Proximal-Benchmark / README.md
wytzzka's picture
Upload 12 files
ad57605 verified
|
Raw History Blame Contribute Delete
18.9 kB
metadata
language:
  - en
  - zh
tags:
  - tabular
  - causal-inference
  - benchmark
  - proximal-causal-inference
  - treatment-effect-estimation
  - evaluation
pretty_name: Proximal ID10 + OOD3 CATE Benchmark
configs:
  - config_name: summary
    data_files:
      - split: test
        path: summary_metrics.csv
  - config_name: tasks
    data_files:
      - split: test
        path: task_metadata.csv
  - config_name: method_results
    data_files:
      - split: test
        path: method_results.csv

Proximal ID10 + OOD3: CATE evaluation data and results

This dataset release contains 2,000 synthetic evaluation datasets, their query-level causal-effect references, and 4,000 recorded outputs comparing pretrained ProximalFM with an RF-assisted proximal series-IV baseline. Each dataset contains 500 observed rows, binary treatment, and one evaluated query point. The estimand is

τ(x)=E[Y(1)−Y(0)∣X=x]. \tau(x)=\mathbb{E}[Y(1)-Y(0)\mid X=x].

Both methods use the same observed dataset and query. The observed inputs are covariates X, treatment proxies Z, outcome proxies W, treatment A, and outcome Y; only X is supplied for the query. Latent confounders, potential outcomes, reference effects, and reference Monte Carlo standard errors are excluded from estimator inputs.

This is a dataset and evaluation-results release. It does not contain a newly trained model, third-party source code or model weights, or the complete environment for rerunning model inference. The included script recomputes the reported aggregate metrics from the released method records. The English card accompanies the 16-page Chinese research report, revised on 2026-11-15.

Main results

The main comparison normalizes each task's error and interval length by that task's observed sample standard deviation of Y. Coverage and inference time retain their original definitions. All 2,000 paired tasks and all 4,000 method outputs are valid; no task was removed after inspecting this evaluation's effects or errors.

Evaluation group Method Tasks Norm. Bias Norm. error variance Norm. RMSE Norm. mean width Coverage Complete time (s)
ID10 overall PFM 1000 0.003962 0.048490 0.220130 0.632507 93.7% 1.193
ID10 overall RF-Bootstrap 1000 0.011580 0.703037 0.838134 2.729626 97.4% 117.440
Cui linear bridge PFM 343 0.022446 0.000925 0.037771 0.220668 99.7% 1.183
Cui linear bridge RF-Bootstrap 343 0.013362 0.109274 0.330354 1.652025 100.0% 119.644
Qi quadratic PFM 321 0.035442 0.031652 0.181133 0.384708 86.3% 1.175
Qi quadratic RF-Bootstrap 321 -0.024272 0.137375 0.370859 2.326158 100.0% 119.034
Qi trigonometric PFM 336 0.000361 0.012076 0.109728 0.499812 97.9% 1.171
Qi trigonometric RF-Bootstrap 336 0.024590 0.179907 0.424235 2.384664 100.0% 118.424

PFM denotes ProximalFM. RF-Bootstrap denotes the RF-assisted baseline with its full B=200 bootstrap. The table contains four evaluation groups: ID10 overall and the three external OOD families. Task counts are shown per method; all attempts succeeded.

The aggregate ID/OOD RMSE comparison is:

Evaluation group Tasks PFM normalized RMSE RF-Bootstrap normalized RMSE
ID10 1000 0.220130 0.838134
OOD3 combined 1000 0.122746 0.376900

Normalized RMSE, normalized interval length, and empirical coverage for the four evaluation groups

These are results for the released task distribution and the two specified estimation procedures. Lower RMSE here does not establish a ranking of all PFN or random-forest methods on all proximal inference problems. Uncertainty calibration remains a limitation: ProximalFM covers 277/321 references in the Qi quadratic family (86.3%, 95% Wilson interval 82.1%-89.6%), while the pooled OOD coverage is 94.8%. Its lowest ID10-family coverage is Softplus, 85/96 (88.5%); family comparisons are descriptive and are not adjusted for multiple testing. The RF intervals are generally wider and have higher empirical coverage.

Dataset construction and the meaning of ID/OOD

ID10: adapted training-support subfamilies

The 1,000 new ID10 datasets use the author's random MLP-SCM generator with a restricted activation construction. Each task first draws one of ten subfamilies uniformly, then draws native random hyperparameters, network weights, sparse masks, noise, and observations. Counts are random outcomes of this sampling rule, rather than fixed quotas.

ID10 subfamily Native activation construction Datasets
id_mlp_identity Identity 97
id_mlp_tanh Tanh 100
id_mlp_relu ReLU 113
id_mlp_leaky_relu LeakyReLU 101
id_mlp_elu ELU 105
id_mlp_softplus Softplus 96
id_mlp_silu SiLU 93
id_mlp_rbf RBFActivation 105
id_mlp_sine SineActivation 100
id_mlp_random_fourier RandomFunctionActivation 90

The original scaled activation factories are retained, including hidden-feature standardization and random scaling. The nominal dimensions are dX=2 and dU=dW=dZ=1; actual delivered dimensions can change through native feature filtering and can be read from the arrays' shapes in each NPZ. Each native episode generates 128 query rows; only query index 0 is evaluated. Its reference CATE is computed using K=4,096 paired Monte Carlo draws, with a saved Monte Carlo standard error.

