#!/usr/bin/env python3
"""Item-based collaborative filtering (item-kNN) with shrinkage, producing top-k lists with explanations.
Input: training interactions CSV (user_id,item_id[,rating]).
Signal:
implicit (default) every interaction counts 1 (repeats add up with --count-repeats)
--min-rating R explicit ratings: only ratings >= R count as positive (weight 1); others are ignored
--weighted use the rating / value itself as the weight (mean-centred per user for explicit data)
Similarity between items i and j (over users):
cosine dot(i, j) / (|i| |j|)
jaccard |users(i) & users(j)| / |users(i) | users(j)|
asymmetric dot(i, j) / (|i|^(2*alpha) |j|^(2*(1-alpha))) with --alpha (0.5 = cosine), which
favours or penalises popular neighbours
then shrunk towards 0 for item pairs with few common users: sim * n / (n + --shrink).
Each item keeps its --neighbors most similar items. A user's score for an unseen item is the sum of the
similarities to the items they interacted with (weighted); the item contributing most is the explanation
("because you liked ...").
Output: CSV user_id,item_id,rank,score,because (the rec_eval.py format), --k items per user, seen items
excluded. Users with no positive history get the most popular items (marked "popular").
Usage:
python itemknn.py train.csv --k 10 -o recs.csv
python itemknn.py ratings_train.csv --min-rating 4 --similarity cosine --shrink 10 --neighbors 50 -o recs.csv
python itemknn.py events.csv --user-col visitor --item-col product --similarity asymmetric --alpha 0.7
Exit codes: 0 written, 2 usage or read error.
"""
import argparse
import csv
import heapq
import math
import sys
from collections import Counter, defaultdict
def num(v):
try:
x = float(v)
return x if math.isfinite(x) else None
except (TypeError, ValueError):
return None
def build(rows, a):
users = defaultdict(dict)
for r in rows:
u, i = r[a.user_col], r[a.item_col]
if a.min_rating is not None:
v = num(r.get(a.rating_col))
if v is None or v < a.min_rating:
continue
w = 1.0
elif a.weighted:
w = num(r.get(a.rating_col)) or 1.0
else:
w = 1.0
if a.count_repeats or a.weighted:
users[u][i] = users[u].get(i, 0.0) + w if a.count_repeats else w
else:
users[u][i] = 1.0
if a.weighted and a.min_rating is None and a.center:
for u, d in users.items():
m = sum(d.values()) / len(d)
for i in d:
d[i] -= m
return users
def similarities(users, a):
items = defaultdict(dict)
for u, d in users.items():
for i, w in d.items():
items[i][u] = w
norm = {i: math.sqrt(sum(w * w for w in d.values())) for i, d in items.items()}
count = {i: len(d) for i, d in items.items()}
dot = defaultdict(float)
common = defaultdict(int)
for u, d in users.items():
lst = list(d.items())
for x in range(len(lst)):
i, wi = lst[x]
for y in range(x + 1, len(lst)):
j, wj = lst[y]
key = (i, j) if i < j else (j, i)
dot[key] += wi * wj
common[key] += 1
neigh = defaultdict(list)
for (i, j), d in dot.items():
n = common[(i, j)]
if a.similarity == "jaccard":
s = n / (count[i] + count[j] - n)
sij = sji = s
elif a.similarity == "asymmetric":
sij = d / ((norm[i] ** (2 * a.alpha)) * (norm[j] ** (2 * (1 - a.alpha))) or 1.0)
sji = d / ((norm[j] ** (2 * a.alpha)) * (norm[i] ** (2 * (1 - a.alpha))) or 1.0)
else:
s = d / (norm[i] * norm[j]) if norm[i] and norm[j] else 0.0
sij = sji = s
shrink = n / (n + a.shrink) if a.shrink else 1.0
neigh[i].append((sij * shrink, j))
neigh[j].append((sji * shrink, i))
return {i: heapq.nlargest(a.neighbors, lst) for i, lst in neigh.items()}, items
def recommend(users, neigh, popular, a, all_users):
out = []
for u in all_users:
hist = users.get(u, {})
scores, because = defaultdict(float), {}
best_contrib = defaultdict(float)
for j, w in hist.items():
for s, i in neigh.get(j, []):
if i in hist or s <= 0:
