CN114692011A - Social media data feature selection method fusing L1 regularization and link attributes - Google Patents

Abstract

The invention belongs to the field of data analysis, and particularly relates to a social media data feature selection method fusing L1 regularization and link attributes, which comprises the following steps of S1, inputting social media data, wherein a behavior sample of the social media data is listed as a feature corresponding to social content; s2, normalizing the social media data; s3, extracting the link relations of 4 common social media data; s4, combining with L1 regularization to realize feature selection under corresponding link relation; and S5, acquiring and outputting the feature subsets obtained by different link relations to obtain a final feature set. The social media data feature selection with the link attribute is realized, the problem of social media data association characteristics which cannot be solved by a traditional feature selection method is solved, and a solid foundation is laid for follow-up operations such as dimension reduction and important feature analysis of large-scale social media data.

Description

A Feature Selection Method for Social Media Data Fusing L1 Regularization and Link Attributes

technical field

The invention belongs to the field of data analysis, and in particular relates to a social media data feature selection method integrating L1 regularization and link attributes.

Background technique

The current development of countless social media services enables people to communicate and express themselves conveniently and easily, such as Weibo. The widespread use of social media has generated massive data at an unprecedented speed. For example, hundreds of millions of microblogs are sent and reposted every day. Massive, high-dimensional social media data poses new challenges for data mining tasks such as classification and clustering. Feature selection is widely used in high-dimensional data mining. Traditional feature selection methods, such as L1 regularization, aim to select relevant features from high-dimensional data to obtain concise and accurate data representation, which can reduce dimensionality. Disaster, speed up the learning process and improve the generalization ability of the learned model. Social media data is mainly composed of (1) traditional high-dimensional attribute value data (such as posts, comments, and images, etc.) and (2) link data describing the relationship between social media users and the person who posted the post, etc. The characteristic of social media data brings new challenges to feature selection, and traditional feature selection methods cannot take advantage of the additional information in linked data. In addition, the nature of social media also determines that its data is massive, noisy and incomplete, which makes the already challenging feature selection problem for social media link data even more severe.

Technical problem: The vast majority of existing feature selection algorithms use "flat" data (samples are represented by the values of multiple features or attributes), which contain uniform attribute value data points and are generally considered to be independent and identically distributed. Considering the attributes of link data in social media, the present invention realizes feature selection for social media data with link attributes by extracting distinguishing relationships from link data and incorporating them into traditional feature selection methods.

SUMMARY OF THE INVENTION

In view of the above problems, the present invention provides a feature selection method for social media data that integrates L1 regularization and link attributes.

The social media data feature selection method that integrates L1 regularization and link attributes disclosed in the present application includes the following steps:

S1. Input social media data, and its behavior samples are listed as the characteristics of the corresponding social content;

S2. Normalized representation of social media data;

S3. Extract the link relationship of 4 common social media data;

S4, combined with L1 regularization to realize feature selection under the corresponding link relationship;

S5. Take and output feature subsets obtained from different link relationships to obtain a final feature set.

Further, in the step S1, normalized representation is performed on the social media data including the link relationship.

Further, in the step S2, under the guidance of the social relevance theory, various link relationships can be extracted from the associated data, and based on the following four link relationships, they can be combined with the existing feature selection methods to model. It is a new feature selection rule, so as to realize feature selection of link data in social media, and the link relationship includes 1) Co-post: that is, multiple posts from the same user or social media instances from the same user are more similar; 2 )Co-following: i.e. if two users follow the same user, then the posts generated by the two users are more similar; 3)Co-followed: i.e. if two users are followed by the same user, their posts are likely Similar; 4) Following: That is, if one user follows another user, the two users may have the same point of interest, so their posts may be similar.

Further, in the step S3, after completing the construction of the link relationship, we take L1 regularization as the basic feature selection model, and combine the above link relationship to construct a new feature selection optimization object. The optimization objective function of L1 regularization is as follows:

Among them, W represents the feature weight, and the parameter α controls the sparsity of W; let L(X, Y) represent

其中，β调整Co-Post的贡献度，T(f_i)＝W^Tf_i；对于Co-following链接关系：

其中，β调整Co-following的贡献度，

对于Co-followed链接关系，首先引入指标矩阵

表示如果u_j是p_i的作者；且FE＝sign(SS^T)，其中sign(x)＝1如果x＞0，否则sign(x)＝0。L_FI表示FI的拉普拉斯矩阵，则优化目标表示为：

其中，B＝XX^T+βXHL_FEH^TX^T,E＝Y^TX^T；对于following链接关系，有：

其中，B＝XX^T+βXHL_SH^TX^T,E＝Y^TX^T,与Co-followed链接关系的区别在于L_S为S的拉普拉斯矩阵。Further, in the step S3, it also includes performing corresponding optimization on the link relationship: for the Co-post link relationship:

Among them, β adjusts the contribution of Co-Post, T(f _i )=W ^T f _i ; for the Co-following link relationship:

Among them, β adjusts the contribution of Co-following,

For the Co-followed link relationship, first introduce the indicator matrix

represents if u _j is the author of _pi ; and FE=sign(SS ^T ), where sign(x)=1 if x>0, otherwise sign(x)=0. L _FI represents the Laplacian matrix of FI, then the optimization objective is expressed as:

Among them, B=XX ^T +βXHL _FE H ^T X ^T , E=Y ^T X ^T ; for the following link relationship, there are:

Among them, B=XX ^T +βXHL _S H ^T X ^T , E=Y ^T X ^T , and the difference from the Co-followed link relationship is that L _S is the Laplacian matrix of S.