“ID10” means a restriction and reweighting within the author's training-generator construction support. It is not the exact original training-prior mixture or its exact conditional distribution. The original layer-wise random activation mixture is replaced by ten equally weighted pure-activation subfamilies. These labels do not assert ten different causal DAGs, mutually disjoint observed distributions, or globally linear equations in the Identity family. Native categorization and effect remixing also affect the observed relationships. Nonlinear completeness remains subject to native heuristics, rather than a general identification guarantee.

The ten-family design, including the exclusion of Square, was fixed before generating and comparing this evaluation's new tasks, informed by an earlier scale diagnostic. No new task was filtered on the basis of its observed outcome scale, reference effect, or model error. The new task numbered 1854 is retained and is distinct from an older task with the same numeric index.

OOD3: external literature-generator families

The 1,000 OOD datasets are retained from the earlier frozen input collection. Their original generation first selected a family with probability 1/3, then drew continuous parameters and a dataset. Their actual counts are:

OOD family Datasets Reference CATE
Cui linear bridge 343 Analytic, constant treatment effect within a task
Qi quadratic 321 Analytic, covariate-dependent treatment effect
Qi trigonometric 336 Analytic, covariate-dependent treatment effect

The parameter hyperdistribution varies effect scale, outcome noise, covariate location/scale, propensity coefficients, bridge coefficients, and proxy correlations; see protocol.json. OOD queries are independently drawn from each task's covariate population. The analytic references have truth_mc_se=0.

“OOD” identifies external literature-derived generation mechanisms relative to the author's randomized MLP-SCM generation law. It is not a proof that their mathematical support is disjoint from the neural prior's support. The released data are synthetic evaluation contexts; the CSV rows described below are task or method records, rather than a new train/validation split.

Estimation procedures

ProximalFM. The existing pretrained checkpoint is frozen. A single in-context inference call provides the native CATE-distribution mean and the central 95% equal-tailed credible interval. There is no task-specific weight update, outer bootstrap, or cross-fitting. Obtain the model and its source from the official model repository and official code repository. The checkpoint used in the recorded evaluation is identified by its content hash in SOURCES.json.

RF-assisted proximal series-IV with full bootstrap. This is a benchmark-specific quadratic-bridge baseline combining random-forest conditional moments, a regularized bridge/GMM step, and a conditional W|X forest. It is not a direct reproduction of the published P-learner(RF). The original-table point estimate is retained; 200 IID whole-row pairs-bootstrap draws jointly resample X,Z,W,A,Y with replacement, at the fixed query. Each draw repeats preprocessing and refits the complete estimator. Interval endpoints are the 2.5% and 97.5% empirical quantiles using linear interpolation.

Only the first-stage conditional moments use two-fold treatment-stratified cross-fitting. Bootstrap copies of an original row inherit its original fold label, preventing copies of that row from appearing in opposing folds. The original split and forest algorithm seeds are fixed. Standardization, the final bridge, and the conditional W|X forest use the complete current table, so the procedure is not fully cross-fitted DML. Any failed prespecified bootstrap draw would make that task's interval unavailable, without replacement draws or success-only quantiles. The frozen evaluation has 0 failed draws out of 400,000.

These compare a native credible interval with a bootstrap confidence interval. A nominal 95% level does not by itself establish 95% frequentist coverage for either procedure.

Metric definitions

For each valid task-method record, let

ei=τ^i−τiref,si=SD⁡(Yi;ddof=1),zi=ei/si. e_i=\widehat\tau_i-\tau_i^{\mathrm{ref}},\qquad s_i=\operatorname{SD}(Y_i;\mathrm{ddof}=1),\qquad z_i=e_i/s_i.

Metric Reported definition
Normalized Bias Mean of the signed errors z_i; positive and negative errors can cancel.
Normalized error variance Sample variance of z_i, with denominator N_valid-1.
Normalized RMSE sqrt(mean(z_i**2)).
Normalized mean interval length mean((upper_i-lower_i)/s_i).
Coverage Fraction of valid intervals containing their task reference CATE.
Inference time Mean recorded complete estimation-and-interval time, in seconds.

The error variance is across evaluation tasks and their references. It is not the variance of the estimates, an individual task's bootstrap variance, or repeated-sampling variance at a fixed DGP and parameter value. Normalization is performed before aggregation; it changes the weighting of heterogeneous task scales and is not equivalent to dividing a final RMSE by one pooled standard deviation. It does not change coverage or the winner of an individual task's absolute-error comparison.

summary_metrics.csv retains both raw and normalized metrics: 32 aggregate rows covering both methods at overall, ID/OOD-group, and generator-family levels. Raw variance_error and normalized_variance_error both use denominator N_valid-1. No confidence interval is provided for the performance difference between methods.

