How do I recommend content based on dwell time and clicks?

A simple way to recommend content using dwell time and clicks is to treat them as signals of user interest, combine them into an engagement score, and then rank content by that score.

1) Define your signals

• Click: user selected the content.
• Dwell time: how long the user stayed on the content after clicking.

Usually:

• Click = strong positive signal
• Longer dwell time = stronger positive signal
• Very short dwell time = weak/negative signal

2) Convert them into a single score

A basic formula:

[ \text{score} = \alpha \cdot \text{click_signal} + \beta \cdot \text{dwell_signal} ]

Where:

• click_signal = 1 if clicked, else 0
• dwell_signal = normalized dwell time
• α and β are weights you choose

Example:

score = 1.0 * click + 0.5 * normalized_dwell

3) Normalize dwell time

Raw dwell time is usually skewed, so normalize it:

• Use log transform: log(1 + dwell_time)
• Or scale it between 0 and 1 using min-max normalization
• Or bucket it:
  • < 5 sec = 0
  • 5–30 sec = 0.3
  • 30–120 sec = 0.7

  • 120 sec = 1.0

4) Use the score to rank content

For each user, compute the score for each candidate content item and recommend the highest scoring items.

5) Example

Suppose:

• Content A: clicked, dwell = 90 sec
• Content B: clicked, dwell = 10 sec
• Content C: not clicked

If you use:

score = click + 0.01 * dwell

Then:

• A = 1 + 0.9 = 1.9
• B = 1 + 0.1 = 1.1
• C = 0

So A ranks highest.

6) Better approach: use feedback labels

You can define interaction quality:

• High engagement: clicked + long dwell
• Medium engagement: clicked + short dwell
• Low engagement: no click or immediate bounce

Then use these as training labels for a learning-to-rank or classification model.

7) Practical recommendation pipeline

1. Collect click and dwell time logs
2. Clean data:
   • remove bot traffic
   • handle outliers
   • ignore tab-switching if possible
3. Create engagement features:
   • click
   • dwell time
   • scroll depth
   • repeat visits
4. Compute engagement score or train a model
5. Rank candidate items
6. Evaluate with offline metrics:
   • CTR
   • average dwell time
   • conversion
   • NDCG / MAP

8) Important caveats

• Dwell time alone can be misleading:
  • long dwell may mean interest, but also confusion
• Clicks are biased by position:
  • items shown higher get more clicks
• You may need to normalize by content type:
  • a 30-second dwell on a news article means something different than on a video

9) Recommended starting formula

A good first version:

engagement_score = 0.7 * click + 0.3 * log(1 + dwell_time)

Then tune weights based on actual user behavior.

If you want, I can also show:

• a Python example
• a SQL implementation
• or a machine learning ranking approach using these signals.