concept ML algorithm in category machine learning

appears as: n ML algorithm, ML algorithms, ML algorithm, ML algorithms, ML algorithm
Real-World Machine Learning

This is an excerpt from Manning's book Real-World Machine Learning.

Figure 1.7. The workflow of real-world machine-learning systems. From historical input data you can build a model using an ML algorithm. You then need to evaluate the performance of the model, and optimize accuracy and scalability to fit your requirements. With the final model, you can make predictions on new data.

In this case, you’d use the Marital status variable as the target, or label, and the remaining variables as features. The job of the ML algorithm will then be to find how the set of input features can successfully predict the target. Then, for people whose marital status is unknown, you can use the model to predict marital status based on the input variables for each individual. Figure 1.9 shows this process on our toy dataset.

Figure 1.9. The machine-learning modeling process

At this point, think of the ML algorithm as a magical box that performs the mapping from input features to output data. To build a useful model, you’d need more than two rows. One of the advantages of machine-learning algorithms, compared with other widely used methods, is the ability to handle many features. Figure 1.9 shows only four features, of which the Person ID and Name probably aren’t useful in predicting marital status. Some algorithms are relatively immune to uninformative features, whereas others may yield higher accuracy if you leave those features out. Chapter 3 presents a closer look at types of algorithms and their performance on various kinds of problems and datasets.

The feature-engineering extension of the workflow allows you to expand on the training data to increase the accuracy of the ML algorithm. To ensure that feature engineering is used properly, you need to run the prediction data through the same feature-engineering pipeline that was applied to the training data. This ensures that predictions are generated by using exactly the same process as applied to the training data.