2 Broad Types of Machine Learning: 5 Real-life Problem Types
#beginners
Big Data, Artificial Intelligence (AI), Machine Learning (ML) … are all trendy terms, and we are hearing more of them than ever. Read this article to understand these cool tools and the broad types of Machine Learning and some real-life problem types.
Many people even think all our problems can now be solved by Big Data, Artificial Intelligence (AI), Machine Learning (ML). But really? If so, these tools would seem to be the greatest discoveries in modern times. In fact, they are not brand-new inventions. So, what are they?
- Big Data is referring to huge datasets and big data analytics are trying to extract useful information from these datasets.
- AI can be thought of trying to mimic how the human mind works in solving problems.
- Machine Learning, which is a subset of AI, applies computer algorithms and mathematical optimization methods to build statistical models to help making prediction.
All the above can be traced back many decades ago but due to the advancement in the computing power, they become more well-known and widely used today. However, can they really solve all our problems? Probably not. However, these tools are especially useful to solve a lot of common business problems.
Two Broad Types of Machine Learning
Let’s focus on machine learning here. For simplicity, there are two broad types of Machine Learning commonly used: Supervised Learning and Unsupervised Learning.
Supervised Learning
Most commonly used in Machine Learning. Training data are available, and they are used to train the model and assess the model performance. Another way to say it, backtesting is possible in the process of model building as we know the “answers” (i.e., what we want to predict) from historical data. For example, credit card companies always want to identify all the fraudulent transactions. Before having a model to make such a prediction, it was a very labour-intensive process. To be more efficient in identifying potential frauds, credit card companies use a set of “training data” which contains all the possible predictors (e.g., age, gender, home location, purchase location, previous purchase history, type of goods/service, etc.) and whether each purchase transaction is fraudulent or not. That is, the “answer” (fraudulent or not) is known for this training data set. This training data set can be used to build a prediction model using machine learning and since the “answer” is known, the performance (e.g., accuracy) of each model built can be assessed.Unsupervised Learning
No training data is available (i.e., backtesting is not possible). Unsupervised learning is attempting to learn something from the dataset using Machine Learning methods. For example, identifying some general grouping of the data (or clusters) based on the overall pattern of data distribution.
Five Common Real-life Machine Learning Problem Types
| Potential Applications | |||
| 1 | Classifying Data into Groups | Supervised | With the help of the training dataset, build a model to triage insurance claims into different level of complexity so that claim adjusters at different level of experience can be more appropriately assigned for each claim. Another application is to classify credit card transactions into different likelihood of becoming fraudulent. |
| 2 | Data Clustering | Unsupervised | An application in insurance industry is to group existing policyholders into clusters based on their similarities, then design and promote different insurance/investment products to different clusters. The above application is an example of customer segmentation which is used in other industries for developing more efficient and focused marketing strategy for each segment. |
| 3 | Finding Fraud or Abnormal Data | Supervised | Identify frauds/outliers/abnormal data by looking at the predictors used in the model (e.g. Insurance claim fraud detection: classify insurance claims into “likely fraud” and “not-likely fraud” assuming actual “fraud” data is not too rare in the training data) Some market tools may use unsupervised learning to handle fraudulent/outlier/abnormal identification model. In that case, training data is not available and no back testing can be done. |
| 4 | Finding the Probability of Each Record in a Particular Group | Supervised | Basically, a classification type problem, but the key objective is to assign a probability of each data record indicating how likely that data record would be in a particular group (e.g., for each claim, estimate the probability of that claim having a mental health condition within 3 months after accepting the claim). |
| 5 | Predicting Numerical Target | Supervised | Claim reserve/liability forecast is a classic application where machine learning model can be applied. For example, the objective can be to predict the total insurance claims costs in the first 2 years after accepting the claim. |
The above listed some simple business problem types which can be solved using Machine Learning. In reality, more complex applications such as image/voice recognition, text mining/analytics can also make use of Machine Learning. The main challenge is usually related to data availability and most of the work is normally on the data preparation stage.
Simplified Process of Machine Learning
From the above table, many problems can be benefited from Machine Learning. However, not everyone is data scientist. Most tools out there require certain level of statistical and/or computer programming knowledge to apply machine learning to problem solving. This is discouraging as a lot of people could have applied machine learning to solve their problems but did not know how. In fact, to apply Machine Learning, one can think of the following very high-level simplified process:
Machine Learning Tool for Layman
For the above simplified process, with a bit of guidance by a simple and user-friendly (i.e., no coding) tool such as Aidy (www.aideaye.com), AI and Machine Learning can be applied by a layman user who is not a data scientist. Can a layman use AI or Machine Learning? Absolutely, as long as the right tool is available. Equally important is the cost – both the price tag and the required learning time. With that thought in mind, tool like Aidy which works well with the commonly used application Excel, would be particular useful as it has the vision of machine learning for everyone, not just for data scientists.
