A dataset is the organized collection of examples from which a machine-learning model learns and is evaluated. Its quality often matters more than model sophistication: useful data should sample the input–output space densely, minimize labeling errors, address missing values, and expose helpful features. section 5.4.1
Prepare a dataset by:
- Collecting and annotating enough data for the problem. section 6.4
- Inspecting inputs, proofreading labels, cleaning records, handling missing values, and selecting useful features. section 5.4.1
- Preparing separate training, validation, and test datasets, then choosing an evaluation protocol and success metric. section 6.4
- Normalizing features using statistics calculated from training data, not from validation or test data. section 13.2.1
Windowed time-series dataset
For temperature forecasting, normalize each feature independently, then create batches containing five days of hourly history and a temperature target 24 hours ahead. Keras's timeseries_dataset_from_array() generates overlapping windows without explicitly storing every redundant sample. With sequence_length = 3, [1, 2, 3, 4, 5, 6] becomes [1, 2, 3], [2, 3, 4], [3, 4, 5], and [4, 5, 6]. section 13.2.1
