Single-Signal Recommenders Feel Dumb: Prune, Don't Rank

Every recommender system begins as a single filter, but many fail to evolve past that stage. Shipping features like colour matching initially looks impressive, but users quickly realize the system feels unintelligent. The author, affiliated with Clad9, investigated this issue and discovered that structural problems, rather than modelling issues, are the root cause.

The domain the author is discussing is wardrobe apps that help users choose outfits based on their clothing. These apps often struggle to gain traction due to two main failures: ingestion cost and reliance on single-signal recommendations.

The first failure point is the "ingestion cliff." Before a recommender can generate suggestions, it needs an inventory of items. Traditionally, this involved taking individual photos of each garment, cropping, tagging with categories, colours, and seasons. For example, a wardrobe with 200 items would require a significant upfront effort before the system could even start recommending.

This leads to a large, unrewarded cost, and most users abandon the process within the first 20 items. To address this problem, the ingestion modality should be changed. Instead of taking individual photos, a video pan across the items, extracting frames, detecting garments, segmenting them, removing backgrounds, and automatically tagging attributes can be used. This approach requires only three minutes of user time and significantly reduces the cost barrier.

The second failure point is the use of single-signal recommendations. Once the inventory is created, the simplest approach is to pick a single signal, rank items based on that signal, and return the top result. Common signals include colour harmony, body proportion, weather, and occasion. However, each of these signals is effective on its own but fails when combined with others.

Users perceive a recommendation that is 75% correct as broken, as the single wrong dimension stands out. The real query is not just ranking items based on one dimension but a conjunction of multiple predicates: wearable (outfit) :- palette_compatible (outfit, user. undertone), silhouette_compatible (outfit, user. proportions), warmth_appropriate (outfit, forecast), formality_appropriate (outfit, calendar.event).

Adding more predicates further narrows down the result set. The insight here is that instead of pruning the vast number of plausible candidates, the focus should be on conjunctive reasoning. A single-signal system often feels arbitrary in its choice, while a conjunctive system can return a single, appropriate outfit. This approach reduces decision fatigue for users.

The author also highlights the importance of not overlooking certain signals. In this example, the occasion signal is often omitted, even though it contains the highest information value. Weather can be easily integrated using a public API, but calendar integration is more complex, requiring OAuth, permissions, and parsing messy event titles.

This leads to teams focusing on weather integration and mistakenly believing it covers context. However, weather is a scalar value that constrains only one dimension, while occasion is categorical, providing far more effective pruning.

Lastly, the author discusses a domain-specific detail: the colour model used in many recommender systems may not be the best fit. The typical approach relies on hue relationships, such as complementary or analogous colours. However, this simplistic model fails to account for undertone, which significantly affects how garments appear on different individuals.

Two complementary colours may look terrible on someone with a specific undertone, even if they are aesthetically pleasing for someone else. The author suggests that a more accurate approach is to decompose colours into warm and cool components relative to the user's own colouring. This more nuanced feature selection can lead to better outfit recommendations.