Future Decomposition Network (FDN)

• Future Decomposition Network is an interpretable spatiotemporal forecasting model that decomposes future trajectories into a compact set of learned patterns.
• It combines a classifier over latent future motifs with an interpolation operator to generate node-level forecasts, enhancing both clarity and performance.
• Empirical results on hydrology, traffic, and energy datasets show FDN outperforms comparable methods with fewer parameters and faster runtime.

Searching arXiv for recent papers on “Future Decomposition Network” and closely related FDN usages to ground the article in current literature.
Future Decomposition Network (FDN) is an interpretable neural architecture for spatiotemporal forecasting that predicts future trajectories by decomposing them into a small set of learned reusable future patterns and generating node-level forecasts as soft interpolations of those patterns [2606.25201]. In its exact formulation, FDN addresses multi-node, multi-step prediction over spatially dependent time series such as hydrologic, traffic, and energy systems, combining a classifier over latent future patterns with an interpolation operator that reconstructs the forecast horizon from those patterns [2606.25201]. The acronym “FDN” is overloaded in the literature and has also been used for unrelated concepts including Feature Decomposition Network, Flow Based Image Denoising Neural Network, and Frequency-aware Decomposition Network; however, the expansion Future Decomposition Network is explicitly associated with an interpretable spatiotemporal forecasting model in 2026 work on forecasting with future decomposition [2302.05031] [2105.04746] [2104.05160] [2604.12905].

1. Definition and scope

FDN is formulated for spatiotemporal systems comprising (N) spatially distributed but interdependent entities, each observed over (F) features and linked by a dependency graph (G=(V,E)) [2606.25201]. At time (\tau), the node features are (\mathbf{X}\tau \in \mathbb{R}^{N \times F}), a sample of (T) contiguous time steps is (X \in \mathbb{R}^{N \times T \times F}), and one feature is selected as the forecast target (\mathbf{Y} \in X) [2606.25201]. The learning problem is to infer the next (O) target values from the previous (I) observations by minimizing
$$
\arg\min{\theta} L!\left(\mathbf{Y}{(\tau+1):(\tau+O)}, \mathcal{F}{\theta}(X_{(\tau-I+1):\tau}; G)\right).
$$

The distinguishing property of FDN is that it does not treat the forecast operator as an opaque decoder. Instead, it assumes that the space of possible future segments is highly redundant and can be represented by a comparatively small library of learned temporal patterns (\hat{\mathbb{F}} \in \mathbb{R}^{O \times K}), where (K) is the number of patterns [2606.25201]. The model then classifies the observed past into a soft distribution over those (K) patterns and reconstructs the forecast as their weighted interpolation. This yields an intrinsically interpretable prediction mechanism because the forecast can be inspected directly in terms of the activated future patterns and their weights [2606.25201].

The paper positions this design against common forecast operators such as identity, fully connected, convolutional, autoregressive, and attention-based mappings, arguing that sophisticated encoders are often paired with comparatively opaque past-to-future transformations [2606.25201]. A plausible implication is that FDN’s main novelty lies less in representation learning alone than in making the future-generation step itself explicit and inspectable.

2. Core forecasting mechanism

At the highest level, FDN consists of a classifier and an interpolator:
$$
\begin{aligned}
\text{Classifier}(X_{(\tau-I+1):\tau}, G) &\rightarrow \mathbf{\mathcal{C} \in \mathbb{R}^{N \times K}} \
\text{Interpolator}(\mathbf{\mathcal{C}, \hat{\mathbb{F}}}) &\rightarrow \hat{\mathbf{Y}}{(\tau+1):(\tau+O)} \in \mathbb{R}^{N \times O}.
\end{aligned}
$$
Here (\mathbf{\mathcal{C}}) is a node-wise soft classification over the (K) learned future patterns, and (\hat{\mathbb{F}}) is the pattern dictionary [2606.25201]. Forecast generation is then performed by
$$
\mathbf{\mathcal{C}\hat{\mathbb{F}}^{T} \rightarrow \hat{\mathbf{Y}}{(\tau+1):(\tau+O)} \in \mathbb{R}^{N \times O}.
$$
For node (n), the forecast is therefore
$$
\hat{\mathbf{Y}}{n,:}=\mathbf{\mathcal{C}}{n,:}\hat{\mathbb{F}}^{T}.
$$`

