Prediction of the Dissemination of Health News on Microblogging Sites Based on Ample Feature Selection and Support Vector Machine

2.1 Identification of influencing factors of news dissemination

Microblog posts that are widely disseminated often resonate with mass concern and may lead to public crisis. Therefore, it is of great interest to understand the causes of the proliferation of health news. The causal analysis is actually a binary classification problem to separate the normal posts from the widely disseminated posts. Before the separation, it is necessary to identify the influencing factors of news dissemination. The news should be represented correctly and completely by a set of rich and diverse features. Otherwise, the detected results will be imprecise or even meaningless. Previous research has shown that the content features of a microblog post, together with contextual information like author features and early-stage social features of the post, may influence the microblog users’ decision on whether to forward the post. Therefore, the dissemination of a health news post may be affected by the combination of these three types of features [3].

(1) Content features

It is widely agreed that content features bear on information dissemination. However, there is no consensus over the conceptualization or operationalization [9]. One reasonable conceptualization is given by persuasion theory: attitude valance and persuasive strength influence information dissemination in the form of persuasive information.

The traditional studies on persuasion have recognized attitude valance, positive or negative, as the key to persuading others. All kinds of decisions, from which candidate to support to which news message to share, are subjected to attitude. Persuasive strength is another common rule for decision-making. It is commonly measured by its relevance, timeliness, accuracy, and comprehensiveness [10].

Assuming all the microblog posts released by mainstream outlets are accurate, this paper measures persuasive strength by relevance, timeliness and comprehensiveness. Relevance refers to how relevant the news is related to public health issues; timeliness stands for how up-to-date the news post is; comprehensiveness means the coverage or depth of the post.

Unlike the traditional linguistic analysis, this paper extends the metrics of persuasion strength to nonlinguistic features in social media. Similar to Twitter, most microblogging sites set a limit on the length of each post. For example, a Weibo post should not exceed 140 characters. Within the restricted length, users are allowed to insert emoticons and URLs, mention others by adding an @ before their usernames, and even attach images to their posts. All these nonlinguistic features have a substantial influence on persuasive effect. Therefore, the stylistic features of microblog were also selected to evaluate the persuasion strength, in addition to the traditional metrics. All the selected content features are listed in Table 1.

Considering the nearly free text and hybrid types of elements in microblog posts, the quantitative content analysis was adopted to explain the linguistic and nonlinguistic content features. Based on preset categories, quantitative content analysis can count and quantify the content of each post. In this way, the unstructured microblog data are transformed into relational forms that can be reasonably analyzed.

Table 1. The content features of microblog posts

(2) Author features

Most Internet users receive news from the media they usually follow. Thus, the features of the author exert a significant impact on the information evaluation by microblog users. As a result, this paper takes account of how author features affect news dissemination. Previous research has demonstrated that the activeness, experience and authority of the author are heuristic factors that represent three kinds of perceptions of the author [11]. The three factors are adopted in this paper to measure the features of the author. Of course, these metrics alone are too simple. For example, “whether an author is verified as an elite member” is not enough to reflect the authority of the author. Therefore, this paper adopts more metrics for the author features. The additional metrics are listed in Table 2. For instance, it is assumed that users are more likely to repost a news released by an author with a high influence score, which is a rating by the microblogging site. The author’s authority is also reflected by the length of self-description and the length of verification information.

(3) Social feature

The forward decisions of microblog users are subjected to social influence. Normative social influence is a type of social influence leading to conformity. In social psychology, it is defined as the influence of others that leads us to conform in order to be liked and accepted by others. Hence, normative social influence can be a very powerful, motivator of behavior, and may overshadow the impact of the authority of the author. Nevertheless, there is little report on the substantive importance of social influence [6]. For microblog users, whether to forward a microblog post depends on their perception of social norms concerning whether to perform the behavior. Therefore, this paper posits that the new posts can be discriminated by social based information. The social feature selected for this research is shown in Table 2.

Table 2. The author features and social feature of microblog posts

2.2 Data collection

The research data were collected from Weibo, the leading microblogging site in China. The news posts from twelve mainstream news outlets covering public health issues were monitored continuously. To ensure the diversity of data sources, the twelve news outlets were carefully selected, including four newspapers, two magazines, five online news sites and one TV news agency. In total, 863 microblog posts were collected, all of which related to public health issues. Details on the dataset can be found in the previous research of Pei [6,12].

Over the years, many thresholds have been designed to evaluate widely disseminated posts. However, few have explained the selection criteria of these thresholds. Pei et al. [6] explored the dissemination of microblog information based on the Pareto Principle or the 80/20 Rule, revealing that 80 % of all the public attention are attracted by roughly 20 % of the posts (posts forwarded more than 766 times). In the light of this finding, the authors set the threshold of forwards to 766, and divided the 863 microblog posts into two classes. The 238 posts that were forwarded 766 times or more were defined as widely disseminated posts and categorized as Class 1, while the remaining 625 posts were defined as normal posts and categorized as Class 2. Then, the two classes were compared to discover the key features of widely disseminated posts, and forecast the types of news posts that will be widely diffused.

