What is machine learning? Understand learning from examples through email classification
Understand training and prediction through email classification, the role of labels and independent tests, and why data shifts and overfitting still matter.
Understand the concept
Machine learning trains models on data to learn relationships useful for prediction or generation. Developers need not write a separate rule for every case, but still define the task, prepare data, and evaluate results. It is part of AI, and deep learning is one family within it. Learned patterns may be useful or biased, so evaluate on data excluded from training.
Training data needs a clearly defined task
To classify incoming emails as refund requests or other messages, supervised learning usually needs correctly labeled examples. The model learns relationships between text and categories for new mail. If annotators disagree about what counts as a refund request, more examples can amplify that inconsistency.
A task can lack predefined categories, such as grouping similar texts to help people discover themes. Unsupervised learning does not remove the need for evaluation; it changes the labels supplied during training. People still decide whether a discovered group is useful.
Using a model on new input is inference
Training adjusts parameters, while inference applies existing parameters to new inputs. They have different resource and evaluation needs. Adding instructions to the current conversation usually changes the input, not the trained model, and does not establish persistence in the next conversation.
A model that succeeds on training examples but fails on new wording may not generalize. Remove near-duplicates from evaluation rather than placing one email under different filenames in both training and test sets, where repeated material can hide real errors.
Recheck results when the data changes
Changes in user language can make old classification patterns less useful. Sample real inputs, record which categories are confused, and decide whether labels or training examples need updating. An API continuing to return results does not establish stable quality.
For narrow tasks with clear rules, compare against a simple rule-based approach. Machine learning requires data and evaluation maintenance; not every automation needs a model. Choose it for reduced important errors and better task outcomes, not a more advanced-sounding label.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.