Dependency Controlled Learning (DCL)

• Dependency Controlled Learning (DCL) is a hierarchical method that explicitly models inter-patch dependencies across multiple temporal scales in time-series data.
• It integrates with Instance-wise Patch Normalization and Instance-level Contrastive Module to balance local stabilization and global instance discrimination.
• Ablation studies show that a two-layer DCL with adaptive dynamic windows significantly boosts accuracy and reduces overfitting compared to simpler dependency models.

Dependency Controlled Learning (DCL) is a latent-level component of the Dependency Controlled Pre-training framework (DeCoP) for time-series representation learning. In this setting, DCL denotes a hierarchical strategy that explicitly models inter-patch dependencies across multiple temporal scales, complementing Instance-wise Patch Normalization (IPN) at the input level and the Instance-level Contrastive Module (ICM) at the global representation level. Its stated role is to address dynamic temporal dependencies that evolve under distribution shifts and multi-scale patterns, thereby improving generalization in both in-domain and cross-domain settings (Wu et al., 18 Sep 2025).

1. Position within the DeCoP framework

Within DeCoP, DCL is not an isolated training recipe but the central latent-level mechanism for modeling temporal structure. The framework is organized around three components: IPN, DCL, and ICM. IPN mitigates distribution shifts while preserving the unique characteristics of each patch; DCL performs hierarchical and controllable modeling of inter-patch dependencies; and ICM enhances global generalization by learning instance-discriminative representations from time-invariant positive pairs (Wu et al., 18 Sep 2025).

Component	Function	Key features
IPN	Patch and instance normalization	Local+global, adaptive
DCL	Latent, hierarchical multi-scale temporal modeling	Dynamic windows, residual, hierarchy
ICM	Instance-level contrastive positive pair generation	Frequency-based time-invariance

This organization is important because DCL operates after patch formation and normalization, rather than directly on raw sequences. A plausible implication is that the method separates input stabilization, dependency modeling, and global alignment into distinct stages, rather than forcing a single module to handle all three requirements simultaneously.

2. Motivation and problem setting

The motivation for DCL is the claim that modeling dynamic temporal dependencies is a critical challenge in time-series pre-training, particularly when dependencies evolve because of distribution shifts and multi-scale patterns. Existing frameworks are described as failing to capture the complex interactions of short- and long-term dependencies, which makes them susceptible to spurious correlations that degrade generalization. DCL is introduced specifically to model inter-patch dependencies at both short and long temporal ranges in a manner that adapts to the evolving, dynamic nature of real-world time series (Wu et al., 18 Sep 2025).

The design is explicitly hierarchical because time series are taken to exhibit multi-scale structure. Learning at both local and global scales is presented as a way to avoid overfitting to spurious, time-localized correlations and to strengthen generalization. The “controlled” aspect refers to the ability to dynamically adjust the temporal window used for modeling, in contrast to fixed-receptive-field alternatives such as global attention or fixed kernels. This suggests that DCL is intended as a structured alternative to indiscriminate global linking, especially when temporal relations are unstable under shift.

3. Architectural formulation

The DCL pipeline begins after patching and normalization. IPN is applied to a time series $X$ to produce patches $X_p$, and ICM generates $(X_p, X'_p)$, described as denoised and masked versions of the time series. DCL then processes each patch sequence through a series of hierarchical blocks, each associated with its own temporal scope (Wu et al., 18 Sep 2025).

The latent embedding stage is given by

$$Z_p = X_p W_p + \text{Bias}, \qquad Z_p \leftarrow Z_p + W_{\text{pos}}$$

where the patches are projected into model space and augmented with learnable relative positional embedding.

