datadoom 0.1.0.dev0__py3-none-any.whl

datadoom/templates/telecom_churn_challenge.datadoom.yaml

@@ -0,0 +1,125 @@
+datadoom_version: "1"
+name: "telecom-churn-challenge"
+description: >
+  HACKATHON — telecom customer churn with realistic records and mixed feature
+  types. Believable identity fields (customer name, email, city, signup date —
+  generated by deterministic providers) sit alongside the real signal: tenure,
+  monthly charges, support-call load, contract type and data usage feed a LATENT
+  `dissatisfaction` score (emit: false) behind the `churned` label. Calibrated to
+  the demanding 'kaggle' baseline-AUROC band [0.62, 0.72] — a genuinely hard
+  churn problem. A light data-quality stack corrupts a copy: heavy-usage records
+  under-report (MNAR), monthly charges have random gaps (MCAR), and 3% of churn
+  labels are noisy. The identity columns are pure realism — leaking no signal,
+  they're a feature-selection / PII-hygiene check (drop them).
+seed: 404
+rows: 8000
+
+features:
+  customer_name:
+    type: text
+    generator: name
+    locale: en
+  email:
+    type: text
+    generator: email
+    locale: en
+  city:
+    type: text
+    generator: city
+    locale: en
+  signup_date:
+    type: datetime
+    start: "2019-01-01"
+    end: "2023-12-31"
+    granularity: day
+  plan:
+    type: categorical
+    categories: [prepaid, basic, plus, premium]
+    weights: [0.2, 0.35, 0.3, 0.15]
+  contract:
+    type: categorical
+    categories: [monthly, annual, two_year]
+    weights: [0.6, 0.25, 0.15]
+  tenure_months:
+    type: numeric
+    dist: normal
+    params: { mean: 30, std: 18 }
+    min: 0
+    dtype: int
+  monthly_charges:
+    type: numeric
+    dist: normal
+    params: { mean: 65, std: 25 }
+    min: 0
+  num_support_calls:
+    type: numeric
+    dist: poisson
+    params: { lam: 1.8 }
+    min: 0
+    dtype: int
+  data_usage_gb:
+    type: numeric
+    dist: lognormal
+    params: { mu: 2.4, sigma: 0.6 }   # median ≈ 11 GB, right-skewed
+    min: 0
+  dissatisfaction:
+    type: numeric          # LATENT churn propensity — the true driver
+    dtype: float
+    emit: false
+  churned:
+    type: boolean          # label (derived): logistic of dissatisfaction
+
+causal:
+  edges:
+    - { from: tenure_months,     to: dissatisfaction, fn: linear, weight: -0.03 }
+    - { from: monthly_charges,   to: dissatisfaction, fn: linear, weight: 0.02 }
+    - { from: num_support_calls, to: dissatisfaction, fn: linear, weight: 0.42 }
+    - { from: data_usage_gb,     to: dissatisfaction, fn: linear, weight: -0.02 }
+    - { from: contract,          to: dissatisfaction, fn: map,    mapping: { monthly: 1.0, annual: -0.3, two_year: -1.0 } }
+    - { from: dissatisfaction,   to: churned, fn: logistic, weight: 1.2, bias: -3.5 }
+  noise:
+    dissatisfaction: { dist: none }
+    churned: { dist: none }
+
+difficulty:
+  target: kaggle            # baseline ROC-AUC calibrated into [0.62, 0.72]
+  label: churned
+  probe: logreg
+  max_iters: 12
+  knobs: [noise, label_noise]
+
+failures:
+  - type: mnar               # heavy users under-report usage
+    column: data_usage_gb
+    rate: 0.08
+    strength: 2.0
+  - type: mcar               # billing gaps
+    columns: [monthly_charges]
+    rate: 0.04
+  - type: label_noise        # 3% noisy churn flags
+    column: churned
+    rate: 0.03
+
+export:
+  formats: [csv]
+  versions: [clean, injected]
+  splits: { train: 0.7, test: 0.3 }
+
+meta:
+  level: hackathon
+  challenge:
+    title: "Predict telecom customer churn"
+    task: classification
+    target: churned
+    metric: ROC-AUC
+    difficulty: kaggle
+    baseline_auroc_band: [0.62, 0.72]
+    train_test_split: "70 / 30"
+    hidden_structure: >
+      dissatisfaction is latent. tenure (−), charges (+), support calls (+),
+      usage (−) and contract type drive it; it drives churn.
+    gotchas:
+      - "customer_name / email / city are realistic but carry NO signal — drop these identifier columns."
+      - "data_usage_gb is MNAR (heavy users under-report) — missingness itself is informative."
+      - "Calibrated to the kaggle band — expect a hard ceiling around 0.72 AUROC; chasing higher means overfitting noise."
+  tags: [telecom, churn, causal, latent, difficulty, realistic-text, missingness, classification]

datadoom/version.py

@@ -0,0 +1,3 @@
+"""Single source of truth for the package version (read by hatchling)."""
+
+__version__ = "0.1.0.dev0"