acceptodds
Under review as a conference paper at ICLR 2027

Forget Before You Connect: Pre-Graph Subspace Projection for Channel-Level Unlearning in Multivariate Time Series

Abstract

Machine unlearning has largely focused on removing training samples or classes, leaving one question underexplored: *How can a deployed model reduce its dependence on an input channel?* In multivariate time-series (MTS) systems, corrupted or privacy-sensitive sensors may need to be removed after deployment. Existing approaches require costly retraining or iterative optimization, while permanent input masking can degrade predictive utility without addressing learned cross-channel dependencies. We introduce *Channel-Centric Subspace Projection* (CCSP), a training-free method that identifies target-associated latent directions via singular value decomposition and attenuates them before cross-channel graph propagation. An optional trainable utility-repair extension is evaluated separately. Across three MTS classification benchmarks, CCSP preserves near-original utility with low residual sensitivity, achieving speedup over iterative baselines. On UCI-HAR, it achieves a mean sensor-wise reduction in linear intervention leakage and reduces mean prediction disagreement from to . On C-MAPSS remaining-useful-life regression, CCSP further reduces target-induced prediction variation by on average across sensors while maintaining comparable predictive utility. These results establish latent subspace intervention as an efficient approach to channel-level forgetting, while distinguishing approximate influence removal from certified information erasure.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Machine Unlearning, Multivariate Time Series, Channel-Level Unlearning, Subspace Projection, Graph Neural Network

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cyibOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CCSP reduces a trained time-series model’s dependence on a selected sensor by estimating dominant directions in that sensor’s pre-graph activations and attenuating them before message passing. The core method keeps model parameters fixed; optional trainable repair is evaluated separately. The paper reports classification results on HAR, NATOPS, and FaceDetection, with additional real-trajectory intervention tests, simultaneous channel edits, and remaining-useful-life regression. Its stated objective is approximate functional dependence reduction, with the target trajectory still available, rather than certified information erasure.

Critical strengths
  1. Editing the target slice before message passing is a sensible and inexpensive intervention point. Retained slices are unchanged at that point, and the core method avoids parameter optimization. The pre/post comparison in Appendix B.5 supports the placement choice. Mean per-sensor runtimes on an A100 are 1.48, 1.98, and 12.20 seconds for HAR, NATOPS, and FaceDetection, respectively (Table 5), making the method practically interesting as a post-training attenuation tool.

  2. The paper goes beyond accuracy differences under zero masking. On HAR, the five-permutation test reduces target-channel prediction disagreement from 9.60% to 8.08%, while retained-channel disagreement changes from 9.60% to 9.87% (Table 14). The intervention-detectability probes and regression extension provide additional evidence of reduced functional sensitivity. The explicit separation between partial dependence reduction and information erasure keeps the conclusions reasonably aligned with these measurements.

Critical weaknesses

The comparison does not yet isolate the benefit of the SVD directions from that of localized, validation-tuned attenuation. Table 3 provides useful evidence: CCSP has lower intervention leakage than INLP-1 (0.0480 versus 0.0670) at nearly identical test accuracy (94.76% versus 94.72%). However, their edit rules differ: CCSP attenuates 1–4 of 32 directions, whereas INLP-1 and INLP-final remove 8 and 32 directions. Appendix B.2/B.4 varies CCSP’s rank and strength, but the manuscript does not report same-rank random-subspace attenuation or a sweep of partial uniform scaling of the target latent slice. These controls would help separate the value of the SVD construction from the broader benefit of a localized, tunable intervention. The specific subspace choice remains less well supported than the practical attenuation method.

Questions for authors
  1. Could the existing shared-donor protocol be extended to same-rank random-subspace attenuation and partial uniform scaling of the target latent slice? With comparable validation search budgets, please compare prediction disagreement or probability total variation at matched utility, using configurations fixed before test evaluation.

  2. Could the same dependence–utility comparison include partial-strength versions of the fitted LEACE and INLP edits, for example by interpolating each edit with the identity and varying INLP iterations where applicable? This would help determine whether CCSP’s subspace construction improves the trade-off beyond the reported operating points.

Reviewer_FNgvOpenReview ↗
Rating
2
Confidence
1(−2)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a training-free method, called CCSP, for removing a sensor’s influence from a trained graph-based time-series model. It applies SVD to the sensor’s representations before graph propagation and dampens the top directions, and the authors report that utility stays close to the original while the method runs much faster than retraining.

Critical strengths

The problem is practically motivated and relatively underexplored, and the method is simple and inexpensive. Also, the paper is commendably transparent about its scope, explicitly distinguishing approximate dependence reduction from certified erasure.

Critical weaknesses

My main concern is that the formulation and evaluation do not fully support the claimed channel-level unlearning.

  1. The paper is motivated by removed, corrupted, or privacy-sensitive sensors, yet CCSP still requires the target trajectory at inference and merely attenuates its latent representation. It therefore cannot handle actual sensor absence or eliminate access to sensitive measurements. This also weakens the practical significance of its speedup comparisons with retraining.

