Contrastive Nonmyopic Objective Cost-Tradeoff Acquisition for Longitudinal Data
Organizations: Department of Computer Science, UNC Chapel Hill · Department of Computer Science and Engineering, University of California San Diego
Abstract
In many critical domains, features are not freely available at inference time: each measurement may come with a cost of time, money, and risk. Longitudinal prediction further complicates this setting because both features and labels evolve over time, and measurements missed at earlier timepoints may become permanently unavailable. We propose CAPE (Contrastive Acquisition-Plan Embeddings), a partitioning acquisition policy for longitudinal active feature acquisition. CAPE first defines NOCT, an oracle-based objective that scores a set of future feature-time acquisitions by its expected predictive loss together with its acquisition cost. Based on NOCT, CAPE learns a contrastive acquisition-utility manifold of masked partial observations, where proximity reflects the expected utility of future acquisitions. Offline, training states are embedded and partitioned at each rollout timepoint; candidate future plans are scored within each partition using plug-in NOCT, and the best partition-level acquisition plans are cached. During inference, CAPE retrieves the cached plan via nearest-centroid lookup, enabling low-latency adaptive acquisitions. Experiments on synthetic and real-world healthcare datasets demonstrate that CAPE outperforms nonparametric, greedy, and RL-based alternatives, achieving higher accuracy at lower acquisition costs. Code is publicly available at https://github.com/dzungdinhh/CAPE.
Figures & tables
| Method | Synthetic | KLG | WOMAC | ILIADD | ||||
| Acc. | cost | AUROC | cost | AUROC | cost | AUROC | cost | |
| Feature Partitioning | ||||||||
| CAPE | ||||||||
| Dataset | W/o perturb. AUROC | W/ perturb. AUROC | AUROC | W/o perturb. cost | W/ perturb. cost | Cost |
| KLG | 0.8164 | 0.8101 | -0.0063 | 9.390 | 9.060 | -0.330 |
| WOMAC | 0.6207 | 0.6246 | 0.0039 | 11.447 | 11.251 | -0.196 |
| ILIADD | 0.8269 | 0.8258 | -0.0011 | 65.126 | 65.010 | -0.116 |
| Dataset | Inliers AUROC | Outliers AUROC | AUROC | Inliers cost | Outliers cost | Cost |
| KLG | 0.8152 | 0.8188 | 0.0037 | 9.522 | 8.861 | -0.661 |
| WOMAC | 0.6151 | 0.6059 | -0.0091 | 11.578 | 10.922 | -0.656 |
| ILIADD | 0.8309 | 0.8145 | -0.0164 | 65.617 | 63.160 | -2.456 |
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | Synthetic | ADNI | WOMAC | KLG | ILIADD | |||||
| Acc. | cost | AUROC | cost | AUROC | cost | AUROC | cost | AUROC | cost | |
| Feature Partitioning | ||||||||||
| CAPE | ||||||||||
| Cost regime | ADNI AUROC | ADNI Cost | WOMAC AUROC | WOMAC Cost | KLG AUROC | KLG Cost | ILIADD AUROC | ILIADD Cost | |
| 0.5 | Low | 0.7468 | 2.788 | 0.6236 | 8.981 | 0.8197 | 9.359 | 0.7926 | 19.543 |
| 1 | Low | 0.7488 | 2.768 | 0.6201 | 8.980 | 0.8168 | 9.638 | 0.7931 | 19.780 |
| 2 | Low | 0.7507 | 2.772 | 0.6224 | 8.822 | 0.8188 | 9.984 | 0.7956 | 19.338 |
| 0.5 | High | 0.7539 | 11.295 | 0.6256 | 18.191 | 0.8363 | 38.084 | 0.8453 | 89.853 |
| 1 | High | 0.7547 | 12.002 | 0.6241 | 17.245 | 0.8353 | 38.299 | 0.8454 | 90.817 |
| 2 | High | 0.7548 | 12.105 | 0.6254 | 17.457 | 0.8351 | 37.550 | 0.8459 | 89.042 |
| Cost regime | ADNI AUROC | ADNI Cost | WOMAC AUROC | WOMAC Cost | KLG AUROC | KLG Cost | ILIADD AUROC | ILIADD Cost | |
