Machine learning tasks or machine learning processes are typically classified into three broad categories, depending on the nature of the learning feedback available to a learning system. Supervised learning, unsupervised learning, and reinforcement learning; these three kinds of machine learning tasks are shown in Figure 3, and will be discussed in this section:
Figure 3: Machine learning tasks.
A supervised learning application makes predictions based on a set of examples, and the goal is to learn general rules that map inputs to outputs aligning with the real world. For example, a dataset for spam filtering usually contains spam messages as well as non-spam messages. Therefore, we could know which messages in a training set are spams or non-spams. Nevertheless, we might have the opportunity to use this information to train our model in order to classify new and unseen messages. Figure 4 shows the schematic diagram of the supervised learning.
In other words, the dataset for training the ML model in this case is labeled with the value of interest and a supervised learning algorithm looks for patterns in those value labels. After the algorithm has found the required patterns, those patterns can be used to make predictions for unlabeled test data. This is the most popular and useful type of machine learning tasks, which is not an exception for Spark as well, where most of the algorithms are a supervised learning technique:
Figure 4: Supervised learning in action.
In unsupervised learning, data points have no labels related or in other words, the correct classes of the training dataset in unsupervised learning are unknown, as shown in Figure 5. As a result, classes have to be inferred from the unstructured datasets, which implies that the goal of an unsupervised learning algorithm is to pre-process the data in some structured ways by describing its structure.
To overcome this obstacle in unsupervised learning, clustering techniques are used typically to group the unlabeled samples based on certain similarity measures, mining hidden patterns towards feature learning. More technically, we can write down a generative model, and then tell the data to find parameters that explain the data to us. Now what will happen next if we are not satisfied with the possibility of this elucidation? The answer is that we should tell the data to do it again until we are using some efficient algorithms or techniques.
Now a new question may arise in your mind, why do we have to put labels on the data? Or cannot we just appreciate the data in its current order recognizing that each datum is unique and pre snowflake? In other words, with a little supervision, our data can grow up to be whatever it wants to be! So why should the unlabeled data be taken into consideration too?
Well, there are some deeper issues regarding this. For example, most of the variation in the data comes from phenomena that are irrelevant to our desired labeling scheme. A more realistic example would be how Gmail classifies e-mails as spam and ham using the supervised learning technique, where the data might use its parameters to explain its semantics, when all we care about is its syntactic properties:
Figure 5: Unsupervised learning.
Reinforcement learning is the technique where the model itself learns from a series of actions or behaviors. Complexity of datasets or sample complexity is very important in the reinforcement learning needed for the algorithms to learn a target function successfully. Moreover, in response to each data point for achieving the ultimate goal, maximization of the reward function should be ensured while interacting with an external environment, as demonstrated in Figure 6. To make the maximization easier, the reward function can either be exploited by penalizing the bad actions or rewarding for the good actions.
In order to achieve the highest reward, the algorithm should be modified with a strategy that also allows the machine or software agent to learn its behavior periodically. These behaviors can be learned once and for all, or the machine learning model can keep adapting as times passes:
Figure 6: Reinforcement learning.
For example, reinforcement learning is common in robotics; the algorithm must choose the robot's next action based on a set of sensor readings. It is also a natural fit for Internet of Things (IoT) applications, where a computer program interacts with a dynamic environment in which it must perform a certain goal, without an explicit mentor. Another example is the game Flappy Bird, which has been trained to play itself.
A recommender system is an emerging application, which is a subclass of information filtering system use for making a prediction of the rating or preference from the users that they usually provide to an item. The concept of recommender systems has become very common in recent years and subsequently applied in different applications. The most popular ones are probably products (for examples, movies, music, books, research articles, news, search queries, social tags, and so on). Recommender systems can be typed into four categories typically:
- Collaborative filtering system, where accumulation of a consumer's preferences and recommendations to other users is based on likeness in behavioral patterns.
- Content-based systems, where the supervised machine learning is used to persuade a classifier to distinguish between interesting and uninteresting items for the users.
- Hybrid recommender systems is a recent research and hybrid approach (that is, combining collaborative filtering and content-based filtering). Netflix is a good example of such a recommendation system that uses Restricted Boltzmann Machines (RBM) and a form of the Matrix Factorization algorithm for large movie databases, such as IMDb. This recommendation, which simply recommends movies or dramas or streaming by comparing the watching and searching habits of similar users, is called rating prediction.
- Knowledge-based systems, where knowledge about users and products is used to reason what fulfills the user's requirements, using the perception tree, decision support systems, and case-based reasoning.
Between supervised and unsupervised learning, there is a small place for semi-supervised learning; where the ML model usually receives an incomplete training signal. More statistically, the ML model receives a training set with some of the target outputs missing. The semi-supervised learning is more or less assumption-based and often uses three kinds of assumption algorithms as the learning algorithm for the unlabeled datasets. The following assumptions are used: smoothness, cluster, and manifold assumption.
In other words, semi-supervised learning can furthermore be denoted as a weakly supervised or bootstrapping technique for using the hidden wealth of unlabeled examples to enhance the learning from a small amount of labeled data. Emerging examples include semi-supervised expectation minimization, and concept learning in human cognition and transitive SVMs.