acceptodds
Desk rejected at ICLR 2027

SIS-NET: Sensor Identity-Aware Set Modeling for Sensor-Configuration-Agnostic Electronic Nose Recognition

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Electronic-nose recognition models are commonly trained with a fixed sensor array and therefore implicitly depend on a fixed number, identity, and ordering of input sensors. This assumption is fragile in deployment, where sensors may be unavailable because of hardware variants, maintenance, or externally identified unreliable measurements. We study whether a single model can recognize samples under different subsets of a known sensor universe without configuration-specific retraining. We propose SIS-NET, a sensor-set model that applies a shared temporal encoder to each available sensor, incorporates an explicit sensor identity embedding, and aggregates the resulting variable-size token set with masked mean pooling. During training, sensor subsets are sampled from multiple availability levels so that the model experiences configuration variation before inference. We evaluate the same checkpoint under full, fixed-subset, random-subset, and held-out-combination conditions. The experiments are designed to distinguish configuration-agnostic recognition from fixed-input zero filling, electronic-nose fault-tolerant ensembles, and multivariate time-series baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Electronic Nose, Sensor-Set Learning, Set Representation Learning, Sensor Configuration Robustness, Multivariate Time Series, Sensor Aggregation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_fm39OpenReview ↗
Rating
1
Confidence
1
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes SIS-Net, a set-based architecture for electronic-nose recognition under varying sensor configurations. The method employs a shared temporal CNN encoder, learnable sensor identity embeddings, and masked mean pooling to process variable-size sensor subsets. Training-time sensor-subset sampling enables a single model to operate under different sensor availability conditions without retraining.

Critical strengths
  • Practical and relevant problem formulation. The paper addresses a meaningful challenge in electronic-nose recognition: handling different sensor configurations using a single model without configuration-specific retraining. This is relevant to practical systems where sensors may become unavailable.

  • Systematic evaluation under varying sensor availability. The evaluation considers fixed, random, and held-out sensor configurations, with group-based cross-validation and component ablations. These experiments provide useful evidence of robustness to sensor removal.

Critical weaknesses

Fundamental limitation of the sensor identity-aware representation. The proposed architecture adds sensor identity embeddings to temporal features before mean pooling:

r=1K∑j∈S(E(xj)+ej).r=\frac{1}{K}\sum_{j\in S}(E(x_j)+e_j).

However, this can be rewritten as:

r=1K∑j∈SE(xj)+1K∑j∈Sej.r=\frac{1}{K}\sum_{j\in S}E(x_j)+\frac{1}{K}\sum_{j\in S}e_j.

Consequently, sensor identity information is aggregated independently of the corresponding sensor measurements. For a fixed sensor subset, permuting the measurements among sensor identities produces exactly the same representation.

Therefore, the architecture cannot explicitly model which specific sensor produced which measurement. The identity embeddings primarily encode the composition of the available sensor subset rather than sensor-specific signal relationships. This substantially weakens the paper's central claim of sensor identity-aware modeling.

Limited novelty and experimental validation. The proposed method is essentially a simple set-based architecture combining a shared CNN, identity embeddings, mean pooling, and subset sampling. These components are well established, and the paper does not demonstrate a substantial methodological advance over existing set-based learning approaches.

Furthermore, evaluation is restricted to a single dataset with 975 processed samples and 16 sensors. The paper lacks direct comparisons against stronger set-based alternatives, such as Deep Sets with identity-conditioned features or Set Transformers, under the same subset-training protocol. Consequently, it remains unclear whether the improvements originate from the proposed architecture or primarily from training-time sensor-subset augmentation.

Questions for authors
  1. How does SIS-Net preserve sensor-specific information when additive identity embeddings are followed by mean pooling?

  2. How does SIS-Net compare with standard Deep Sets or Set Transformers using the same subset-training strategy?

  3. Can the authors validate the approach on additional datasets or real sensor failure scenarios?

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer fm39

We thank the reviewer for the careful mathematical analysis and constructive feedback. We address the three main concerns below.

(1) Sensor identity representation. We agree with the reviewer's mathematical observation. With additive identity embeddings followed by mean pooling, the representation decomposes into two independent terms:

ri=1Ki∑j∈SiEθ(xi,j)+1Ki∑j∈Siej.r_i = \frac{1}{K_i}\sum_{j\in\mathcal{S}i} E\theta(x_{i,j}) + \frac{1}{K_i}\sum_{j\in\mathcal{S}_i}e_j.

Consequently, for a fixed available sensor subset, permuting measurements among sensor identities leaves the pooled representation unchanged. The identity embeddings encode the composition of the available sensor subset but do not explicitly preserve the correspondence between each measurement and its physical sensor identity.

We have revised the Abstract, Introduction, Methodology (Section 3.6), ablation discussion (Section 4.2), and Conclusion to clarify this limitation. The Sensor ID ablation in Table 3 shows that the embeddings provide useful available-set information in the evaluated settings. For example, Avg-4 Macro-F1 decreases from 0.7522 to 0.6546 when identity embeddings are removed. However, this does not demonstrate explicit signal–identity interaction modeling. We acknowledge that the original description overstated this capability.

