Machine Learning models are trained using data with specific features. The way in which the data is structured helps the models to learn and develop relationship between these features. A well-processed training set is required to build a robust model which in turn generates accurate results. In this article we shall look at some of the ways in which one can build a structured dataset for training.
How To Build The Data
To build a robust model, one has to keep in mind the flow of operations involved in building a quality dataset. The data should be accurate with respect to the problem statement. For example, while trying to determine the height of a person, feature such as age, sex, weight, or the size of the clothes, among others, are to be considered. Here, the person’s clothes will account for his/her height, whereas the colour of the clothes and the material will not add any value in this case. Hence these features have very low weightage for predicting the height of a person. A golden rule of machine learning is: Larger the data better the results.
There are several steps included in this process:
1. Data Selection
In this step, one should be concerned about opting the right number of features for the particular dataset. The data should be consistent and should have least number of missing values. If a feature has more than 25 to 30 percent missing values then it is usually considered not fit to be a part of the training set. But there are instances where the relationship between this feature and the Y feature is high. In that case, one has to impute and handle the missing values for better results.
For example, let us say an institution has borrowed a loan from a bank. A feature containing the GDP value of the particular country is available with 30 percent missing values. If one infers that the particular feature has a very high importance to predict whether the institution is able to repay the loan or not, then this feature has to be considered.
If the feature does not hold high importance for developing the AI model, one should exclude the data. At the end of this particular step, one should have an idea about how to deal with the preprocessing data.
2. Data Preprocessing
Once the right data is selected, preprocessing includes selection of the right data from the complete dataset and building a training set. Here, some of the common steps are:
- Organise And Format: The data might be scattered in different files, for example, classroom datasets of various grades in a school which needs to be clubbed together to form a dataset. One has to find the relation between these datasets and preprocess to form a dataset of required dimensions. Also if the datasets are in different language they have to be transformed into a universal language before proceeding.
- Data Cleaning: This is one of the major steps in data preprocessing. Cleaning refers to mainly dealing with the missing values and removal of unwanted characters from the data. For example, if a feature consists of age of a person, with 4 percent missing values, it can either deleted or replaced. Here, is an in-depth article of how to handle missing in machine learning.
- Feature Extraction: This step involves analysis and optimisation of the number of features. One has to find out which features are important for prediction and select them for faster computations and low memory consumption. For example, while dealing with an image classification problem, images with noise (irrelevant images with respect to the dataset) should be removed.
3. Data Conversion
- Scaling: This is necessary when the dataset is placed. Considering a linear dataset — bank data. If the feature containing the transaction amount is important, then the data has to be scaled in order to build a robust model. By default in correlation matrix, the Pearson method is used to find the relationship. This might lead to a misunderstanding of the data if it is not scaled by a definite value.
- Disintegration And Composition: This step is considered when one needs to split a particular feature to build a better training data for the model to understand. One of the best examples of the data disintegration is splitting up the time-series feature. Where one can extract the days, months, year, hour, minutes, seconds, etc. from a particular sample. And also let us say, the Project ID is IND0002. Here the first three characters refer to the country code and 0002 refer to a categorical value. Separating and processing may result in better accuracy.
- Composition: This process involves combining different features to a single feature for more meaningful data. For example, in the Titanic dataset, the prefix of the passengers with Dr, Mr, Miss. etc can be clubbed into a particular age groups of categorical data which adds more weight in predicting the passengers’ survival.
In this article, one can understand how processed training set helps a machine learning to develop the relationship between the features. This process involves a lot of time, analysis and examination of the data. With a well-structured data, machine learning model can train faster and give robust results.
Try deep learning using MATLAB