Imagine building a machine learning model to detect human stress using wearable sensors. You might collect electrodermal activity (EDA), heart rate, and movement data, extract standard statistical features, and train a classifier. In a controlled laboratory setting, this model might achieve excellent accuracy. But what happens when the user steps outside? A sudden spike in EDA might indicate cognitive stress during a difficult exam, but the exact same physiological spike could simply be the result of walking up a steep hill on a hot day.
Traditional machine learning pipelines treat feature importance as a static property: a feature is either globally useful or it is noise. However, a growing body of research demonstrates that for human-centric sensing—spanning physiology, biomechanics, and behavior—the predictive utility of a feature is fundamentally tied to the context in which it is observed. Context-dependent feature selection is the paradigm of dynamically shifting, weighting, or filtering the input space based on environmental, physical, or cultural states to maintain robust model performance.
The Physical Dimension: Task Dynamics and Sensor Mechanics
Why do fixed feature sets fail when physical conditions change? The answer lies in how physical tasks alter the underlying data-generating distributions.
Consider biomechanical estimation. When predicting human energy expenditure during walking, Slade et al. demonstrated that the relevance of specific biological signals shifts dramatically based on the environment. On flat ground, electromyography (EMG) signals tracking muscle activation provide critical information. However, when a person walks on an incline or carries an external load, ground reaction forces become significantly more predictive. A static model that heavily weights EMG might fail to capture the increased metabolic cost induced by gravity and load.
Similarly, in human-in-the-loop optimization for exoskeletons, Grimmer et al. found that the optimal mathematical representation of an EMG signal depends entirely on task complexity. Simple, predictable effort increases can be accurately captured using straightforward amplitude-based features. But when the task introduces complex, dynamic muscle activation patterns—such as interacting with an active exoskeleton—amplitude features lose their edge, and mean- or variance-based features become necessary to capture the fluctuating signal dynamics.
This variability extends to basic human activity recognition and stress monitoring. As shown by Li et al., the optimal feature set for classifying activities is highly sensitive to sensor placement and the specific characteristics of the data stream, requiring segmentation methods that adapt to variable-duration activity bouts. When detecting mental strain, Zhu et al. observed that fixed feature extraction pipelines degrade when transitioning between different types of stressful activities or environments. A curated set of low-dimensional statistical metrics might suffice in one scenario, while a high-dimensional feature space is required in another to maintain accuracy.
The Socio-Cultural Dimension: Shifting Meanings of Behavioral Signals
If physical context alters feature utility through mechanics and physiology, socio-cultural context alters it through human behavior and societal habits. When models process digital or behavioral "breadcrumbs," the mapping from feature to target state is rarely universal.
For instance, smartphone-based mood inference models often rely on digital signals like app usage or mobility patterns. However, Meegahapola et al. discovered that features highly indicative of a user's mood in one country completely lose their predictive power when applied to users in another. This is not merely a data-shifting issue; it is a fundamental shift in what the feature means. Expanding on this, Assi et al. demonstrated that digital markers like health app usage can accurately predict physical activity in regions with high technology adoption, but become entirely irrelevant in cultural contexts where such apps are rarely integrated into daily routines.
This cultural dependency is even more pronounced in computer vision and social robotics. In multimodal depression prediction, Cameron et al. found that behavioral markers like gaze patterns and facial expressions manifest differently across cultures. While eye gaze might be a primary differentiator for depression in the USA, individuals in China may exhibit depression through distinct facial action units or head orientations.
As argued by Gjaci et al., the relationship between features and outcomes is governed by cultural frameworks; an intelligent system cannot accurately interpret a social cue without explicitly modeling the cultural logic that produced it. Even in automated surveillance, Nikouei et al. noted that behavioral markers effective for video indexing in one social context fail to translate to others, necessitating flexible frameworks that adapt to environmental variability.
Why Context-Dependent Selection Works: Reducing Structural Variance
To understand why context-dependent feature selection works, it helps to view it through the lens of the bias-variance tradeoff.
When we force a single machine learning model to use a fixed set of features across highly diverse contexts, we force it to learn a highly complex, non-linear mapping function $f(X) \to Y$. The model must figure out that feature $X_1$ matters immensely when context $C_1$ is active, but should be ignored when context $C_2$ is active. If the model lacks sufficient capacity or data, it suffers from high bias or learns spurious correlations—such as confusing a hot day (high EDA) with psychological stress.
By explicitly conditioning the feature selection process on context, we effectively partition the problem. Instead of learning one massive, universal function, we select a subset of features $X_{\text{context}}$ that have a stable, reliable relationship with the target $Y$ within that specific domain.
[Raw Sensor Data] ---> [Context Detection]
|
v
[Raw Sensor Data] ---> [Context-Specific Feature Filter] ---> [Target Predictor]
This approach provides three core benefits:
- Dimensionality Reduction: By filtering out features that are irrelevant or noisy in the current context, we improve the signal-to-noise ratio and protect the downstream model from overfitting.
- Sample Efficiency: Models trained on context-specific features do not need to waste parameters learning when to ignore a feature; the pipeline handles it upstream, allowing for lighter, faster models.
- Fairness and Generalization: Accounting for cultural and environmental context prevents models from misinterpreting behavioral differences as pathological or anomalous, leading to more equitable and generalizable systems.
Key Foundational Work
Applying Multivariate Segmentation Methods to Human Activity Recognition From Wearable Sensors’ Data (2018) — contributed 1 concept elements to this direction.
Rapid energy expenditure estimation for ankle assisted and inclined loaded walking (2019) — contributed 1 concept elements to this direction.
Real-Time Index Authentication for Event-Oriented Surveillance Video Query using Blockchain (2018) — contributed 1 concept elements to this direction.