ML approaches are based on a set of statistical and mathematical algorithms in order to carry out tasks such as classification, regression analysis, concept learning, predictive modeling, clustering, and mining of useful patterns. Using ML, we aim to improve the whole learning process automatically such that we may not need complete human interactions, or we can at least reduce the level of such interactions as much as possible.
Overview of ML
Working principles of a learning algorithm
Tom M. Mitchell explained what learning really means from a computer science perspective:
"A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E."
Based on this definition, we can conclude that a computer program or machine can do the following:
- Learn from data and histories
- Improve with experience
- Iteratively enhance a model that can be used to predict outcomes of questions
Since the preceding points are at the core of predictive analytics, almost every ML algorithm we use can be treated as an optimization problem. This is about finding parameters that minimize an objective function, for example, a weighted sum of two terms such as a cost function and regularization. Typically, an objective function has two components:
- A regularizer, which controls the complexity of the model
- The loss, which measures the error of the model on the training data
On the other hand, the regularization parameter defines the trade-off between minimizing the training error and the model's complexity, in an effort to avoid overfitting problems. Now, if both of these components are convex, then their sum is also convex. So, when using an ML algorithm, the goal is to obtain the best hyperparameters of a function that return the minimum error when making predictions. Therefore, by using a convex optimization technique, we can minimize the function until it converges toward the minimum error.
Given that a problem is convex, it is usually easier to analyze the asymptotic behavior of the algorithm, which shows how fast it converges as the model observes more and more training data. The task of ML is to train a model so that it can recognize complex patterns from the given input data and can make decisions in an automated way.
Thus, inferencing is all about testing the model against new (that is, unobserved) data and evaluating the performance of the model itself. However, in the whole process and for making the predictive model a successful one, data acts as the first-class citizen in all ML tasks. In reality, the data that we feed to our machine learning systems must be made up of mathematical objects, such as vectors, so that they can consume such data. For example, in the following diagram, raw images are embedded into numeric values called feature vectors before feeding in to the learning algorithm:
Depending on the available data and feature types, the performance of your predictive model can vacillate dramatically. Therefore, selecting the right features is one of the most important steps before the inferencing takes place. This is called feature engineering, where the domain knowledge about the data is used to create only selective or useful features that help prepare the feature vectors to be used so that a machine learning algorithm works.
For example, comparing hotels is quite difficult unless we already have a personal experience of staying in multiple hotels. However, with the help of an ML model, which is already trained with quality features out of thousands of reviews and features (for example, how many stars does a hotel have, size of the room, location, room service, and so on), it is pretty feasible now. We'll see several examples throughout the chapters. However, before developing such an ML model, knowing some ML concepts is also important.
General machine learning rule of thumb
The general machine learning rule of thumb is that the more data there is, the better the predictive model. However, having more features often creates a mess, to the extent that the performance degrades drastically, especially if the dataset is high-dimensional. The entire learning process requires input datasets that can be split into three types (or are already provided as such):
- A training set is the knowledge base coming from historical or live data that is used to fit the parameters of the ML algorithm. During the training phase, the ML model utilizes the training set to find optimal weights of the network and reach the objective function by minimizing the training error. Here, the back-prop rule or an optimization algorithm is used to train the model, but all the hyperparameters are needed to be set before the learning process starts.
- A validation set is a set of examples used to tune the parameters of an ML model. It ensures that the model is trained well and generalizes toward avoiding overfitting. Some ML practitioners refer to it as a development set or dev set as well.
- A test set is used for evaluating the performance of the trained model on unseen data. This step is also referred to as model inferencing. After assessing the final model on the test set (that is, when we're fully satisfied with the model's performance), we do not have to tune the model any further, but the trained model can be deployed in a production-ready environment.
A common practice is splitting the input data (after necessary pre-processing and feature engineering) into 60% for training, 10% for validation, and 20% for testing, but it really depends on use cases. Sometimes, we also need to perform up-sampling or down-sampling on the data based on the availability and quality of the datasets.
This rule of thumb of learning on different types of training sets can differ across machine learning tasks, as we will cover in the next section. However, before that, let's take a quick look at a few common phenomena in machine learning.
General issues in machine learning models
When we use this input data for the training, validation, and testing, usually the learning algorithms cannot learn 100% accurately, which involves training, validation, and test error (or loss). There are two types of error that one can encounter in a machine learning model:
- Irreducible error
- Reducible error
The irreducible error cannot be reduced even with the most robust and sophisticated model. However, the reducible error, which has two components, called bias and variance, can be reduced. Therefore, to understand the model (that is, prediction errors), we need to focus on bias and variance only:
- Bias means how far the predicted value are from the actual values. Usually, if the average predicted values are very different from the actual values (labels), then the bias is higher.
- An ML model will have a high bias because it can't model the relationship between input and output variables (can't capture the complexity of data well) and becomes very simple. Thus, a too-simple model with high variance causes underfitting of the data.
The following diagram gives some high-level insights and also shows what a just-right fit model should look like:
Variance signifies the variability between the predicted values and the actual values (how scattered they are).
Identifying high bias and high variance: If the model has a high training error as well as the validation error or test error is the same as the training error, the model has high bias. On the other hand, if the model has low training error but has high validation or high test error, the model has a high variance.
An ML model usually performs very well on the training set but doesn't work well on the test set (because of high error rates). Ultimately, it results in an underfit model. We can recap the overfitting and underfitting once more:
- Underfitting: If your training and validation error are both relatively equal and very high, then your model is most likely underfitting your training data.
- Overfitting: If your training error is low and your validation error is high, then your model is most likely overfitting your training data. The just-rightfit model learns very well and performs better on unseen data too.
Bias-variance trade-off: The high bias and high variance issue is often called bias-variance trade-off, because a model cannot be too complex or too simple at the same time. Ideally, we would strive for the best model that has both low bias and low variance.
Now we know the basic working principle of an ML algorithm. However, based on problem type and the method used to solve a problem, ML tasks can be different, for example, supervised learning, unsupervised learning, and reinforcement learning. We'll discuss these learning tasks in more detail in the next section.
