EHHR: an efficient evolutionary hyper-heuristic based recommender framework for short-text classifier selection.

Area of Science:

• Machine Learning
• Natural Language Processing
• Data Science

Background:

• Classifying short-text data from social media is complex due to heuristic limitations.
• Existing classifier recommendation algorithms are not designed for the nuances of short-text data.
• The No Free Lunch theorem highlights the need for problem-specific heuristic selection.

Purpose of the Study:

• To propose an efficient classifier recommender framework for short-text data.
• To address the limitations of existing algorithms in handling diverse short-text classification tasks.
• To develop a system that leverages past performance data to predict optimal heuristics for new problems.

Main Methods:

• Introduction of the Efficient Evolutionary Hyper-heuristic based Recommender Framework (EHHR).
• Utilization of a Hybrid Adaptive Genetic Algorithm (HAGA) for dataset-level feature optimization and performance prediction.
• Reusing previous solutions to predict heuristic performance on unseen short-text classification problems.

Main Results:

• EHHR identifies key features for heuristic recommendation: average entropy, mean word string length, adjective variation, verb variation II, and average hard examples.
• The Hybrid Adaptive Genetic Algorithm (HAGA) demonstrated 80% higher accuracy compared to the standard Genetic Algorithm (GA).
• EHHR successfully clusters datasets and ranks heuristics cluster-wise, correctly classifying 9 out of 10 problems.

Conclusions:

• EHHR provides an effective solution for recommending appropriate machine learning heuristics for short-text classification.
• The framework's ability to cluster datasets and rank heuristics offers improved performance and efficiency.
• This approach overcomes the limitations of generic classifier recommendation systems for specialized short-text data.