CN110867250B - A social media self-harm behavior detection method based on strong robust feature selection - Google Patents

Abstract

The invention discloses a social media self-harm behavior detection method based on strong robust feature selection, 1) obtaining multi-dimensional heterogeneous information from online social media websites; Perform feature extraction in each aspect to construct a dataset of self-mutilation content and a dataset of normal content; 3) Construct a supervised self-mutilation detection model based on strong robust feature selection through the loss function and regularization term of l_2,1 norm; 4) Feature extraction is performed on the target data to be detected, and the constructed detection model is used for self-harm detection. Compared with the traditional self-mutilation detection method, the social media-oriented self-mutilation detection method disclosed in the present invention can more widely contact the self-mutilation subjects, explore the behavior patterns of the self-mutilation subjects more deeply, and discover the self-mutilation behaviors more efficiently and in a timely manner. application advantages.

Description

A social media self-harm behavior detection method based on strong robust feature selection

technical field

The invention belongs to the field of social media data mining, in particular to a social media self-harm behavior detection method based on strong robust feature selection.

Background technique

In recent years, self-harm has gradually become a major challenge in the field of social public health. Being able to detect self-harm in society in a timely and effective manner is a realistic need to meet this challenge. Since the existing traditional self-mutilation detection strategies based on self-mutilation subjects themselves and their family members and friends have shortcomings such as difficulty in execution and low efficiency, a new self-mutilation detection strategy is urgently needed. With the popularity of online social media, more and more people tend to post their thoughts and record their lives on social media, which makes it possible to use social media to detect self-harm behavior. Compared with traditional self-harm detection methods, self-harm behavior detection using social media can more efficiently detect more self-harm behaviors.

There has been a lot of research on the health status of network users using various network data mainly based on social media. The prior art proposes a psychological stress detection method based on heart rate and social media microblogs to discover the stress interval and stressor events of the target individual, which mainly includes: first, detecting the abnormal heart rate of the individual to reflect The nervous system tension level of the individual during the test period; then, the abnormal interval of individual microblogs is detected to find the abnormal situation of the frequency of users posting positive microblogs during the test period; finally, the abnormal heart rate and the abnormality of microblog posting are matched , so as to determine the pressure interval, and discover the pressure source events through the microblog data.

The prior art proposes an early warning method for the psychological crisis of social media users. The method mainly includes: first, acquiring text data published by users on social media, and preprocessing the data to obtain a data set composed of words; The word frequency statistics of negative words are used to quantify the sentiment analysis and calculation of the text, and the sentiment feature vector of the text published by the user is obtained; finally, the obtained feature vector is input into the neural network to obtain the user's negative emotional strength, and the user's psychological state. rating.

The above-mentioned data analysis methods based on social media only choose to use homogeneous information sources, and do not make full use of the rich heterogeneous information sources on social media for comprehensive data mining. At the same time, the data mining algorithms of the above methods are too simple, and cannot fully mine valuable information in media data and adapt to the complex data full of noise in practical applications.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a social media self-harm behavior detection method based on strong robust feature selection to solve the above problems.

To achieve the above object, the present invention adopts the following technical solutions:

A social media self-harm behavior detection method based on strong robust feature selection, comprising the following steps:

Step 1, social media data collection: take the historical data of online social media sites as the data source, obtain text information, user behavior information, time information and picture information of self-harm related posts and non-self-harm posts, and obtain a post collection consisting of several posts. ; remember a post collection consisting of n posts

Step 2, data feature extraction and data set construction: extract the features of four heterogeneous information sources from the posts p _i (i=1, 2,..., n) obtained from the data collection, and obtain the post feature vector fi _i = { _wi , _ui , t _i , _pi }, where _wi represents text features, _ui represents user behavior features, _ti represents temporal features of posts, and _pi represents image features of posts, thus constructing self-mutilation features respectively. post dataset and normal post dataset;

Step 3, establish a self-harm detection model: extract training samples from the data set constructed in step 2, and build and train a supervised self-harm detection model based on the objective function selected by strong robust features;

Step 4, self-mutilation content detection: For the target post p to be detected, construct its feature vector f according to the feature extraction method in step 2, and then input its feature vector f into the detection model trained in step 3 for feature selection. Determine if it is a self-harm related post.

