Configurational patterns for COVID-19 related social media rumor refutation effectiveness enhancement based on machine learning and fsQCA

Abstract

Infodemics are intertwined with the COVID-19 pandemic, affecting people's perception and social order. To curb the spread of COVID-19 related false rumors, fuzzy-set qualitative comparative analysis (fsQCA) is used to find configurational pathways to enhance rumor refutation effectiveness. In this paper, a total of 1,903 COVID-19 related false rumor refutation microblogs on Sina Weibo are collected by a web crawler from January 1, 2022 to April 20, 2022, and 10 main conditions affecting rumor refutation effectiveness index (REI) are identified based on “three rules of epidemics”. To reduce data redundancy, five ensemble machine learning models are established and tuned, among which Light Gradient Boosting Machine (LGBM) regression model has the best performance. Then five core conditions are extracted by feature importance ranking of LGBM. Based on fsQCA with the five core conditions, REI enhancement can be achieved through three different pathway elements configurations solutions: “Highly influential microblogger * high followers’ stickiness microblogger”, “high followers’ stickiness microblogger * highly active microblogger * concise information description” and “high followers’ stickiness microblogger * the sentiment tendency of the topic * concise information description”. Finally, decision-making suggestions for false rumor refutation platforms and new ideas for improving false rumor refutation effectiveness are proposed. The innovation of this paper reflects in exploring the REI enhancement strategy from the perspective of configuration for the first time.

1. Introduction

With the rapid development of information technology, online social networking platforms have become a new way for people to interact and collect information, but they have also become a medium for raging online rumors. In 2003, Rothkopf (2003) first proposed the concept of “infodemic” to describe the phenomenon of mass panic caused by the mixing of facts and rumors during the SARS outbreak. At the end of 2019, the novel coronavirus pneumonia (COVID-19) broke out. During the pandemic, the public has relied more on obtaining information from the Internet to cope with an environment of uncertainty and rapid differentiation (Ball & Maxmen, 2020). However, various misinformation may confuse or mislead public behaviors, where the infodemics are intertwined with the COVID-19 (Briand, 2021; Yan, 2022). Against this background, online rumors permeated the entire cyberspace, which not only distorted people's perception of the pandemic, but also disrupted social order and public security. The current situation of the parallel spread of infodemics and diseases has caused social panic and brought serious challenges for curbing COVID-19 related rumors (Ali, O., & Hamedat, 2022; Jana, 2022).

There are mainly two methods of curbing rumors, namely, preventing the spread of rumors and spreading the truth to clarify the rumors (Wen, 2014). Expanding the spread of truth has better long-term results than preventing rumors. Due to the open nature of the Internet, it is increasingly difficult to prevent the spread of rumors, blocking rumors directly may have a backfire effect if there is no appropriate social support (He, 2015). Blocking information will cause concerns and doubts among Internet users, which may lead to a decline in the false rumor refutation effectiveness and social volatility. Therefore, spreading the truth has become the preferred way to curb false rumors. From the perspective of quantity and quality evaluation of truth dissemination, Li et al. (2021) proposed rumor refutation effectiveness index (REI), which makes it possible to quantitatively evaluate the dissemination effectiveness of rumor refutation information. In this paper, fsQCA is used for the first time to explore the COVID-19 related false rumor refutation effectiveness from the unique perspective of configuration path.

QCA has three main variants according to the different set form: crisp set QCA (csQCA), multi-value QCA (mvQCA) and fuzzy-set QCA (fsQCA). CsQCA, the first variants of the QCA designed by Charles Ragin and Kris Drass in 1987 (Roig-Tierno, 2017), is a method of converting a variable into a dichotomous variable of “0″ or “1” (Civera, 2020). MvQCA can be seen as an extension of csQCA, which treats variables as multi-valued variables rather than dichotomous variables (Cronqvist, 2003). FsQCA allows variables to take partial membership scores between “0″ and “1”, which solves the dichotomous variables limitation of csQCA. The advantages of fsQCA are as follows: (1) FsQCA uses fuzzy set logic to minimize information loss in data analysis (Fiss, 2007). (2) FsQCA can be used for both small samples and large samples. FsQCA was originally developed specifically for small sample study situations (10 to 40 cases) to ensure sufficient heterogeneity between cases for comparison (Ragin, 2008), but some recent studies showed that QCA with large samples can also be used for hypothesis testing and deductive reasoning (Greckhamer, 2008, 2013; Ham, 2019). (3) FsQCA analyzes the causal relationship of variables from a holistic perspective. The configuration analysis of fsQCA breaks through the traditional isolated analysis perspective and analyzes the causal relationship between variables from the whole perspective. The conditions of management practice are interdependent, which means that the effect of a condition depends on its configuration relationship with other conditions (Rihoux & Ragin, 2009). Aiming at the sample size and data volume in this paper, variables are not suitable for simple processing as 0, 1 or multi-valued forms. Therefore, fsQCA is used to explore configuration solutions to enhance the false rumor refutation effectiveness, which provides decision support for relevant authorities and false rumor refutation platforms.

To curb the spread of COVID-19 related false rumors, fsQCA is used to find the configuration pathways to improve false rumor refutation effectiveness. The contributions of this paper are: (1) FsQCA is used to explore COVID-19 related REI enhancement configuration pathways for the first time. (2) Machine learning methods are applied to extract core conditions and reduce data redundancy, which makes condition determination more objective and fsQCA pathways more feasible. (3) Three configurational pathways for REI enhancement are obtained, providing decision-making suggestions and method guidance for false rumor refutation platforms.

The overall research framework of this paper is shown in Fig. 1 . The remainder of this paper is organized as follows: Section 2 is literature review. Section 3 presents the experimental setup and methodology. Section 4 is data construction. Section 5 is mainly about the experimental results. Section 6 discusses our main findings, theoretical and practical contributions. Section 7 states the conclusion and future research.

Fig. 1.

Research framework.

2. Literature review

This section discusses the related work of false rumor refutation in social media and fsQCA applications in social media.

2.1. False rumor refutation in social media

On qualitative side, current research on false rumor refutation mainly focused on case-oriented study (Chen, 2021; Zhang & Xu, 2021), the exploration from the perspective of psychology (Berinsky, 2017; Manyika, 2013) and the perspective of news communication and management (Dai, 2020; Paek & Hove, 2019). On quantitative side, researchers were more concerned about the identification of rumors and their influence factors of rumor spreading (Liu, 2019; Lu, 2019), and the establishment of rumor spreading models (Guilherme, 2022; Hosseini & Zandvakili, 2022). Lewandowsky et al. (2012) and Zhang et al. (2022) explored determinants and cognitive factors governing the effectiveness of misinformation refutations. Kim and Dennis (2019) found influencing factors on social media article believability. Besides, a small number of papers focused on rumor refutation effectiveness (Ben, 2022), some studied effectiveness of different strategies for rumor control (Agarwal, 2022; Huang, 2020) or analyzed specific cases of rumor control (Ding, 2020; Paek & Hove, 2019).

Notably, the rumor refutation effectiveness index (REI) proposed by Li et al. (2021) provided the possibility for quantitative evaluation of false rumor-refuting effectiveness. REI is proposed from the qualitative and quantitative perspectives, and conditions influencing REI remarkably are found through content and context factors. Li et al. (2022) focused on factors influencing the lifetime, peak and dynamic change of REI, which is an innovation compared with previous study. However, both studies analyzed the effectiveness of false rumor refutation from the perspective of influencing factors, instead of the perspective of configuration patterns. This paper will make up with this research gap.

2.2. FsQCA applications in social media

FsQCA has been widely used in transportation and sustainable development (Hartmann, 2022; Llopis-Albert, 2021), enterprise management (Kusa, 2021), economic consumption (Abbasi, 2022; Li, 2022), education (Nistor, 2019) and other fields. However, there are still few applications in social media related research. Zhang et al. (2022) used fsQCA to obtain emotional stimulation, altruism and relationship management motivation as the key factors of brand social media rumor propagation and their coexistence between conditions. Capatina et al. (2018) similarly used fsQCA to mine the factors influencing brand, which determined the core impact factors and causality of social media on accommodation branding. The key factors and configurational pathways influencing travel experiences recommendation on social media could also be extracted by fsQCA (Wang, 2022). For social media users research, Gunawan and Huarng (2015) analyzed the factors that may influence consumers’ purchase intentions and engage in word-of-mouth on social media. Xie and Tsai (2021) studied the reasons for the decline of social media users and confirmed the significance of discontinuance intention determinants by fsQCA.

FsQCA was widely used in factors identification of brand rumor management, transportation, economic consumption and network platforms management related to social media. However, there is still a lack of research in the field of social media false rumor refutation effectiveness by fsQCA.

3. Experimental setup and methodology

This section states key experimental setup and methodology: definition and calculation of REI, topic extraction and sentiment analysis, core conditions extraction and obtaining approach of configurational patterns for REI enhancement.

3.1. Outcome and conditions identification

This paper adopts REI as the outcome of configurational patterns. REI is an index for social media false rumor refutation effectiveness evaluation (Li, Z., 2021) from both the transmission quantity and quality of the information. The calculation formula of REI is as follows:

$$REI=-{\log }_2\frac{1}{r(1+k)+p+l+1}$$ (1)

Where $r$, $k$, $p$, $l$ denote the number of the microblog's retweets, the ratio of influential users of the retweeters, the number of the microblog's positive comments and the number of the microblog's likes, respectively.

To get the pathways of REI enhancement, conditions affecting REI need to be identified. Gladwell (2006) proposed three rules of epidemics: (1) The law of few. The rapid spread of information relies heavily on the participation of people with some particular and rare social talents. They can be defined as key opinion leader (KOL) (Lazarsfeld, 1948). Obviously, microbloggers’ attributes significantly impact the diffusion and effect of information. Therefore, the law of few are corresponding to microbloggers’ conditions, including identity of microbloggers, influence of microbloggers, followers’ stickiness of microbloggers and activity of microbloggers. (2) The power of context. Our inner states, preferences and emotions are actually powerfully and subtly influenced by the environment. Topic is one of the important factors that constitutes social environment. The context of a microblog is the topic. That is, whether the false rumor refutation microblog is related to specific topic types, popular topics, or topics involved with some directions of public emotional tendencies should be considered. Therefore, the power of context is corresponding to contextual conditions, including type of the topic, popularity of the topic and sentiment tendency of the topic. (3) The stickiness factor. Information with splashy pictures and videos is often preconceived, attracting more people. In addition, information published at different time reaches different numbers of people. Therefore, the contents and forms of the false rumor refutation microblogs play a critical role on information's dissemination. The richer the content, the more frequent the user interaction, and the higher the user stickiness. To some extent, the content conditions are stickiness factors (Wen, 2014), including length of information descriptions, the number of pictures and videos and release time.