Further, in the step S4, the feature selection modeling of the social media data is realized through the obtained link relationship feature selection criteria, and for the feature subset obtained for each relationship, the final feature set is obtained in the form of a union, Complete follow-up social media data analysis.

The present invention has the following advantages: it proposes to integrate the known social media data association relationship to solve the problem of feature selection of this type of data, based on four types of relationships extracted from the linked data, and their constraints are integrated into commonly used feature selection methods In L1 regularization, the feature selection of social media data with link attributes is realized, which solves the social media data association characteristics that cannot be solved by traditional feature selection methods, and lays a solid foundation for subsequent operations such as dimensionality reduction of large-scale social media data and analysis of important features. The basics.

Description of drawings

Figure 1 is a schematic diagram of social media data;

Fig. 2 is the social media data feature selection framework with link attribute;

FIG. 3 is a flowchart of a feature selection method for social media data that integrates L1 regularization and link attributes.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The social media data feature selection method that integrates L1 regularization and link attributes disclosed in the present application includes the following steps: S1. First, the social media data containing the link relationship needs to be represented in a normalized manner; considering the social media data, its data points or instances Interconnected in nature, without loss of generality, Figure 1 presents a simple example construction of social media data with two data representations. Figure 1(a) has four users (u1,..., u4), each user follows some other users (eg, u1 follows u2 and u4) and has some posts or blogs (eg, posts p1 and p2 by user u1) ). Figure 1(b) is a conventional representation of attribute value data: the row is a post, which in social media represents each tweet or post published, and the column is the feature or term of the text, which is the main content carrier of social media data. In the context of social media, there is additional information in the form of link data, as shown in Figure 1(c), including the relationship who made the post, who followed whom, etc. Let p={p1,p2,…,pn} denote posts or blogs in social media data, n denote the number of posts or blogs; f={f1,f2,…,fm} denote the number of posts or blogs in social media data Feature attribute description, m represents the number of features. X represents a feature data matrix consisting of all samples (p) and their corresponding feature attribute sets, and the size is m*n. c={c1,c2,...,ck} represents the tag corresponding to each post or each sample, and k represents the number of tag categories. Y represents a matrix containing all samples (p) and their corresponding labels, and the size is n*k. u={u1,u2,...,ut} represents the set of users, and t represents the number of users. The following relationship between users and users is modeled as a graph and adjacency matrix S, where S(i,j)=1, indicating that there is a concern relationship from uj to ui (user j has followed user i), otherwise it is 0, For example shown in Figure 1 as S(:,1)=[0,1,0,1] ^T . At the same time, Q(i,j)=1 means that the post pj is sent by user ui, otherwise it is 0, for example, it is shown as Q(1,:)=[1,1,0,0,0,0,0 in Figure 1 ,0]. Therefore, the traditional supervised feature selection is to select a subset of features from m features based on the mapping of {X, Y}; while the feature selection of social media data considering the link relationship is expressed as considering the user relationship S and the user and post. On the basis of the relation Q, a subset feature is selected from the m features according to the mapping of {X, Y}. Among them, S and Q can be regarded as a unique constraint relationship when social media data is used for feature selection.

S2. After completing the representation of social media data, it is necessary to integrate the difference relationships of different types of samples to model the feature selection process. Here, under the guidance of social relevance theory, various relationships can be extracted from linked data, that is, several common known link relationships between users and social media in social media data:

1) Co-post: i.e. multiple posts from the same user or social media instances from the same user are more similar, in other words, a user's posts on topics (like "sports", "music") are more similar than those that are random The selected posts are more similar, such as {p3, p4, p5} in Fig. 1(a);

2) Co-following: that is, if two users follow the same user, then the posts generated by the two users are more similar. In Figure 1(b), u1 and u3 follow u4, then their posts {p1, p2} and {p6, p7} are likely to be similar themes;

3) Co-followed: that is, if two users are followed by the same user, their posts may be similar. For example, in Figure 1(a), two users u2 and u4 are followed by user u1, then their posts {p3, p4, p5} and {p8} may have relatively similar topics;

4) Following: That is, if a user follows another user, the two users may have the same point of interest, and thus their posts may be similar. For example, in Figure 1(a), u1 follows u2, then their posts {p1,p2} and {p3,p4,p5} may be similar.

Based on the above four link relationships, it can be combined with the existing feature selection method to model as a new feature selection rule, so as to realize the feature selection of link data in social media.

