← Ancestry Interview Insights

Ancestry·Software Engineer·Onsite - System Design / Architecture·Senior

Senior
Apr 2026

Summary

Ancestry ML-adjacent system design round, single question but it sprawled into a lot of territory fast. The feature engineering angle was expected but the depth they wanted on encoding choices and evaluation caught me a bit off guard.

Questions Asked (1)

Q1

How would you define features for finding similar users or similar products? Walk through the feature families you'd consider, how you'd encode them, and how you'd evaluate whether your similarity approach actually works.

System DesignTechnical Trade-offsA/B Testing & Experimentation
Author's notes

I started with behavioral features because that felt safest for a genealogy product (search history, tree edits, record views) and then worked outward to demographic and content signals.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business goal and data available, then outline feature families (demographic, behavioral, content-based, graph-based) and encoding methods, and finally describe offline and online evaluation strategies. Emphasize trade-offs between feature complexity, scalability, and interpretability, and how you'd iterate using A/B tests.

Pro tip: Anchor your answer in Ancestry's domain: genealogy data is inherently graph-structured (family trees), so graph-based features like shared ancestors or co-occurrence in records are powerful and differentiate you from generic answers.

1. Clarify the problem and data

Ask about the definition of 'similar' (e.g., similar users by behavior or similar products by attributes), the available data sources, and the scale. This ensures your feature design aligns with business objectives.

2. Identify feature families

List categories such as demographic (age, location), behavioral (purchase history, clicks), content-based (product descriptions, user profiles), and graph-based (social connections, co-occurrence). Explain why each is relevant.

3. Encode features for similarity

Describe encoding methods: one-hot, embeddings (e.g., word2vec for products, node2vec for graphs), normalization, and dimensionality reduction. Discuss how to handle categorical, numerical, and text data.

4. Choose similarity metrics and model

Select metrics like cosine similarity, Jaccard, or Euclidean distance based on feature types. Mention techniques like matrix factorization, nearest neighbors, or clustering, and trade-offs (e.g., scalability vs. accuracy).

5. Evaluate and iterate

Define offline metrics (precision@k, recall, NDCG) using labeled data or proxy labels. Then design online A/B tests measuring engagement (CTR, conversion) and guardrail metrics. Iterate based on results.

Key Points to Mention

  • Graph-based features leveraging family trees and record co-occurrence (unique to Ancestry)
  • Embedding techniques for high-cardinality categorical data (e.g., user IDs, product IDs)
  • Handling sparsity and scalability with approximate nearest neighbors (e.g., FAISS, Annoy)
  • Offline evaluation with precision@k and recall, and online A/B testing with engagement metrics
  • Trade-offs between interpretability (e.g., rule-based) and performance (e.g., deep learning)
  • Cold-start problem and how to incorporate content-based features for new users/products

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