Few preprocessing techniques are not deterministic

[rabsr]

Describe the bug

Following preprocessing components are not deterministic. Output generated is different every time. I have tested these components individually.

• QuantileTransformer
• SelectPercentile (Classification and regression) when mutual info is used as score
• SelectRates (Classification and regression) when mutual info is used as score
• TruncatedSVD

Expected behaviour

Given constant random state/seed, the output must be same each time for a given technique.

Environment and installation:

Please give details about your installation:

• Python version - 3.7.3
• Auto-sklearn version - 0.12.6
• Scikit-learn - 0.24.1

[mfeurer]

Thank you very much for debugging this. Indeed, we're missing a random state for those components. Would you like to create a PR to add them for quantile regressor and truncated SVD? They should be simple and easily mergable. For select percentile and select rates there is no random state option, so I assume this is an issue with scikit-learn, and you would first have to open an issue there.

[rabsr]

@mfeurer I will raise a PR. There is already a similar issue raised with scikit-learn. I do have simple solution in mind for select rates and select percentile until it is fixed in scikit-learn. I will include it in PR. I will look for your feedback.

[rabsr]

Fixed via #1172