Further, in the social media data collection in step 1, through the tag information of different social media posts, the web crawler or the application program interface provided by social media is used to crawl the subject of self-harm related and non-self-harm posts, and the main information obtained for each post is obtained. content include:

(1) Text information: Get the title of the post, the list of hashtags, the body text and all the comment texts it contains;

(2) User behavior information: obtain the total number of posts posted by the user, the time when the user joined the social media platform, the number of followers and followers of the user;

(3) Time information: get the posting time of the post and the shooting time of the pictures in the post;

(4) Picture information: Get all pictures attached in the post.

Further, in step 2 feature extraction and data set construction, it mainly includes:

(1) Text features: distribution features of text parts of speech, calculating the proportion of different parts of speech in the text content of each post; readability features, using the readability calculation formula in linguistics to calculate the readability index of the text; Sentiment tendency feature, using text sentiment analysis to judge whether the sentiment tendency of the post is positive, neutral or negative; the word vector representation of the text, using the deep model to calculate its vector representation for the text of each post; the above features are expressed as w={w _ling ,w _read ,w _sent ,w _vec };

(2) User behavior characteristics: Calculate the user's average posting volume according to the user's total posting volume and the time of using the social platform; use the user's total number of posts and the number of replies to calculate the average response to the user's posts. rate; plus the number of followers and fans of the user, its characteristics can be expressed as u={u _post , u _rep , u _fol , u _fan };

(3) Time characteristics: Divide each day into 24 time periods by hour, and count the time period of the posting time of the post and the shooting time of the attached picture, and its characteristics can be expressed as t={t _post , t _pic };

(4) Picture features: characterize the color pattern in the picture, and use color information to quantitatively analyze the emotional dimension of the picture; extract the local features of the picture according to the algorithm in the image processing and use the neural network to characterize the picture. It can be expressed as p={p _col , p _sent , p _local , p _net }.

来表示训练数据中可用的标注信息，其中，对

中帖子p_i，当{Y_i1＝1,Y_i2＝0}时，该帖子为自残内容帖子，反之，当{Y_i1＝0,Y_i2＝1}时，该帖子为正常帖子；Further, in the establishment of the self-damage detection model in step 3, an efficient and robust feature selection method with strong robustness is used: first, use

to represent the annotation information available in the training data, where

Medium post p _i , when {Y _i1 = 1, Y _i2 = 0}, the post is a post with self-harm content, otherwise, when {Y _i1 = 0, Y _i2 = 1}, the post is a normal post;

表示训练数据的数据矩阵，其中l_i为第i个异质信息源所抽取的特征数量；Then, use

represents the data matrix of training data, where li is the number of features extracted from the _i -th heterogeneous information source;

将数据矩阵X映射到标注信息矩阵Y，训练方式为：Finally, the purpose of strong robust feature selection is achieved by using the loss function of the l _2,1 norm and the regularization term; the constructed supervised model is to train a coefficient matrix

Map the data matrix X to the annotation information matrix Y, and the training method is:

为正则化项的参数，具体的训练过程为：in,

is the parameter of the regularization term, and the specific training process is:

其中，

为单位矩阵，

同时，初始化矩阵

为单位矩阵，并设置训练过程收敛的终止阈值为∈；(1) Build a matrix

in,

is the identity matrix,

At the same time, initialize the matrix

is the identity matrix, and the termination threshold of the training process convergence is set to ∈;

(2) Calculate U=D ^-1 A ^T (AD ^-1 A ^T ) ^-1 Y;

(3) Update the diagonal matrix D, and the diagonal elements are d _ii =1/(2‖u _i ‖ ₂ ), where ui is the _ith row of U;

(4) Construct W=(u ₁ , u ₂ ,..., _{um mn} ), and judge whether the decrease of the objective function is less than ∈ is true, if not, return to the process (2) to continue training; otherwise, exit the training and save the coefficient matrix W .

Further, in the self-mutilation content detection in step 4, the target post p to be detected can be mapped to obtain its indication vector y∈R^2 through the learned coefficient matrix W, and when y_1>y_2, it is determined that the post is self-mutilation content; Otherwise, the post is determined to be normal content, and normal content is non-self-harm content.

