04. Algorithms Overview
Each splitter returns two sets of indices (train, test), partitioning samples according to the chosen strategy. Choose based on dataset size, desired diversity, or grouping needs.
Kennard–Stone Split
Description: Iteratively selects samples by choosing the point farthest from all previously chosen points in the feature space. This approach ensures the training set uniformly covers the distribution of predictors.
Use Cases: Ideal for small to medium datasets where uniform feature coverage improves model generalization.
Pros:
- Deterministic selection ensures reproducible splits.
- Provides diverse representation of feature space.
Cons:
- Requires storing and computing full distance matrix, which can be memory-intensive.
Lazy Kennard–Stone Split
Description: A streaming variant of Kennard–Stone that computes distances on the fly, avoiding the full matrix. Suitable when memory is constrained.
Use Cases: Large datasets where memory is a bottleneck.
Pros:
- Reduced memory overhead.
Cons:
- May incur extra computational overhead due to repeated distance computations.
SPXY Split
Description: Extends Kennard–Stone to both features (X) and target variable (y), selecting samples that maximize combined spread. Ensures training set represents both predictors and response.
Use Cases: Regression tasks where preserving target distribution is critical.
Pros:
- Balances feature and response diversity.
Cons:
- Still requires pairwise distance computations; may overweight target variation.
OptiSim Split
Description: Applies iterative swapping of samples between train and test sets to minimize within-set dissimilarity, guided by a dissimilarity measure.
Use Cases: When split quality is paramount and computational resources allow.
Pros:
- Can yield high-quality, low-dissimilarity splits.
Cons:
- Computationally intensive; sensitive to initial split.
Minimum/Maximum Dissimilarity Split
Description: Greedy cluster-based strategy that adds clusters to the training set based on minimal or maximal dissimilarity criteria until the desired fraction is reached.
Use Cases: When a fast, cluster-aware splitting heuristic is sufficient.
Pros:
- Simpler and faster than full optimization.
Cons:
- Greedy nature may miss globally optimal arrangement; cluster sizes may vary.
Sphere Exclusion Split
Description: Forms clusters by selecting a sample and excluding all neighbors within a specified radius. Entire clusters are then assigned to train or test to meet the split fraction.
Use Cases: Spatial or similarity-based data where locality grouping matters.
Pros:
- Intuitive radius control over cluster sizes.
Cons:
- Sensitive to radius choice; can produce imbalanced clusters.
Cluster Shuffle Split
Description: Shuffles cluster labels and sequentially collects whole clusters into the training set until reaching the specified fraction, preserving inherent group structure.
Use Cases: Grouped or clustered data where splitting within groups is undesirable.
Pros:
- Maintains group integrity; introduces randomness.
Cons:
- May overshoot or undershoot desired fraction if cluster sizes vary.
TargetPropertySplit
Description: Partitions data into train/test sets by sorting samples according to a user-specified property (e.g., a column, a function of the sample, or a target value). The order argument (or alias) controls whether the largest/smallest property values are placed in the training set.
Use Cases: Useful for extrapolation or interpolation splits, e.g., when you want to train on the lowest (or highest) values of a property and test on the rest.
Pros:
- Simple and interpretable
- Works with any property (target, feature, etc.)
- Flexible via property function and order aliases
Cons:
- Not diversity-based; may not represent the full data distribution
TimeSplit
Description: Splits a 1D array of dates/times into train/test sets, grouping by unique date/time values. No group (samples with the same date) is split between train and test. The actual fraction may be slightly above the requested one, but never below.
Use Cases: Useful for time series or temporal data where you want to avoid splitting samples with the same timestamp.
Pros:
- Respects temporal order
- Never splits samples with the same date/time
Cons:
- Fraction may not be exact due to grouping