Long-Tail Items

Long-tail items are the many items in a catalog that receive few interactions, in contrast to the small set of popular items that receive most of them. Abdollahpouri et al. (2019) state that rating data are generally skewed toward popular items: a few items have thousands of ratings, and the majority of items combined have fewer.

Where the tail begins

There is no universal cutoff. Abdollahpouri et al. treat an item as popular if its popularity falls within the top 20% of item popularities and regard the rest as non-popular. Other analyses split by a share of interactions or by a fixed rank. The split should be stated and kept the same across compared models.

A hypothetical example: ten items have 500, 300, 120, 40, 15, 10, 8, 4, 2, and 1 interactions, 1,000 in total.

The top two items, 20% of the catalog, hold 80% of the interactions. Under the 20% rule they are the head and the other eight items are the tail, which together hold 20% of the interactions.

Why it matters for recommenders

A model trained on interactions can learn to favor what is already popular. Abdollahpouri et al. describe popularity bias as popular items being recommended too frequently while the majority of other items do not get the attention they deserve, and report that recommendations can be extremely concentrated on popular items even for users interested in non-popular ones.

Measuring exposure of the tail

• The share of recommendation slots filled by head items or by tail items.
• Item coverage and the concentration of recommendation counts.
• Popularity-based novelty metrics. Vargas and Castells (2011) describe their popularity-based novelty metrics as measuring the ability of a system to recommend relevant long-tail items.

Cautions

• Tail items are not automatically good ones. Tail exposure only helps if the items are relevant, so it should be read together with an accuracy metric.
• Tail items have little data, so metrics computed on them are noisy.
• A new item may initially have few interactions because it has had little exposure. This is a recommendation cold-start problem: there is not yet enough interaction history to estimate its appeal. Some items gain interactions later, while others remain in the tail.

References:

• Abdollahpouri, Mansoury, Burke, and Mobasher, The Unfairness of Popularity Bias in Recommendation (RMSE workshop, RecSys 2019), arXiv:1907.13286
• Vargas and Castells, Rank and Relevance in Novelty and Diversity Metrics for Recommender Systems (RecSys 2011)