Wrapping up Kaggle’s Home credit competition
Despite being busy with my schedule recently, I managed to obtain a silver medal (top 5% placement) in this competition. I did struggle because i was doing it alone, which meant that I had to juggle the entire competition pipeline with other commitments / work. But I am glad the effort did pay off eventually. :)
However, I felt that I could do better. I will elaborate more when I explain some of the methodologies that I have adopted in this competition in another post. At the same time, I made a costly mistake: not selecting my best submission. Luckily, this did not cost me heavily.
I am definitely more adventurous in this competition, utilizing more tools (LibFM as well as Keras). But due to time/ resource constraints, I am not able to do everything I had planned out.
On a side note, the cost of this silver medal is pretty hefty. Looking at how much AWS cost I racked up…
Most of the time, I am using t2.2xlarge since it is the most cost efficient instance in my opinion. ( If you guys have better / cheaper options, do share with me! ) If I am getting memory error or I need some of the big guns (i.e GPU), I will either use c5.4xlarge (computation), r4.2xlarge (memory) or the g3.4xlarge (cheap gpu).
In conclusion, this is definitely a competition that is very intense and fun. I definitely learned more regarding credit scoring. On hindsight, I definitely should use all the variables in the data, which should give a stronger model. I will be posting the unique strategies that I had utilized in another post once I have cleared and clean everything up.
