#!/usr/bin/env python3
"""Compare feature importances from several methods and explain where they disagree.
Input: a wide CSV with a `feature` column and one column per method, e.g.
feature,gain,split,permutation,permutation_std,mean_abs_shap,drop_column
Method kinds are recognised from the column name: gain/split/weight/cover = impurity-based (training data),
permutation/drop = held-out loss-based, shap = mean |SHAP|, coef = |standardised coefficient|. A column
`<method>_std` gives the standard deviation over permutation repeats.
Reports:
table each method as a share of its total (negative permutation values kept as-is), the rank per
method, and the consensus rank (mean rank across methods)
agreement Spearman correlation between every pair of methods
flags IMPURITY-ONLY high gain/split share but little held-out (permutation/drop) importance: split bias
towards continuous or high-cardinality features, or redundancy with a
correlated feature (permuting one is compensated by the other)
NO-EFFECT held-out importance within 2 std of zero (or <= 0 without std): no measurable
contribution on new data; candidate to drop (check correlated groups first)
ID-LIKE name looks like an identifier/hash/timestamp: should not be a feature
DOMINANT one feature holds > 50% of a held-out method: check for leakage
SHAP-VS-PERM mean |SHAP| share at least 3x the held-out share: the model uses it, but a
correlated feature carries the same information
With --groups groups.csv (written by gen_importance_script.py: group, group_permutation,
group_permutation_std, sum_single_permutation), groups whose joint permutation exceeds the sum of their
single-feature permutations by > 30% are reported as SHARED-INFO: the members carry overlapping information,
so report the group's importance rather than each member's.
The recommended headline method is held-out permutation (or drop-column) importance, with SHAP for
direction and impurity importance only as a cheap first look.
Usage:
python compare_importances.py importances.csv
python compare_importances.py importances.csv --groups groups.csv
python compare_importances.py importances.csv --top 10 --json
Exit codes: 0 compared, 1 a DOMINANT or ID-LIKE flag was raised, 2 usage or read error.
"""
import argparse
import csv
import json
import math
import re
import sys
ID_LIKE = re.compile(r"(^id$|_id$|^id_|uuid|guid|hash|_key$|index|row_?num|timestamp|^ts$|created_at|updated_at)", re.I)
def kind_of(col):
c = col.lower()
if "shap" in c:
return "shap"
if "perm" in c or "drop" in c:
return "heldout"
if "coef" in c:
return "coef"
if any(k in c for k in ("gain", "split", "weight", "cover", "impurity", "gini", "mdi")):
return "impurity"
return "other"
def ranks_desc(xs):
order = sorted(range(len(xs)), key=lambda i: -xs[i])
r = [0.0] * len(xs)
i = 0
while i < len(order):
j = i
while j + 1 < len(order) and xs[order[j + 1]] == xs[order[i]]:
j += 1
for k in range(i, j + 1):
r[order[k]] = (i + j) / 2 + 1
i = j + 1
return r
def spearman(a, b):
ra, rb = ranks_desc(a), ranks_desc(b)
n = len(a)
ma, mb = sum(ra) / n, sum(rb) / n
num = sum((x - ma) * (y - mb) for x, y in zip(ra, rb))
den = math.sqrt(sum((x - ma) ** 2 for x in ra) * sum((y - mb) ** 2 for y in rb))
return num / den if den else 0.0
def load(path):
with open(path, newline="", encoding="utf-8") as f:
rows = list(csv.DictReader(f))
if not rows:
raise ValueError("no rows")
cols = list(rows[0])
fcol = next((c for c in cols if c.lower() in ("feature", "features", "name", "variable")), None)
if fcol is None:
raise ValueError("need a 'feature' column")
methods = [c for c in cols if c != fcol and not c.lower().endswith("_std")]
if not methods:
raise ValueError("no method columns")
feats = [r[fcol] for r in rows]
if len(set(feats)) < len(feats):
raise ValueError("features repeat: this looks like a long (run, feature, importance) file; pivot it to one "
"row per feature, or check run-to-run stability with importance_stability.py")
vals, stds = {}, {}
for m in methods:
try:
vals[m] = [float(r[m]) if r[m] not in ("", None) else 0.0 for r in rows]
except ValueError:
raise ValueError(f"column {m!r} is not numeric")
sc = next((c for c in cols if c.lower() == f"{m.lower()}_std"), None)
if sc:
stds[m] = [float(r[sc] or 0) for r in rows]
return feats, methods, vals, stds
