acceptodds
Under review as a conference paper at ICLR 2027

FSA-ILR: Explicit Relative-Scale Structure for Dynamic Odor Recognition

Abstract

Identifying food ingredients from the dynamic responses of an electronic nose can provide odor-based evidence for ingredient identity verification. This task requires jointly using the magnitude of change, relative configurations across scales, and temporal dynamics in multichannel responses. Existing time-series methods typically encode these types of information jointly, leaving the contribution of cross-scale ratios to class discrimination largely to implicit learning. We propose the Finite-Structure Additive ILR Network (FSA-ILR), which uses an amplitude–composition reparameterization and an additive class-score readout to explicitly structure how different types of dynamic information contribute to classification. Amplitude anchors and full isometric log-ratio (ILR) coordinates jointly preserve the positive scale energies of active channels. Comparisons with learned references for each class–channel pair, constrained by an activity mask, produce composition-matching scores that are added classwise to temporal and activity/amplitude scores. On the 50-class food ingredient identification task in SmellNet-Base, FSA-ILR outperforms the comparison methods in Acc@1, Acc@5, and macro-F1 under all four sensor-only input settings, with an Acc@1 gain of up to 3.96 percentage points over the strongest comparator in each setting. Representation probes and independently retrained controls further support the joint roles of complete ratio information, scale identity, and structured class-score readouts.

Primary area:
learning on time series and dynamical systems
Keywords:
dynamic odor recognition, electronic nose, food ingredient identification, multivariate time-series classification, compositional data analysis

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_168WOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates a method for food ingredient recognition based on the dynamic response of an electronic nose. For this task, this paper proposes that the amplitude of changes, cross-scale structure, and dynamic representations must be considered simultaneously, whereas existing time-series methods tend to mix information encoding.

This paper proposes the Finite-Structure Additive ILR Network (FSA-ILR), which explicitly organizes the contributions of different types of dynamic information to classification through amplitude–composition reparameterization and additive category score outputs. Finally, the score is obtained by summing the composition match score and the temporal score.

The proposed method was evaluated on the SmellNet-Base dataset using four sensor signals and achieved the best results when compared to MLP, CNN, ScentFormer, and BiLSTM.

Critical strengths
  1. The method described in this paper achieved the best accuracy and F1 score when the data was grouped by collection date. Ablation experiments demonstrate that all component paths contribute to the results.

  2. This paper proposes a method for constructing scores by summing three different paths, taking into account temporal dynamics, amplitude, and sensor-scale scaling. This transparent structural composition is beneficial for sensor tasks.

Critical weaknesses
  1. Figure 1 indicates that this is a schematic diagram. The claim that “Existing time-series methods typically encode these types of information jointly, leaving the contribution of cross-scale ratios to class discrimination largely to implicit learning.” is not supported by actual experiments or data.

  2. The claimed contributions of this method lack empirical evidence, and its novelty is limited. The performance improvement may stem from the architecture itself or from the manually designed multiscale features.

  3. The generalizability of the method is not sufficiently demonstrated. Given that the study is based on only one dataset and includes few comparison methods, the conclusion that this method can serve as a general approach for dynamic response recognition lacks supporting evidence, considering the scope of this study.

Questions for authors

I have a few questions I’d like the author to answer and clarify.

  1. Figure 1(b) is an illustrative example. Is this supported by actual experiments and data?
  2. The paper states that the core idea is that existing models rely on implicit learning of cross scale proportional information. I would like to see additional experimental explanations regarding the validity of this claim for such tasks.
  3. The core method of this paper relies on the combined effects of different modules. As shown in Table 2, when considering only the temporal path, the results are essentially consistent with those of ScentFormer. If other methods were to incorporate the statistical path mentioned by the authors, would their performance also improve? If so, is this improvement due to the concatenation of additional feature paths? I would like to see more fair comparative experiments.
  4. The abstract states that “FSA-ILR outperforms the comparison methods in Acc@1, Acc@5, and macro-F1 under all four sensor-only input settings,” but the main text does not include experimental results for Acc@5.
  5. This paper uses only one experimental dataset. Additionally, it lacks comparisons with baselines such as TimesNet, TimeMixer and others.
  6. Please review the results in Table A2.
Reviewer_85AoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the odor class prediction from a sensor array input using a combination of Transformer and engineered features. The method computes a change energy for each sensor channel at five temporal scales {1, 2, 5, 10, 25} and it splits the log of these energies into two parts: a log-amplitude anchor (the finest-scale energy k_1 = 1) and four isometric log-ratio (ILR) coordinates. The ILR coordinates are scored against learned per-class, per-channel reference points using squared distance, which equals the Aitchison distance. That score is added to the logits of a Transformer and an MLP over activity flags and amplitudes. On Day 6 of SmellNet-Base with 50 classes, the proposed pipeline outperforms MLP, CNN, BiLSTM and ScentFormer across four input settings.

