feat(recommender): nightly ALS with offline metrics and CF candidates in the wave (M6)

- Implicit scores from play events and reactions per the signals table
  (90-day half-life, weaker skips outside the wave, dislikes override).
- numpy ALS for implicit feedback; time-split evaluation on the last
  7 days (NDCG@10, Recall@50) against a popularity baseline; a new model
  is activated only if NDCG@10 does not regress by more than 2%.
- Vectors are written to track_embeddings.cf and user_profiles_rec.taste_cf,
  versions to model_versions; NightlyTrainer runs at TRAIN_HOUR and on
  first start, `pipu-rec train` runs it manually.
- Wave: cf and item_cf candidate sources and a 0.10 cf term in the v1
  score; impressions are tagged heuristic-v1+cf.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

services/recommender/pipu_rec/wave/engine.py

@@ -10,6 +10,8 @@ from numpy.typing import NDArray
 from pipu_rec.wave.contexts import FEATURES, Rule, fit
 
 MODEL_VER = "heuristic-v1"
+# С коллаборативным сигналом (ALS) — суффикс к версии модели в показах.
+CF_SUFFIX = "+cf"
 
 Vec = NDArray[np.float32]
 
@@ -24,10 +26,15 @@ class CatalogIndex:
     has_emb: NDArray[np.bool_]
     features: dict[str, NDArray[np.float64]]  # NaN — нет признака
     popularity: NDArray[np.float64]  # 0..1
+    cf: Vec = field(default_factory=lambda: np.zeros((0, 0), dtype=np.float32))  # (n, k) ALS, нормирован
+    has_cf: NDArray[np.bool_] = field(default_factory=lambda: np.zeros(0, dtype=bool))
     pos: dict[str, int] = field(default_factory=dict)
 
     def __post_init__(self) -> None:
         self.pos = {t: i for i, t in enumerate(self.ids)}
+        if self.has_cf.shape[0] != len(self.ids):
+            self.cf = np.zeros((len(self.ids), 1), dtype=np.float32)
+            self.has_cf = np.zeros(len(self.ids), dtype=bool)
 
     def __len__(self) -> int:
         return len(self.ids)
@@ -40,6 +47,7 @@ class CatalogIndex:
         features: dict[str, list[float | None]],
         plays: list[int],
         dim: int = 256,
+        raw_cf: list[Vec | None] | None = None,
     ) -> CatalogIndex:
         n = len(ids)
         emb = np.zeros((n, dim), dtype=np.float32)
@@ -59,7 +67,17 @@ class CatalogIndex:
         }
         p = np.log1p(np.array(plays, dtype=np.float64))
         pop = p / p.max() if n and p.max() > 0 else np.zeros(n)
-        return CatalogIndex(ids=ids, artist=artist, emb=emb, has_emb=has, features=feats, popularity=pop)
+        cf_dim = next((len(v) for v in raw_cf or [] if v is not None), 0)
+        cf = np.zeros((n, max(cf_dim, 1)), dtype=np.float32)
+        has_cf = np.zeros(n, dtype=bool)
+        for i, v in enumerate(raw_cf or []):
+            if v is not None and len(v) == cf_dim:
+                norm = float(np.linalg.norm(v))
+                if norm > 0:
+                    cf[i], has_cf[i] = v / norm, True
+        return CatalogIndex(
+            ids=ids, artist=artist, emb=emb, has_emb=has, features=feats, popularity=pop, cf=cf, has_cf=has_cf
+        )
 
 
 @dataclass
@@ -74,6 +92,8 @@ class Profile:
     liked_recent: set[str]  # лайкнутые, игравшие за 24 ч — штраф
     rediscover: set[str]  # любимые, не игравшие > 60 дней
     exploration: float = 0.15
+    taste_cf: list[float] | None = None  # фактор пользователя ALS
+    recent_likes: list[str] = field(default_factory=list)  # последние лайки, новые первыми (для item_cf)
 
 
 @dataclass
@@ -145,6 +165,21 @@ def update_session(
     session.vec = vec.tolist() if np.any(vec) else None
 
 
+def cf_scores(index: CatalogIndex, profile: Profile) -> NDArray[np.float64] | None:
+    """Косинус треков к фактору пользователя ALS; None — нет модели или пользователь в неё не попал."""
+    if profile.taste_cf is None or not index.has_cf.any():
+        return None
+    u = np.array(profile.taste_cf, dtype=np.float32)
+    if u.shape[0] != index.cf.shape[1]:
+        return None
+    u = u / (float(np.linalg.norm(u)) or 1.0)
+    return np.where(index.has_cf, index.cf @ u, 0.0).astype(np.float64)
+
+
+def model_version(index: CatalogIndex, profile: Profile) -> str:
+    return MODEL_VER + (CF_SUFFIX if cf_scores(index, profile) is not None else "")
+
+
 def _top(scores: NDArray[np.float64], mask: NDArray[np.bool_], k: int) -> NDArray[np.intp]:
     idx = np.nonzero(mask)[0]
     if len(idx) <= k:
@@ -189,6 +224,7 @@ def build_batch(
     sim_session = index.emb @ svec if svec is not None else np.zeros(size)
     sim_taste = np.where(index.has_emb, sim_taste, 0.0)
     sim_session = np.where(index.has_emb, sim_session, 0.0)
+    cf_score = cf_scores(index, profile)
 
