help

Twitter Sentiment Analysis of Real-time Customer Experience Feedback for Predicting Growth of Indian

Telecom Companies

Sandeep Ranjan1, Sumesh Sood2, Vikas Verma3* 1Research Scholar, Inder Kumar Gujral Punjab Technical University, Kapurthala, India and Assistant Professor, Lyallpur

Khalsa College of Engineering, Jalandhar, [email protected] 2Inder Kumar Gujral Punjab Technical University, Kapurthala

3Lovely Professional University, Phagwara, India, [email protected]

Abstract— Corporations have always desired prompt customer experience feedback about their products for amending current pricing and policies to stay ahead of their competitors. A positive customer experience can be created by analyzing customer sentiments and acting on them promptly. Social networks like Twitter represent collective intelligence and opinion of the general public and hence can be harnessed for real-time feedback. They have evolved as a resource for extracting sentiments for applications in various fields. Sentiment analysis can be used to obtain the overall customer experience of a large customer base on a real time. In this research, a total of 153,651 distinct tweets for Twitter handle of 5 popular telecom brands in India: Aircel, Bharti Airtel, Idea Cellular, Reliance Jio and Vodafone India were extracted for five months to develop a prediction model for telecom subscriber addition using the sentiment score. The results were validated statistically using correlation analysis. Positive customer sentiments about the brand which they prefer is reflected by higher growth rate of new subscribers added with that brand in the study period. The sentiment analysis results can be used by managements to take timely actions for improving the future customer experience and avoiding customer churn.

Keywords- Social network, text mining, customer experience, sentiment analysis, opinion mining.

I. INTRODUCTION

The world is getting more and more interconnected both economically and socially thanks to the recent advancements in the Internet and networking technologies [1]. There has been a remarkable rise in the last few years in the number of Internet users. Around 4.1 billion people have access to the Internet representing about half of the global population and this number has been growing at an enormous rate since the year 2000, primarily due to the smartphone revolution in the telecom sector [2]. People create their own content, share videos, images or repost other user’s content [3]. This content is related to the user’s social activities and personal experiences about different

services, products and events. Social networks are helping businesses grow and make profits around the globe [4]. It helps them advertise their upcoming services and products at much lower cost compared to traditional marketing models and maintain a close relationship with customers. Upcoming festivals, music albums, movies, sporting events and new product releases create a buzz among the general public much ahead of the release and this can be gauged through social network responses related to them [3]. Thus, social networks are finding their way in business models as they are helpful in recognizing new opportunities and threats.

Users have long friend lists, and they participate in discussions of mutual interest. This way a lively, complex and enormous network of relationships evolves on a social network where influences are created depending upon the number and nature of ties among entities. For each and every popular event or opinion shared by the users there can be a large number of connected responses by other users. This leads to an accumulation of a huge collection of text data posted on social networks, which represents the collective wisdom of the general public.

Corporations spend a huge amount of money and time on brand monitoring and collection of real-time customer experience feedback to keep a watch on their sales and revenues. Most of the traditional methods of monitoring and feedback have a huge cost involved with them and have latency issues. The feedback results are summarized and presented to competent authorities of higher management after a delay of a few days to a couple of weeks from the actual start of the feedback gaining process. Within this span, market scenarios and equations get changed. The delayed planning can lead to drop in market position and further corrective actions occur after a further lag of time. Sales figures for Fast Moving Consumer Goods (FMCG), telecom service providers, box office collections, etc., which are dependent on a large consumer population spread on a vast geographic area are difficult to be monitored day in and day out.

The paper proposes a model for predicting the growth of

Indian Telecom Companies in terms of subscriber addition

166

2018 4th International Conference on Computing Sciences

978-1-5386-8025-4/18/$31.00 ©2018 IEEE DOI 10.1109/ICCS.2018.00035

by analyzing sentiments from tweets posted on their Twitter handles. This paper has been organized as follows: Section 2 of the covers the review of the related work. Section 3 covers the dataset creation. Section 4 covers the process of pre-processing, community detection, sentiment analysis and the proposed model of prediction.

