ImgurClassifier

This project classifies imgur posts as the likelihood of making it to the front page of imgur.

Dataset is multimodal, consisting of (3x) images, various text, categorical, and numeric features.

Purpose

Primary purpose is to demonstrate advanced feature engineering techniques in ML.NET.

Techniques demonstrated:

Model stacking
- k-means cluster distance featurizer
- Text target encoding (stacked linear model on text ngrams)
- Alt-label -- stacking of a regression model towards an alternative label
- Model stacking using only subset of features (images, text, ...)
- Stacking of AutoML models
Multi-threaded parallel model creation
Expression transform to create a Weight column
Image featurization
Label rotation to allow a regression model to train on a multi-class dataset
String statistics (length, number of vowels, word count, ...) using a Custom Mapper
Explainability -- feature importance from model weights via training a proxy model