Imagine trying to determine the statistical importance of wearing a heavy winter coat. If you analyze a dataset spanning the entire globe over a full year, the overall global correlation between wearing a coat and survival might appear negligible or noisy. In tropical climates, a coat is irrelevant; in arctic climates, it is a matter of life and death.
Traditional feature selection and model interpretation methods suffer from exactly this blind spot: they seek a single, globally optimal set of features for an entire dataset. As Roquero Gimenez et al. observe, this "one-size-fits-all" approach fails to capture complex systems where a variable's relevance is contingent on the state of other variables. When we evaluate features strictly on their average global performance, we inadvertently misidentify highly predictive, locally relevant features as noise (Roquero Gimenez et al.).
Local explanation frameworks resolve this by revealing how feature importance shifts dynamically based on context, allowing us to discover local exceptions and intricate feature interactions that global metrics obscure (Baehrens et al.).
Partitioning the Feature Space
How do we systematically uncover these hidden, localized drivers? The foundational mechanism relies on partitioning the feature space into distinct, interpretable regions where feature behaviors stabilize.
Instead of scoring variables across a whole dataset, Roquero Gimenez et al. calculate feature importance within specific partitions of the feature space, isolating sub-populations where niche variables become highly predictive. Tree-based models are naturally suited for this task. For instance, Sutera et al. utilize random forests to calculate the absolute difference between the information a variable provides globally versus what it provides within a specific subset. This allows the authors to classify variables into two intuitive categories:
- Context-complementary variables, which gain predictive power when a specific context is known.
- Context-redundant variables, which lose their relative importance because the context already captures the necessary information (Sutera et al.).
However, searching for these combinations can be computationally prohibitive. To address this under memory constraints, Sutera et al. developed a sequential subspace method that iteratively maintains and prioritizes previously identified relevant features, ensuring the search algorithm preserves the necessary context to recognize highly interdependent variables. Furthermore, Ithapu et al. demonstrate that feature selection must account for a variable's role within a larger hierarchical cluster rather than treating it in isolation, preventing models from overlooking relative leverage within specific structural groups.
This localized behavior is deeply embedded within neural architectures as well. Gori et al. show that deep ReLU networks function as piecewise linear machines. By grouping data points that trigger identical activation clusters, the authors reveal that a feature's importance can drastically change or even completely reverse direction depending on the specific region of the input space.
Dynamic and Instance-Wise Selection
While partitioning divides data into static regions, a more granular approach evaluates feature relevance dynamically for each individual data point.
To achieve this, Sristi et al. introduce Conditional Stochastic Gates (c-STG), a deep learning architecture that leverages a hypernetwork to dynamically select different features based on situational context variables (such as age or environmental conditions). This allows the model to maintain high predictive accuracy across heterogeneous datasets where the underlying data-generating process shifts (Sristi et al.). Similarly, for tabular data, Liu et al. propose a framework called DIWIFT that utilizes Influence Functions to optimize instance-wise feature selection, capturing fine-grained conditional dependencies that global models miss.
This dynamic perspective fundamentally changes how we approach model robustness. Traditional machine learning techniques often categorize features into a strict binary: universally reliable or universally spurious. However, Ghosal et al. demonstrate that models capable of dynamically re-evaluating feature importance on a case-by-case basis significantly outperform static robustness methods, proving that effective generalization requires adapting to shifting environmental contexts.
This dynamic pairing is also vital in recommendation systems; as Hong et al. point out, the relevance of a specific feature is never static, but varies fluidly based on the exact combination of features it is paired with during a given interaction.
Resolving Sign Reversals and Redundancy
When feature importance becomes context-dependent, standard interpretability tools often break down or yield conflicting results. This occurs because local regions introduce complex phenomena like sign reversals and non-stationary sensitivity.
To resolve these conflicts, Herbinger et al. developed the GRANITE framework, which recursively partitions the feature space into regions where different explanation methods agree. The authors show that feature interactions can undergo complete sign reversals—where the direction of an effect flips entirely depending on the data subset—explaining why global explanation methods frequently mislead users.
Another challenge is feature redundancy: if two features share the same information, standard attribution methods often split the credit, underestimating both. Catav et al. address this by proposing Marginal Contribution Feature Importance (MCI). Instead of averaging contributions, MCI evaluates a feature by its maximum marginal gain across all possible feature subsets, successfully pinpointing the exact context where a variable is most informative.
This non-stationary sensitivity is equally critical when dealing with missing data. In non-linear models, the value of acquiring new data points is highly dependent on the features we have already observed for that specific instance. By fixing these observed features during attribution, Wu et al. introduce observation-dependent Shapley values (OD SHAP) to intelligently prioritize data collection based on the model's localized sensitivity.
Contextual Relevance Across Domains
The necessity of context-aware feature evaluation spans across diverse data modalities, from text and graphs to high-dimensional embeddings:
- Natural Language Processing: Classifiers that rely on global averages fail when a feature's meaning changes entirely based on its local context. Akbik et al. demonstrate this in semantic role labeling, where a specific syntactic path can indicate entirely different semantic roles depending on the predicate it accompanies.
- Computer Vision: In human-object interaction detection, treating all visual cues with equal importance leads to suboptimal performance. Zhong et al. show that the diagnostic value of specific features (such as human pose versus object appearance) shifts dramatically depending on the interaction type.
- Graph Neural Networks: Within graph structures, He et al. demonstrate that an attribute's significance is highly context-dependent, changing based on the topology and characteristics of its local neighborhood.
- Network Analysis: Real-world networks exhibit local homogeneities where a single attribute might drive link formation in one community but be completely irrelevant in another, meaning universal similarity metrics inevitably degrade performance (Xu et al.).
- Embedding Spaces: Even within continuous vector spaces, Zheng et al. show that the significance of metadata attributes in defining clusters varies across different regions of the space, with certain features only emerging as primary differentiators after the data has been hierarchically partitioned.
Key Foundational Work
Classifying the Hungarian Web (2003) — contributed 1 concept elements to this direction.
Feature Selection for Clustering (2000) — contributed 1 concept elements to this direction.
Model-Based Hierarchical Clustering (2000) — contributed 1 concept elements to this direction.