November 22

2016-12-22

TODO(MatthewJA)

• Discuss milestones
• Discuss design decisions from 0.1(?)
• Set up recurring meetings

Design Decisions from 0.1

• Committee should be generalised
• Committees should be part of the Predictor environment rather than the Recommender environment
• Predictors should wrap scikit-learn lightly; write a function that takes a scikit-learn predictor (class or instance) to a Predictor object.
• Swap Predictor interface to use IDs instead of NumPy arrays.

Things for Future

• Implement the above.
• -w flag that mutates input file e.g. -wo colname.
• Add an example folder that includes an example comparing a command-line application of Acton to a Python script-based application.
• Factor out Matplotlib code.
• Write down why we can't just use scikit-learn's API — why do we need to use Predictor?
• Port some of Alasdair's recommenders. #30
• Make crowdastro galaxy classification task work. #31
• Write a README.
• Write a quick-start guide.
• Milestone 0.1.1: Support binary classification. Evaluation by reproducing Figure 5.1 from my thesis.
• Milestone 0.2: Support multiclass classification. Evaluation by reproducing Figure 5.3 from Alasdair's thesis.