Compared with the prior art, the present invention has the following technical effects:

(1) Since it is difficult for the self-mutilation subject to find a reliable person for help in the physical world due to trust issues, it is more inclined to confide in social media with relative anonymity, so that the present invention can reach the self-mutilation subject more widely;

(2) The traditional research on self-harm behavior has shortcomings such as few samples and long tracking and observation period, and the large amount of social data accumulated by the great popularity of social media contains a large number of self-harm cases, so that the present invention can further explore and understand the behavior patterns of self-harm subjects ;

(3) Due to the concealment of self-mutilation behavior, the traditional method of discovering self-mutilation subjects' relatives and friends is difficult and lagging behind, while the self-mutilation detection mathematical model constructed based on social media data enables the present invention to detect self-mutilation behaviors more timely and effectively.

Description of drawings

FIG. 1 is a block diagram of the social media self-harm behavior detection method based on strong robust feature selection according to the present invention.

Figure 2 is a flow chart of the data acquisition process.

Figure 3 is a flow chart of the feature analysis process.

Figure 4 is a flow chart of the detection model training process.

Figure 5 is a flow chart of the self-harm detection process.

Figure 6 is an example of self-harm related posts on social media.

Detailed ways

The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples. It should be noted that the embodiments described herein are only used to explain the present invention, and are not used to limit the present invention. In addition, the technical features involved in the embodiments of the present invention may be combined with each other without conflict.

The specific implementation process of the present invention includes a data acquisition process, a feature analysis process, a model establishment process, and a self-harm detection process. FIG. 1 is a block diagram of the social media self-harm behavior detection method based on strong robust feature selection according to the present invention.

1. Data collection process

Figure 6 is an example of network social media data. The specific process of data acquisition is as follows:

(1) Through crawling technology, subject crawling is performed according to the tags of each data post in social media. When crawling self-harm related content, you can use self-harm related tags such as "selfharm", "selfinjury", "suicide" to crawl; when crawling normal posts, you can saturate the target web page without a theme Crawling;

(2) For posts related to self-harm, considering that the sets of posts returned by different tags may overlap, the posts related to self-harm are de-duplicated. Afterwards, in order to prevent non-self-harm users from accidentally posting self-harm related content, remove users with less than 5 posts and their related posts in the user set to which the self-harm posts belong;

(3) For normal content posts, filter by tags to remove posts with self-harm related tags. After that, random sampling is performed according to the demand of normal posts.

其中，爬取的内容信息如图6所示，包括：The process of the above steps is shown in Figure 2, so as to obtain a set of posts

Among them, the crawled content information is shown in Figure 6, including:

(1) Text information: Get the title of the post, the list of hashtags, the body text and all the comment texts it contains;

(2) User behavior information: obtain the total number of posts posted by the user, the time when the user joined the social media platform, the number of followers and followers of the user;

(3) Time information: get the posting time of the post and the shooting time of the pictures in the post;

(4) Picture information: Get all pictures attached in the post.

2. Feature analysis process

Feature analysis and extraction are performed on the posts obtained from the data collection process. Extract the features of its 4 heterogeneous information sources from the post p _i (i=1,2,...,n), and obtain the post feature vector f _i ={ _wi ,u _i ,t _i , _pi }. Its main process includes:

(1) Text feature extraction: text part-of-speech distribution features, calculate the proportion of nouns, verbs, adjectives and adverbs in the text content of each post, which can be calculated using social media-oriented text analysis tools such as CMUTweetTagger; readability Features, using the readability calculation formula in linguistics to calculate the readability index of text, such as Flesch readability calculation formula, Linsear Write readability calculation formula, Fog readability calculation formula and Dale-Chall readability calculation formula. Readability calculation formula; Sentiment tendency feature, use text sentiment analysis to determine whether the post's emotional tendency is positive, neutral or negative, and can use the corpus MPQA to calculate; the word vector representation of the text, use the word2vec model to calculate the text of each post. vector representation. The above features are expressed as w _i ={w _ling ,w _read ,w _sent ,w _vec };

(2) User behavior feature extraction: Calculate the average posting volume of the user according to the total posting volume of the user and the time of using the social platform; use the total number of posts of the user and the number of replies to calculate the average posting volume of the user's posts Reply rate; plus the number of followers and fans of the user, its characteristics can be expressed as u _i ={u _post ,u _rep ,u _fol ,u _fan };

(3) Temporal feature extraction: Divide each day into 24 time periods by hour, and count the time period of the posting time of the post and the shooting time of the attached picture, which can be characterized by a vector of {0,1} ²⁴ , and its features can be Expressed as t _i ={t _post ,t _pic };

