Introduction to the MathOverflow Tag Recommendation Problem
Here is the first paragraph of a recent post on the front page of MathOverflow:
More …Data Exploration and Preprocessing
We take a look at the data which comes from the quarterly Stack Exchange data dump. We explore the data to understand how it is structured and clean the data.
More …Multilabel Stratified Split
Most train/valid/test split tools are not optimized for multilabel problems. The tool MultilabelStratifiedShuffleSplit from iterstrat.ml_stratifiers (see the github page) implements the algorithm from Konstantinos Sechidis, Grigorios Tsoumakas & Ioannis Vlahavas (2011).
More …PMI Model
Andrej Karpathy makes a distinction between what he calls software 1.0 and software 2.0. Software 1.0 consists of explicit instructions for transforming inputs into desired outputs. Software 2.0 is machine learning: we provide a model with a ton of parameters and minimize a loss function. The trained model then transforms inputs into desired outputs in a way which performs well on the training data, and which (we hope!) will generalize to novel data.
More …Distilbert with a Simple Classifier Head
We summarize the work done in this Colab notebook.
More …