II. RELATED WORK A. Social Networks

The diffusion of information on social media is at an enormous scale and real-time analysis techniques focus on it to provide predictive inferences [5]. People prefer the social network endorsement of brand over other types of endorsements [6], [7]. Social networks provide electronic word of mouth option to the public, which influences them more and its measurement is also feasible [8], [9]. Researchers discuss how people seek opinions from others be it everyday chores or national elections emphasizing the need for opinion mining. Respondents may not be comfortable with the traditional feedback methods or in some cases don’t have time for feedback surveys. Respondents may not report their true opinions while filling written questionnaires, but they let out their feelings on social networks when they like or dislike certain brands and companies. In the research carried out, social network sentiment analysis has been used to provide a solution to this problem. Social network data can be fetched real-time and requires less time and effort to summarize and lead to conclusions.

Twitter is one of the most popular social network

websites. It is a microblogging service that lets registered users Tweet (post) their views on any topic [10], [11]. It has become a popular mode of sharing one’s views about current trends and events. The ‘following and being followed’ relationship on Twitter does not require reciprocation as is the case with many other social networking websites. Twitter was selected for mining user opinions as it has millions of registered users who every day more than 500 million Tweets and also that the tweets are made available to the general public using APIs within specified rate limits [12]. Twitter uses Open Authentication (OAuth) for APIs requesting information from it [13]. The crawled data includes pairs of userids, tweet, the relation between userids, relationship date, hashtags in Tweet etc. B. Sentiment Analysis

Sentiment analysis is the process of analyzing opinions and views of authors about a particular entity [19], [20]. The main objective of the sentiment analysis process is to identify opinions, identify their sentiments and classify the polarity [21], [22]. It is one of the fastest evolving research areas in computer science. The decision making of individuals is affected by the opinions formed by collective reviews from others. This decision making has shifted from a few personal references to a huge number of social media

user opinions from across the globe. With the growth of social media websites, the size of the opinion dataset has grown enormously that it can no longer be manually parsed to generate an inference leading to the evolution of sentiment analysis processes using specialized software.

Sentiment analysis software or sentiment analysis systems use natural language processing (NLP) for processing text articles to identify the underlying relationships [23]. Sentiment analysis focuses on how sentiments are expressed in a text and also if the sentiment points to a positive (favorable) opinion or a negative (unfavorable) opinion on the subject in question [24]. Therefore the analysis focuses on capturing expressions with sentiments, their polarity and relationship with the subject. Consider an expression “A is taller than B”. Here the sentiment “taller” signifies a positive sentiment for A but a negative sentiment for B. Social media sentiment analysis consists of three processes: fetching data, analyzing the data and finally presenting meaningful interpretations of the data [25]. The work presents an overview of more than 50 articles on Twitter sentiment analysis and discussed popular methods for opinion retrieval, irony detection and tracking sentiment over time [26]. Open issues in sentiment analysis like multilingual content, data sparsity, use of deep learning, lack of benchmarks and multidisciplinary research were discussed.

Figure 1 describes the lexical analysis of a phrase of the

given text. The tokens representing an entity or a concept are identified and are given due weight in determining positive or negative opinion. Here the tokens “ABC” and “DEF” represent entities, “better” and “lower” help in comparing the identified entities ABC and DEF; and “33” and “55” are used in assigning numerical values to weights.

Figure 1. Text lexical analysis

Sentiment analysis process comprises of identifying the

polarity of user-defined concepts and entities by letting users define their own dictionary, imparting flexibility to be applied to any scenario. Additionally, it determines if the given text being processed is subjective or objective. It is also checked if the text contains some distinctive text keywords at a global level, giving additional information about the reliability of the polarity.