According to Gladwell's idea, three conditions may affect REI are considered: microbloggers’ conditions, contextual conditions and content conditions, which are decomposed to 10 conditions shown in Table 1 .

Table 1.

Conditions affecting REI.

	Condition categories		Conditions
A	Microbloggers’ conditions	A1	Identity of microbloggers
		A2	Influence of microbloggers
		A3	Followers’ stickiness of microbloggers
		A4	Activity of microbloggers
B	Contextual conditions	B1	Type of the topic
		B2	Popularity of the topic
		B3	Sentiment tendency of the topic
C	Content conditions	C1	Length of information description
		C2	The number of pictures and videos
		C3	Release time

3.1.1. Microbloggers’ conditions

KOL is a high-influence group with high activity and common interests with followers Lazarsfeld (1948). Information often flows first to the KOL, and then to the public, and the involvement of opinion leaders accelerates the dissemination of information and expands its influence. Microbloggers with these characteristics can make information more authentic. Therefore, this paper considers four conditions for the attributes of microbloggers: identity, influence, followers’ stickiness and activity.

• •Identity of microbloggers (A1)

Different types of information spreaders contribute differently to information dissemination. Official government media with a wide audience plays a vital role in spreading false rumor refutation information. Celebrities and local business microbloggers continue to facilitate the dissemination of information by spreading false rumor refutation information to audiences far from the core of the network. Although individual microbloggers account for the largest part of Weibo, they are relatively passive in disseminating information (Cha, 2012). Based on this, the paper divides the identity of microbloggers into four categories: International media microbloggers, national media microbloggers, celebrities and local business microblogges, individual microbloggers.

• •Influence of microbloggers (A2)

The influence of the microblogger is very important to the credibility of the false rumor-refuting information (Choi & Lee, 2017; Zareie, 2019). When people are faced with a new piece of information, prior knowledge affects people's assessment of the information's credibility. They prefer to trust information from sources they trust rather than other unknown sources (Kim & Dennis, 2019). Therefore, this paper evaluates the influence of microbloggers according to the number of followers.

• •Followers’ stickiness of microbloggers (A3)

Stickiness is the willingness of followers to visit a microblogger's homepage and stay there longer (Lin, 2007). Users’ stickiness is a powerful indicator of influencers’ value (Hu, 2020). Proactive stickiness keeps followers on the microbloggers’ home pages longer which improves the effectiveness of false rumor refutation. The more time followers are willing to stay on the homepage, the more likely they are to comment; and the more comments, the stronger the interaction. Therefore, this paper evaluates the microbloggers’ stickiness with followers according to the number of comments.

• •Activity of microbloggers (A4)

The number of followers of a microblogger is positively correlated with his information dissemination behavior (Suh, 2010). If a microblogger is highly active, the possibility of interaction and the number of followers will increase. Although frequent broadcasting to followers on social media does not necessarily lead to greater engagement (Suh, 2010), microbloggers with high activity are more likely to encourage users to share the microblogs and promote the dissemination of false rumor-refuting information (Li, 2021; Sinha, 2013). The number of microblogs published by the microblogger is used to measure microbloggers’ activity.

3.1.2. Contextual conditions

The context has a powerful impact on information dissemination (Gladwell, 2006), while the topic is an important contextual component of social environment. Therefore, we consider three conditions for the topic features of false rumor-refuting microblogs: type of the topic, popularity of the topic and sentiment tendency of the topic.

• •Type of the topic (B1)

Type of the topic means the main idea of the information. Although users’ attention is affected by different microblogs’ topics, they are more likely to pay attention to the more important information they need and interact with it (Li, 2021; Nah & Davis, 2002). Therefore, different topics of false rumor refutation have different degrees of uses’ attention, which may further affect the effectiveness of false rumor refutation.

• •Popularity of the topic (B2)

Though users read little or part of the details, in fact, they always focus on the information they are interested in (Nah & Davis, 2002). Therefore, microblogs with high popularity contents are more likely to catch users’ attention. Easier for users to retweet, comment, and like these kinds of information can improve the effectiveness of false rumor refutation. The popularity of a microblog increases with the importance of the topic. Therefore, information importance is used to measure topic popularity. Through expert interviews, the importance ranking is obtained according to Grounded Theory.

• •Sentiment tendency of the topic (B3)

Social media provides a channel for emotional sharing. When users are aware of a new false rumor-refuting microblog, they will generate emotions under the condition of subjective cognition (Hu, 2020). Emotions further influence the dissemination of the false rumor-refuting information. Emotional choices of users play a vital role in mitigating disinformation and guiding public opinion (Yin, 2022).

3.1.3. Content conditions

The content and form of information affect the information propagation (Gladwell, 2006). Therefore, three conditions for the description of false rumor-refuting microblogs are considered, namely, the length of information description, the number of pictures and videos and the release time.

• •Length of information description (C1)

Based on psychological theory, the correct format of microblogs (e.g. the convincingness of the interpretation, the brevity of the presentation) is conducive to the correction of misinformation (Lewandowsky, 2012). Microblogs with long content often interfere with users’ reading and understanding, achieving a backfire result, thereby affecting the false rumor-refuting effectiveness. Total number of words and punctuation marks are used to measure the length of the information description.

