How do I use click and purchase data to rank recommendations?

A common way to rank recommendations with click and purchase data is to treat them as different strengths of feedback and train a model to predict a weighted relevance score.

1) Decide what “good” means

Usually:

• Purchase = strong positive signal
• Click = weaker positive signal
• No interaction = unknown or negative, depending on exposure

A simple ranking target might be:

[ \text{score} = 1 \cdot \text{click} + 5 \cdot \text{purchase} ]

or any weighting that matches your business value.

2) Build training examples

For each user-item impression, create features such as:

• user features
• item features
• context features
• historical interactions

And label the outcome:

• clicked = 1/0
• purchased = 1/0

If both happen, the item gets a stronger label.

3) Use a weighted loss

You can train a ranking model with:

• weighted classification: purchase examples get higher weight
• pairwise ranking: items purchased/clicked rank above non-clicked items
• learning-to-rank methods like LambdaMART, XGBoost ranker, LightGBM ranker

Example weights:

• purchase: 5
• click only: 1
• no click/no purchase: 0 or small negative weight

4) Predict a recommendation score

At serving time, produce a score like:

[ \text{final score} = P(\text{click}) + \alpha \cdot P(\text{purchase}) ]

or train one model to directly estimate expected value:

[ \text{expected value} = P(\text{click}) \times v_{click} + P(\text{purchase}) \times v_{purchase} ]

where (v) is business value.

5) Rank by the score

Sort candidate items by the final score descending.

---

Practical approaches

A. Simple weighted labels

Best if you want something easy:

• label purchase = 5
• click = 1
• no interaction = 0

Train a regression or ranking model.

B. Two-stage model

1. Candidate generation using clicks
2. Re-rank using purchases more heavily

This works well because purchases are rarer and more valuable.

C. Multi-objective ranking

Model click probability and purchase probability separately, then combine: [ \text{rank score} = w_1 P(click) + w_2 P(purchase) ]

---

Important caveat: exposure bias

A missing click does not always mean the item was bad — the user may never have seen it. Ideally, train only on impressions or use logged exposure data.

---

Good starting recipe

If you want a strong baseline:

1. Collect impression-level data
2. Label click and purchase separately
3. Train a ranker with:
   • purchase weight = 5
   • click weight = 1
4. Evaluate with offline ranking metrics:
   • NDCG
   • MAP
   • Recall@K
5. Tune the weights based on online A/B tests

If you want, I can also show:

• a Python example with LightGBM/XGBoost ranking
• a scoring formula for click + purchase data
• or a full pipeline for recommendation ranking.