Increase training size for feature level classifier

[bkowshik]

Ref https://github.com/mapbox/gabbar/issues/43

---

• We currently use 5,269 changesets for training our feature level classifier.
• From changesets reviewed on osmcha with one feature modifications, it looks like we can potentially add upto 4,000 changesets.
• This increase in the samples in the training dataset should in-turn improve the model.

Next actions

• [ ] Update dataset with the additional 4,000 changesets - @bkowshik

---

cc: @batpad @geohacker

[bkowshik]

Curious to see the effect training size of the model has on the metrics, we have the following:

Notes / Questions

• The metrics although diminishing have a significant positive slope.
• If roc_auc score is 0.8 with 6,000 samples, what would it look like with 10,000 samples?
• When do we know that we have enough samples?

---

cc: @anandthakker

[bkowshik]

Workflow

1. Set number of samples to use for the current run
2. Use only this subset of samples from the labelled training data
3. Train a model on this subset of training data
4. Get predictions from model for the entire validation dataset
5. Extract metrics on validation dataset
6. Increase number of samples to use for the next run and go again

[bkowshik]

Before we had 8,620 labelled samples out of which 6,036 was used for training and 2,584 for validation. With the backfill done, we now have 10,165 out out which we use 7115 for testing and 3050 for validation.

• In total we added 1,545 new changesets to the labelled dump. 🎉

Interestingly, the nice upward graph now has become something like below. I don't understand why this is happening though.

We are 💯 to close here.