Watson – finally!

I was at an IBM conference on Watson Analytics yesterday. This is, I think, the fourth briefing I have attended on Watson Analytics and I have been at rather more on Watson itself. At the reception afterwards one of the other analysts, who has probably been to at least as many of these things as I have, commented that they still didn’t understand the difference between Watson and Watson Analytics. I can’t say I blame them: IBM has been appalling at explaining the difference.

However, there was a comment made yesterday – almost an aside – which means that I think I do, finally, understand the difference.

The first point to understand is that some of the Watson technology is built into Watson Analytics. So they both understand structured and unstructured data and they both support natural language processing. Thus both products respond to natural language queries and, in the case of Watson Analytics, the software will generate natural language queries whose results you might be interested in. That is, it will recognise correlations and other relationships within the data and then propose likely analyses that you might want to examine. Machine learning capabilities mean that it will get better at generating these queries over time, as it learns the sorts of things that you are interested in.

The difference, however, is in the “analytics”. IBM obviously doesn’t feel the need to define what it means by this. What it actually means is that when you are doing “analytics” the results you can expect will typically be represented by a chart or a graph – a histogram, a bar chart, a bubble chart or some other form of visualisation that represents a collection of quantitative results.

Watson, on the other hand, derives answers to questions for which there is a single answer. Actually, it will typically generate multiple answers, giving you a confidence level for each answer and, if you wish, it will provide you with the reasoning behind these answers. But the point is that each suggested answer is a single thing. It might be numeric and it might not, but the answer won’t be quantitative in the sense of required representation in a bar chart. To take a simple example, there is a thing called Watson Chef – it’s a toy really but fun nevertheless – you tell it what you have in your fridge and it will generate recipes for you. Watson Oncology – a much more serious offering – is designed to help doctors with treatment choices. In either case, the software generates likely choices from which the practitioner makes a selection.

A good example to explain the difference would be a question such as “does the weather affect the types of drinks I sell” and the answer would be “yes” if you asked Watson, but that isn’t very useful when you really want to know what drinks are affected and by how much. Ask Watson Analytics and you would get a bunch of figures that related sales of different types of drinks depending on whether it was hot or cold, rainy or sunny, and so on. Thus you might deduce that you sell more hot chocolate (and how much more) when it is cold.

So, if you want to ask traditional analytic type questions then you need Watson Analytics. If you want to ask questions that have specific answers use Watson.