In an ideal world, time-series data arrives on a perfectly predictable, uniform grid. Every sensor ticks at exactly the same millisecond, every satellite passes over a field at a fixed interval, and every patient visits the clinic on the first day of the month.
In reality, temporal data is messy. Satellites encounter cloud cover, as noted by Marc Rußwurm et al. in BREIZHCROPS: A TIME SERIES DATASET FOR CROP TYPE MAPPING. Wearable devices drop packets due to motion artifacts, a problem studied by Davide Morelli et al. in Analysis of the Impact of Interpolation Methods of Missing RR-intervals Caused by Motion Artifacts on HRV Features Estimations. Mobile hardware clocks drift, leading to highly irregular sampling intervals, as shown by Zhongjie Ba et al. in Learning-based Practical Smartphone Eavesdropping with Built-in Accelerometer.
This creates a fundamental mismatch: standard machine learning architectures—such as Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), and Transformers—mathematically expect a clean, uniformly sampled grid. When data is missing or irregularly spaced, these models break down. To fix this, a powerful paradigm has emerged: transforming discrete, fragmented observations into continuous mathematical paths via temporal interpolation.
The Intuition of Linear and Padding Imputation
The simplest way to bridge a gap between two observations is to assume the system changes at a constant rate (linear interpolation) or stays exactly the same until a new measurement occurs (temporal padding or forward-fill).
While these methods might seem overly simplistic, they work remarkably well when the underlying phenomenon exhibits slow variance or high local correlation.
Linear Interpolation: (o)-------------(o)
\ /
\ /
(x)-[Estimated Point]
Forward-Fill (Padding): (o)------------->(x)-[Estimated Point]
For instance, in digital health tracking, physiological states do not fluctuate instantaneously. Shkurta Gashi et al. demonstrate in Handling Missing Data For Sleep Monitoring Systems that carrying the last observed value forward is highly effective for recovering sleep-wake classifier accuracy. Because human sleep states possess strong temporal inertia, forward-filling maintains local continuity without the computational overhead of complex predictive models.
In remote sensing, filling gaps with constant dummy values introduces artificial "shocks" that confuse neural networks. Paolo Bertellini et al. show in Binary Classification of Agricultural Crops Using Sentinel Satellite Data and Machine Learning Techniques that linear interpolation preserves the smooth, continuous trajectory of vegetation indices across cloud-obscured periods. By drawing a straight line through the missing days, the model can still trace the underlying lifecycle of the crop, significantly improving classification recall.
Capturing Curvature and Higher-Order Dynamics
Linear paths fail when a system involves acceleration, momentum, or strict geometric constraints. If you draw a straight line between two distant points on a curve, you cut the corner, losing vital information about the system's dynamics.
Consider heart rate variability (HRV) analysis. If a wearable sensor misses a series of heartbeats due to a user's movement, interpolating the duration between beats linearly distorts the frequency features of the cardiac signal. Instead, Davide Morelli et al. find that applying quadratic interpolation directly to the timestamps of the heartbeats preserves the underlying physical properties of the cardiovascular system. The quadratic curve accounts for the acceleration and deceleration of the heart rate, maintaining the integrity of downstream statistical features.
Linear (Cuts Corners): (o)---------(o)
\ /
\ /
\ /
(x)
Higher-Order (Preserves): (o) . . (o)
. (x) .
When dealing with periodic signal loss—such as a camera losing sight of a user's eyes during a blink—the transition boundaries matter immensely. The authors of Stress Estimation Using Multimodal Biosignal Information from RGB Facial Video utilize Hermitian interpolation to solve this.
Unlike linear methods, Hermitian interpolation matches both the values and the first derivatives (the rates of change) at the boundaries of the missing data block. By ensuring that the slope of the reconstructed wave matches the slope of the signal right before and right after the blink, they prevent artificial discontinuities. This reconstruction allows downstream classifiers to analyze autonomic stress responses without being tripped up by sudden edge spikes.
From Preprocessing to Continuous-Time Models
Traditionally, interpolation is treated as a separate preprocessing step: you fill the gaps first, bake the data into a rigid grid, and then hand it to a model. However, a more elegant approach integrates interpolation directly into the model's architecture, treating the data as a truly continuous-time process.
This is the core mechanism behind Neural Controlled Differential Equations (NCDEs). Jeongwhan Choi leverages this paradigm in Graph Neural Controlled Differential Equations for Traffic Forecasting to handle highly irregular spatio-temporal traffic data.
Instead of forcing missing sensor readings into discrete time steps, the framework uses interpolation to construct a continuous path, $X(t)$, from the sparse observations. The hidden state of the neural network, $h(t)$, is then modeled as a continuous process driven by the derivative of this path:
$$\frac{dh(t)}{dt} = f(h(t)) \frac{dX(t)}{dt}$$
Discrete Input: (o) (o) (o) (o)
| | | |
v v v v
[Broken/Irregular Grid - Standard Models Fail]
Continuous Path: (o)~~~~~~~(o)~~~~~~~~~~~~~(o)~~~~~~~(o)
\_________________ X(t) _________________/
|
v
[Neural CDE Engine dh/dt]
Why does this work so well? Because the neural network evaluates the continuous path $X(t)$ rather than discrete data points, it becomes entirely agnostic to when or how much data was missing. If a sensor drops offline for an hour, the continuous path remains defined over that interval.
Jeongwhan Choi demonstrates that this approach allows traffic forecasting models to remain stable and accurate even when up to 50% of the data points are missing. The model simply samples the underlying continuous representation at whatever timestamp the prediction requires, turning a data-integrity nightmare into a smooth, mathematically tractable curve.
Key Foundational Work
Semantics of Time-Varying Information (1996) — contributed 1 concept elements to this direction.
An Energy-Efficient Framework for Multirate Query in Wireless Sensor Networks (2007) — contributed 1 concept elements to this direction.