Social media has emerged as a platform for brand managers to generate and work on brand specific content [27]. Sentiment analysis can be applied to automate

167

categorization of Tweets as neutral, positive or negative. In recent years, the trend of sharing health-related issues on social networks has increased analyzing user posted content has come up as a challenge in semantic analysis [28]. Adverse Drug Reactions (ADRs) which are a leading cause of deaths in patients admitted in hospitals. Patients on their own reported on social media about ADRs on social media. Twitter and other forums focusing on Spanish respondents were explored by the TrendMiner project to monitor the content related to drugs and related reactions. Social networks are vast repositories of public created data that contain information representing a general mood about brands and events [26]. In the airline industry, collecting feedback from customers is an important task for which the conventional methods are inefficient and incorrect [29]. The study created a dataset of 107,866 for the major airlines in North America. Tweets were labeled as positive, negative and neutral for the segregation process.

Performance of various sentiment classification approaches was compared to develop an ensemble approach. As per the study, in the airline industry, sentiment-analysis and classification is accurate enough to be used for customer satisfaction investigation [30]. The researchers used social media opinion mining to Poland’s 2010 political crisis discussion dataset containing a high conflict level and deep polarization [31]. The study also analyzed the relation between online sentiment and its impact on Java Governing Board’s open sourcing decision of Java.

The researchers studied the application of sentiment analysis process in economic and financial modeling and concluded that Tweet sentiment analysis is one of the best methods to automate the sentiment analysis process [32]. They proposed to develop an ontology-based method for segmenting individual Tweet rather than treating each Tweet as a single expression being assigned a sentiment score. The research suggested that a brand’s social network is an effective way of attracting consumers and attaching them emotionally to the brand [33]. Also, existing consumer’s bonding with the brand is reflected in the brand’s social network on different social network websites. Sentiment analysis helps organizations to scrutinize social media content on a real-time basis and act accordingly [34]. Sentiment analysis of social media can help in picking promising stocks for better returns.

III. RESEARCH METHOD

Social networks have become a valuable source of knowledge in the form of sentiments for a number of sectors such as public opinion management, brand building customer relationship management [35]. Social networks allow APIs to query them for a particular keyword generally superseded by # (hashtag). As most of the social networks have users of the order of millions, any current news item, trend, product, organization or service is flooded with numerous user opinions on them.

This information is of great use for governments and corporate managers to assess the general public’s opinion and capture experience feedback. Current trends, marketing promotions and media release get quick attention and the user feedback can be accessed and monitored on a real-time basis.

a. DATASET CREATION

In the research, tweets for #aircel, #airtelindia, #ideacellular, #reliancejio, #vodafonein the Twitter handles for the Telecomm companies, were extracted on a daily basis from 1st March 2017 to 31st July 2017. This period is selected for research as Reliance Jio had announced to start charging its customers for all its services from 1st April, 2017 (though actual charging of services started on 15th April). This could be a real test for Reliance Jio to analyze if the general public is inclined towards its services or not amidst the presence of operators like Aircel, Bharti Airtel, Idea Cellular and Vodafone India which were operating for many years in the past. The combined sentiment of the general public needed to be compared to check if the general public is showing interest towards the new entrant Reliance Jio or existing market leaders, when all of them were now offering mobile services at similar competitive tariffs.

The number of tweets extracted on a particular instance is restricted by the Twitter rate limiting. During the said period, 18,457 tweets for #aircel, 37,546 tweets for #airtelindia, 14,592 tweets for #ideacellular, 51,478 tweets for #reliancejio and 31,578 tweets for #vodafonein were extracted. Then, the filter was applied to this dataset on the Tweet column containing Tweet content to remove duplicate tweets generated from retweets. As a result, distinct tweets for #aircel, #airtelindia, #ideacellular, #reliancejio, #vodafonein were 4045, 7674, 5346, 10,591 and 7894 respectively. The tweets in an individual dataset of #aircel, #airtelindia, #ideacellular, #reliancejio, #vodafonein have each other’s mention in some cases where the user wants to compare them in terms of different parameters like network usage charges, coverage, call drops, Internet speed etc. Figure 2 shows a snapshot of dataset of #reliancejio dataset. b. DATASET PROCESSING

This process discards the incomplete, incorrect or irrelevant data. Tweets contain advertisement links of different companies involved in paid marketing, and other irrelevant data. Many users post tweets in Non-English language text which cannot be used in the general text processing or Natural Language Processing. It is required to remove such data from the dataset. This module identifies and removes the unwanted data and the cleaned data are passed to the next module.