| 0.5 | Low | 0.7495 | 2.775 | 0.6217 | 8.966 | 0.8177 | 9.272 | 0.7941 | 19.425 |
| 1 | Low | 0.7487 | 2.760 | 0.6230 | 8.781 | 0.8164 | 9.390 | 0.7933 | 19.582 |
| 2 | Low | 0.7506 | 2.772 | 0.6228 | 9.055 | 0.8165 | 9.574 | 0.7965 | 19.868 |
| 0.5 | High | 0.7607 | 11.785 | 0.6250 | 18.074 | 0.8310 | 33.251 | 0.8448 | 90.956 |
| 1 | High | 0.7520 | 11.640 | 0.6248 | 17.484 | 0.8374 | 37.567 | 0.8460 | 88.172 |
| 2 | High | 0.7515 | 11.845 | 0.6231 | 17.391 | 0.8341 | 33.699 | 0.8459 | 88.761 |
| Cost regime | ADNI AUROC | ADNI Cost | WOMAC AUROC | WOMAC Cost | KLG AUROC | KLG Cost | ILIADD AUROC | ILIADD Cost | |
| 16 | Low | 0.7487 | 2.763 | 0.6192 | 9.024 | 0.8195 | 9.766 | 0.7937 | 19.496 |
| 32 | Low | 0.7488 | 2.768 | 0.6235 | 8.965 | 0.8101 | 8.845 | 0.7956 | 19.441 |
| 64 | Low | 0.7487 | 2.760 | 0.6183 | 8.964 | 0.8178 | 9.582 | 0.7940 | 19.454 |
| 128 | Low | 0.7537 | 2.772 | 0.6234 | 8.902 | 0.8147 | 9.284 | 0.7980 | 19.328 |
| 16 | High | 0.7579 | 11.490 | 0.6232 | 17.288 | 0.8322 | 37.052 | 0.8460 | 88.529 |
| 32 | High | 0.7537 | 11.922 | 0.6268 | 16.859 | 0.8353 | 39.903 | 0.8460 | 89.421 |
| Cost regime | ADNI AUROC | ADNI Cost | WOMAC AUROC | WOMAC Cost | KLG AUROC | KLG Cost | ILIADD AUROC | ILIADD Cost | |
| 3 | Low | 0.7493 | 2.808 | 0.6197 | 8.853 | 0.8134 | 9.190 | 0.7948 | 19.365 |
| 5 | Low | 0.7488 | 2.768 | 0.6254 | 9.104 | 0.8169 | 9.294 | 0.7918 | 19.248 |
| 10 | Low | 0.7468 | 2.788 | 0.6237 | 8.943 | 0.8189 | 9.719 | 0.7945 | 19.552 |
| 3 | High | 0.7545 | 11.700 | 0.6209 | 17.150 | 0.8359 | 36.646 | 0.8464 | 89.641 |
| 5 | High | 0.7551 | 11.935 | 0.6265 | 17.277 | 0.8374 | 38.735 | 0.8466 | 87.570 |
| 10 | High | 0.7545 | 11.787 | 0.6251 | 17.433 | 0.8354 | 35.513 | 0.8425 | 85.558 |
| Synthetic Acc. | Synthetic Cost | ADNI AUROC | ADNI Cost | KLG AUROC | KLG Cost | ILIADD AUROC | ILIADD Cost | |
| 100 | 0.8709 | 14.498 | 0.7485 | 11.772 | 0.8348 | 42.087 | 0.8367 | 105.824 |
| 500 | 0.8673 | 14.113 | 0.7505 | 11.312 | 0.8349 | 35.249 | 0.8411 | 95.380 |
| 1,000 | 0.8805 | 13.780 | 0.7498 | 11.677 | 0.8381 | 40.366 | 0.8401 | 94.587 |
| 5,000 | 0.8845 | 13.647 | 0.7520 | 11.640 | 0.8374 | 37.567 | 0.8460 | 88.172 |
| 10,000 | 0.8838 | 13.865 | 0.7503 | 11.738 | 0.8364 | 34.542 | 0.8481 | 86.389 |
| RAS AUROC | RAS Cost | ASAC AUROC | ASAC Cost | AS AUROC | AS Cost | ACO AUROC | ACO Cost | CAPE AUROC | CAPE Cost | ||
| 10 | 34 | 0.5780 | 41.097 | 0.6328 | 28.600 | 0.7415 | 204.497 | 0.7789 | 53.147 | 0.7862 | 16.672 |
| 20 | 34 | 0.7507 | 283.176 | 0.6309 | 109.795 | 0.7308 | 109.639 | 0.7517 | 111.830 | 0.8294 | 42.338 |
| 30 | 34 | 0.7579 | 441.695 | 0.6710 | 110.000 | 0.7224 | 109.756 | 0.7468 | 110.168 | 0.8376 | 56.791 |
| 50 | 34 | 0.7492 | 1508.057 | 0.5641 | 40.645 | 0.6969 | 109.785 | 0.6872 | 33.475 | 0.7621 | 25.844 |
| 20 | 68 | 0.7146 | 442.117 | 0.6592 | 109.992 | 0.6567 | 187.338 | 0.6539 | 101.002 | 0.7764 | 49.605 |
| 20 | 102 | 0.6555 | 725.170 | 0.6372 | 109.930 | 0.6008 | 267.578 | 0.6618 | 54.408 | 0.6912 | 52.451 |
| Dataset | Search grid for | Selected |
| Synthetic | ||
| ADNI | ||
| WOMAC | ||
| KLG | ||
| ILIADD |
Active Feature Acquisition
Active Feature Acquisition studies how models can selectively obtain costly or missing inputs to improve predictions within limited budgets.
See all 22 papers