← Cohere Interview Insights

Cohere·Software Engineer·Hiring Manager Screen·Intermediate

Intermediate
Apr 2026

Summary

45-minute live session with the hiring manager at Cohere for a data engineer role. The whole thing was a walkthrough of actual work, which felt more like a code review than a traditional interview.

Questions Asked (1)

Q1

You're currently filling missing values with 0. Why not just drop those records entirely? And why not leave them as NULL instead of replacing them?

Technical Trade-offsData ModelingRoot Cause Analysis
Author's notes

This tripped me up a bit because it sounds like a cleanup question but it's really asking you to justify a design decision under pressure.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Acknowledge that the choice depends on the data's meaning and the model's requirements, then explain the trade-offs of each option (drop, NULL, zero-fill) in terms of bias, information loss, and downstream compatibility. Finally, justify your current approach by tying it to the specific context and mention alternatives you considered.

Pro tip: Show that you understand missing data mechanisms (MCAR, MAR, MNAR) and that zero-filling can be a deliberate choice for sparse features, but always validate with domain knowledge and consider model sensitivity.

1. Clarify the goal and data context

Start by stating that the right approach depends on why values are missing, what the feature represents, and how the model will use it. Mention that dropping, NULL, and zero-fill each have different implications for bias and performance.

2. Evaluate dropping records

Explain that dropping can be valid if missingness is random and the dataset is large, but it risks losing valuable information and introducing bias if missingness is systematic. Also note that dropping may not be feasible if many features have missing values.

3. Evaluate leaving as NULL

Discuss that NULL preserves the missingness but many models (e.g., linear models, neural networks) cannot handle NULL directly, requiring imputation or special handling. NULL is useful for tree-based models that can treat it as a separate category, but may complicate feature engineering.

4. Evaluate zero-filling

Explain that zero-filling is a simple imputation that works when zero is a meaningful value (e.g., counts, sparse features) and when missingness likely means absence. However, it can distort distributions and introduce bias if zero is not semantically appropriate.

5. Justify your choice and mention alternatives

Conclude by stating why zero-filling was chosen in this case (e.g., domain knowledge, model compatibility, simplicity) and mention that you considered dropping or NULL but rejected them for specific reasons. Optionally, suggest more sophisticated imputation methods if needed.

Key Points to Mention

  • Missing data mechanisms: MCAR, MAR, MNAR and how they affect the choice.
  • Information loss and potential bias from dropping records.
  • Model compatibility: many models cannot handle NULL, requiring imputation.
  • Semantic meaning of zero: zero may represent absence or a true zero, which can be valid for certain features.
  • Impact on downstream tasks: imputation can affect feature distributions and model performance.
  • Alternatives: mean/median imputation, model-based imputation, or using missingness as a feature.

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