Experimental results dataset for the master's dissertation: "A quantitative and comparative analysis of single-pass stream-based active learning query algorithms"

Use email button above to contact.

1. Overview

This dataset contains the aggregated and structured results of a large-scale benchmark evaluating twelve single-pass stream-based active learning query strategies. This is the experimental results dataset for the master's dissertation: "A Quantitative and Comparative Analysis of Single-Pass Stream-Based Active Learning Query Algorithms".

The experiments span:

• 82 datasets
• 5 machine learning models
• 12 stream-based query strategies
• 5 labeling budgets: 5%, 10%, 20%, 50%, and 100%
• 20,000+ experimental runs

Each row represents a single experimental configuration, defined by:

(dataset, model, hyperparameters, query strategy, labeling budget)

This file is designed for statistical analysis, ranking, and comparative evaluation of strategies under constrained labeling scenarios.

2. File Structure

• Granularity: One row per experimental run
• Primary metric: Final model accuracy
• Evaluation setting: Single-pass stream-based active learning

3. Column Dictionary

Below is the semantic definition of each column in the dataset.

dataset

• Type: String
• Description: Dataset used in the experiment.
• Scope: 82 unique datasets.
• Purpose: Enables cross-dataset robustness analysis.

model_name

• Type: String
• Description: Machine learning algorithm used.
• Scope: 5 model families.
• Purpose: Allows studying model–strategy interaction.

model_params

• Type: String (serialized dictionary)
• Description: Hyperparameters used for the model.
• Example:

{'C': 0.01}

• Recommendation: Parse into dictionary for reproducibility or hyperparameter grouping.

query_strategy

• Type: String
• Description: Active learning strategy used in the stream.
• Scope: 12 strategies.
• Purpose: Main variable of interest for comparative evaluation.

budget

• Type: Float
• Values:
• 0.05
• 0.10
• 0.20
• 0.50
• 1.00
• Description: Fraction of instances allowed to be labeled.
• Interpretation: Controls labeling cost.

initial_score

• Type: Float
• Description: Baseline performance before applying active learning.
• Purpose: Reference point for measuring improvement.

percentage_queried

• Type: Float
• Description: Actual fraction of instances labeled.
• Note:
• May slightly differ from the defined budget due to stream dynamics.
• Reflects real labeling consumption.

final_accuracy

• Type: Float
• Description: Final model performance after active learning.
• Metric: Classification accuracy.
• Primary evaluation metric.

4. Summary

experiment_results.csv is a large-scale benchmark dataset for evaluating stream-based active learning strategies under varying labeling budgets.

It supports:

• Cross-dataset comparisons
• Strategy ranking
• Budget sensitivity analysis
• Model–strategy interaction studies
• Efficiency and robustness evaluation

The structure is analysis-ready and designed for statistical benchmarking and research publication purposes.