UPY Notes”
Section 1: Course Introduction
Section 2: Environment Set-Up
Section 3: Jupyter Overview
Section 4: Python Crash Course
Section 5: Python for Data Analysis - NumPy
Section 6: Python for Data Analysis - Pandas
Section 7: Python for Data Analysis - Pandas Exercises
Section 8: Python for Data Visualization - Matplotlib
Section 9: Python for Data Visualization - Seaborn
Section 10: Python for Data Visualization - Pandas Built-in Data Visualization
Section 11: Python for Data Visualization - Plotly and Cufflinks
Section 12: Python for Data Visualization - Geographical Plotting
Section 13: Data Capstone Project
Section 14: Introduction to Machine Learning
78. Supervised Learning Overview
- Supervised learning is machine learning with labels
79. Evaluating Performance - Classification Error Metrics
- Accuracy = number correctly predicted. As a measure of fit, is only good when the labeled data is balanced (e.g. half dogs, half cats – NOT people with disease vs people no disease)
- Recall
- Ability to find ALL relevant cases
number of true positives / (number of true positives + number of false negatives)
- Precision
- Ability to identify ONLY relevant data points
number of true positives / (number of true positives + number of false positives)
- F1 is a measure of the harmonic mean of precision and recall, taking both metrics into account
- NOTE: There is no “best measure of fit” – it 100% depends on what you’re trying to test for. (e.g. a false positive is much better than a false negative in cancer tests)
80. Evaluating Performance - Regression Error Metrics
- MAE – mean absolute error. PROBLEM: Does not punish outliers
- MSE – mean squared error. Solves the outlier problem, HOWEVER, the metric isn’t easily translatable (as it is squared)
- RMSE – Root mean squared error. Takes the root of the MSE and solves both problems.
81. Machine Learning with Python
- scikit-learn tries to be the same across all modeling families (see below)
- some methods will be available only to supervised learning algorithms (and others only for unsupervised)
- X and Y are passed to supervised learning algos (because that is LABELED data) and only X is passed to unsupervised learning algos
On ALL ESTIMATORS/ALGOS
model.fit()– fit training data (for supervised, this accepts X and y (model.fit(X,y)) – for unsupervised, just X (model.fit(X)))- SUPERVISED:
model.predict()(this takes the new datamodel.predict(X_new)and returns the learned label for each object in the passed data) - SUPERVISED:
model.predict_proba()for clssification rpoblems, some estimators use this to return the probability the new observation has each categorical label. Label with highest probability is returned by model.predict() - SUPERVISED:
model.score()(scores are betwen 0 and 1, closer to 1 being better fit) - UNSUPERVISED:
model.predict()as well (predict labels in clustering algos) - UNSUPERVISED:
model.transform()accepts X_new - UNSUPERVISED:
model.fit_transform()
Section 15: Linear Regression
Originally “regression” came from a study in the 1800s about a father’s and the height of his sons. It was shown that while a father’s height is important, the more important factor is the overall mean of the population. The son’s height tended to be closer to the overall average height of all the people. Meaning height REGRESSES towards the mean.
Our goal with linear regression, is to draw a line that’s as close as possible to every single data point. If we have only two points, this line will simply hit both of those points. If there are many points, our goal is to find THE BEST line. Well, what’s THE BEST line? There are a few ways to measure this but the overall goal is to minimize the distance between ALL the points and their distance to our line.
- Things tend to regress towards the mean
- Linear regression is finding a line that is as close as possible to every data point
- How do we find “the best” line? We try to minimize the distance between each point and our line
- What is our “distance measure”? – If we are using the Least Squares Method (sum of squares of the residuals)
NOTE:
from sklearn.cross_validation import train_test_split
has been changed to :
from sklearn.model_selection import train_test_split
Linear Regression with Python Pt 1:
Split into X and y, toss out language
df.columns X = df[[columns (without predictor column – PRICE – or address bc it’s NLP)]]
y = column we want
use documentation to tuples for traintest split X_train, X_test, etc. (UPY has test_size=0.4 and random_State=101)
from sklearn.linear_model (use tab) import LinearRegression model
then instantiate it
lm = LinearREgression()
lm. HIT TAB to see all available methods on the model
we want lm.fit() (we only want to train the training data)
PRINT COEFF = lm.coef_
PRINT INTERCEPT: lm.intercept_
make a df of COEF
cdf = pd.DataFrame(lm.coef_, X.columns)
In English – a one unit increase (holding all others constant) results in the COEFF increase in hosue price
QUICK REVIEW:
- Grab data
- Do quick EDA
- Separate our data into X and y (features and what we are trying to predict)
- Import the model (in this case, Linear Regression)
- Fit that model to the training data
Section 16: Cross Validation and Bias-Variance Trade-Off
Section 17: Logistic Regression
Section 18: K Nearest Neighbors
Section 19: Decision Trees and Random Forests
Section 20: Support Vector Machines
Section 21: K Means Clustering
Section 22: Principal Component Analysis
Section 23: Recommender Systems
GOAL – take notes while NOT WATCHING LECTURE
Then, recreate what I learned with my notes
- import pandas and numpy
- create column_names [‘userid’, itemid, rating, timestamp]
- df = pd.read_csv u,data sep=”\t” (tab separated data)
- check head
- MOVIE LENS dataset
- grab the movie titles from a separate csv
- pd.merge(df, movie_titles, on=”item_id”)
- Now explore the data
- Import matplotlib.pyplot as plt
- import seaborn as sns
- sns.set_Style(‘white’)
- matplotlib line
- ratings df
- avg rating number of ratings
- groupby title rating mean()
- sort_values(ascending=False)
- get MOST ratings (count)
- Add number of ratings column
- `sns.jointplot(x=’rating’,y=’num of ratings’, data=ratings, alpha=0.5)
PART TWO!
- Create a matrix that has user ids on one axis and movie titles on another axis (each cell has rating the user gave that movie)
- Use PIVOT TABLE to get this matrix
PART TWO, AGAIN! 02-03-21
- Make a matrix with
df.pivot_table - Pick movies to test our recommender with!
- using our ratings df, sort_values (ascending=False)
- get JUST the ratings for those selected movies (e.g. star wars and liar liar)
- our_matrix[‘Star Wars (1977)’]
- Find movies similar to our test movies with the
corrwithfunction! - Makd a df out of this, drop the na,
- But wait?! These don’t really coorelate with Star Wars!?
- We should drop anything that has fewer than 100 ratings 7.
NOTE: We use JOIN instead of MERGE when we have to dfs with the same index (in this case, title)
HTML
Testing this thing
PYTHON
Testing this thing
Section 24: Natural Language Processing
Section 25: Neural Nets and Deep Learning
STARTED: 4/5/21
127. Perceptron Model
- Modeled after an actual neuron – has dendrites (inputs) that feed into the neuron (the place where the computation happens) and outputs to the axon (the output)
- We can add weights and biases to our dendrite inputs
129. Activation Functions
NOTES:
- Remember, w = weight, x is our variable and b is our bias.
- Think about bias as what w _ x would have to overcome to be “counted” so if the bias = -10, w _ x would have to overcome 10 to be “counted”
What if we want to limit our w _ x + b? ENTER ACTIVATION FUNCTIONS!! We input our w _ x + b into our activation functions to limit what their output can be.