Katkat
Back

Predicting the Next Elevator Call Location

CoupangData Scientist · Technical Screen
MediumData ScientistTechnical ScreenMachine Learning

Last updated

This question tests the ability to design a spatio-temporal forecasting system to predict future elevator call locations, focusing on feature engineering, model selection, and handling data sparsity.

Problem Statement

You are working for a logistics firm that manages large-scale smart warehouses. Each warehouse is equipped with a network of autonomous elevators. To reduce wait times, your team wants to implement a predictive model that triggers elevators to move to specific floors before a call button is pressed.

You have access to a historical dataset calls with the following schema: [timestamp, warehouse_id, floor_level, user_type].

Your Tasks:

  1. Problem Framing: How would you define the target variable for a machine learning model to predict the next elevator call location? Would you treat this as a classification or regression problem?
  2. Feature Engineering: Propose at least four features that capture the temporal and spatial patterns of elevator usage. How would you handle the 'cold-start' problem for a newly opened floor or a new warehouse?
  3. Model Selection & Evaluation: Recommend a model architecture suitable for this task. Define an appropriate offline metric and explain how you would bridge the gap between offline performance and the business goal of 'reduced customer wait time'.

Constraints & Considerations:

  • Consider that the warehouse environment is highly dynamic (e.g., peak lunch hours vs. late-night operations).
  • The prediction horizon should be short (e.g., next 5-10 minutes).
  • Assume the model will run on a rolling basis every minute.

Hints

Hint 1

Consider discretizing time into small intervals and treating the problem as a multi-class classification task where the classes are the floors.

Hint 2

Think about cyclical features for time, such as sine and cosine transforms for hours of the day and days of the week.

Hint 3

For evaluation, distinguish between the model's predictive accuracy and the actual reduction in latency, which may require a simulation environment.

Follow-up questions

  • How would you handle concept drift, for example, if a company shifts its working hours or adds new departments to specific floors?
  • If the computational budget for inference is extremely limited, how would you optimize the model serving strategy?
  • How would you incorporate 'external' signals, such as scheduled warehouse shift changes or planned maintenance, into your model?