Most deep learning architectures for time series—from LSTMs to Transformers—share a hidden assumption: data arrives on a uniform grid. They expect a sequence of observations $x_1, x_2, \dots, x_n$ where the time elapsed between $x_t$ and $x_{t+1}$ is always constant.
In the real world, this assumption is frequently violated. Medical sensors fail, satellite revisits are blocked by clouds, and financial trades occur at millisecond intervals or not at all for hours. When we force this "messy" data into a rigid grid through imputation or resampling, we introduce bias and distort the underlying signal.
A growing body of research is moving away from discrete-time sequences toward models that treat time as a continuous coordinate. By reframing time series as continuous functions, graphs, or sets, these methods can naturally handle missing values and asynchronous sensors without the "error compounding" common in traditional models.
The Problem with the Uniform Grid
Traditional Recurrent Neural Networks (RNNs) update their hidden state $h_t$ based on the previous state $h_{t-1}$ and the current input. As Thomas Demeester notes in System Identification with Time-Aware Neural Sequence Models, this fails to capture the physics of continuous systems because the model has no concept of the "gap" between observations. An observation arriving after one second is treated the same as one arriving after one hour.
Furthermore, multivariate data is often asynchronous. In a hospital, a patient's heart rate might be recorded every minute, while blood glucose is checked every four hours. Standard models struggle to align these different frequencies, often requiring "canonical pre-alignment" which, as Ziyu Zhou argues in Revitalizing Canonical Pre-Alignment for Irregular Multivariate Time Series Forecasting, is historically computationally expensive and prone to data distortion.
From Discrete Steps to Continuous Flows
One powerful solution is to view neural networks through the lens of Differential Equations (DEs). Instead of discrete layers, the model's state evolves according to a continuous-time dynamical system. Hongjue Zhao and colleagues in Understanding the Theoretical Foundations of Deep Neural Networks through Differential Equations explain that this allows models to "query" the state at any arbitrary timestamp, effectively bridging the gaps between irregular observations.
This "physics-inspired" approach extends to generative modeling. Carl Remlinger and Joseph Mikael introduced the Conditional Euler Generator (CEGEN), which uses Stochastic Differential Equations (SDEs) to capture the temporal "rhythm" of data. Similarly, Marin Biloš and Kashif Rasul proposed Process Diffusion, which treats time series as samples from underlying continuous functions (like Gaussian Processes) rather than discrete vectors.
Time Series as Graphs and Sets
If we stop viewing time series as a rigid sequence, we can represent them as more flexible structures.
Graphs: Kiran Madhusudhanan and Randolf Scholz developed GraFITi, which represents irregularly sampled data as a sparse bipartite graph. In this view, forecasting becomes an edge-weight prediction task. This bypasses the need for heavy differential equation solvers while maintaining the ability to handle messy intervals. For multivariate data, HyperIMTS uses hypergraphs to model dependencies between different sensors that are not synchronized in time.
Sets: Another approach is to treat the time series as an unordered set of $(time, value)$ tuples. Siyuan Shan and Junier B. Oliva demonstrate in NRTSI that by using Transformer-based architectures on these sets, models can avoid the "error compounding" that happens when RNNs try to step through long gaps. Vijaya Krishna Yalavarthi further refined this with the Tripletformer, which processes $(time, channel, value)$ triplets to interpolate missing data with high precision.
Learning the Underlying Function
A recent breakthrough involves Implicit Neural Representations (INRs) or Neural Fields. Instead of predicting the next value in a sequence, the model learns a function $f(t) = v$ that maps any time $t$ to a value $v$.
Wencheng Zhang and Long Li proposed DualTimesField, which splits a time series into two neural fields: one for smooth, long-term trends and another for sparse, transient events. This separation allows the model to reconstruct irregularly sampled data with significantly lower error than discrete models. In urban computing, Tong Nie used meta-learned INRs in MetaTSI to "collaboratively" impute missing traffic data across different cities, leveraging shared dynamics even when local data is sparse.
Intuition: The Query-Based View
The shift in this research area can be summarized as moving from "What is the next value?" to "What is the state of the system at time $T$?"
Junghoon Lim encapsulates this in QuITE, a query-based embedding module. Instead of forcing observations into a grid, QuITE uses learnable query tokens to aggregate information from asynchronous events. This allows high-performance "backbone" models (like standard Transformers) to process raw, irregular data without needing to change their internal architecture.
By treating time as a continuous feature rather than a discrete index, these models—whether based on kernels (Benjamin M. Marlin), attention (Loïc Landrieu), or differential equations—transform "missing data" from a nuisance into a natural part of the signal.