Skip to content

Latest commit

 

History

History
85 lines (59 loc) · 3.9 KB

README.md

File metadata and controls

85 lines (59 loc) · 3.9 KB
Raw

didbits

An unsorted bunch of ideas on practical aspects of machine learning. This is aimed at the scientist (biologist, astronomer) who is starting to use machine learning tools.

If you have a good description for any of the concepts in Ideas please send me a pull request. Use the issue tracker for suggested topics.

Concepts with more detailed descriptions

Mostly in python.

  • Double loop cross validation CrossVal When a classifier has hyperparameters that need to be set, a separate validation set of data needs to be set aside. The double loop refers to the need for an inner loop to find good hyperparameters and an outer loop to compute test error.

  • Normalising features Normalisation/normalise_vector.ipynb Various vector transformations to make learning work well. Could still have bugs with transpose of matrix

  • Normalising the kernel matrix Normalisation/normalise_kernel.ipynb In contrast to normalising features, we can normalise the kernel matrix. This is particularly useful when combining kernels, e.g. with MKL.

  • Reshaping data with pandas DataWrangling/process_abs.ipynb shows an example of converting data into a convenient tabular format.

  • Performance measures. Accuracy/evaluation.py It is worth repeating things like balanced error, area under the Receiver Operating Characteristic curve, etc. Many of the concepts are best explained by considering the confusion matrix.

Ideas

  • No peeking. It is not obvious what generalisation error is, and how to not do things like report variance explained. At one level, this is education about cross validation and bootstrap. At another level, in long complicated pipelines, test data may be used in training. The famous example is Vapnik using some feature selection method to choose two genes on the whole dataset, and showing the SVM works well with only two features.

  • Data wide predictions. Related to above. Very often in biology, there are finite spaces (e.g. the genome). So they want to make predictions genome wide. Of course the question is then how to train since you are predicting everywhere. One option is to do leave one chromosome out prediction. But then it is hard to explain that you actually have 23 different classifiers, and not just one.

  • Visualisation of data and results. Things like PCA and clustering. See also t-SNE.

  • How to subsample data. Perhaps for computational reasons, you may want to sanity check your pipeline on a small sample. Taking the first k items from your file is generally not a good idea. There are also other good reasons, e.g. cross validation, quick debugging, visualisation. Good ways to do single pass sampling are not obvious.

  • There is no statistical test for equality. Hard to convince an experimentalist that there is no p-value for showing that two things are the same.

  • Predictive densities. For the more technically savvy crowd, getting a distribution is much better than getting point estimates.

  • Normalisation. It is worth repeating that normalisation changes results for classifiers such as the SVM. How to deal with it? Who knows

  • Feature selection. Biologists are very keen on finding predictive models that they can understand. So feature selection is an end goal.

  • Linear correlation vs rank correlation

  • correlation does not equal causality

  • Dangers of thresholding. Very often for quality control, people throw away data, discretise values, or set something to zero. This introduces complex changes due to the cliff in the distributions. There are other reasons for thresholding, e.g. prior knowledge that values are positive, or removing outliers from the data.

  • Dependent random variables. Graphical models, etc.

  • Computing entropy and mutual information practical tricks for counting, accounting for zeros, and binning.

  • Open science