Revisiting Sequential Information Bottleneck: New Implementation and Evaluation.

Area of Science:

• Computer Science
• Data Mining
• Machine Learning

Background:

• Clustering algorithms are essential for data analysis.
• Existing methods may lack efficiency or optimal quality.
• Sequential Information Bottleneck (SIB) is a promising clustering approach.

Purpose of the Study:

• Introduce a modern, optimized, and publicly available implementation of the SIB algorithm.
• Enhance computational efficiency, especially for sparse data.
• Facilitate real-world text clustering applications.

Main Methods:

• Developed an optimized implementation of the sequential Information Bottleneck algorithm.
• Incorporated specific optimizations for efficient computation on sparse data.
• Conducted extensive evaluations comparing the algorithm against alternatives.

Main Results:

• The optimized SIB implementation achieves a competitive balance between clustering quality and speed.
• Outperforms considered alternatives in text clustering tasks, even with basic Term-Frequency representation.
• Demonstrates efficiency gains, particularly for sparse data representations.

Conclusions:

• The new SIB implementation offers a highly attractive trade-off between quality and speed.
• It is well-suited for practical text clustering applications.
• The optimized algorithm facilitates broader adoption in real-world scenarios.