Some time ago we got an offer from Shlomi Hassan to join Kenshoo hackathon (http://kenshoo.com/data-hackathon/). The idea of hackathon was to try to perform a kind of user value prediction based on history of events.

It has to be said that predictive models are not exactly our specialisation – we are working on anomaly detection in big data sets. So why did we participate?
First of all, working with data which values up to billions of dollars was too tempting. There is famous quote from one old Soviet time’s movie: “it happens once in a lifetime, but not to everyone”.
Secondly, we were interested in trying out our system, in visualizing anomalies in the data, in a bid to improve our predictions.
So, we brought the best members of our team to compete. Note that the organisation was perfect, help received from AWS team was timely and handy, and the overall atmosphere was nice. We worked, applied tool to the data, etc.
We learned a lot of things during this competition. We tried to apply very complex model of prediction. We implemented it, but it was too far from the right answer. We did something oversimplified, and it also fall shot.
The task was so interesting that, even after exhausting 2.5 days of hackaton, we continued our work and finally produced a quite good result.
It was also was very fascinating to observe such great people trying other approaches.
After the competition, I got to the following conclusions:
– Data scientists are very, very under-tooled for big data. Most of the teams had very hard times to do anything with big data sets. We were in a better position due to the fact that it is our main business…
– Complex models, without long tuning, regularization, data cleansing etc can be 10 times worse than simple ones. So we have to select models complexity by time we can realistically invest into tuning.

It has to be said that predictive models are not exactly our specialisation – we are working on anomaly detection in big data sets. So why did we participate?
First of all, working with data which values up to billions of dollars was too tempting. There is famous quote from one old Soviet time’s movie: “it happens once in a lifetime, but not to everyone”.
Secondly, we were interested in trying out our system, in visualizing anomalies in the data, in a bid to improve our predictions.
So, we brought the best members of our team to compete. Note that the organisation was perfect, help received from AWS team was timely and handy, and the overall atmosphere was nice. We worked, applied tool to the data, etc.
We learned a lot of things during this competition. We tried to apply very complex model of prediction. We implemented it, but it was too far from the right answer. We did something oversimplified, and it also fall shot.
The task was so interesting that, even after exhausting 2.5 days of hackaton, we continued our work and finally produced a quite good result.
It was also was very fascinating to observe such great people trying other approaches.
After the competition, I got to the following conclusions:
– Data scientists are very, very under-tooled for big data. Most of the teams had very hard times to do anything with big data sets. We were in a better position due to the fact that it is our main business…
– Complex models, without long tuning, regularization, data cleansing etc can be 10 times worse than simple ones. So we have to select models complexity by time we can realistically invest into tuning.

I really want to thank Kenshoo for providing us with this great opportunity.