Pith. sign in

REVIEW 3 major objections 5 minor 82 references

A single self-calibrating rule lets IoT sensors suppress about 95% of transmissions while keeping reconstruction error near 0.35°C, by normalizing prediction errors against local signal volatility.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-01 12:18 UTC pith:22UDNS6A

load-bearing objection Useful empirical idea with a broken theoretical guarantee — the self-calibrating claim should be fixed or dropped before this is publishable. the 3 major comments →

arxiv 2607.19590 v1 pith:22UDNS6A submitted 2026-07-21 cs.IT math.IT

Learning to Transmit: Volatility-Aware Predictive Communication for Energy-Efficient IoT Networks

classification cs.IT math.IT
keywords IoT networkspredictive communicationstudentized residualvolatility-aware thresholdrecursive least squaresdual predictionenergy efficiencydata reduction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that communication in IoT networks should be driven by information novelty relative to local volatility, not by fixed intervals or fixed thresholds. It proposes a rule: transmit a reading only when the prediction residual, divided by a rolling estimate of signal variability, exceeds a single constant alpha. The authors argue that this studentized-residual criterion automatically adapts to seasons, sites, and sensing modalities without per-deployment tuning, and they support this with experiments on three datasets totaling over 2.4 million observations. The strongest reported result is simultaneous Pareto dominance: the method achieves the highest data-reduction ratio and the lowest reconstruction error against six baselines, and an online RLS variant maintains 12-18% error improvements under drift. The paper further claims that the same alpha controls false-suppression probability across all regimes through a Gaussian approximation, a guarantee that fixed or moving-average thresholds cannot provide.

Core claim

The paper's central claim is that a volatility-normalized transmission rule—transmit when |x(t)-x_hat(t)|/sigma(t) > alpha, where sigma(t) is a rolling standard deviation of recent readings—makes communication decisions self-calibrating. Under the assumption that prediction residuals are locally Gaussian, the probability that a reading is suppressed at epoch t is Phi(alpha), independent of the signal's scale or regime, so a single alpha works across seasons and deployment sites without manual tuning. Empirically, the authors report that this rule, paired with a Ridge predictor and dual-prediction reconstruction, achieves up to 94.7% transmission reduction and 0.352°C MAE on outdoor temperatu

What carries the argument

The studentized residual r(t)=|x(t)-x_hat(t)|/sigma(t), where sigma(t) is the rolling standard deviation over the previous h epochs, is the engine of the method. It normalizes the prediction error into a dimensionless quantity that is approximately scale-invariant, so a single threshold alpha acts as a universal false-suppression knob. The paper pairs this with a closed-form Ridge predictor for the static version and Recursive Least Squares with exponential forgetting for the online version, and with dual-prediction reconstruction at the sink, where the receiver runs a synchronized copy of the predictor and fills suppressed slots with its forecasts.

Load-bearing premise

The load-bearing premise is that prediction residuals divided by the rolling standard deviation behave like standard normal random variables independent of the estimated scale, so that one alpha controls the false-suppression probability identically in every season and site; if that normality/independence assumption fails, the self-calibrating guarantee in Eq. (6) is unsupported.

What would settle it

Compute the empirical distribution of r(t)=|x(t)-x_hat(t)|/sigma(t) on the paper's own test sets: if P(r<=1.0) deviates substantially from Phi(1.0)≈0.841, or if this probability varies across the three Chicago stations, seasons, or the three datasets, then the universal-alpha guarantee fails even if the reported DRR/MAE numbers hold.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Operators can set one alpha without per-site tuning, and the rule automatically tightens or loosens as the signal becomes more or less volatile; no threshold re-calibration is needed across seasons or deployments.
  • Transmission count drops to roughly 5-13% of periodic sampling on the tested datasets, implying a comparable reduction in communication energy, which is the dominant energy cost in battery-powered IoT nodes.
  • The online RLS variant keeps reconstruction error bounded under sensor drift and non-stationarity, preserving DRR above 93% even when a linear bias is injected or real air-quality sensor drift is present.
  • Dual-prediction reconstruction cuts MAE by about 41% relative to hold-last-value, bringing reconstruction error close to the theoretical lower bound of the predictor's own error.
  • The rule scales linearly across 1-50 independent nodes with no cross-node coordination, making it a drop-in protocol for existing sensor fleets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The Gaussian assumption behind Eq. (6) is not validated in the paper, and the rule actually uses |r(t)| with r(t) defined without an absolute value; if residuals are heavy-tailed or scale-dependent, alpha will not be perfectly universal. A safer production variant would estimate the empirical quantile of r(t) online and map alpha to that distribution instead of relying on Phi.
  • The rule's self-correction under packet loss—DRR rises as losses increase—suggests it behaves like a feedback controller: missed packets inflate the next residual and trigger retransmissions. This could be analyzed formally as an event-triggered control system, potentially yielding stability and convergence guarantees.
  • Because the threshold responds to a second-order volatility statistic, the method is a lightweight change-point detector; it could be extended to detect anomalies or regime shifts in other streaming contexts (e.g., industrial monitoring, battery health) where communication is also energy-constrained.
  • A direct testable extension is to hold the predictor fixed and vary only the volatility window h; the paper's results use h=24, but the self-calibrating claim should hold for any reasonable h, and the sensitivity of the alpha guarantee to h could be measured empirically.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes ADAPTIVEML and ADAPTIVEML-RLS, predictive communication protocols for IoT in which a sensor transmits only when the one-step-ahead prediction residual exceeds a threshold proportional to the rolling standard deviation of the raw signal. The authors claim that this volatility-normalized rule is self-calibrating (the same alpha controls false-suppression probability across regimes), that the online RLS variant handles drift, and that the methods Pareto-dominate six baselines on three real-world datasets, achieving DRR around 94% with low reconstruction error. The manuscript reports extensive experiments and an energy proxy.

