Smart Product Pricing — Amazon ML Challenge 2025

Price prediction for e-commerce products using text-only features and a lightweight 2-model ensemble.

• Public LB Score: 54.6631 SMAPE
• Final file submitted: sub_ensemble-v1.csv
• Metric: SMAPE (Symmetric Mean Absolute Percentage Error, lower is better)

---

1) Problem Statement

Given a catalog of products with text fields, predict the price of each product in the hidden test set.

• Input: sample_id, catalog_content (title/description text)
• Output: sample_id, price for all items in test.csv
• Goal: minimize SMAPE between predictions and true prices

---

2) Dataset & Submission Format

Put raw CSVs (not committed to git) in the local dataset/ directory:

dataset/
├─ train.csv    # 75,000 rows: sample_id, catalog_content, price
└─ test.csv     # 75,000 rows: sample_id, catalog_content

Columns (train.csv)

column	type	notes
sample_id	int/str	unique row id
catalog_content	text	title/description text
price	float	target to predict

Submission CSV

sample_id,price
10000001,123.45
10000002,89.99
...

All predicted prices are positive (clipped at 0.01) and row order follows test.csv.

---

3) Approach (High Level)

A fast, robust text-only pipeline that avoids leakage:

1. Normalize text (lowercase, collapse spaces).
2. Build two complementary feature families:
   • TF-IDF n-grams (word 1–2, char 3–5; ~300k dims).
   • TextNumerics: regex/engineered stats from text (lengths, digit counts/ranges, pack × qty, etc.).
3. Train Level-0 models with grouped 5-fold CV:
   • Ridge (L2) on TF-IDF.
   • XGBoost on TextNumerics (trained on log1p(price)).
4. Median-ratio calibration in price space.
5. Weighted blend of the two Level-0 predictions → final submission.

Architecture

flowchart LR
  A[train.csv / test.csv] --> B[Normalize Text]
  B --> C1[TF-IDF word 1-2 char 3-5]
  B --> C2[TextNumerics len words digits ranges IPQ pack-qty qty_count]
  C1 --> D1[Ridge L2]
  C2 --> D2[XGBoost hist]
  subgraph CV[GroupKFold by text-hash]
    D1 --> OOF1[OOF Ridge preds]
    D2 --> OOF2[OOF XGB preds]
  end
  OOF1 --> E[Blend + Calibrate]
  OOF2 --> E
  E --> F[sub_ensemble-v1.csv]

Loading

---

4) Cross-Validation & Calibration

• Grouping: group = SHA1(normalized_catalog_content)[:12]
  Ensures near-duplicates stay in the same fold (no leakage).
• Target space: Train models on log1p(price); convert back with expm1.
• Calibration: scale = median(train_price) / median(expm1(OOF_log)) applied to test predictions.

---

5) How to Reproduce

Environment (Windows PowerShell)

python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -r requirements.txt

Place train.csv and test.csv into dataset/.

---

A) Build TF-IDF & Train Ridge

import os, numpy as np, pandas as pd, scipy.sparse as sp, hashlib
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import Ridge
from sklearn.model_selection import GroupKFold

BASE = "."
DATA = f"{BASE}/dataset"
FEAT = f"{BASE}/features/tfidf"
SUB  = f"{BASE}/submissions"

os.makedirs(FEAT, exist_ok=True)
os.makedirs(SUB,  exist_ok=True)


def norm(s):
    s = s.fillna("").str.lower()
    return s.str.replace(r"\s+", " ", regex=True).str.strip()


def groups(text):
    t = norm(text)
    return t.map(lambda x: hashlib.sha1(x.encode()).hexdigest()[:12]).values


# Load data
train = pd.read_csv(f"{DATA}/train.csv")
test  = pd.read_csv(f"{DATA}/test.csv")
tr    = norm(train["catalog_content"])
te    = norm(test["catalog_content"])

# Build TF-IDF features
w = TfidfVectorizer(ngram_range=(1, 2), max_features=200_000, min_df=2)
c = TfidfVectorizer(analyzer="char", ngram_range=(3, 5), max_features=100_000, min_df=2)