     # Привязанность к исполнителю — сглаженное среднее оценок его треков.
     totals: dict[str, list[float]] = {}
@@ -217,6 +253,15 @@ def build_batch(
         never = np.array([t not in profile.played for t in index.ids])
         add(_top(sim_taste + 0.3 * sim_session, allowed & index.has_emb & never, 50), "fresh")
         add(_top(query, allowed & index.has_emb, 150), "audio_sim")
+    # Коллективный опыт (ALS): похожие по вкусу других слушателей и «с этим слушают» к последним лайкам.
+    if cf_score is not None:
+        add(_top(cf_score, allowed & index.has_cf, 100), "cf")
+        likes = [
+            index.pos[t] for t in profile.recent_likes[:10] if t in index.pos and index.has_cf[index.pos[t]]
+        ]
+        if likes:
+            item_sim = (index.cf @ index.cf[likes].T).max(axis=1).astype(np.float64)
+            add(_top(np.where(index.has_cf, item_sim, -1.0), allowed & index.has_cf, 60), "item_cf")
     liked_artists = {a for a, v in affinity.items() if v > 0.3}
     if liked_artists:
         by_artist = [int(i) for i in np.nonzero(allowed)[0] if index.artist[i] in liked_artists]
@@ -234,6 +279,7 @@ def build_batch(
         + 0.25 * sim_taste[cand]
         + 0.15 * artist_aff[cand]
         + 0.05 * index.popularity[cand]
+        + (0.10 * cf_score[cand] if cf_score is not None else 0.0)
         + 0.02 * rng.random(len(cand))  # разбивает ничьи, особенно при холодном старте
     )
     liked_recent = np.array([index.ids[i] in profile.liked_recent for i in cand])

services/recommender/pipu_rec/wave/service.py

@@ -9,7 +9,7 @@ import psycopg
 from pydantic import BaseModel, Field
 
 from pipu_rec.wave.contexts import Contexts, Rule
-from pipu_rec.wave.engine import MODEL_VER, Pick, build_batch, taste_centroids, update_session
+from pipu_rec.wave.engine import Pick, build_batch, model_version, taste_centroids, update_session
 from pipu_rec.wave.store import IndexHolder, SessionStore, load_profile, seed_tracks
 
 MAX_SERVED_MEMORY = 300
@@ -89,5 +89,6 @@ class WaveService:
         session.recent_artists = (session.recent_artists + artists)[-10:]
         self.sessions.put(req.session_id, session)
         return BatchResponse(
-            tracks=[BatchTrack(id=p.id, source=p.source, score=p.score) for p in picks], model_ver=MODEL_VER
+            tracks=[BatchTrack(id=p.id, source=p.source, score=p.score) for p in picks],
+            model_ver=model_version(index, profile),
         )

services/recommender/pipu_rec/wave/store.py

@@ -39,7 +39,8 @@ def load_index(conn: psycopg.Connection[Any]) -> CatalogIndex:
                 WHERE ta.track_id = t.id AND ta.role = 'main' ORDER BY ta.position LIMIT 1),
                 al.album_artist_id::text),
             e.audio::text, f.bpm, f.energy, f.danceability, f.valence, f.acousticness,
-            COALESCE((SELECT sum(play_count) FROM user_track_stats s WHERE s.track_id = t.id), 0)
+            COALESCE((SELECT sum(play_count) FROM user_track_stats s WHERE s.track_id = t.id), 0),
+            e.cf::text
         FROM tracks t
         JOIN albums al ON al.id = t.album_id
         JOIN media_files mf ON mf.id = t.primary_file_id AND mf.missing_since IS NULL
@@ -60,6 +61,7 @@ def load_index(conn: psycopg.Connection[Any]) -> CatalogIndex:
         raw_emb=[parse_vector(r[2]) for r in rows],
         features=feats,
         plays=[int(r[8]) for r in rows],
+        raw_cf=[parse_vector(r[9]) for r in rows],
     )
 
 
@@ -123,11 +125,15 @@ def load_profile(conn: psycopg.Connection[Any], user_id: str) -> Profile:
         scores[tid] = raw * math.exp(-math.log(2) * (now - ts) / 86400 / HALF_LIFE_DAYS)
     disliked: set[str] = set()
     liked: set[str] = set()
+    recent_likes: list[str] = []
     for tid, reaction in conn.execute(
-        "SELECT track_id::text, reaction::text FROM track_reactions WHERE user_id = %s", (user_id,)
+        "SELECT track_id::text, reaction::text FROM track_reactions WHERE user_id = %s "
+        "ORDER BY created_at DESC",
+        (user_id,),
     ):
         if reaction == "like":
             liked.add(tid)
+            recent_likes.append(tid)
             scores[tid] = scores.get(tid, 0.0) + 3.0
         else:
             disliked.add(tid)
@@ -149,8 +155,11 @@ def load_profile(conn: psycopg.Connection[Any], user_id: str) -> Profile:
             (user_id,),
         )
     }
-    row = conn.execute("SELECT exploration FROM user_profiles_rec WHERE user_id = %s", (user_id,)).fetchone()
+    row = conn.execute(
+        "SELECT exploration, taste_cf::text FROM user_profiles_rec WHERE user_id = %s", (user_id,)
+    ).fetchone()
     exploration = float(row[0]) if row else 0.15
+    taste_cf = parse_vector(row[1]) if row else None
     if len(played) < 30:
         exploration = max(exploration, 0.4)  # холодный старт: больше исследования на первых треках
     day = 86400
@@ -163,6 +172,8 @@ def load_profile(conn: psycopg.Connection[Any], user_id: str) -> Profile:
         liked_recent={t for t in liked if now - last.get(t, 0) < day},
         rediscover={t for t, s in scores.items() if s > 1.5 and now - last.get(t, now) > 60 * day},
         exploration=exploration,
+        taste_cf=taste_cf.tolist() if taste_cf is not None else None,
+        recent_likes=recent_likes,
     )