168

A. COMMUNITY DETECTION

Popular tweets attract user attention and get shares, retweets and responses. Multiplicity of edges amongst nodes leads to community evolution and growth. The communities are the channels for online word of mouth propagation. The

research used Python libraries to detect communities in the datasets by using the betweenness centrality measure of network graphs. The betweenness centrality of a node v, BC(v) is defined as:

Figure 2 Snapshot of the tweets for #reli ancejio

Semantic analysis was performed on each of the

hashtag's datasets to classify the Tweets into five levels of polarity N+, N, NEU, P and P+. N+ and N represent a negative polarity generated for a negative opinion in the text, whereas P and P+ represent positive polarity. NEU polarity is generated for neutral opinion or when the polarity cannot be calculated. Table I shows weights assigned to the standard polarities. NEU category of polarities Tweets contains neutral opinion about the concept or entity and hence they have been given a 0 (zero) weight. Negative opinion Tweets (N and N+) have been given -1 and -2 weight depending upon the magnitude of negative opinion.

Similarly positive opinion Tweets (P and P+) have +1 and +2 weights. A sentiment dictionary is a critical part of sentiment analysis to recognize the sentiment tokens in any document [36]. This dictionary contains words, phrases, related concepts and sentiment polarity and valuable information used to identify certain significant phrases from the source and establish agreements and weight for them. Table II shows the custom dictionary created for performing sentiment analysis of the Tweet datasets. Some of the most influential keywords for describing opinion about likes and dislikes for a telecom brand have been included in this dictionary to capture the essence of user comments.

TABLE I. POLARITY AND WEIGHTS Polarity N+ N NEU P P+

Weight -2 -1 0 1 2

TABLE II. SENTIMENT ANALYSIS DICTIONARY

Keyword App Call drop Cost Kbps Mbps Network coverage Port to Signal Support

Type Entity Concept Concept Concept Concept Concept Concept Entity Concept

Figure 3 describes the overall architecture of the sentiment analysis based prediction model for monthly subscriber addition of Telecomms. The sentiment analysis was separately performed on datasets of distinct tweets of #aircel, #airtelindia, #ideacellular, #reliancejio, #vodafonein. The sentiment score of a particular Tweet is calculated by multiplying its identified polarity with the weight assigned to that polarity. The total sentiment score of a hashtag is obtained by summing up the score of all tweets and the following results were obtained.

Table III data and figure 4 show the polarity distribution for tweets of #aircel, #airtelindia, #ideacellular, #reliancejio, #vodafonein for the base month March, 2017. It is evident that there are a significant number of Tweets with polarity value 0. These are the neutral sentiment Tweets for which positive or negative sentiment could not be determined. The frequency of neutral sentiment does not contribute to the overall sentiment score.

169

Figure 3 Sentiment based subscriber addition data prediction architecture

TABLE III. Polarity distribution for Hashtags for March 2017

Hashtag N+ N NEU P P+ Total tweets #aircel 641 728 650 1187 839 4045

#airtelindia 1178 1252 897 2245 2102 7674 #ideacellular 947 474 1127 1875 923 5346 #reliancejio 921 621 1458 5189 2402 10591 #vodafonein 1315 947 874 2187 2571 7894

Figure 4. Sentiment count for positive and negative tweets of #aircel, #airtelindia, #ideacellular, #reliancejio, #vodafonein

Validation (Correlation Analysis)