continue
c = s * w
scores[i] += c
if c > best_contrib[i]:
best_contrib[i], because[i] = c, j
ranked = heapq.nlargest(a.k, scores.items(), key=lambda kv: kv[1])
seen = set(a.seen.get(u, ())) | set(hist)
if len(ranked) < a.k: # fill with popular unseen items
have = {i for i, _ in ranked}
for i in popular:
if len(ranked) >= a.k:
break
if i not in have and i not in seen:
ranked.append((i, 0.0))
because[i] = "popular"
for rank, (i, s) in enumerate(ranked, 1):
out.append((u, i, rank, round(s, 6), because.get(i, "popular")))
return out
def main():
ap = argparse.ArgumentParser(description="Item-kNN collaborative filtering.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("train")
ap.add_argument("--user-col", default="user_id")
ap.add_argument("--item-col", default="item_id")
ap.add_argument("--rating-col", default="rating")
ap.add_argument("--min-rating", type=float)
ap.add_argument("--weighted", action="store_true")
ap.add_argument("--no-center", dest="center", action="store_false", help="do not mean-centre --weighted ratings")
ap.add_argument("--count-repeats", action="store_true")
ap.add_argument("--similarity", choices=["cosine", "jaccard", "asymmetric"], default="cosine")
ap.add_argument("--alpha", type=float, default=0.5)
ap.add_argument("--shrink", type=float, default=10.0)
ap.add_argument("--neighbors", type=int, default=50)
ap.add_argument("--k", type=int, default=10)
ap.add_argument("--users", help="CSV with a user_id column: only recommend for these users")
ap.add_argument("-o", "--output")
a = ap.parse_args()
try:
with open(a.train, newline="", encoding="utf-8") as f:
rows = list(csv.DictReader(f))
if not rows:
raise ValueError("no rows")
for c in (a.user_col, a.item_col):
if c not in rows[0]:
raise ValueError(f"column {c!r} not found (have: {', '.join(rows[0])})")
if (a.min_rating is not None or a.weighted) and a.rating_col not in rows[0]:
raise ValueError(f"rating column {a.rating_col!r} not found")
target = None
if a.users:
with open(a.users, newline="", encoding="utf-8") as f:
target = sorted({r[a.user_col] for r in csv.DictReader(f)})
except (OSError, ValueError) as e:
print(f"error: {e}", file=sys.stderr)
return 2
a.seen = defaultdict(set)
for r in rows:
a.seen[r[a.user_col]].add(r[a.item_col])
users = build(rows, a)
neigh, items = similarities(users, a)
popular = [i for i, _ in Counter(r[a.item_col] for r in rows).most_common()]
all_users = target or sorted(a.seen)
recs = recommend(users, neigh, popular, a, all_users)
f = open(a.output, "w", newline="", encoding="utf-8") if a.output else sys.stdout
w = csv.writer(f)
w.writerow(["user_id", "item_id", "rank", "score", "because"])
w.writerows(recs)
if a.output:
f.close()
n_pop = sum(1 for r in recs if r[4] == "popular")
print(f"wrote {a.output}: {len(all_users)} users x {a.k}, {len(neigh)} items with neighbours, "
f"{n_pop} popularity fill-ins", file=sys.stderr)
return 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.
counterfactual-explanation-generator
Generate counterfactual explanations and algorithmic recourse: the smallest realistic change to a row's features that changes the model's decision. Solves exact minimal-cost recourse for linear and logistic models (MAD-scaled L1 cost, immutable features, increase/decrease directions, bounds, integer and binary features, diverse alternatives, features every counterfactual must change, and a no-recourse finding when immutable features decide), generates DiCE counterfactuals for any scikit-learn model with constraints turned into permitted ranges and every result re-predicted, and validates counterfactuals from any method for constraint violations, implausible values, off-manifold rows, broken derived-feature relations and failure to flip the prediction.