Description
Finds items customers tend either to love or hate by computing entropy over their feedback ratings, then uses responses to those items to learn individual preferences.
Some items split opinion: most people rate them very high or very low. Because a customer's reaction to such an item reveals a lot about their taste, these polarising items are valuable probes. The method computes an entropy-based measure over each item's rating distribution after the fact to flag polarising items, then deliberately includes them in recommendations to learn about each customer, accepting a short-term drop in purchase likelihood for better long-term fit. The authors liken it to a batter who makes a pitcher throw many pitches to learn what they have. It is not an A/B test but a learning procedure, justified only where the key metric rewards long-term relationship rather than single-order conversion. The source gives no formula or thresholds, so the entropy measure and cut-offs must be designed and validated locally.
Attributes
- Discipline
- Product Analytics & Strategy
- Loss type
- Latent
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Weeks
Steps
- Collect explicit feedback ratings per item across customers.
- Compute an entropy-based dispersion measure over each item's rating distribution, checking also for bimodality so that uniformly mixed ratings are not mistaken for polarisation.
- Flag items above a locally chosen threshold as polarising and review them qualitatively.
- Include polarising items deliberately for customers whose preferences are uncertain.
- Feed the responses into preference profiles and recommendation models.
- Monitor the short-term conversion cost against long-term retention or satisfaction to confirm the trade-off pays.
Sources
- Designing with Data (King et al., 2017)
Also known as
- Polarising style analysis
- Love-or-hate item analysis
Categories
- Analytics & Behaviour
- Product Analytics & Strategy