Skip to content

Latest commit

 

History

History
82 lines (46 loc) · 2.59 KB

File metadata and controls

82 lines (46 loc) · 2.59 KB

Simdi

  • Looking into error of possible different agg models for ensemble purposes

  • main thing that works: training and testing on the 13 data!

  • Possible questions that it brings

    • What is its performace on lower dimensional data?
    • how to transfer this into lower dimensional data?
  • Eror Analysis:

    • the main postion of the error comes from the recall.
      • Q: why do I have this much error in the recall? And how to decrease knowing that we only take 4% recall?
      • An option maybe to feed the data into a graph! why graph you ask? Or models that take into account a dependence between all the data that we have at the moment of the prediction
  • sample 50-50 data from the dataset for taining (did not improve the performance)

  • the cat model with the simplest agg model -> using the cat variables as a cont var with LabelEncoder

  • use the sum of the cat features and create and train a cat model

  • ivnestigate possible different aggregations

  • run the lightgbm, xgb with a variaty of different features

  • SFA SFA suggests that some feature may be more informative when transformed. Currently investigating combinations of features

    • What does the results of SFA imply?

    • possibly come up with a distribution of Weak learners sampling from the dist of sinfle factor analysis

  • Different precisons from the official pytorch

Bilinenler

  • I have a convolutional model that works pretty good for the C13
  • I know that recall is the main bottleneck in its error metric
  • I have started training weak models. -> a possible path is to train so many weak models give it to a ensemble learner that inegrates the preds

Roadmap

  • train SFA

    • on c13 data
  • evaluate errors

  • later propogate the results into the rest of the dataset

  • Add the predictions from other models to the features that you have. Initial attmepts with XGB did not perform the best! This is the way to go..

  • Investigate ranked probabilities

Results

  • ContCols + Cat cols transformed -> 793 (could possibly have been improved with CV and early stopping)

  • K79 + 6 best 13 overfits significantly with 799 on validation data and somehow 785 on real test data!!

Current repeat procedure

FE

Main finding here: Different features result in a different complementary predictions - CNN with SFA

Models

HPO

  • consider using the max_bin_by_feature of the lightgbm

CV

Ensemble

How do we combine all the predictions (that are complementory into each other) into a final prediction

How do we add a prediction from one model into the other for complementory predictions?