Abstract
dc:description.abstractThis work presents a scalable and interpretable link prediction framework embedded nativelywithin the Multi-Agent Spatial Simulation (MASS) library. By extending MASS’s distributed graph infrastructure and property-aware computation model, we implement both classical topological heuristics and embedding-based approaches—most notably Fast Random Projection (FastRP) combined with k-Nearest Neighbors (kNN)—to infer potential connections in graph-structured data. Topological algorithms such as Adamic-Adar and Resource Allocation, implementedas distributed primitives, demonstrate parity with Neo4j in accuracy and outperform it in execution time on large-scale query workloads. FastRP embeddings are generated via an agent-driven propagation pipeline that mirrors adjacency-based diffusion, enabling fullgraph vector generation in distributed environments. Though the current FastRP + kNN pipeline in MASS exhibits higher latency due to agent overhead and synchronization, it achieves competitive recall, especially at higher K values, validating its utility for applications that prioritize coverage over ranking precision. Experimental results on the Cora citation network show that MASS supports interactiveand batch link prediction tasks at scale, offering a memory-local alternative to centralized systems like Neo4j. This project transforms MASS from a simulation-only platform into a programmable, graph-native AI engine—capable of powering graph reasoning tasks for knowledge graphs, recommendations, and retrieval-augmented generation.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Hotchandani, Sumit
- Advisor dc:contributor.advisor
-
- Fukuda, Munehiro
Subjects
dc:subject × 8Rights
dc:rights- Statement dc:rights
-
- none
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1773/53501