Greedy learning of binary latent trees.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Mining

Background:

• Inferring latent structures aids in understanding data generation processes.
• Hierarchical latent class (HLC) models represent tree-structured distributions with latent variables.
• Existing Bayesian network-based methods for HLC model learning can be computationally intensive.

Purpose of the Study:

• To develop computationally efficient algorithms for learning HLC models.
• To investigate greedy bottom-up procedures for HLC model inference.
• To compare the performance and speed of new methods against existing approaches.

Main Methods:

• Introduced BIN-G: a greedy algorithm for joint determination of tree structure and latent variable cardinality.
• Introduced BIN-A: an algorithm that first uses hierarchical clustering for tree structure and then BIN-G for cardinality.
• Focused on binary tree structures for computational efficiency.

Main Results:

• BIN-G and BIN-A achieve HLC model quality comparable to Zhang's method (based on cross-validated log-likelihood).
• The proposed greedy methods are generally faster to compute than existing algorithms.
• Demonstrated interpretable latent structure inference on real-world datasets like 20 newsgroups and PASCAL VOC 2007.

Conclusions:

• Greedy algorithms like BIN-G and BIN-A offer an efficient alternative for learning HLC models.
• These methods provide comparable model quality with significant speed improvements.
• The learned tree-structured models offer insights into topic modeling and object co-occurrence in images.