def analyse(feats, methods, vals, stds):
n = len(feats)
kinds = {m: kind_of(m) for m in methods}
shares, rank = {}, {}
for m in methods:
tot = sum(abs(v) for v in vals[m]) or 1.0
shares[m] = [v / tot for v in vals[m]]
rank[m] = ranks_desc(vals[m])
consensus = [sum(rank[m][i] for m in methods) / len(methods) for i in range(n)]
agree = {f"{a} vs {b}": spearman(vals[a], vals[b]) for i, a in enumerate(methods) for b in methods[i + 1:]}
held = [m for m in methods if kinds[m] == "heldout"]
imp = [m for m in methods if kinds[m] == "impurity"]
shap = [m for m in methods if kinds[m] == "shap"]
flags = {f: [] for f in feats}
for i, f in enumerate(feats):
heldout_null = None
for m in held:
v = vals[m][i]
sd = stds.get(m, [0.0] * n)[i]
null = v <= 2 * sd if sd > 0 else v <= 1e-12 or shares[m][i] < 0.005
heldout_null = null if heldout_null is None else (heldout_null and null)
if shares[m][i] > 0.5:
flags[f].append(("DOMINANT", f"{shares[m][i]:.0%} of {m}: check for target leakage"))
if heldout_null:
flags[f].append(("NO-EFFECT", "held-out importance indistinguishable from zero"))
for m in imp:
if shares[m][i] >= 0.1 and held and all(shares[h][i] < shares[m][i] / 3 for h in held):
flags[f].append(("IMPURITY-ONLY", f"{shares[m][i]:.0%} of {m} but "
+ ", ".join(f"{shares[h][i]:.0%} of {h}" for h in held)
+ ": split bias (continuous/high-cardinality) or redundancy with a correlated feature"))
break
for m in shap:
if shares[m][i] >= 0.08 and held and all(shares[h][i] < shares[m][i] / 3 for h in held):
flags[f].append(("SHAP-VS-PERM", f"{shares[m][i]:.0%} of mean |SHAP| but "
+ ", ".join(f"{shares[h][i]:.0%} of {h}" for h in held)
+ ": the model uses it, but a correlated feature carries the same information"))
break
if ID_LIKE.search(f):
flags[f].append(("ID-LIKE", "identifier/hash/timestamp-like name: remove it from the features"))
order = sorted(range(n), key=lambda i: consensus[i])
return {"methods": methods, "kinds": kinds, "features": [
{"feature": feats[i], "consensus_rank": consensus[i],
"shares": {m: shares[m][i] for m in methods}, "ranks": {m: rank[m][i] for m in methods},
"flags": [{"flag": k, "detail": d} for k, d in flags[feats[i]]]} for i in order],
"agreement": agree}
def load_groups(path):
with open(path, newline="", encoding="utf-8") as f:
rows = list(csv.DictReader(f))
out = []
for g in rows:
try:
joint, single = float(g["group_permutation"]), float(g["sum_single_permutation"])
except (KeyError, ValueError):
raise ValueError("groups file needs group, group_permutation, sum_single_permutation columns")
sd = float(g.get("group_permutation_std") or 0)
shared = joint > 1.3 * max(single, 0) and joint - max(single, 0) > 2 * sd
out.append({"group": g["group"], "joint": joint, "single_sum": single, "std": sd, "shared": shared})
return out
def report(r, top):
ms = r["methods"]
w = max(12, max(len(m) for m in ms) + 2)
lines = [f"{'consensus':>9} {'feature':<22}" + "".join(f"{m:>{w}}" for m in ms) + " (share of each method)"]
for d in r["features"][:top]:
lines.append(f"{d['consensus_rank']:>9.1f} {d['feature'][:22]:<22}"
+ "".join(f"{d['shares'][m]:>{w}.1%}" for m in ms))
if len(r["features"]) > top:
lines.append(f" ... {len(r['features']) - top} more")
lines.append("")
lines.append("method agreement (Spearman):")
for k, v in r["agreement"].items():
lines.append(f" {k:<40} {v:>6.2f}")
lines.append("")
flagged = [d for d in r["features"] if d["flags"]]
for d in flagged:
for fl in d["flags"]:
lines.append(f"{fl['flag']:<14} {d['feature']}: {fl['detail']}")
if not flagged:
lines.append("no flags: the methods tell a consistent story")
for g in r.get("groups", []):
if g["shared"]:
lines.append(f"{'SHARED-INFO':<14} {g['group']}: joint permutation {g['joint']:.4g} vs {g['single_sum']:.4g} "
"summed singly: overlapping information, report the group's importance")
else:
lines.append(f"{'group':<14} {g['group']}: joint {g['joint']:.4g} vs {g['single_sum']:.4g} summed singly")
kinds = set(r["kinds"].values())
lines.append("")
if "heldout" not in kinds:
lines.append("NOTE: no held-out method (permutation or drop-column): impurity and SHAP importances describe "
"how the model uses features on training data, not what generalises")
else:
lines.append("Headline: use the held-out method's ranking; SHAP for direction; impurity only as a first look.")
