AWS Certified Machine Learning – SpecialtyModelingHard
A retail company is developing a recommendation system for clothing based on customer purchase history and browsing behavior. They observe that the system frequently recommends popular, generic items and struggles to surface diverse or niche products, even for customers with eclectic tastes. This leads to a lack of personalization and potential 'filter bubbles'. Which type of bias is the system exhibiting, and what is the primary cause?
- AMeasurement bias, caused by inaccuracies in tracking customer browsing behavior.
- BSelection bias, caused by non-random sampling of customer data.
- CPopularity bias, caused by over-representation of frequently purchased items in training data.
- DConfirmation bias, caused by reinforcing existing user preferences rather than exploring new ones.
Show answer & explanationAnswer & explanation
Correct answer: C. Popularity bias, caused by over-representation of frequently purchased items in training data.
Popularity bias occurs when a recommendation system disproportionately recommends popular items due to their higher frequency in the training data. This leads to a 'rich get richer' phenomenon, where popular items become even more popular, and niche items are rarely surfaced, reducing diversity and personalization.
Why the other options are wrong
- A. Measurement bias would relate to errors in data recording, not the algorithmic tendency to favor popular items once data is collected.
- B. Selection bias relates to how data is collected, not directly to the recommendation of popular items over niche ones in a well-collected dataset.
- D. While confirmation bias can be a related issue in recommender systems, the core problem described (generic items, lack of diverse recommendations) is most precisely characterized as popularity bias stemming from data imbalance.
Popularity Bias in Recommenders
The tendency of recommendation systems to disproportionately recommend items that are already popular, thereby reinforcing their popularity and neglecting less popular items.
- Arises from imbalanced training data where popular items are more frequent.
- Leads to lack of diversity and 'filter bubbles' for users.
- Mitigation involves re-ranking, exposure control, or debiasing techniques.
Memory trick: Popularity bias makes the popular more popular.