Significance. If the empirical results hold, the core idea is practically valuable: replacing a fixed or first-order adaptive threshold with a volatility-normalized rule is a simple, plausible improvement for energy-constrained IoT sensing. The multi-dataset evaluation against six baselines is a useful contribution, and the paper provides a clear protocol that could be reproduced from public datasets. However, the theoretical guarantee attached to the threshold rule (Eq. 6) is incorrect as stated, and the experimental design does not isolate the contribution of the volatility-aware threshold from the choice of predictor. The Pareto-dominance claim is also stronger than what a single operating point can support. The empirical tables may stand, but the paper's central conceptual claim needs substantial revision.

major comments (3)
  1. [Section III-C, Eq. (6)] The claimed false-suppression probability is wrong. r(t) is defined with an absolute value, so even under Gaussian residuals, P(r<=alpha) is not Phi(alpha) but the folded-normal tail probability. More importantly, sigma(t) in Eq. (4) is the rolling standard deviation of the raw signal x, not of the prediction residuals. If e(t) ~ N(0, sigma_e^2), then r(t) ~ |N(0, sigma_e^2)| / sigma_x, giving P(r<=alpha)=2*Phi(alpha*sigma_x/sigma_e)-1, which equals Phi(alpha) only in the special case sigma_e=sigma_x. Thus the statement that 'the same alpha controls this probability consistently across all seasons and deployment sites' is unsupported. This is load-bearing because the self-calibrating property is presented as the main advantage over fixed and moving-average threshold schemes.
  2. [Section IV, ablation of threshold mechanism] The experiments do not isolate the volatility-aware threshold. For prediction-based baselines (B3-B6), the same adaptive rule (5) is used, so the comparison reflects predictor quality, not the threshold mechanism. The Static Threshold baseline (B2) uses |x(t)-x(t-1)|>delta, which confounds threshold rule and predictor. To support the claim that volatility normalization improves over fixed or moving-average thresholds, the authors should include ablations with the same Ridge/RLS predictor under (a) a fixed residual threshold and (b) a moving-average of residual magnitudes, on the same datasets. Without this, the central advantage over prior adaptive-threshold schemes is not established.
  3. [Section IV-F, Tables II-IV and Fig. 4] The Pareto-dominance claim is based on a single operating point (alpha=1.0) for each method. The alpha-sweep (Fig. 4) is shown only for ADAPTIVEML. Since alpha trades DRR against MAE, a method that performs best at alpha=1.0 may not dominate at other operating points. To claim Pareto dominance, the authors should present DRR-MAE trade-off curves for all methods, or compare at matched DRR or matched MAE. The current statement 'simultaneously outperforming all six baselines in both efficiency and fidelity' is an overstatement without this evidence.
minor comments (5)
  1. [Section III-C] Calling r(t) a 'studentized residual' is misleading because the denominator is the rolling standard deviation of the raw signal, not an estimate of the residual standard deviation. Please rename or clarify.
  2. [Section IV-H, Table V] The injected drift magnitudes are physically implausible: mu_d=0.1 degrees C per epoch over 2208 epochs gives a cumulative drift of 220.8 degrees C. Please clarify whether the drift is normalized, applied per window, or perhaps the units are degrees C per 100 epochs.
  3. [Section II-B] The reference block [19]-[53] is dominated by self-citations that are not directly relevant to the predictive-communication contribution. Please prune to the cited works that actually motivate the design.
  4. [Section IV-C] State explicitly that test data are normalized using the training partition's mean and standard deviation; otherwise the preprocessing could leak information from the test set.
  5. [Section IV-F] The claim that 'DRR variance is less than 2%' across stations would be more informative as standard deviation or a full per-station table with confidence intervals; the current Table III reports only averages.

Circularity Check

0 steps flagged

No significant circularity: the central empirical claims are validation-tuned evaluations, and the theoretical guarantee is an explicitly assumed probabilistic statement rather than a fitted or self-referential reduction.

full rationale

The paper's load-bearing empirical results (DRR/MAE across three datasets) are computed with hyperparameters fixed on a held-out validation split (Section IV-E) and evaluated on disjoint test partitions; there is no test-set fit relabeled as a prediction, and the baselines share the same adaptive-threshold mechanism (Section IV-A), so the comparisons are not forced by construction. The theoretical self-calibration claim in Eq. (6) is introduced as "Under the assumption that prediction residuals are locally Gaussian" — an explicit modeling assumption, not a consequence of a fitted parameter. The equation is numerically questionable (r(t) uses an absolute value and sigma(t) is signal volatility rather than residual scale), but that is a soundness/validity defect, not circularity, because the target conclusion is not used as an input to its own derivation. The large [19]-[53] citation block in Section II-B is self-referential and irrelevant to the claim it accompanies, but it is not load-bearing: removing it would not alter any equation, algorithm, or experimental number, so under the rule that self-citation alone is not circularity it does not raise the score. The numerical complementarity between the Introduction's "approximately 5.3%" novel-observation statistic and the reported 94.7% DRR is suggestive, but the paper never defines the Introduction statistic as the output of rule (5), so there is no exhibited by-construction reduction. Accordingly, no circular step is identified.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The method's central operation is thresholding studentized residuals; it introduces no new physical entities. The main theoretical free parameters are α, window sizes, regularization, and RLS constants, all tuned on validation. The Gaussian-scale assumption behind Eq. (6) is the main unvalidated domain axiom.

