Pinterest·Machine Learning Engineer·Technical Phone Screen·Senior

Senior

Jun 2026

Summary

Pinterest ML engineer interview with a classic scale question that sounds straightforward but has a few layers worth thinking through.

Questions Asked (1)

Given a dataset with a million data points, would you choose a deep neural network or a KNN classifier, and why?

Technical Trade-offsAlgorithms & Data Structures

Author's notes

KNN at a million points is basically a no-go unless you have serious infrastructure for it.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the problem context (data characteristics, latency, accuracy, and scalability needs) before choosing a model. Then compare KNN and deep neural networks on computational complexity, memory, and performance for a million-point dataset, and justify your choice with trade-offs. Conclude with a recommendation that aligns with Pinterest's production constraints.

Pro tip: Mention that KNN's inference cost scales linearly with dataset size, making it impractical for low-latency serving at Pinterest scale, while DNNs offer constant inference time after training. This shows you think about production deployment, not just accuracy.

1. Clarify the problem and constraints

Ask about data dimensionality, feature types, latency requirements, and whether the task is online or batch. This ensures your answer is tailored to the actual use case.

2. Analyze computational complexity

Compare training and inference costs: KNN requires O(N) per query and O(N^2) for pairwise distances, while DNNs have fixed inference cost after training. Highlight that a million points makes KNN slow and memory-intensive.

3. Evaluate accuracy and scalability

Discuss how DNNs can learn complex patterns and scale with data, whereas KNN's accuracy may degrade in high dimensions (curse of dimensionality) and cannot easily leverage distributed training.

4. Consider practical deployment factors

Mention memory footprint, model updates, and hardware acceleration. DNNs can be optimized with GPUs/TPUs and compressed, while KNN stores the entire dataset and is hard to update incrementally.

5. Make a recommendation with trade-offs

Choose DNN for large-scale, low-latency, high-dimensional tasks, but acknowledge KNN's simplicity for small, low-dimensional, or baseline scenarios. Tie back to Pinterest's need for scalable, real-time recommendations.

Key Points to Mention

KNN inference time grows linearly with dataset size, making it impractical for a million points in production.
Deep neural networks offer constant inference time after training and can be optimized for latency.
Curse of dimensionality: KNN performance degrades in high-dimensional spaces, common in image or text data.
DNNs can leverage distributed training and hardware accelerators, while KNN is memory-bound and hard to scale.
Pinterest's scale requires models that can handle real-time serving and frequent updates, favoring DNNs.
Trade-offs: KNN is simple, interpretable, and effective for small, low-dimensional datasets; DNNs require more data and tuning but scale better.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.