feat(eda): funciones de agregación/OLAP para AutomaticEDA (groupby/pivot push-down + selección LLM)

Cuatro funciones nuevas del grupo eda que nutren el capítulo AGREGACION: - select_groupby_keys (pure): elige categóricas agrupables + numéricas medida desde el TableProfile. - groupby_stats_duckdb (impure): GROUP BY push-down en DuckDB (count/mean/median/std/min/max por grupo). - pivot_table_duckdb (impure): pivot A×B push-down, limitado a top filas/cols para no cortar. - suggest_aggregations_llm (impure): el LLM elige las agregaciones interesantes con fallback determinista. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-30 15:33:55 +02:00
parent 415154d9a3
commit 96da9e3015
13 changed files with 2146 additions and 0 deletions
@@ -0,0 +1,213 @@
+"""Tests para select_groupby_keys (grupo eda, dominio datascience)."""
+
+import os
+import sys
+
+sys.path.insert(0, os.path.dirname(__file__))
+
+from select_groupby_keys import select_groupby_keys
+
+
+def _cat_col(name, card, *, imbalance=2.0, flags=None, null_pct=0.0):
+    """ColumnProfile categorico minimo con bloque categorical."""
+    return {
+        "name": name,
+        "inferred_type": "categorical",
+        "distinct_count": card,
+        "unique_pct": card / 1000.0,
+        "null_pct": null_pct,
+        "flags": flags or [],
+        "numeric": None,
+        "categorical": {"imbalance": imbalance, "mode_pct": 0.5, "n_distinct": card},
+    }
+
+
+def _num_col(name, *, std, cv, flags=None, unique_pct=0.1):
+    """ColumnProfile numerico minimo con bloque numeric."""
+    return {
+        "name": name,
+        "inferred_type": "numeric",
+        "distinct_count": 200,
+        "unique_pct": unique_pct,
+        "null_pct": 0.0,
+        "flags": flags or [],
+        "numeric": {"std": std, "cv": cv},
+        "categorical": None,
+    }
+
+
+def _titanic_like_profile() -> dict:
+    """Perfil estilo titanic: 2 categoricas buenas, 2 numericas, 1 id, 1 constante."""
+    return {
+        "n_rows": 891,
+        "key_candidates": ["passenger_id"],
+        "columns": [
+            _cat_col("sex", 2, imbalance=1.8),
+            _cat_col("pclass", 3, imbalance=2.5),
+            _num_col("age", std=14.5, cv=0.49),
+            _num_col("fare", std=49.7, cv=1.54),
+            # id secuencial: flag possible_id + unique_pct alto.
+            {
+                "name": "passenger_id",
+                "inferred_type": "numeric",
+                "distinct_count": 891,
+                "unique_pct": 1.0,
+                "null_pct": 0.0,
+                "flags": ["possible_id"],
+                "numeric": {"std": 257.4, "cv": 0.58},
+                "categorical": None,
+            },
+            # columna constante: flag constant + std 0.
+            {
+                "name": "embarked_const",
+                "inferred_type": "categorical",
+                "distinct_count": 1,
+                "unique_pct": 0.001,
+                "null_pct": 0.0,
+                "flags": ["constant"],
+                "numeric": None,
+                "categorical": {"imbalance": 1.0},
+            },
+        ],
+    }
+
+
+def test_titanic_picks_good_cats_excludes_id_and_constant():
+    out = select_groupby_keys(_titanic_like_profile())
+
+    # Elige las dos categoricas buenas.
+    chosen_cols = {g["col"] for g in out["group_keys"]}
+    assert chosen_cols == {"sex", "pclass"}
+
+    # Excluye la constante y el key_candidate.
+    assert "embarked_const" not in chosen_cols
+    assert "passenger_id" not in chosen_cols
+
+    # Cada group key trae col, cardinality y score.
+    for g in out["group_keys"]:
+        assert set(g.keys()) == {"col", "cardinality", "score"}
+        assert isinstance(g["score"], float)
+    by_col = {g["col"]: g for g in out["group_keys"]}
+    assert by_col["sex"]["cardinality"] == 2
+    assert by_col["pclass"]["cardinality"] == 3
+
+    # Ordenadas por score descendente.
