In implicit feedback data, why is an unobserved user-item cell not a negative example?

## The shape of the problem In implicit feedback you have a huge user-item matrix in which a tiny fraction of cells carry an observed interaction and everything else is blank. In a catalog of a million tracks, a heavy listener might touch a few thousand. So more than 99.9% of every user's row is blank, and how you interpret those blanks decides what the model learns. ## Why blank is not negative A blank cell is the union of at least three distinct realities: 1. **Never exposed.** The system never put the track in front of the listener. It carries zero information about taste — the listener had no opportunity to express anything. 2. **Exposed and ignored.** The track was surfaced 40 times and never opened. This *is* evidence of disinterest, though still soft: wrong moment, wrong mood, bad artwork. 3. **Known and deliberately avoided.** The listener recognises it and does not want it. This is a genuine negative, and it is invisible in the log. All three land in the same blank cell. Calling the union "negative" asserts a dislike for millions of items the person has never heard of. Since the catalog is enormous and exposure is tiny, that assertion is wrong for the overwhelming majority of cells. ## Why you cannot just drop them either The tempting alternative is to train only on observed interactions. That fails immediately. With positives and nothing else, the loss is minimised by predicting the maximum score for every user-item pair. The model has no reason to rank anything below anything else, because it has never been shown a pair where the answer is "lower". Ranking is inherently comparative: you need something to rank *against*. This is why implicit feedback is called a **one-class problem** — the negatives are not merely rare, they are structurally absent, and the training procedure has to manufacture them. ## The two standard treatments **Whole-matrix weighting.** Keep every cell. Observed cells get preference 1 with a confidence that grows with the interaction count. Unobserved cells get preference 0 with a small baseline confidence — the model is nudged toward zero everywhere it has no evidence, but only weakly, so a single positive elsewhere can easily outweigh it. This treats the blank as "probably not, but we barely know", which is the honest reading. The cost is that the matrix is enormous, so the fitting procedure has to exploit the structure of the all-zeros background rather than materialise it. **Negative sampling.** For each observed interaction, draw a handful of unobserved items and treat them as negatives for this training step, resampling as training proceeds. This keeps each update cheap and scales to very large catalogs. Because the drawn items are unobserved rather than known dislikes, some fraction of them are false negatives — items the listener would actually love — and the sampling design determines how damaging that is. Both encode the same belief: unobserved means *probably not interacted with for a reason we do not know*, and the strength of that belief must be low. ## Using exposure to disambiguate If the product logs impressions — what was actually rendered on screen — you can split the blanks. Shown-and-not-opened is a much stronger negative than never-shown, and can be trained with higher weight. Never-shown items remain genuinely uninformative and are the safest pool to sample from when you need cheap negatives. Most teams that have impression logs use them exactly this way, and the ones that do not are stuck treating the entire unobserved region uniformly. ## Consequences for evaluation The same trap appears at evaluation time. If you score a model by how well it avoids recommending unobserved items, you reward it for suppressing everything the system has never surfaced, which is a self-fulfilling loop. Held-out evaluation should ask whether the items a user interacted with in a **later** time window are ranked highly, not whether unobserved items are ranked low. ## How to say it in an interview One sentence: unobserved is *missing data with an unknown reason*, and the two failure modes are symmetric — labelling it negative asserts dislikes that were never expressed, and discarding it removes the only thing that makes ranking well-posed. The craft is in choosing how much weight the blanks carry and, where impression data exists, in splitting them by whether the user ever had the chance to say no.