A resilience index of online group opinion

Abstract

The sudden change in online group opinion under the attack of external stimulus is what managers of any social organization should strive to avoid. To predict the sudden change in online group opinion, the evolution of online group opinion is considered as a complex system. The resilience index, which measures the accumulated pressure of this complex system, is developed as an indicator to predict the sudden change in group opinion. The employees’ opinion of Dada, a crowdsourcing logistics platform in China, is taken as an example. Text data of group opinion are treated. The methods of fitting the catastrophe model and building the resilience index model for group opinion of Dada forum are demonstrated. Verification is performed. The method of searching thresholds of the resilience index is designed, and two thresholds are obtained. The factors that influence the thresholds are investigated. The proposed resilience index and its modeling method have contributions to researches on online group opinion as well as complex systems, and have practical implications for business enterprises in face of customer group opinion in marketing and governments confronted with public group opinion.

Appendices

Appendix: 9 Data processing

1.1 9.1 A sample of a post

Each post is taken as a sample, including tittle, time, original poster statement and information, follower comment and many other contents; Fig. 17 shows an example of a post.

Fig. 17

A sample of a post

All the samples are freely available for readers in website: https://github.com/221BC/Dada-forum.

1.2 9.2 Text minning

Each follower comment of a post is taken for data processing and text mining. Text mining methods include NLP and semantic network analysis. SnowNLP is a Chinese text processing library of Python. We use the SnowNLP functions to calculate the sentiment value of texts. Jieba, a Python module, is employed to conduct word segmentation of Chinese texts. After word segmentation process of text data, we get frequency, keywords, etc. (see Fig. 18). Figure 19 shows the semantic network of all follower comments of a post. If there are co-occurrence keywords between follower comments, there will be a connection between the two nodes, thus forming a complex network by this rule.

Fig. 18

Keyword and frequency of all follower comments in a post

Fig. 19

Semantic network of all follower comments in a post

Sentiment value, keywords and frequency can be obtained from NLP, and complex network characteristics can be obtained from text semantic networks. An example of data processing result from these processing methods is shown in Fig. 20.

Fig. 20

An example of text data processing result

After the data analysis and text mining processing, we obtain the data processing results, shown in Fig. 21. Finally, we pick out one dependent variable $y$ and eight valid independent variables ${x}_{1},\dots ,{x}_{8}$, which is explained in Table 1.

Fig. 21

Text analysis results

Appendix: 10 Model fitting

2.1 10.1 Fitting results of the catastrophe mode

We input standardized independent variables ${x}_{1},\dots ,{x}_{8}$ and dependent variable $y$ into the cusp function of the cusp-package and obtain the following fitting results. The $w$ item in the fitting result is the coefficient of the linear relationship between system state $f$ and dependent variable $y$, while the $a$ and $b$ are the coefficients of the linear relationship between control variables $u$, $v$ and independent variables ${x}_{1},\dots ,{x}_{8}$, respectively, as shown in Eq. (2). The P-value of each item in the fitting result is at the end of each item. The results show that the P-value of each independent variable and the dependent variable is less than 0.001, indicating that these variables are significant. Comparing the three models of linear, logistic and cusp, the logistic model is the most suitable for the system, for it has the smallest AIC and BIC and the largest goodness-of-fit R square. Figure 22 shows the fitting results of the model. Thus, we choose the logistic model to build the resilience index model.

Fig. 22

Fitting results of the linear, logistic and cusp catastrophe model

2.2 10.2 Modeling results of the logistic model

The cusp.logist function of the cusp-package is applied to fitting of the logistic model using the whole 11,802 valid samples. The fitting results of the logistic model are shown in Fig. 23. Similarly, the $a$ and $b$ items are the coefficients of the linear relationship between control variables $u,v$ and independent variables ${x}_{1},\dots ,{x}_{8}$, respectively, as shown in Eq. (2).

Fig. 23

Fitting results of the logistic model

After that, we calculated the control variables of each sample according to the fitting results in Eq. (8). The results of control variables $u$ and $v$ are shown in Fig. 24.

Fig. 24

Results of control variables $v$, $u$

Appendix: 11 Results of resilience index

According to the resilience index model in Eq. (9), we calculated the resilience index of each sample. The results of the resilience index are shown in the last column of Fig. 25.

Fig. 25

Results of the resilience indexes

Appendix: 12 Resilience index analysis

4.1 12.1 Search algorithm for the thresholds of resilience index

As we have discussed in Sect. 6.1, the first threshold of resilience index locates where the sentiment change gradient is 0, and the second threshold locates where the sentiment change gradient is 1. According to this rule, we develop an integrated algorithm to search the thresholds of resilience index, shown in Fig. 26. The inputs of the algorithm are the samples containing the control variables $u,$ $v$ and $\mathrm{resilience}\, \mathrm{index}$, such as Fig. 25. Besides, list $X$ are used for the calculation of thresholds. The function $\mathrm{logistic}$ and external stimulus $S$ are important parameters for the calculation of sentiment change. The main idea of the algorithm is to search the thresholds according to sentiment change gradient. And the output ${t}_{1}$ and ${t}_{2}$ are accordingly the early waring and danger signal of the two thresholds.

Fig. 26

Search algorithm for the thresholds

4.2 12.2 Analysis on sample size

9000 samples are randomly selected from entire samples for fitting. The fitting results of 9000 samples are similar to that of the whole samples. Then the resilience index was recalculated, the fitting results are shown in Figs. 27 and 28. The resilience index surface is illustrated in Fig. 29.

Fig. 27

Fitting results of the catastrophe model for 9000 samples

Fig. 28

Modeling results of the logistic model for 9000 samples

Fig. 29

Resilience index surface for 9000 samples

4.3 12.3 Analysis on external influence

The standardized statement sentiment of the first experiment is increased by 0.5 and the second experiment is decreased by 0.5, respectively. In the third and fourth experiment, statement sentiment is increased by 1.0 and decreased by 1.0, respectively. And the resilience index surfaces of four experiments are depicted in Fig. 30.

Fig. 30

Resilience index surfaces of four experiments on statement sentiment