AOCL-Data Analytics is a new data analytics library providing optimized building blocks for data analysis and classical machine learning.
New Features
- Approximate nearest neighbors (ANN) with IVFFlat algorithm: Scalable high-dimensional similarity search — useful for recommendation systems, image retrieval, and vector embedding lookup.
- Radius neighbors classification and regression: Finds all samples within a given distance radius, suited for density-based classification and regression tasks.
- Model persistence: Trained models can be saved to disk and reloaded, enabling deployment of models without retraining.
- 1D cubic spline interpolation solver for smooth, continuous curve fitting.
Performance Improvements
- SVM training and inference stages are faster with optimized math operations and parallel sorting.
- DBSCAN clustering algorithm improved in multithreaded scenarios.
- Row/column-major data handling improved in k-means clustering training phase.
- Tuning of train_test_split kernels.
- Basic statistics now use recursive summation to deliver improved performance in both serial and parallel workloads.
Bug Fixes
- Moment statistics: Improved accuracy of summation phase for large datasets.
- Linear Models: Added a fallback solver to handle edge cases where the primary solver fails due to numerical conditions.