free parameters (7)
  • α (sensitivity threshold) = 1.0
    Transmission rule |x−x̂| > α σ(t); tuned via grid search on held-out validation (Section IV-E); central knob for DRR/MAE tradeoff.
  • w (sliding window length) = 24
    Number of lags in Ridge/RLS predictor; selected by grid search; affects prediction error.
  • h (volatility window length) = 24
    Rolling standard deviation window in Eq. (4); default equals w; affects threshold scale.
  • λ (Ridge regularization) = 1.0
    Grid-selected on validation; L2 penalty in Eq. (3).
  • γ (RLS forgetting factor) = 0.98
    Exponential forgetting in Eq. (10); tuned; lower trades accuracy for adaptation speed.
  • ρ (RLS covariance initialization) = 10.0
    Initial P0 in Algorithm 1; chosen by hand/default; affects early adaptation.
  • Baseline hyperparameters (δ, ARIMA order, Q, R, EMA β, LMS µ) = δ=1.2°C, ARIMA(2,1,1), Q=0.01, R=0.1, β=0.9, µ=0.01
    Chosen per standard calibration or grid search (Section IV-A/E); central claim of superiority over six baselines depends on these being representative.
axioms (4)
  • domain assumption Local Gaussianity of prediction residuals and independence of the rolling standard deviation estimate.
    Section III-C Eq. (6) asserts P(r≤α)≈Φ(α); this requires residuals to be standard normal and σ(t) to be the true scale. Neither is established, and the absolute value in r makes the stated CDF wrong.
  • domain assumption Rolling standard deviation σ(t) over h=24 samples is a consistent estimator of local signal volatility.
    Eq. (4) uses a short trailing window; if the signal is strongly non-stationary within the window, σ(t) is biased, so the studentized residual is not scale-free.
  • domain assumption The dual-prediction sink remains synchronized with the node predictor.
    Section III-E Eq. (11) and the dual-prediction gain in Section IV-H3 assume no model divergence; packet-loss experiments are run without a resynchronization protocol, which the paper lists as future work.
  • standard math RLS with exponential forgetting remains stable and tracks drift at the chosen γ and ρ.
    Section III-D invokes standard RLS stability (Goodwin–Sin [64]); the specific choices γ=0.98, ρ=10 are assumed to keep P_t positive definite and β_t bounded.

pith-pipeline@v1.3.0-alltime-deepseek · 17075 in / 15258 out tokens · 127017 ms · 2026-08-01T12:18:00.143321+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Learning to Transmit: Volatility-Aware Predictive Communication for Energy-Efficient IoT Networks." pith.science (2026). https://pith.science/paper/22UDNS6A

@misc{pith2026260719590,
  author       = {Pith},
  title        = {Pith review of: Learning to Transmit: Volatility-Aware Predictive Communication for Energy-Efficient IoT Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/22UDNS6A}},
  note         = {Machine review of arXiv:2607.19590}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Communication is the dominant source of energy consumption in Internet-of-Things (IoT) networks, yet many sensed measurements exhibit strong temporal correlations and provide little new information to the receiver. This paper introduces \textsc{ADAPTIVEML}, a volatility-aware predictive communication framework that enables IoT devices to intelligently decide when communication is necessary. Each sensor maintains a lightweight machine learning predictor and transmits only when the prediction residual exceeds an adaptive threshold proportional to the local signal volatility. By normalizing prediction errors using a rolling estimate of signal variability, the proposed transmission policy automatically adapts to changing environmental conditions, seasonal variations, and deployment-specific dynamics without manual threshold tuning. To address long-term non-stationarity, we further propose \textsc{ADAPTIVEML-RLS}, an online learning extension based on Recursive Least Squares (RLS) with exponential forgetting, allowing continuous adaptation to sensor drift and evolving signal characteristics. Extensive experiments are conducted on three heterogeneous real-world datasets comprising more than 2.4 million sensor observations from outdoor environmental monitoring, indoor wireless sensor networks, and urban air-quality sensing. Compared with six representative baselines, including periodic transmission, static-threshold suppression, ARIMA, Kalman filtering, EMA, and LMS filtering, \textsc{ADAPTIVEML} achieves up to 94.7\% transmission reduction while maintaining a reconstruction error of 0.352$^\circ$C. \textsc{ADAPTIVEML-RLS} further reduces reconstruction error by 12--18\% under drift conditions while preserving transmission reduction above 93\%. These results demonstrate the effectiveness of volatility-aware predictive communication for energy-efficient and adaptive IoT networks.

Figures

Figures reproduced from arXiv: 2607.19590 by Ifrat Ikhtear Uddin, John Kangethe, Longwei Wang.