Web

www.twitter.com

Subscriber monthly addition data Telecom hashtag tweets

www.coai.com

Text preprocessing and cleaning

Subscriber addition prediction

Sentiment Analysis

Natural Language Processing

Ontology assignment

170

Figure 4 shows the number of positive and negative sentiment tweets for #aircel, #airtelindia, #ideacellular, #reliancejio, #vodafonein. For #airtelindia and #vodafonein, the number of negative sentiment is greater than the positive sentiment and hence the overall score of #airtelindia and #vodafonein is negative. The overall sentiment score of #reliancejio is positive as the positive sentiment count is greater than the negative sentiment count. Sentiment count without weight or polarity does not carry the essence of the dataset. Sentiment score is calculated by multiplying number of tweets having a particular polarity with the weight of that polarity.

Table IV data shows that the overall sentiment score of #aircel, #airtelindia, #ideacellular, #reliancejio, #vodafonein is 855, 2841, 1353, 7530 and 3752 respectively.. The score indicates that the general public has a comparatively stronger positive opinion about Reliance Jio, the new entrant

in the Telecomm field. It is a reflection of respondent’s individual experience about the brand in terms of different parameters based on different issues related to mobile network service and pricing. This affects their bonding with the brand they are using or the decision making for the brand to which they want to switch over to.

Table V shows the month wise overall sentiment score for the study duration. Table VI shows the month wise subscriber addition data. The data for the number of subscribers added each month for Aircel, Bharti Airtel, Idea Cellular, Reliance Jio and Vodafone India was downloaded from the website of the Cellular Operators Association of India [37]. The data for the March month has been used as base data. The difference of sentiment score of an operator and its previous month sentiment score gives the growth rate percentage.

TABLE IV. TELECOMM SENTIMENT SCORES FOR MARCH 2017

Polarity → N+ N NEU P P+ Total

Weight (W) → -2 -1 0 1 2

No of tweets for #aircel (AC) →

(W*AC) →

641 728 650 1187 839 4045

-1282 -728 0 1187 1678 855

No of tweets for #airtelindia (AT) →

(W*A) →

1178 1252 897 2245 2102 7674

-2356 -1252 0 2245 4204 2841

No of tweets for #ideacellular (I) →

(W*A) →

947 474 1127 1875 923 5346

-1894 -474 0 1875 1846 1353 No of tweets for

#reliancejio (R) → (W*R) →

921 621 1458 5189 2402 10591

-1842 -621 0 5189 4804 7530 No of tweets for

#vodafonein (V) → (W*V) →

1315 947 874 2187 2571 7894

-2630 -947 0 2187 5142 3752

TABLE V. TELECOMM TWEET COMMUNITY SENTIMENT SCORE (PREDICTED GROWTH RATE) APRIL MAY JUNE JULY

March Base

Sentiment Score

Growth rate (%)

Sentiment Score

Growth rate (%)

Sentiment Score

Growth rate (%)

Sentiment Score

Growth rate (%)

#aircel 855 850 -0.35 855 0.35 851 -0.47 847 -0.47

#airtelindia 2841 2871 1.06 2895 0.84 2916 0.73 976 -66.53

#ideacellular 1,353 1347 -0.44 1350 0.22 1351 0.07 1334 -1.26

#reliancejio 7,530 10722 42.39 11676 8.90 12868 10.21 1554 -87.92

#vodafonein 3752 3767 0.40 3788 0.56 3806 0.48 3779 -0.71

TABLE VI. TELECOMM MONTHLY SUBSCRIBER DATA (ACTUAL GROWTH RATE)

APRIL MAY JUNE JULY

March Base Subscriber Addition

Growth rate (%)

Subscriber Addition

Growth rate (%)

Subscriber Addition

Growth rate (%)

Subscriber Addition

Growth rate (%)

