Cut Through the Noise: How Discriminant Analysis Empowers Product Managers to Make Data-Driven Decisions

When dealing with large data sets that have multiple variables, the sheer amounts of data can obscure decision-making. As a product manager your primary objective is to ensure the best customer experience while maintaining your competitive edge.  Typically, to ensure that the feature under consideration meets user needs, we use reviews, feedback, surveys, ticket analysis etc. we may spend hours trying to discern noise from insight. the sheer number of disparate sources of data creates a paradox. We have a flood of metrics, KPIs at our analytical disposal and yet identifying which data point can point us to the needle mover can be a struggle. In these cases, i have always relied on Discriminant Analysis as a means to separate the wheat from the chaff. Discriminant analysis is often ignored by product teams who consider that the methodology is only applicable for ML intensive use cases.

On a project that I worked on with an e-commerce company, I grappled with why a specific category of items had lower than expected conversion rate when items were added to cart after a session, while the same user experience generated a higher conversion rate on other categories. A high-level glance at the data appeared to not show any distinction except conversion rate between the categories. I decided to use discriminant analysis to reveal hidden patterns in customer segments in my dataset.

To begin with, I separated the data into meaningful classes. Despite this separation, my dataset had vast amounts of data . I began filtering and interpreting the data in context to the conversion rate. I filtered by sub category, price thresholds, customer type (broadly based on loyalty status) and user feedback/rating. I further separated the data into 4 quadrants. based on their relevance to KPI and accuracy (clean and complete data). In other words, I selected highly relevant data and highly accurate as quadrant one followed by highly relevant but low accuracy in quadrant two and so on. to avoid analysis paralysis , i eliminated data from sub-categories that had a high return rate as these items could potentially have reasons for low conversion beyond the scope of the program. Within the quadrants, I also eliminated data that was low in accuracy and relevancy (fourth quadrant). I now focused my efforts on quadrants 1,2 and 3. The data now revealed hidden customer segments of purchasing patterns associated to price per unit on specific categories and specific customer demographic. Armed with these discriminant variables, we then enabled an A/B testing, that targeted categories with low conversion rate for customers who met the purchasing pattern criterion. Our Test was successful, and we saw a +28% uptick in conversion rates for our treatment against the control segments.

Now is discriminant analysis applicable everywhere ? No! When you have complex and non- linear datasets intertwined with customer psychology [for example: browsing patterns] discriminate analysis often produces poor classification and obscures insight generation. In these cases, it may be valuable to first distill the data using funnel analysis techniques and reduce the heterogeneity in variances before applying any discriminate analysis methodologies.

When used wisely, Discriminate Analysis empowers product managers to uncover unknown customer insights and fuel data driven decisions that optimize product strategy and drive growth.

Good luck turning your insights into your next big win !!