Multi-scale modeling is then realized through a set of window sizes $\{W_k\}$, with typical sizes satisfying $W_1 < W_2 < \dots < W_K$. At layer $k$, the latent sequence is partitioned into temporal windows of length $W_k$, using

$$Z_w = \text{Flatten}(\text{Reshape}(\text{Padding}(Z_p), W_k)).$$

This yields windowed blocks that are then processed by a local temporal learner $\mathcal{E}_{enc}$, for example a Linear or MLP module: $X_p$0 A residual update is applied as

$X_p$1

Stacking multiple such blocks with progressively increasing $X_p$2 forms what is described as a hierarchical, global-to-local dependency pyramid. In operational terms, DCL therefore alternates between window construction, within-window temporal learning, and residual propagation. The paper’s ablation summary identifies a two-layer DCL with window sizes $X_p$3 as achieving the best tradeoff between accuracy and efficiency, indicating that the hierarchy is intended to be controllable rather than arbitrarily deep (Wu et al., 18 Sep 2025).

4. Training objective and interaction with IPN and ICM

DCL is optimized jointly with reconstruction and contrastive objectives. Its outputs are used both for masked patch reconstruction and for instance-level contrastive learning. The joint loss is

$X_p$4

where $X_p$5 is the MSE on masked patch reconstruction and $X_p$6 is the instance-level contrastive loss computed from the representations produced by the final DCL block (Wu et al., 18 Sep 2025).

The interaction among modules is central to the formulation. Upstream, DCL benefits from IPN because patch features are made more stable and distribution-aligned before dependency modeling begins. It also benefits from ICM, which supplies positive, time-invariant counterparts used to learn robust general representations. Downstream, the DCL-encoded sequence becomes the shared representation consumed by both reconstruction and contrastive objectives. This suggests that DCL is intended to mediate between local patch regularization and global instance discrimination, rather than serving solely as a forecasting backbone.

5. Empirical behavior, ablations, and generalization

DeCoP is reported to achieve state-of-the-art results on ten datasets with lower computing resources, including a 3% MSE improvement on ETTh1 over PatchTST using only 37% of the FLOPs. The contribution attributed to DCL in the detailed summary is its explicit, dynamic modeling of dependencies under domain and label shift, with particularly strong performance in cross-domain settings (Wu et al., 18 Sep 2025).

The ablation evidence is framed around replacing DCL with simpler dependency assumptions. A patch-independent variant, which “forgets dependencies,” and a patch-dependent variant, which treats all patches as equally connected, both degrade performance. In-domain, the degradation reaches up to $X_p$7 for the patch-dependent design on classification. In cross-domain evaluation, the gaps widen to $X_p$8 for patch-independent and $X_p$9 for patch-dependent variants on classification tasks. The reported interpretation is that controllable, hierarchical dependency modeling is more effective than either ignoring dependencies or imposing uniform global dependence.

A further reported observation is that DCL reduces overfitting, as indicated by a smaller gap between training and validation loss. The summary also states that IPN plus DCL preserve both global statistical consistency and patch-level semantic variation, aiding adaptation under distribution shift. This suggests that DCL’s empirical role is not only to improve accuracy, but also to stabilize representation learning when temporal correlations are non-stationary.

6. Terminological ambiguity and related uses of “DCL”

The acronym “DCL” is highly polysemous in the arXiv literature. It has been used for “Dirichlet Continual Learning” in continual learning for task-oriented dialogue (Zeng et al., 2023), “Deep Controlled Learning” in inventory control (Temizöz et al., 2020), “Dynamic Curriculum Learning” for imbalanced classification (Wang et al., 2019), “Distributional Constraint Learning” in mixed-integer stochastic optimization (2211.11392), “Deep Collaborative Learning” for visual recognition (Wang et al., 2017), “Deep Coherence Learning” in ultrasound beamforming (Cho et al., 2023), and “Deep Companion Learning” for generalization through historical consistency (Zhu et al., 2024).

In that context, “Dependency Controlled Learning” refers specifically to the latent-level hierarchical dependency-modeling strategy inside DeCoP, rather than to a standalone family of methods spanning multiple domains. A common misconception is therefore to treat DCL as a single established method class across machine learning. The evidence instead indicates that the same acronym has been independently reused for unrelated frameworks, and that Dependency Controlled Learning is the time-series-specific usage associated with DeCoP (Wu et al., 18 Sep 2025).