0408happyfeet/books-tabular-autogluon

Books Tabular — AutoGluon (Classical ML)

Task: Predict Is_Textbook (Yes/No) from physical book attributes.

Dataset: EricCRX/books-tabular-dataset

Training: AutoGluon Tabular with presets="best_quality" on the augmented split, evaluated on the original split.

Results (Original split)

• Accuracy: 1.0000
• Weighted F1: 1.0000
• Weighted Precision: 1.0000
• Weighted Recall: 1.0000
• ROC-AUC (Yes vs No): 1.0000

Best Model

• Overall: CatBoost/T18
• Best base model: CatBoost/T18

Hyperparameters

Best overall (leaf): CatBoost/T18 (CatBoostModel)

{
  "iterations": 10000,
  "learning_rate": 0.13950397023933195,
  "random_seed": 0,
  "allow_writing_files": false,
  "eval_metric": "Accuracy",
  "depth": 7,
  "l2_leaf_reg": 3.4522538315365296
}

Repro steps

pip install autogluon.tabular datasets huggingface_hub cloudpickle
from huggingface_hub import hf_hub_download
from autogluon.tabular import TabularPredictor

# Download (native dir zip) from 0408happyfeet/books-tabular-autogluon
zip_path = hf_hub_download(repo_id="0408happyfeet/books-tabular-autogluon", repo_type="model", filename="autogluon_predictor_dir.zip")
# Unzip and load:
#   unzip autogluon_predictor_dir.zip -d ./model_dir
#   predictor = TabularPredictor.load("./model_dir")

# Or quick (pickle):
pkl_path = hf_hub_download(repo_id="0408happyfeet/books-tabular-autogluon", repo_type="model", filename="autogluon_predictor.pkl")
import cloudpickle
predictor = cloudpickle.load(open(pkl_path, "rb"))

Inference

import pandas as pd
X = pd.DataFrame([{
    "Length_cm": 24.0,
    "Width_cm": 16.0,
    "Thickness_cm": 3.0,
    "Pages": 400,
    "Hardcover": "Yes",
    "Cover_Color": "Blue"
}])
pred = predictor.predict(X)
proba = predictor.predict_proba(X)
print(pred.iloc[0], proba)

---

Purpose

Educational AutoML exercise for classical ML on a peer’s Homework 1 dataset.

Data origin & splits

Source dataset: EricCRX/books-tabular-dataset (classmate). Splits used:

• augmented (~300 rows) for training/HPO with an 80/20 holdout.
• original (30 rows) held out as an external test set.

Features & target

Features: Length_cm, Width_cm, Thickness_cm, Pages, Hardcover, Cover_Color. Target: Is_Textbook ∈ {Yes, No}.

Preprocessing

AutoGluon Tabular defaults (type inference, category handling, missing-value processing). No manual feature engineering.

Training setup

• Framework: AutoGluon Tabular (preset: best_quality).
• Model families: GBM, XGB, RF, XT, KNN, CAT (classical only).
• HPO: hyperparameter_tune_kwargs='auto' (AutoGluon scheduler/strategy).
• Budget: time_limit=1200 seconds.
• Seed: 42.
• Compute: Google Colab, Tesla T4 GPU; RAM ~12.7 GB.
• Wall-clock: unknown.

Hyperparameters & search space

• Best model: CatBoost/T18 (CatBoostModel).
• Search space: AutoGluon defaults under best_quality with automatic HPO.

Metrics (original split)

Units: fraction (0–1), where 1.0 = 100%.

• Point estimates: Accuracy / F1 / Precision / Recall (+ ROC‑AUC if available) as reported above.
• Uncertainty (bootstrap):
• Accuracy: 1.0000 (95% CI [1.0000, 1.0000])
• F1_weighted: 1.0000 (95% CI [1.0000, 1.0000])
• Precision_weighted: 1.0000 (95% CI [1.0000, 1.0000])
• Recall_weighted: 1.0000 (95% CI [1.0000, 1.0000])
• ROC-AUC: 1.0000 (95% CI [1.0000, 1.0000])

Limitations & ethical notes

• Very small real test set (n=30); synthetic augmentation can misrepresent real‑world distribution.
• No PII; low ethical risk in this domain (books).

License

MIT (matching the source dataset’s license).

Hardware & compute budget

• HPO budget: 1200s.
• Runtime hardware: Google Colab; GPU Tesla T4 (15 GB VRAM); CPU x86_64 (2 vCPU); RAM ~12.7 GB.
• Training wall‑clock: unknown.
• Peak memory: ~650.0 MB.

AI usage disclosure

• AutoML & HPO performed by AutoGluon.
• Some documentation/code scaffolding assisted by generative AI (ChatGPT); human‑reviewed.