Critical strengths
  • The derivations are correct and clearly presented.

  • There are thorough ablations, and in particular the matched-random-coordinates control is well designed.

  • The ablations and comparison to baselines show that the statistics pathways carry real information, and the authors report SOTA performance in their evaluation procedure based on Day 6 test set.

Critical weaknesses
  • The evaluation in all settings except representation probes in Table A2 is done on a single test set coming from Day 6. It has only one session per class, and windows within a session are strongly correlated, so the test set is effectively 50 sessions, not 581 windows. Thus, gains of 2-4 pp correspond to roughly 10-25 windows, and many ablation gaps (about 1-1.5 pp) to only 6-8 windows. Reported standard deviations also reflect only training randomness and not data variation, which limits assessing the statistical significance of the FSA-ILR gains over the baselines.

  • The authors cite several related works in the "Multivariate time series and multiscale representations" paragraph but do not evaluate and compare any of these models to the proposed FSA-ILR.

  • Several reported numbers have unexpected regularities. Table A2 reports constant differences between macro-F1 and Acc@1 (2.80) in all leave-one-day-out settings, Table A1 reports constant s.d. for macro-F1 of each model (0.55, 0.58, 0.52, 0.60, 0.56), and Table A3 reports in Δ\Delta[range] minimum and maximum which is perfectly symmetric around the median (i.e., min/max = median ±\pm const.). One possible explanation for the last phenomenon is the narrow test set, but then this underlines the need to evaluate the methods in leave-one-day-out on all available days.

  • Baselines' hyperparameters and training details are missing. Considering the small advantage of FSA-ILR in some settings, how much hyperparameter search budget was invested in each baseline?

  • FSA-ILR has a preprocessing advantage in 50/0 and 100/0 settings. FSA-ILR’s scale energies are built from differences of block means and are offset-invariant, so its statistics pathway effectively includes multi-lag differencing that the baselines lack in the p = 0 settings. As can be seen in Table 1, differencing strongly benefits CNN and BiLSTM (increase between 14 to 27 pp), and the headline 3.96 pp gain comes from 50/0. Therefore, part of the advantage in the p = 0 settings may come from preprocessing rather than the compositional readout. This also makes the p = 25 settings the fairer comparison.

  • (Minor) The motivation is loosely connected to the task. Figure 1a’s adulteration statistics motivate food authenticity, but the task is clean single-ingredient classification. Adulteration is a contamination problem that requires identification of a change in characteristic odorant molecules, and the paper doesn’t address it.

Questions for authors
  • Please report all evaluations on leave-one-day-out cross-validation using 5 training and 1 test day. Currently, the majority of claims (except representation probes in Table A2) are tested on Day 6, which limits the statistical assessment of the results (see weaknesses).

  • Please comment on the regularities in Tables A1, A2, and A3 and what is their origin (see weaknesses).

  • Please evaluate multivariate time series baselines, at least MiniRocket and InceptionTime.

  • To isolate the contribution from the reference point-learning parametrization, please add one ablation that replaces the class-reference score ρμ,y(a,U)\rho_{\mu,y}(a,U) with an unconstrained linear layer on the ILR coordinates, keeping all other components and training settings identical.

  • Could the authors specify the chosen hyperparameters and training procedure of the baselines?

  • How often is a channel actually inactive (ac=0a_c = 0) in training and test windows? If I understand well, κc\kappa_{c} equals 1 whenever channel c varies in at least one training window and τc\tau_{c} is close to zero because the median t is multiplied by 64×2−5264\times 2^{-52} which is unlikely to be outpaced by any data-driven values of the median.

  • How were the qualitative examples (Fig. 4) chosen, and can you show a failure case where the composition pathway reverses a correct prediction? Also given the advantage of FSA-ILR in 50/0 and 100/0 evaluations (see weaknesses), can you report such examples in 50/25 or 100/25 settings?

