Machine Learning Study Resource Notepad

ML study memo of de amateur. Irregular updates.

Python resource for ML

• Awesome Pysci by bafoo
• https://github.com/bakfoo/awesome-pysci
• A machine learning version of that awesome python. With a polite Japanese explanation.

webinar

• Machine Learning by Stanford Univ.
• https://www.coursera.org/learn/machine-learning/
• Famous guy. Free, easy to understand, with Japanese (subtitles).
• Basically just talking about math. It is more effective to touch numpy / scikit-learn / tensorflow etc. after seeing this and understanding the mathematical part.
• take time. It takes 1-2 hours every day. You have to do the test seriously.
• [Impressions of those who have completed] Coursera's machine learning course has been completed

Book

• Python Machine Learning Programming Theory and Practice by Expert Data Scientists
• https://www.amazon.co.jp/dp/4844380605

data set

• UCI Machine Learning Repository
• http://archive.ics.uci.edu/ml/
• Various including Iris
• Kaggle Datasets
• https://www.kaggle.com/datasets
• It looks fun with an interesting dataset like the European Soccer Database (https://www.kaggle.com/hugomathien/soccer)

Web article

• Machine Learning in a Week
• https://medium.com/learning-new-stuff/machine-learning-in-a-week-a0da25d59850#.y88amjan7

# 1: Linear regression

• scikit-learn video #6: Linear regression (plus pandas & seaborn)

• http://blog.kaggle.com/2015/05/28/scikit-learn-video-6-linear-regression-plus-pandas-seaborn/

• You should subscribe to kaggle's blog. (Http://blog.kaggle.com/)

• A friendly introduction to linear regression (using python)

• http://www.dataschool.io/linear-regression-in-python/

• Practice: Bike Sharing Demand (kaggle)

• https://www.kaggle.com/c/bike-sharing-demand

• Since the deadline has been closed, I will not get a formal rank, but it will be evaluated

• It seems easy but not so

• If you write the code easily and submit it, it will be about 3000/3200 (total number of participants during the deadline)

• After grouping in season, holiday, workingday, and weather, each is trained, and test data is predicted using the learning results of the same group.

• Use temp, atemp, windspeed, humidity to learn

• Commentary: https://www.analyticsvidhya.com/blog/2015/06/solution-kaggle-competition-bike-sharing-demand/

• Note

• How to choose the learning rate: (too small: slow convergence) ..., 0.001, 0.003, 0.01, 0.03, 0.1, 0.3, 1, ... (too large: cost function does not become small or does not converge) → 3 times each

• Feature Scailing: normalized_value = (original_value --average) / (max --min)

application

Online game cheat measures

• Probabilistic Approaches to Cheating Detection in Online Games
• http://www.hamilton.ie/dwmalone/CIG2010.pdf