First of all: congrats to the winners and thank you to the organizers!!!
The competition was an exciting experience and gave us the possibility to have a lot of fun. On the way, we have tested some naive models built only on the aggregates, a couple of variants of target encoding connected with linear and boosting models. Also, Bayesian data generation seems to work, but we didn't have enough time to explore this direction fully.
Till the next one :-)
Team Baldur
One of the biggest challenges in the online marketing data is its sparsity and huge dimensionality that we were trying to reduce by looking for a continuous, dense representation. For that purpose, we have exploited aggregated data. We have been experimenting with target and weight of evidence encodings. We have also tried neural embeddings. The finally used encodings were corrected for the noise present in the data.
Next, a bunch of models (logistic regressions, gradient boosting methods, neural networks) on one hot encoded and dense representation encoded non-aggregated data were trained and stacked. Stacking appeared to be a very efficient way to improve final prediction results. It worked the best when stacking poor with superb component models using a simple liner approach.