Industrial sensor streams, such as water treatment plants, server fleets, and spacecraft telemetry, are labeled by tired humans. Analysts confirm or reject flagged windows, and active learning chooses which windows they see. Standard active learning sends them the most ambiguous windows, which are exactly the ones they get wrong most often. The wrong labels then distort the model, and the distorted model picks even worse windows to query. This proposal asks whether a different notion of similarity between windows can absorb annotator error by construction rather than by after-the-fact cleaning.
That notion is the behavioral pseudometric from quantitative bisimulation. It is a distance under which two system states count as close if they produce similar outputs now and lead to similar futures. It is defined as the fixed point of a recursion that discounts and averages over next states. Our central observation is simple. If the "output" in that recursion is the analyst's label rate, and analysts err with class-dependent rates, then noisy labels rescale every behavioral distance by the same constant and leave the geometry unchanged. We would build an active labeling loop on this fact. It queries by behavioral coverage instead of boundary ambiguity. It pools noisy labels within behavioral neighborhoods. It handles annotator fatigue explicitly, because fatigue is the one case where the invariance breaks.
Core Claim
Discretize multivariate windows into $K$ abstract states $s$ with empirically estimated transitions $P(\cdot\mid s)$. Let $\tilde y(s)$ be the probability that an annotator labels a window from $s$ as anomalous. The label-driven behavioral metric is the fixed point of
$$d(s,s') = |\tilde y(s)-\tilde y(s')| + c\, W_d\big(P(\cdot\mid s),P(\cdot\mid s')\big),$$
where $c\in(0,1)$ is a discount and $W_d$ is the Wasserstein-1 (earth-mover) distance whose ground cost is $d$ itself. This follows the construction of Ferns et al. and Desharnais et al..
Suppose binary labels flip with stationary class-conditional rates $\rho_0$ (normal to anomalous) and $\rho_1$ (anomalous to normal). Then $\tilde y = \rho_0 + (1-\rho_0-\rho_1)\,y$, where $y$ is the clean anomaly probability. Because the recursion is positively homogeneous, with $W_{\alpha d}=\alpha W_d$, the noisy fixed point is exactly
$$d_{\text{noisy}} = (1-\rho_0-\rho_1)\, d_{\text{clean}}.$$
Every ranking, every nearest-neighbor set and every $\epsilon$-ball aggregation, after rescaling $\epsilon$, is unchanged. Annotator noise in this regime costs only estimation variance, not bias.
The invariance fails in two cases. The first is time-varying noise $Q(t)$, which Learning from Time Series under Temporal Label Noise shows is the realistic case. The second is instance-dependent noise, where some anomaly types are systematically missed. In both cases the distortion is bounded by the sup-norm deviation of the per-state noise map from a single affine map, times $1/(1-c)$, which is the standard contraction bound. This tells us exactly what must be estimated: the time-indexed affine factor, not a full clean posterior.
What would be new is the following. The proposal uses a behavioral metric as the object that absorbs label noise. It derives when this is exact and when it fails. It then builds query and relabeling rules around those conditions.
The direction does not claim to unify the three fields. The bisimulation metric is the mechanism, multivariate sensor streams are the regime, and the fallible oracle is the constraint being stress-tested.
Research Program
Representation and abstraction. We first build the abstract chain. Windows are embedded with a causal, channel-clustered encoder in the style of Cluster-Aware Causal Mixer for Online Anomaly Detection in Multivariate Time Series. That model masks future time steps and groups correlated channels so that spurious cross-channel correlations do not drive the embedding. We reuse its encoder, not its thresholding.
Embeddings are clustered into $K$ states (we sweep $K$ from 100 to 1000), and transition counts give $P$. Estimation error in $P$ is tolerable because behavioral pseudometrics are Lipschitz in the transition kernel, as shown by Feng et al. and Distribution-based bisimulation for labelled Markov processes. We will report the empirical sensitivity to both $K$ and window count.