(2) Comparison with stronger set-based models. We agree that SIS-Net is closely related to established Deep Sets formulations. Its contribution lies primarily in investigating electronic-nose recognition under changing sensor configurations, training-time subset sampling, and evaluation using one frozen checkpoint across Full, Fixed, Random, and HeldOut conditions. We have not compared against identity-conditioned Deep Sets or Set Transformers under the same protocol and therefore do not claim architectural superiority over them.

(3) Generalization and physical sensor failures. Our current evaluation uses one electronic-nose dataset with sensor unavailability simulated through controlled masking. We have not evaluated additional datasets or actual hardware failures. These limitations are now explicitly acknowledged in the revised Conclusion.

We appreciate the reviewer for highlighting these issues. Developing representations that preserve sensor–measurement correspondence and evaluating them across broader sensing conditions remain important directions for future work.

Reviewer_fm39OpenReview ↗

Response to author

The author haven't address any of my concern. I will keep my decision.

Reviewer_Y5bxOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

To address challenges such as sensor failures, maintenance requirements, or hardware configuration variations encountered during the practical deployment of e-nose multi-sensor systems, this paper defines the sensor-configuration-agnostic e-nose recognition task. For this task, the authors propose the SIS-Net model. The model first employs a three-layer 1D CNN—with parameters shared across sensors—to independently encode the temporal responses of each available sensor. The resulting temporal representations are then combined with learnable identity embeddings corresponding to each physical sensor, forming a variable-sized set of sensor tokens. Finally, these tokens are mapped into a fixed-dimensional global representation via masked average pooling and fed into a classifier. Using a dynamic gas mixture dataset, the authors establish a unified evaluation protocol—encompassing Full16, Fixed-k, Random-k, and HeldOut-k scenarios—to demonstrate that SIS-Net outperforms existing approaches, including zero-padding baselines, fault-tolerant ensemble methods, and the multivariate time-series model RAINDROP.

Critical strengths

None

Critical weaknesses

The manuscript has significant writing issues and is replete with redundant or meaningless expressions, particularly in the methodology section. In that section, the authors unnecessarily complicate what are, in fact, extremely simple definitions by introducing formulas. Furthermore, there are clear problems throughout both the methodology and experimental sections of the manuscript.

Questions for authors

(1) In the experiments section, the authors compare their approach against four baselines. Except for the final one, RAINDROP (a 2022 study), the sources for these methods are not provided. Furthermore, the number of baselines is quite limited, and the methods themselves are rather outdated.

(2) In the methodology section, the presentation is unnecessarily complex, incorporating numerous redundant or ineffective formulas.

(3) There are several issues with the references; for instance, the second entry lacks any publication information.

(4) In Table 1, why does SIS-Net perform significantly better on the Fixed12 configuration than on the Full16 configuration (which utilizes all 16 sensors)? Conversely, why is the performance on Fixed8 lower than that on Fixed4 (which uses only 4 sensors)?

(5) Given that the original dataset consists of continuous gas concentration monitoring data, what are the specific thresholds used to define the four concentration categories, and what is the preprocessing pipeline for the 975 samples? Additionally, what are the specific network architecture and the hyperparameters used for training?

(6) Are Equation (8) and Equation (21) identical?

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer Y5bx

We thank the reviewer for the detailed comments. We have made targeted revisions to improve the clarity, accuracy, and presentation of the manuscript. Our responses are provided below.

(1) Baselines and references. We have clarified the origins of the baseline methods in Section 4.1. Full-Sensor Reference and Zero-Fill Fixed Input are study-defined controls. Sensor Swarm Ensemble is a study-specific implementation motivated by previous sensor-swarm research (Wang et al., 2023), while Multi-Kernel Fault-Tolerant E-nose is a study-specific comparison rather than an exact reproduction of a published architecture. RAINDROP is identified as the published graph-based time-series baseline (Zhang et al., 2022). We acknowledge that comparisons with additional recent methods would strengthen the evaluation.

(2) Methodological presentation. We agree that several equations in the original manuscript were unnecessarily detailed. We have simplified Section 3 by converting elementary definitions into inline expressions and removing redundant descriptions, while retaining the essential model and training equations.

(3) Reference formatting. We have corrected the REVE reference to include its NeurIPS 2025 publication information. We have also checked and completed relevant bibliographic entries, including LUNA, BrainOmni, and several journal references.

(4) Non-monotonic Fixed-k results. We agree that these results require explanation. Table 1 evaluates particular fixed sensor combinations, whose informativeness may differ. Therefore, Fixed12 (0.8698) can outperform Full16 (0.8397), and Fixed4 (0.7892) can outperform Fixed8 (0.7734), without implying that fewer sensors are generally better. This is a possible explanation rather than an established causal mechanism. Importantly, Table 2 averages over ten fixed combinations and reports Macro-F1 values of 0.8172, 0.8003, and 0.7482 for Fixed12, Fixed8, and Fixed4, respectively, showing a decreasing average trend. We have added this clarification to Section 4.2.

