Cold start, popularity bias and feedback loops

A recommender learns from what users did after seeing its recommendations. That creates three problems. New users and items have no history to learn from. Popular items gather so much data that they crowd out the rest. And the model's own output becomes its next training set. Each problem has a practical fix, and none of them is solved by a better architecture alone.

Cold start

Cold-start users have little or no history. Useful fixes:

• Ask for tastes during onboarding.
• Use coarse priors, such as location, device or language.
• Lean on content features and popularity until history builds up.
• Explore more in the first few sessions.

Cold-start items have no interactions yet, such as a new video or a new product. Useful fixes:

• Build content embeddings from the title, thumbnail, audio or description.
• Boost new items in retrieval, so they get exposure.
• Train the item tower on content features. A new item then has a usable vector on day one.

Popularity bias

Collaborative filtering and CTR models both favor popular items. The more an item was shown, the more data it has, and the more confidently the model ranks it. Without a correction, the long tail gets starved.

The corrections are inverse-propensity weighting in training, popularity-normalized labels, exploration traffic, and diversity limits in re-ranking.

Feedback loops

The loop runs like this:

1. The model recommends item A more than item B.
2. A gets more impressions, so it collects more clicks.
3. Those clicks become training data.
4. The model sees more evidence for A and ranks it higher still.

Suppose A and B are equally good. A now looks better in the data only because it was shown more. Left alone, this narrows the catalog into a filter bubble.

Countermeasures:

• Log the position of each item and whether it was shown at all.
• Train with propensity weights so exposure is not mistaken for preference.
• Keep a holdout of exploration traffic that ignores the model.
• Monitor inventory diversity and creator concentration, not only click metrics.

Common mistakes

• Assuming more data fixes cold start. Cold items have no data to add. You need content features or exploration.
• Reading popularity as quality. A popular item may simply have had more exposure.
• Measuring only clicks. A loop can raise click rate while narrowing what users see.

Check yourself

1. Why does a new item need content features on day one? Answer: it has no interactions, so its vector can only come from its content.
2. How does a feedback loop make equal items look unequal? Answer: the item shown more gathers more clicks and training data, so the model ranks it higher.

Key takeaways

• Cold start needs content features, onboarding and exploration.
• Popularity bias starves the long tail. Correct it in training and in re-ranking.
• Feedback loops turn exposure into apparent quality. Log positions and keep exploration traffic.
• Monitor diversity and concentration alongside clicks.

Further learning

Next, see how one large video platform combined these ideas in its 2016 recommender in YouTube's 2016 recommender.