Metric computation. The metric is computed with Sinkhorn Value Iteration from Bisimulation Metrics are Optimal Transport Distances, and Can be Computed Efficiently. That work recasts the fixed point as an entropic optimal-transport problem, with per-iteration cost $O(n^2S^2)$ for sparse kernels. Sparse kernels are exactly what windowed sensor chains produce. Warm-starting across labeling rounds keeps recomputation cheap.
Collapse in sparse labels. Labels are extremely sparse, and Towards Robust Bisimulation Metric Learning shows that identical outputs collapse such metrics. We therefore add an observation term $\beta\,\|o(s)-o(s')\|$ built from label-free behavioral-primitive features: drift, spike, oscillation, repetition and switching, as defined in Between Zeros and Ones. This term breaks pure homogeneity. Noise now acts as a change of weight from $\beta$ to $\beta/(1-\rho_0-\rho_1)$, which we undo using an estimate of the affine factor obtained from a small set of repeatedly labeled anchor windows.
Active loop. Each round queries windows from the abstract states that maximize uncovered probability mass within behavioral $\epsilon$-balls. This rule selects representatives of behavioral regions rather than boundary windows, which directly targets the boundary paradox described in An active semi-supervised deep learning model for human activity recognition.
Relabeling is triggered only where labels inside a ball disagree more than the estimated noise rate predicts. This is a cheap, local version of the diversity-versus-certainty tradeoff from Re-Active Learning. It avoids that paper's global impact computation, which took over 7 hours per selection on relation extraction.
For fatigue, each label is tagged with its query time. We estimate $Q(t)$ with the continuous estimator of Learning from Time Series under Temporal Label Noise and map each label back through its own affine factor before it enters $\tilde y$. This restores the invariance. The anomaly score of a new window is the metric-kernel-weighted label rate of its abstract neighbors.
Theory. The analysis has three pieces. The first is the exact scaling result. The second is the deviation bound for time-varying and instance-dependent noise. The third is a finite-sample bound that combines label-count variance per ball with transition-estimation error, building on the empirical estimation bounds in A Generalized Bisimulation Metric of State Similarity between Markov Decision Processes.
Baselines
Naive reference points. Two simple references come first. One is random querying with a single label per window and the same kernel scorer. The other is the label-free causal mixer itself, which shows what labels buy at all. Cluster-Aware Causal Mixer for Online Anomaly Detection in Multivariate Time Series reports best-F1 improvements of 8.71% on WADI and 8.65% on PSM over the next-best models. These are relative gains, and no absolute values appear in the file, so we will rerun it ourselves on every split.
Strongest active-learning competitor. This is RLAD from RLAD: Time Series Anomaly Detection through Reinforcement Learning and Active Learning. RLAD combines a Q-learning detector, margin-sampling queries and label propagation to neighboring segments. It is the closest label-efficient anomaly detector and it already propagates labels, so it is a real competitor on pooling. Its reported numbers are nearly 0.8 F1 on Yahoo A1 and 1.0 on A2 with 10% labels, and a 3x higher F1 than unsupervised methods on KPI with 0.1% labels. Those numbers are univariate and noise-free. We will run it with the identical multivariate encoder, label budget and simulated annotators.
Strongest fallible-oracle competitor. This is SoQal with its consistency-based acquisition from SoQal. SoQal gates oracle queries with a learned misclassification predictor. It reports an AUC of about 0.59 under 80% injected noise, versus about 0.50 for traditional methods, on cardiac data. We also include a margin-sampling detector trained with Forward Temporal Loss, so that temporal noise correction is available to the competitors and not only to us.
Fair comparison. All methods share encoder weights, label budget, annotator simulator and hyperparameter search budget. Beating them means two things. First, our best-F1 under fatigue noise is higher on at least three of SWaT, WADI, HAI, PSM and SMD. Second, the gap between our noisy-annotator and clean-annotator F1 is materially smaller than theirs.
Ablations that isolate the key ingredient. There are three. The first replaces the behavioral metric with Euclidean distance on the same embeddings while keeping coverage queries and ball pooling. The second sets $c=0$, so that distances ignore the future. The third removes the per-label temporal affine correction.
Why not just…?
