Dmitry Gordeev

Data Scientist, Kaggle Grandmaster

[Dmitry Gordeev @dott1718]

@StatsbyLopez @dj_mosfett @NextGenStats @ph_singer We haven't looked at the maps, we rather focused on what variables and network connections make most impact on the accuracy. Judging by that, what matters most is relative location and velocity of each defender vs offensive players and vice versa.

[Dmitry Gordeev @dott1718]

@a_erdem4 @ph_singer Even I used to like qop. 4 years since I played it last time...

[Dmitry Gordeev @dott1718]

Together with @ph_singer @StackNet_ @Kagglenizer and @srisatish we have published an article on the importance of backtesting the covid-19 forecasting models https://t.co/Xz558UUKTP and open sourced the code https://t.co/imOmx1kNHR

[Philipp Singer @ph_singer]

Proud that NFL implemented our (\w @dott1718) #kaggle winning model and that they will use it throughout the 2020 season. I hope the success of this competition can be an inspiration to many others to follow. Details: https://t.co/IYT89EVYAD (click also the second tab)

[Dmitry Gordeev @dott1718]

@WalterReade I am so looking forward to Kaggle days getting back on track though. Tokyo trip was great...

[T. Ott @thomasottio]

The Zoo, aka Philipp Singer and Dmitry Gordeev, are two Senior Data Scientists here at @h2oai who are well-known for winning many Kaggle competitions. On May 26th at 11AM CET there will be a live Q&A with them. Sign up below and ask away! https://t.co/A3R6ruYQqD

[Dmitry Gordeev @dott1718]

@StatsInTheWild @CaioBrighenti @devinpleuler @ph_singer I think Z was around 256... x3 stands for 3 consecutive convolution layers (with ReLU activation after each).

[Dmitry Gordeev @dott1718]

@StatsInTheWild @CaioBrighenti @devinpleuler @ph_singer You are correct, but the output is Zx10x11 where Z is number of CNN units to be chosen when you build the model. Then you pool across the 11 players, making it Zx10.

[Dmitry Gordeev @dott1718]

@StatsInTheWild @CaioBrighenti @devinpleuler @ph_singer I am not really following a strict notation either :)

[Dmitry Gordeev @dott1718]

@StatsInTheWild @CaioBrighenti @devinpleuler @ph_singer You might have gotten it right.. but hard to be precise with the word "across" :D
With tensor of shape 5x10x11 when you run a 2d convolution with kernel of 1 it will act as a transformation of the 5 inputs (channels) applied to each of 10x11 vectors independently.

[Dmitry Gordeev @dott1718]

@StatsInTheWild @CaioBrighenti @devinpleuler @ph_singer We applied both 2d and 1d convolutions. First a 2d which runs through all 10x11 pairs and then, after the first pooling, a 1d through 10 defenders. But in both cases convolutions had kernel sizes of 1, meaning they didn't take any "neighbouring" pairs or players into account.

[Dmitry Gordeev @dott1718]

@CaioBrighenti @devinpleuler @StatsInTheWild @ph_singer The result is then "pooled" by taking average (and max), which are also independent of the order. So we explicitly limited ourselves to these operations, which give the same result regardless of the order.

[Dmitry Gordeev @dott1718]

@CaioBrighenti @devinpleuler @StatsInTheWild @ph_singer Yeah, in a nutshell it's about a convolution with kernel size 1, which a transformation applied to each individual player, so the order of the players in the tensor doesn't matter.

[Dmitry Gordeev @dott1718]

As @ph_singer has joined the team as well, we can announce it as a career move for the whole Zoo ;)

[Dmitry Gordeev @dott1718]

I am happy to announce that I joined the Grandmasters team of https://t.co/522bNAfja5! It was an exciting first day for me today, thanks to everyone for such a warm welcome!

[Dmitry Gordeev @dott1718]

@VRaoRao That's the plan for me, hope to say exactly the same thing after a while! The decision to make such a career move was easy for me, but the move itself took years...

[Dmitry Gordeev @dott1718]

I have been working for banks, insurance and consulting for more than 10 years now. Today was my last day employed in finance.. It's time to explore new areas, I hope now it will be as fun as I imagine it.

[Dmitry Gordeev @dott1718]

@mhviraf No, I am not. But will announce my new job tomorrow!

[Dmitry Gordeev @dott1718]

@olgaiv39 @kaggle @ph_singer Teaming up helps. And last week there was more spare time with the quarantine.. But you are right, it takes a lot of efforts

[Dmitry Gordeev @dott1718]

Last night The Zoo finished 2nd in Bengali @kaggle competition! It was tough, but a fun learning experience in computer vision for both me and @ph_singer. Great work from participants and great solutions!