Figure 1
Figure 1. Figure 1: The ADAPTIVEML/ADAPTIVEML-RLS five-stage system pipeline. Step 01: the IoT node reads the raw sensor value x(t). Step 02: a Ridge regression or RLS predictor generates xˆ(t) from the w=24-lag sliding window. Step 03: the rolling standard deviation σ(t) is computed over the preceding h epochs, forming the adaptive threshold. Step 04: the studentised residual |x(t) − xˆ(t)|/σ(t) is compared against α; the re… view at source ↗
Figure 2
Figure 2. Figure 2: Visual summary of Table II at 63rd Street Station. (a) Ntx; (b) DRR; (c) MAE (solid) and RMSE (hatched). ADAPTIVEML (gold-bordered bar) achieves the best result on all three metrics simultaneously, confirming Pareto dominance. A particularly telling comparison is with the Kalman Filter. Despite being the Bayesian-optimal linear estimator under Gaus￾sian dynamics [58], it trails ADAPTIVEML by 4.2× in MAE. T… view at source ↗
Figure 3
Figure 3. Figure 3: (a) Average Ntx (bars, left axis) and average MAE (diamond markers, right axis) across all three Chicago stations for all seven methods. ADAPTIVEML (gold-bordered navy bar) records the lowest Ntx=117 and the lowest MAE=0.352 ◦C simultaneously, Pareto-dominating every competing approach. (b) Per￾station breakdown for ADAPTIVEML only, confirming geographically consistent performance across 63rd Street (Ntx=1… view at source ↗
Figure 4
Figure 4. Figure 4: α-sensitivity analysis at 63rd Street for w ∈ {6, 12, 24, 48}. (a) DRR–MAE Pareto frontier: larger windows achieve lower MAE at the same DRR; w=24 provides the best accuracy–memory balance. (b) DRR versus α: larger windows reach high DRR at lower α (less aggressive suppression). The dashed vertical marks the default α=1.0. G. Multi-Dataset Generalisability Table IV summarises performance across all three d… view at source ↗
Figure 5
Figure 5. Figure 5: DRR–MAE Pareto analysis across all three evaluation datasets ( [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: (a) Rolling MAE (30-day window) over the deployment year under injected drift µd=0.05 ◦C/epoch. ADAPTIVEML (red dashed) accumulates error monotonically once drift exceeds the training distribution. ADAPTIVEML-RLS (blue solid, γ=0.98) continuously re-estimates βt , keeping MAE substantially lower throughout. The shaded region quantifies the RLS benefit; the dotted grey line is the no-drift reference (µd=0).… view at source ↗
Figure 7
Figure 7. Figure 7: Robustness evaluation across all three datasets ( [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

82 extracted references · 8 linked inside Pith

  1. [1]

    Internet of things (iot): A vision, architectural elements, and future directions,

    J. Gubbi, R. Buyya, S. Marusic, and M. Palaniswami, “Internet of things (iot): A vision, architectural elements, and future directions,”Future Generation Computer Systems, vol. 29, no. 7, pp. 1645–1660, 2013

  2. [2]

    The internet of things: A survey,

    L. Atzori, A. Iera, and G. Morabito, “The internet of things: A survey,” Computer Networks, vol. 54, no. 15, pp. 2787–2805, 2010

  3. [3]

    Internet of things for smart cities,

    A. Zanella, N. Bui, A. Castellani, L. Vangelista, and M. Zorzi, “Internet of things for smart cities,”IEEE Internet of Things Journal, vol. 1, no. 1, pp. 22–32, 2014

  4. [4]

    Internet of things: A survey on enabling technologies, protocols, and applications,

    A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and M. Ayyash, “Internet of things: A survey on enabling technologies, protocols, and applications,”IEEE Communications Surveys & Tutorials, vol. 17, no. 4, pp. 2347–2376, 2015

  5. [5]

    Wireless sensor network survey,

    J. Yick, B. Mukherjee, and D. Ghosal, “Wireless sensor network survey,” Computer Networks, vol. 52, no. 12, pp. 2292–2330, 2008

  6. [6]

    Energy conservation in wireless sensor networks: A survey,

    G. Anastasi, M. Conti, M. Di Francesco, and A. Passarella, “Energy conservation in wireless sensor networks: A survey,”Ad Hoc Networks, vol. 7, no. 3, pp. 537–568, 2009. 11

  7. [7]

    Wireless sensor networks: A survey,

    I. F. Akyildiz, W. Su, Y . Sankarasubramaniam, and E. Cayirci, “Wireless sensor networks: A survey,”Computer Networks, vol. 38, no. 4, pp. 393–422, 2002

  8. [8]

    Wireless sensor networks: A survey on recent developments and potential synergies,

    P. Rawat, K. D. Singh, H. Chaouchi, and J. M. Bonnin, “Wireless sensor networks: A survey on recent developments and potential synergies,”The Journal of Supercomputing, vol. 68, no. 1, pp. 1–48, 2014

  9. [9]

    Data transmission reduction schemes in wsns for efficient iot systems,

    A. Jarwan, A. Sabbah, and M. Ibnkahla, “Data transmission reduction schemes in wsns for efficient iot systems,”IEEE Journal on Selected Areas in Communications, vol. 37, no. 6, pp. 1307–1324, 2019

  10. [10]