• •The number of pictures and videos (C2)

Multimedia content includes images, texts, videos, etc., reflecting the richness of information (Amato, 2019). The higher the information richness, the higher the information quality, which is more likely to influence the interaction behavior of users. Compared with plain text, multimedia content can provide supplementary information, allowing users to better understand the false rumor-refuting information, which is easier to grab the users’ attention and promote their retweeting behavior (Yan & Huang, 2014).

• •Release time (C3)

According to the research on users’ time pattern released by Sina Weibo enterprise (SocialBeta, 2011), the higher the user activity during the microblog release time, the better the false rumor refutation effectiveness. The research found that users’ interaction enthusiasm in different time periods from low to high is: 2:00–7:00, 0:00–2:00, 14:00–17:00, 7:00–13:00, 13:00–14:00, 17:00–24:00. Therefore, microbloggers publish false rumor-refuting microblogs at high user activity time period can get more users’ feedback.

3.2. Topic extraction and sentiment analysis

LDA model is used for topic extraction (Blei et al, 2003). Perplexity is used to determine the number of topics, which refers to the uncertainty of the model to identify the topic of the documents in text analysis. Therefore, the lower the perplexity and the smaller the uncertainty, the better the final clustering result. Coherence refers to whether the high-frequency words under each topic are consistent in semantics. The larger the coherence, the better the classification of the topic (Stevens, 2012).

To evaluate the sentiment tendency of the topic according to the microblog content, sentiment analysis is applied to convert text content into a value between 0 and 1. The closer the value is to 1, the more positive the text emotional attitude. Baidu AipNLP (Baidu, 2019), as one of the most intelligent Chinese text analysis techniques, can convert text into corresponding sentiment positive probabilities. Therefore, Baidu AipNLP is used to calculate the positive sentiment probability of false rumor-refuting microblogs texts.

3.3. Core conditions extraction

As the number of conditions examined increases, so does the difficulty of interpreting the results (Greckhamer, 2013; Hafner & Pidun, 2022). To reduce the redundancy of conditions, five ensemble machine learning regression models are used to extract core conditions influencing REI: Adaptive Boosting regression model (AdaBoost), Gradient Boosting Decision Tree regression model (GBDT), Extreme Gradient Boosting regression model (XGBoost), Light Gradient Boosting Machine regression model (LGBM) and Random Forest regression model (RF). Grid search is used to optimize the parameters of each model to maximize the prediction effect.

AdaBoost obtains a strong classifier by linear combination of multiple weak classifiers. During the training process, it constantly changes the distribution of training data weights, making the misclassified data receive greater attention in the subsequent round of classification (Ghorbanali, 2022). GBDT, the combination of GB (Gradient Boosting) and DT (Decision Tree), can be regarded as a generalization of AdaBoost, constructing multiple decision trees for data prediction and improving the model by calculating negative gradients (Delgado-Panadero, 2021; Wang, 2022). XGBoost is a further improvement of GBDT, which performs second-order Taylor expansion on the loss function (Sagi & Rokach, 2021). XGBoost's regularization helps avoid overfitting problems. LGBM uses a leaf-wise algorithm with depth constraints, which has higher efficiency and better accuracy than GBDT's level-wise strategy and it has a shorter training time than XGBoost (Kilincer, 2022). RF is a bagging algorithm with a simple structure (Herce-Zelaya, 2020). Unlike the boosting algorithm, there is no dependency between each weak learner, so it can be easily trained in parallel. The above five ensemble algorithms have strong data processing ability and stability. We applied these models by Scikit-learn (Abraham, 2014), XGBoost (Mitchell & Frank, 2017), Lightgbm (Barfar, 2022) libraries in Python.

Error analysis indicators used to evaluate machine learning models’ performance include the mean square error (MSE), the mean absolute error (MAE) and the correlation coefficient ($R^2$), which are respectively defined as follows:

$$MSE=\frac{1}{N}\sum _{n=1}^N{(y_n-{\widehat{y}}_n)}^2$$ (2)

$$MAE=\frac{1}{N}\sum _{n=1}^N\left|y_n-{\widehat{y}}_n\right|$$ (3)

$$R^2=1-\frac{{\sum }_{n=1}^N{(y_n-{\widehat{y}}_n)}^2}{{\sum }_{n=1}^N{(y_n-{\overline{y}}_n)}^2}$$ (4)

Where $N$ is the number of samples, $y_n$, ${\widehat{y}}_n$ and ${\overline{y}}_n$ ($n$=1, 2,$\dots ,N$) are the actual, predicted and average value of REI, respectively.

3.4. Obtaining approach of configurational patterns for REI enhancement

The conceptual model in Fig. 2 depicts specific methodology that this paper applies to explore REI enhancement configurational patterns. First, core conditions extraction. Machine learning regression models are used to extract the core conditions affecting REI enhancement to reduce the redundancy of data, because the increase of conditions can lead to nonsensical theoretical models which increase the difficulty of interpretation (Greckhamer, 2013; Ragin, 2008). Second, data calibration for core conditions. Among the core conditions affecting REI enhancement, the calibration of classification variables adopts manual calibration and the calibration of continuous variables adopts algorithm-based calibration (Fainshmidt, 2020). Third, necessary conditions analysis. Core conditions affecting REI enhancement with consistency greater than 0.9 and adequate coverage are considered necessary conditions (Douglas, 2020), which will be applied in the analysis process later in this paper. Finally, standard analysis. There are three solutions: complex solution, parsimonious solution and intermediate solution (Rihoux & Ragin, 2009), where the intermediate solution only considers simple counterfactual analysis which has moderate complexity (Greckhamer, 2011). Therefore, intermediate solutions are preferred to analyze the pathways of REI enhancement. The closer the solution consistency of the intermediate solution is to 1, the better the false rumor refutation effectiveness.

