features The higher the similarity.
  • determining the target feature set according to the connectivity diagram may specifically include:
  • the core nodes in each connected graph are respectively determined, and the features corresponding to the core nodes are recorded in the target feature set.
  • the feature top map corresponding to the eight features is segmented, and two connected graphs (the dotted circles) with the labels A and B respectively can be obtained; further, the connected graph can be determined in step S14.
  • the core node of A is node 3
  • the core node of connection graph B is node 7, and the features corresponding to core nodes 3 and 7 are recorded as representative features in the target feature set, that is, the features corresponding to core node 3 are nodes 1 to 4.
  • the feature corresponding to the core node 7 is the representative feature of the four features corresponding to the nodes 5 to 8.
  • the method for determining the core node of each connected graph is multiple, and at least includes: determining the core node according to any one of a Node Degree, a median value, and a Kcore coefficient of each node; Different determination methods can be selected according to different application scenarios.
  • the node degree value of a node refers to the number of edges associated with the node in the topology map; correspondingly, the node with the largest node degree value may be selected as the core node.
  • the median value of a node refers to the ratio between the number of shortest paths passing through the node in the topology graph and the total number of shortest paths in the topology map; correspondingly, the node with the largest median value can be selected as the core node.
  • the Kcore coefficients of each node on the connected graph structure are calculated.
  • the edge nodes in the connected graph are filtered, and the points of the relative core positions in the connected graph are found and their associations are found.
  • the step of determining the core node according to the node degree value is: in the connected graph A, the node 3 has a total of 4 edges, that is, the node 3 has a node degree of 4, and the connected graph A
  • the node degrees of other nodes are all less than 4, so it can be determined that node 3 is the core node of connected graph A; correspondingly, the corresponding feature of node 3 is taken as the representative feature of connected graph A, and is recorded in the target feature set.
  • the present application clusters and selects sample data features based on graph theory knowledge, and does not need to rely on the classification labeling of sample data (ie, no supervision), and the difference of each representative feature in the final target feature set is compared. Large, there will be no information redundancy.
  • e-commerce data for the user characteristics such as “user's star rating”, “15-day browsing times”, “15-day purchases”, and “15-day total purchase amount”, “15 days can be identified by this embodiment.
  • the number of views, the number of purchases in 15 days, and the total amount of purchases in 15 days are similar features, and there are at most one of the three in the target feature set.
  • step S13 The segmentation of the feature topological map described in the above step S13 is actually implementing feature clustering, that is, achieving unsupervised classification of all features.
  • the specific implementation method of step S13 may be various.
  • the feature topology map may be directly segmented according to the similarity between the features, and the specific steps are as follows:
  • S1311 respectively compare the similarity corresponding to each edge in the feature topology graph with a preset similarity threshold
  • S1313 The nodes connected by the edge are divided into the same connected graph to obtain one or more connected graphs.
  • the similarity between the two features ranges from [0, 1], and the similarity of 0 indicates that the two features are completely dissimilar.
  • a similarity of 1 indicates that the two features are identical.
  • the preset similarity threshold may be set according to a specific application scenario, and the higher the preset similarity threshold, the more dispersed the node distribution in the feature topology graph, so that the The more connected graphs and the smaller the scale of each connected graph, the more representative features can be obtained in the target feature set.
  • the lower the preset similarity threshold the denser the node distribution in the feature topology graph. The fewer the number of connected graphs mined and the larger the scale of each connected graph, the fewer the target feature set will be.
  • the preset similarity threshold should not be less than 0.5.
  • the feature topography corresponding to the eight features labeled 1 to 8 is segmented, and the preset similarity threshold is set to 0.5, and the similarity of each edge in the feature topology is compared with 0.5. , delete the edge with the similarity less than 0.5, such as the edge between nodes 3 and 7 (indicated by the dotted line), so that two connected graphs A and B can be obtained.
  • step S13 may further implement segmentation of the feature topology map based on the label propagation algorithm, and the specific steps are as follows:
  • S1321 Set a unique label for each node in the feature topology map; specifically, the node ID may be used as its label.
  • S1322 Propagating a label of each node to a corresponding neighbor node.
  • Step S1324 Determine whether the label of each node changes (that is, determine whether there is a node whose label changes). If there is a change of the label of at least one node, return to step S1322, and if the labels of all nodes have not changed, execute Step S1325.
  • the label with the smallest value selected in the above step S1323 is a new label of the node.
  • the label with the largest value may also be selected, as long as the selection criteria adopted by all nodes in the same feature topology map are consistent.
  • the above label propagation algorithm is especially suitable for scenarios where the feature topology map is relatively simple, and the label propagation and comparison workload are not too large, and it is simple and easy.
  • a community discovery algorithm may also be used to complete segmentation (feature clustering) of the feature topology map to obtain a connected graph.
  • the embodiment of the present application further provides a computer storage medium, such as a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device.
  • the computer storage medium stores a program.
  • the program in the storage medium is executed by a corresponding processor of the data analysis platform, the data analysis platform is enabled to perform some or all of the steps of the unsupervised feature selection method described in the above method embodiments.
  • the embodiment of the present application further provides an unsupervised feature selection device.
  • 4 is a structural block diagram of the apparatus, specifically including: a similarity calculation unit 301, a topology map construction unit 302, a connectivity diagram mining unit 303, and a target feature selection unit 304.
  • the similarity calculation unit 301 is configured to calculate the similarity between all the features of the sample data.
  • the topology map construction unit 302 is configured to construct a feature topology map according to the similarity between the nodes by using the feature as a node.
  • the connected graph mining unit 303 is configured to segment the feature topology map to obtain one or more connected graphs.
  • the target feature selection unit 304 is configured to determine a target feature set according to the connected graph.
  • the feature topological map is firstly constructed according to the similarity between the features, and then the feature topological map is segmented, so that the feature nodes with higher similarity are divided into the same connected graph, and the pair is implemented.
  • Characteristic clustering of sample data so that the characteristics of the sample data can also be selected in a scene in which the classification of the sample data cannot be predetermined; after the feature clustering is completed through the connected graph, one of each connected graph is further selected.
  • the node, and the feature corresponding to the node is a representative feature, and is recorded in the target feature set, thereby obtaining a comprehensive and non-repeating representative feature corresponding to the entire sample data.
  • the embodiment does not need to rely on the classification indication of the sample data, can realize unsupervised feature selection, and ensures that two or more similar features do not appear in the target feature set, so that the target feature set can be more intuitively described.
  • Sample data to avoid information redundancy.
  • the topology map construction unit 302 is specifically configured to: add an edge between two nodes whose degree of similarity is greater than zero, to obtain the feature topology map.
  • the connectivity graph mining unit 303 may include a first mining module;
  • the first mining module is configured to: compare the similarity corresponding to each edge in the feature topology graph with a preset similarity threshold, and delete the similarity in the feature topology map to be less than the preset similarity
  • the edge of the threshold the nodes connected by the edge are divided into the same connected graph to obtain one or more connected graphs.
  • the connectivity graph mining unit 303 may also include a second mining module; the second mining module is specifically configured to: set a unique label for each node in the feature topology map. Transmitting the label of each node to the corresponding neighbor node; for each node, comparing its own label with all received labels from the neighbor, selecting the label with the smallest value as the new label of the corresponding node; Whether the label of the node changes; if there is a change of the label of at least one node, returning to perform the step of propagating the label of each node to the corresponding neighbor node; if all the labels of the node have not changed, the label is The same nodes are divided into the same connected graph to obtain one or more connected graphs.
  • the target feature selection unit 304 is specifically configured to: respectively determine core nodes in each connected graph, and record features corresponding to the respective core nodes into the target feature set.
  • the target feature selecting unit is specifically configured to: determine, in each connected graph, according to any one of a node degree value, a median value, and a Kcore coefficient of each node. Core node.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)