This mechanism is interpretable because each column of (\hat{\mathbb{F}}) is an explicit (O)-step temporal pattern and each row of (\mathbf{\mathcal{C}}) is an explicit probability distribution over those patterns [2606.25201]. Hard classification would choose one future prototype; FDN instead uses soft classification so that the final forecast is a smoother interpolation among multiple plausible futures [2606.25201].

The future-decomposition hypothesis is motivated by the empirical redundancy of training futures. The paper considers a matrix of all known future segments,
$$
\mathbb{F} \in \mathbb{R}^{N \cdot O \times B},
$$
and argues that it can be approximated by a much smaller set of fundamental patterns [2606.25201]. In Wabash River, about (K=200) patterns are reported to represent 70 years of localized streamflow futures across 1,276 subbasins [2606.25201]. This suggests that FDN is effectively a learned low-rank future model, though the paper learns the pattern matrix by SGD rather than by a fixed preprocessing method such as SVD [2606.25201].

3. Classifier architecture and conditioning

FDN does not classify directly from raw input windows. Its classifier builds node-specific spatiotemporal embeddings through learned node embeddings, Localized Dynamic Attention (LDA), graph convolution, periodic embeddings, and a GRU [2606.25201]. The pipeline is summarized as
$$
\begin{aligned}
\sigma(\mathbf{E}\mathbf{E}^{T}) &\rightarrow G \
\text{LDA}(X_{(\tau-I+1):\tau}, \mathbf{E}) &\rightarrow X^{*}_{(\tau-I+1):\tau} \in \mathbb{R}^{N \times I \times F} \
[X^{*}_{(\tau-I+1):\tau}, \mathbf{E}] &\rightarrow X^{*}_{(\tau-I+1):\tau} \in \mathbb{R}^{N \times I \times (F+D)} \
\text{Chebyshev-GCN}(X^{*}_{(\tau-I+1):\tau}, G) &\rightarrow \mathcal{E}^{s} \in \mathbb{R}^{N \times I \times H} \
[\mathcal{E}^{s}, \mathbf{P}_{(\tau-I+1):\tau}] &\rightarrow \mathcal{E}^{s} \in \mathbb{R}^{N \times I \times (H+\rho)} \
\text{GRU}(\mathcal{E}^{s}) &\rightarrow \mathbf{\mathcal{E}^{st} \in \mathbb{R}^{N \times H}} \
\sigma(\text{FC}(\mathbf{\mathcal{E}^{st}})) &\rightarrow \mathbf{\mathcal{C} \in \mathbb{R}^{N \times K}}.
\end{aligned}
$$

The learned node embeddings (\mathbf{E} \in \mathbb{R}^{N \times D}) serve three roles: they condition the attention mechanism, they are concatenated with filtered node features as spatial context, and they induce the graph through (\sigma(\mathbf{E}\mathbf{E}^{T}) \rightarrow G) [2606.25201]. This makes FDN usable both with pre-defined graphs and with learned inter-node dependency structure.

LDA is introduced for multivariate inputs with (F>1), where the relevance of features varies across nodes and time. It computes
$$
\begin{aligned}
W &\in \mathbb{R}^{D \times I \times F} \
\sigma(\mathbf{E}W) &\rightarrow \hat{A} \in \mathbb{R}^{N \times I \times F} \
\hat{A} \odot X_{(\tau-I+1):\tau} &\rightarrow X^{*}_{(\tau-I+1):\tau} \in \mathbb{R}^{N \times I \times F}.
\end{aligned}
$$
Because (D \ll N), this produces node-conditioned feature-time attention without a full (N \times I \times F) parameterization [2606.25201].

