3 Approaches to Machine Learning and How to Choose One
‘Information is not knowledge’, Albert Einstein once said. Information gives us knowledge only when yielded meaningfully.
Earlier, organizations employed people to study available information and organize it to get insightful information. Now, after years of advancement, we have finally come to an age where the machine is more than capable of accessing, analyzing, and finally predicting the future, without any human intervention.
The reason ML is so quick to adapt to all the changes is that it uses logic over memory.
The same-day shipping feature from Amazon is available because of machine learning. In fact, their current ML algorithm has decreased the ‘click-to-ship’ time by 225%.
Machine Learning is a perfect epitome for organizations to learn more from little availability. With pre-data, ML can predict anything in the future and thereby helps an organization gain a better understanding of their survival in the market.
Basically, there are 3 ways an organization can approach machine learning and leverage it. And you just need to understand the concepts of mathematics in order to get a clear view of how these approaches work and how you can implement them in your organization.
3 Approaches to Machine Learning
Machine learning coupled with data science aids in the math in AI. ML approaches are a vast set of conceptual algorithms and are fully based on mathematical concepts.
1. Logical Expression
As the name suggests, this approach uses logic to give or predict an appropriate answer.
An algorithm is built using a set of data and the machine studies the given data and then uses logic to make a decision. There are two types of logical expressions.
A tree model is simple. A set of data follows the boolean expression to predict a decision i.e., TRUE or FALSE.
The concept is relatively easy to understand and produces accurate decisions. A tree model is a simple representation of a flowchart in the form of a tree with each branch (node) representing an attribute or a feature to predict a possible outcome.
For understanding the tree model better, consider a simple example. A person named ‘Mr. J’ wants to go out for the weekend. Going out mainly depends on the weather and depending on it are the activities he’ll do. Here, predicting an activity depends on the weather.
So, if the weather is sunny and bright, Mr. J will go out to play golf and if the weather is rainy and damp, then he’d play video games indoors. There is also another possibility where he simply chooses not to go out, even if the weather is sunny.
Here, ‘going out for the weekend’ is called the ‘root node’, the yes or no is called the ‘branches’ and the weather (sunny or rainy) is called the ‘internal nodes’. Depending on the branch nodes, a prediction can be made.
Because of the ease with which tree models can be used, various industries like finance, healthcare, energy, pharmaceutical, education, construction, etc are using it to predict the possibilities.
Rule model, simply put, uses an instruction, ‘If Else’ to make a decision. The machine using a stored set of rules identifies and utilizes a set of relational rules that collectively represent the knowledge captured by a system. The instructions or the rules in any given criteria are called the attributes and these play an important role in prediction.
Let’s assume a person ‘Mr. M’ is late to his office and he needs a vehicle that’ll get him faster to his workplace. So, he uses the maps to know if the roads are traffic-free and congested. ‘If’ the roads are traffic-free, ‘then’ he would go in a car or ‘else’ he’d choose a motorbike.
Here, choosing the vehicle is the problem, and depending on the conditions i.e., ‘traffic or congestion’, the person will choose the possible outcome (choosing a motorbike or a car).
2. The Geometry of the Instance Space:
For a given problem, the collection of all possible outcomes represents the instance space. The objects can have any features that are constructed by a set of elements like lines, curves, points, etc. In the geometric models, the features can be set in two or three-dimensional spaces.
There are two types of models in the geometric models.
A linear model uses planes or spaces to classify an instance space and uses a quadratic equation, y=mx + c or y=ax+ b, to predict an outcome. Here ‘y’ will be depending on the value of ‘x’.
Based on the given data, the learning process computes each feature to form a model that can predict the target value.
Complex behavior can be analyzed and predicted using a linear model and systems can analyze financial and biological data.
For example, let us consider that you want to predict the future value (high or low) of something. For solving the problem, we use the above-given equation, y=mx + c.
Here, ‘y’ is the output
‘m’ Slope or gradient descent
C is y-intercept
Linear models are parametric i.e., they have a ﬁxed small number of numeric parameters that need to be learned from data.
Based on this, we take the values of x and m from the graph and make the predictions of the future value(y). This model is quite stable and easy to use when compared to the logic expressions.
The distance-based model simply predicts the value of an object based on its nearness that is the distance between two objects (same or otherwise).
Here, the concept of distance is not just based on the physical distance between two points. We consider the mode of transport to know the distance between two points.
Distance-based models calculate the physical distance between features(x) and labels(y). These distances are usually calculated using the Euclidean, Minkowski, Manhattan, and Mahalanobis metrics.
For instance, you want to find a particular fruit (say apples ‘y’) that has been mixed up with another fruit (say oranges ‘x’). KNN classifiers can be used to classify and forecast based on Euclidean or Minkowsky and other distance measures.
The attributes of the fruit (x), such as color, shape, etc., are taken from the data and the fruits are thus separated by comparing the nearest neighbor of the fruit (y).
3. The Probability to Classify the Instance Space:
This is the third classification of machine learning algorithms and is extensively used in data science. This model uses probability to classify the new entities. For this, Bayes theorem by Naive Bayes is used.
P(y/x) = P(x/y) * P(y) / P(x)
The algorithm is based on the idea of conditional probability, which is finding the probability that something will happen in the future on a condition that something has already happened.
The best example to understand the concept of probability classification is flipping a coin and predicting the outcome (whether it is heads or tails). Though this classification is hard to grasp, it is quite intuitive and works well in a multiclass prediction.
Guide to Choosing the Right Approach to Machine learning:
The first thing to understand when choosing the right approach to machine learning is knowing that every approach is different and depends on the data available. You need to understand, know, clean, and augment your data for better outcomes.
Apart from this, the approaches depend also on the accuracy, the available computational time, and finally the features of the approach.
Machine learning can be used to predict the future of an organization and can be used to enhance various fields in the organization like sales & marketing, product quality, manufacturing process, and also improves the overall workflow.
Effective use of ML will help organizations gain a substantial ROI and put them on another realm in terms of competitive advantage.
According to StatWolf, Netflix saved $1 billion in the year 2017 as it used the machine learning algorithm which recommends personalized TV shows and movies to subscribers.
Narwal Inc. is a technology company that deals with various aspects like data migration, provides end-to-end automation solutions for better quality and workflows, and also helps other organizations solve complex problems by analyzing and predicting data-driven solutions using data and analytics.