Improve Booking via Cover Photo Selection (ML Design)

Pick the cover photo for each listing to maximize click-through and booking. Open-ended ML design — the candidate drives the framing. Requirements Functional For each listing with >= 5 photos, choose the cover photo shown on the search results page. Optimize for click-through to the listing page; secondary for booking conversion. Non-functional Batch inference acceptable; daily or weekly recompute. Latency at serve time is negligible (cover photo is materialized in the listing record). Notes Problem framing. Cast as either (a) supervised ranking over a listing's own photo set, labeled by per-photo click-through rate from a historical experiment that randomized photo order, or (b) absolute photo-quality scoring trained on a labeled photogenicity dataset. Training data. Hardest part. Either run a position-randomized experiment to collect unbiased per-photo CTR, or bootstrap from an external image-quality model (CLIP-style scoring) and refine with a small labeled set. Features. Image embedding (CLIP / vision tower), room-type classifier output, photo composition score (rule-of-thirds, color harmony), text from listing title / description for relevance check. Model. Two stages: (1) an image-quality model producing a per-photo score, (2) a per-listing ranker that re-orders the photo set using listing context. Both are typically GBDT / MLP on top of the embedding. Eval. Offline: AUC against the next click on a held-out cohort. Online: A/B test with click-through and booking as primary, time-to-first-click as guard. Failure modes. Over-fitting to seasonality (winter cabins vs summer beaches), bias toward stock-photo aesthetics, missing photos for new listings. Cold start. New listings get the default first-photo-uploaded behavior until the model has confidence. Preparation Pre-script the position-randomized-experiment framing — the senior signal is recognizing CTR is biased by current photo order. Drill the two-stage model (per-photo quality → per-listing rank) on paper. Pre-write the A/B test plan with primary + guard metrics. Be ready for the "what if we have no clickstream data" fallback (CLIP-style external scoring).