  2. The primary metric, (|G|), measures only the aggregate accuracy gap under zero-masking and is also directly minimized during hyperparameter selection. A near-zero gap does not imply that individual predictions are independent of the target sensor, making it an insufficient basis for claiming effective unlearning.

  3. Under real-trajectory replacement on HAR, mean prediction disagreement only decreases from 9.54% to 8.12%. For sensor 6, it barely changes. Linear intervention leakage also remains (0.0813 to 0.0480). These results suggest partial sensitivity suppression rather than convincing channel-level forgetting.

Questions for authors
  1. The paper frames CCSP as channel-level unlearning, but the method still requires the target sensor at inference and does not establish removal of its learned training influence. What distinguishes CCSP from inference-time feature suppression, and which of the motivating sensor-removal or privacy scenarios can it actually address?

  2. How do the authors reconcile near-zero (|G|) with substantial residual PDR? What happens if hyperparameters are selected using PDR or probe leakage instead?

  3. How does CCSP compare with training-free alternatives, such as mean/median imputation, in terms of the forgetting–utility trade-off under real-trajectory interventions?

Reviewer_P85ZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Channel-Centric Subspace Projection (CCSP), a channel-level unlearning method for multivariate time series. The method employs singular value decomposition (SVD) to identify the dominant directions in the target sensor's latent representations and attenuates these directions before graph message passing, aiming to reduce the model's dependence on the specified sensor while preserving predictive performance. Experiments are conducted on three time-series classification datasets and further extended to multi-channel interventions and remaining useful life regression.The problem addressed in this paper is practically relevant, and the proposed method is simple and computationally efficient. However, the current work has notable limitations regarding the theoretical motivation behind its core design, the evaluation of unlearning effectiveness, and comparisons with key baselines. In particular, the paper does not sufficiently demonstrate that attenuating dominant SVD directions offers meaningful advantages over simpler alternatives, such as uniform feature scaling or random subspace projection.

Critical strengths

The paper addresses channel-level unlearning, a relatively novel problem with practical relevance. The proposed method does not require updating pretrained model parameters and incurs relatively low computational overhead. Furthermore, the authors consider not only predictive accuracy but also representation leakage and prediction disagreement to assess residual target-channel influence. The evaluation is further extended to multi-channel interventions and regression tasks. These efforts demonstrate the authors' awareness of the complexity of the problem.

Critical weaknesses
  1. The proposed method uses SVD to identify dominant directions in the target sensor's latent representations and treats them as the subspace to be suppressed. However, larger singular values do not necessarily imply that the corresponding directions have greater influence on downstream predictions. The paper lacks sufficient theoretical justification for this design choice. Moreover, comparisons with simpler alternatives, such as uniform feature scaling and random subspace projection, are missing. Consequently, the necessity and distinct contribution of SVD-based subspace selection remain unclear.
  2. The paper primarily measures unlearning effectiveness through the accuracy difference before and after zero-masking the target channel. However, a small accuracy difference does not necessarily indicate that the model has become independent of the target channel. Although the authors additionally report representation leakage and prediction disagreement metrics, these experiments are mainly conducted on UCI-HAR and do not adequately reveal potential nonlinear information remaining in the latent representations. Therefore, the current evidence is insufficient to comprehensively validate the proposed method's ability to suppress target-channel dependence.
  3. The proposed method is primarily evaluated using FC-STGNN, while the results on other architectures are not sufficiently convincing to establish its generalizability. Furthermore, comparisons with LEACE, INLP, and other baselines do not adequately control for differences in subspace dimensionality and utility-loss budgets, raising concerns about fairness. Some experimental tables also contain inconsistencies in the reported performance metrics, and statistical analyses across multiple random seeds are lacking. These issues weaken the reliability of the experimental conclusions.
Questions for authors
  1. The proposed method assumes that the dominant singular directions of the target sensor's latent representations are the most relevant directions for suppressing its influence. However, high representation energy does not necessarily imply high predictive importance. What evidence supports this assumption? Can the authors clarify whether CCSP provides meaningful advantages over simpler alternatives, such as uniform latent feature scaling or random subspace projection?
  2. The paper motivates channel-level unlearning through sensor failures, privacy concerns, and post-deployment sensor removal. However, CCSP still requires the target sensor's input to be processed by the encoder before applying the projection. How would the proposed method operate when the target sensor is physically unavailable or its data can no longer be collected? What practical advantages does CCSP offer over simply masking the unavailable sensor in such scenarios?
  3. The primary unlearning metric measures the accuracy difference before and after masking the target channel. However, a small accuracy difference does not necessarily indicate that the model has become independent of that channel. Given that the reported prediction disagreement remains non-negligible after CCSP, how can the authors substantiate the effectiveness of channel-level unlearning beyond the masking-based metric? Are the observed improvements consistent across different intervention strategies and datasets?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.