Upload folder using huggingface_hub

README.md

@@ -32,10 +32,10 @@ cannot tell whether last-click / linear / time-decay / Markov / Shapley attribut
 ## Files
 - `touchpoints.parquet` — one row per touchpoint: `journey_id, step, channel, t_hours, cost`.
 - `journeys.parquet` — one row per journey: `journey_id, n_touch, n_channels, converted, conv_prob`, and the ground-truth credit `gt_shapley_<channel>` (0 if the channel is absent or the journey did not convert).
-- `ground_truth.json` — the DGP parameters (channel weights, base logit, half-life) and the aggregate true channel contribution. Note: the aggregate is computed over **realized conversions** (Bernoulli draws), so it varies slightly with the seed; it is not the unconditional expected-credit share.
+- `ground_truth.json` — the DGP parameters (channel weights, base logit, half-life) and the aggregate true channel contribution.
 
 ## Channels & true weights (DGP)
-| channel | weight | arrival_rate | cost |
+| channel | weight | mix_weight | cost |
 |---|---|---|---|
 | paid_search | 1.1 | 0.9 | 1.2 |
 | social | 0.7 | 0.85 | 0.8 |
@@ -45,76 +45,21 @@ cannot tell whether last-click / linear / time-decay / Markov / Shapley attribut
 | direct | 1.0 | 0.4 | 0.0 |
 | referral | 0.5 | 0.3 | 0.1 |
 
+`mix_weight` is a **categorical mixture weight**, not an independent arrival rate.
+Each touchpoint in a journey is sampled from a single categorical draw over all channels,
+with probabilities proportional to these weights — not from independent Poisson processes
+per channel. Journeys that happen to contain a channel multiple times do so by repeated
+draws, not by a separate channel-level process.
+
 Base logit = -3.2; exposure half-life = 168h; journey window = 720h.
 Conversion ~ Bernoulli(sigmoid(base + Σ_channel weight·saturate(decayed_exposure))).
 Ground-truth credit = exact Shapley value of each channel under that value function.
 
-## Model comparison (demo notebook)
-
-`attribution_demo.ipynb` runs Last-Click, First-Click, Linear, Time-Decay and exact Shapley
-against the known ground truth. Key results (MAE = mean absolute error on channel credit share):
-
-| Model | MAE vs ground truth |
-|---|---|
-| **Shapley** | **~0.000** (reads exact labels) |
-| Last-Click | ~0.040 |
-| Time-Decay | ~0.042 |
-| Linear | ~0.042 |
-| First-Click | ~0.050 |
-
-Shapley recovers causal truth with ~15× lower MAE than Last-Click.
-
-**Key finding:** `email` is the most distorted channel across position-based heuristics —
-high causal weight (0.90), very low cost (0.05), but arrives early in the journey and gets
-starved of budget when Last-Click or Time-Decay is used.
-
-![Attribution model comparison](attribution_comparison.png)
-
-![MAE by model](mae_comparison.png)
-
-![Bias heatmap by model × channel](bias_heatmap.png)
-
-## Uniqueness
-
-As of 2026-06-14, this is the **only dataset on Hugging Face** combining:
-1. Interpretable channel names (not hashed)
-2. Exact per-journey Shapley ground truth (`gt_shapley_*` columns)
-3. Fully documented, reproducible DGP
-
-Real-world datasets (Criteo, RetailRocket) have hashed features and no ground truth,
-making it impossible to evaluate whether any attribution model is *correct*.
-
 ## Suggested uses
 - **Benchmark attribution models**: score last-click/first-click/linear/time-decay/Markov
   against `gt_shapley_*` (MAE on channel credit shares).
 - **tabular-classification**: predict `converted` from journey features.
 - Teaching material for CRO / marketing attribution.
 
-## GA4 DDA vs. Exact Shapley — comparison study
-
-`ga4_dda_comparison.ipynb` runs three GA4-style approximations against the exact Shapley
-ground truth on 500,000 journeys (`dda_comparison_results.json`).
-
-| Method | MAE vs exact Shapley | Error source |
-|---|---|---|
-| Temporal removal (GA4 modern / Shender 2023) | **0.0218** | Order bias — negligible |
-| LOO marginal | 0.0228 | Ignores synergies — negligible |
-| **Zhao 2018 (Ads Data Hub `SHAPLEY_VALUES`)** | **0.0429** | **2× worse — confuses frequency with causality** |
-
-![DDA credit comparison](credit_comparison.png)
-
-![DDA bias heatmap](bias_heatmap_dda.png)
-
-![Main finding: frequency vs causal weight](main_finding.png)
-
-**Key finding:** The order-bias hypothesis (GA4 uses temporal order instead of all
-permutations) does not explain the error. The real problem is Zhao 2018's R(S): it counts
-converting journeys that visited *all* channels in S, which conflates high-frequency channels
-with high-impact channels. `display` (freq=0.80, causal weight=0.35) is over-attributed
-by +0.095; `paid_search` (freq=0.90, causal weight=1.10) is under-attributed by -0.116.
-
-Counterintuitively, GA4's modern TTE algorithm (free product) is more accurate than the
-Ads Data Hub `SHAPLEY_VALUES` API (paid product, uses Zhao 2018).
-
 ## License
 Synthetic data, no PII. CC BY 4.0.