Raw errors are highly concentrated in a few ID tasks. Task 1106 contributes 43.1% / 59.3% of ID raw squared-error sum for PFM/RF; each method's own five largest errors contribute 95.6% / 87.9%. These are SSE shares, not shares of Bias or centered error variance. The report includes common-task tail-removal and family-weighting sensitivity checks as diagnostics; they do not replace the formal results or remove any released task.

Timing and historical-result reuse

The recorded evaluation used 16 CPU workers, one Torch/BLAS thread per worker, and RF n_jobs=1. Warm inference time includes preprocessing, fitting or forward prediction, interval computation, and output extraction. RF time includes its original-table estimate and all 200 bootstrap refits. Model initialization/loading, independent warmup, input-file reads, DGP/oracle computation, and common result/checkpoint I/O are excluded. Cached bootstrap draws retain their original recorded compute time when resuming.

Of the 2,000 complete method pairs, 335 completed OOD pairs and their original timing measurements were reused, and 1,665 pairs were newly evaluated. Reused times are neither zero nor newly measured. These are complete per-task compute durations under a shared-CPU protocol, not exclusive-resource latency or the wall-clock duration of one fresh run. Timing ratios also reflect the different interval procedures.

Reference and calibration limitations

ID references retain finite paired-MC error. Their native standard errors are recorded, not treated as zero. Native feature calibration and effect remixing are batch-dependent: changing K with the same seed can change the realized SCM and is not a valid same-SCM precision experiment. Exact native replays and unit-transformation checks support computational reproducibility; they do not prove oracle convergence, nonlinear completeness, unbiased estimation, or nominal coverage. The report keeps these limitations and the mechanism-specific coverage results visible.

Files and record structure

All public files are at the repository root:

File Contents
README.md This English dataset card.
Proximal_v3report.pdf Formal 16-page Chinese report, including raw and normalized results and completed diagnostic reviews.
normalized_metrics.png Four-group normalized RMSE, interval-length, and coverage figure.
summary_metrics.csv 32 aggregate rows with raw and normalized metrics.
task_metadata.csv 2,000 tasks: group/family, query, reference effect/MC-SE, outcome scale, seed metadata, NPZ filename and NPZ-byte SHA-256.
method_results.csv 4,000 method records: point estimate, interval, reference, status, timing, bootstrap count, uncertainty type and input hash.
observed_datasets.zip 2,000 original observation archives, 0000.npz through 1999.npz, preserved byte-for-byte.
protocol.json Public evaluation, sampling, estimation and timing definitions.
SOURCES.json External source references and recorded asset identities.
recompute_metrics.py Offline aggregation and consistency check using the Python standard library.
FILES_SHA256.json Public-file integrity inventory.
PUBLIC_RELEASE.json Public-release scope and provenance metadata.

Each observation NPZ contains only X, Z, W, A, Y, and query_x. The first three arrays are observed feature/proxy matrices with 500 rows; A and Y are length-500 vectors; query_x is the evaluated covariate vector. Its corresponding reference effect and MC-SE are in the CSV records, rather than additional model-input columns. Even task indices are the new ID10 collection; odd indices are the retained OOD collection.

method_results.csv uses method IDs proximalfm and rf_proximal_sieve_iv_bootstrap. Coverage is stored as a fraction in the summary CSV; the tables in this card display percentages. query_x in the task CSV is a JSON-array string. observed_npz_sha256 identifies the NPZ file bytes, whereas input_sha256 identifies the canonical input arrays and query; these are different hash definitions. Seeds and hashes identify the frozen tasks; native parameter and observation draws share a recorded joint RNG stream, rather than two independently controlled native streams.

The public selection does not duplicate the complete B200 draw trajectories or checkpoint/recovery logs. Those remain in the original audit archive. The files here support inspection of the released observations, recorded predictions/intervals, and their aggregation; they do not supply the full inference implementation or independently reconstruct the saved bootstrap intervals.

Recompute the aggregate metrics offline

Download the public files into one directory. With Python available, run:

python recompute_metrics.py

This uses only the Python standard library: no NumPy, model weights, Torch, random forests, or model-execution environment is required. It reads method_results.csv, recomputes the aggregate metrics, compares all 32 rows field by field with summary_metrics.csv, and prints a pass result if they agree. The default command does not write files, generate data, fit models, rebuild intervals, or remeasure time.

To explicitly export the derived table:

python recompute_metrics.py --output derived_metrics.csv

For observation-level analysis, extract observed_datasets.zip. Loading the NPZ arrays separately requires an NPZ-capable reader such as NumPy; use allow_pickle=False. This optional observation analysis is separate from the standard-library aggregate check.

Sources and attribution

This release evaluates an existing model and literature-derived synthetic generators. It does not claim authorship of those models or papers, and does not redistribute their source code or weights. Source identities and links are recorded in SOURCES.json.

The original model and external source materials retain their respective terms. No license for third-party components is assigned by this dataset card.