add materials for classes 17 and 18

README.md

@@ -468,8 +468,6 @@ Tuesday | Thursday
 * These examples may help you to better understand the process of feature engineering: predicting the number of [passengers at a train station](https://medium.com/@chris_bour/french-largest-data-science-challenge-ever-organized-shows-the-unreasonable-effectiveness-of-open-8399705a20ef), identifying [fraudulent users of an online store](https://docs.google.com/presentation/d/1UdI5NY-mlHyseiRVbpTLyvbrHxY8RciHp5Vc-ZLrwmU/edit#slide=id.p), identifying [bots in an online auction](https://www.kaggle.com/c/facebook-recruiting-iv-human-or-bot/forums/t/14628/share-your-secret-sauce), predicting who will [subscribe to the next season of an orchestra](http://blog.kaggle.com/2015/01/05/kaggle-inclass-stanfords-getting-a-handel-on-data-science-winners-report/), and evaluating the [quality of e-commerce search engine results](http://blog.kaggle.com/2015/07/22/crowdflower-winners-interview-3rd-place-team-quartet/).
 * [Our perfect submission](https://www.kaggle.com/c/restaurant-revenue-prediction/forums/t/13950/our-perfect-submission) is a fun read about how great performance on the [public leaderboard](https://www.kaggle.com/c/restaurant-revenue-prediction/leaderboard/public) does not guarantee that a model will generalize to new data.
 
-<!--
-
 -----
 
 ### Class 17: Decision Trees
@@ -503,8 +501,11 @@ Tuesday | Thursday
 * [Not Even the People Who Write Algorithms Really Know How They Work](http://www.theatlantic.com/technology/archive/2015/09/not-even-the-people-who-write-algorithms-really-know-how-they-work/406099/) argues that the decreased interpretability of state-of-the-art machine learning models has a negative impact on society.
 * For an intuitive explanation of Random Forests, read Edwin Chen's answer to [How do random forests work in layman's terms?](http://www.quora.com/Random-Forests/How-do-random-forests-work-in-laymans-terms/answer/Edwin-Chen-1)
 * [Large Scale Decision Forests: Lessons Learned](http://blog.siftscience.com/blog/2015/large-scale-decision-forests-lessons-learned) is an excellent post from Sift Science about their custom implementation of Random Forests.
+* [Unboxing the Random Forest Classifier](http://nerds.airbnb.com/unboxing-the-random-forest-classifier/) describes a way to interpret the inner workings of Random Forests beyond just feature importances.
 * [Understanding Random Forests: From Theory to Practice](http://arxiv.org/pdf/1407.7502v3.pdf) is an in-depth academic analysis of Random Forests, including details of its implementation in scikit-learn.
 
+<!--
+
 -----
 
 ### Class 19: Advanced scikit-learn and Clustering

code/17_bikeshare_exercise_nb.py

@@ -0,0 +1,68 @@
+# # Exercise with Capital Bikeshare data
+
+# ## Introduction
+# 
+# - Capital Bikeshare dataset from Kaggle: [data](https://github.com/justmarkham/DAT8/blob/master/data/bikeshare.csv), [data dictionary](https://www.kaggle.com/c/bike-sharing-demand/data)
+# - Each observation represents the bikeshare rentals initiated during a given hour of a given day
+
+import pandas as pd
+import numpy as np
+from sklearn.cross_validation import cross_val_score
+from sklearn.linear_model import LinearRegression
+from sklearn.tree import DecisionTreeRegressor, export_graphviz
+
+
+# read the data and set "datetime" as the index
+url = 'https://raw.githubusercontent.com/justmarkham/DAT8/master/data/bikeshare.csv'
+bikes = pd.read_csv(url, index_col='datetime', parse_dates=True)
+
+
+# "count" is a method, so it's best to rename that column
+bikes.rename(columns={'count':'total'}, inplace=True)
+
+
+# create "hour" as its own feature
+bikes['hour'] = bikes.index.hour
+
+
+bikes.head()
+
+
+bikes.tail()
+
+
+# - **hour** ranges from 0 (midnight) through 23 (11pm)
+# - **workingday** is either 0 (weekend or holiday) or 1 (non-holiday weekday)
+
+# ## Task 1
+# 
+# Run these two `groupby` statements and figure out what they tell you about the data.
+
+bikes.groupby('workingday').total.mean()
+
+
+bikes.groupby('hour').total.mean()
+
+
+# ## Task 2
+# 
+# Run this plotting code, and make sure you understand the output. Then, separate this plot into two separate plots conditioned on "workingday". (In other words, one plot should display the hourly trend for "workingday=0", and the other should display the hourly trend for "workingday=1".)
+
+bikes.groupby('hour').total.mean().plot()
+
+
+# ## Task 3
+# 
+# Fit a linear regression model to the entire dataset, using "total" as the response and "hour" and "workingday" as the only features. Then, print the coefficients and interpret them. What are the limitations of linear regression in this instance?
+
+# ## Task 4
+# 
+# Use 10-fold cross-validation to calculate the RMSE for the linear regression model.
+
+# ## Task 5
+# 
+# Use 10-fold cross-validation to evaluate a decision tree model with those same features (fit to any "max_depth" you choose).
+
+# ## Task 6
+# 
+# Fit a decision tree model to the entire dataset using "max_depth=3", and create a tree diagram using Graphviz. Then, figure out what each leaf represents. What did the decision tree learn that a linear regression model could not learn?