    A new data aggregation scheme based on sparse representation for wsns,

    G. B. Tayeh, A. Makhoul, D. Laiymani, and J. Demerjian, “A new data aggregation scheme based on sparse representation for wsns,”Ad Hoc Networks, vol. 87, pp. 110–121, 2019

  11. [11]

    A survey about prediction- based data reduction in wireless sensor networks,

    G. M. Dias, B. Bellalta, and S. Oechsner, “A survey about prediction- based data reduction in wireless sensor networks,”ACM Computing Surveys, vol. 49, no. 3, pp. 1–35, 2016

  12. [12]

    Adaptive model selection for time series prediction in wireless sensor networks,

    Y .-A. Le Borgne, S. Santini, and G. Bontempi, “Adaptive model selection for time series prediction in wireless sensor networks,”Signal Processing, vol. 87, no. 12, pp. 3010–3020, 2007

  13. [13]

    An improved adaptive dual prediction scheme for reducing data transmission in wireless sensor networks,

    H. Liazid, M. Lehsaini, and A. Liazid, “An improved adaptive dual prediction scheme for reducing data transmission in wireless sensor networks,”Wireless Networks, vol. 25, no. 6, pp. 3545–3555, 2019

  14. [14]

    An energy- efficient dual prediction scheme using lms filter and lstm in wireless sensor networks for environment monitoring,

    T. Shu, J. Chen, V . K. Bhargava, and C. W. de Silva, “An energy- efficient dual prediction scheme using lms filter and lstm in wireless sensor networks for environment monitoring,”IEEE Internet of Things Journal, vol. 6, no. 4, pp. 6736–6747, 2019

  15. [15]

    Predictive data reduction in wireless sensor networks using selective filtering for engine monitoring,

    D. J. McCorrie, E. Gaura, K. Burnham, N. Poole, and R. Hazelden, “Predictive data reduction in wireless sensor networks using selective filtering for engine monitoring,” inWireless Sensor and Mobile Ad-Hoc Networks. Springer, 2015, pp. 129–148

  16. [16]

    Evaluation of deep learning methods in a dual prediction scheme to reduce transmission data in a wsn,

    A. Rodrigues da Silva, M. Coutinho, L. A. Amaral, J. Araujo, and J. N. d. Souza, “Evaluation of deep learning methods in a dual prediction scheme to reduce transmission data in a wsn,”Sensors, vol. 21, no. 21, 2021

  17. [17]

    The impact of dual prediction schemes on the reduction of the number of transmissions in sensor networks,

    G. M. Dias, B. Bellalta, and S. Oechsner, “The impact of dual prediction schemes on the reduction of the number of transmissions in sensor networks,”Computer Communications, vol. 112, pp. 109–122, 2017

  18. [18]

    Practical data prediction for real-world wireless sensor networks,

    U. Raza, A. Camerra, A. L. Murphy, T. Palpanas, and G. P. Picco, “Practical data prediction for real-world wireless sensor networks,”IEEE Transactions on Knowledge and Data Engineering, vol. 27, no. 8, pp. 2231–2244, 2015

  19. [19]

    Representation learning and nature encoded fusion for heterogeneous sensor networks,

    L. Wang and Q. Liang, “Representation learning and nature encoded fusion for heterogeneous sensor networks,”IEEE Access, vol. 7, pp. 39 227–39 235, 2019

  20. [20]

    Congestion aware dynamic user association in heterogeneous cellular network: A stochastic decision approach,

    L. Wang, W. Chen, and J. Li, “Congestion aware dynamic user association in heterogeneous cellular network: A stochastic decision approach,” in 2014 IEEE International Conference on Communications (ICC). IEEE, 2014, pp. 2636–2640

  21. [21]

    Enhanced robustness by symmetry enforcement,

    L. Wang, A. Ghimire, K. Santosh, Z. Zhang, and X. Li, “Enhanced robustness by symmetry enforcement,” inIEEE Conference on Artificial Intelligence (IEEE CAI) 2024, 2024

  22. [22]

    Partial interference alignment for heterogeneous cellular networks,

    L. Wang and Q. Liang, “Partial interference alignment for heterogeneous cellular networks,”IEEE Access, vol. 6, pp. 22 592–22 601, 2018

  23. [23]

    Optimization for user centric massive mimo cell free networks via large system analysis,

    ——, “Optimization for user centric massive mimo cell free networks via large system analysis,” in2016 IEEE Global Communications Conference (GLOBECOM). IEEE, 2016, pp. 1–1

  24. [24]

    Explaining the behavior of neuron activations in deep neural networks,

    L. Wang, C. Wang, Y . Li, and R. Wang, “Explaining the behavior of neuron activations in deep neural networks,”Ad Hoc Networks, vol. 111, p. 102346, 2021

  25. [25]

    Exploration vs exploitation for distributed channel access in cognitive radio networks: A multi-user case study,

    L. Wang, X. Chen, Z. Zhao, and H. Zhang, “Exploration vs exploitation for distributed channel access in cognitive radio networks: A multi-user case study,” in2011 11th International Symposium on Communications & Information Technologies (ISCIT). IEEE, 2011, pp. 360–365

  26. [26]

    Deep reinforcement learning based computation offloading for mobility-aware edge comput- ing,

    M. Shi, R. Wang, E. Liu, Z. Xu, and L. Wang, “Deep reinforcement learning based computation offloading for mobility-aware edge comput- ing,” inInternational conference on communications and networking in china. Springer International Publishing Cham, 2019, pp. 53–65

