Machine Learning-based Lead Generator (or Recommender System)

Description:

At geoimpact, I developed a lead generator that suggests promising buildings for the sale of a renewable energy product based on a list of past sales of the product.

Methods:

• Thompson sampling: This is a method for addressing the exploration- exploitation tradeoff in contextual bandit problems, where we want to maximise a certain reward (i.e., contacting the most promising building owners) and at the same time to continuously improve our model predicting the reward.

• Extra trees classifier: I used a simple implementation of Thompson sampling by sampling different trees of an extra tree classifier.

• Supervised clustering: The extra tree classifier can also be used to create cluster of buildings that behave “similar” with respect to this sales-problem. These cluster were used for a stratified sampling approach that helps to enhance the diversity in the potential customer exploration. The plot below shows a similarity matrix clustered into ten clusters of “similar” buildings.

Specials:

We tested this method with a company selling photovoltaic systems. The project was stopped after the testing phase, as the cold acquisition process was too tedious. Here, I learned that a method can be theoretically very elaborated, but in the end it still needs to fit well into an end-to-end business workflow in order to be practical.

Technology:

Python, Scikit-Learn