To state the obvious, online reviews are important to consumers. The star ratings, as well as the individual opinions, influence the decisions consumers make about brands. For the brands, however, the real question are: how do we leverage the information inside of reviews? With so many open-ended answers, how do we decide which data points to ignore? And what insights can we gain by aggregating opinions?
This is the question Kimberly-Clark hopes to help solve with IBM Bluemix. IBM Bluemix can best be viewed as a toolbox, with one whole drawer occupied by Watson and Watson-derived tools. During the annual Hackathon on February 23rd, Kimberly-Clark set up a group of Nielsen and Supply Chain students with their Amazon data for several products – primarily from Huggies and Cottonelle – to gather any new observations we could using the Bluemix Watson Discovery tool.
The Hackathon gave fascinating insight into three major areas: the importance of training your robots carefully, how to confuse Supply Chain students (and, conversely, how to utilize their knowledge), and the power of harnessing consumer habit.
One of Watson Discovery’s main skills is its ability to analyze sentiment, bucketing reviews into “positive,” “negative,” or “neutral”, allowing for observation of general sentiment trends over time. This sentiment analysis can be carefully refined and taught, but the student version we were given approached everything using the default settings, leading it to think that reviews that said “expensive” were all negative sentiment (even when they were praising Kimberly-Clark brands for being less expensive than others), as well as it being very wary of any casual language used around fecal matter. Teaching Watson to understand what is and is not important to your brand is vital in getting a clear read of what’s happening – especially when studying diapers and toilet paper.
The Nielsen students were mostly paired with students from the Grainger Center for Supply Chain Management during the Hackathon, which lead to interesting conversations. Some of these were simply amusing, like the confusion expressed by Supply Chain students for how long people could discuss toilet products (or that people would put Huggies on their dogs), but some proved to be a microcosm of what is likely a larger issue in corporations: the trade-offs between what consumers want, and what it is feasible to give them. For example, toilet paper rolls that don’t have a central cardboard tube appear to have a somewhat higher tendency to become misshapen during shipping than standard toilet paper, which led to many of the complaints for that product. Supply Chain students pointed out that to reduce the squished effect a non-minor amount of extra packaging would have to be included, the extra costs likely negating any benefit.
The most interesting insight to me, however, came primarily from observing pure sales data and double-checking it with reviews. There was a slow rise in Cottonelle toilet paper over the period, a dramatic one-week spike in 2015, and a faster rise thereafter (weekly sales before the spike were roughly 60% that of weekly sales after the spike). I used the reviews to find that the spike occurred during Amazon Prime Day, and that consumers hadn’t realized Cottonelle was available on Amazon prior to that promotion – and afterward defaulted to using Cottonelle on Amazon as the low-energy shopping option. This ability to make purchases habitual strongly decreases the likelihood of switching behavior “at shelf,” and shows the power of harnessing Amazon (and similar marketplaces) correctly.
As Bob Loos, Global Director of Omni-Channel Analytics at Kimberly-Clarke said several times, “effortless style takes a lot of work”. It looks like someday soon, machine learning will be fluent enough in English to help.