🤖 AI / ML

Recommendation Diversity

recommendation diversity · diversity · intra-list diversity · ild · serendipity · diversification

Definition

Recommendation diversity is a property of a recommendation list that reflects how unlike each other the suggested products are, usually measured as the mean pairwise distance between their embeddings.

What recommendation diversity is

A recommender optimised for accuracy alone always converges on the same outcome: showing the
shopper whatever most closely resembles what they have already viewed. That is logical, and it is a
problem. The shopper enters a bubble of uniform products that stop being interesting.

Diversity is the antidote. It measures how unlike each other the items in a recommendation list are.

Two ways of measuring it:
– Inter-list diversity — how unlike each other the items inside a single recommendation block are
– Intra-user diversity — how far the recommendations step outside the shopper’s own interaction
history (closely related to novelty)

How diversity is measured

The standard metric is intra-list diversity (ILD):

ILD = mean pairwise distance between items in the list
ILD = Sum dist(i, j) / [n x (n-1) / 2]

Here dist(i, j) is the cosine distance between the item embeddings. ILD runs from 0 (identical
items) to 1 (maximally different).

Adjacent metrics:
– Coverage — the share of unique catalogue positions that appear in recommendations at least
once. Low coverage means concentration on bestsellers and a tail that never works.
– Serendipity — the unexpectedness of a recommendation, a subjective measure of discovery.

The trade-off: accuracy versus variety

This is the classic compromise in recommender systems. There is no correct value for diversity; it
depends on what the block is for:

Block type	Diversity needed	Reasoning
Similar items	Low to moderate	The goal is alternatives, not other categories
You may also like	High	The goal is discovery, widening the horizon
Frequently bought together	Mandatory	The block complements rather than repeats
Trending products	Moderate	It reflects the breadth of audience interest

Tools for steering diversity

Maximal marginal relevance (MMR) is a post-processing algorithm. When adding the next position
to a list, it selects the item that best balances relevance against dissimilarity from the items
already selected.

Merchandising rules are the simple, transparent manual lever:
– No more than 2 items from the same brand in a block
– No more than 3 items from the same category
– Exclude items already viewed

Blending algorithms — combine collaborative filtering (accuracy) with content-based filtering
(diversity) in a fixed proportion.

Tip: do not optimise diversity through an A/B test on CTR alone — clicks are not sales. Use
attributed revenue and repeat visit rate, the metrics that reflect the long-term effect of variety
in recommendations.

Frequently asked questions

Why can recommendations that are too accurate be harmful?

An algorithm optimised purely for relevance drifts toward recommending ever more similar items — the same category, adjacent brands, one price band. The shopper ends up inside a bubble of near-identical suggestions, loses interest and clicks less. Moderate diversity widens the horizon and holds attention.

How is recommendation diversity measured?

The primary metric is intra-list diversity (ILD): the mean pairwise cosine distance between the embeddings of the items in a recommendation list, running from 0 (all identical) to 1 (maximally different). Two companions are useful: coverage, the share of catalogue positions that ever appear in recommendations, and novelty, how new the recommended items are to that shopper.

How do you tune the balance between relevance and diversity?

Through post-processing, that is reranking: maximal marginal relevance (MMR) picks each next position as a compromise between its relevance and its dissimilarity from the items already chosen. The second route is merchandising rules — no more than X items from one category, or no more than X from one brand.

Is high diversity always desirable?

No, it depends on context. A similar items block on a product page needs moderate diversity, since completely different categories would be wrong. A you may also like block on the home page needs more, because the goal is discovery. A frequently bought together block requires it, since the point is complementing rather than repeating the item. Tune diversity per block, not globally.

Gravity Field