Abstract

本申请公开一种无监督的特征选择方法、装置，其首先根据特征之间的相似度构建特征拓扑图，然后将该特征拓扑图进行分割，使得相似度较高的特征节点划分在同一个连通图中，实现对样本数据的特征聚类，从而在无法预先确定样本数据的分类标示的场景下也可以对样本数据的特征进行选择；在通过连通图完成特征聚类后，进一步可以从每个连通图中选出一个节点，并将该节点对应的特征为代表特征，记入目标特征集，从而得到整个样本数据对应的全面且不重复的代表特征。因此，本申请不需要依赖样本数据的分类标示，可以实现无监督的特征选择，且保证目标特征集中不会出现两个或两个以上相类似的特征，使得目标特征集可以更直观地描述样本数据，避免信息冗余。

Description

无监督的特征选择方法、装置
本申请要求2016年04月05日递交的申请号为201610205887.7、发明名称为“无监督的特征选择方法、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
本申请涉及计算机数据处理技术领域，尤其涉及无监督的特征选择方法、装置。
背景技术
目前，基于大数据分析的应用不断涌现，如信息检索、挖掘用户的消费倾向、消费模式等；通过对样本数据进行分析，提取出能够描述样本数据的属性信息，即样本的特征。例如，从某商品的网络销售数据中，可以提取出购买用户等级、购买次数、购买金额等特征。一般的，样本数据都具有多个特征，但未必每个特征都对数据的分析有用处，因此需要从大量特征中选出有效的特征，以减少特征集中特征的个数，进而减小具体应用中基于特征集的计算过程的复杂度，提高预测精度。
针对样本数据的特征选择问题，相关技术中已经给出多种实现方法，这些实现方法大多都是需要依赖样本数据分类标示的有监督方式，即各条样本数据的类别应当为已知，从而根据这些已知类别调整特征选择模型的参数，以保证特征选择的准确率。
但是有些应用场景下样本数据不具备分类标示，从而难以通过现有特征选择方法进行特征选择。另外，应用现有特征选择方法会出现所选出的两个或两个以上特征所代表的信息相似的现象，即最终得到的特征集存在信息冗余。
发明内容
为克服相关技术中存在的问题，本申请提供一种无监督的特征选择方法、装置。
本申请第一方面，提供一种无监督的特征选择方法，包括：
计算样本数据的所有特征之间的相似度；
以特征为节点，根据节点之间的相似度构建特征拓扑图；
对所述特征拓扑图进行分割，得到一个或多个连通图；
根据所述连通图，确定目标特征集。
结合第一方面，在第一方面第一种可行的实施方式中，根据节点之间的相似度构建 特征拓扑图，包括：
在任意相似度大于零的两个节点之间添加一条边，得到所述特征拓扑图。
结合第一方面，或者第一方面第一种可行的实施方式，在第一方面第二种可行的实施方式中，对所述特征拓扑图进行分割，得到一个或多个连通图，包括：
分别将所述特征拓扑图中每条边对应的相似度与预设相似度阈值进行比较；
删除所述特征拓扑图中相似度小于所述预设相似度阈值的边；
将通过边连接的节点划分在同一个连通图中，得到一个或多个连通图。
结合第一方面，或者第一方面第一种可行的实施方式，在第一方面第三种可行的实施方式中，对所述特征拓扑图进行分割，得到一个或多个连通图，包括：
为所述特征拓扑图中的每个节点设置唯一的标签；
将每个节点的标签传播至相应的邻居节点；
对于每个节点，将自己的标签和接收到的所有来自邻居的标签进行比较，选其中值最小的标签作为相应节点的新标签；
判断各个节点的标签是否发生变化；
如果存在至少一个节点的标签发生变化，则返回执行所述将每个节点的标签传播至相应的邻居节点的步骤；
如果所有节点的标签均未发生变化，则将标签相同的节点划分在同一个连通图中，得到一个或多个连通图。
结合第一方面，或者第一方面第一种可行的实施方式，在第一方面第四种可行的实施方式中，根据所述连通图，确定目标特征集，包括：
分别确定每个连通图中的核心节点，并将各个核心节点对应的特征记入目标特征集。
结合第一方面第四种可行的实施方式，在第一方面第五种可行的实施方式中，分别确定每个连通图中的核心节点，包括：
根据各个节点的节点度值、介数值和Kcore系数中的任意一种确定每个连通图中的核心节点。
本申请第二方面，提供一种无监督的特征选择装置，包括：
相似度计算单元，用于计算样本数据的所有特征之间的相似度；
拓扑图构建单元，用于以特征为节点，根据节点之间的相似度构建特征拓扑图；
连通图挖掘单元，用于对所述特征拓扑图进行分割，得到一个或多个连通图；
目标特征选择单元，用于根据所述连通图，确定目标特征集。
结合第二方面，在第二方面第一种可行的实施方式中，所述拓扑图构建单元具体被配置为：
在任意相似度大于零的两个节点之间添加一条边，得到所述特征拓扑图。
结合第二方面，或者第二方面第一种可行的实施方式，在第二方面第二种可行的实施方式中，所述连通图挖掘单元包括：第一挖掘模块；
所述第一挖掘模块被配置为：
分别将所述特征拓扑图中每条边对应的相似度与预设相似度阈值进行比较；
删除所述特征拓扑图中相似度小于所述预设相似度阈值的边；
将通过边连接的节点划分在同一个连通图中，得到一个或多个连通图。
结合第二方面，或者第二方面第一种可行的实施方式，在第二方面第三种可行的实施方式中，所述连通图挖掘单元包括：第二挖掘模块；
所述第二挖掘模块被配置为：
为所述特征拓扑图中的每个节点设置唯一的标签；
将每个节点的标签传播至相应的邻居节点；
对于每个节点，将自己的标签和接收到的所有来自邻居的标签进行比较，选其中值最小的标签作为相应节点的新标签；
判断各个节点的标签是否发生变化；
如果存在至少一个节点的标签发生变化，则返回执行所述将每个节点的标签传播至相应的邻居节点的步骤；
