Add GET /v1/users/{id}/feed/for-you endpoint

[dylanjeffers]

Summary

Adds GET /v1/feed/for-you, a personalized track feed modeled on Twitter's open-sourced 2023 algorithm (the-algorithm / the-algorithm-ml). The pipeline is candidate-retrieval → ranking → filtering+diversity, the same three-stage shape Twitter uses on top of a learned heavy ranker. Audius doesn't yet have a trained ranker, so the heavy ranker is approximated by a hand-tuned linear blend; the candidate retrieval and diversity passes carry over directly so a learned model can drop in later.

Client consumer: AudiusProject/apps#14237.

Endpoint

GET /v1/feed/for-you?user_id=<hashId>&limit=25&offset=0&max_per_artist=3

user_id is required (the handler 400s without it — "For You" without a "you" degenerates into trending+underground). limit defaults to 25 (max 100), offset to 0 (max 200), max_per_artist to 3 (max 10).

Algorithm

1. Candidate retrieval (UNION across 4 capped sources)

Source	What it pulls	Cap
in_network	Tracks uploaded in the last 14 days by users I follow	200
trending	Top week-trending from track_trending_scores (mirrors /tracks/trending)	100
underground	Week-trending whose owner has < 1500 follower & following count	50
similar	Recent uploads (60d) by artists saved by users who also save my saved-artists (1-hop saves-graph CF)	100

DISTINCT ON (track_id) ORDER BY track_id, prio keeps the strongest source for each track, so an in-network track that's also trending keeps the in-network weight.

2. Ranking

recency_score    = exp(-ln(2) * age_hours / 48)
                   // 48h half-life: 48h → 0.5, 96h → 0.25
engagement_score = ln(1 + 3*saves + 2*reposts + plays) / 12
                   // preserves saves > reposts > plays, log-compressed
social_boost     = 1.0 + min(ln(1 + my_engagement_with_artist) / 4, 1)
                   // up to ~2x for artists I already engage with often
source_weight    = { in_network: 1.20, trending: 1.00,
                     underground: 0.95, similar: 0.90 }

final_score = (0.55 * recency_score + 0.45 * engagement_score)
              * social_boost * source_weight

3. Filters (applied once after the union)

• Track liveness: is_current, is_delete=false, is_unlisted=false, is_available=true, stem_of IS NULL
• Owner liveness: is_current, is_deactivated=false, is_available=true (same shape as v1_events_remix_contests.go)
• Access-gating: ungated, or caller's wallet is on access_authorities (matches the v1_users_feed authed-wallet pattern)
• Already-saved by caller (don't resurface)
• Caller's own uploads

4. Diversity

• SQL hard cap: ROW_NUMBER() OVER (PARTITION BY owner_id ORDER BY score DESC, track_id DESC) filtered to <= max_per_artist (default 3) — prevents a single hot artist from filling the page.
• Go greedy pass: walks the ranked pool keeping global rank order, but if the next track shares an owner with the one just emitted, prefers the next non-same-owner candidate within a 5-position lookahead. Soft penalty on consecutive-same-artist runs without computing a second ranker.

Pagination is offset/limit applied on the diversity-ordered list, so pages are stable as long as underlying scores haven't shifted.

Test plan

• TestV1FeedForYou_Basic — in-network + trending + underground all surface; deleted/unlisted/deactivated/own/saved tracks are excluded
• TestV1FeedForYou_RequiresUserId — 400 without user_id
• TestV1FeedForYou_ExcludesAlreadySavedTracks — already-saved exclusion works
• TestV1FeedForYou_MaxThreePerArtist — 3-per-artist cap enforced
• TestV1FeedForYou_DiversityPassNoConsecutiveSameArtist — Go greedy pass interleaves artists
• TestV1FeedForYou_PaginationDoesNotRepeat — pagination doesn't repeat ids across pages
• TestV1FeedForYou_InvalidParams — limit/offset out-of-range → 400
• TestV1FeedForYou_RecencyAndEngagementRanking — fresh+engaged outranks low-engagement and old (joint signal test)
• Smoke-test against staging once deployed: hit /v1/feed/for-you?user_id=… for a real account, eyeball the mix of in-network vs trending vs underground
• Cross-check with apps#14237 client integration

🤖 Generated with Claude Code