Final project submission by Adil Haris
Andrew ID: aharisib
Email: [email protected]
Twitter has become the go-to platform for addressing consumer complains. Consumers often find it very convenient to tag a company on Twitter with the complaint they have and post their tweet, thus attracting the eye of their connections as well as the followers of the company. Attempting to minimize damage to digital presence, brands have become very proactive in addressing such complains on Twitter. This paper uses techniques in Machine Learning to classify the various types of airline-related negative complaints posted on Twitter given the text of the tweet. The output of this paper may be of interest to brands who are keen on exploring ways to quickly and effectively classify and address customer complaints.
Social media over the past several years has grown to various functions over and above the realm of digital networking. Going beyond its primary functions of connecting friends and family, allowing people to share pictures of one another, platforms such as Twitter and Facebook have become core to the functions of many industries. Some of these functions include e-commerce (where social media channels are used in the promotion and sale of goods online), advertising (where social media channels are used to reach new groups of users with the placement of targeted ads), and customer service complaints (where brands maintain social media teams to address grievances and complaints targeted against them). This paper explores the function of customer complaints in social media, specifically Twitter, with regards to how Airline brands tackle them.
Twitter has become one of the first places people go to address customer service concerns. Some of these concerns may be positive, such as saying a Thank you, while other times it's about having questions answered and concerns addressed. Taking to Twitter to address such issues is often much more easier than having to go on long customer service calls. Thereby, Twitter has changed the dynamics of customer service by enabling brands and customers to directly connect with each other on a public platform. A survey conducted by Twitter UK indicates that 61% of residents use Twitter as a means to address complaints due to its public nature [4]. This is a prime example of how Twitter has transformed customer service from a 1-to=1 to a 1-to-many interaction.
Research has identified that retail and travel are the two major industries that customers often take to Twitter for. 40% of those who had recently used the platform for customer service had done so for retail and 33% for travel [4]. Customer service interactions via Twitter present an opportunity for the brand to strengthen its relationship with the customer. In addition to their problem being solved, speed, friendliness, and personalization are three other qualities that customers look for in their interaction. Companies have been quite successful in this front as well; over 83% of users from the same research indicated that they were successful in getting their problem resolved by using the platform.
The above findings clearly indicate that addressing issues on twitter has a large impact on the brand's image and even their bottom lines. It helps build customer trust and loyalty. One study found that 96% of users who turned to Twitter for customer service and had a friendly experience with a brand would buy from that brand again [4]. Likewise, an impressive 83% of customers would recommend that brand to others with such a positive experience. A recent study [5] conducted in the US shows how these statistics specifically ring true for the airline industry. The study found that customers were willing to pay nearly $20 more to travel with an airline that had responded to their Tweet in under six minutes. When they got a response over 67 minutes after their Tweet, they would only pay $2 more to fly with that airline.
Hence being cognizant of the above and establishing processes to tackle customer service issues as they arise is key to airline brands. The first step in such a process is to quickly classify the tweets by types so that an airline can quickly delegate the service responsibility to the team in charge. This first step is the primary focus of this paper, wherein we attempt to classify negative tweets by the type of service issue. In section 2, we look at some of the related work that has been conducted in this space. Afterwards, in section 3, we look at the data and the process that was used to collect the data. We outline a procedure for how to tackle this specific Machine Learning problem in section 4 followed by the description and results of the process itself.
One of the papers that was used as reference for the research conducted was Uncovering customer service experiences with Twitter: the case of airline industry [1]. **In this paper, the authors parsed 67,953 public tweets targeted at four airline companies to identify the general customer sentiment about their services. They used a lexicon approach and a vector space model for assessing the polarity of the posts.
With the above approach, the authors were able to identify the major reasons that cause both customer service delight as well as dissatisfaction. Positive sentiments were largely linked to online and mobile check-in services, favorable pricing, and in-flight experiences that were beneficial to the customer. Some of these sentiments also arose out of delightful experiences in airport lounges. Negative sentiments usually revealed problems with the usability of the companies' websites, flight elays, and lost luggage. The paper was quite successful in demonstrating how a sentiment analysis on Twitter data can be used to research on customer service experiences and an alternative to manual classification via a Kano or SERVQUAL model.
Another paper that was closely referenced was Sentiment Classification System of Twitter Data for US Airline Service Analysis [2]. The paper uses the same dataset obtained from Kaggle [3] and performs a Multi-class sentiment analysis. The paper attempts to create a label function for any given tweet by conducting a phrase-level analysis using Doc2vec. The analysis was carried out using 7 different classification strategies: Decision Tree, Random Forest, SVM, K-Nearest Neighbors, Logistic Regression, Gaussian Naïve Bayes and AdaBoost. The output of the test set was a set of tweet sentiments (positive, negative or neutral). These results were compared, benchmarked and visualized against each classification approach.
Both papers offer a lot to learning, opportunities for enhancement, and further research. The first paper cited above introduces some of the major clusters that could be expected to aggregate for reasons for negative sentiments. Some of these could possibly be validated with a Principal Component Analysis (PCA) that may uncover insights into the formations of similar clusters. The paper also introduces more room for thought into the study of airline experiences via Twitter as some positive or negative tweets can be related to a customer's experience at an airport lounge which may not directly be under the purview of the airline company. The second paper cited, offers various techniques that can be used in pre-processing the data. The paper recommends techniques such as removing stop-words and lemmatization (the process where a word is reduced to its base form with the use of vocabulary). The paper also recommends restricting the feature space by removing symbols and punctuations as well as non-english words.
The data was collected in February of 2015 where Twitter data from 6 major airlines were scraped: United Airlines, US Airways, American Airlines, Delta, Southwest and Virgin America. The data was labeled by Crowdflower workers and released under Crowdflower's Data For Everyone program. Contributors were asked to first classify positive, negative, and neutral tweets, followed by categorizing negative reasons (such as "late flight" or "rude service").