如果所有节点的标签均未发生变化，则将标签相同的节点划分在同一个连通图中，得到一个或多个连通图。
结合第二方面，或者第二方面第一种可行的实施方式，在第二方面第四种可行的实施方式中，所述目标特征选择单元具体被配置为：
分别确定每个连通图中的核心节点，并将各个核心节点对应的特征记入目标特征集。
结合第二方面第四种可行的实施方式，在第二方面第五种可行的实施方式中，所述目标特征选择单元具体被配置为：根据各个节点的节点度值、介数值和Kcore系数中的任意一种确定每个连通图中的核心节点。
由以上技术方案可知，本申请实施例基于图论知识完成对样本数据的分析，首先根据特征之间的相似度构建特征拓扑图，然后将该特征拓扑图进行分割，使得相似度较高的特征节点划分在同一个连通图中，实现对样本数据的特征聚类，从而在无法预先确定 样本数据的分类标示的场景下也可以对样本数据的特征进行选择；在通过连通图完成特征聚类后，进一步从每个连通图中选出一个节点，并将该节点对应的特征为代表特征，记入目标特征集，从而得到整个样本数据对应的全面且不重复的代表特征。因此，本实施例不需要依赖样本数据的分类标示，可以实现无监督的特征选择，且保证目标特征集中不会出现两个或两个以上相类似的特征，使得目标特征集可以更直观地描述样本数据，避免信息冗余。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，并不能限制本申请。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，示出了符合本发明的实施例，并与说明书一起用于解释本发明的原理。
图1是本申请实施例提供的一种无监督的特征选择方法的流程图；
图2是本申请实施例提供的无监督的特征选择方法中挖掘连通图的原理示意图；
图3是本申请实施例提供的无监督的特征选择方法中分割特征拓扑图的一种方法示意图；
图4是本申请实施例提供的一种无监督的特征选择装置的结构框图。
具体实施方式
这里将详细地对示例性实施例进行说明，其示例表示在附图中。下面的描述涉及附图时，除非另有表示，不同附图中的相同数字表示相同或相似的要素。以下示例性实施例中所描述的实施方式并不代表与本发明相一致的所有实施方式。相反，它们仅是与如所附权利要求书中所详述的、本发明的一些方面相一致的装置和方法的例子。
图1示出的一种无监督的特征选择方法的流程图。如图1所示，该方法包括以下步骤。
S11、计算样本数据的所有特征之间的相似度。
具体的，本申请实施例可以根据欧氏距离、余弦距离和杰卡德相似系数中的任意一种计算两个特征之间的相似度。
其中，欧氏距离表征空间两点之间的绝对距离。假设两个特征分别为X、Y，可以视为多维空间中的两个点，坐标分别为X(x1,x2,...)和Y(y1,y2,...)，则其欧氏距离计 算公式为：
Figure PCTCN2017077997-appb-000001
计算结果Eu(X,Y)即可作为特征X和Y之间的相似度；其中xi表示特征X在多维坐标系中第i维的坐标，yi表示特征Y在多维坐标系中第i维的坐标，i＝1，2,……。
余弦距离，也称为余弦相似度，是用向量空间中两个向量夹角的余弦值作为衡量两个个体间差异的大小的度量。特征X和Y的余弦距离，即X和Y对应的向量
Figure PCTCN2017077997-appb-000002
和
Figure PCTCN2017077997-appb-000003
的夹角余弦值，计算公式为：
Figure PCTCN2017077997-appb-000004
计算结果cos(X,Y)即可作为特征X和Y之间的相似度。相比欧氏距离，余弦距离更加注重两个特征在方向上的差异。
杰卡德相似系数，两个集合的交集元素个数与这两个集合的并集元素个数之间的比值，称为这两个集合的杰卡德系数；对于特征X和Y，其对应的向量
Figure PCTCN2017077997-appb-000005
和
Figure PCTCN2017077997-appb-000006
为多维向量，将每个维度的数值分别视为一个元素，多维向量的各个维度值构成一个集合，故特征X和Y也可以视为两个集合，进而可以计算其杰卡德相似系数，公式为：
Figure PCTCN2017077997-appb-000007
计算结果J(X,Y)即可作为特征X和Y之间的相似度。
当然，除了上述欧氏距离、余弦距离和杰卡德相似系数以为，本申请实施例还可以采用其他方法计算两个特征之间的相似度，实际应用中可以根据具体场景的不同，选择不同的公式来计算特征之间的相似性。
S12、以特征为节点，根据节点之间的相似度构建特征拓扑图。
S13、对所述特征拓扑图进行分割，得到一个或多个连通图。
在拓扑图中，若任意两个节点间均有路径相连(既包括通过一条边直接相连，也包括通过至少一个节点和至少两条边间接相连)，则称为连通图。根据上述建图的逻辑可知，相对于连通图外的节点，处于同一连通图中的各节点之间的连接关系更紧密，即处于同一连通图中的各节点对应的特征之间的相似度更高，也即步骤S13实际是将相似特征对应的节点聚合到同一个连通图中，实现对样本数据的特征聚类，从而不需要预知样本数据的分类标示。
S14、根据所述连通图，确定目标特征集。
由于每个连通图都相当于一类相似特征的集合，故本实施例步骤S14可以从每个连通图中选择一个节点，将该节点对应的特征作为该连通图所对应的特征集合的代表特征，记入目标特征集，使得目标特征集中每种类型的特征仅包含一个，保证目标特征集中的 特征之间的差异性。
