Data Science Applications in E-commerce: Real-World Examples

from collections import Counter
from itertools import combinations

orders = [
    ["phone", "case", "charger"],
    ["phone", "case"],
    ["laptop", "mouse"],
    ["phone", "charger"],
    ["laptop", "mouse", "bag"],
]
pairs = Counter()
for items in orders:
    for a, b in combinations(sorted(set(items)), 2):
        pairs[(a, b)] += 1

def bought_with(item, top=2):
    scores = Counter({(b if a == item else a): n
                      for (a, b), n in pairs.items() if item in (a, b)})
    return scores.most_common(top)

print(bought_with("phone"))

import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

# rows = users, columns = items (1 = viewed)
items = ["kurta", "jeans", "sneakers", "saree", "t-shirt"]
views = np.array([
    [1, 0, 0, 1, 0],
    [0, 1, 1, 0, 1],
    [1, 0, 0, 1, 0],
    [0, 1, 1, 0, 0],
])
sim = cosine_similarity(views.T)          # item-item similarity
idx = items.index("jeans")
ranked = sorted(zip(items, sim[idx]), key=lambda x: -x[1])
print([name for name, s in ranked if name != "jeans"][:2])

import numpy as np
import pandas as pd
from sklearn.ensemble import GradientBoostingRegressor

rng = np.random.default_rng(0)
dates = pd.date_range("2025-01-01", periods=400, freq="D")
df = pd.DataFrame({"date": dates})
df["dow"] = df["date"].dt.dayofweek
df["month"] = df["date"].dt.month
df["sale_day"] = (df["date"].dt.day == 10).astype(int)
df["units"] = (50 + 10 * (df["dow"] >= 5) + 80 * df["sale_day"]
               + rng.normal(0, 5, len(df))).round()

train, test = df.iloc[:360], df.iloc[360:]
X = ["dow", "month", "sale_day"]
m = GradientBoostingRegressor(random_state=0).fit(train[X], train["units"])
mae = (m.predict(test[X]) - test["units"]).abs().mean()
print("MAE on last 40 days:", round(mae, 1))

import pandas as pd

# Always compare with a simple baseline:
# predict this week's demand = same day last week
sales = pd.Series([52, 49, 51, 55, 60, 72, 75,
                   53, 50, 52, 56, 61, 70, 78])
baseline = sales.shift(7)
mae = (sales - baseline).abs().dropna().mean()
print("seasonal naive MAE:", round(mae, 2))
# A model is only useful if it beats this.

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity

products = [
    "cotton kurta for women blue",
    "running shoes for men lightweight",
    "blue denim jeans slim fit",
    "women cotton saree printed",
]
vec = TfidfVectorizer()
P = vec.fit_transform(products)

query = "blue cotton kurta"
scores = cosine_similarity(vec.transform([query]), P)[0]
for score, name in sorted(zip(scores, products), reverse=True)[:3]:
    print(round(score, 2), name)

import pandas as pd
from sklearn.linear_model import LogisticRegression

orders = pd.DataFrame({
    "is_fashion": [1, 1, 0, 1, 0, 1, 0, 1],
    "size_differs_from_usual": [1, 0, 0, 1, 0, 1, 0, 0],
    "first_purchase": [1, 0, 1, 1, 0, 0, 0, 1],
    "returned": [1, 0, 0, 1, 0, 1, 0, 0],
})
X, y = orders.drop(columns="returned"), orders["returned"]
m = LogisticRegression().fit(X, y)
new = pd.DataFrame([{"is_fashion": 1, "size_differs_from_usual": 1, "first_purchase": 1}])
print("return risk:", round(m.predict_proba(new)[0][1], 2))
# Use: show a size guide prompt, not a penalty.

1. Bought-together recommender from public order data
   Metric: hit rate on held-out orders vs popular-items baseline

2. Demand forecast for top 20 products
   Metric: MAE vs seasonal naive baseline, time-based split

3. Review analysis for one category
   Output: top complaint themes with examples

# Bought-Together Recommender

## Problem
Suggest relevant add-on items on the product page.

## Data
Public order dataset: orders, items, categories.

## Approach
Co-purchase counts, compared with popular items.

## Results
Hit rate: baseline vs model (held-out orders).

## Limitations and next steps

[ ] Business question written first
[ ] Simple baseline built
[ ] Time-based split for forecasting
[ ] Metric linked to business value
[ ] Charts that a manager can read
[ ] Limitations stated honestly
[ ] Code runs from a clean environment