A multimodal recommender trained on the Amazon Office catalogue never sees 674 of its 2,420 products. They are dropped before training because their modality data is incomplete, the standard preprocessing step across the field. So the model is scored on a catalogue edited to fit its assumptions, then deployed against one that was not.

Fashion catalogues are never edited that way. They arrive incomplete, stay incomplete, and turn over faster than any cleanup project can close them. The same research group that measured those deletion rates asked at CIKM 2024 whether dropping the items is defensible at all, and found it was not: pre-filtering, they concluded, is “not only unnecessary but also harmful to the performance.” Their own fix is graph-based imputation, filling the gap from an item’s co-purchase neighbours rather than deleting the item. That stops the discarding. It does not train the model to work without the data in the first place, which is the harder and more useful problem this piece is about.

The assumption that breaks is not that every product has a photo. Across six Amazon categories audited in February, images are almost always present and descriptions frequently are not: Office is missing zero images against 674 descriptions. The photograph is the asset someone was paid to shoot, scheduled, and signed off; the description is the field a supplier left blank at six on a Friday. Sports, the largest category examined, repeats the pattern at scale, missing 70 images and 2,701 descriptions. Multimodal fusion assumes vision and text fail independently. In a real catalogue they fail at rates an order of magnitude apart, and nearly always on the same side.

This publication has twice reached the opposite conclusion, and it is worth naming where that was wrong. Covering Tstars-Tryon in May and the outfit-generation datasets two days later, we argued that model quality was solved and the bottleneck had moved to catalogue metadata. The observation holds; the implied remedy does not. “The bottleneck is your catalogue” licenses a cleanup project with a completion date, and the missing-modality literature is the evidence that no such date exists.

A catalogue is not a stock that can be completed; it is a flow that is always incomplete at the front.

The items most likely to be incomplete are the ones the business most needs recommended. A new arrival enters the system with a hero image from the campaign shoot and a description that will be written when someone gets to it, which means the product carrying the most margin and the least behavioural history is also the one the recommender understands worst. Google’s guidance to merchants is blunt about the consequence at the other end of the pipe: a product “will generally have lower data quality if its required attributes are missing from your product data,” which “may lead to reduced performance of your offer in search results.” The gap that suppresses a listing in Shopping also starves the on-site recommender. Retailers treat the first as a feed problem and the second as a model problem. They are the same missing sentence.


Training for absence outperforms reconstructing for presence, and the margin is measurable. I³-MRec, accepted to ACM Multimedia 2025, simulates modality-missing conditions during training instead of assuming them away. MoDiCF, a diffusion method that reconstructs the absent features, holds 98% of its full-modality performance when only a tenth of the catalogue is incomplete. That is the easy case, and the one least likely to describe a live retail feed; as the gap widens, MoDiCF sheds close to a fifth of its recall. I³-MRec, which treats each modality as its own training environment and limits how much it leans on any single one, gives up far less at the same extreme. The method that rebuilds missing data loses more than the method that learns to work without it.

The architectural version of this argument is narrower and more damaging. Most multimodal recommenders use late fusion: encode each modality separately, align the representations, then combine them. EFMRec, presented at WSDM 2026, argues that alignment is the fragile step, because a modality that is absent has no representation to align and the machinery built to reconcile two views has nothing to reconcile. Its alternative projects whatever modalities exist into a shared semantic space and aggregates them early, using what is there and ignoring what is not. No reconstruction, no auxiliary losses, no imputation stage to tune. The decision that determines whether a recommender survives an incomplete catalogue is made before any data arrives.

The strongest objection is that this problem has an expiry date. A vision-language model can read a product photograph and write the missing description, which is what catalogue-enrichment vendors are selling right now; if descriptions can be generated faster than products are onboarded, robustness research is an elaborate workaround for a solvable problem. The condition that would have to hold is that completeness is a state a catalogue reaches rather than a rate it runs at. It is not, and the generated description makes the deeper problem worse. Text derived from an image is not an independent signal; it is the image restated, and a fusion layer built to exploit disagreement between two views of a product now receives one view twice. A catalogue enriched this way scores as complete while quietly halving the information the recommender was designed to use.

If catalogue turnover keeps outrunning catalogue cleanup, and nothing suggests the reverse, the retailers who get value from multimodal recommendation will not be the ones with the cleanest data. They will be the ones who wrote into procurement that the model must be evaluated on an incomplete catalogue. The question is answerable in a sentence: what is your recall at 50% missing text? A vendor trained on complete data will not have the number. A vendor trained for absence will have measured it, because it is the only condition the model was built for. The choice was never between clean data and dirty data; it is between a system that assumes the catalogue will be fixed and one that assumes it will not.