由以上技术方案可知，本申请实施例基于图论知识完成对样本数据的分析，首先根据特征之间的相似度构建特征拓扑图，然后将该特征拓扑图进行分割，使得相似度较高的特征节点划分在同一个连通图中，实现对样本数据的特征聚类，从而在无法预先确定样本数据的分类标示的场景下也可以对样本数据的特征进行选择；在通过连通图完成特征聚类后，进一步从每个连通图中选择一个节点，并将该节点对应的特征为代表特征，记入目标特征集，从而得到整个样本数据对应的全面且不重复的代表特征。因此，本实施例不需要依赖样本数据的分类标示，可以实现无监督的特征选择，且保证目标特征集中不会出现两个或两个以上相类似的特征，使得目标特征集可以更直观地描述样本数据，避免信息冗余。
在本申请一个可行的实施例中，步骤S12所述的根据节点之间的相似度构建特征拓扑图，具体可以包括：
在任意相似度大于零的两个节点之间添加一条边，得到所述特征拓扑图。
如图2所示，假设有标号为1至8的八个特征，分别表示某购物网站中某个商品7天浏览次数、14天浏览次数、30天浏览次数、60天浏览次数、7天购买次数、14天购买次数、30天购买次数、60天购买次数。构建这八个特征对应的特征拓扑图的过程为：经过相似度计算得知，标号为1和2的两个特征之间的相似度为0.7，则将二者对应的节点通过一条边连接，标号为3和7的两个特征之间的相似度为0.2，则将二者对应的节点通过一条边连接……依此类推，可以得到图2所示特征拓扑图。另外，为了更直观的展现特征之间的相似度高低，可以设定相似度越小，相应的边长度越大，使得相应的两个节点之间距离越大(例如图2中，由于0.7＞0.2，故标号1和2之间的边比标号3和7之间的边短)，从而可以根据节点密集程度确定相应特征的相似度高低，即节点之间越密集度，相应的特征之间相似度越高。
在本申请一个可行的实施例中，上述步骤S14所述的根据所述连通图，确定目标特征集，具体可以包括：
分别确定每个连通图中的核心节点，并将各个核心节点对应的特征记入目标特征集。
仍参照图2，根据上述步骤S13对八个特征对应的特征拓扑图进行分割，可以得到标号分别为A和B的两个连通图(虚线圈表示)；进而，在步骤S14中可以确定连通图A的核心节点为节点3，连通图B的核心节点为节点7，并将该核心节点3和7对应的特征作为代表特征记入目标特征集，即核心节点3对应的特征为节点1至4对应的四个特 征的代表特征，核心节点7对应的特征为节点5至8对应的四个特征的代表特征。
可选的，上述确定各个连通图的核心节点的方法有多种，至少包括：根据各个节点的节点度值(Node Degree)、介数值和Kcore系数中的任意一种确定所述核心节点；具体可以根据应用场景的不同，选择不同的确定方法。
其中，某个节点的节点度值，指在拓扑图中该节点相关联的边的条数；相应的，可以选择节点度值最大的节点作为核心节点。
某个节点的介数值，指拓扑图中经过该节点的最短路径的数目与拓扑图中最短路径的总数之间的比值；相应的，可以选择介数值最大的节点作为核心节点。
根据KCore算法计算各节点在连通图结构上的Kcore系数，本质上是过滤连通图中的边缘节点，找出连通图中相对核心位置的点并发现他们的关联。
仍以图2所示情境为例，根据节点度值确定核心节点的步骤为：连通图A中，节点3相关联的边共4条，即节点3的节点度值为4，且连通图A中其他节点的节点度值都小于4，故可以确定节点3为连通图A的核心节点；相应的，节点3对应的特征作为连通图A的代表特征，被记入目标特征集。
由以上技术方案可知，本申请基于图论知识对样本数据特征进行聚类、选择，不需要依赖样本数据的分类标示(即无监督)，且最终得到的目标特征集中的各个代表特征差异都较大，不会存在信息冗余。以电商数据为例，对于“用户的星级”、“15天浏览次数”、“15天购买次数”、“15天购买总金额”等用户特征，通过本实施例可以识别出“15天浏览次数”、“15天购买次数”和“15天购买总金额”三者为相似特征，目标特征集中最多存在三者中的一个。
上述步骤S13所述的对特征拓扑图进行分割，实际是实现特征聚类，也即实现对所有特征的无监督分类。实际应用中，步骤S13的具体实施方法可以有多种。在本申请一个可行的实施例中，可以直接根据特征之间的相似度进行特征拓扑图的分割，具体步骤如下：
S1311、分别将所述特征拓扑图中每条边对应的相似度与预设相似度阈值进行比较；
S1312、删除所述特征拓扑图中相似度小于所述预设相似度阈值的边；
S1313、将通过边连接的节点划分在同一个连通图中，得到一个或多个连通图。
具体的，两个特征之间的相似度取值范围为[0,1]，相似度为0表示两个特征完全不相似，相似度为1表示两个特征完全相同。相应的，上述预设相似度阈值可以根据具体应用场景设置，预设相似度阈值越高，特征拓扑图中节点分布越分散，从而可以挖掘出 的连通图个数越多且每个连通图的规模越小，从而使得目标特征集中可以得到更多的代表特征；相反，预设相似度阈值越低，特征拓扑图中节点分布越密集，可以挖掘出的连通图个数越少且每个连通图的规模越大，从而最终得到的目标特征集中代表特征越少。一般的，预设相似度阈值不宜小于0.5。
仍参照图2，对标号为1至8的八个特征对应的特征拓扑图进行分割，假设预设相似度阈值设置为0.5则将特征拓扑图中每条边对应的相似度分别与0.5进行比较，删除相似度小于0.5的边，如节点3和7之间的边(以虚线表示)，从而可以得到A和B两个连通图。
参照图3，在本申请另一个可行的实施例中，步骤S13中还可以基于标签传播算法实现对特征拓扑图的分割，具体步骤如下：
S1321、为所述特征拓扑图中的每个节点设置一个唯一的标签；具体的，可以将节点ID作为其标签。
S1322、将每个节点的标签传播至相应的邻居节点。
S1323、对于每个节点，将自己的标签和接收到的所有来自邻居的标签进行比较，选其中值最小的标签作为相应节点的新标签。
