File size: 13,569 Bytes
97b9b33
 
 
 
 
 
 
 
 
5223365
 
2330821
 
 
 
ac6d643
97b9b33
 
 
 
 
 
 
 
 
2330821
97b9b33
 
 
ac6d643
 
 
2330821
 
 
ac6d643
2330821
ac6d643
 
2330821
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5223365
2330821
 
 
 
97b9b33
 
2330821
97b9b33
 
 
2330821
97b9b33
2330821
97b9b33
 
 
 
 
 
 
 
 
 
2330821
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
97b9b33
2330821
 
97b9b33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2330821
 
 
97b9b33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2330821
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
97b9b33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2330821
 
 
 
 
97b9b33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2330821
97b9b33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2330821
97b9b33
 
 
 
 
 
 
 
 
 
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
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
from __future__ import annotations

import json
from pathlib import Path
from typing import Any, Dict, List, Optional

import joblib
import numpy as np
import pandas as pd
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel, Field
import xgboost as xgb

import os
from huggingface_hub import hf_hub_download

# Compatibility shim for pickles created with newer sklearn that include _RemainderColsList
import sklearn.compose._column_transformer as _ct  # type: ignore
if not hasattr(_ct, "_RemainderColsList"):
    class _RemainderColsList(list):  # type: ignore
        pass
    _ct._RemainderColsList = _RemainderColsList

ROOT = Path(__file__).resolve().parents[1]
MODEL_DIR = Path(__file__).resolve().parent / "model"
# MODEL_PATH = MODEL_DIR / "xgboost_pipeline.pkl"
BOOSTER_PATH = MODEL_DIR / "xgboost_booster.json"
META_PATH = MODEL_DIR / "explain_meta.json"

HF_MODEL_REPO = os.getenv("HF_MODEL_REPO", "Gutema/frankscore-model-artifact")
HF_MODEL_REVISION = os.getenv("HF_MODEL_REVISION", "main")

try:
    MODEL_PATH = Path(
        hf_hub_download(
            repo_id=HF_MODEL_REPO,
            filename="xgboost_pipeline.pkl",
            revision=HF_MODEL_REVISION,
        )
    )
except Exception as e:
    raise RuntimeError(f"Failed to download model artifact from HF repo={HF_MODEL_REPO}: {e}") from e

if not META_PATH.exists():
    raise FileNotFoundError(f"Explainability meta missing at {META_PATH}")
if not BOOSTER_PATH.exists():
    raise FileNotFoundError(f"Booster file missing at {BOOSTER_PATH}")

if not MODEL_PATH.exists():
    raise FileNotFoundError(f"Model file missing at {MODEL_PATH}")
if not META_PATH.exists():
    raise FileNotFoundError(f"Explainability meta missing at {META_PATH}")
if not BOOSTER_PATH.exists():
    raise FileNotFoundError(f"Booster file missing at {BOOSTER_PATH}")

PIPELINE = joblib.load(MODEL_PATH)
META = json.loads(META_PATH.read_text())

EXPECTED_FEATURES = list(getattr(PIPELINE, "feature_names_in_", []))
PREPROCESS = PIPELINE.named_steps.get("preprocess") if hasattr(PIPELINE, "named_steps") else None
if PREPROCESS is None:
    raise RuntimeError("Pipeline missing 'preprocess' step; cannot infer columns.")

if not EXPECTED_FEATURES:
    EXPECTED_FEATURES = list(getattr(PREPROCESS, "feature_names_in_", []))
    if not EXPECTED_FEATURES:
        raise RuntimeError("Unable to determine expected feature names from the pipeline.")

_col_map = {name: cols for name, _, cols in getattr(PREPROCESS, "transformers_", [])}
NUM_FEATURES = list(_col_map.get("num", []))
CAT_FEATURES = list(_col_map.get("cat", []))
PRE_FEATURE_NAMES = META.get("pre_feature_names") or list(getattr(PREPROCESS, "get_feature_names_out", lambda: [])())
RAW_FEATURE_SET = set((META.get("raw_num_cols") or []) + (META.get("raw_cat_cols") or []))
FEATURE_GROUPS = {
    "Repayment Activity": [
        "num_previous_defaults",
        "past_default_rate",
        "repayment_consistency",
        "repayment_intensity",
    ],
    "Loan Amount & Burden": [
        "Total_Amount",
        "Total_Amount_to_Repay",
        "amount_bucket",
        "burden_percentile",
        "daily_burden",
        "duration",
        "duration_bucket",
        "interest_rate",
        "amount_ratio",
        "burden_ratio",
        "lender_exposure_ratio",
    ],
    "Borrowing History": [
        "account_age_days",
        "avg_past_amount",
        "avg_past_daily_burden",
        "avg_time_bw_loans",
        "borrower_history_strength",
        "days_since_last_loan",
        "loan_frequency_per_year",
        "num_previous_loans",
        "std_past_amount",
        "std_past_daily_burden",
        "trend_in_amount",
        "trend_in_burden",
        "lender_id",
        "lender_risk_profile",
    ],
    "Spending & Transactions": [
        "latest_amount_ma3",
        "days_to_local_festival",
        "days_to_salary_day",
        "month",
        "quarter",
        "week_of_year",
    ],
}
FEATURE_GROUP_LOOKUP: Dict[str, str] = {}
for group, variables in FEATURE_GROUPS.items():
    for var in variables:
        FEATURE_GROUP_LOOKUP[var] = group