Why not just average labels over Euclidean nearest neighbors, as RLAD's label propagation does? This is the main threat. Averaging labels over any fixed neighborhood is also invariant to stationary class-conditional noise, so the scaling result alone does not favor bisimulation. The claim that must be earned is that dynamics-aware neighborhoods pool windows that genuinely share labels. One example is a slow drift that looks normal pointwise but leads into an attack state, which point-wise distances miss according to Between Zeros and Ones. Another is that the label term feeds back into the geometry. The Euclidean and $c=0$ ablations settle this directly. If they match the full method, we will report the contribution as noise-aware coverage pooling and drop the bisimulation claim.
Why not just relabel every query several times and take a majority vote? Repeated labeling works but spends budget linearly. Re-Active Learning shows that choosing when to relabel matters, but its impact-based choice is very expensive. Majority vote also does nothing against systematic bias. Evaluating the Impact of Coder Errors on Active Learning shows such bias makes active learning worse than random sampling. We include fixed three-times relabeling at equal total budget as a baseline.
Why not just train a standard detector with a temporal noise-corrected loss? Forward Temporal Loss lowers error on the moving dataset to about 4.2%, versus about 14.9% for the best static method, in Learning from Time Series under Temporal Label Noise. It is strong offline. However, it has no acquisition rule, and margin sampling still routes boundary windows to the annotator. Our margin-plus-temporal-loss baseline tests whether correction alone suffices once queries are active.
Proposed Experiments
Datasets. We use SWaT, WADI and HAI as the primary datasets, with PSM and SMD as secondary ones.
Simulated annotators. We simulate four noise regimes. The first is stationary class-conditional flips, where the theory predicts exact invariance. The second is fatigue noise whose flip rate rises within each labeling session and resets between sessions. The third is primitive-dependent noise, where annotators under-report drift and spike attacks more than oscillation. The fourth is a mixture of the three.
The primitive labels come from the automatic assignment in Between Zeros and Ones, which agrees with manual annotation at 98.3% (Cohen's κ 0.978). This makes the structured noise realistic rather than uniform.
Multiclass task. We also run a multiclass primitive-identification task on SWaT. There, Between Zeros and Ones reports that a Random Forest drops from 85.44% binary macro F1 to 37.84% multiclass macro F1, which leaves clear headroom. The scaling result is stated for binary labels, so the multiclass case is an empirical extension.
Metrics. We report best-F1 and AUC versus label budget (0.1% to 10% of windows), the clean-to-noisy F1 gap, the fraction of queries landing on windows the annotator gets wrong, and wall-clock time per acquisition round.
Decisive checks. There are three, in order. First, under stationary noise the measured correlation between noisy and clean metric matrices should be near 1, with slope near $1-\rho_0-\rho_1$. If not, the abstraction or estimation step is broken. Second, under fatigue noise the uncorrected metric should degrade in proportion to the variance of $\rho(t)$, and the temporal correction should recover it. Third, the full method must beat the Euclidean-neighborhood ablation at equal budget on multivariate data. The third check is the one that determines whether quantitative bisimulation earns its place.
References
- Metrics for Finite Markov Decision Processes
- Metrics for Labeled Markov Systems
- Learning under Temporal Label Noise
- Cluster-Aware Causal Mixer for Online Anomaly Detection in Multivariate Time Series
- When Equivalence and Bisimulation Join Forces in Probabilistic Automata
- Distribution-Based Bisimulation for Labelled Markov Processes
- Bisimulation Metrics are Optimal Transport Distances, and Can be Computed Efficiently
- Towards Robust Bisimulation Metric Learning
- Between Zeros and Ones: Behavioral Characterization Beyond Binary Labeling Across Public ICS Datasets
- An active semi-supervised deep learning model for human activity recognition
- Re-Active Learning: Active Learning with Relabeling
- A Generalized Bisimulation Metric of State Similarity between Markov Decision Processes: From Theoretical Propositions to Applications
- RLAD: Time Series Anomaly Detection through Reinforcement Learning and Active Learning
- SoQal: Selective Oracle Questioning for Consistency Based Active Learning of Cardiac Signals
- Evaluating the Impact of Coder Errors on Active Learning