But What If We’re Not Good At Math?

Recently, a colleague of mine (Nathan Gilliatt) posted a great article about a crucial and usually absent skill in social media analysis: boolean logic expertise (Can Analytics be Taught?)  The crux of his article - at least from my standpoint - is that the quality of an analytics effort relies upon the ability of the analyst to define his search parameters, and those parameters are usually defined using some interface to boolean logic. Boolean logic is a systematic way of defining complex, “and:or” descriptions, but it is generally only taught as part of computer science, mathematics, and engineering disciplines.

I have used, demoed and toured most of the popular social media analytics platforms out there, and pretty much every system offers some interface, “wizard”, or helpful manual staff-time solution to help would-be analysts overcome their inability to define their search and monitoring profiles in a useful albeit a priori way.  The different platforms are definitely not equal in how well they accomplish this task.  Poorly defined profiles end up measuring and generating great visualizations for, to a greater or lesser degree, irrelevant content.

Hey, We’ve Already Solved This Problem!

However, it has struck me that setting up analytics profiles like these are really problems of classification: are the social media items being examined in Category A (relevant to my brand) or Category B (not relevant to my brand)? It seems an obvious solution that most future analytics platforms will improve the quality of their search profiles through machine learning tools like Naive Bayesian classifiers. The process might look like this:

1. Define initial search profile using boolean logic or a logic wizard.
2. The analyst vets initial results, confirming or denying results as being correctly relevant to the brand in question.
3. Once a sufficient training set has been built (1000+ posts?), the system switches to “full automatic” classification
4. (optional) the classifier can continue to be trained over time, or sensitivity tests can be run to understand which terms are most significant.

A well-trained Bayesian (or other) classifier can reach incredible accuracy rates, 99% and higher.  But here’s the kicker: successfully defining a search profile this way will itself produce some deep insight about your brand’s reputation and about the things people associate with your brand! In other words, the classifier works well because it has come up with an accurate “definition” of a brand based upon the keywords that it has trained itself upon.

Coming Soon To A Theater Near You

I’m certain there’s a few analytics systems providers out there that already have a head start on this; I recall that at least Biz360’s dashboard implemented  machine learning tactics to improve analysis of industry-specific lexicons and jargon.  Given the continuation of Moore’s law and its extension to general technology growth, we can expect that features (like Biz360’s) of high-end systems will become increasingly available in middle-market and entry-level analytics solutions.

| Permalink | Trackback