Welcome to our first Time Series Learning Competition: Earthquakes!

marcmuc · November 21, 2018, 4:16pm

So, here is a baseline for the competition : Accuracy 0.748

I wanted to play around with the tabular model with block api, as I had not done that before. This dataset of course is not ideal for that, but I got it working and get consistent accuracy of 0.748201. This number comes up no matter how deep/wide the architecture and after fiddling with the hps a bit. It always seems to converge on this (I had one freakvalue of 78.4% but that seemed random and not reproducible.

Have a look:

gist.github.com

https://gist.github.com/mprostock/c0f5d2b3151d6c3a3984e73c4c048301

earthquakes_tabular.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [

This file has been truncated. show original

BTW: I have used the TEST file as validation set, because the number of examples in total is very limited. So training a naked neural net with this is actually not ideal. So I am looking forward to seeing better results with CNN approaches! Let’s go! Beat me!