Coding.
Remember the Bayes Dice app we build many weeks ago? Let's revisit that app, but with a twist.
You have 3 coins with the following probabilities. P(H|1) = 0.3, P(H|2) = 0.45, P(H|3) = 0.75.
That is read as the probability of heads for coin 1 is 30%, etc.
Write a small app, using Object Oriented Python, that allows you to randomly select a coin (without looking) and then repeatedly flip it about 10 times or so until you are fairly certain as to the type of coin you selected.
Questions.
In general, what makes the Naive Bayes Classifier so naive?
- It is naive because it assumes all features are independent.
What is the difference between the Bernoulli, Gaussian and Multinomial Naive Bayes Classifiers?
- Bernoulli is when features are 0 or 1
- Multinomial is when features are counts
- Gaussian is when features are continuous and normally distributed
Can you use the Naive Bayes Classifier if your features are not independent?
- You shouldn't, as that's the primary assumption before using this classifier.
Models.
Take this data. https://github.com/gSchool/dsi-logistic-regression/blob/g79/data/grad.csv
Predict whether someone will get into grad school. Use the following models.
- Logistic Regression
- Random Forest
- Naive Bayes (you will need to figure out what type works best for this data)
Which model performed the best?
Text Classification.
Remember this assignment.
https://github.com/data-science-ml/tweets-nlp-assignment/blob/master/nlp-assignment.md
Take the above tweets and turn them into a bag of words. Use a Naive Bayes classifier to figure out if a particular tweet is Neutral, Negative or Positive. Remember to split your data.
What is the accuracy of your model?
Compare this model to a KNN model (neighbors == 3) where each tweet is a 300 dimensional vector.
Which model performs better?