How Data Science Is Used in Indian Telecom Companies Today

import pandas as pd

# usage.csv: subscriber_id, month, data_gb, call_mins, recharge_amt, complaints
u = pd.read_csv("usage.csv").sort_values(["subscriber_id", "month"])

g = u.groupby("subscriber_id")
features = pd.DataFrame({
    "avg_data_gb": g["data_gb"].mean(),
    "last_vs_avg_data": g["data_gb"].last() / g["data_gb"].mean().clip(lower=0.1),
    "avg_recharge": g["recharge_amt"].mean(),
    "complaints_3m": g["complaints"].apply(lambda s: s.tail(3).sum()),
    "tenure_months": g["month"].count(),
})
# A sharp drop in last month's usage vs average is often a churn signal
print(features.sort_values("last_vs_avg_data").head())

import pandas as pd
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import KMeans

# per-subscriber averages
profile = pd.read_csv("subscriber_profile.csv")   # data_gb, call_mins, sms
X = StandardScaler().fit_transform(profile[["data_gb", "call_mins", "sms"]])
profile["segment"] = KMeans(n_clusters=4, n_init=10, random_state=0).fit_predict(X)

print(profile.groupby("segment")[["data_gb", "call_mins", "sms"]].mean().round(1))
# Name segments, e.g. 'heavy data', 'voice-first', 'light users'
# and match each to a suitable plan.

import pandas as pd

# hourly.csv: tower_id, hour, drop_rate
h = pd.read_csv("hourly.csv", parse_dates=["hour"]).sort_values(["tower_id", "hour"])

# Compare each hour with that tower's recent normal behaviour
roll = h.groupby("tower_id")["drop_rate"]
h["mean_24h"] = roll.transform(lambda s: s.shift(1).rolling(24).mean())
h["std_24h"] = roll.transform(lambda s: s.shift(1).rolling(24).std())
h["z"] = (h["drop_rate"] - h["mean_24h"]) / h["std_24h"]

alerts = h[h["z"] > 3]
print(alerts[["tower_id", "hour", "drop_rate", "z"]].head())

import pandas as pd
from statsmodels.tsa.holtwinters import ExponentialSmoothing

# daily traffic for one tower: date, traffic_gb
ts = (pd.read_csv("tower_101_daily.csv", parse_dates=["date"])
        .set_index("date")["traffic_gb"].asfreq("D").interpolate())

model = ExponentialSmoothing(ts, trend="add", seasonal="add",
                             seasonal_periods=7).fit()
forecast = model.forecast(30)
print(forecast.tail())
# Planning teams compare forecasts with tower capacity limits.

import pandas as pd

# cdr.csv (call detail records): msisdn, calls_out, unique_callees,
#                                  avg_call_secs, calls_in
cdr = pd.read_csv("cdr_daily.csv")

# SIM-box style pattern: many outgoing calls to many numbers,
# very few incoming calls, short-to-medium durations
flags = cdr[
    (cdr["calls_out"] > cdr["calls_out"].quantile(0.99)) &
    (cdr["unique_callees"] > cdr["unique_callees"].quantile(0.99)) &
    (cdr["calls_in"] < 5)
]
print(len(flags), "numbers flagged for review")
# Flags go to an investigation team - never auto-block on one rule.

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import make_pipeline

texts = ["no network in my area since morning",
         "recharge done but data pack not activated",
         "calls dropping every few minutes",
         "charged twice for the same recharge"]
labels = ["network", "recharge", "network", "billing"]

clf = make_pipeline(TfidfVectorizer(), LogisticRegression(max_iter=1000))
clf.fit(texts, labels)
print(clf.predict(["data not working after recharge"]))
# Real systems train on many labelled complaints in several languages.

Project: Predicting subscriber churn

Data: public telecom churn dataset (credit the source)
Question: Which subscribers are likely to leave next month?
Decision supported: Who receives a retention call or offer

Steps:
1. EDA: usage, tenure, complaints, plan type
2. Baseline: logistic regression
3. Better model: gradient boosting
4. Choose threshold based on retention budget
5. Explain top drivers in plain language
6. One-page recommendation

Choosing a churn metric

If retention calls are expensive:
  -> favour precision (call fewer, more certain churners)

If losing a customer is very costly:
  -> favour recall (catch more churners, accept some extra calls)

Always report:
- Precision and recall at the chosen threshold
- ROC-AUC or PR-AUC
- Expected number of calls per month at that threshold

Telecom data handling checklist

[ ] Use public or synthetic data for portfolio work
[ ] Remove or mask phone numbers and personal details
[ ] Split by time (train on earlier months)
[ ] Check class imbalance in churn labels
[ ] Avoid features recorded after the churn event
[ ] Document assumptions and limitations