house-prices-tabular-models / INTERPRETABILITY.md
t22000t's picture
v3: add SHAP + Captum IG interpretability artefacts
df1ec6d verified
|
Raw
History Blame Contribute Delete
4.47 kB

Interpretability summary: house_prices_8arch_interp

Performance ranking

Rank Model Test Gini Test MAE A/E ratio n params Train time
1 xgboost 0.2049 17203.89 0.999 462 0.4s
2 stacked_ensemble 0.2049 17203.89 0.999 9 0.0s
3 catboost 0.1996 29223.45 1.161 499 2.7s
4 localglmnet 0.1991 23419.59 0.988 22,620 6.3s
5 drn 0.1962 27927.62 0.981 53,010 6.6s
6 cann 0.1941 24906.06 1.024 52,815 6.5s
7 cann_gbm 0.1940 32932.26 1.193 52,815 5.8s
8 ft_transformer 0.0368 187771.09 3337.260 483,267 340.8s
9 tabm 0.0331 187802.11 7436.716 410,364 176.6s

Top-10 features per architecture

Rank catboost xgboost
1 TotalSF (0.096) TotalSF (0.138)
2 OverallQual (0.043) OverallQual (0.103)
3 GrLivArea (0.043) GrLivArea (0.032)
4 YearRemodAdd (0.038) YearRemodAdd (0.025)
5 LotArea (0.030) OverallCond (0.025)
6 KitchenQual (0.028) HouseAge (0.021)
7 OverallCond (0.028) GarageCars (0.020)
8 MSZoning (0.022) LotArea (0.019)
9 HouseAge (0.022) YearBuilt (0.012)
10 GarageCars (0.021) KitchenQual (0.011)

Cross-method agreement

Features that appear in top-5 across every model with importance scores:

  • GrLivArea

  • OverallQual

  • TotalSF

  • YearRemodAdd

This cross-method agreement is a strong signal - when both tree-based SHAP and gradient-based Captum IG identify the same feature as critical, the finding is unlikely to be a method-specific artefact.

LocalGLMnet coefficient analysis

LocalGLMnet emits one row of coefficients per test record. We summarise the distribution of each feature's coefficient across the sampled test set (mean ± std):

Feature Mean coef Std Sign stability
LotArea 2.912e-04 1.847e-04 98%
YearBuilt 3.377e-04 2.296e-04 99%
YearRemodAdd 3.083e-04 2.715e-04 97%
TotalBsmtSF 2.400e-04 1.464e-04 99%
1stFlrSF 2.361e-04 1.171e-04 100%
2ndFlrSF -2.467e-04 2.454e-04 96%
GrLivArea -5.616e-05 1.236e-04 72%
FullBath 1.693e-04 2.968e-04 86%
BedroomAbvGr -9.640e-05 1.435e-04 80%
TotRmsAbvGrd -1.148e-04 1.200e-04 84%
GarageCars 3.736e-04 4.435e-04 98%
GarageArea 2.725e-04 1.512e-04 99%
OverallQual 1.848e-04 1.240e-04 98%
OverallCond 2.792e-04 3.282e-04 99%
TotalSF 2.274e-04 1.413e-04 99%
HouseAge -3.340e-04 2.015e-04 100%
LotArea_raw 3.759e-08 2.385e-08 98%
YearBuilt_raw 1.106e-05 7.519e-06 99%
YearRemodAdd_raw 1.493e-05 1.314e-05 97%
TotalBsmtSF_raw 5.483e-07 3.345e-07 99%
1stFlrSF_raw 6.213e-07 3.082e-07 100%
2ndFlrSF_raw -5.686e-07 5.655e-07 96%
GrLivArea_raw -1.075e-07 2.365e-07 72%
FullBath_raw 3.079e-04 5.398e-04 86%
BedroomAbvGr_raw -1.181e-04 1.757e-04 80%
TotRmsAbvGrd_raw -7.042e-05 7.359e-05 84%
GarageCars_raw 5.017e-04 5.955e-04 98%
GarageArea_raw 1.298e-06 7.199e-07 99%
OverallQual_raw 1.338e-04 8.976e-05 98%
OverallCond_raw 2.500e-04 2.938e-04 99%
TotalSF_raw 2.768e-07 1.720e-07 99%
HouseAge_raw -1.091e-05 6.583e-06 100%

Sign stability is the fraction of test records where the coefficient has the same sign as the mean. Values close to 100% mean the model is confident about that feature's direction; values closer to 50% mean the feature's effect flips across records (which is exactly what LocalGLMnet was designed to detect).

Artefacts on disk

  • dashboard_dl_interpretability.html - interactive Plotly dashboard with SHAP beeswarm plots, Captum IG heatmaps, FT-Transformer attention matrices, CANN residual histograms.

  • localglmnet_coefficients.csv - per-test-record coefficients from LocalGLMnet (one row per test record, one column per continuous feature).

  • feature_importance.csv - consolidated importances (CatBoost / XGBoost native importance scores).

  • drn_distributional_outputs.csv - DRN's predictive distribution moments (mean, variance, quantiles) per test row.