Distance Dependent Infinite Latent Feature Models.

Area of Science:

• Machine Learning
• Statistical Modeling
• Bayesian Nonparametrics

Background:

• Latent feature models decompose data into components.
• Bayesian nonparametric models use the Indian buffet process (IBP) to infer the number of features.
• Existing models often assume data exchangeability, limiting their application.

Purpose of the Study:

• To generalize the Indian buffet process (IBP) for non-exchangeable data.
• To introduce the distance dependent Indian buffet process (dd-IBP).
• To model dependencies based on distances between data points, such as temporal or spatial relationships.

Main Methods:

• Developed the distance dependent Indian buffet process (dd-IBP).
• Theoretically characterized the feature-sharing properties of the dd-IBP.
• Derived a Markov chain Monte Carlo sampler for a linear Gaussian model with a dd-IBP prior.

Main Results:

• The dd-IBP allows data points to share features based on their proximity.
• The original IBP is shown to be a special case of the dd-IBP.
• Demonstrated the performance of the dd-IBP on real-world non-exchangeable datasets.

Conclusions:

• The dd-IBP provides a flexible framework for latent feature modeling of non-exchangeable data.
• This approach can capture complex dependencies like temporal and spatial structures.
• The dd-IBP extends the capabilities of Bayesian nonparametric methods in data analysis.