← Grab Interview Insights

Grab·Machine Learning Engineer·Technical Phone Screen·Senior

Senior
Jun 2026

Summary

Grab ML engineer round where they drop a Kaggle link on you and expect you to build a POI deduplication model from scratch. Less about writing code yourself, more about whether you actually understand the problem well enough to direct an AI agent toward something sensible.

Questions Asked (1)

Q1

Given pairs of POI records, how would you determine whether two records refer to the same real-world location? Walk through your full ML approach including feature engineering, model choice, and evaluation.

Technical Trade-offsData ModelingAlgorithms & Data Structures
Author's notes

The setup is a Kaggle-style problem so it feels approachable at first, but the feature engineering part is where they actually probe you.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Frame the problem as a binary classification task: given a pair of POI records, predict whether they refer to the same real-world location. Walk through the end-to-end ML pipeline: feature engineering from raw attributes, model selection with trade-offs, and evaluation using appropriate metrics and business impact.

Pro tip: Emphasize that in production, precision often matters more than recall for deduplication to avoid merging distinct locations, and discuss how to handle class imbalance and scalability to millions of pairs.

1. Problem Definition and Data Understanding

Clarify the goal: determine if two POI records refer to the same physical place. Discuss data sources (name, address, coordinates, phone, category) and challenges like noise, missing values, and class imbalance.

2. Feature Engineering

Create features that capture similarity and agreement between record pairs: string similarity (Jaro-Winkler, Levenshtein) for names and addresses, geospatial distance (Haversine), phone number match, category match, and possibly embedding-based semantic similarity.

3. Model Selection and Training

Choose a model suitable for pairwise classification: gradient boosted trees (XGBoost/LightGBM) for tabular features, or a Siamese neural network for learned embeddings. Discuss handling class imbalance via sampling or class weights.

4. Evaluation and Threshold Tuning

Use precision, recall, F1, and AUC-ROC, but prioritize precision if false merges are costly. Tune the decision threshold based on business requirements and validate with cross-validation or a held-out set.

5. Deployment and Scalability

Address how to scale to millions of pairs: blocking/indexing to reduce candidate pairs, distributed training, and online serving with low latency. Monitor performance and retrain periodically.

Key Points to Mention

  • Feature engineering: string similarity metrics (Jaro-Winkler, Levenshtein), geospatial distance (Haversine), exact matches on phone/category, and semantic embeddings for names.
  • Model choice: gradient boosted trees (XGBoost/LightGBM) for tabular features, or Siamese networks for learned representations; trade-offs between interpretability and performance.
  • Class imbalance: use of techniques like negative sampling, class weighting, or focal loss to handle the rarity of true matches.
  • Evaluation metrics: precision, recall, F1, AUC-ROC; emphasize precision if false positives (incorrect merges) are costly.
  • Scalability: blocking/indexing (e.g., geohash, phonetic blocking) to avoid O(n^2) comparisons, and distributed processing (Spark) for large-scale training.
  • Business impact: aligning model threshold with operational goals, such as reducing duplicate listings or improving search relevance.

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