File size: 6,051 Bytes
a91323c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
from __future__ import annotations

from dataclasses import dataclass
from typing import Dict, List, Tuple, Iterable, Set, Any

import math

from .alignment import AlignmentEngine
from .models import StrategicObjective, ActionTask


@dataclass
class EvalConfig:
    top_k: int = 5


def precision_recall_at_k(
    pred_ids: List[str], truth_ids: Set[str], k: int
) -> Tuple[float, float]:
    preds = pred_ids[:k]
    hits = sum(1 for pid in preds if pid in truth_ids)
    precision = hits / max(1, len(preds))
    recall = hits / max(1, len(truth_ids))
    return precision, recall


def average_precision(pred_ids: List[str], truth_ids: Set[str]) -> float:
    hits = 0
    ap_sum = 0.0
    for i, pid in enumerate(pred_ids, start=1):
        if pid in truth_ids:
            hits += 1
            ap_sum += hits / i
    return ap_sum / max(1, hits)


def ndcg_at_k(pred_ids: List[str], truth_ids: Set[str], k: int) -> float:
    # Relevance is binary: 1 if in truth, else 0
    dcg = 0.0
    for i, pid in enumerate(pred_ids[:k], start=1):
        rel = 1.0 if pid in truth_ids else 0.0
        dcg += rel / math.log2(i + 1)
    # Ideal DCG assumes all relevant items are ranked first
    ideal_rel_count = min(len(truth_ids), k)
    idcg = sum(1.0 / math.log2(i + 1) for i in range(1, ideal_rel_count + 1))
    return dcg / idcg if idcg > 0 else 0.0


@dataclass
class StrategyEval:
    strategy_id: str
    precision_at_k: float
    recall_at_k: float
    ap: float
    ndcg: float


@dataclass
class EvalSummary:
    top_k: int
    macro_precision: float
    macro_recall: float
    map: float
    mean_ndcg: float
    per_strategy: List[StrategyEval]
    similarity_summary: Dict[str, float] | None = None


def evaluate_alignment(
    engine: AlignmentEngine,
    strategies: Iterable[StrategicObjective],
    actions: Iterable[ActionTask],
    ground_truth: Dict[str, List[str]],
    config: EvalConfig | None = None,
) -> EvalSummary:
    cfg = config or EvalConfig()
    # Run alignment retrieval
    result = engine.align(
        strategies=list(strategies), actions=list(actions), top_k=cfg.top_k
    )

    per_strategy: List[StrategyEval] = []
    p_list: List[float] = []
    r_list: List[float] = []
    ap_list: List[float] = []
    ndcg_list: List[float] = []

    for sres in result["strategy_results"]:
        sid = sres["strategy_id"]
        preds = [m["action_id"] for m in sres.get("top_matches", [])]
        truth = set(ground_truth.get(sid, []))
        p, r = precision_recall_at_k(preds, truth, cfg.top_k)
        ap = average_precision(preds, truth)
        nd = ndcg_at_k(preds, truth, cfg.top_k)

        per_strategy.append(
            StrategyEval(
                strategy_id=sid,
                precision_at_k=p,
                recall_at_k=r,
                ap=ap,
                ndcg=nd,
            )
        )
        p_list.append(p)
        r_list.append(r)
        ap_list.append(ap)
        ndcg_list.append(nd)

    summary = EvalSummary(
        top_k=cfg.top_k,
        macro_precision=sum(p_list) / max(1, len(p_list)),
        macro_recall=sum(r_list) / max(1, len(r_list)),
        map=sum(ap_list) / max(1, len(ap_list)),
        mean_ndcg=sum(ndcg_list) / max(1, len(ndcg_list)),
        per_strategy=per_strategy,
        similarity_summary=None,
    )
    return summary


def precision_at_k(pred_ids: List[str], truth_ids: Set[str], k: int) -> float:
    p, _ = precision_recall_at_k(pred_ids, truth_ids, k)
    return p


def recall_at_k(pred_ids: List[str], truth_ids: Set[str], k: int) -> float:
    _, r = precision_recall_at_k(pred_ids, truth_ids, k)
    return r


def run_evaluation(
    alignment_result: Dict[str, Any], ground_truth_path: str | None, top_k: int = 5
) -> Dict[str, Any]:
    """Compute Precision@K, Recall@K and similarity summaries given alignment results.

    Ground truth format: {"S1": ["A3","A9"], "S2": ["A2"], ...}
    """
    import json
    from pathlib import Path

    truth_map: Dict[str, List[str]] = {}
    if ground_truth_path:
        p = Path(ground_truth_path)
        if p.exists():
            with p.open("r", encoding="utf-8") as f:
                data = json.load(f)
            if isinstance(data, dict):
                truth_map = {str(k): list(v or []) for k, v in data.items()}

    per_strategy: List[Dict[str, Any]] = []
    p_list: List[float] = []
    r_list: List[float] = []

    retrieved_sims: List[float] = []
    relevant_sims: List[float] = []

    for sres in alignment_result.get("strategy_results", []):
        sid = sres.get("strategy_id")
        preds = [m.get("action_id") for m in sres.get("top_matches", [])]
        sims = [float(m.get("similarity", 0.0)) for m in sres.get("top_matches", [])]
        truth = set(truth_map.get(str(sid), []))

        p, r = precision_recall_at_k(preds, truth, top_k)
        ap = average_precision(preds, truth)
        nd = ndcg_at_k(preds, truth, top_k)

        per_strategy.append(
            {
                "strategy_id": sid,
                "precision_at_k": p,
                "recall_at_k": r,
                "ap": ap,
                "ndcg": nd,
            }
        )
        p_list.append(p)
        r_list.append(r)

        # Similarity summaries
        retrieved_sims.extend(sims)
        # Relevant sims: similarity of matches that are in ground truth
        for m in sres.get("top_matches", []):
            if m.get("action_id") in truth:
                relevant_sims.append(float(m.get("similarity", 0.0)))

    eval_summary = {
        "top_k": top_k,
        "macro_precision": sum(p_list) / max(1, len(p_list)),
        "macro_recall": sum(r_list) / max(1, len(r_list)),
        "per_strategy": per_strategy,
        "similarity_summary": {
            "retrieved_mean": (sum(retrieved_sims) / max(1, len(retrieved_sims)))
            if retrieved_sims
            else 0.0,
            "relevant_mean": (sum(relevant_sims) / max(1, len(relevant_sims)))
            if relevant_sims
            else 0.0,
        },
    }
    return eval_summary