Discriminative STL Mining Procedure

• The paper develops a framework integrating STL template enumeration, parameter optimization, and statistical testing to achieve high discrimination in time-series classification.
• Discriminative STL mining is a method that constructs formulas using real-valued predicates and temporal operators to detect anomalies in cyber-physical systems and other applications.
• Empirical performance demonstrates that beam search, permutation tests, and embedding techniques enhance scalability, interpretability, and classification accuracy across diverse domains.

A discriminative STL mining procedure is a data-driven method for extracting Signal Temporal Logic (STL) formulas that optimally distinguish between time-series (signal) classes, typically for purposes such as anomaly detection, interpretability, or requirement synthesis. These procedures integrate formal STL grammar enumeration, parameter optimization, statistical hypothesis testing, and search-space pruning to yield interpretable, minimal, and highly discriminative classifiers directly expressible in temporal logic. Methods in this family are increasingly central to cyber-physical systems, AI reasoning analysis, and controls engineering (Saveri et al., 2024, Mohammadinejad et al., 2020, Mao et al., 19 Jan 2026).

1. Formal Foundations of Discriminative STL Mining

Signal Temporal Logic (STL) extends first-order logic with temporal operators parameterized by real intervals. STL formulas $\varphi$ are constructed from atomic predicates over real-valued signals, Boolean connectives, and temporal operators such as $\mathcal{G}_{I}$ ("always") and $\mathcal{F}_{I}$ ("eventually") over intervals $I=[a,b]$:

$$\varphi ::= tt \mid p(x) \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \mathcal{U}_{I} \varphi_2$$

with atomic predicates $p(x) \equiv f(x) \geq 0$. Quantitative semantics further assign each $(\varphi, x, t)$ a "robustness" score $\rho(\varphi, x, t)$ indicating the "distance" to satisfaction/violation.

The positive class $\mathcal{D}_+$ and negative class $\mathcal{D}_-$ are finite datasets of temporal traces, typically multidimensional, possibly arising as input or output signals of a cyber-physical system. The task is to find STL formulas $\varphi^*$ such that:

• $\varphi^*$ accepts most of $\mathcal{D}_+$ and rejects most of $\mathcal{D}_-$,
• $\varphi^*$ is interpretable or minimal (structure/topological size),
• All template/parameter choices are statistically justified.

Discriminative STL mining methods explicitly search over parameterized STL template classes (possibly full grammars), select numeric values for all thresholds/bounds, and optimize a discrimination objective—typically formulated as classification error, robust-statistics separation, or information gain (Saveri et al., 2024, Mohammadinejad et al., 2020, Mao et al., 19 Jan 2026).

2. Core Methodological Components

The discriminative STL mining pipeline typically comprises the following elements:

A. STL template enumeration:

All candidate STL formula structures are recursively generated up to predefined bounds on parse-tree depth, operator count, or number of signal variables. Both future and past-time STL operators may be included. Candidate templates may be filtered via normalization heuristics to eliminate semantic duplicates (Mohammadinejad et al., 2020, Saglam et al., 2019).

B. Parameter grid search or optimization:

Each formula's numeric constants (predicate thresholds, interval bounds) are treated as tunable parameters. Search strategies range from coarse-to-fine gridding and greedy beam search (Mao et al., 19 Jan 2026), to decision-tree induction over robustness signatures (Mohammadinejad et al., 2020), to global gradient or Bayesian optimization in continuous formula embedding spaces (Saveri et al., 2024).

C. Discriminative objective function:

Formulas are scored via classification metrics (accuracy, recall, precision), robust discrimination metrics (mean robustness separation divided by combined standard deviation), or negative log-likelihood over a train/validation split (Saveri et al., 2024, Mao et al., 19 Jan 2026). For class-separating tasks, AUROC or $F_\beta$ scores may be employed (Saglam et al., 2019).

