An online store has purchase and customer data but no ready-made categories for its customers. It can use clustering to group customers with similar behavior, then examine whether those groups help with market segmentation. The input is an unlabeled dataset; the immediate objective is to find structure in it, not to assign categories chosen in advance. section 12.0 section 12.1.1
Machine learning develops mathematical models that describe a system by learning characteristics from data. The aim is often to describe the available dataset while also generalizing to new inputs. It is a branch of artificial intelligence; deep learning, which uses deep neural networks, is a narrower part of machine learning. section 12.1 section 16.4
The kind of feedback available during training distinguishes three approaches:
- Supervised learning uses entries paired with labels or values. Training adjusts model parameters to improve predictions for new entries, such as detected object classes or estimated prices. section 12.1.1
- Reinforcement learning supplies feedback about the outcome of a task rather than a label for each entry. Learning to play chess or Go and teaching a robotic arm to juggle are examples. section 12.1.1
- Unsupervised learning receives no feedback on the data. Clustering looks for hidden structure by bringing similar points together and separating dissimilar ones. section 12.1.1
Gradient descent can be one component of learning: it helps find a low-cost point for a chosen, differentiable cost function. That does not make it a complete machine-learning method. The model, cost function, and optimization goal still matter; some clustering algorithms also use gradient descent. section 16.4
For an unlabeled segmentation task, algorithms such as k-means, DBSCAN, and OPTICS offer different clustering approaches. Implementation is not always necessary: a trusted library may be practical for a common algorithm such as k-means, while unusual requirements or less readily available algorithms can make understanding the internals more important. No single running-time or memory bound applies to machine learning as a whole; assess those costs for the chosen algorithm and system. section 12.0 section 1.2 section 11.5
