Evaluate existing model experiments on why they perform how they do.

[StefanKennedy]

ACs:

• Classifiers missing important metrics have metrics computed
• Assess result of classifiers in a report in the notebook
• Include metrics decided in #146 in notebook report

[StefanKennedy]

Found a resource that explains how to plot useful graphs, it certainly helps with explaining why logistic regression / SVMs can divide the samples: https://www.dummies.com/programming/big-data/data-science/how-to-visualize-the-classifier-in-an-svm-supervised-learning-model/

For example:

[StefanKennedy]

Logistic regression can be visualized using something like this:

https://stackoverflow.com/questions/46085762/sklearn-logistic-regression-plotting-probability-curve-graph

[StefanKennedy]

First visualization of how LinearSVC divides the data:

[StefanKennedy]

Evaluations:

• How is the variance with unseen data? A high change from validation data to unseen data can indicate overfitting.
• Bias error. (High in Linear Discriminant Analysis and Logistic Regression, low in k-Nearest Neighbors and Support Vector Machines) More likely to oversimplify a problem.
• Variance error (trade off with bias, so the high/low classifier variance list is the invert of the list of high/low classifier bias) Should see a large change when switching training data.

[StefanKennedy]

These graphs were created by performing a lot of dimensionality reduction, and you can see that a linear divide is not going to be able to separate the samples at this dimensionality.

[StefanKennedy]

Assessing features with logistic regression curves:

From these basic visualisations we can see the following:

• As the max number of reviews in one day increases, the probability of a review being deceptive increases
• As the average length of a reviewer's reviews increases, probability of a review being deceptive decreases
• As the reviewer's review standard deviation increases, the probability of a review being deceptive decreases

[StefanKennedy]

Bayesian graphs: https://dataconomy.com/2015/02/introduction-to-bayes-theorem-with-python/

[StefanKennedy]

Naive bayes classification, on a subset of the samples:

[StefanKennedy]

Naive bayes feature selection :fireworks: