Thinking-data-science-a-data-science-practitioners-guide

4 1 Data Science Process
operations in real-time data streams. Today, we have machine learning applications
like traffic analysis that work on live data streams. All these new models made
machine learning engineers (in the short term, we call them ML engineers) and data
scientists to develop and learn new methods for processing image data. Finally, our
computers understand only binary data; though image data is binary data, we still
need to transform this data into a machine-understandable format. Note that each
single image comprises several binary data items (pixels) and each pixel in an image
is a RGB representation in an image data file.
The older classical ML technology failed to meet these new requirements, and the
industry started looking at the alternative approaches. The ANNs (artificial neural
networks) technology, which was invented many decades ago, came to the rescue.
The modern computing resources made the use of this technology workable in
developing such models. The neural network training requires several gigabytes of
memory, GPUs, and many hours of training. Tech giants had those kinds of
resources, and they trained networks, which we can reuse and extend their functionalities
for our own purpose. We call this new technology Transfer Learning. I have
given an exhaustive coverage later in the book on this new technology—transfer
learning.
Model Development on Text Datasets
After seeing the success of ANN/DNN (deep neural networks) technology in
building image applications, researchers started exploring its application to text
data. There came the new term and the field of its own—natural language processing
(NLP). Preparing text data for machine learning requires a different kind of approach
as compared to numeric. Numeric data is contained in databases having a few
columns. Each column is a potential candidate for a feature. Thus, in numeric
machine learning models, the number of features is typically very low.
Now, consider the text data, where each word or a sentence is a potential
candidate for a feature. For email-spam applications, you use every word in your
text as a feature, while for a document summarization model, you will use every
sentence as a feature. Considering the vocabulary of words that we have, you can
easily imagine the number of features. The traditional dimensionality reduction
techniques that we have for numeric data would not apply for a text dataset in
reducing its dimensions. You need to cleanse the entire text corpus. For cleansing
text data, you would need several steps, removing punctuations, stop words, removing/converting
numbers, lowercasing, and so on. Further to this data cleaning
operation, to reduce the features count, we need to apply a few more techniques
like building a dictionary of unique words, stemming, lemmatization, and so
on. Finally, you need to understand tokenization so that these words/sentences
would be transformed into machine-understandable formats.
For text data, the context in which a word appears also plays an important role in
its analytics. So, there came a new branch called natural language understanding