  27. [27]

    Performance analysis of cooperative multicell precoding with global csi and local individual csi in the large dimensional regime,

    L. Wang and Q. Liang, “Performance analysis of cooperative multicell precoding with global csi and local individual csi in the large dimensional regime,”IEEE Transactions on Vehicular Technology, vol. 67, no. 4, pp. 3229–3238, 2017

  28. [28]

    Low complexity optimization for user centric cellular networks via large dimensional analysis,

    ——, “Low complexity optimization for user centric cellular networks via large dimensional analysis,”Physical Communication, vol. 25, pp. 412–419, 2017

  29. [29]

    Improving robustness of deep neural networks via large-difference transformation,

    L. Wang, C. Wang, Y . Li, and R. Wang, “Improving robustness of deep neural networks via large-difference transformation,”Neurocomputing, vol. 450, pp. 411–419, 2021

  30. [30]

    Looking beyond content: Modeling and detection of fake news from a social context perspective

    K. Xiao, L. Wang, A. Gupta, and X. Qin, “Looking beyond content: Modeling and detection of fake news from a social context perspective.” inProceedings of the 55th Hawaii International Conference on System Sciences 2022, 2022, pp. 1–10

  31. [31]

    Large system analysis for densification of cellular networks with massive mimo,

    L. Wang and Q. Liang, “Large system analysis for densification of cellular networks with massive mimo,” in2016 IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS). IEEE, 2016, pp. 956–961

  32. [32]

    Collaborative spectrum sharing based on information pooling for cognitive radio networks with channel heterogeneity,

    L. Wang, X. Chen, Z. Zhao, and H. Zhang, “Collaborative spectrum sharing based on information pooling for cognitive radio networks with channel heterogeneity,” in2011 11th International Symposium on Communications & Information Technologies (ISCIT). IEEE, 2011, pp. 483–488

  33. [33]

    Dense cross-connected ensemble convolutional neural networks for enhanced model robustness,

    L. Wang, X. Li, and Z. Zhang, “Dense cross-connected ensemble convolutional neural networks for enhanced model robustness,”arXiv preprint arXiv:2412.07022, 2024

  34. [34]

    Information theory and representation learning inspired multimodal data fusion,

    L. Wang and Y . Li, “Information theory and representation learning inspired multimodal data fusion,”IEEE MMTC Frontier, 2019

  35. [35]

    Bridging interpretability and robustness using lime-guided model refinement,

    N. Nayyem, A. Rakin, and L. Wang, “Bridging interpretability and robustness using lime-guided model refinement,”arXiv preprint arXiv:2412.18952, 2024

  36. [36]

    Enhancing adversarial robustness of deep neural networks through supervised contrastive learning,

    L. Wang, N. Nayyem, and A. Rakin, “Enhancing adversarial robustness of deep neural networks through supervised contrastive learning,”arXiv preprint arXiv:2412.19747, 2024

  37. [37]

    Explainability- driven defense: Grad-cam-guided model refinement against adversarial threats,

    L. Wang, I. I. Uddin, X. Qin, Y . Zhou, and K. Santosh, “Explainability- driven defense: Grad-cam-guided model refinement against adversarial threats,” inProceedings of the AAAI Symposium Series (AAAI) 2025, vol. 6, no. 1, 2025, pp. 49–57

  38. [38]

    Multi-scale unrectified push-pull with channel attention for enhanced corruption robustness,

    R. N. Ranabhat, L. Wang, X. Qin, Y . Zhou, and K. Santosh, “Multi-scale unrectified push-pull with channel attention for enhanced corruption robustness,” inProceedings of the AAAI Symposium Series 2025, vol. 6, no. 1, 2025, pp. 34–41

  39. [39]

    Expert-guided explainable few- shot learning for medical image diagnosis,

    I. I. Uddin, L. Wang, and K. Santosh, “Expert-guided explainable few- shot learning for medical image diagnosis,” inMICCAI Workshop on Data Engineering in Medical Imaging 2025. Springer Nature Switzerland, 2025, pp. 95–104

  40. [40]

    Ecologically valid benchmarking and adaptive attention: Scalable marine bioacoustic monitoring,

    N. R. Rasmussen, R. Rizk, L. Wang, and K. Santosh, “Ecologically valid benchmarking and adaptive attention: Scalable marine bioacoustic monitoring,”arXiv preprint arXiv:2509.04682, 2025

  41. [41]

    Shape- aware thoracic edge map chest x-ray representation for pulmonary ab- normality screening,

    S. Chataut, A. Ghimire, A. Thakur, L. Wang, and K. Santosh, “Shape- aware thoracic edge map chest x-ray representation for pulmonary ab- normality screening,” inInternational Conference on DATA ANALYTICS & LEARNING. Springer, 2024, pp. 209–220

  42. [42]

    Expert-guided explainable few- shot learning with active sample selection for medical image analysis,

    L. Wang, I. I. Uddin, and K. Santosh, “Expert-guided explainable few- shot learning with active sample selection for medical image analysis,” IEEE Journal of Biomedical and Health Informatics, 2026

  43. [43]

    Coswin: Convolution enhanced hierarchical shifted window attention for small-scale vision,

    P. Khadka, R. Rizk, L. Wang, and K. Santosh, “Coswin: Convolution enhanced hierarchical shifted window attention for small-scale vision,” arXiv preprint arXiv:2509.08959, 2025

