What Are The Practical Benefits of Machine Learning

 

Learn about the practical use-cases of, and the impact of machine learning and data science in our daily lives from a non-technical perspective.

Some of my previous introductory posts to machine learning and data science were a bit technical. However, my purpose of this post is to explain some of the practical use-cases of ML solely from a non-technical savvy layman’s perspective who has had nil exposure to it previously. To satisfy your curiosity, I will also mention the specific ML algorithms that are generally applicable to each use-case if you want to learn more about them.

What type of problems does ML help us with? Irrespective of the specific domain, what answers or actionable insights it offers? Instead of the ‘how’, our focus here will be more on ‘what’ and ‘why’.

What is This? A or B?

This family of ML algorithms predicts in which one of the only two possible categories an observation belongs to. There is no other third potential option. Consider that the management wants to predict which of your existing customers will churn. The answer can only be whether a specific customer will churn or not. Other practical examples include:

• Is this email spam or not?
• Will this customer default or not?
• Are these symptoms symptomatic of a specific disease or not?
• Will this customer continue with a purchase or not?
• Is this an image of a boy or a girl?

Formally known as Binary Classification, the relevant algorithms include:

• Logistic Regression
• Support Vector Machine
• k-Nearest Neighbor
• Classification Decision Tree

What is This? A or B or C or D (Or Something Else)?

An extension of binary classification, here, the number of potential categories can be more than two. Consider that you are working on a face recognition model; the person in a specific picture can be any of the individuals in your database. The number of possible correct answers is only limited to the amount of data used during model development. Other practical examples include:

• Optical Character Recognition: which character is this?
• Which animal is in this image?
• Which genre does this movie belong to?
• Sentiment Analysis: what is the feeling associated with this tweet?
• Whose voice is it in this audio recording?

Formally known as Multi-Class Classification, the relevant algorithms include:

• Random Forests
• Classification Decision Tree
• XGBoost
• k-Nearest Neighbor
• Artificial Neural Networks

How Much or How Many of Something To Expect?

This family of ML algorithms predicts quantities of something as a continuous output or number (i.e., the prediction can be any of the unlimited numbers of possible outcomes). There are no fixed possible categories that can be predicted — for example, predicting sales volume for the next quarter. That sales prediction can be 1,000 units, 10,000 units, 1,200 units, or any other positive real number.

The output of these algorithms can be any real number (positive, negative, zero, fractions); however, your specific use-case will determine whether negatives or fractions can be expected and accepted. For example, a sales forecast cannot be negative.

Other practical use-cases of this class of algorithms include:

• What will be tomorrow’s temperature?
• How many prospects can we sign up as customers in the next quarter?
• What will be our energy consumption next month?
• How long will it take for an event to occur?

Formally known as Regression, the relevant algorithms include:

• Linear Regression
• Regression Decision Tree
• XGBoost
• Artificial Neural Networks

Is This Data Normal or Abnormal?

Oftentimes, we are more interested in whether a specific observation is atypical, abnormal, or anomaly. Or is it merely a normal and usual observation. We can have historical observations classified as abnormal or not. Or it could be the case that such historical classification does not exist, and an ML algorithm will be used to detect any outliers.

Typical use-cases include:

• Is this purchase materially different from the customer’s past purchases?
• Is this traffic pattern from a computer network typical?
• Are these outputs from a piece of industrial equipment atypical?

Formally known as Outlier or Anomaly Detection, the relevant algorithms include:

• Isolation Forest
• Density-Based Spatial Clustering of Applications with Noise (DBSCAN)
• Z-Scores (not technically an ML algorithm, instead a statistical test to identify outliers)
• One-Class Support Vector Machine

How Can We Organize this Data?

Are there any underlying identifiable characteristics that can be used to categorize and organize data into specific groups (also known as clusters or segments)? These unique characteristics are not known to us, and often, even the number of potential clusters is unknown. Clustering or organizing your data may assist you with further analysis or developing cluster-specific strategies.

For example, we may segment our customers into distinct groups based on their age, gender, purchase history, etc. to devise segment-specific sales, marketing, or promotion strategies.

Other practical use-cases of this class of algorithms include:

• Which of our subscribers like similar movies or songs?
• How can we categorize several text documents or audio recordings?
• How can we better segment our products or services?
• Which model of a specific machine is more prone to breakdowns?

Formally known as Clustering, the relevant algorithms include:

• k-Means Clustering
• Mean-Shift Clustering
• Density-Based Spatial Clustering of Applications with Noise (DBSCAN)
• Agglomerative Hierarchical Clustering
• Balanced Iterative Reducing and Clustering using Hierarchies (BIRCH)

What To Do Next?

This is where ML gets really interesting, whereby the ML algorithm not only predicts but also tells us what to do given its prediction. This family of ML algorithms might not be mature enough yet for all use-cases; however, substantial progress has been made recently in the light of advanced deep learning algorithms and the greater processing power available to us.

These algorithms rely on trial and error and multiple feedback loops while not being as heavily dependant upon data as other algorithms. Mostly applicable in automated systems, the recommended action is usually taken by the machine.

Formally known as Reinforcement Learning, it is usually implemented through deep neural networks.

Some practical applications of reinforcement learning include:

• What should the robot do next in its situation in an industrial concern?
• Should we adjust the temperature or leave it untouched?
• How should a self-driving car react (accelerate, decelerate, apply brakes, etc.) given the hazard ahead?

Conclusion

There you have it: a practical, no-nonsense introduction to functional scenarios where ML assists us in plain language.

Free free to comment or reach out to me if you would like to discuss anything further related to machine learning, data analytics, risk scoring, and financial analysis.

Till next time, code on!