research: tune the classifier inside `score_leadership_prospects` on a fiscal-year grouped fold

オープン
  • 何か問題がありますか？
#349 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
4/5
見積もり時間
3〜5日
初心者へのやさしさ
35/100
issue の種類
機能追加
明瞭さ
おおむね明確
活発さ
活発
技術スタック
python, scikit-learn
領域
machine-learning

調査の方向性

Start in philanthropy/models/_upgrade.py at score_leadership_prospects to see which inner classifier settings are exposed, then read philanthropy/model_selection/_temporal_donor_splitter.py for FiscalYearGroupedSplitter. Run a small grid with that splitter, not a random split, and compare hit rates on the KDD98, DonorsChoose and PSID Results files via scripts/make_results_pages.py. Done means a changed default only if it wins on validation and loses on no test fold, or a recorded rejection with numbers. Comment on the issue before starting.

索引モデルが issue の本文から書いたものです。

説明

enhancement help wanted roadmap
Goal

Check whether the leadership-upgrade model's inner classifier settings can be improved without leaking future years.

Acceptance criteria
  • Tune a small grid of MajorGiftClassifier settings with FiscalYearGroupedSplitter, never a random split. Start from tree shape (max_leaf_nodes, min_samples_leaf) and class_weight: the MajorGiftClassifier docstring already records that learning_rate=0.05, max_iter=300 scored the same as the defaults on a KDD Cup 1998 validation fold (response target, five seeds), because early stopping kicks in above 10,000 rows. Re-test those two only on the leadership target, where files are smaller.
  • Tune on the same rows the function trains on: build them with philanthropy.ingest.build_leadership_snapshots and fit MajorGiftClassifier directly. score_leadership_prospects takes no classifier settings (it builds MajorGiftClassifier(random_state=...) at philanthropy/models/_upgrade.py around lines 461 and 530).
  • Compare against the current defaults on the existing Results files (KDD98, DonorsChoose, PSID) on top-10% hit rate with intervals.
  • Change the default only if it wins on the validation fold and doesn't lose on the test fold of any file; otherwise record the rejection with numbers. A winning setting changes the two MajorGiftClassifier(...) calls in _upgrade.py, not MajorGiftClassifier's own defaults, which the response Results also depend on.
Scope

Tuning only. The report keys and the CLI output stay the same.

Where to start
  • philanthropy/models/_upgrade.py (score_leadership_prospects)
  • philanthropy/model_selection/_temporal_donor_splitter.py
  • scripts/make_results_pages.py
Dependencies
  • None

Phase order is a release plan, not a rule against picking up independent work early. Comment on this issue before you start so two people don't build the same thing.

主要言語
Python
スター
11
フォーク
23
平均マージ
2日 19時間
マージ済み PR（30日）
31

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

PhilanthroPy-Project/PhilanthroPy のほかの issue

PhilanthroPy-Project/PhilanthroPy の issue をすべて見る

似ている issue

Python の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。