Temporal Dictionary Ensemble in HIVE-COTE 2.0

• The paper introduces TDE as a novel classifier component in HIVE-COTE 2.0 that boosts accuracy by over 1% via randomized symbolic representations.
• TDE employs a bag-of-words approach by discretizing time series into symbolic words, ensuring both phase-independent and shape-sensitive discrimination.
• Empirical results confirm TDE’s effectiveness on univariate and multivariate datasets, reinforcing its role in state-of-the-art time series classification.

The Temporal Dictionary Ensemble (TDE) is a constituent classifier introduced in HIVE-COTE 2.0, an advanced meta-ensemble for time series classification. TDE is designed to enhance accuracy via a sophisticated bag-of-words based dictionary approach, complementing other classifiers within the ensemble. It operates by constructing ensembles over symbolic representations extracted from temporal sequences, addressing the need for both phase-independent and shape-sensitive discrimination. TDE replaced previous dictionary-based components in HIVE-COTE and contributes to the ensemble’s improved empirical performance on large benchmark datasets, demonstrating efficacy for both univariate and multivariate series (Middlehurst et al., 2021).

1. Overview and Context

TDE, as part of HIVE-COTE 2.0, serves a fundamental role by exploiting symbolic (dictionary-based) methods for time series classification. Dictionary-based classifiers represent local subsequences (words) from a time series using discrete symbols, modeling frequency and distributional properties within a classification framework. TDE, together with interval-based, shapelet-based, and convolutional modules, provides a diverse set of inductive biases, yielding the observed gains in ensemble accuracy over earlier state-of-the-art systems (Middlehurst et al., 2021).

2. Dictionary-Based Classification in Time Series

In dictionary-based approaches, time series are partitioned into overlapping or non-overlapping subsequences, typically using a sliding window. Each subsequence is transformed into a symbolic word via quantization or representation techniques (such as SAX, SFA, or others). The resulting bag-of-words (histogram of symbolic words) encodes the local structural information. Classifiers are then trained on features derived from word occurrence frequencies.

The utility of such methods in time series arises from their phase-independence and robustness to temporal distortions, as well as their ability to summarize repetitive patterns and motifs—crucial in domains such as bioinformatics, finance, and sensor analysis. TDE advances these principles by assembling an ensemble over multiple randomized dictionary representations, providing increased diversity and mitigating overfitting (Middlehurst et al., 2021).

3. Algorithmic Construction of the Temporal Dictionary Ensemble

While the implementation details specific to TDE are not enumerated in the citation, the typical workflow of a temporal dictionary ensemble classifier includes the following high-level steps (as informed by the function of dictionary components in related work):

1. Subsequence Extraction: The time series is segmented into multiple local windows via a sliding window mechanism.
2. Symbolic Transformation: Each subsequence is discretized to a symbolic representation, potentially using methods like SAX (Symbolic Aggregate approXimation) or SFA (Symbolic Fourier Approximation).
3. Word Histogram Construction: For each time series, the frequency histogram of generated words is assembled.
4. Ensemble Learning: Multiple models, each potentially using different parameterizations, window sizes, or symbolic mapping strategies, form the ensemble. Their outputs are aggregated, typically by averaging probabilities or voting.

TDE is designed to maximize classifier diversity through randomized dictionary construction choices, parameter perturbations, and potentially by leveraging multiple symbolization techniques. This strategy aligns with the ensemble principle underpinning HIVE-COTE 2.0.

4. Role Within HIVE-COTE 2.0

HIVE-COTE 2.0 is a hierarchical, transformation-based ensemble, composed of several base classifiers operating over different representations:

• Phase-independent shapelet methods
• Bag-of-words (dictionary-based) classifiers (TDE)
• Phase-dependent interval-based classifiers (e.g., DrCIF)
• An ensemble of ROCKET classifiers (the Arsenal)

TDE replaces earlier dictionary-based constituents and is one of the core modules contributing to overall accuracy. The ensemble structure provides both competitive and complementary classification capabilities. Ablation studies in HIVE-COTE 2.0 indicate that the synergy enabled by diverse modules like TDE and DrCIF is key to outperforming prior meta-ensembles; removal of TDE or any such module leads to statistically significant reductions in accuracy (Middlehurst et al., 2021).

5. Empirical Results and Performance Characteristics

Empirical evaluation of HIVE-COTE 2.0, including its TDE component, demonstrates statistically significant improved accuracy over prior systems across 112 univariate UCR archive datasets and 26 multivariate UEA datasets. TDE, in replacing prior dictionary modules, contributes to this observed state-of-the-art performance. Empirical studies reveal that ensembles containing TDE, combined with other module types, outperform any subset of modules, with over 1% average accuracy gain compared to HIVE-COTE 1.0 (Middlehurst et al., 2021).

6. Significance, Applications, and Future Directions

The introduction of TDE in HIVE-COTE 2.0 underscores the continued relevance of dictionary-based methods for time series classification, particularly when embedded in a carefully diversified ensemble framework. TDE's symbolic approach complements interval, shapelet, and convolutional representations, leading to superior generalization. Applications are wide-ranging and include domains where temporal noise, phase shifts, and pattern recurrence are prevalent.

A plausible implication is that further advances in dictionary-based representation, symbolization schemes, or more sophisticated ensemble learning strategies may continue to yield incremental improvements, especially when tightly integrated with other diverse base classifiers. The documented gains of TDE within HIVE-COTE 2.0 support ongoing research into richer symbolic and hybrid ensemble approaches for temporal data (Middlehurst et al., 2021).