Fig. 2.

Conceptual model.

4. Data construction

This section states data collection of false rumor information, topic extraction results, condition assignment and REI calculation example.

4.1. Rumors collection

Sina Weibo is one of the most popular social platform applications in China, which is widely used and has strong interactivity (Liu, 2022). Since the outbreak of COVID-19 at the end of 2019, online false rumors have emerged in large numbers. Python crawler is used to collect the latest COVID-19 related false rumor microblogs. A total of 1903 related microblogs from Sina Weibo published or retweeted by “Weibo Rumor Refutation” (the official account of Sina Weibo for false rumor refutation) between January 1, 2022 and April 20, 2022 are obtained. The social media rumor refutation effectiveness lifecycle (SMRREL) shows a long tail distribution. With the rapid update of information, few rumor refutation microblogs attract people's attention for more than 200 h. More specifically, 57.77% of SMRREL lifespan does not exceed 50 h. SMRREL with a lifespan of 50 to 100 h accounted for 27.49%. However, the percentage of SMRREL with a lifespan more than 100 h is only 14.74% (Li, 2022). To avoid inaccurate analysis due to data fluctuations, we crawled the microblog after it was posted 300 h (about two weeks) later. At this moment, the probability of users being attracted by the microblog is less than 1% (Li, 2022), which means the number of retweets, comments and likes is basically stable.

The crawled elements include: the microblogger, the microblog content, the number of retweets, comments, likes and pictures, videos and links attached to the microblog, the number of positive comments and their likes, the number of microblogs published by the microbloggers, the number of microbloggers’ followers.

4.2. Topics extraction

According to the LDA perplexity curve shown in Fig. 3 , the more categories, the less perplexity. However, too many categories may cause over-fitting phenomenon. Combined with the LDA coherence curve, coherence value is the highest when the number of categories is 6. This paper sets 6 types of false rumor-refuting microblogs. The specific type names are summarized by the high-frequency keywords of each type as shown in Table 2 . Topics are ranked from lowest to highest according to the total number of retweets, comments, likes and positive comments.

Fig. 3.

Perplexity and coherence curve.

Table 2.

Type of the topic and the number of different elements.

	Topic types	Amounts	Retweets	Comments	Likes	Positive comments
1	Police reports on rumormongers	448	284,659	70,451	581,930	3176
2	Carrier of virus transmission	202	53,940	49,807	933,184	1597
3	Semester arrangement and work resumption	306	75,409	112,918	1134,031	3139
4	City lockdown and traffic control	182	1234,199	20,152	203,875	1487
5	False rumor refutation of new confirmed cases	478	299,204	214,638	3175,392	3987
6	Personal protection and prevention	287	118,508	378,465	3273,484	2277

4.3. Assignment rules of conditions and calculation example

To extract core conditions by machine learning and perform fsQCA, data assignment in this paper adopts the conditional measurement rules proposed in Section 3.1.1 to Section 3.1.3 and "mean anchor method" in QCA technology. Table 3 shows descriptions and assignment rules of conditions. Table 4 shows an example for microblog REI calculation and conditions assignment.

Table 3.

Descriptions and assignment rules of conditions.

	Conditions	Description	Value
A1	Identity of microbloggers	Individual microblogger	0
		Celebrities and local business microblogger	0.33
		National media microblogger	0.67
		International media microblogger	1
A2	Influence of microbloggers	The number microbloggers’ of followers	/
A3	Followers’ stickiness of microbloggers	The number of comments	/
A4	Activity of microbloggers	The number of microbloggers’ microblogs	/
B1	Type of the topic	Police reports on rumormongers	0
		Carrier of virus transmission	0.2
		Semester arrangement and work resumption	0.4
		City lockdown and traffic control	0.6
		False rumor refutation of new confirmed cases	0.8
		Personal protection and prevention	1
B2	Popularity of the topic	Unimportant	0
		Medium	0.33
		Important	0.67
		Extremely important	1
B3	Sentiment tendency of the topic	Calculated by AipNLP, which converts the text to a value between 0 and 1	Baidu AipNLP
C1	Length of information description	The sum of nouns, verbs, punctuation and links	/
C2	The number of pictures and videos	The sum of pictures and videos	/
C3	Release time	2:00–7:00	0
		0:00–2:00	0.2
		14:00–17:00	0.4
		7:00–13:00	0.6
		13:00–14:00	0.8
		17:00–24:00	1

Table 4.

Example for microblog REI calculation and conditions assignment.

	Conditions	Data	Description (value)
A1	Identity of microbloggers	Hangzhou Daily	Celebrities and local business microblogger (0.33)
A2	Influence of microbloggers	The number microbloggers’ of followers: 2.809 million	/
A3	Followers’ stickiness of microbloggers	The number of comments: 22	/
A4	Activity of microbloggers	The number of microbloggers’ microblogs: 44,392	/
B1	Type of the topic	Topic: #Rumor refutation: “Hangzhou East Station restricts passengers from leaving the province” is false rumors!”#	City lockdown and traffic control (0.6)
B2	Popularity of the topic	Grounded Theory	Important (0.67)
B3	Sentiment tendency of the topic	0.0381	/
C1	Length of information description	The sum of nouns, verbs, punctuation and links: 188	/
C2	The number of pictures and videos	1	/
C3	Release time	2022.01.28 11:06	7:00–13:00 (0.6)
The number of retweets		5	/
The number of positive comments		5	/
The number of likes		7	/
The ratio of influential users of the retweeters		40%	/
Calculation: $REI=-{\log }_2\frac{1}{r(1+k)+p+l+1}=-{\log }_2\frac{1}{5\times (1+40\%)+5+7+1}=4.32$

5. Experimental results

This section analyzes conditions extraction by machine learning and configurational patterns for REI enhancement by fsQCA.

