[scikit-learn] custom estimator with more than two arguments to fit()
Hi all, I'm fairly new to scikit-learn, but have been using a predictive model for a while now that would benefit from scikit-learn's estimator API. However, I could use some advice on how best to implement this. Briefly, the model is a combination of dimension reduction and nearest neighbors, but the dimension reduction step (canonical correspondence analysis - CCA) relies on two matrices to create the synthetic feature scores for the candidates in the nearest neighbor step. The two matrices are a "species" matrix (spp) and an "environmental" matrix (env) which are used to create orthogonal CCA axes that are linear combinations of the environmental features. In reading through the documentation on creating new estimators, it seems that every estimator should provide a fit(X, y) method. Somehow I need my X parameter to be both the spp and env matrices together. I got a lot of good inspiration from this post on Stack Overflow: https://stackoverflow.com/questions/45966500/use-sklearn-gridsearchcv-on-custom-class-whose-fit-method-takes-3-arguments and can mostly understand how the OP implemented this, basically by creating a DataHandler class that packs together the two matrices, such that the call to fit would look like: estimator.fit(DataHandler(spp, env), y) I'm wondering if this is the best way to handle the design or if I'm not fully understanding how I could use a Pipeline to accomplish the same goal. Thanks for any guidance - boilerplate sample code would be most appreciated! matt ___ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn
Re: [scikit-learn] custom estimator with more than two arguments to fit()
Hi Matt, We do have CCA and other PLS-related transformers / regressors in scikit-learn. They are able to do dimensionality reduction on both X and Y (which I believe correspond to spp and env), so you might want to have a look at these. However, they're not fully compatible with the whole ecosystem unfortunately: for example our Pipeline objects assume that only X can be transformed, not Y. Nicolas On 7/31/20 12:02 PM, Gregory, Matthew wrote: Hi all, I'm fairly new to scikit-learn, but have been using a predictive model for a while now that would benefit from scikit-learn's estimator API. However, I could use some advice on how best to implement this. Briefly, the model is a combination of dimension reduction and nearest neighbors, but the dimension reduction step (canonical correspondence analysis - CCA) relies on two matrices to create the synthetic feature scores for the candidates in the nearest neighbor step. The two matrices are a "species" matrix (spp) and an "environmental" matrix (env) which are used to create orthogonal CCA axes that are linear combinations of the environmental features. In reading through the documentation on creating new estimators, it seems that every estimator should provide a fit(X, y) method. Somehow I need my X parameter to be both the spp and env matrices together. I got a lot of good inspiration from this post on Stack Overflow: https://stackoverflow.com/questions/45966500/use-sklearn-gridsearchcv-on-custom-class-whose-fit-method-takes-3-arguments and can mostly understand how the OP implemented this, basically by creating a DataHandler class that packs together the two matrices, such that the call to fit would look like: estimator.fit(DataHandler(spp, env), y) I'm wondering if this is the best way to handle the design or if I'm not fully understanding how I could use a Pipeline to accomplish the same goal. Thanks for any guidance - boilerplate sample code would be most appreciated! matt ___ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn ___ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn
Re: [scikit-learn] custom estimator with more than two arguments to fit()
Hi Nicolas, Nicolas Hug wrote: > We do have CCA and other PLS-related transformers / regressors in > scikit-learn. They are able to do dimensionality reduction on both > X and Y (which I believe correspond to spp and env), so you might > want to have a look at these. However, they're not fully > compatible with the whole ecosystem unfortunately: for example our > Pipeline objects assume that only X can be transformed, not Y. Just to clarify, I'm only seeing canonical *correlation* analysis and not canonical *correspondence* analysis (ter Braak) in scikit-learn? https://scikit-learn.org/stable/modules/generated/sklearn.cross_decomposition.CCA.html But your point is taken - I can use this for inspiration because it has both X and Y matrices. But if I'm understanding correctly, there is no way to couple this with a further step of NearestNeighbors into a pipeline? I will only need the transformed scores coming out of CCA to feed into the NearestNeighbors step. Sorry if I'm not understanding this correctly. matt ___ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn
