The big data challenge is still considered an open challenge. Recently, researchers have developed techniques and tools to address the problem; one way to overcome the challenges is to use Machine Learning techniques for big data analytics. In most Machine Learning tasks, it is usual to train models on a dataset and examine their generalization capabilities.
Therefore, this research aims to analyze the impact of datasets’ on the skills and performance of machine learning models, a practical approach is utilized to demonstrate that datasets’ have a huge effect on the skills and performance of machine learning models. Hence, three datasets (with different qualities) are picked for the experiment and promising results are shown. Finally, if an ideal dataset exists, almost always machine learning models perform identical regardless of their complexity.
Foremost, access to datasets and then should preprocess to achieve the perfect result using by cleaning text, tokenization method (which split each unique words of sentences), then stem the datasets after preprocessing the datasets can classify them in Binary classification and train them in the case of Train Test Split and split them in 80% for training and 20% for testing, after training the datasets vectorized them because Machine Learning Model can work properly datasets and can use TF-IDF vectorization.
However, when preprocessing the Persian datasets should be careful because Persian datasets do not work with some libraries of pre-build for preprocessing, one can use another library that was created to preprocess the Persian datasets, finally apply the Machine Learning Models.
When applying Machine Learning Models (Logistic Regression, Random Forest, SVM with SGD, SVM) on English text or clear Dataset the result of F1 Score shows completely similar that the clear dataset is working with these Machine Learning Models properly and the results are similar with a bit of difference. And when applying the same Machine Learning models on Persian text which is fetched from GitHub or noisy dataset the result of the F1 Score is different from English text which shows that the noisy dataset works well the result of each model is different. But if apply the same machine learning models on Persian text or noisy dataset fetched from Facebook Page. The result of the F1 Score is completely different which shows that the noisy dataset cannot work properly with all machine learning models, and the results of each model are not similar they do not predict correctly, and have fewer data.
These experiences proved that the datasets' effects on skills and performance of machine learning models and machine learning models are depended on different datasets, which can change the behavior of a Machine Learning model. When a quality dataset work with Machine Learning models does not matter which model is, but when incompetent dataset work with Machine Learning models it takes a long time to fit and the time complexity will be increased and the performance of models goes down also, their speed becomes slower.
In my experience, the Support Vector Machine (SVM) Machine Learning model works properly with each dataset, but better works with the English dataset.
For more details:
Email: [email protected]
GitHub: https://github.com/aketman
You must be logged in to post a comment.