S1324、判断各个节点的标签是否发生变化(即判断是否存在标签发生变化的节点)，如果存在至少一个节点的标签发生变化，则返回至步骤S1322，如果所有节点的标签均未发生变化，则执行步骤S1325。
S1325、将标签相同的节点划分在同一个连通图中，得到一个或多个连通图。
仍参照图2所示情境，对于节点1，将其标签label1分别传播至其邻居节点2和3，并接收邻居节点2、3传播来的标签label2和label3，将label1、label2和label3进行比较，得到的比较结果为label3的值最小，则将节点1的标签由label1更改为label3；其他节点也依照与节点1相同的方式进行标签传播、比较及更新。由于至少节点1对应的标签发生了变化，故返回步骤S1322，继续在八个节点之间进行标签传播、比较及更新；依此循环，直至相对于上一次循环结果，八个节点的标签都无变化，停止循环并执行步骤S1325；例如，假设此时节点1至4对应的标签都为label3，节点5至8对应的标签都为label7，则将标签同为label3的节点1至4划分在同一个连通图A中(实际意义为表征浏览次数的特征都划分在了连通图A中)，标签同为label7的节点5至8划分在同一个连通图B中(实际意义为表征购买次数的特征都划分在了连通图B中)。
需要说明的是，上述步骤S1323中选择值最小的标签作为节点的新标签仅仅是一种 具体实施方式，也可以选择值最大的标签，只要保证同一特征拓扑图中的所有节点采用的选择标准一致即可。
上述标签传播算法尤其适用于特征拓扑图较简单的场景，标签传播、比较工作量不会太大，且简单易行。在本申请其他实施例中，对于较复杂的特征拓扑图，也可以采用社区发现算法完成对特征拓扑图的分割(特征聚类)，得到连通图。
另外，本申请实施例还提供了一种计算机存储介质，例如可以是ROM、随机存取存储器(RAM)、CD-ROM、磁带、软盘和光数据存储设备等；该计算机存储介质中存储有程序，当所述存储介质中的程序由数据分析平台的相应处理器执行时，使得该数据分析平台能够执行上述方法实施例中记载的无监督的特征选择方法的部分或全部步骤。
相应于上述无监督的特征选择方法，本申请实施例还提供了一种无监督的特征选择装置。图4为该装置的一种结构框图，具体包括：相似度计算单元301、拓扑图构建单元302、连通图挖掘单元303和目标特征选择单元304。
其中，该相似度计算单元301用于，计算样本数据的所有特征之间的相似度。
该拓扑图构建单元302用于，以特征为节点，根据节点之间的相似度构建特征拓扑图。
该连通图挖掘单元303用于，对所述特征拓扑图进行分割，得到一个或多个连通图。
该目标特征选择单元304用于，根据所述连通图，确定目标特征集。
由以上装置结构可知，本申请实施例首先根据特征之间的相似度构建特征拓扑图，然后将该特征拓扑图进行分割，使得相似度较高的特征节点划分在同一个连通图中，实现对样本数据的特征聚类，从而在无法预先确定样本数据的分类标示的场景下也可以对样本数据的特征进行选择；在通过连通图完成特征聚类后，进一步从每个连通图中选出一个节点，并将该节点对应的特征为代表特征，记入目标特征集，从而得到整个样本数据对应的全面且不重复的代表特征。因此，本实施例不需要依赖样本数据的分类标示，可以实现无监督的特征选择，且保证目标特征集中不会出现两个或两个以上相类似的特征，使得目标特征集可以更直观地描述样本数据，避免信息冗余。
在本申请一个可行的实施例中，上述拓扑图构建单元302具体被配置为：在任意相似度大于零的两个节点之间添加一条边，得到所述特征拓扑图。
在本申请一个可行的实施例中，上述连通图挖掘单元303可以包括第一挖掘模块； 该第一挖掘模块具体被配置为：分别将所述特征拓扑图中每条边对应的相似度与预设相似度阈值进行比较；删除所述特征拓扑图中相似度小于所述预设相似度阈值的边；将通过边连接的节点划分在同一个连通图中，得到一个或多个连通图。
在本申请另一个可行的实施例中，上述连通图挖掘单元303也可以包括第二挖掘模块；该第二挖掘模块具体被配置为：为所述特征拓扑图中的每个节点设置唯一的标签；将每个节点的标签传播至相应的邻居节点；对于每个节点，将自己的标签和接收到的所有来自邻居的标签进行比较，选其中值最小的标签作为相应节点的新标签；判断各个节点的标签是否发生变化；如果存在至少一个节点的标签发生变化，则返回执行所述将每个节点的标签传播至相应的邻居节点的步骤；如果所有节点的标签均未发生变化，则将标签相同的节点划分在同一个连通图中，得到一个或多个连通图。
在本申请一个可行的实施例中，上述目标特征选择单元304具体被配置为：分别确定每个连通图中的核心节点，并将各个核心节点对应的特征记入目标特征集。
可选的，为确定每个连通图中的核心节点，所述目标特征选择单元具体被配置为：根据各个节点的节点度值、介数值和Kcore系数中的任意一种确定每个连通图中的核心节点。
关于上述实施例中的装置，其中各个单元执行操作的具体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。
本领域技术人员在考虑说明书及实践这里公开的发明后，将容易想到本发明的其它实施方案。本申请旨在涵盖本发明的任何变型、用途或者适应性变化，这些变型、用途或者适应性变化遵循本发明的一般性原理并包括本申请未公开的本技术领域中的公知常识或惯用技术手段。说明书和实施例仅被视为示例性的，本发明的真正范围和精神由下面的权利要求指出。
应当理解的是，本发明并不局限于上面已经描述并在附图中示出的精确结构，并且可以在不脱离其范围进行各种修改和改变。本发明的范围仅由所附的权利要求来限制。

