Shell, Unilever & Tesco: Data Roles (Bengaluru)

• Gradient-boosted decision trees: add trees sequentially, each fitting the gradient (and Hessian) of the loss with respect to the current predictions, using a second-order Taylor approximation.
• Regularised objective: loss + γ·(number of leaves) + ½λ‖leaf weights‖², giving the optimal leaf weight −G/(H + λ) and a split gain formula.
• Engineering: histogram/approximate split finding, sparsity-aware missing-value handling (a learned default direction), column and row subsampling, shrinkage (η), parallel tree building, early stopping.
• Key hyperparameters: max_depth, eta, n_estimators, subsample, colsample_bytree, min_child_weight, lambda/alpha, scale_pos_weight.

SELECT store_id, SUM(sales) AS total
FROM transactions
WHERE txn_date >= '2025-01-01'          -- filters rows before grouping
GROUP BY store_id
HAVING SUM(sales) > 1000000;            -- filters groups after aggregation

• WHERE cannot use aggregates; HAVING can. Put row filters in WHERE for performance.

from abc import ABC, abstractmethod
class Sensor(ABC):                         # abstraction
    def __init__(self, sid): self._sid = sid          # encapsulation (protected attribute)
    @abstractmethod
    def read(self) -> float: ...
class PressureSensor(Sensor):              # inheritance
    def read(self): return 101.3
class TempSensor(Sensor):
    def read(self): return 36.6
for s in (PressureSensor("p1"), TempSensor("t1")):   # polymorphism
    print(type(s).__name__, s.read())

• Decompose: distribution (number of stores stocking it) × rate of sale × price; channel (modern trade, general trade, e-commerce).
• Check: competitor launches or promotions, price increases, out-of-stocks at distributors, seasonality, ad spend changes, a pack-size change.
• Data: secondary sales (distributor → retailer), retail audit data, e-commerce share, promotion calendars.
• Recommend: fix the distribution gaps, targeted promotions, pricing adjustments; track weekly.

Answer (STAR):

• "Two stakeholders wanted conflicting definitions of 'active customer' for the same dashboard."
• "I documented both definitions, showed the numerical impact of each, and facilitated a 30-minute alignment meeting. We agreed on one primary metric plus one secondary metric."
• "The dashboard launched on time, and the definition became a company standard."

import pandas as pd
df = pd.read_csv("sales.csv", parse_dates=["date"])
weekly = (df.set_index("date").groupby("store_id")["sales"].resample("W").sum().reset_index())
weekly["wow_pct"] = weekly.groupby("store_id")["sales"].pct_change() * 100
print(weekly.head())