S3. After completing the construction of the link relationship, we use L1 regularization (Lasso regression) as the basic feature selection model, and combine the above four link attribute relationships to construct a new feature selection optimization object. The optimization objective function of L1 regularization is as follows:

则上述4种关系对应的优化问题为：Among them, W represents the feature weight, and the parameter α controls the sparsity of W; let L(X, Y) represent

Then the optimization problems corresponding to the above four relations are:

For Co-post link relationship:

Among them, β adjusts the contribution of Co-Post, T(fi )=W ^T _{f i .}

For Co-following link relationships:

对于Co-followed链接关系，首先引入指标矩阵

表示如果u_j是p_i的作者；且FE＝sign(SS^T)，其中sign(x)＝1如果x＞0，否则sign(x)＝0。L_FI表示FI的拉普拉斯矩阵。则优化目标表示为：Among them, β adjusts the contribution of Co-following,

For the Co-followed link relationship, first introduce the indicator matrix

represents if u _j is the author of _pi ; and FE=sign(SS ^T ), where sign(x)=1 if x>0, otherwise sign(x)=0. L _FI represents the Laplacian matrix of FI. Then the optimization objective is expressed as:

Wherein, B=XX ^T +βXHL _FE H ^T X ^T , E=Y ^T X ^T ;

Similarly, for the following link relationship, there are:

Among them, B=XX ^T +βXHL _S H ^T X ^T , E=Y ^T X ^T , and the difference from the Co-followed link relationship is that L _S is the Laplacian matrix of S. By combining the four common link relationships in social media data and the L1 regularization feature selection method, feature selection for social media data containing link relationships can be achieved. The specific algorithm principle is shown in Figure 2.

S4. Based on the above-obtained link relationship feature selection criteria, we can implement feature selection modeling for social media data. For the feature subset obtained from each relationship, the final feature set can be obtained in the form of a union to complete subsequent social media data analysis, such as classification.

S5. After completing the construction of various classifiers, we can predict the attributes of samples by inputting methylation data for unknown samples to be evaluated.

Only the preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to the above-mentioned embodiments, and within the scope of knowledge possessed by those of ordinary skill in the art, various aspects can also be made without departing from the purpose of the present invention. Various changes should be included within the protection scope of the present invention.

Claims (6)

1. A social media data feature selection method integrating L1 regularization and link attributes, characterized in that: S1. Input social media data, and its behavior samples are listed as the characteristics of the corresponding social content; S2. Normalized representation of social media data; S3. Extract the link relationship of 4 common social media data; S4, combined with L1 regularization to realize feature selection under the corresponding link relationship; S5. Take and output feature subsets obtained from different link relationships to obtain a final feature set. 2 . The feature selection method for social media data combining L1 regularization and link attributes according to claim 1 , wherein in the step S1 , normalized representation is performed on social media data including link relationships. 3 . 3. The feature selection method for social media data combining L1 regularization and link attributes according to claim 2, wherein in the step S2, under the guidance of social relevance theory, various Link relationship, based on the following four link relationships, it can be combined with the existing feature selection method and modeled as a new feature selection rule, so as to realize the feature selection of link data in social media, the link relationship includes 1) Co-post: that is, multiple posts from the same user or social media instances from the same user are more similar; 2) Co-following: that is, if two users follow the same user, then the posts generated by the two users are more similar Similar; 3) Co-followed: that is, if two users are followed by the same user, their posts may be similar; 4) Following: that is, if one user follows another user, the two users may have the same points of interest, and thus their posts may be similar. 4. The social media data feature selection method of merging L1 regularization and link attributes according to claim 3, characterized in that: in the step S3, after completing the construction of the link relationship, we take L1 regularization as the basic feature Select the model and combine the above link relationship to construct a new feature selection optimization object. The optimization objective function of L1 regularization is as follows:

Among them, W represents the feature weight, and the parameter α controls the sparsity of W; let L(X, Y) represent

5. The social media data feature selection method of integrating L1 regularization and link attributes according to claim 4, characterized in that: in the step S3, it also includes performing corresponding optimization on the link relationship: for Co-post Link relationship:

Among them, β adjusts the contribution of Co-Post, T(f _i )=W ^T f _i ; for the Co-following link relationship:

Among them, β adjusts the contribution of Co-following,

For the Co-followed link relationship, first introduce the indicator matrix

represents if u _j is the author of _pi ; and FE=sign(SS ^T ), where sign(x)=1 if x>0, otherwise sign(x)=0. L _FI represents the Laplacian matrix of FI, then the optimization objective is expressed as:

Among them, BXX ^T +βXHL _FE H ^T X ^T , E=Y ^T X ^T ; for the following link relationship, there are:

Among them, B=XX ^T +βXHL _S H ^T X ^T , E=Y ^T X ^T , and the difference from the Co-followed link relationship is that L _S is the Laplacian matrix of S. 6. The feature selection method for social media data combining L1 regularization and link attributes according to claim 5, characterized in that: in the step S4, feature selection of social media data is realized through the obtained link relationship feature selection criteria Modeling, for the feature subset obtained for each relationship, the final feature set is obtained in the form of union, and the subsequent social media data analysis is completed.

Images