acceptodds
Withdrawn from ICLR 2027

ECRI: Evidence-based Interaction-Aware Time-Series Variable Selection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time-series variable selection for multivariate forecasting must distinguish redundant variables from complementary ones while accounting for recurring variable costs. Selection rules that require each variable to provide an individual gain can miss jointly predictive variables, while predictive relevance alone does not determine whether a variable is worth acquiring. We propose Evidence-based Contextual Relevance Inference (ECRI), a framework that evaluates variable acquisition as a conditional, cost-aware joint decision. ECRI uses interaction features in a lightweight Bayesian surrogate and scores individual and joint acquisition actions by their conditional marginal evidence under separate variable and interaction costs. This permits a variable pair to satisfy the acquisition criterion without either variable first satisfying a single-variable criterion, while avoiding repeated training of downstream forecasting models. We characterize this mechanism under orthogonal-design conditions, explicitly accounting for the evidence and structural penalties of joint acquisition. The resulting variable-selection mask can be reused across forecasting backbones. Across three real-world sensor-network datasets and four forecasting backbones, ECRI removes 47.0–97.4% of optional variables and achieves the lowest prediction time and FLOPs among the six evaluated variable selectors, with competitive forecasting accuracy. These results support interaction-aware variable selection as a practical approach to reducing recurring forecasting computation.

Primary area:
learning on time series and dynamical systems
Keywords:
Variable selection, Multivariate time-series forecasting, Bayesian model evidence.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ehRwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The author proposed a variable selection algorithm using a greedy algorithm and applied it to this time series prediction. From the experimental results, it can be seen that this method is quite good. However, in terms of the author's writing and presentation style, there is still room for improvement.

Critical strengths
  1. The author used the greedy algorithm to construct the optimization function, which is great.
  2. The experimental results indicate that this method significantly improves the efficiency of variable utilization.
Critical weaknesses
  1. Algorithm 1 in the main text describes the complete version of iterating through all actions one by one, but in reality, it runs a shortlist version of 4 actions per round, with multiple scores mixed together, making it impossible for readers to reproduce according to the main text.
  2. Although optimizing the problem into a greedy algorithm is good, the essence of the problem may not be a greedy algorithm problem, but a combinatorial optimization problem. Search only adds without subtracting, without deleting, swapping, or restarting. If you make a mistake in the early stages, you can never change it back.
  3. The pre-screening of Equation (18) only retains variables that are linearly related to the residuals, while the paper is most looking for variables that are not useful individually but are useful in combination. These types of variables will be screened out in the first step.
Questions for authors
  1. Can the author clearly write down the actual algorithm? The current algorithm itself and its actual implementation cannot be understood, and the author has not provided the code for implementation.
  2. If the optimal solution is to choose B first and then choose C, but according to the principle of greed, it may choose A first and then choose B, and then find that C has a negative effect. This algorithm cannot avoid this problem, we cannot determine whether greed can reach a relatively optimal state, and the author does not compare where the upper bound of the algorithm is.
Reviewer_2cVGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks whether a target series can be forecast using fewer contextual variables while keeping prediction quality and reducing inference cost. ECRI represents each candidate variable by its historical mean, adds pairwise product features, and selects variables using evidence from a Bayesian linear surrogate together with cost penalties. It first screens the candidate variables, then considers a small shortlist of actions at each search step. The selected variables can be used with different forecasting models. Across three sensor-network datasets and four forecasting models, the authors report a 47.0%–97.4% reduction in optional variables and lower prediction time and FLOPs than the other evaluated selectors.

Critical strengths

The problem has practical value: using fewer input variables can reduce computation when a forecasting model is deployed. Being able to reuse the selected variables across forecasting models is also useful. The authors report error, context size, selection time, and inference cost across three datasets and four models, which gives a reasonably broad view of the results. Appendices G–H also examine which actions were used during search and explain what those counts can establish. These are useful parts of the paper.

Critical weaknesses

The experiments provide limited support for the claimed advantage. In Table 2, adding interactions reduces MAE from 2.5651 to 2.5557, about 0.37%, but no variability estimates are reported, so the stability of the improvement is unclear. Comparisons at matched context sizes or computational costs are also missing, leaving ECRI’s advantage at the same budget uncertain. Incomplete baseline tuning and implementation details further make the results difficult to assess and reproduce.

Questions for authors

The MAE difference between full ECRI and the version without interactions is about 0.37%. Could you report the mean and standard deviation across random seeds to help assess whether this improvement is consistent? Could you also compare the methods at matched context sizes or similar FLOPs to determine whether ECRI retains an advantage under the same resource budget?

Reviewer_1YVMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ECRI, an interaction-aware variable-selection framework for multivariate time-series forecasting. It uses a Bayesian linear regression surrogate with interaction features to evaluate individual and joint variable acquisitions under explicit structural costs. Residual-based screening and action shortlisting reduce selection overhead. Experiments on three sensor-network datasets and four forecasting backbones show that ECRI reduces optional variables by 47.0%-97.4% and achieves lower prediction time and FLOPs than the evaluated baselines while maintaining competitive forecasting accuracy.

Critical strengths
  1. Interaction-aware selection: ECRI can jointly select variables whose predictive value may not be apparent individually, while accounting for variable and interaction costs.

  2. Promising empirical efficiency: Experiments across three datasets and four forecasting backbones show substantial reductions in context size, prediction time, and FLOPs, with competitive forecasting accuracy.

Critical weaknesses
  1. Assumption-dependent theoretical justification: The connection between Bayesian surrogate evidence and downstream forecasting loss relies on uniform approximation and alignment assumptions. Their practical validity, particularly for nonlinear forecasting backbones, is not sufficiently established.

  2. Limited validation of interaction screening: ECRI evaluates only a small shortlist of candidate actions, potentially missing complementary variables. The paper does not sufficiently establish screening reliability or robustness across broader datasets and settings.

Questions for authors
  1. How well do Bayesian evidence rankings correlate with actual downstream forecasting performance across the evaluated backbones?

  2. How often does the screening procedure retain useful complementary variable pairs that would be selected by exhaustive ECRI search?

  3. How sensitive are the results to the cost coefficients and shortlist size, and do the gains generalize to additional datasets and forecasting horizons?

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

Related papers

Discussion (0)

Sign in to comment.