This is the folder for PetFinder.my Adoption Prediction competition, team "BaMoOaAl"
Team Components:
- https://www.kaggle.com/alessandrosolbiati, SolbiatiAlessandro
- https://www.kaggle.com/oanaflorescu, flores-o
Contributors:
- https://github.com/ferrucc-io, ferrucc-io (dramatically helpful in fixing typos)
We will follow along Stanford CS231-n assignments and implement them here on the competition. note: stage1 ended on 09/04/2019 competition results: we arrived 359th out of 1805, 18% percentile
Models:
- KNN: best validation score: 0.20 + 0.14 public LB score: 0.279 BIMODEL
- NB: best validation score: 0.10249 public LB score: 0.172
- implement SVM: best validation score: public LB score:
- implement NN (ResNet transfer learning)
- LGBM: best validation score: 0.17435 public LB score: 0.278
- CATBOOST: best validation score: 0.20133 public LB score: 0.349
Framework:
- write standard PredictiveModel
- write test
- write benchmark/execution scripts ( we are using notebooks )
- write docs with model performance and insight
- add code coverage
- [X] write implementation for Google Cloud Machine Learning Engine to run models on cloud (inside GCP)
Exploratory Data Analysis + Feature Engineering
- Adoption Speed
- Name
- Age
- Breed
- Color
- Size
- Country
- Images
This stage consist in finishing what we were planning to do during competition, and set up better framework/knoweldge on this known competition instead of jumping in on new competitions.
- set up a consistent GCP framework/pipeline for future competition
- explore and implement the recent autoML approaches to Kaggle
- get insight on the specific competition
GCP FRAMEWORK/PIPELINE
- build a running prototype running on cloud
kaggle-pets/models/TREES/XGBOOST/gcp_training/ - write a document about GCP
- start creating a template repo to clone for every new competition
autoML
- explore existing solution out there and evaluate wheter to implement myself
Competition
- read solution and replicate them