Introduction to Machine Learning Datasets
Machine learning dataset is defined as the collection of data that is needed to train the model and make predictions. These datasets are classified as structured and unstructured datasets, where the structured datasets are in tabular format in which the row of the dataset corresponds to record and column corresponds to the features, and unstructured datasets corresponds to the images, text, speech, audio etc. which is acquired through Data Acquisition, Data Wrangling and Data Exploration, during the learning process these datasets are divided as training, validation and test sets for the training and measuring the accuracy of the mode.
Following are the three main steps needed in data analysis:
- Data Acquisition
- Data Wrangling or Data Pre-Processing
- Data Exploration
As an output of data analysis, we will be having a relevant dataset that can be used in the training of the model.
Types of Datasets
In Machine Learning while training a model we often encounter the problem of over-fitting and underfitting. In order to overcome the situation, we need to divide our dataset into 3 different parts:
- Training Dataset
- Validation Dataset
- Test Dataset
The division of the dataset into the above three categories is done in the ratio of 60:20:20.
- Training Dataset: This data set is used to train the model i.e. these datasets are used to update the weight of the model.
- Validation Dataset: These types of a dataset are used to reduce overfitting. It is used to verify that the increase in the accuracy of the training dataset is actually increased if we test the model with the data that is not used in the training. If the accuracy over the training dataset increase while the accuracy over the validation dataset decrease, then this results in the case of high variance i.e. overfitting.
- Test Dataset: Most of the time when we try to make changes to the model based upon the output of the validation set then unintentionally we make the model peek into our validation set and as a result, our model might get overfit on the validation set as well. To overcome this issue we have a test dataset that is only used to test the final output of the model in order to confirm the accuracy.
Dataset structure and properties are defined by the various characteristics, like the attributes or features. Dataset is generally created by manual observation or might sometimes be created with the help of the algorithm for some application testing. Data available in the dataset can be numerical, categorical, text or time series. For example, in predicting the car price the values will be numerical. In the dataset, each row corresponds to an observation or a sample.
Types of Data
Let’s understand the type of data available in the datasets from the perspective of machine learning.
1. Numerical Data
Any data points which are numbers are termed as numerical data. Numerical data can be discrete or continuous. Continuous data has any value within a given range while the discrete data is supposed to have a distinct value. For example, the number of doors of cars will be discrete i.e. either two, four, six, etc. and the price of the car will be continuous that is might be 1000$ or 1250.5$. The data type of numerical data is int64 or float64.
2. Categorical Data
Categorical data are used to represent the characteristics. For example car color, date of manufacture, etc. It can also be a numerical value provided the numerical value is indicating a class. For example, 1 can be used to denote a gas car and 0 for a diesel car. We can use categorical data to forms groups but cannot perform any mathematical operations on them. Its data type is an object.
3. Time Series Data
It is the collection of a sequence of numbers collected at a regular interval over a certain period of time. It is very important, like in the field of the stock market where we need the price of a stock after a constant interval of time. The type of data has a temporal field attached to it so that the timestamp of the data can be easily monitored.
4. Text Data
Text data is nothing but literals. The first step of handling test data is to convert them into numbers as or model is mathematical and needs data to inform of numbers. So to do so we might use functions as a bag of word formulation.
Various Sources of Dataset
It is quite often hard to find the dataset for the machine learning application. Following are the few lists of datasets along with their descriptions which can be used for experimentation.
Google Dataset Search Engine
Google has its own search engine for the dataset. Their objective was to unify almost all the available dataset repositories and make it discoverable. One can easily search for the dataset based upon the application of their learning model.
Microsoft has Microsoft Research Open Data. It is a data repository that makes the dataset created by the researchers at Microsoft available to the data scientists. Over here one can get a bunch of curated datasets.
Computer Vision Dataset
This source provides a dataset of images. If you plan to work on image processing, deep learning or computer vision you can use this source. There are great visual datasets that are available to build computer vision models.
It contains numerous amounts of data with different shapes and sizes. Most of the available dataset has kernels associated with them, where many data scientist has provided their notebooks to analyze the dataset.
It contains a dataset from the field of public transport, satellite images, etc. These datasets are available on the Amazon Web Service resource like Amazon S3. It becomes handy if you plan to use AWS for machine learning experimentation and development.
Conclusion – Machine Learning Datasets
In this article, we understood the machine learning database and the importance of data analysis. We have also seen the different types of datasets and data available from the perspective of machine learning. In the end, you have the various source which can be used to avail the dataset for the experimentation and development of machine learning models.
This is a guide to Machine Learning Datasets. Here we discuss different types of datasets and data along with the various source of machine learning datasets. You may also look at the following articles to learn more –