D. Pruning and search control:

To control combinatorial explosion, most approaches employ beam search, monotonicity-based branch-and-bound, or template pool expansion constrained by structural or heuristic scores. Numerical parameterizations yielding redundant prototype formulas (e.g., near-identical robustness signatures) are filtered to promote diversity and parsimony (Saveri et al., 2024).

E. Statistical significance assessment:

Some frameworks incorporate permutation-based hypothesis testing to ensure strict type-I error control. For example, the Westfall–Young max-T permutation scheme selects p-value thresholds to guarantee family-wise error rate (FWER) $\leq \alpha$ across the whole mined formula set (Duy et al., 2019, Bunker et al., 2023).

3. Statistical Discriminative Sub-trajectory Mining (SDSM)

SDSM is a canonical instance of discriminative STL/sub-trajectory mining tailored for comparing raw spatial trajectories across two groups (Duy et al., 2019). For a set of $n$ trajectories $\{\tau_i\}$ each with $m_i$ spatial points and group labels $g_i \in \{+1,-1\}$, SDSM extracts all sub-trajectories $T_i^{(s,e)}$ (consecutive points, minimal length $L$) that statistically discriminate between classes.

Key SDSM algorithmic steps:

• Similarity definition:

For equal-length sub-trajectories, use top-$K$ average of the largest pointwise Euclidean distances as $dist_K(T,T')$. Define $\epsilon$-neighborhoods as all subtrajectories of matching length within $dist_K \leq \epsilon$.

• Support computation:

For any candidate $T$, compute $sup_+(T)$ (number of $+$ group trajectories with an $\epsilon$-neighbor) and $sup_-(T)$ analogously.

• Contingency and significance:

Each $T$ induces a 2×2 table; the null $H_0$ posits label independence. Fisher's exact test with calculated hypergeometric probabilities yields exact p-values. Adjustments for multiple testing via Westfall–Young permutations determine the significance cutoff $\delta^*$, controlling FWER at prescribed $\alpha$.

• Efficient enumeration and pruning:

Sub-trajectories are indexed via a tree structure. Monotonicity ensures non-increasing support down the tree, allowing monotonic lower bounds on $p(T)$ for aggressive branch pruning. Complexity is $O(NB)$ in the worst case, but effective pruning typically yields much smaller search times and space (Duy et al., 2019).

SDSM was demonstrated on datasets with up to $10^6$ trajectories and over $10^{10}$ subtrajectories, showing practical scalability and domain-conformant pattern discovery (e.g., bottlenecks, hurricane tracks).

4. Framework Variants and Recent Extensions

Several advance discriminative STL mining via alternative search, fitting, and representation paradigms:

A. Retrieval-Augmented Bayesian Optimization:

Retrieval-augmented mining frameworks precompute a dense vector database of millions of STL formula embeddings via an STL-robustness-based kernel and use Bayesian optimization to guide formula search in this continuous semantic space (Saveri et al., 2024). Each iteration proposes a candidate semantic embedding, retrieves nearest formulae via approximate nearest neighbor (ANN) search (e.g., FAISS), and evaluates discrimination metrics $G(\varphi)$:

$$G(\varphi) = \frac{\mathbb{E}[R_\varphi|X_p] - \mathbb{E}[R_\varphi|X_n]}{\sigma(R_\varphi|X_p) + \sigma(R_\varphi|X_n)}$$

This loop proceeds without any template restriction, allowing extraction of highly interpretable, optimally discriminative formulas.

B. Decision-tree-based STL Fitting:

Template enumeration combined with robustness-based feature construction is followed by supervised decision tree induction over the set of instantiated robustness signatures, yielding STL classifiers of the form “conjunction of paths to leaf” and “disjunction over positive leaves” (Mohammadinejad et al., 2020). Accuracy is further improved by employing counterexample-guided falsification: any counterexample input violating the output requirements is incorporated and triggers retraining.

