Product Recommendations in the Bloomreach Engagement platform have got performance and algorithm/model limitations of which you need to be aware. Here is the list of restrictions:

1. Max 500 products/items are returned in a metric-based engine in a single request.
2. Max 500 products/items are returned in a metric-based category engine in a single request with one category in categoryNames attribute.
3. Catalog filter supports neither regex filters, nor list attributes and filters.
4. Catalog filter serves as a post-filtering. Meaning that the recommendation model might return 10 suitable items, but if you specified a very strict catalog filter, it can filter out some of those recommended products.
5. In Filter-based template, the dynamic catalog filter from the get-recommendation request is ignored.
6. In Advanced template request latency stack up with more recommendation engines added in a single advanced template instance. Please consider it when designing a recommendation engine.
7. In all AI models, which are using events for training, there's a minimum threshold of 2 events, which are considered for training per customer. e.g. customer with 1 purchase_item during the last 90 days will not be considered, a customer with 1 purchase_item and 1 view_item, will be considered for model training. This mechanism is introduced to mitigate model skewness.
8. Max 1000 products could be requested in the parameter size for AI models.
9. Textual similarity is trained only on a model with less than 1 million rows.

Challenges & Trade-offs

While developing/using recommendations, one deals with several challenges:

• Exploitation/Exploration: We need to recommend items the user already interacted with (exploitation), but we also need to show something new (exploration).
• Popularity bias: Most users view very few items and most items are seen by a minority of users. This creates huge “data gaps” and we need to cover them.
• Cold-start problem: When a fresh client starts to use Bloomreach Engagement, we are usually not provided with enough historical user-item interactions. Thus our personalized RS cannot perform well. The more data we collect, the better the performance will be observed. The same idea applies when a new product is added to the catalog.