I looked into the hierarchy of both py and scala sides, and found that they are quite different, which damage the parity and make the codebase hard to maintain.
The main inconvenience is that most models in pyspark do not support any param getters and setters.
In the py side, I think we need to do:
1, remove setters generated by _shared_params_code_gen.py;
2, add common abstract classes like the side side, such as JavaPredictor/JavaClassificationModel/JavaProbabilisticClassifier;
3, for each alg, add its param trait, such as LinearSVCParams;
4, since sharedParam do not have setters, we need to add them in right places;
Unfortunately, I notice that if we do 1 (remove setters generated by _shared_params_code_gen.py), all algs (classification/regression/clustering/features/fpm/recommendation) need to be modified in one batch.
The scala side also need some small improvements, but I think they can be leave alone at first