5.1. Conditions extraction based on machine learning

Based on 1903 pieces of microblogs data, 80% of the data is used as the training set and 20% is used as the testing set. The fitting accuracy of five models in Table 5 shows that LGBM has the best performance. The learning curve of LGBM shown in Fig. 4 , which illustrates that the $R^2$ scores of training and testing sets gradually converge with the increase of sample size. This indicates that there is no overfitting problem in the model. Therefore, LGBM is used to extract core conditions influencing REI. Besides, Fig. 5 shows the ranking order and feature values of feature importance. To simplify the expression, the conditions in Fig. 5 are represented by alphabetical numbers used in Table 1. QCA generally selects 4–8 conditions to analysis configuration solutions (Greckhamer, 2013). According to the feature importance ranking of LGBM in Fig. 5, feature value over 200 is the most significant combined with QCA rules. Therefore, five conditions are the appropriate number of conditions to obtain the feasible pathways. The core conditions affecting REI are: length of information description (C1), influence of microbloggers (A2), activity of microbloggers (A4), followers’ stickiness of microbloggers (A3) and sentiment tendency of the topic (B3), respectively.

Table 5.

Fitting accuracy of five models.

Model	Training set (80% of the data)			Testing set (20% of the data)
	MSE	MAE	$R^2$	MSE	MAE	$R^2$
LGBM	1.2917	0.8061	0.8491	2.2968	0.9783	0.8001
AdaBoost	1.9952	1.0404	0.7669	2.6634	1.1007	0.7681
GBDT	1.4237	0.8660	0.8337	2.5610	1.0045	0.7771
XGBoost	1.5796	0.8883	0.8154	2.3938	0.9872	0.7916
RF	1.4955	0.8777	0.8253	2.3691	0.9885	0.7938

Note: Bold font indicates the model with the best performance.

Fig. 4.

The learning curve of LGBM.

Fig. 5.

Feature importance ranking of LGBM.

To verify the accuracy of the results, SHapley Additive exPlanations (SHAP) is used in feature ranking. As it can be seen from Fig. 6 , the top five core conditions are the same as the results of LGBM. The comparison shows the robustness of the core conditions extraction results. Using machine learning for feature ranking is more objective and provides the basis for fsQCA pathways configuration (Rezapouraghdam, 2021).

Fig. 6.

Feature importance ranking of SHAP.

5.2. Configurational patterns for REI enhancement

FsQCA consists of four steps: data calibration, necessary conditions analysis, standard analysis and pathway elements configurations. This section describes the process and criteria for each step mentioned above, the final intermediate solution is the pathways solution for REI enhancement.

5.2.1. Data calibration

The purpose of the calibration is to select a threshold raw data score by which the researcher's judges will reflect whether respondents are “fully in” or “fully out” of the membership set (Douglas, 2020). To achieve this goal, the data for conditions should be converted into fuzzy scores of 0–1 (Ragin, 2008). The data calibration of continuous variables adopts algorithm-based calibration and it is performed according to the standards of 5% (fully out), 95% (fully in) and 50% (crossover point) proposed by Ragin (2000). The extracted conditions calibration information is shown in Table 6 . After converting the fuzzy scores, fsQCA 3.0 is used to analyze combinations of conditions affecting REI.

Table 6.

Calibration information of extracted conditions.

	Conditions	95%Full in	50%Crossover point	5%Full out
A2	Influence of microbloggers	1473.187	14.855	0.550
A3	Followers’ stickiness of microbloggers	19.9	7	0
A4	Activity of microbloggers	12.520	1.148	0.211
B3	Sentiment tendency of the topic	0.573	0.5	0.00046
C1	Length of information description	401.9	216	33

5.2.2. Necessary conditions analysis

Necessary conditions refer to the conditions that must be met for the existence of the interest (Rihoux & Ragin, 2009). A combination of consistency greater than 0.9 can be regarded as the necessary condition for the occurrence of the event (Douglas, 2020). Table 7 shows the analysis results of necessary conditions. The consistency of all conditions is less than 0.9, there is no necessary condition for REI enhancement.

Table 7.

Analysis of Necessary Conditions.

Outcome variable: REI	Consistency	Coverage
A2	0.672271	0.733009
∼A2	0.655408	0.529101
A3	0.841772	0.821960
∼A3	0.471918	0.416978
A4	0.658254	0.711763
∼A4	0.669460	0.543818
B3	0.529175	0.703315
∼B3	0.796850	0.567781
C1	0.610525	0.602864
∼C1	0.715889	0.626241

Note: Symbol “∼” is negation (“NOT”). Symbol “*” indicates the coexistence of the linked variables.

5.2.3. Configurational patterns for REI enhancement based on fsQCA

To examine the configurational patterns for REI enhancement, in the truth table algorithm, raw consistency remains above 0.9 and proportional reduction in inconsistency (PRI) consistency remains above 0.7 (Greckhamer, 2018). An intermediate solution was obtained from the standard analysis based on the five core conditions extracted by LGBM as shown in Table 8 . Intermediate solution is considered to be the first choice for interpretation in QCA research (Rihoux & Ragin, 2009). Compared with complex solution and parsimonious solution, the intermediate solution is reasonable, well-founded and moderate complexity. The necessary conditions will not be eliminated. The raw coverage represents the proportion of cases that can be explained by the conditions’ combination and is generally used as an indicator to examine the interpretation ability of the conditions’ combination for the results (Ham, 2019). The unique coverage represents the number of cases that can only be explained by this combined pathway. As shown in Table 8, there are 3 condition combination pathways for REI enhancement. The solution coverage is 0.680966 and the solution consistency is 0.911039, which shows that these pathways can comprehensively cover and analyze the pattern of REI enhancement during the COVID-19. From Table 8, the raw coverage of all pathways is higher than the unique coverage, indicating that there are cases supporting multiple causal pathways. In addition, the raw coverage of both pathway 1 and pathway 2 is higher than 40%, meaning that these two pathways are the core pathways for REI enhancement.

Table 8.

Pathway elements configurations for REI enhancement.

Pathway	Intermediate Solution	Description	Raw coverage	Unique coverage	Consistency
1	A2*A3	Highly influential microblogger * high followers’ stickiness microblogger	0.588479	0.153295	0.917212
2	A3*A4*∼C1	High followers’ stickiness microblogger * highly active microblogger * concise information description	0.446075	0.0169705	0.928789
3	A3*B3*∼C1	High followers’ stickiness microblogger * the sentiment tendency of the topic * concise information description	0.383284	0.0636859	0.938368
Solution coverage: 0.680966
Solution consistency: 0.911039

Note: Symbol “∼” is negation (“NOT”). Symbol “*” indicates the coexistence of the linked variables.

Table 9 shows the configuration solution more intuitively. Although “high followers’ stickiness microblogger” (A3) is present in every configuration solution, it is not a necessity condition in the causal necessity analysis (Table 7). It just shows that the universality of the condition A3 (Bol & Luppi, 2013). According to the three configuration solutions, analysis results are as follows.

Table 9.

Configuration solution for REI enhancement.

Condition			Configuration solution
			Pathway 1	Pathway 2	Pathway 3
Microbloggers’ conditions	A2	Influence of microbloggers
	A3	Followers’ stickiness of microbloggers
	A4	Activity of microbloggers
Contextual condition	B3	Sentiment tendency of the topic
Content condition	C1	Length of information description		∇	∇

Note: “” indicates the core condition; “” indicates the peripheral condition; “∇” indicates the absent condition; The blank space indicates that the condition may exist or not (Ham, 2019).

For the microbloggers, followers’ stickiness is the core condition in determining REI enhancement. Netizens’ persistent motivation of comments and sharing behaviors had greatly strengthened the users’ stickiness of Weibo (Chiang & Hsiao, 2015). Sharing and interacting behaviors simultaneously contribute to the enhancement of REI. Highly influential microbloggers tend to post more authoritative information, meanwhile netizens are more willing to repost and comment on their microblogs (Li, 2021). Highly active microbloggers are more likely to attract users’ intentions, encouraging them to retweet, comment and like (Li, 2021). For information ontology, the continuous spread of negative emotions leads to psychological stress among netizens and even triggers panic behavior reflected in offline activities, in turn, more positive information may improve the interactive mood of netizens to enhance the effectiveness of false rumor refutation (Yin, 2022). For information description, the length of microblogs is negatively correlated with correction effectiveness and concise false rumor-refuting information is more effective (Zhang, 2022). In conclusion, REI enhancement can be achieved through three different configuration solutions.

6. Discussion and implications

This section describes the importance of core conditions extraction, configurations of pathway elements for REI enhancement, theoretical and practical implications.

6.1. The importance of core conditions extraction

According to the three rules of epidemics proposed by Gladwell (2006), ten influencing conditions of false rumor refutation were initially identified. LGBM with the highest fitting accuracy among five ensemble machine learning regression models is used to extract core features from 10 conditions that affect REI. If core conditions are not extracted by machine learning, as a comparison, Table 10 shows the fsQCA analysis results without feature extraction. It can be seen from the Table 10 that although the solution consistency is close to 1, elements making up the pathways are complex, which is almost difficult to achieve in practical applications. The extraction of core conditions not only reduces data redundancy, but also provides actionability for the final fsQCA analysis results.

Table 10.

Pathway elements configurations for REI enhancement without conditions extraction.

Pathway	Intermediate Solution	Raw coverage	Unique coverage	Consistency
1	A1*∼A2*A3*∼B1*∼B2*∼B3*∼B4*∼C1*C2*C3	0.0401411	0.0113208	0.990745
2	A1*∼A2*A3*∼A4*∼B1*B2*∼B3*∼B4*∼C1*∼C2*C3	0.0369997	0.00885911	0.984328
3	∼A1*A2*A3*A4*∼B1*∼B2*∼B3*∼B4*∼C1*C2*C3	0.0194628	0.00748833	1
4	A1*A2*A3*A4*B1*B2*∼B3*∼B4*∼C1*C2*C3	0.0379331	0.0165559	0.994978
solution coverage: 0.0774852
solution consistency: 0.990156

Note: Symbol “∼” is negation (“NOT”). Symbol “*” indicates the coexistence of the linked variables.