Claims (12)

  1. 一种无监督的特征选择方法，其特征在于，包括：
    计算样本数据的所有特征之间的相似度；
    以特征为节点，根据节点之间的相似度构建特征拓扑图；
    对所述特征拓扑图进行分割，得到一个或多个连通图；
    根据所述连通图，确定目标特征集。
  2. 根据权利要求1所述的方法，其特征在于，根据节点之间的相似度构建特征拓扑图，包括：
    在任意相似度大于零的两个节点之间添加一条边，得到所述特征拓扑图。
  3. 根据权利要求1或2所述的方法，其特征在于，对所述特征拓扑图进行分割，得到一个或多个连通图，包括：
    分别将所述特征拓扑图中每条边对应的相似度与预设相似度阈值进行比较；
    删除所述特征拓扑图中相似度小于所述预设相似度阈值的边；
    将通过边连接的节点划分在同一个连通图中，得到一个或多个连通图。
  4. 根据权利要求1或2所述的方法，其特征在于，对所述特征拓扑图进行分割，得到一个或多个连通图，包括：
    为所述特征拓扑图中的每个节点设置唯一的标签；
    将每个节点的标签传播至相应的邻居节点；
    对于每个节点，将自己的标签和接收到的所有来自邻居的标签进行比较，选其中值最小的标签作为相应节点的新标签；
    判断各个节点的标签是否发生变化；
    如果存在至少一个节点的标签发生变化，则返回执行所述将每个节点的标签传播至相应的邻居节点的步骤；
    如果所有节点的标签均未发生变化，则将标签相同的节点划分在同一个连通图中，得到一个或多个连通图。
  5. 根据权利要求1或2所述的方法，其特征在于，根据所述连通图，确定目标特 征集，包括：
    分别确定每个连通图中的核心节点，并将各个核心节点对应的特征记入目标特征集。
  6. 根据权利要求5所述的方法，其特征在于，分别确定每个连通图中的核心节点，包括：
    根据各个节点的节点度值、介数值和Kcore系数中的任意一种确定每个连通图中的核心节点。
  7. 一种无监督的特征选择装置，其特征在于，包括：
    相似度计算单元，用于计算样本数据的所有特征之间的相似度；
    拓扑图构建单元，用于以特征为节点，根据节点之间的相似度构建特征拓扑图；
    连通图挖掘单元，用于对所述特征拓扑图进行分割，得到一个或多个连通图；
    目标特征选择单元，用于根据所述连通图，确定目标特征集。
  8. 根据权利要求7所述的装置，其特征在于，所述拓扑图构建单元具体被配置为：
    在任意相似度大于零的两个节点之间添加一条边，得到所述特征拓扑图。
  9. 根据权利要求7或8所述的装置，其特征在于，所述连通图挖掘单元包括：第一挖掘模块；
    所述第一挖掘模块被配置为：
    分别将所述特征拓扑图中每条边对应的相似度与预设相似度阈值进行比较；
    删除所述特征拓扑图中相似度小于所述预设相似度阈值的边；
    将通过边连接的节点划分在同一个连通图中，得到一个或多个连通图。
  10. 根据权利要求7或8所述的装置，其特征在于，所述连通图挖掘单元包括：第二挖掘模块；
    所述第二挖掘模块被配置为：
    为所述特征拓扑图中的每个节点设置唯一的标签；
    将每个节点的标签传播至相应的邻居节点；
    对于每个节点，将自己的标签和接收到的所有来自邻居的标签进行比较，选其中值最小的标签作为相应节点的新标签；
    判断各个节点的标签是否发生变化；
    如果存在至少一个节点的标签发生变化，则返回执行所述将每个节点的标签传播至相应的邻居节点的步骤；
    如果所有节点的标签均未发生变化，则将标签相同的节点划分在同一个连通图中，得到一个或多个连通图。
  11. 根据权利要求7或8所述的装置，其特征在于，所述目标特征选择单元具体被配置为：
    分别确定每个连通图中的核心节点，并将各个核心节点对应的特征记入目标特征集。
  12. 根据权利要求11所述的装置，其特征在于，为确定每个连通图中的核心节点，所述目标特征选择单元具体被配置为：
    根据各个节点的节点度值、介数值和Kcore系数中的任意一种确定每个连通图中的核心节点。
PCT/CN2017/077997 2016-04-05 2017-03-24 无监督的特征选择方法、装置 Ceased WO2017173929A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201610205887.7A CN107291760A (zh) 2016-04-05 2016-04-05 无监督的特征选择方法、装置
CN201610205887.7 2016-04-05