  44. [44]

    Bridging symmetry and robustness: On the role of equivariance in enhancing adversarial robustness,

    L. Wang, I. I. Uddin, C. Zhang, X. Qin, and Y . Zhou, “Bridging symmetry and robustness: On the role of equivariance in enhancing adversarial robustness,”Advances in Neural Information Processing Systems (NeurIPS), vol. 38, pp. 159 102–159 129, 2025

  45. [45]

    Channel- selected stratified nested cross-validation for clinically relevant eeg-based parkinson’s disease detection,

    N. R. Rasmussen, R. Rizk, L. Wang, A. Singh, and K. Santosh, “Channel- selected stratified nested cross-validation for clinically relevant eeg-based parkinson’s disease detection,” in2026 IEEE Conference on Artificial Intelligence (CAI). IEEE, 2026, pp. 91–97

  46. [46]

    Winsor-cam: Human-tunable visual explanations from deep networks via layer-wise winsorization,

    C. Wall, L. Wang, R. Rizk, and K. Santosh, “Winsor-cam: Human-tunable visual explanations from deep networks via layer-wise winsorization,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

  47. [47]

    Promoting shape bias in cnns: Frequency-based and contrastive regularization for corruption robustness,

    R. N. Ranabhat, L. Wang, A. K. Patel, and K. Santosh, “Promoting shape bias in cnns: Frequency-based and contrastive regularization for corruption robustness,” inInternational Conference on Intelligent Systems and Pattern Recognition. Springer, 2025, pp. 16–26

  48. [48]

    Learning to select like humans: Explainable active learning for medical imaging,

    I. I. Uddin, L. Wang, X. Qin, Y . Zhou, and K. Santosh, “Learning to select like humans: Explainable active learning for medical imaging,” in 2026 IEEE Conference on Artificial Intelligence (CAI). IEEE, 2026, pp. 458–463

  49. [49]

    Explainable novel category discovery in semantic concept space,

    I. I. Uddin, Y . Zhou, K. Santosh, and L. Wang, “Explainable novel category discovery in semantic concept space,”arXiv preprint arXiv:2607.04548, 2026

  50. [50]

    Mechanistic inter- pretability of llm jailbreaks via internal attribution graphs,

    A. Wagle, I. I. Uddin, C. Zhang, and L. Wang, “Mechanistic inter- pretability of llm jailbreaks via internal attribution graphs,”arXiv preprint arXiv:2607.07903, 2026

  51. [51]

    Frequency- aware contrastive learning for robust shape-biased convolutional neural networks,

    R. N. Ranabhat, L. Wang, A. K. Patel, and K. Santosh, “Frequency- aware contrastive learning for robust shape-biased convolutional neural networks,”Pattern Recognition Letters, 2026. 12

  52. [52]

    Explainability-guided defense: Attribution-aware model refinement against adversarial data attacks,

    L. Wang, M. N. Nayyem, A. Al Rakin, K. Santosh, C. Zhang, and Y . Zhou, “Explainability-guided defense: Attribution-aware model refinement against adversarial data attacks,” in2025 IEEE International Conference on Data Mining (ICDM). IEEE, 2025, pp. 1585–1592

  53. [53]

    Large dimensional analysis of cooperative multicell precoding with local individual csi,

    L. Wang and Q. Liang, “Large dimensional analysis of cooperative multicell precoding with local individual csi,” in2016 IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS). IEEE, 2016, pp. 89–94

  54. [54]

    An adaptive method for data reduction in the internet of things,

    Y . Fathy, P. Barnaghi, and R. Tafazolli, “An adaptive method for data reduction in the internet of things,” inProceedings of the IEEE World Forum on Internet of Things (WF-IoT), 2018, pp. 1–6

  55. [55]

    An adaptive strategy for quality-based data reduction in wireless sensor networks,

    S. Santini and K. Römer, “An adaptive strategy for quality-based data reduction in wireless sensor networks,” inProceedings of the 3rd International Conference on Networked Sensing Systems (INSS), 2006, pp. 29–36

  56. [56]

    Adaptive stream resource management using kalman filters,

    A. Jain, E. Y . Chang, and Y .-F. Wang, “Adaptive stream resource management using kalman filters,” inProceedings of the ACM SIGMOD International Conference on Management of Data, 2004, pp. 11–22

  57. [57]

    Energy efficient information collection with the arima model in wireless sensor networks,

    C. Liu, K. Wu, and M. Tsao, “Energy efficient information collection with the arima model in wireless sensor networks,” inProceedings of the IEEE GLOBECOM, vol. 5, 2005, pp. 2470–2474

  58. [58]

    A new approach to linear filtering and prediction problems,

    R. E. Kalman, “A new approach to linear filtering and prediction problems,”Journal of Basic Engineering, vol. 82, no. 1, pp. 35–45, 1960

  59. [59]

    R. G. Brown,Statistical Forecasting for Inventory Control. McGraw-Hill, 1959

  60. [60]

    Haykin,Adaptive Filter Theory, 4th ed

    S. Haykin,Adaptive Filter Theory, 4th ed. Prentice Hall, 2002

  61. [61]

    A survey on concept drift adaptation,

    J. Gama, I. Žliobait ˙e, A. Bifet, M. Pechenizkiy, and A. Bouchachia, “A survey on concept drift adaptation,”ACM Computing Surveys, vol. 46, no. 4, pp. 1–37, 2014