Reviewer_N4LBOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

FSA-ILR is an additive network for identity recognition of multi-channel dynamic response signals. Its core idea is to explicitly disentangle the "magnitude of change," the "cross-scale ratio structure," and the "temporal dynamics" of a signal. On a fixed scale set, the change energy at each scale is extracted. And then, for active channels, the magnitude of change is represented by an amplitude anchor, while the full degrees of freedom of the cross-scale ratio are represented by complete ILR coordinates, and the two jointly constitute a recoverable reparameterization. Subsequently, a reference is learned for each class-channel pair, and an activity-mask-gated Aitchison distance converts the proximity between the observed ratio and the class reference into a matching score. Finally, this score is additively fused at the class level with the outputs of a Transformer temporal branch and an activity/amplitude branch, and the whole model is trained jointly under a single cross-entropy objective. Later experiments in this paper compare FSA-ILR against multiple baselines under four sensor-only input settings, and controlled experiments including leave-one-day-out probing, structure substitution, and readout substitution verify the role of each design. The main contributions of this work are as follows: Firstly, it proposes a recoverable reparameterization that jointly uses amplitude anchors and complete ILR coordinates, enabling the magnitude of change and the ratio structure to be clearly separated in roles without information loss. Secondly, it designs an activity-aware class-reference readout that explicitly converts the full log-ratio structure into class evidence and additively combines it with the temporal and amplitude pathways. By conducting systematic mechanism experiments, the authors separately examine the roles of relative information, scale identity, and readout organization, providing complementary evidence for the joint design of the fixed statistics & learnable discriminative readout. The effectiveness of the paper is reflected in the following. In terms of performance, under four sensor-only input settings on a 50-class food ingredient recognition task, FSA-ILR outperforms the baselines in Acc@1, Acc@5, and macro-F1. In terms of mechanism, independently retrained experiments such as leave-one-day-out representation probing, structure substitution, and readout substitution show that the gain brought by complete ILR coordinates is significantly higher than that of a single coarse-fine log-ratio and of matching random features, and that disrupting scale identity or substituting random coordinates both lead to performance degradation, while removing the component pathway or changing the readout manner likewise reduces accuracy, indicating that the relative structure, scale consistency, and additive readout organization each play a role. In terms of evidence completeness, the overall comparison, pathway ablation, window-level score analysis, and class-level recall analysis complement one another, supporting the effectiveness of the method design while faithfully presenting the limitation that performance is not consistent across all classes.

Critical strengths

This work leaves the impression on me for its clear mathematical modeling of the research problem. It explicitly decomposes the "magnitude of change," the "cross-scale proportional structure," and the "temporal dynamics" into three pathways. The combination of amplitude anchors and full ILR coordinates constitutes a recoverable reparameterization, with the ILR distance being strictly equivalent to the Aitchison distance—these designs have a clear geometric and algebraic basis. At the same time, the authors have designed a relatively complete validation system with targeted control experiments: beyond overall performance comparison, multiple groups of independently retrained controls separately test "whether relative information is useful," "whether scale identity matters," and "whether readout organization is crucial," and the chain of evidence is mutually complementary. All in all, the innovation of this work lies in its reasonable modeling and its signal-processing pipeline design. The ablation experiments separately demonstrate the gain contributed by each design detail to the overall results, which is sufficiently rigorous. At last, the fact that all metrics show good improvement over the baselines indicates that this work is effective and promising.

Critical weaknesses

In my points of view, the critical weaknesses of this work lie in the experimental setup and the analysis of corresponding results. The experiments are conducted on only one dataset—the 50-class food ingredient task with 6 sensor channels—and are not validated on other multi-channel dynamic response datasets or on more general time-series classification benchmarks, which somewhat weakens the persuasiveness of the results. At the same time, the comprehensive analysis of the experimental results is not sufficiently complete or in-depth. For example, Figure A2 shows that 73 of the 400 class × setting comparison units still exhibit negative gains, indicating that the overall advantage does not extend to multiple classes. However, the paper merely describes this result without conducting an in-depth analysis of the failure modes, proposing possible causes, or offering corresponding improvement strategies.

Questions for authors

Based on the above weaknesses, I would like to raise the following questions to the authors: First, could the effectiveness of FSA-ILR be further validated on other datasets or benchmarks, so as to strengthen the persuasiveness of the experimental results? Second, would it be possible to conduct an in-depth analysis of the failure modes of the specific classes that exhibit negative gains in metrics such as recognition accuracy in the experiments, and to propose possible explanations?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.