app = FastAPI(title="FrankScore", version="1.0.0")


class PredictionRequest(BaseModel):
    records: List[Dict[str, Any]] = Field(..., description="List of borrower feature dictionaries")


class PredictionResponse(BaseModel):
    probabilities: List[float]


class ScoreRequest(BaseModel):
    probabilities: List[float] = Field(..., description="Probabilities of default (0-1)")


class ScoreResponse(BaseModel):
    scores: List[float]


class ExplainRequest(BaseModel):
    records: List[Dict[str, Any]]
    top_k: Optional[int] = Field(default=10, ge=1, le=100, description="Number of top features to return per record")


class FeatureContribution(BaseModel):
    feature: str
    shap_value: float


class GroupContribution(BaseModel):
    group: str
    total_shap_value: float
    percentage: float
    direction: str
    label: str
    features: List[FeatureContribution]


class ExplainItem(BaseModel):
    probability: float
    base_value: float
    group_contributions: List[GroupContribution]


class ExplainResponse(BaseModel):
    explanations: List[ExplainItem]


class PredictExplainItem(BaseModel):
    probability: float
    score: float
    base_value: float
    group_contributions: List[GroupContribution]


class PredictExplainResponse(BaseModel):
    results: List[PredictExplainItem]


def prepare_frame(records: List[Dict[str, Any]]) -> pd.DataFrame:
    if not records:
        raise HTTPException(status_code=400, detail="No records provided.")
    df = pd.DataFrame(records)
    for col in EXPECTED_FEATURES:
        if col not in df.columns:
            df[col] = np.nan
    df = df[EXPECTED_FEATURES]
    if NUM_FEATURES:
        df[NUM_FEATURES] = df[NUM_FEATURES].apply(pd.to_numeric, errors="coerce")
    if CAT_FEATURES:
        df[CAT_FEATURES] = df[CAT_FEATURES].astype("object")
    return df


def pd_to_score(p: np.ndarray, base_score: float = 50, base_odds: float = 9, pdo: float = 20) -> np.ndarray:
    p = np.clip(p, 1e-6, 1 - 1e-6)
    B = pdo / np.log(2)
    A = base_score - B * np.log(base_odds)
    odds = (1 - p) / p
    score = A + B * np.log(odds)
    return np.clip(score, 0, 100)


def _sanitize_feature_name(name: str) -> str:
    sanitized = name
    for ch, repl in {"[": "", "]": "", "<": "lt", ">": "gt", " ": "_", ",": "_", "=": "_"}.items():
        sanitized = sanitized.replace(ch, repl)
    return sanitized


def _base_feature_name(name: str) -> str:
    base = name
    if "__" in base:
        base = base.split("__", 1)[1]
    if base in RAW_FEATURE_SET:
        return base
    parts = base.split("_")
    while len(parts) > 1:
        candidate = "_".join(parts[:-1])
        if candidate in RAW_FEATURE_SET:
            return candidate
        parts = parts[:-1]
    return base


def _label_for_percentage(pct: float) -> str:
    if pct >= 30:
        return "Exceptional"
    if pct >= 20:
        return "Very Good"
    if pct >= 10:
        return "Good"
    if pct >= 5:
        return "Bad"
    return "Very Bad"


def _direction_for_value(val: float) -> str:
    if val > 0:
        return "raises risk"
    if val < 0:
        return "reduces risk"
    return "neutral"


def _build_group_contribs(
    group_totals: Dict[str, float],
    group_details: Dict[str, List[FeatureContribution]],
    top_k: Optional[int],
) -> List[GroupContribution]:
    denom = sum(abs(v) for v in group_totals.values())
    if denom == 0:
        denom = 1e-12  # avoid division by zero; all percentages become ~0
    group_contribs: List[GroupContribution] = []
    for grp, total in sorted(group_totals.items(), key=lambda kv: abs(kv[1]), reverse=True):
        feats = sorted(group_details.get(grp, []), key=lambda fc: abs(fc.shap_value), reverse=True)
        if top_k:
            feats = feats[:top_k]
        pct = abs(total) / denom * 100
        group_contribs.append(
            GroupContribution(
                group=grp,
                total_shap_value=total,
                percentage=pct,
                direction=_direction_for_value(total),
                label=_label_for_percentage(pct),
                features=feats,
            )
        )
    return group_contribs


def get_booster():
    if not hasattr(get_booster, "_booster"):
        booster = xgb.Booster()
        booster.load_model(str(BOOSTER_PATH))
        base_score = booster.attr("base_score")
        if base_score:
            try:
                float(base_score)
            except ValueError:
                cleaned = base_score.strip("[]")
                try:
                    cleaned_val = str(float(cleaned))
                except Exception:
                    cleaned_val = "0.5"
                booster.set_param({"base_score": cleaned_val})
                booster.set_attr(base_score=cleaned_val)
        get_booster._booster = booster
    return get_booster._booster


