Implicit feedback as labels.

A buyer types "air max 90" into Loftmarket and gets 20 listings. She taps one of them (L4, fourth from the top), reads it for 40 seconds, messages its seller about the sole, and pays for it two days later. Nobody asked her to rate anything, yet the logs now hold five different things a training pipeline could call a positive.

Each one answers its own question. A click says the photo and price looked worth a look. A long stay on the page says the listing held her attention. A message or an offer says she was seriously considering it. A purchase says it was the right item. A purchase she did not return within 14 days of the order adds that the item matched its description. Walking down that list, the signal gets closer to what Loftmarket actually wants, and it also gets rarer and later.

Explicit feedback exists too: star ratings after delivery, a "not relevant" button. It is unambiguous but tiny next to the logs. Covington and colleagues made the same call for YouTube's recommender: thumbs and surveys existed, but they trained on watches because the implicit history was orders of magnitude larger and reached videos deep in the tail where explicit ratings barely exist.

Joachims and colleagues tracked searchers' eyes. The first two results were looked at about equally often, yet the first drew far more clicks. To separate relevance from rank they had judges decide which of the top two abstracts was better. When the better one sat at rank 1, 19 of the 20 searches with a single click on the top two went to it. When the better one sat at rank 2, only 2 of 7 did; the other 5 still went to rank 1. People trust the order they are given. The same study found that viewing drops with every rank down, and sharply at the edge of the screen.

A simple model folds both effects into one factor per rank: the click rate at rank k ≈ position factor(k) × P(click | relevance). The factor mixes how often rank k is looked at with how much buyers trust whatever sits there. The classic examination model assumes the click, once a result is seen, depends on relevance alone; trust bias breaks that, since ranks 1 and 2 were seen equally but clicked unequally (a WWW 2019 paper by Agarwal and colleagues models trust as its own term). The randomized bucket below measures the combined factor, which is what the correction needs. That factor, the propensity, belongs to the page layout, not to the listing. Inverse propensity scoring (Joachims, Swaminathan and Schnabel, 2017) weights each click in the loss by 1 ÷ propensity of the rank where it happened, so a click earned deep in the page counts more than one handed to rank 1. In expectation this removes the position effect, provided the propensities are right and every rank has some chance of being seen.

Every row in the training table is a listing the current ranker chose to show. A good listing it always buried on page 3 never gets a click, so it never gets a positive. Position weighting cannot rescue it: its propensity at rank 60 is effectively zero, and 1 ÷ 0 is not a weight. How that gap compounds as each model trains on the last one's choices is the subject of feedback loops, and spending a little traffic on purpose to show unproven listings is exploration slots. Here the question is narrower: where else can labels come from?

Two remedies come from practice. Rules of ML #34 advises that when a model filters things out, you hold back a small share of traffic, around 1%, from the filter so the filtered items still earn honest labels. And Covington and colleagues built YouTube's training examples from all watches, including videos played embedded on other sites, not only from its own recommendations, so new videos had a way in. Loftmarket's equivalent is to also learn from listings reached through category browsing, saved searches and shared links, where the ranker did not decide what the buyer saw.

Engagement is not permission. Buyers happily tap a counterfeit bag priced at a tenth of retail, and a scam listing can collect messages all day. No weighting scheme turns those clicks into the answer to "should this be allowed?"; that question needs people reading a policy, covered in human labeling. When you have no labels at all yet, rules and heuristics can bootstrap them, as in weak supervision. And every source, implicit ones included, leaves some rows simply wrong; measuring and cleaning those is label noise. When the same logged signal is used to judge a launch rather than to train, the concerns shift to proxy metrics.