6.2. Configurations of pathway elements for REI enhancement

According to the configuration solutions shown in Table 10, to enhance the REI, the influence of microbloggers (A2), the activity of microbloggers (A4) and the sentiment tendency of the topic (B3) severally has different degree of importance in different pathways. The length of information description (C1) should be strictly limited, because concise information description facilitates REI enhancement.

The constituent elements of the three paths indicate that the fsQCA analysis results contribute to REI enhancement from three aspects: (1) microbloggers, (2) microbloggers and features of false rumor-refuting microblogs and (3) the combination of the three condition categories.

Pathway 1

: highly influential microblogger (A2) * high followers’ stickiness microblogger (A3)

This pattern can be interpreted as “highly authoritative and interactive information source”. Highly influential microbloggers tend to be extremely authoritative, such as CCTV News, CGTN, Xinhua News Agency and other national certified accounts. By providing professional and high-quality false rumor-refuting content, coupling with timely comment interaction and question communication with users, the REI enhancement will be greatly improved. Therefore, increasing the user stickiness of highly influential microbloggers can effectively improve the effect of refuting false rumors. High followers’ stickiness means more positive comments which can promote the interaction of users and the dissemination of false rumor refutation information.

Pathway 2

: high followers’ stickiness microblogger (A3) * highly active microblogger (A4) * concise information description (C1)

This pattern can be interpreted as “highly interactive and active information source with concise expression”. Highly active microbloggers can ensure that false rumor-refuting information is released in a timely manner and have a high coverage rate for various false rumor rebuttals. On this basis, the followers’ stickiness should be ensured at the same time so that users can get official answers to their questions about different kinds of misinformation in time. Furthermore, concise information expression means describing the complete information content in a concise length which is recommended to be less than 500 words (Zhang, 2022). The false rumors refutation effectiveness drops sharply when the length of the microblog content is too long. This is the reason why the concise language is more accessible and eye-catching for REI enhancement.

Pathway 3

: high followers’ stickiness microblogger (A3) * sentiment tendency of the topic (B3) * concise information description (C1)

This pattern can be interpreted as “highly interactive information source with concise and positive expression”. Negative expressions may cause psychological pressure on netizens. Besides, negative emotions are infinitely amplified due to the echo chamber effect (Wang, 2022). The large-scale accumulation of negative emotions mainly depends on the user's participation and replication of emotional choices (Yin, 2022). In consequence, highly interactive microbloggers should spread positive emotions to users and maintain the stability of the network environment. With concise information description, REI enhancement can achieve better results.

6.3. Theoretical and practical implications

The theoretical contributions are as follows. First, this paper collects 10 conditions that may affect REI. To prevent data redundancy and complexity of final results, LGBM, the model with the best performance of five machine learning models, is used to extract core conditions. This lays the data foundation for the subsequent fsQCA analysis. Second, in this study, fsQCA is selected as the method to analyze REI enhancement. It is a research method between case-oriented and variable-oriented, which combines qualitative and quantitative methods to analyze causality. Through fsQCA, this paper identifies three configuration pathways for REI enhancement and then interprets the findings.

On the practical front, five core conditions are extracted: influence of microbloggers, followers’ stickiness of microbloggers, activity of microbloggers, sentiment tendency of the topic and length of information description. In the subsequent work of false rumor refutation, highly authoritative microbloggers should continuously improve followers’ stickiness and activity, and also publish positive and concise false rumor-refuting microblogs to improve the effectiveness. In addition, this paper provides method guidance for false rumor-refuting platform, which can use machine learning models to accurately locate the core conditions that affect the effectiveness of false rumor refutation. It not only reduces the workload in the future, but also increases the feasibility of improving the rumor refutation effectiveness. Besides, although the COVID-19 has become one of the world's major public health challenges, and its related rumors would also continue to affect the public, this study may provide a reference for the implementation of the COVID-19 related false rumor refutation plans.

7. Conclusion

To curb the rapid spread of infodemics during the COVID-19, this paper explored REI enhancement from the perspective of configuration patterns. There are two key steps in this paper: the extraction of core conditions based on machine learning algorithms and configurations of pathway elements. The extraction of core conditions based on LGBM reduces the data redundancy and increases the rationality of the results. According to the data calibration, necessary conditions analysis and truth table algorithm of fsQCA, the configurations of pathway elements is obtained for REI enhancement.

Despite the contributions and implications of this study, there are still some limitations. (1) There is great potential to apply the proposed machine learning and fsQCA method to other social media platforms besides Weibo or other areas of rumor refutation, which may have different findings compared with Weibo. (2) More conditions will be considered when measuring microbloggers’ influence. For example, Weibo view volume, the number of interaction. (3) This paper identifies three pathways patterns of REI enhancement. However, other configuration patterns are likely to exist in reality and we need to study further. (4) Only ten conditions are initially determined, but five core conditions are finally extracted. In the complex network environment, there may be some other important conditions that have not been mined, and more exploration is needed in the future.

CRediT authorship contribution statement

Zongmin Li: Conceptualization, Funding acquisition, Writing – original draft, Methodology, Writing – review & editing. Ye Zhao: Data curation, Formal analysis, Writing – original draft. Tie Duan: Data curation, Formal analysis. Jingqi Dai: Funding acquisition, Supervision, Writing – review & editing.

Data availability

• No data was used for the research described in the article.