Pattern Recognition analytics shell
Nodal.PatternRecognition is the provider-neutral analytics layer of Nodal
Framework. It sits above Nodal.Core and provider packages, consumes canonical
graph results, and produces explainable similarity, community, sequence, and
temporal-transition evidence.
The package is under active alpha development. The bitset similarity kernel is the first executable slice; path extraction, candidate generation, community detection, and temporal APIs will follow behind explicit limits and benchmarks.
The dependency and data flow runs upward: providers → canonical Nodal graph results → Pattern Recognition analytics shell → application evidence. Optional native pushdown travels back down as a capability-checked optimization.
Why it is a shell
The database still owns storage, indexes, transactions, and native query execution. Pattern Recognition surrounds those capabilities with a portable analysis pipeline:
- Extract bounded canonical paths through any provider.
- Encode typed structural, positional, property, and temporal features.
- Generate a bounded candidate set instead of materializing all pairs.
- Score candidates with exact bitsets, sparse features, or optional vectors.
- Build similarity communities and directed temporal transitions.
- Return evidence, configuration, provider execution, and version metadata.
Provider-native algorithms can accelerate a stage only when the capability matrix certifies equivalent semantics. They never redefine the portable result silently.
First executable experiment
The first kernel packs typed multi-hot features into contiguous ulong words.
It computes XOR difference, AND intersection, OR union, normalized Hamming,
Jaccard, and binary cosine scores without allocating during comparison. A scalar
oracle validates unrolled and hardware-vector candidates across feature widths
and densities before any runtime dispatcher is selected.
See the P3 Journal entry for the benchmark design and roadmap for the remaining slices.