Temporal context is supplemented by periodic embeddings (\mathbf{P} \in \mathbb{R}^{M \times \rho}), retrieved through a periodic index function (p(t)) that maps timestamps to seasonal positions [2606.25201]. The retrieved sequence
$$
\mathbf{P}_{(\tau-I+1):\tau} \in \mathbb{R}^{I \times \rho}
$$
is concatenated to the spatial encoding before GRU processing [2606.25201]. The paper reports that periodic embeddings generally help, with dataset-dependent sensitivity to periodic resolution (M) and embedding dimension (\rho) [2606.25201].

4. Interpretability and pattern semantics

FDN’s interpretability is intrinsic rather than post hoc because its prediction is explicitly the combination of a pattern library and pattern probabilities [2606.25201]. For every forecast, the model exposes three interpretable objects: the learned pattern matrix (\hat{\mathbb{F}}), the classifier output (\mathbf{\mathcal{C}}), and the final interpolated forecast (\mathbf{\mathcal{C}}\hat{\mathbb{F}}^{T}) [2606.25201]. This permits pattern-level attribution: one can inspect which future patterns were active, how strongly they contributed, and how their weighted combination produced the forecast.

The learned patterns are directly visualizable future trajectories. The paper reports that these patterns correspond to domain-meaningful activities such as flood dissipation in streamflow, traffic relief after congestion, and sunset-related shutdown in photovoltaic generation [2606.25201]. Real-time case studies display the observed preamble, the model forecast, and the top-ranked patterns, with visual emphasis on higher-confidence patterns [2606.25201]. In streamflow, the model identifies a descent pattern when a node lies on the descending side of a high-flow event; in traffic, it predicts the return of flow near the end of congestion; in solar energy, it predicts power drop associated with sunset around 4:30 pm in late December in Alabama [2606.25201].

This interpretability differs from feature attribution. FDN explains predictions in terms of future-pattern mixtures rather than by assigning causal importance to input features [2606.25201]. A common misconception would be to equate its classification step with a standard discrete class label. In fact, the paper emphasizes soft classification, and the forecast is the interpolation of patterns rather than the selection of a single prototype [2606.25201].

Another qualitative analysis projects learned patterns, predicted futures, and ground-truth futures by PCA and reports that the convex hull of the learned pattern set nearly covers both train and test future samples, particularly on E-PEMS-BAY [2606.25201]. This suggests that the pattern dictionary behaves as a usable future space rather than merely memorizing training windows.

5. Empirical performance and efficiency

FDN is evaluated on three public datasets spanning hydrology, traffic, and energy systems [2606.25201]. Wabash River contains 1,276 subbasins with five features and daily prediction targets of streamflow at horizons of 1, 4, and 7 days [2606.25201]. E-PEMS-BAY contains 325 traffic sensors with five features and horizons of 5, 30, and 60 minutes from 1 hour of observations [2606.25201]. Solar-Energy contains 137 solar plants with univariate power observations and horizons of 10, 60, and 120 minutes from 6 hours of history [2606.25201].

The model is compared against 11 forecasting models; the text specifically identifies AGCRN, StemGNN, and MTGNN as strong comparators on the three datasets [2606.25201]. Evaluation uses MAE, MAPE, and RMSE, with masking to exclude imputed values [2606.25201]. The paper states that FDN matches or exceeds state of the art overall and consistently outperforms the next-best model for longer prediction horizons across all datasets [2606.25201]. Reported representative improvements include a 9.1% reduction in MAPE and a 2.5% reduction in RMSE for the longest horizon on E-PEMS-BAY, a 6.3% reduction in MAPE on the largest Wabash horizon while nearly matching RMSE, and a 1% reduction in MAPE together with a 12% reduction in RMSE on the largest Solar-Energy horizon [2606.25201].

Qualitative comparisons attribute these gains to event sensitivity. FDN captures high streamflow peaks better than AGCRN on Wabash River, overpredicts less than MTGNN during peak afternoon solar output, and better predicts rush-hour stoppage and late-evening recovery than StemGNN on E-PEMS-BAY [2606.25201]. The paper also notes that node-level plots show stronger gains on higher-variance nodes, suggesting improved modeling of large changes or events [2606.25201].

