Feature engineering means selecting, transforming and creating attributes (features) from raw data so that a learning algorithm can detect patterns more easily. Feature extraction derives compact, meaningful values from complex data such as text, images or signals. Example: individual purchase records are turned into features like ‘days since last purchase’ or ‘average order value’, which can noticeably improve a churn prediction.
Feature extraction is the process of deriving meaningful values from raw data so that a model can process them efficiently. Instead of millions of individual pixels, a system might use edges, colours or shapes; texts are turned into word weights or embedding vectors. Good features reduce data volume, lower computing effort and often improve accuracy. Feature extraction is a core part of feature engineering.
Dimensionality reduction lowers the number of features in a data set without losing the essential information. Well-known methods include principal component analysis (PCA) as well as t-SNE and UMAP, which are often used for visualisation. It speeds up computation, reduces the risk of overfitting and makes data easier to grasp – for example when hundreds of customer attributes are mapped onto a two-dimensional chart on which similar customers sit close together.
Data cleansing is the detection and correction of incorrect, duplicate or inconsistent entries in data sets. Typical steps include standardising spellings and date formats, removing superfluous spaces and merging duplicates such as 'Smith Ltd' and 'Smith Limited'. Clean data is a prerequisite for reliable analyses, mailings and AI applications.
Machine learning is a branch of artificial intelligence in which computers derive patterns from data instead of having every rule explicitly programmed. Depending on the task, models learn from labelled examples (supervised), from unlabelled data (unsupervised) or through rewards (reinforcement). Applications range from spam filters and demand forecasts to product recommendations. Quality depends heavily on the quantity and quality of the training data.