  62. [62]

    Incremental on-line learning: A review and comparison of state-of-the-art algorithms,

    V . Losing, B. Hammer, and H. Wersing, “Incremental on-line learning: A review and comparison of state-of-the-art algorithms,”Neurocomputing, vol. 275, pp. 1261–1274, 2018

  63. [63]

    Learning from time-changing data with adaptive windowing,

    A. Bifet and R. Gavaldà, “Learning from time-changing data with adaptive windowing,” inProceedings of the SIAM International Conference on Data Mining (SDM), 2007, pp. 443–448

  64. [64]

    G. C. Goodwin and K. S. Sin,Adaptive Filtering, Prediction and Control. Prentice Hall, 1984

  65. [65]

    Evaluating short-term forecasting of multiple time series in iot environments,

    C. Tzagkarakis, P. Charalampidis, S. Roubakis, A. Fragkiadakis, and S. Ioannidis, “Evaluating short-term forecasting of multiple time series in iot environments,” in2022 30th European Signal Processing Conference (EUSIPCO). IEEE, 2022, pp. 1116–1120

  66. [66]

    Sliding- window signatures for time series: Application to electricity demand forecasting,

    N. Drobac, M. Brégère, J. de Vilmarest, and O. Wintenberger, “Sliding- window signatures for time series: Application to electricity demand forecasting,”arXiv preprint arXiv:2510.12337, 2025

  67. [67]

    Friedman, T

    J. Friedman, T. Hastie, and R. Tibshirani,The Elements of Statistical Learning, 2nd ed. Springer, 2009

  68. [68]

    Benchmarking tinyml systems: Challenges and direction,

    C. Banbury, V . J. Reddi, M. Lam, W. Fu, A. Fazel, J. Holleman, S. Natara- jan, O. Regev, P. Whatmough, and D. Wu, “Benchmarking tinyml systems: Challenges and direction,”arXiv preprint arXiv:2003.04821, 2020

  69. [69]

    Warden and D

    P. Warden and D. Situnayake,TinyML: Machine Learning with Tensor- Flow Lite on Arduino and Ultra-Low-Power Microcontrollers. O’Reilly Media, 2019

  70. [70]

    Design and evaluation of a versatile and efficient receiver-initiated link layer for low-power wireless,

    P. Dutta, S. Dawson-Haggerty, Y . Chen, C.-J. M. Liang, and A. Terzis, “Design and evaluation of a versatile and efficient receiver-initiated link layer for low-power wireless,” inProceedings of the ACM Conference on Embedded Networked Sensor Systems (SenSys), 2010, pp. 1–14

  71. [71]

    Telos: Enabling ultra-low power wireless research,

    J. Polastre, R. Szewczyk, and D. Culler, “Telos: Enabling ultra-low power wireless research,” inProceedings of the International Symposium on Information Processing in Sensor Networks (IPSN), 2005, pp. 364–369

  72. [72]

    Taming the underlying challenges of reliable multihop routing in sensor networks,

    A. Woo, T. Tong, and D. Culler, “Taming the underlying challenges of reliable multihop routing in sensor networks,” inProceedings of the ACM Conference on Embedded Networked Sensor Systems (SenSys), 2003, pp. 14–27

  73. [73]

    Ridge regression: Biased estimation for nonorthogonal problems,

    A. E. Hoerl and R. W. Kennard, “Ridge regression: Biased estimation for nonorthogonal problems,”Technometrics, vol. 12, no. 1, pp. 55–67, 1970

  74. [74]

    Regression shrinkage and selection via the lasso,

    R. Tibshirani, “Regression shrinkage and selection via the lasso,”Journal of the Royal Statistical Society: Series B, vol. 58, no. 1, pp. 267–288, 1996

  75. [75]

    On field calibration of an electronic nose for benzene estimation in an urban pollution monitoring scenario,

    S. De Vito, E. Massera, M. Piga, L. Martinotto, and G. Di Francia, “On field calibration of an electronic nose for benzene estimation in an urban pollution monitoring scenario,”Sensors and Actuators B: Chemical, vol. 129, no. 2, pp. 750–757, 2008, dataset: UCI Machine Learning Repository Air Quality. [Online]. Available: https://archive.ics.uci.edu/ml/dat...

  76. [76]

    R. J. Hyndman and G. Athanasopoulos,Forecasting: Principles and Practice, 3rd ed. OTexts, 2021, [Online]. Available: https://otexts.com/ fpp3

  77. [77]

    G. E. P. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung,Time Series Analysis: Forecasting and Control, 5th ed. Wiley, 2015

  78. [78]

    J. D. Hamilton,Time Series Analysis. Princeton University Press, 1994

  79. [79]

    Beach weather stations — automated sensors,

    City of Chicago, “Beach weather stations — automated sensors,” Chicago Data Portal, 2016, [Online]. Available: https://data.cityofchicago. org/Parks-Recreation/Beach-Weather-Stations-Automated-Sensors/ k7hf-8y75

  80. [80]

    Intel lab data,

    S. Madden, “Intel lab data,” CRAWDAD Dataset, UC Berkeley, 2004, [Online]. Available: https://db.csail.mit.edu/labdata/labdata.html

Showing first 80 references.