In many real-world applications, from tracking professional athletes to monitoring medical patients, data is rarely a continuous stream. Sensors fail, players move off-camera, and clinical tests are performed at irregular intervals. Traditionally, researchers treated these gaps as a forecasting problem: given the past, predict the missing values.
However, a growing body of research suggests that forecasting is the wrong tool for filling gaps. Instead, the field has shifted toward bidirectional temporal modeling, a paradigm that uses both the past (preceding observations) and the future (succeeding observations) to "anchor" the missing data. By treating the gap as a bridge to be built from both sides, these models achieve a level of physical and statistical plausibility that unidirectional models cannot match.
The Power of the Future Anchor
The fundamental intuition behind bidirectional modeling is the reduction of the search space. In a unidirectional model, errors compound over time; a small miscalculation in velocity at the start of a ten-second gap leads to a massive positional drift by the end.
As Daniel Hennes and colleagues demonstrate in their work on the Graph Imputer, knowing the "reappearance point"—where an agent is first seen after a gap—provides a critical constraint. By performing "retrodiction" (predicting the past from the future) alongside forward prediction, the model anchors the estimated path at both ends. This bidirectional approach effectively "sandwiches" the uncertainty, preventing the drift that plagues standard autoregressive models.
This concept is particularly vital in multi-agent environments like football. In Multiagent off-screen behavior prediction in football, Shayegan Omidshafiei and the DeepMind team show that bidirectional temporal modeling allows for the reconstruction of full-pitch dynamics from limited broadcast feeds, enabling advanced analytics like "pitch control" even when players are off-screen.
Beyond Time: Multi-directional Correlations
While time is the primary axis for imputation, real-world data often contains hidden correlations across different variables. Jinsung Yoon and Mihaela van der Schaar introduced the Multi-directional Recurrent Neural Network (M-RNN) to exploit this.
The M-RNN doesn't just look forward and backward in time; it simultaneously interpolates within a single stream and imputes across multiple streams. For example, in a medical setting, a missing blood pressure reading can be inferred not just from previous blood pressure readings, but also from current heart rate or oxygen levels. By capturing these synchronous features alongside temporal trends, the model prevents the loss of complex patterns that occurs when variables are imputed in isolation.
Enforcing Physical and Structural Laws
A "statistically likely" path might still be physically impossible. To solve this, researchers are integrating domain-specific constraints into the bidirectional framework.
In the DBHP framework, Hanjun Choi and Changjo Kim use a "self-ensemble" that enforces derivative-based consistency. By ensuring that the imputed positions respect the laws of velocity and acceleration, the model produces trajectories that look like human movement rather than jittery noise.
Similarly, Harry Bendekgey and Erik B. Sudderth argue for Structured Variational Autoencoders. By embedding probabilistic graphical structures directly into deep learning models, they allow the system to represent multiple plausible outcomes for a missing segment while remaining bound by global temporal constraints.
The Evolution of Architecture: From RNNs to Mamba
The field is rapidly moving away from traditional Recurrent Neural Networks (RNNs), which struggle with long-range dependencies and slow training. New architectures like the State Space Model (SSM), specifically the "Mamba" architecture, are being adapted for bidirectional tasks.
Zhong Wang and Zengyu Wan demonstrated this with MambaPupil, which uses a bidirectional selective recurrent model for eye tracking. During an eye blink (a total signal dropout), the model uses the "future" context of where the eye reappears to interpolate the movement during the blink.
Furthermore, Yubo Yang and Bo Jiang introduced AdaSTI, which utilizes bidirectional SSMs within a diffusion framework. This approach allows the model to capture long-range dependencies from both directions simultaneously, resolving local ambiguities in sequence completion that would baffle a causal, forward-only model.
Conclusion
The shift from "predicting the next step" to "filling the gap" represents a fundamental change in how we handle incomplete data. By leveraging the future as an anchor, bidirectional models transform the impossible task of long-term forecasting into a constrained, solvable problem of interpolation. Whether through physical laws, cross-variable correlations, or efficient state-space architectures, the goal remains the same: using the full context of a sequence to reconstruct the parts we cannot see.