(4) Picture feature extraction: The color mode in the picture is characterized by the cylindrical coordinate color space HSV, and the chroma (Hue), color saturation (Saturation) and brightness (Brightness) of the picture are obtained. At the same time, using color information to quantitatively analyze the emotional dimension of the picture, the calculation formula is:

In addition, the SURF algorithm, LBP algorithm and GIST algorithm in image processing are used to extract the local features of the picture, and the AlexNet neural network that has been pre-trained on the ImageNet dataset is used to extract the features of the picture. Its characteristics can be expressed as p _i ={p _col , p _sent , p _local , p _net }.

A flowchart of the process is shown in Figure 3. After that, a dataset of self-harm posts and a dataset of normal posts can be constructed separately.

3. Model building process

其中，对

中帖子p_i，当{Y_i1＝1,Y_i2＝0}时，该帖子为自残内容帖子；反之，当{Y_i1＝0,Y_i2＝1}时，该帖子为正常帖子。使用训练数据的特征向量组成的数据矩阵

(l_i为第i个异质信息源所抽取的特征数量)，所构建的基于强鲁棒性特征选择的有监督模型为训练一个系数矩阵

将数据矩阵X映射到标注信息矩阵Y，训练方式为：Define the self-harm content annotation information in the training data set as

Among them, to

Medium post p _i , when {Y _i1 =1,Y _i2 =0}, the post is a post with self-harm content; on the contrary, when {Y _i1 =0,Y _i2 =1}, the post is a normal post. A data matrix composed of eigenvectors using the training data

(l _i is the number of features extracted from the i-th heterogeneous information source), the constructed supervised model based on strong robust feature selection is to train a coefficient matrix

Map the data matrix X to the annotation information matrix Y, and the training method is:

为正则化项参数，具体的训练过程为：in,

is the regularization parameter, and the specific training process is as follows:

其中，

为单位矩阵，

同时，初始化矩阵

为单位矩阵，并设置训练收敛的终止条件为(‖XW-Y‖_2,1+αW2,1<∈；(1) Build a matrix

in,

is the identity matrix,

At the same time, initialize the matrix

is the identity matrix, and set the termination condition of training convergence as (‖XW-Y‖ _2,1 +αW2,1<∈;

(2) Calculate U=D ^-1 A ^T (AD ^-1 A ^T ) ^-1 Y;

(3) Update the diagonal matrix D, and the diagonal elements are d _ii =1/(2‖u _i ‖ ₂ ), where ui is the _ith row of U;

(4) Construct W=(u ₁ , u ₂ , ..., _{um mn} ), and judge whether the termination condition is established, if not, return to process (2) to continue training; otherwise, exit the training and save the coefficient matrix W.

The flow chart of the training process of the above supervised model is shown in Figure 4.

4. Self-harm detection process

当y₁>y₂时，则判定该帖子为自残内容；否则，判定该帖子为正常内容。该检测过程流程图如图5所示。For the target post p to be detected, construct its feature vector f according to the feature extraction method in the feature analysis, and then input its feature vector f into the detection model trained during the model establishment process to determine whether it is a self-harm related post. Through the mapping of the supervised self-harm detection model coefficient matrix W, the indicator vector of the target post p can be obtained

When y ₁ >y ₂ , the post is determined to be self-harm content; otherwise, the post is determined to be normal content. The flow chart of the detection process is shown in FIG. 5 .

Claims (1)

1. A social media self-disabling behavior detection method based on strong robustness feature selection is characterized by comprising the following steps:

step 1, social media data acquisition: taking historical data of a network social media website as a data source, and acquiring text information, user behavior information, time information and picture information of self-disabled related posts and non-self-disabled posts to obtain a post set consisting of a plurality of posts; post set composed of n posts

