Ultralearning Data Science - Week 1

published on October 27th, 2019

How the First Week Went

I'm writing this post at the tail end of my first "ultralearning" week where I'm trying to learn data science.

Overall, the week went pretty well. I spent about 14 hours on the project - a bit less than what I had intended, but there were some unforseen things which made it impossible to meet my full committment. I've worked pretty concentratedly, being aware of the limitied time and setting timer has worked in the first week - so I'll keep doing it.

I've watched three lectures of the fast.ai machine learning introduction course and have practiced the question book method and open recall which worked really well to remember what was in the lectures.

I also submitted to three Kaggle challenges as planed. I really stumbled through the kaggle submissions and as a result, the submissions were really low quality - at least it should not be too hard to improve them. My biggest challenge is that I don't know at all what to do to improve my predictions. I fumbled around with some random forest parameters to improve the results on the validation set, but this was really just try and error. I'm also stumbling more than I thought working with numpy and pandas, getting a routine when working on the kaggle challenges will make this work faster.

I'm happy with my progress, given that I've only worked on this for approximately 14 hours, which is less than two days of full time work. There is of course a lot to be done.

Results

In terms of tangible results, I have the following to show

Submitted to Three Kaggle Challenges

Titanic: categorization accuracy 0.77033, position 7499 / 12930, 58th percentile.
Rossmann: RMSPE 0.20313, position 2783 / 3303, 85th percentile.
House Prices: RMSLE of 0.15678, position 3229 / 4854, 67th percentile

Watched the First Three Lectures

I've watched the lectures and practiced retrieval via the question book method and open recall. This worked well.

Concepts and Anki Flash Cards

I haven't written anything about learned concepts yet and haven't created any Anki flash cards.

Log

Here are some very brief notes I took on the days of the week:

Monday 2019-10-21

First Kaggle submission
Installed anaconda, got jupyter to run

Tuesday 2019-10-22

Went through the jupyter notebook of lecture one
Open recall on the notes
Tried to apply the learnings from the lecture to the Titanic challenge
- Hard time to figure out why I was getting continuous values, needed to use RandomForestClassifier instead of RandomForestRegressor
- then had a hard time why my random forest classifier was working very badly - interestingly when splitting the training data into training and validation it worked much better.

Wednesday 2019-10-23

Figured out that the prediction on titanic was bad because it used the passenger id as one of the independent variables, after dropping it
watched the last 20 minutes of lecture 1
started watching lecture 2
pretty busy day, maria throwing up in the evening, so didn't get to much (1hr overall)
joined a bunch of Kaggle competitions, among them home prices and rossmann sales predictions

Thursday 2019-10-24

At home in the morning, so couldn't do anything there
finished watching lecture 2

Friday 2019-10-25

Open recall on lecture 2
Hit a problem running the notebooks on my machine and lost some time on this
did some reading on xgboost and the rossmann kaggle challenge

Saturday 2019-10-26

Briefly worked and submitted the first entry on the Rossmann challenge

Sunday 2019-10-27

Watched lecture 3
Worked on and submitted the House Prices: Advanced Regression Techniques
Wrote achievements and planning blog post (this thing ;))

My Plan for Next Week

Here are my goals for the comming second week:

Improve my results in the three challenges
Participate in three new Kaggle challenges
Watch lectures 4 and 5 of the fast ai ML lectures and practice retrieval
Create Anki flash cards on concepts I've learned.

To a successful second week of ultralearning!