A Memory-Efficient Encoding Method for Processing Mixed-Type Data on Machine Learning.

Area of Science:

• Machine Learning
• Data Science
• Information Theory

Background:

• Machine learning often struggles with datasets containing both numerical and categorical data.
• Existing encoding methods like one-hot and feature hashing increase dataset dimensionality, leading to challenges with excessive variables and noisy data.

Purpose of the Study:

• To propose a novel encoding approach for mixed-type data that addresses the limitations of current methods.
• To map mixed-type data into an information space using Shannon's Theory to model information content.

Main Methods:

• Developed a new encoding technique based on Shannon's Theory to represent information in mixed-type data.
• Evaluated the proposed method on ten UCI repository datasets and two real-world datasets.
• Applied the encoding for classification, regression, and clustering tasks.

Main Results:

• The novel encoding approach demonstrated promising results across various datasets and tasks.
• Achieved superior memory efficiency compared to one-hot and feature-hashing encoding.
• Successfully preserved the information content of the original mixed-type data.

Conclusions:

• The proposed encoding method offers a significant improvement over traditional techniques for handling mixed-type data.
• This approach enhances memory efficiency and information preservation in machine learning preprocessing.
• The method is effective for classification, regression, and clustering tasks.