Step 2, data feature extraction and data set construction: for posts p obtained from data collection _i Wherein i =1,2, ·, n; extracting the characteristics of 4 heterogeneous information sources to obtain a post characteristic vector f _i ＝{w _i ，u _i ，t _i ，p _i In which w _i Representing a text feature, u _i Representing a user behavior feature, t _i Representing the temporal characteristics of the post, p _i Representing picture characteristics of the posts, and respectively constructing a self-disabled post data set and a normal post data set;

step 3, establishing a self-residual detection model: extracting training samples from the data set constructed in the step 2, and constructing and training a supervised self-residual detection model based on a target function selected by the strong robustness characteristics;

step 4, self-disabled content detection: constructing a feature vector f of a target post p to be detected according to the feature extraction method in the step 2, inputting the feature vector f into the detection model trained in the step 3 for feature selection, and judging whether the target post p is a self-mutilation related post;

in step 1, in social media data acquisition, through tag information of different social media posts, topic crawling of self-residual related posts and non-self-residual posts is performed by using an application program interface provided by a web crawler or a social media, and main contents acquired by each post comprise:

(1) Text information: acquiring a title, a subject label word list, a text and all comment texts contained in the text;

(2) User behavior information: acquiring the total posting volume of posting users, the time for the users to join the social media platform, and the attention number and the fan number of the users;

(3) Time information: acquiring the publishing time of the post and the shooting time of the picture in the post;

(4) Picture information: acquiring all pictures attached to the posts;

step 2, in feature extraction and data set construction, the method mainly comprises the following steps:

(1) Text characteristics: text part-of-speech distribution feature w _ling Calculating the proportion of different parts of speech in the text content of each post; readability characteristic w _read Calculating the readability index of the text by using a readability calculation formula in linguistics; emotional tendency characteristics w _sent Judging whether the emotional tendency of the post is positive, neutral or negative by utilizing text emotional analysis; word vector representation w of text _vec Calculating vector representation of the text of each post by using a depth model; the above feature is represented by w = { w = { [ w ] _ling ，w _read ，w _sent ，w _vec }；

(2) The user behavior characteristics are as follows: calculating the average posting volume u of the user according to the total posting volume of the user and the time of using the network social media _post (ii) a Calculating the average reply rate u of the user posts by using the total number of the posts of the user and the number of the posts with replies _rep (ii) a Plus the user's attention number u _fol Number u of vermicelli made from bean starch _fan Its characteristic can be expressed as u = { u = { u = _post ，u _rep ，u _fol ，u _fan }；

(3) Time characteristics: dividing each day into 24 time periods by hour, and counting the post publishing time t _post And a time period t of the shooting time of the attached picture _pic Characterized in that it can be represented as t = { t = { t = } _post ，t _pic }；

(4) Picture characteristics: for color pattern p in picture _col Representing and utilizing color information to carry out emotion dimension p on the picture _sent Carrying out quantitative analysis; local characteristic p of picture according to algorithm in image processing _local Extracting and representing the picture p by using a neural network _net Its characteristics can be expressed as p = { p = _col ，p _sent ，p _local ，p _net }；

Step 3, in the self-residual detection model establishment, a high-efficiency and steady feature selection method with strong robustness is used: first, use

To represent the annotation information available in the training data, wherein

Chinese post p _i When { Y } _i1 ＝1，Y _i2 =0} the post is a self-spoiled content post, otherwise, when { Y } _i1 ＝0，Y _i2 =1}, the post is a normal post;

then, use

A data matrix representing training data, wherein _i The feature quantity extracted for the ith heterogeneous information source;

finally, by using l _2，1 The loss function and the regularization term of the norm achieve the purpose of selecting the strong robustness characteristics; the constructed supervision model is used for training a coefficient matrix

Mapping the data matrix X to a labeling information matrix Y, wherein the training mode is as follows:

wherein,

for the parameters of the regularization term, the specific training process is:

(1) Constructing matrices

Wherein,

is a matrix of the units,

at the same time, the matrix is initialized

Setting a termination threshold value of the convergence of the training process as an element matrix;

(2) Computing

(3) Updating diagonal matrix D with diagonal elements D _ii ＝1/(2||u _i || ₂ ) Wherein u is _i Is Uth line;

(4) Configuration W = (u) ₁ ，u ₂ ，...，u _m-n ) Judging whether the descending amplitude of the target function is less than the epsilon or not, if not, returning to the process (2) to continue training; otherwise, quitting training and saving the coefficient matrix W;

step 4, in the self-disabled content detection, a target post p to be detected can be mapped through a learned coefficient matrix W to obtain an indication vector y ∈ R ^2, and when y _1 is larger than y _2, the post is judged to be self-disabled content; otherwise, judging that the post is normal content, wherein the normal content is non-self-disabled content.

Images