Doubt

Get uncertainty estimates of machine learning models — blog post

A Python package that implements various algorithms that allows machine learning practitioners to get uncertainty estimates from their machine learning models. This includes wrapping their existing models in a bootstrap wrapper, as well as implementing stand-alone models such as linear quantile regression and quantile regression trees.

Scholarly

Classification of scientific papers — blog post

This project aims to classify the category of scientific papers, based on the subject classification on arxiv.org. Try the live demo here:

ttilt.dk/scholarly

From the ArXiv API I scraped ~1.3 million titles and abstracts, which was stored in a SQLite database. On this text corpus I trained my own FastText bigram word vectors. My classification model was an attention-based recurrent neural network, similar to SHA-RNN, to classify the texts into the 148 subject categories.

The subject categories can be organised in 6 different ‘master categories’, such as Mathematics and Physics, and I developed my own loss function to take this into account, so that even when the model mislabels a text then it will assign a label within the same master category. The loss function also takes class imbalance into account, both on the category- and master category level.

Every text can be assigned to multiple categories, making this a multilabel classification problem. A metric relevant to such a task is the sample-average F1 score, which computes the F1 score of the predictions associated to every sample, with the final score being the average of these F1 scores. In a multiclass scenario this would coincide with the accuracy, but with multiple labels it gives more information as it rewards the model for partial results.

The best model achieved a ~93% and ~65% validation sample-average F1 score on the master categories and all the categories, respectively. This was trained on the University of Bristol’s compute cluster BlueCrystal Phase 4 compute cluster, on a single GPU node for a day.

AutoPoet

Build poems from text sources — blog post

Automatically build Haiku poems from a given text source. To ensure that the syllables are counted correctly a recurrent neural network has been trained to syllabise English words with a 97% accuracy. This model was trained on The Gutenberg Moby Hyphenator II dataset, consisting of ~170k hyphenated English words.

NaturalSelection

An all-purpose Pythonic genetic algorithm — blog post

A Python package which uses genetic algorithms to maximise a given function. A prime example of this is to hyperparameter optimise neural networks. See the readme for several examples and a more in-depth explanation of the algorithm.

Image showing an example of the evolution of a population

Athelstan

A comparative analysis of constituencies in England and Wales

This project was a first attempt at a cluster analysis of constituencies in England and Wales based on eight different variables related to demographics, business, qualifications and amenities. A use case would be to find similar constituencies as a given location in one of these two countries. Data on constituencies and their gps coordinates are from doogal.co.uk, and all the other data are from from Office for National Statistics.

This project was the capstone project in the IBM applied data science specialisation on Coursera.