Board Classification - ML Model to Classify Pinterest Board Topics

Question

## Round 1 - ML Design

## Problem

You are building a classifier to automatically assign topic labels to Pinterest boards. A board contains a list of pins, each with an image URL, title text, and description text. Design an end-to-end ML system for this task.

**Part A:** Feature Engineering
- What features would you extract from pins to represent a board?
- How do you aggregate pin-level features to board-level?

**Part B:** Model Choice
- Would you use text features, image features, or both?
- Compare a simple TF-IDF + logistic regression baseline vs. a fine-tuned BERT or CLIP approach.

**Part C:** Evaluation
- How would you split training data given that boards can span multiple topics?
- What metrics matter most: precision, recall, or F1?

```
Example input:
board = {
  "pins": [
    {"title": "Sourdough recipe", "desc": "Easy starter guide"},
    {"title": "Cast iron skillet tips", "desc": "Seasoning and care"}
  ]
}
predicted_label -> "Food & Cooking"
```

## Follow-ups

1. How do you handle boards with mixed topics (e.g. Travel + Food)?
2. How would you build a training set without manual labeling?
3. What signals would you use to detect when the model needs retraining?
4. How do you serve multi-label predictions at Pinterest scale?

Board Classification - ML Model to Classify Pinterest Board Topics

Question Details

Full Details

Topics

More from Pinterest