Machine learning
Machine learning is the use of data to fit a model that can make predictions or identify patterns. A developer specifies a task and a way to assess errors, then a training procedure adjusts the model. The useful test is how it performs on cases it did not see during training.
Common approaches
In supervised learning, examples include a target: a message and its category, or a building's characteristics and its measured energy consumption. Classification predicts a category; regression predicts a numerical value. In unsupervised learning, practitioners look for structure without supplying a target for each example, for instance by grouping similar records.[1]
Reinforcement learning concerns an agent that chooses actions and learns from rewards through interaction. It differs from assigning labels to a fixed collection of records. Self-supervised learning, common in language models, derives training targets from the data themselves, such as predicting a hidden or subsequent part of a sequence.
Example: routing support requests
A support team could train a classifier on past requests labelled by destination team. Start with a baseline, such as keyword rules, and compare it with a small statistical model. Reserve recent requests for testing so that the evaluation resembles deployment. Review mistakes by category: sending an urgent security report to the wrong queue matters more than misrouting a routine question.
Keep evaluation data separate from training and tuning. If preprocessing learns from the complete dataset before the split, information can leak into the model and inflate the reported result.[2]
After deployment
Languages, products and working practices change. Monitor errors and retain a route for staff to correct a classification. Retrain when evidence justifies it.
For Sustainable IT, compare solutions at the quality level the service needs. Include training, routine predictions and human correction in the assessment. A complex model that wins a benchmark by a small margin may offer little benefit in the actual workflow.
See also
References
- ↑ scikit-learn, User guide, supervised and unsupervised learning.
- ↑ scikit-learn, Common pitfalls and recommended practices, section on data leakage.
Further reading
- scikit-learn, Model selection and evaluation: choosing metrics and evaluating models.