@app.post("/predict", response_model=PredictionResponse)
def predict(req: PredictionRequest) -> PredictionResponse:
    frame = prepare_frame(req.records)
    probas = PIPELINE.predict_proba(frame)[:, 1]
    return PredictionResponse(probabilities=probas.tolist())


@app.get("/health")
def health() -> Dict[str, str]:
    return {"status": "ok", "model_path": str(MODEL_PATH)}


@app.post("/score", response_model=ScoreResponse)
def score(req: ScoreRequest) -> ScoreResponse:
    if not req.probabilities:
        raise HTTPException(status_code=400, detail="No probabilities provided.")
    arr = np.array(req.probabilities, dtype=float)
    scores = pd_to_score(arr)
    return ScoreResponse(scores=scores.tolist())


@app.post("/explain", response_model=ExplainResponse)
def explain(req: ExplainRequest) -> ExplainResponse:
    if not req.records:
        raise HTTPException(status_code=400, detail="No records provided.")
    frame = prepare_frame(req.records)
    probas = PIPELINE.predict_proba(frame)[:, 1]
    booster = get_booster()
    X_proc = PREPROCESS.transform(frame)
    feat_names = np.array(PRE_FEATURE_NAMES) if PRE_FEATURE_NAMES else np.array([f"f{i}" for i in range(X_proc.shape[1])])
    sanitized_names = [_sanitize_feature_name(n) for n in feat_names]
    dmat = xgb.DMatrix(X_proc, feature_names=sanitized_names)
    contribs = booster.predict(dmat, pred_contribs=True)
    if contribs.shape[1] != X_proc.shape[1] + 1:
        raise HTTPException(status_code=500, detail="Unexpected contribution shape from booster.")
    base_vals = contribs[:, -1]
    feat_contribs = contribs[:, :-1]
    explanations: List[ExplainItem] = []
    for i in range(feat_contribs.shape[0]):
        row_vals = feat_contribs[i]
        group_totals: Dict[str, float] = {}
        group_details: Dict[str, List[FeatureContribution]] = {}
        for name, val in zip(feat_names, row_vals):
            base = _base_feature_name(str(name))
            group = FEATURE_GROUP_LOOKUP.get(base, "Other")
            group_totals[group] = group_totals.get(group, 0.0) + float(val)
            group_details.setdefault(group, []).append(
                FeatureContribution(feature=str(name), shap_value=float(val))
            )
        group_contribs = _build_group_contribs(group_totals, group_details, req.top_k)
        explanations.append(
            ExplainItem(
                probability=float(probas[i]),
                base_value=float(base_vals[i]),
                group_contributions=group_contribs,
            )
        )
    return ExplainResponse(explanations=explanations)


@app.post("/predict_explain", response_model=PredictExplainResponse)
def predict_explain(req: ExplainRequest) -> PredictExplainResponse:
    if not req.records:
        raise HTTPException(status_code=400, detail="No records provided.")
    frame = prepare_frame(req.records)
    probas = PIPELINE.predict_proba(frame)[:, 1]
    booster = get_booster()
    X_proc = PREPROCESS.transform(frame)
    feat_names = np.array(PRE_FEATURE_NAMES) if PRE_FEATURE_NAMES else np.array([f"f{i}" for i in range(X_proc.shape[1])])
    sanitized_names = [_sanitize_feature_name(n) for n in feat_names]
    dmat = xgb.DMatrix(X_proc, feature_names=sanitized_names)
    contribs = booster.predict(dmat, pred_contribs=True)
    if contribs.shape[1] != X_proc.shape[1] + 1:
        raise HTTPException(status_code=500, detail="Unexpected contribution shape from booster.")
    base_vals = contribs[:, -1]
    feat_contribs = contribs[:, :-1]
    items: List[PredictExplainItem] = []
    for i in range(feat_contribs.shape[0]):
        row_vals = feat_contribs[i]
        group_totals: Dict[str, float] = {}
        group_details: Dict[str, List[FeatureContribution]] = {}
        for name, val in zip(feat_names, row_vals):
            base = _base_feature_name(str(name))
            group = FEATURE_GROUP_LOOKUP.get(base, "Other")
            group_totals[group] = group_totals.get(group, 0.0) + float(val)
            group_details.setdefault(group, []).append(
                FeatureContribution(feature=str(name), shap_value=float(val))
            )
        group_contribs = _build_group_contribs(group_totals, group_details, req.top_k)
        score_val = int(round(float(pd_to_score(np.array([probas[i]]))[0])))
        items.append(
            PredictExplainItem(
                probability=float(probas[i]),
                score=score_val,
                base_value=float(base_vals[i]),
                group_contributions=group_contribs,
            )
        )
    return PredictExplainResponse(results=items)