Sanjeevni Wanchoo
Feb 19th, 2016

Sun Ray in City web

Many companies are sitting on an asset and aren’t realizing its value.

This asset is the data you have from customer call transcripts and complaints.  A lot of such call transcripts and customer complaint databases have existed for a long time without drawing much interest as a source of broader insights. In most cases, the data has been passively captured over the years – often without much structure or thought. Now, equipped with better analytical techniques for handling unstructured data, several companies have started looking at this data as an asset, and mining it for insights that drive their key initiatives. Three ways in which this data is being used is discussed below:


1.  Topic-modeling to zoom in on issues: Most call center transcripts or complaints databases have a taxonomy of their own when it comes to reporting issues. For example, when calling for help regarding an Amazon product, you may be asked to select from a list of categories you may need help with. If you choose to return an item, you are further asked for a specific reason for doing so. While these broader categories may capture a lot of the common issues that customers face, they may miss the point.

In recent years, an automotive company had to recall vehicles in response to a safety glitch that resulted in several accidents. In retrospect, they found that several complaints had been made over the course of the previous months that pointed to this problem well in advance. However, these went largely unnoticed in the plethora of complaints. An analysis was later conducted to see if topic modeling using Latent Dirichlet Allocation (LDA), or other techniques could be used to diagnose these new complaint topics. It was found that these issues were, in fact, completely predictable, and could have saved several lives (not to mention, prevented monetary losses). Now, the company uses these techniques on the complaints database to understand what kind of keywords and topics seem to show up increasingly.

In the past, most of the analysis of complaints has focused on the bigger picture (“Increase in the number of injuries from accidents this month versus last”), rather than pinpointing specific issues (“The airbags failed to inflate/are defective”). The former comes from issues observed in the past, but may not capture those that have never happened. The latter takes a different approach. Using text mining techniques, this data can, instead, be used to proactively predict any red flags.


2.  Continuous Monitoring: The above analysis can be extended and integrated into a continuous monitoring system which tracks and watches out for unusual increases in the frequency of certain words/topics, or looks out for other patterns. This can not only be used proactively to watch out for any new issues, but also to help prioritize areas of the company’s customer service that need improvement.


3.  Feature Selection: One of the challenges of using any supervised learning technique is ensuring that its variables – or features – capture most, if not all, of the variability in data. As predictive analytics is being used in increasingly novel ways to answer questions that have never been answered through data before, it becomes increasingly challenging to know what the sources of variability might be.

For example, if a utility company is trying to understand what kind of customers are choosing paper over paperless billing, it is important for the data scientist to ensure that the model captures all kinds of information about the customer. This includes looking at their usage data, billing data, demographic information, etc. However, these factors may or may not capture most of the variability in data. In this case, call center transcripts may provide a lead. For example, if many of the calls that result in a switch from paperless to paper bill also discuss login and password retrieval issues, the reason for switching to paper bills may actually be rooted in a negative user experience on the website. The data scientist can then choose to extract and include features that capture the user’s web interaction in the final model.



A side note on sentiment analysis: Sentiment analysis helps capture the degree of positive and negative sentiment in a given text. We must be wary of using some of this data on interaction with customers as a way to gauge the company’s popularity among customers, since complaint databases and even call transcripts are, by their nature, negative. People are unlikely to call a call center to congratulate them on a job well done. For gauging customer sentiment, it is better to analyze news articles or relevant outside blogs for a more complete picture.