Publications (1)

Publication Number Publication Date
WO2017173929A1 true WO2017173929A1 (zh) 2017-10-12

Family

ID=60000861

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/077997 Ceased WO2017173929A1 (zh) 2016-04-05 2017-03-24 无监督的特征选择方法、装置

Country Status (3)

Country Link
CN (1) CN107291760A (zh)
TW (1) TW201738774A (zh)
WO (1) WO2017173929A1 (zh)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110362603A (zh) * 2018-04-04 2019-10-22 北京京东尚科信息技术有限公司 一种特征冗余性分析方法、特征选择方法和相关装置
CN111325246A (zh) * 2020-02-06 2020-06-23 腾讯科技（深圳）有限公司 区域选择方法、装置、计算机设备及存储介质
CN111401392A (zh) * 2019-01-02 2020-07-10 中国移动通信有限公司研究院 聚类集成方法及装置、电子设备及存储介质
CN111597373A (zh) * 2020-05-19 2020-08-28 清华大学 基于卷积神经网络和连通图的图片归类方法及相关设备
CN112132444A (zh) * 2020-09-18 2020-12-25 北京信息科技大学 一种互联网+环境下文化创新企业知识缺口的识别方法
CN112925800A (zh) * 2021-02-26 2021-06-08 平安普惠企业管理有限公司 数据依赖判断方法、装置、计算机设备及存储介质
CN114187457A (zh) * 2021-12-16 2022-03-15 北京大学 一种迭代式图对齐方法
CN114428871A (zh) * 2021-12-20 2022-05-03 深圳云天励飞技术股份有限公司 图像聚类方法、装置、计算机设备及存储介质
CN114912505A (zh) * 2022-03-15 2022-08-16 国能四川华蓥山发电有限公司 一种锅炉燃烧中煤质识别的装置与方法
CN116883758A (zh) * 2023-07-20 2023-10-13 海南大学 基于多模态特征相似性的脑部疾病分类系统、方法、设备
CN119048843A (zh) * 2024-10-28 2024-11-29 南昌大学 一种基于图拓扑学习的ai生成图像检测方法
CN119358182A (zh) * 2024-10-23 2025-01-24 国网黑龙江省电力有限公司伊春供电公司 一种基于图论的电网拓扑建模方法、计算机设备及系统
CN121051414A (zh) * 2025-10-29 2025-12-02 乐山永鑫科技有限责任公司 基于深度学习的负载流量数据识别方法及系统

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111340082B (zh) * 2020-02-19 2024-08-13 深圳市商汤科技有限公司 数据处理方法及装置、处理器、电子设备、存储介质
CN111639189B (zh) * 2020-04-29 2023-03-21 西北工业大学 一种基于文本内容特征的文本图构建方法
CN112364902B (zh) * 2020-10-30 2022-11-15 太原理工大学 一种基于自适应相似性的特征选择学习方法
CN116719994B (zh) * 2023-06-09 2025-09-30 深圳思谋信息科技有限公司 媒体对象推荐方法、装置、计算机设备及存储介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101470731A (zh) * 2007-12-26 2009-07-01 中国科学院自动化研究所 一种可个性化定制的网页过滤方法
US20130177244A1 (en) * 2012-01-09 2013-07-11 General Electric Company Image concealing via efficient feature selection
CN103853831A (zh) * 2014-03-10 2014-06-11 中国电子科技集团公司第二十八研究所 一种基于用户兴趣的个性化搜索实现方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN100495405C (zh) * 2007-01-10 2009-06-03 复旦大学 大规模文本逐次二分的层次聚类方法
US8683389B1 (en) * 2010-09-08 2014-03-25 The New England Complex Systems Institute, Inc. Method and apparatus for dynamic information visualization
CN104035983A (zh) * 2014-05-29 2014-09-10 西安理工大学 基于属性权重相似度的分类变量聚类方法
CN104504264B (zh) * 2014-12-08 2017-09-01 深圳市华傲数据技术有限公司 虚拟人建立方法及装置

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101470731A (zh) * 2007-12-26 2009-07-01 中国科学院自动化研究所 一种可个性化定制的网页过滤方法
US20130177244A1 (en) * 2012-01-09 2013-07-11 General Electric Company Image concealing via efficient feature selection
CN103853831A (zh) * 2014-03-10 2014-06-11 中国电子科技集团公司第二十八研究所 一种基于用户兴趣的个性化搜索实现方法