(5) Dataset processing and implementation details. The manuscript reports 975 processed samples of shape [16, 300], four concentration-based classes, five group-based folds, a shared three-layer 1D CNN, and training-time sensor-count sampling from {16, 12, 8, 4}. However, we acknowledge that the exact concentration thresholds, raw-data processing procedure, detailed preprocessing steps, and complete training hyperparameters remain insufficiently documented. These reproducibility issues have not been fully resolved in the current revision, and we do not introduce unverified details.

(6) Duplicate equations. The reviewer is correct that Equations (8) and (21) in the original submission were identical. We have removed the redundant expression and retained a single complete model mapping, now numbered Equation (2) in the revised manuscript.

We appreciate the reviewer's comments, which have helped us identify and correct several presentation and documentation issues.

Thank you for the author's response. Although the author has explained my doubts, after reading the manuscript and reviewing the content, I believe that the existing manuscript is not sufficient for publication in ICLR. Sorry, I will maintain my original score.

Reviewer_DxFYOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

In this work, the authors proposed SIS-NET, a sensor-set model for electronic-nose recognition that operates across different subsets of a known sensor universe without configuration-specific retraining. Conventional e-nose models are trained on a fixed sensor array and thus implicitly assume a fixed number, identity, and ordering of inputs, an assumption that breaks down in deployment when sensors become unavailable due to hardware variants, maintenance, or detected unreliability. SIS-NET addresses this by applying a shared temporal encoder to each available sensor, adding an explicit sensor identity embedding, and aggregating the resulting variable-size token set through masked mean pooling. During training, sensor subsets are sampled at multiple availability levels so the model is exposed to configuration variation before inference. A single checkpoint is then evaluated under full, fixed-subset, random-subset, and held-out-combination conditions. The experimental design is intended to separate true configuration-agnostic recognition from alternatives such as fixed-input zero filling, fault-tolerant e-nose ensembles, and multivariate time-series baselines.

Critical strengths
  • The problem introduced in the apper is interesting. Electric nose data focus on a unique subfield of tabular/time series data learning where, like many other time series input in the health domain, many inputs are not always reliable or even available. This gives a unique opportunies for learning with partial data.

  • The paper is clear and easy to read. The evaluation protocol is clean, and the results shows substantial improvements over the baselines.

Critical weaknesses
  • The methods are very similar to prior works, with insufficient discussion/comparison over them. The central equation (eq. 8) is not much different from Deep Sets (https://arxiv.org/abs/1703.06114), with same shared encoder with symmetric pooling. The encoder id with tokenized time series input is also no different from iTransformer (https://arxiv.org/abs/2310.06625). In related works, the author described it as could only take in data with fixed channel configurations. However, I think it's trivial for the model to support variable number of channels. Learning under incomplete sensor data is also widely explored, with Google's LSM-2 (https://arxiv.org/abs/2506.05321) being a prominant example.

  • While the channel ablation experiments are complete, the whole paper relies on a single dataset for all evaluations. The conlusion/implications therefore may only be applicable to this single dataset, which gives very limited impacts. It would be nice if the authors could add more datasets, or apply the methods to other time series domains (i.e. EEG, as discussed in introduction).

Questions for authors

Please see weaknesses.

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer DxFY

We thank the reviewer for recognizing the practical relevance of our problem formulation and the clarity of the evaluation protocol.

Regarding novelty and related work, we agree that shared sensor-wise encoding and symmetric mean pooling are established components closely related to Deep Sets. Our contribution is therefore not a fundamentally new set-learning operator, but the investigation of sensor-configuration-agnostic electronic-nose recognition using a single checkpoint across Full, Fixed, Random, and HeldOut sensor configurations, without configuration-specific retraining.

We also acknowledge the relevance of iTransformer and LSM-2. Our original discussion was overly broad in characterizing existing time-series models as requiring fixed channel configurations. Some architectures can accommodate variable channels or incomplete observations. We have revised the manuscript to distinguish architectural flexibility from demonstrated robustness under our evaluation protocol. We acknowledge that direct comparisons with these stronger alternatives remain necessary.

Regarding experimental scope, we agree that evaluation on a single electronic-nose dataset limits the generalizability of our findings. We have clarified in the revised manuscript that sensor unavailability is simulated through masking and that neither real hardware failures nor cross-domain generalization has been evaluated. We do not claim that our current results establish superiority across other sensing domains.

We appreciate these comments and have revised the manuscript to present its contributions and limitations more precisely.

Reviewer_DxFYOpenReview ↗

Thanks for response and keeping my score

Thanks authors for the response. As all weaknesses have been acknowledged but not addressed, I'm keeping my current score.

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

Related papers

Discussion (0)

Sign in to comment.