What are some good machine learning programming exercises? [closed]

Question

Ideally, they would have the following characteristics:

1. They can be completed in just an evening of coding. It will not require a week or more to get interesting results. That way, I can feel like I've learned and accomplished something in just one (possibly several hour long) sitting.

2. The problems are from the real world, or they are at least toy versions of a real world problems.

3. If the problem requires data to test the solution, there are real-world datasets readily available, or it is trivial to generate interesting test data myself.

4. It is easy to evaluate how good of a job I've done. When I test my solution, it will be clear from the results that I've accomplished something nontrivial, either by simple inspection, or by a quantifiable measure of the quality of the results.

Answer 1

Implement the following algorithms:

• Perceptron, margin perceptron: you can try to detect images of faces (classify images of faces and non-faces) using any face database. Try for example the MIT CBCL face database. You can also try the MNIST data and write a poor man's OCR system.
• LVQ, Kohonen map: you can try to compress images. You can download large images from any wallpaper site.
• Naive bayes classifier: you can classify spam and not spam. There are also more scientific datasets, such as Reuters and Newsgroups, etc. which you have to determine the topic, given the article.
• Backpropagation, multi layer perceptron: you can try this with the faces, or with the spam, or with the text/histogram data.
• Primal SVM linear learning using SGD: you can try this with MNIST digits, for example.

There are a bunch of projects, some of them take a couple hours, some a couple of days, but you will definitely learn a lot.

Answer 2

Check the UCI machine learning repository out for real datasets.

The Breast Cancer Wisconsin (Diagnostic) Data Set for example. Check the data set description for more information about it.

Even the Naive Bayes classifier will give great results on this dataset (over 95% cross-validated accuracy). With some variable selection you can even get to 100%, if I remember correctly.