+    scores = [g["score"] for g in out["group_keys"]]
+    assert scores == sorted(scores, reverse=True)
+
+
+def test_titanic_measures_exclude_id_constant_and_keep_numerics():
+    out = select_groupby_keys(_titanic_like_profile())
+
+    # Solo nombres (strings) de numericas informativas, sin el id secuencial.
+    assert all(isinstance(m, str) for m in out["measures"])
+    assert "passenger_id" not in out["measures"]
+    assert set(out["measures"]) == {"age", "fare"}
+
+    # fare tiene mayor cv (1.54 > 0.49) -> primero.
+    assert out["measures"][0] == "fare"
+
+
+def test_titanic_generates_one_pivot():
+    out = select_groupby_keys(_titanic_like_profile())
+
+    # Con 2 group keys -> exactamente 1 pivot.
+    assert len(out["pivots"]) == 1
+    pivot = out["pivots"][0]
+    assert set(pivot.keys()) == {"index", "columns", "value"}
+    assert {pivot["index"], pivot["columns"]} == {"sex", "pclass"}
+    # El valor es la primera measure (fare).
+    assert pivot["value"] == "fare"
+
+
+def test_empty_profile_returns_all_empty_and_does_not_crash():
+    out = select_groupby_keys({})
+    assert out["group_keys"] == []
+    assert out["measures"] == []
+    assert out["pivots"] == []
+    assert isinstance(out["note"], str)
+
+
+def test_none_profile_does_not_crash():
+    out = select_groupby_keys(None)
+    assert out == {
+        "group_keys": [],
+        "measures": [],
+        "pivots": [],
+        "note": out["note"],
+    }
+    assert isinstance(out["note"], str)
+
+
+def test_only_numerics_yields_empty_group_keys_and_no_pivots():
+    profile = {
+        "n_rows": 500,
+        "key_candidates": [],
+        "columns": [
+            _num_col("price", std=12.0, cv=0.6),
+            _num_col("weight", std=3.0, cv=0.2),
+        ],
+    }
+    out = select_groupby_keys(profile)
+    assert out["group_keys"] == []
+    assert out["pivots"] == []
+    # Las numericas si se eligen como measures.
+    assert set(out["measures"]) == {"price", "weight"}
+    assert out["measures"][0] == "price"  # mayor cv.
+
+
+def test_high_cardinality_and_max_card_are_excluded():
+    profile = {
+        "n_rows": 1000,
+        "key_candidates": [],
+        "columns": [
+            _cat_col("city", 50, flags=["high_cardinality"]),  # flag -> fuera.
+            _cat_col("zone", 35),  # card 35 > max_card 20 -> fuera.
+            _cat_col("region", 5),  # valida.
+        ],
+    }
+    out = select_groupby_keys(profile, max_card=20)
+    assert {g["col"] for g in out["group_keys"]} == {"region"}
+
+
+def test_max_keys_limits_group_keys():
+    profile = {
+        "n_rows": 1000,
+        "key_candidates": [],
+        "columns": [
+            _cat_col("a", 4, imbalance=1.0),
+            _cat_col("b", 5, imbalance=1.2),
+            _cat_col("c", 6, imbalance=1.5),
+            _cat_col("d", 7, imbalance=2.0),
+        ],
+    }
+    out = select_groupby_keys(profile, max_keys=2)
+    assert len(out["group_keys"]) == 2
+    # Hasta 2 pivots con >=2 keys (aqui exactamente 1 par posible entre 2 keys).
+    assert len(out["pivots"]) == 1
+
+
+def test_three_keys_cap_pivots_to_two():
+    profile = {
+        "n_rows": 1000,
+        "key_candidates": [],
+        "columns": [
+            _cat_col("a", 4, imbalance=1.0),
+            _cat_col("b", 5, imbalance=1.1),
+            _cat_col("c", 6, imbalance=1.2),
+            _num_col("m", std=10.0, cv=0.5),
+        ],
+    }
+    out = select_groupby_keys(profile, max_keys=3)
+    assert len(out["group_keys"]) == 3
+    # 3 keys -> 3 pares posibles, capado a 2.
+    assert len(out["pivots"]) == 2
+    for p in out["pivots"]:
+        assert p["value"] == "m"
+
+
+def test_does_not_mutate_input():
+    profile = _titanic_like_profile()
+    before = repr(profile)
+    select_groupby_keys(profile)
+    assert repr(profile) == before