Aircel 90,899,868 -339427 -0.4 173091 0.2 -410128 -0.5 (391,321) -0.4 Bharti Airtel 273648383 2853044 1.0 2098006 0.8 2047884 0.7 602771 0.2

Idea Cellular 195,368,847 683858 0.4 190043 0.1 35541 0.0 -2320480 -1.2

Reliance Jio 72,157,644 30683394 42.5 9309227 9.1 11213605 10.0 5216737 4.2

Vodafone India 209062866 756637 0.4 1131435 0.5 988424 0.5 -1389854 -0.7

171

TABLE VII. MONTH-WISE PREDICTED AND ACTUAL GROWTH RATE: CORRELATION ANALYSIS

(a) (b)

(c) (d)

IV. VALIDATION OF RESULTS

The experimental results obtained from the proposed model were validated by applying correlation analysis. Correlation analysis was carried out on the month wise predicted growth rate (sentiment score growth rate %) and actual growth rate (subscriber addition growth rate %) for the operators using IBM SPSS 24. The results are shown in Table VII (a), (b), (c) and (d). The calculated correlation between the two tested variables is within the significance limits validating the results of the prediction model. The Twitter sentiment score obtained from the dataset communities which is the online word of mouth representing wisdom of the crowds accurately predicts the popularity and success of the Telecomm operators. The success of Telecomm operators is reflected in the month wise subscriber addition data.

CONCLUSION

Monitoring customer sentiments for real-time customer experience feedback is a crucial task for planning and maintaining their business models for competitive success and survival. The research experiment used Twitter mining and community sentiment analysis to fetch real-time sentiments of the masses. During the test period, the proposed model using sentiment score obtained from

sentiment analysis of Twitter hashtag datasets of Indian Telecomm operator handles successfully predicted the growth rate of the Telecomm operators in terms of subscriber addition. Online word of mouth helps formation of dense network communities and affects both existing and prospective customers of a brand. Custom dictionary based sentiment analysis of community datasets is used to obtain the overall sentiment score which forms the basis of growth rate prediction.

The predicted growth rate and actual growth rate show a high degree of correlation validating the prediction model. The results were consistent for all the four months of study duration. Reliance Jio was the biggest gainer amongst the operators in terms of the sentiment score as well as subscriber addition. The addition of a large number of new subscribers to a company’s subscriber base can be seen as customer churn of subscribers shifting away from other companies which failed to monitor the customer sentiment and act in time to prevent it.

A positive sentiment score of a company is an indicator

of the brand preference of the public and a negative sentiment score indicates customer dissatisfaction or inclination towards any other company which is better

APRIL MONTH Predicted

growth rate (%)

Actual growth rate (%)

Predicted growth rate (%)

Pearson Correlation 1 .993

Significant value .0001

Number of companies 5 5

Actual growth rate (%)

Pearson Correlation .993 1

Significant value .0001

Number of companies 5 5

MAY MONTH Predicted

growth rate (%)

Actual growth rate (%)

Predicted growth rate (%)

Pearson Correlation 1

.949

Significant value .025

Number of companies 5 5

Actual growth rate (%)

Pearson Correlation .924 1

Significant value .025

Number of companies 5 5

JULY MONTH Predicted

growth rate (%)

Actual growth rate (%)

Predicted growth rate (%)

Pearson Correlation 1 .893

Significant value .041

Number of companies 5 5

Actual growth rate (%)

Pearson Correlation .893 1

Significant value .041

Number of companies 5 5

JUNE MONTH Predicted

growth rate (%)

Actual growth rate (%)

Predicted growth rate (%)

Pearson Correlation 1 .865

Significant value .058

Number of companies 5 5

Actual growth rate (%)

Pearson Correlation .865 1

Significant value .058

Number of companies 5 5

172

suited to their requirements. Data mining and sentiment analysis techniques can be used by managers to take timely actions to predict and prevent such customer churn [38], [39].

REFERENCES

[1] J. Poushter, “Smartphone Ownership and Internet

Usage Continues to Climb in Emerging Economies,” Pew Research Center, pp. 1–5, 2016.

