feedrank

Fluxa-PostTopics (FluxRank)

FluxRank is the open-source feed-ranking algorithm behind Fluxa's "For You" timeline — a transparent, deterministic alternative to black-box recommendation systems.

日本語: Fluxaの「おすすめ」フィードを並べているアルゴリズム本体です。機械学習でも ブラックボックスでもなく、誰でも読める1つの数式と公開された重みで動きます。同じ 入力からは常に同じフィードが再現され、全ての投稿に「なぜ表示されたか」の内訳が付きます。

Why open

• Transparent — one scrutable formula; every weight lives in a versioned config, not in code.
• Deterministic — pure functions, no I/O, no clock, seeded randomness only. Identical inputs always produce identical feeds, so the algorithm is auditable and golden-testable.
• Explainable — every ranked post carries Reasons (followed_author, topic:music, mentions_you, …) and a full per-term score Breakdown, enough to render a "why am I seeing this?" panel.
• Privacy-respecting — no dwell-time tracking, no DM signals, no follower-count worship. Earned reach comes only through log-damped engagement.

See ALGORITHM.md for the full specification.

Install

go get github.com/Fluxa-org/Fluxa-PostTopics

Pure Go, standard library only, go >= 1.22.

Quick start

import feedrank "github.com/Fluxa-org/Fluxa-PostTopics"

// Map hashtags (EN/JA aliases included) onto your topic taxonomy.
topics := feedrank.PostTopics("新曲できた！ #音楽", nil, canonical, feedrank.DefaultAliases(), 3)

user := feedrank.UserContext{
    UserID:    "viewer",
    Interests: []string{"music"},
    Follows:   map[string]bool{"alice": true},
}
candidates := []feedrank.Candidate{{
    Post: feedrank.Post{
        ID: "post-1", AuthorID: "alice", CreatedAt: createdAt,
        Likes: 12, Replies: 3, Topics: topics,
    },
    Source: feedrank.SourceInNetwork,
}}

page := feedrank.Rank(feedrank.DefaultConfig(), user, candidates, time.Now(), 20)
for _, r := range page {
    fmt.Println(r.ID, r.Score, r.Reasons) // e.g. post-1 0.32 [followed_author topic:music]
}

The caller owns all I/O: gather candidates from your stores, map them into Candidates, and render the returned page. Rank never touches a database.

The formula

score = freshness × (0.35·engagement + 0.30·affinity + 0.20·topic + 0.15·social_proof + 0.10·search_trend) × modifiers

Modifiers: seen ×0.1 · not-interested topic ×0.2 · stranger-reply ×0.3 · moderation labels (stackable) · prolific-author damp sqrt(threshold/count) · "show more like this" ×1.5 · mentions-you ×2 · language mismatch ×0.5.

Page rules: max 2 posts per author, ~50 % in-network quota, no 3-in-a-row same topic, and every 10th slot reserved for exploration (seeded, stable within the hour) so new authors get discovered.

Feed profiles

One engine, several published weight sets — algorithmic choice in the spirit of Bluesky's feed marketplace:

Low volume is a first-class case: when fresh posts can't fill a page the ranker readmits older ones instead of serving an empty feed, and with few posts it degrades gracefully toward a recency feed.

cfg := feedrank.BuiltinProfiles()["discover"]
cfg, err := feedrank.ConfigFromJSON(raw) // or define your own (unknown fields rejected)

Topics, hashtags, mentions, and search — i18n included

Ranking never consumes raw hashtags. PostTopics maps them onto a canonical taxonomy — ~350 built-in aliases across ten language groups (en, ja, ko, zh, es, hi, vi, fr, de, pt: #音楽→music, #맛집→food, #fútbol→sports) — and unmapped tags contribute nothing, so tag spam cannot game the feed. Extraction is Unicode-aware, so hashtags in any language extract.

• ExtractMentions pulls @handles (Unicode) so callers can set the mentions-you boost.
• MapQuery turns free-text search queries into topics: feed a user's own searches back as SearchInterests, and aggregate everyone's searches into TrendingTopics so the feed reflects what the network is looking for (aggregate counts only — individual queries never reach the ranker).
• Post Language + viewer Languages down-rank posts the viewer can't read; unknown languages are never penalized.

Testing

~35 table-driven tests, randomized invariant tests, fuzzing, and benchmarks; 98 % coverage, race-clean. Ranking 600 candidates into a 50-post page takes ~0.4 ms on an M1 Pro — your database is the bottleneck, not this.

go test -race -cover ./...
go test -run='^$' -fuzz=FuzzRank -fuzztime=30s .
go test -run='^$' -bench=. -benchmem .

What is deliberately absent

Dwell time, profile-visit tracking, DM signals, follower counts, and ML prediction. The pipeline is shaped so a learned scorer could replace the formula later without touching sourcing, filters, or page rules — but v1 stays fully auditable.

License

MIT