Encoding Dissimilarity Data for Statistical Model Building.

Area of Science:

• Statistics
• Machine Learning
• Computational Biology

Background:

• Statistical model building often faces challenges with discrete, noisy, incomplete, and scattered pairwise dissimilarity data.
• Existing methods may struggle to effectively incorporate such data into complex models, limiting their applicability.

Purpose of the Study:

• To review and comment on three papers addressing the use of challenging dissimilarity data in statistical modeling.
• To present a new algorithm for embedding new objects into a pre-defined Euclidean space derived from dissimilarity information.
• To demonstrate how this embedding facilitates the integration of dissimilarity data into various machine learning models.

Main Methods:

• Utilizing convex cone optimization codes to embed objects into a Euclidean space that respects dissimilarity information while controlling dimensionality.
• Developing a "newbie" algorithm for embedding new objects into this established space.
• Integrating the dissimilarity information into Smoothing Spline ANOVA penalized likelihood models, Support Vector Machines, and other models admitting Reproducing Kernel Hilbert Space components.

Main Results:

• Successfully demonstrated a framework for kernel regularization applicable to problems like protein clustering.
• Showcased the utility of flexible risk models in examining covariate influences, including familial and genetic factors.
• Developed a robust manifold unfolding technique with kernel regularization for complex data structures.

Conclusions:

• The presented methods provide a robust framework for incorporating discrete, noisy, and incomplete dissimilarity data into statistical and machine learning models.
• The "newbie" algorithm offers a practical solution for extending existing embeddings with new data points.
• Future research directions and open questions in this domain are identified, paving the way for further advancements.