return "\n".join(lines)
def main():
ap = argparse.ArgumentParser(description="Compare feature importance methods.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("csv")
ap.add_argument("--groups", help="groups.csv with grouped permutation results")
ap.add_argument("--top", type=int, default=20)
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
try:
feats, methods, vals, stds = load(a.csv)
except (OSError, ValueError) as e:
print(f"error: {e}", file=sys.stderr)
return 2
r = analyse(feats, methods, vals, stds)
if a.groups:
try:
r["groups"] = load_groups(a.groups)
except (OSError, ValueError) as e:
print(f"error: {e}", file=sys.stderr)
return 2
print(json.dumps(r, indent=2) if a.json else report(r, a.top))
bad = any(fl["flag"] in ("DOMINANT", "ID-LIKE") for d in r["features"] for fl in d["flags"])
return 1 if bad else 0
if __name__ == "__main__":
sys.exit(main())
anomaly-detection-model-builder
Build and evaluate anomaly detection for tabular, sensor and transaction data with few or no labels: define the anomaly type (point, contextual, collective), choose detectors (robust z-score/MAD and IQR rules, Mahalanobis with robust covariance, k-NN distance, LOF, Isolation Forest, one-class SVM, PCA reconstruction / autoencoders, forecasting residuals for time series), combine them by rank, explain each alert by its deviating features, set thresholds from an alert budget or the score distribution, and evaluate with precision@k, PR-AUC and recall on known incidents, or detector agreement when no labels exist. Includes pure-Python detectors (robust z, Mahalanobis, kNN, Isolation Forest, ensemble) and an evaluation script.
catboost-model-builder
Build CatBoost models that use its strengths correctly: raw categorical columns as cat_features (ordered target statistics, no manual encoding), text_features, Pool construction with NaN-safe categories, ordered boosting and has_time for temporal data, symmetric (oblivious) trees, overfitting detector with eval_set and use_best_model, losses for classification, regression (Quantile, Poisson, Tweedie) and ranking (YetiRank), class weights, monotone constraints, feature importance (PredictionValuesChange, LossFunctionChange) and SHAP values, GPU training and model export. Plans cat/text/numeric columns from a CSV and generates a runnable training script, and checks parameter sets for the conflicts CatBoost rejects (iterations vs n_estimators, Bayesian bootstrap with subsample, multiple class-weight options, depth > 16, use_best_model without eval_set).
categorical-encoding-selector
Choose and implement the right encoding for every categorical column given the model family and the data: one-hot (with infrequent-level grouping and handle_unknown), ordinal with an explicit order, cross-fitted target encoding (scikit-learn TargetEncoder, CatBoost ordered statistics, category_encoders inside CV), frequency/count encoding, feature hashing, native categorical support in LightGBM, HistGradientBoosting, XGBoost and CatBoost, prefix levels for codes (zip3, NAICS-2), and embeddings. Measures cardinality, rare-level share, unseen-level rate on a time-ordered holdout and target signal per column, emits the ColumnTransformer, and quantifies target-encoding leakage (in-sample vs out-of-fold vs holdout AUC).
classification-model-selector
Choose the right classification algorithm for tabular, text or mixed data and prove the choice with a fair bake-off: logistic regression (L2, L1, elastic net), linear SVM, naive Bayes, k-NN, RBF SVM, decision trees, random forest, extra trees, HistGradientBoosting, LightGBM, XGBoost, CatBoost, explainable boosting machines (EBM/GAM), points scorecards and MLPs, matched to data size, width, categorical and high-cardinality share, missingness, class imbalance, text, and constraints such as interpretability, monotonic constraints, calibrated probabilities, latency budgets and tiny deployable models. Ranks candidates with starting hyperparameters and cautions, then generates a bake-off script running every candidate on identical CV folds with per-fold metrics, fit time and prediction latency.
clustering-algorithm-selector
Choose the right clustering algorithm and check that clustering makes sense at all: cluster tendency (Hopkins statistic with a correlated-features caveat, nearest-neighbour isolation), then k-means / MiniBatchKMeans, Gaussian mixtures with BIC, agglomerative (Ward, average, complete, single) with dendrograms, DBSCAN, HDBSCAN, OPTICS, spectral clustering, BIRCH, mean shift, k-prototypes and k-modes for mixed or categorical data, and Gower distance, matched to data size, expected shapes, noise, varying density, known k, soft memberships, hierarchy and streaming. Covers preprocessing (scaling, encoding, dimensionality, distance metrics), validation (internal indices, bootstrap stability, external labels, business profiling) and assigning new points.
collaborative-filtering-implementer
Implement collaborative filtering recommenders that beat popularity: item-based kNN with cosine, Jaccard or asymmetric similarity, shrinkage, neighbourhood size, positive-only or mean-centred rating signals and 'because you liked X' explanations; implicit ALS with confidence weighting, BPR pairwise ranking and the closed-form EASE model in plain NumPy, tuned on a temporal validation split by recall@k and NDCG@k and refit on all data; plus user-kNN, explicit matrix factorisation, the implicit/LightFM libraries, hyperparameter ranges, cold-start fallbacks, scaling with sparse matrices and approximate nearest neighbours. Produces top-k lists in the format of the recommendation evaluator.