Abstract
Multi-Dimensional Classification (MDC) involves assigning multiple labels to an object across distinct semantic dimensions. Traditional MDC methods often struggle with complex label dependencies and the use of a uniform feature set for all dimensions. This paper introduces PIST (Pairwise dImension-Specific feaTures), a novel framework that extracts dimension-specific features by modeling pairwise label dependencies. PIST utilizes a Pairwise Dimension Encoding module to learn latent label embeddings, which then guide an Attention Network to filter relevant features for each dimension pair. Extensive evaluations on 11 real-world datasets demonstrate that PIST significantly outperforms state-of-the-art MDC methods, particularly on large-scale data.
Introduction
In Multi-Dimensional Classification (MDC), a single instance is associated with multiple class labels simultaneously, each belonging to a different semantic dimension. For example, a landscape image might be categorized by time of day, weather conditions, and scene type. Effectively solving MDC requires addressing two core challenges: capturing the intricate dependencies between labels across different dimensions and identifying the specific features relevant to each dimension.
Existing approaches often rely on a shared feature space for all dimensions, which is suboptimal as different classification tasks typically depend on different subsets of features. Furthermore, modeling high-order label dependencies can lead to computational intractability and overfitting.
The PIST Framework
The proposed PIST framework addresses these issues through a three-stage process: Pairwise Dimension Encoding, Dimension-Specific Feature Extraction, and Classification Integration.
Pairwise Dimension Encoding
Instead of modeling the full joint distribution of all dimensions, PIST focuses on pairs of dimensions. This "pairwise" approach strikes a balance between capturing essential dependencies and maintaining computational efficiency. The module learns latent label embeddings for every possible label combination within a pair, optimized to minimize intra-class covariance and maximize inter-class separation.
Dimension-Specific Feature Extraction
This is the core innovation of PIST. An Attention Network takes the learned label embeddings as input to generate a feature importance vector. This vector acts as a mask, applied via a Hadamard product to the general feature embedding of the input data. This process effectively filters the features, highlighting only those relevant to the specific semantic task of a given dimension pair.
Classification and Integration
The model predicts joint probabilities for these dimension pairs. To obtain the final label for a single dimension, PIST aggregates the joint probabilities from all pairs containing that dimension, ensuring that final predictions are informed by the learned dependencies.
Experimental Evaluation
PIST was evaluated against 13 state-of-the-art MDC approaches across 11 real-world datasets from diverse domains, including computer vision, bioinformatics, and social data.
Key Results
The experimental results demonstrate the superiority of PIST, which outperformed competing methods in the majority of test cases across three key metrics: Hamming Score, Exact Match, and Sub-Exact Match.
As shown in the figure above, the performance of PIST remains relatively stable across various dimensions of label embeddings ($t$), indicating the robustness of the learned representations. The model's performance is particularly noteworthy on large-scale datasets, where its deep learning architecture can effectively leverage higher data volumes.
Conclusion
PIST introduces a paradigm shift in MDC by using label semantics to guide feature selection. By modeling pairwise dependencies and extracting dimension-specific features, it achieves state-of-the-art performance while remaining scalable. Future research may explore automated dimension grouping strategies to further enhance efficiency in extremely high-dimensional scenarios.
Summary updated