C. Discriminative Mining for LLM Signals:

For stepwise signals such as LLM confidence traces, the mining process emphasizes discriminating STL patterns with interpretable structure and numeric thresholding (e.g., forbidding sharp confidence drops late in the reasoning chain). Formulas are parameterized for maximal discrimination, further conditioned by hypernetworks to capture context sensitivity (Mao et al., 19 Jan 2026).

5. Statistical Guarantees and Parameter Tuning

Proper control of statistical error is central in discriminative STL mining:

• Family-wise error rate control:

The Westfall–Young permutation/max-T correction ensures that the likelihood of one or more false positives (i.e., sub-trajectories falsely deemed significant) is $\leq \alpha$, even for massive trajectory sets (Duy et al., 2019, Bunker et al., 2023).

• Parameter recommendations:
  • $\alpha$ is typically set at 0.05.
  • Number of permutations $B \geq 1000$ for robust significance estimation.
  • Minimum candidate sub-trajectory or formula length $L$ is chosen to balance interpretability and search tractability.
  • Discriminative objectives leverage distances (e.g., $dist_K$ with $K$ up to $L$), thresholds ($\epsilon$) derived from empirical or domain-specific tolerances.

6. Empirical Performance and Applications

Discriminative STL mining procedures have been validated across diverse domains:

Application Area	Typical Output Type	Performance
Cyber-physical systems (Mohammadinejad et al., 2020, Saveri et al., 2024)	Environment assumptions, anomaly detectors	99–100% accuracy, concise formulas
Large-scale trajectory analysis (Duy et al., 2019)	Statistically discriminative sub-trajectories	100k+ discoveries, sub-24h runtime on 256 cores
Multi-agent movement analysis (Bunker et al., 2023)	Discriminative spatiotemporal submatrices	Intuitive multi-agent motifs (NBA)
LLM confidence calibration (Mao et al., 19 Jan 2026)	STL-calibrated confidence estimators	ECE < 0.04, Brier < 0.17

Qualitative analysis reveals that mined STL patterns are often compact, directly interpretable, and match known or hypothesized domain phenomena (e.g., bottlenecks, domain-specific attack signatures in sport analytics, high-confidence subchains in LLM reasoning).

7. Limitations and Future Directions

Despite their strengths, discriminative STL mining procedures face several intrinsic and practical limitations:

• Scalability:

Template enumeration and parameter grid search are exponential in formula size and parameter count; practical approaches mitigate this with aggressive pruning, beam search, or embedding-based retrieval, but memory and runtime can still grow rapidly (Saveri et al., 2024, Mohammadinejad et al., 2020).

• Coverage:

Precomputed databases may miss entirely new pattern structures; a dynamic or online extension could incorporate newly discovered formulae (Saveri et al., 2024).

• Parameter sensitivity:

Numeric thresholds and temporal bounds may exhibit task-specific or instance-specific sensitivity, motivating joint or adaptive parameter tuning (e.g., hypernetworks) (Mao et al., 19 Jan 2026).

• Adaptation to new domains:

Formula kernels or prior measures in embedding constructions may need to be retrained for domain-specific signal distributions (Saveri et al., 2024).

Planned extensions include online adaptation to streaming data, expansion to multi-class or regression settings, and the development of syntax-guided or user-influenced enumeration to focus search towards domain-relevant patterns.

• (Duy et al., 2019) Statistically Discriminative Sub-trajectory Mining
• (Saveri et al., 2024) Retrieval-Augmented Mining of Temporal Logic Specifications from Data
• (Mohammadinejad et al., 2020) Mining Environment Assumptions for Cyber-Physical System Models
• (Mao et al., 19 Jan 2026) Confidence over Time: Confidence Calibration with Temporal Logic for LLM Reasoning
• (Saglam et al., 2019) Cause Mining and Controller Synthesis with STL
• (Bunker et al., 2023) Multi-agent statistical discriminative sub-trajectory mining and an application to NBA basketball