Predictive Targeting Tool

Model Accuracy Results
Model Accuracy (AUC = Area under curve)

At Dell I was assigned to help create a lead profile for a particular product they were selling. At the time sales was picking customers using some basic heuristics. Once I got a hold of the customer database, I realized that with some data analysis, I could determine which customer attributes (that we collected) were more likely to buy. Ultimately there were too many factors and too much data for me to figure out on my own, but I realized that this was a good use case for machine learning. Using Python I did a factor analysis, and identified the key variables. I then created a predictive model from the key variables and hosted it on Microsoft Azure. This way anyone can upload a (properly formatted) data set and get the results of the model back. The results would be appended to the end with a percentage likelihood to buy chance.

Evaluating the effectiveness of this model is complicated. Essentially I would split the data, one section for training, and one section for evaluating. The model would learn the attributes that made likely to buy on the training section. Then I would run it on the evaluating section and it would give predictions. I would compare those predictions against the actual results. It turned out to correctly predict who would buy 91% or 99% of the time. What is complicated about this number however, is that most people don't buy anything, so even if the model predicted "Will not buy" for every single person it would still be very accurate, just not useful. This is why the population model despite being 99% accurate was less useful than the sample model, because the sample model found far more customers. Ultimately the model identified around $130 Million in untapped customers that the model predicts Dell would win.

Other projects: 

Glad you reached the end of the page!