A General Framework for Association Analysis of Heterogeneous Data

Gen Li, Irina Gaynanova

Abstract

Multivariate association analysis is of primary interest in many applications. Despite the prevalence of high-dimensional and non-Gaussian data (such as count-valued or binary), most existing methods only apply to low-dimensional data with continuous measurements. Motivated by the Computer Audition Lab 500-song (CAL500) music annotation study, we develop a new framework for the association analysis of two sets of high-dimensional and heterogeneous (continuous/binary/count) data. We model heterogeneous random variables using exponential family distributions, and exploit a structured decomposition of the underlying natural parameter matrices to identify shared and individual patterns for two data sets. We also introduce a new measure of the strength of association, and a permutation-based procedure to test its significance. An alternating iteratively reweighted least squares algorithm is devised for model fitting, and several variants are developed to expedite computation and achieve variable selection. The application to the CAL500 data sheds light on the relationship between acoustic features and semantic annotations, and provides effective means for automatic music annotation and retrieval.

Introduction

In modern data science, researchers frequently encounter "multi-view" data, where multiple sets of measurements are taken from the same subjects. A prominent example is music information retrieval, where a single song is characterized by both raw acoustic features (continuous signals) and semantic tags like "mellow" or "jazz" (binary labels). Understanding the relationship between these disparate data types is crucial for tasks like automatic annotation and recommendation.

The standard statistical tool for linking two data views is Canonical Correlation Analysis (CCA). However, CCA faces two significant challenges in contemporary applications. First, it assumes data is continuous and normally distributed, making it ill-suited for binary or count data. Second, in high-dimensional settings where the number of variables ($p$) exceeds the number of samples ($n$), CCA is prone to overfitting, often identifying spurious correlations that do not generalize.

The Generalized Association Study (GAS) Framework

To address these limitations, we introduce the Generalized Association Study (GAS) framework. GAS is designed to handle heterogeneous data types in high-dimensional settings by modeling the underlying parameters that generate the data rather than the raw observations themselves.

Exponential Family Modeling

The core innovation of GAS is the use of exponential family distributions to unify different data types. By mapping continuous, binary, and count data to their respective "natural parameters," the framework can analyze them within a consistent mathematical structure.

Table 1: Summary of exponential family distributions used in the GAS framework, including Gaussian, Poisson, and Bernoulli models.

As shown in the table above, Gaussian, Poisson, and Bernoulli distributions can all be expressed through a natural parameter $\theta$ and a link function $g(\mu)$. This allows GAS to relate a "1" in a binary set to a continuous value in another set in a principled manner.

Structured Decomposition

GAS decomposes the natural parameter matrix of each dataset into three distinct components:

  1. Intercept: Captures the baseline mean of the variables.
  2. Joint Structure: Represents the shared patterns or "bridge" between the two datasets.
  3. Individual Structure: Accounts for variation unique to a single dataset.

By explicitly modeling individual variation, GAS avoids the common pitfall of mistaking dataset-specific noise for a joint signal, a frequent issue in traditional CCA.

Optimization and Implementation

Model fitting is achieved through an Alternating Iteratively Reweighted Least Squares (IRLS) algorithm. To ensure scalability for large-scale datasets, we developed a "one-step approximation" variant. This approach significantly accelerates computation—reducing processing time to under three minutes for standard benchmark datasets—without compromising the accuracy of the association analysis.

Experimental Results

We evaluated the GAS framework using both synthetic simulations and the CAL500 (Computer Audition Lab 500-song) dataset. The CAL500 dataset consists of 502 popular songs, each described by 68 continuous acoustic features and 174 binary semantic annotations.

Performance in Music Annotation

On the CAL500 dataset, the GAS framework achieved an auto-tagging precision of 0.438, significantly outperforming established methods such as MixHier (0.265) and Autotagger (0.312). The model prioritizes precision, ensuring that suggested tags are highly accurate, which is particularly valuable for recommendation systems.

Quantifying Association

We introduced a new association coefficient ($\rho$), bounded between 0 and 1, to measure the strength of the link between heterogeneous sets. For the CAL500 data, we found a statistically significant but moderate association ($\rho = 0.265$). This result suggests that while acoustic features are informative, they only capture a portion of the semantic meaning attributed to music by listeners.

Conclusion and Future Work

The GAS framework provides a robust toolkit for high-dimensional association analysis across diverse data types. While highly effective, future extensions could address "over-dispersion" in real-world data where variance exceeds model expectations. Additionally, extending the framework from pairs of datasets to "multi-block" data (three or more views) remains a promising direction for future research. Finally, the moderate association found in music suggests that developing more sophisticated acoustic features may be necessary to fully bridge the gap between signal and semantics.

Summary updated

Abstract

Multivariate association analysis is of primary interest in many applications. Despite the prevalence of high-dimensional and non-Gaussian data (such as count-valued or binary), most existing methods only apply to low-dimensional data with continuous measurements. Motivated by the Computer Audition Lab 500-song (CAL500) music annotation study, we develop a new framework for the association analysis of two sets of high-dimensional and heterogeneous (continuous/binary/count) data. We model heterogeneous random variables using exponential family distributions, and exploit a structured decomposition of the underlying natural parameter matrices to identify shared and individual patterns for two data sets. We also introduce a new measure of the strength of association, and a permutation-based procedure to test its significance. An alternating iteratively reweighted least squares algorithm is devised for model fitting, and several variants are developed to expedite computation and achieve variable selection. The application to the CAL500 data sheds light on the relationship between acoustic features and semantic annotations, and provides effective means for automatic music annotation and retrieval.
BibTeX
@article{li2018general,
  author = {Gen Li and Irina Gaynanova},
  title = {{A General Framework for Association Analysis of Heterogeneous Data}},
  journal = {The Annals of Applied Statistics},
  year = {2018},
  doi = {10.1214/17-aoas1127},
}
Figure