[2] S. I. World, “World Internet Users Statistics and 2018 World Population Stats,” Internet world stats, 2018. [Online]. Available: https://www.internetworldstats.com/stats.htm. [Accessed: 26-May-2018].

[3] S. Ranjan and S. Sood, “Exploring Twitter for Large Data Analysis,” International Journal of Advance Research in Computer Science and Software Engineering, vol. 6, no. 7, pp. 325–330, 2016.

[4] Y. M. Li and Y. L. Shiu, “A diffusion mechanism for social advertising over microblogs,” Decision Support System, vol. 54, no. 1, pp. 9–22, 2012.

[5] I. Taxidou and P. Fischer, “Realtime analysis of information diffusion in social media,” Proc. VLDB Endowment, vol. 6, no. 12, pp. 1416–1421, 2013.

[6] J. Prager, “Open-Domain Question–Answering,” Foundations and Trends® in Information Retrieval, vol. 1, no. 2, pp. 91–231, 2006.

[7] L. Harris and A. Rae, “Social networks: The future of marketing for small business,” Journal of Business Strategy, vol. 30, no. 5, pp. 24–31, 2009.

[8] S. K. Janane, M. S. Keerthana, and B. Subbulakshmi, “Hybrid Classification for Sentiment Analysis of Movie Reviews,” International Journal of Engineering Sciences and Research, vol. 7, no. 4, pp. 724–728, 2018.

[9] S. Ranjan, “Online Word of Mouth Communication in Bollywood Tweet Dataset,” International Journal for Research in Applied Science & Engineering Technology, vol. 5, no. 12, pp. 1442–1449, 2017.

[10] Kwak, H., Lee, C., Park, H., Moon, S., 2010. What is Twitter, a Social Network or a News Media?, in proc: The International World Wide Web Conference Committee (IW3C2), 2010, pp. 1–10. https://doi.org/10.1145/1772690.1772751

[11] W. X. Zhao et al., “Comparing Twitter and Traditional Media Using Topic Models,” in Proc. European Conf. on Information Retrieval, 2011, pp. 338–349.

[12] “Twitter Usage Statistics - Internet Live Stats,” 2018. [Online]. Available: http://www.internetlivestats.com/twitter-statistics/. [Accessed: 07-May-2018].

[13] S. Kumar, F. Morstatter, and H. Liu, “Twitter Data Analytics,” Springer, p. 89, 2013.

[14] T. Press Release, “Telecom Regulatory Authority of India,” July, no. 65, pp. 1–19, 2017.

[15] “Statistics.” [Online]. Available: http://www.itu.int/en/ITU- D/Statistics/Pages/stat/default.aspx. [Accessed: 29-

Jun-2017]. [16] “About Bharti Airtel.” [Online]. Available:

http://www.airtel.in/about-bharti/about-bharti-airtel. [Accessed: 25-May-2018].

[17] “Company History | About Vodafone India.” [Online]. Available: https://www.vodafone.in/about- us/company-history?section=consumer. [Accessed: 23-May-2018].

[18] “Products & Brands :: Reliance Industries Limited.” [Online]. Available: http://www.ril.com/OurCompany/ProductsAndBrand s.aspx. [Accessed: 20-May-2018].

[19] S. Bharathi and A. Geetha, “Sentiment Analysis for Effective Stock Market Prediction,” International Journal Intell. Eng. Syst., vol. 10, no. 3, pp. 146– 153, 2017.

[20] S. Bharathi, A. Geetha, and R. Sathiynarayanan, “Sentiment analysis of twitter and RSS news feeds and its impact on stock market prediction,” International Journal Intell. Eng. Syst., vol. 10, no. 6, pp. 68–77, 2017.

[21] W. Medhat, A. Hassan, and H. Korashy, “Sentiment analysis algorithms and applications: A survey,” Ain Shams Eng. J., vol. 5, no. 4, pp. 1093–1113, 2014.