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
MITRA, P.: "Unsupervised Feature Selection Using Feature Similarity", IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, vol. 24, no. 3, 31 March 2002 (2002-03-31), pages 301 - 312, XP011094210, ISSN: 0162-8828, DOI: doi:10.1109/34.990133 *
XIA ET AL: "Applied research of node similarity label propagation in social networks", COMPUTER ENGINEERING AND APPLICATIONS, no. 14, 31 July 2014 (2014-07-31), ISSN: 1002-8331 *
ZHANG: "Research and Application of Large Scale Machine Learning Theory", ELECTRONIC TECHNOLOGY & INFORMATION SCIENCE, no. 12, 15 December 2013 (2013-12-15), ISSN: 1674-022X *

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110362603A (zh) * 2018-04-04 2019-10-22 北京京东尚科信息技术有限公司 一种特征冗余性分析方法、特征选择方法和相关装置
CN111401392A (zh) * 2019-01-02 2020-07-10 中国移动通信有限公司研究院 聚类集成方法及装置、电子设备及存储介质
CN111401392B (zh) * 2019-01-02 2023-05-09 中国移动通信有限公司研究院 聚类集成方法及装置、电子设备及存储介质
CN111325246B (zh) * 2020-02-06 2023-04-18 腾讯科技（深圳）有限公司 区域选择方法、装置、计算机设备及存储介质
CN111325246A (zh) * 2020-02-06 2020-06-23 腾讯科技（深圳）有限公司 区域选择方法、装置、计算机设备及存储介质
CN111597373A (zh) * 2020-05-19 2020-08-28 清华大学 基于卷积神经网络和连通图的图片归类方法及相关设备
CN111597373B (zh) * 2020-05-19 2023-06-20 清华大学 基于卷积神经网络和连通图的图片归类方法及相关设备
CN112132444A (zh) * 2020-09-18 2020-12-25 北京信息科技大学 一种互联网+环境下文化创新企业知识缺口的识别方法
CN112132444B (zh) * 2020-09-18 2023-05-12 北京信息科技大学 一种互联网+环境下文化创新企业知识缺口的识别方法
CN112925800A (zh) * 2021-02-26 2021-06-08 平安普惠企业管理有限公司 数据依赖判断方法、装置、计算机设备及存储介质
CN114187457A (zh) * 2021-12-16 2022-03-15 北京大学 一种迭代式图对齐方法
CN114428871A (zh) * 2021-12-20 2022-05-03 深圳云天励飞技术股份有限公司 图像聚类方法、装置、计算机设备及存储介质
CN114428871B (zh) * 2021-12-20 2025-10-31 深圳云天励飞技术股份有限公司 图像聚类方法、装置、计算机设备及存储介质
CN114912505A (zh) * 2022-03-15 2022-08-16 国能四川华蓥山发电有限公司 一种锅炉燃烧中煤质识别的装置与方法
CN116883758A (zh) * 2023-07-20 2023-10-13 海南大学 基于多模态特征相似性的脑部疾病分类系统、方法、设备
CN116883758B (zh) * 2023-07-20 2025-10-24 海南大学 基于多模态特征相似性的脑部疾病分类系统、方法、设备
CN119358182A (zh) * 2024-10-23 2025-01-24 国网黑龙江省电力有限公司伊春供电公司 一种基于图论的电网拓扑建模方法、计算机设备及系统
CN119048843A (zh) * 2024-10-28 2024-11-29 南昌大学 一种基于图拓扑学习的ai生成图像检测方法
CN121051414A (zh) * 2025-10-29 2025-12-02 乐山永鑫科技有限责任公司 基于深度学习的负载流量数据识别方法及系统
CN121051414B (zh) * 2025-10-29 2026-02-03 乐山永鑫科技有限责任公司 基于深度学习的负载流量数据识别方法及系统

Also Published As

Publication number Publication date
TW201738774A (zh) 2017-11-01
CN107291760A (zh) 2017-10-24

Similar Documents

Publication Publication Date Title
WO2017173929A1 (zh) 无监督的特征选择方法、装置
Chen et al. Density level sets: Asymptotics, inference, and visualization
Post et al. The state of the art in flow visualisation: Feature extraction and tracking
Panagiotakis et al. Interactive image segmentation based on synthetic graph coordinates
CN104573130B (zh) 基于群体计算的实体解析方法及装置
CN105354578B (zh) 一种多目标物体图像匹配方法
TW201504829A (zh) 圖像搜尋、獲取圖像文字資訊的方法及裝置
KR102039244B1 (ko) 반딧불 알고리즘을 이용한 데이터 클러스터링 방법 및 시스템
Cai et al. Spatial outlier detection based on iterative self-organizing learning model
Ren et al. Saliency integration driven by similar images
Schulze et al. Sets of globally optimal stream surfaces for flow visualization
Yin et al. Study and application of improved level set method with prior graph cut in PCB image segmentation
CN104835174B (zh) 基于超图模式搜索的鲁棒模型拟合方法
CN103955524A (zh) 一种基于超图模型的与事件相关的社会化图像查找算法
CN113763061B (zh) 相似物品聚合的方法和装置
WO2020147259A1 (zh) 一种用户画像方法、装置、可读存储介质及终端设备
CN108804635A (zh) 一种基于属性选择的相似性度量方法
CN103793504B (zh) 一种基于用户偏好与项目属性的聚类初始点选择方法
CN103065302A (zh) 一种基于离群数据挖掘的图像显著性检测方法
CN105162648A (zh) 基于骨干网络扩展的社团检测方法
CN108427730A (zh) 一种基于随机游走和条件随机场的社会标签推荐方法
CN115757673B (zh) 一种网络地图文本标注投放的适宜性评估方法
CN113852629B (zh) 基于自然邻的自适应加权核密度的网络连接异常识别方法及计算机存储介质
Hu et al. Multi-level trajectory learning for traffic behavior detection and analysis
CN104123382A (zh) 一种社会媒体下的图像集摘要生成算法

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

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

Ref document number: 17778593

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 17778593

Country of ref document: EP

Kind code of ref document: A1