Machine Learning/Courera's Competitive Data Science Course: Difference between revisions

From Federal Burro of Information
Jump to navigationJump to search
 
Line 39: Line 39:


for example hashed out data
for example hashed out data
=== Build a quick baseline ===
ref: https://hub.coursera-notebooks.org/user/hnlrfyqblxjsscsosbsvlx/notebooks/readonly/reading_materials/EDA_video3_screencast.ipynb
<pre>
from sklearn.ensemble import RandomForestClassifier
# Create a copy to work with
X = train.copy()
# Save and drop labels
y = train.y
X = X.drop('y', axis=1)
# fill NANs
X = X.fillna(-999)
# Label encoder
for c in train.columns[train.dtypes == 'object']:
    X[c] = X[c].factorize()[0]
   
rf = RandomForestClassifier()
rf.fit(X,y)
RandomForestClassifier(bootstrap=True, class_weight=None, criterion='gini',
            max_depth=None, max_features='auto', max_leaf_nodes=None,
            min_impurity_split=1e-07, min_samples_leaf=1,
            min_samples_split=2, min_weight_fraction_leaf=0.0,
            n_estimators=10, n_jobs=1, oob_score=False, random_state=None,
            verbose=0, warm_start=False)
plt.plot(rf.feature_importances_)
plt.xticks(np.arange(X.shape[1]), X.columns.tolist(), rotation=90);
</pre>

Latest revision as of 17:40, 14 May 2018

Exam notes

I

Suppose that you have a credit scoring task, where you have to create a ML model that approximates expert evaluation of an individual's creditworthiness. Which of the following can potentially be a data leakage? Select all that apply.

1. First half of the data points in the train set has a score of 0, while the second half has scores > 0. - This should be selected
Is a leak -
2. Among the features you have a company_id, an identifier of a company where this person works. It turns out that this feature is very important and adding it to the model significantly improves your score. = Un-selected is correct
not a source of leak
3. An ID of a data point (row) in the train set correlates with target variable.
Is a leak
Explainiation: Data was not shuffled, this information can not be used in real-world scenario

II

What is the most foolproof way to set up a time series competition?

Split train, public and private parts of data by time. Remove all features except IDs (e.g. timestamp) from test set so that participants will generate all the features based on past and join them themselves.
Correct Answer
Make a time based split for train/test and a random split for public/private.
This should not be selected - WRONG!
Explanation: Vulnerable to leaderboard probing.
Split train, public and private parts of data by time. Remove time variable from test set, keep the features.
Also Ronge - dunno why


Exploratory Data Analysis (EDA)

notepad https://hub.coursera-notebooks.org/user/hnlrfyqblxjsscsosbsvlx/notebooks/readonly/reading_materials/EDA_video2.ipynb

also see Big_Data#Exploratory_Data_Analysis_.28EDA.29

Handling analysing annoymoused data

for example hashed out data

Build a quick baseline

ref: https://hub.coursera-notebooks.org/user/hnlrfyqblxjsscsosbsvlx/notebooks/readonly/reading_materials/EDA_video3_screencast.ipynb

from sklearn.ensemble import RandomForestClassifier
​
# Create a copy to work with
X = train.copy()
​
# Save and drop labels
y = train.y
X = X.drop('y', axis=1)
​
# fill NANs 
X = X.fillna(-999)
​
# Label encoder
for c in train.columns[train.dtypes == 'object']:
    X[c] = X[c].factorize()[0]
    
rf = RandomForestClassifier()
rf.fit(X,y)
RandomForestClassifier(bootstrap=True, class_weight=None, criterion='gini',
            max_depth=None, max_features='auto', max_leaf_nodes=None,
            min_impurity_split=1e-07, min_samples_leaf=1,
            min_samples_split=2, min_weight_fraction_leaf=0.0,
            n_estimators=10, n_jobs=1, oob_score=False, random_state=None,
            verbose=0, warm_start=False)

plt.plot(rf.feature_importances_)
plt.xticks(np.arange(X.shape[1]), X.columns.tolist(), rotation=90);