Do Your Customers Love You? Find Out with Text Analytics
04/15/2019 by Reid Baughman Machine Learning, Modernization - Analytics
Lasting relationships are built on good communication. If you didn’t listen to your significant other, how long do you think that relationship would last? The same is true for your customer relationships. Whether you like it or not, your customers are saying things about you online that deserve your attention. In some cases, they’re offering constructive feedback.
Many companies don’t listen – not because they don’t want to, but because they don’t know how to. This is often due to the sheer volume of comments, reviews and tweets out there which overwhelm conventional methods of analysis. Fortunately, SAS® Viya® Text Analytics or SAS Enterprise Text Miner can help you quickly sift through all the data to find actionable insights that can help companies truly listen to their stakeholders.
To illustrate, let’s explore some complaint data available at the CFPB website. In the wake of the 2007 financial crisis, the Consumer Protection Financial Bureau (CFPB) was created to “promote fairness and transparency for mortgages, credit cards, and other consumer financial products and services.” As part of that mission, they established a system for consumers to log complaints with any bank regarding financial products and made the database of complaints available to the public. Turns out, people have A LOT to say about their banks, and in this dataset, no bank is spared. The dataset captures both the complaint itself as well as how the company officially responded to the complaint. Two actual complaints (misspellings and all) are shown below:
This complaint text provides the opportunity to listen to the voice of the customer and understand common reasons for dissatisfaction. Using that information in conjunction with the company’s response to the customer provides insight into which complaint topics are potentially avoidable. The assumption here is that if the company offered relief, then they accept wrongdoing and that the complaint should have been avoidable.
Below are a couple of techniques to help companies better understand their customers and create actionable results. The dataset used analyzes roughly 5,000 customer complaints from one financial institution (though these techniques easily scale to hundreds of thousands of comments).
The topic clustering approach analyzes just the text of the complaint and derives common topics. Topics are defined as words that appear together frequently across all the complaints. Below are the most common topics that appear across the 5,000 complaints referenced above.
One of the most dominant topics you can see in the table is home loan modification and foreclosure. Others include complaints related to credit/debit cards accounts and overdraft fees (one area I’m willing to admit to you that I was keenly aware of during my college years). With these results, we can quickly see the dominant themes and use this information in areas such as product development, marketing, and customer service.
Say we were concerned with our online image and wanted to take action to reduce the number of complaints popping up online. It’s unlikely that we would be able to stop all complaints (some people are just cranky), but what if we could identify avoidable complaints and train our customer service representatives to resolve those types of issues before they escalate? Since the dataset contains information on which complaints received relief from the bank, we could use this information to identify which complaint topics were most likely to receive relief (i.e., were avoidable).
In the previous method, we were just trying to understand our data by creating topic clusters to reveal common themes from customer complaints. Now we’ll attempt to predict customer feedback using our data – in this case, which complaint topics have the highest likelihood of being avoidable. We’ll rely on a rule-based method that creates a variety of AND/AND NOT combinations of individual words from the complaints (rules) and uses them to predict whether that complaint received relief (i.e., was avoidable). The table below shows some of the rules generated from this method.
In this table, we can see which combinations of words appear to be associated with relief being provided or not. The “~” means that a particular term was NOT included in that rule. The precision and recall columns are useful to organizations that are concerned about false positives or false negatives with prediction and may want to minimize one or the other (or both). Below are the top five terms/topics associated with relief:
Most of these “avoidable” complaints appear to be related to fees charged to consumers. Since removing fees would likely cause a non-trivial impact on the bank’s profit, perhaps a better solution might be to train customer service staff to remove fees from customers’ accounts when they call in to complain. A brief investigation into the text of some of these claims shows that many customers contacted the bank regarding the fees prior to filing the complaint with the CFPB so the problem could have been taken care of before the complaint became public!
Instead of treating your customers like your significant other by buying flowers and chocolate, take the time to truly listen to them and show your love by acting on what they tell you.
If you have a lot of catching up to do and feel overwhelmed by the number of comments, reviews and tweets out there, we at Zencos have your back and can apply actionable analytics to your data.