2.3 Measurement

(1) Attitude valance

The attitude valance of a microblog post is generally measured through sentiment analysis, which mainly detects the emotional sentiment reflected in the text for various purposes. An important basis of sentiment analysis is to identify positive and negative words. In this paper, the attitude valance of a microblog post is evaluated by metrics like the number of positive words, the number of neutral emotions, the number of negative emotions and the number of negative words.

(2) Persuasive strength

Relevance, a feature of persuasive strength, describes how relevant a news post is to public health issues. It can be measured by the number of public health words appearing in the text [6]. Hence, this feature is measured here by the frequencies of “food safety” terms, “medical” terms and “environmental protection” terms.

Timeliness, another determinant of persuasive strength, was evaluated by two metrics: the hour the post was published and whether the post was published in peak hours.

Comprehensiveness was assessed by various stylistic features of the post, such as “does the news post contain images”. Specifically, the order and co-occurrence of words were used to judge the comprehensiveness. This is realized with the aid of n-gram, a contiguous sequence of n words from a given sequence of text. The n-gram technique, especially 2-gram words, can effectively characterize a message [13], and improve the prediction of the popularity of news posts. Comprehensiveness was also evaluated by microblog specific features like the topic length and number of mentions. Similar to twitter, Weibo users can indicate the topic of each post with a pair of hashtags, and mention others by placing an @ before their usernames.

Overall, 23 contents features, 11 author features and 1 social feature were selected for further analysis.

2.4 Feature reduction

The huge number of features may complicate the binary classification problem, suppressing the efficiency of the learning algorithms. Therefore, the predictive ability of each selected feature was investigated using the information gain method, aiming to find out the features that best reflect the attributes of news posts [14]. The information gain can be interpreted as follows [15].

Let Z be a discrete class variable with m alternative values. Then, entropy H(Z) can be defined as:

$\mathrm{H}(\mathrm{Z})=-\sum_{\mathrm{i}=1}^{\mathrm{m}} \mathrm{P}\left(\mathrm{z}_{\mathrm{i}}\right) \log _{2} \mathrm{P}\left(\mathrm{z}_{\mathrm{i}}\right)$     (1)

For a given X with k alternatives, the conditional entropy of Z can be expressed as:

$\mathrm{H}(\mathrm{Z} | \mathrm{X})=\sum_{\mathrm{j}=1}^{\mathrm{k}} \mathrm{H}\left(\mathrm{Z} | \mathrm{x}_{\mathrm{j}}\right) \mathrm{P}\left(\mathrm{x}_{\mathrm{j}}\right)$     (2)

The information gain method aims to select features based on the information contribution related to the class variable, i.e. the amount of additional information about Z provided by X:

$I G(Z ; X)=H(Z)-H(Z | X)$     (3)

After obtaining the information gain of each feature, a minimum limit was set to filter out the features whose information gain is below the limit.

2.5 Classification

After feature reduction, each news post was transformed into a number of features that are represented as numerals. This lays the basis for the application of machine learning methods [16-17]. Support vector machine (SVM) is one of the most popular machine learning algorithms for classification and regression problems. The core idea of the SVM is to separate different classes with a hyperplane. This algorithm has been proved to have excellent effect on linear data classification in many applications [18]. If the original data are only nonlinearly separable, the SVM can be coupled with kernel functions to map the original data the feature space to a higher-dimensional space, such as to separate the data linearly. This subsection briefly describes the SVM algorithm adopted for our binary classification problem.

For linearly separable data, the decision function can be written as:

$\mathrm{f}(\mathrm{x})=\mathrm{w} \mathrm{Tx}+\varepsilon=0$    (4)

where, w is the weight vector; ε is the bias; x is the dataset. The hyperplane described by formula (4) divides one space into two parts: a positive part for samples in the positive class (+) and a negative part for samples in the negative class (-). Since the problem is to determine the values of w and ε, so that the hyperplane can be as far as possible from all the samples. More specifically, the SVM algorithm sets up hyperplanes, HP₁ and HP₂, as follows:

$H P_{1} \rightarrow w^{T} x_{i}+b=+1$ for $y_{i}=+1$

$H P_{2} \rightarrow w^{T} x_{i}+b=-1$ for $y_{i}=-1$    (5)

where, $w^{T} x_{i}+\varepsilon \geq+1$ gives the hyperplane for the positive class; $w^{T} x_{i}+\varepsilon \leq-1$ gives the hyperplane for the negative samples. The two equations in formula (5) can be combined into:

$y_{i}\left(w^{T} x_{i}+b\right)-1 \geq 0 \quad \forall_{i}=1,2, \ldots, \mathrm{n}$    (6)

The SVM margin represents the sum of $d_{1}$ and $d_{2}$ as:

$y_{i}\left(w^{T} x_{i}+b\right)-1 \geq 0 \quad \forall_{i}=1,2, \ldots, \mathrm{n}$    (7)

where, $d_{1}$ and $d_{2}$ are the distance of the samples from the first and second hyperplanes, respectively. In the SVM algorithm, the margin width needs to be maximized as:

$\begin{aligned} \min \frac{1}{2}|| w||^{2} & \\ \text {s.t. } y_{i}\left(w^{T} x_{i}+b\right)-1 \geq 0 & \forall_{i}=1,2, \ldots, n \end{aligned}$    (8)

Combining the objective function $\left(\min _{2} \frac{1}{2}|| w||^{2}\right)$ and the constraint $y_{i}\left(w^{T} x_{i}+b-1 \geq 0\right),$ the binary classification problem can be formalized into the following Lagrange formula:

$\min L_{p}=\frac{\|w\|^{2}}{2}-\sum_{I} a_{i}\left(y_{i}\left(w^{T} x_{i}+b\right)-1\right)=\frac{\|w\|^{2}}{2}-\sum_{I} a_{i}\left(y_{i}\left(w^{T} x_{i}+b\right)\right)+\sum_{i=1}^{N} a_{i}$    (9)

where, $a_{i}$ is the Lagrange multiplier for $x_{i} ; L_{p}$ is the primary problem. The values of $w, \varepsilon$, and $a$ that minimize $L_{p}$ in formula ( $9)$ were calculated by differentiating $L_{p}$ with respect to w and ε and setting the derivatives to zero as:

$\frac{\partial L_{p}}{\partial w}=0 \Rightarrow \mathrm{w}=\sum_{i-1}^{N} \alpha_{i} y_{i} x_{i}$     (10)

$\frac{\partial L_{p}}{\partial b}=0 \Rightarrow \sum_{i=1}^{N} \alpha_{i} y_{i}=0$    (11)

Substituting formulas (10) and (11) into formula (9), the binary classification problem can be transformed as:

$\max L_{D}=\sum_{i=1}^{N} \alpha_{i}-\frac{1}{2} \sum_{i, j} \alpha_{i} \alpha_{j} y_{i} y_{j} x_{i}^{T} x_{j}$

s.t. $\alpha_{i} \geq 0, \Sigma_{i=1}^{N} \alpha_{i} y_{i}=0 \forall_{i}=1,2, \ldots, N$    (12)

where, $L_{D}$ is the dual form of $L_{p}$. Then, the values of $\mathrm{w}, \varepsilon,$ and $\alpha$ can be determined by finding out a solution to formulas $(10) \sim(12) .$ In the SVM, most of the $\alpha_{i}$ are zeros, an evidence of the sparseness of the algorithm. The nonzero  values are the samples closest to the hyperplane, corresponding to support vectors (SV_s). Hence, the SV_s achieved the maximum width margin.

For nonlinear data, the key of the SVM is to use a nonlinear mapping function, i.e. the kernel function, to map samples into a higher-dimensional linear space. After the mapping, linear classification can be performed in the higher-dimensional space. After kernel transformation by the kernel function K_{(x, y)}, the decision function becomes:

$f_{x}=\mathrm{w} \cdot \Phi_{x}+\varepsilon$    (13)

2.6 Optimization

Then, several components of the SVM algorithm was optimized, including the penalty factor C, the kernel function width q and the insensitive band loss function g. The commonly used optimization methods include particle swarm optimization (PSO), and genetic algorithm (GA). After that, the classification effect of our model was evaluated by mean absolute percentage error (MAPE), mean square error (MSE) and Theil’s coefficient of inequality.

2.7 Comparative experiment

Finally, the SVM classification performance of the research data were compared with traditional intelligent algorithms like Naive Bayes classifier, logistic regression, J48 tree classifier, and radial basis function (RBF) network. Many scholars have utilized one of these methods to predict information cascades, but few explained the reason to choose a specific method [19-20]. That is why the authors decided to compare the effectiveness of all these methods and verify if the SVM outperform the contrastive methods in predicting the dissemination of health news.

Social media like microblogging sites have already developed into 24/7 dissemination platform of real-world events. Despite the research efforts in information dissemination, it remains a huge challenge to monitor or detect the risk information from social media services. For crisis managers, it is extremely important to understand the news propagation on microblogging sites and predict future crisis based on microblog news posts. Therefore, this paper explores deep into the features of microblog posts on health news and their dissemination trends.

This research makes several theoretical contributions. First, the authors identified the inconsistencies of conceptualization in the literature on content research, and re-conceptualized content features according to the persuasion theory. Second, this research incorporates the new realities of microblogging sites into the conceptualization of factors related to attitude strength. Third, this research highlights the importance of social information and the key role of author features in news evaluation.

There are also direct practical implications of our research. For example, the authors developed a set of ample, diverse features and an optimal algorithm to predict news dissemination. These results provide a valuable reference to researchers on information cascades. Furthermore, our approach targets specifically the news on public health issues. Thus, our analysis informs crisis managers about public reaction towards specific health news. Finally, our method is computationally feasible in near real-time scenarios and can be utilized to capture rapidly changing dynamics of microblog news dissemination.