To personae, or not to personae

To ride the data wave everyone was intrigued and excited by the idea of customer centric product design and the 1st step to achieve that was to know who our customers are.

To give you a little background, LexisNexis is a corporation providing computer-assisted legal research where we have a portfolio of products to address specific needs of users associated with the legal system.

It is a very niche set of customers and as our product is subscription based, the 1st problem was to define what success means. Unlike retail or e-commerce we couldn’t assess customer satisfaction by number of purchases/revenue generated.

So our 1st hypothesis was segment/role doesn’t matter when users are using our product. They have a “style of working” and the segment/role may not have anything to do with it. We wanted to break out of the traditional representation of our users based on segments or roles and define groups based on their interaction with our products. 

1st analysis to achieve that was to take the user behavioral data we collect and see if we can find any pattern. We used data points like count of search, no. of document views, no. of document delivered etc. aggregated to user level over a period of time and took a random sample. We were hit hard when 90% of the user base went into a single cluster.

This particular problem was handled by transforming the sparse and skewed variables and using advanced clustering techniques. 

We also trained the model with multiple combinations of independent variables, different segment of users and different timeframe where it became evident that the initial hypothesis we made regarding user behavioral pattern is not true at all. It looked like users’ preference change based on the task in their hand rather than their personal preference while using our product!

We were already aware from separate analyses that only few in the plethora of features that we offer are actually being used by our userbase, so when we focused on one such widely used feature “search” we were able to get better clusters of users based on their preference of search language.

But after going through these multiple iterations we started to think if these traditional clustering methods are suitable for our niche customer base. As I have mentioned before our products are not like any e-commerce or retail business where the path to success is usually one. Now that we have established our users don’t display very distinctive straightforward patterns, we are trying to find alternatives which will help us understand our users better. 

The next steps in that journey is trying to humanize a single user properly than to group multiple user types based on some characteristics. 

Stay tuned for more details about the next part of my journey.

Looking forward to hearing your thoughts/comments/experience in similar use cases.