X_tr = sp.hstack([w.fit_transform(tr), c.fit_transform(tr)], format="csr")
X_te = sp.hstack([w.transform(te),     c.transform(te)],     format="csr")

sp.save_npz(f"{FEAT}/X_tr_tfidf.npz", X_tr)
sp.save_npz(f"{FEAT}/X_te_tfidf.npz", X_te)
np.savez_compressed(
    f"{FEAT}/labels_ids.npz",
    y      = train["price"].astype("float32").values,
    sid_tr = train["sample_id"].values,
    sid_te = test["sample_id"].values,
)

# Cross-validated Ridge
y      = np.load(f"{FEAT}/labels_ids.npz")["y"]
sid_te = np.load(f"{FEAT}/labels_ids.npz")["sid_te"]
g      = groups(train["catalog_content"])
gkf    = GroupKFold(5)

oof        = np.zeros_like(y, dtype="float32")
test_preds = []

for fold, (tr_idx, va_idx) in enumerate(gkf.split(X_tr, y, g)):
    m = Ridge(alpha=1.0, random_state=42)
    m.fit(X_tr[tr_idx], np.log1p(y[tr_idx]))
    oof[va_idx] = m.predict(X_tr[va_idx])
    test_preds.append(m.predict(X_te).astype("float32"))

pred = np.clip(np.expm1(np.mean(test_preds, axis=0)), 0.01, None)
pd.DataFrame({"sample_id": sid_te, "price": pred}).to_csv(
    f"{SUB}/sub_tfidf_ridge.csv", index=False, float_format="%.6f"
)
print("Wrote:", f"{SUB}/sub_tfidf_ridge.csv")

---

B) Train XGBoost on TextNumerics

python src/models/run_textonly_baseline.py \
    --data_dir dataset \
    --out_dir  artifacts \
    --nfolds   5
# writes: submissions/sub_textnum_xgb.csv

---

C) Blend to Final Submission

import pandas as pd, numpy as np, os

BASE = "."
SUB  = f"{BASE}/submissions"
os.makedirs(SUB, exist_ok=True)

xgb = pd.read_csv(f"{SUB}/sub_textnum_xgb.csv")
rid = (
    pd.read_csv(f"{SUB}/sub_tfidf_ridge.csv")
    .set_index("sample_id")
    .reindex(xgb["sample_id"])
    .reset_index()
)

W_XGB, W_RIDGE = 0.6, 0.4
ens = np.clip(W_XGB * xgb["price"] + W_RIDGE * rid["price"], 0.01, None)

pd.DataFrame({"sample_id": xgb["sample_id"], "price": ens}).to_csv(
    f"{SUB}/sub_ensemble-v1.csv", index=False, float_format="%.6f"
)
print("Wrote:", f"{SUB}/sub_ensemble-v1.csv")

Submit submissions/sub_ensemble-v1.csv.

---

6) Repo Layout

amazon-ml-2025/
├─ README.md
├─ LICENSE
├─ requirements.txt
├─ .gitignore
├─ submissions/
│  ├─ sub_textnum_xgb.csv
│  ├─ sub_tfidf_ridge.csv
│  └─ sub_ensemble-v1.csv      # final submitted
├─ src/
│  └─ models/
│     └─ run_textonly_baseline.py
├─ dataset/                    # put train.csv & test.csv here (git-ignored)
├─ artifacts/
├─ features/
├─ meta/
├─ lvl0/
└─ logs/                       # all local-only (git-ignored)

---

7) Notes & Future Work

• Replacing/augmenting TextNumerics with Sentence-Transformers or OpenCLIP embeddings (with TTA) is a natural next step.
• A Level-1 meta-learner (e.g., LightGBM on OOF predictions) can further stabilize ranking.
• The current pipeline is intentionally lightweight and fully reproducible on Colab or a local CPU.

---

8) License

Released under the MIT License — see LICENSE.