[22] T. Wilson, J. Wiebe, and P. Hoffman, “Recognizing contextual polarity in phrase level sentiment analysis,” Acl, vol. 7, no. 5, pp. 12–21, 2005.

[23] M. Kanakaraj and R. M. R. Guddeti, “NLP based sentiment analysis on Twitter data using ensemble classifiers,” in 2015 3rd International Conference on Signal Processing, Communication and Networking, ICSCN 2015, 2015.

[24] T. Nasukawa, T. Nasukawa, J. Yi, and J. Yi, “Sentiment analysis: Capturing favorability using natural language processing,” Proc. 2nd International Conf. Knowl. capture, pp. 70–77, 2003.

[25] W. Fan and M. D. Gordon, “The Power of Social Media Analytics,” Commun. ACM, vol. 57, no. 6, pp. 74–81, 2014.

[26] A. Giachanou and F. Crestani, “Like it or not: A survey of Twitter sentiment analysis methods,” ACM Comput Surv, vol. 49, no. 2, p. Article 28; 1- 41, 2016.

[27] X. Liu, A. C. Burns, and Y. Hou, “An Investigation of Brand-Related User-Generated Content on Twitter,” J. Advert., vol. 46, no. 2, pp. 236–247, Apr. 2017.

[28] P. Martínez, J. L. Martínez, I. Segura-Bedmar, J. Moreno-Schneider, A. Luna, and R. Revert, “Turning user generated health-related content into actionable knowledge through text analytics services,” Comput. Ind., vol. 78, pp. 43–56, 2016.

[29] Y. Wan and Q. Gao, “An Ensemble Sentiment Classification System of Twitter Data for Airline Services Analysis,” Proc. - 15th IEEE International Conf. Data Min. Work. ICDMW 2015, pp. 1318– 1325, 2016.

[30] M. M. Mostafa, “An emotional polarity analysis of consumers’ airline service tweets,” Soc. Netw. Anal. Min., vol. 3, no. 3, pp. 635–649, Sep. 2013.

173

[31] P. Sobkowicz, M. Kaschesky, and G. Bouchard, “Opinion mining in social media: Modeling, simulating, and forecasting political opinions in the web,” Gov. Inf. Q., vol. 29, no. 4, pp. 470–479, 2012.

[32] E. Kontopoulos, C. Berberidis, T. Dergiades, and N. Bassiliades, “Ontology-based sentiment analysis of twitter posts,” Expert Syst. Appl., vol. 40, no. 10, pp. 4065–4074, 2013.

[33] H. Park and Y. K. Kim, “The role of social network websites in the consumer-brand relationship,” J. Retail. Consum. Serv., vol. 21, no. 4, pp. 460–467, 2014.

[34] R. Feldman, “Techniques and applications for sentiment analysis,” Commun. ACM, vol. 56, no. 4, pp. 82–89, 2013.

[35] M. M. Mostafa, “More than words: Social networks’ text mining for consumer brand sentiments,” Expert Syst. Appl., vol. 40, no. 10, pp. 4241–4251, 2013.

[36] A. C. Tsai, “Building a Concept- Level Sentiment on Commonsense Knowledge,” Ieee Intell. Syst., vol. MARCH/APRI, pp. 22–30, 2013.

[37] “Cellular Operators Association of India.” [Online]. Available: https://coai.com/#research-and-reports. [Accessed: 17-Jul-2017].

[38] A. M. Almana, M. S. Aksoy, and R. Alzahrani, “A Survey On Data Mining Techniques In Customer Churn Analysis For Telecom Industry,” J. Eng. Research Appl., vol. 4, no. 5, pp. 165–171, 2014.

[39] W. Verbeke, D. Martens, and B. Baesens, “Social network analysis for customer churn prediction,” Appl. Soft Comput. J., vol. 14, no. PART C, pp. 431–446, 2014.

174