The model is also presented as efficient. Relative to AGCRN on Wabash River, FDN uses from (1/12) to (1/2) as many parameters and shows about a 2.4× runtime speed-up [2606.25201]. Relative to StemGNN on E-PEMS-BAY, it uses about (1/38) as many parameters while nearly matching runtime [2606.25201]. Relative to MTGNN on Solar-Energy, it uses from (1/276) to (1/78) as many parameters and shows about a 10.2× speed-up in single-step forecasting, though it is slower for some multi-step settings [2606.25201]. This efficiency stems primarily from the compact pattern dictionary (\hat{\mathbb{F}}) and the fact that final decoding is a matrix product rather than a heavy autoregressive or attention-based decoder [2606.25201].

6. Training, ablations, and relation to broader decomposition models

FDN is trained with MAE / L1 loss using PyTorch L1Loss, Adam, a learning rate of 0.003, Kaiming normal initialization, an epoch limit of 200, patience of 15, and batch size 64, with results averaged over three random seeds [2606.25201]. The appendix further notes a pattern-regularization term inspired by SVD orthogonality: the learned patterns are constrained to be dissimilar by adding pattern similarity to the forecast loss, although the exact analytic form is not given in the supplied text [2606.25201]. This suggests that FDN does not merely learn any basis, but is encouraged to learn a diverse basis.

Ablation results are central to understanding the architecture. Increasing the number of learned patterns (K) improves performance only up to a point, and error saturates quickly, with useful performance even at about 8 patterns [2606.25201]. This supports the low-rank future hypothesis. LDA significantly improves E-PEMS-BAY, less so Wabash River, and is not relevant for univariate Solar-Energy [2606.25201]. Dependency embedding and graph convolution significantly improve results, node embedding performance saturates around approximately (D=10), pattern regularization helps especially on Wabash River and Solar-Energy, Wabash River benefits from periodic resolution around 3-month granularity, and E-PEMS-BAY and Solar-Energy benefit from high-resolution daily moments [2606.25201].

The term “FDN” requires disambiguation because it has been used for unrelated architectures. In recommender-system multi-task learning, "Feature Decomposition Network" denotes a model for reducing negative transfer via decomposition into task-specific and task-shared features [2302.05031]. In facial expression recognition, FDN denotes the decomposition stage inside “Feature Decomposition and Reconstruction Learning,” producing facial action-aware latent features from a ResNet-18 backbone [2104.05160]. In image denoising, FDN denotes a flow-based invertible denoiser that splits latent variables into clean and noise components [2105.04746]. In robotic wrench forecasting, it denotes "Frequency-aware Decomposition Network," which separates wrench into low-frequency trend and high-frequency residual with asymmetric deterministic and probabilistic heads [2604.12905]. This terminological overlap is important because “Future Decomposition Network” is not a generic historical meaning of FDN; it is the specific name of the spatiotemporal forecasting architecture introduced in 2026 [2606.25201].

FDN also differs from closely related decomposition-based forecasting proposals that decompose the computation rather than the future. "FDNet: Focal Decomposed Network for Efficient, Robust and Practical Time Series Forecasting" decomposes input processing into focal local paths and explicitly notes that it is not literally presented as an FDN nor as a future decomposition method [2306.10703]. A plausible implication is that FDN and FDNet belong to a broader decomposition-oriented design family, but they decompose different objects: FDN decomposes the future space into reusable patterns, whereas FDNet decomposes the input-processing computation graph.

The main limitation identified in FDN is its assumption that futures can be represented well by a relatively small shared pattern dictionary [2606.25201]. The learned patterns are shared across nodes, which improves efficiency and interpretability but may constrain expressivity if subsets of nodes exhibit fundamentally different future shapes [2606.25201]. Its interpretability is pattern-level rather than feature-causal, and the exact regularization formula for pattern dissimilarity is not provided in the supplied text [2606.25201]. Even with these caveats, the model’s central contribution is clear: it turns spatiotemporal forecasting into the problem of recognizing likely future patterns and interpolating among them, making the future operator itself explicit, compact, and interpretable [2606.25201].