DTMiner: A Data-Centric System for Efficient Temporal Motif Mining

Key findings from this study

• The study found that data accesses in temporal motif mining exhibit strong spatial similarity and temporal monotonicity enabling optimization opportunities.
• The researchers demonstrate that DTMiner's Load-Explore-Synchronize execution model reduces redundant graph traversals by coordinating task execution around sequentially loaded temporal graph chunks.
• The authors report performance improvements of 1.14× to 11.98× over existing solutions through improved cache utilization and restricted memory fragmentation.

Overview

Temporal motif mining identifies recurring patterns in temporal graphs, which are critical for applications across multiple domains. Existing algorithms suffer from inefficiency due to redundant graph traversals and fragmented memory access patterns caused by irregular search tree expansion during motif matching tasks.

Methods and approach

DTMiner proposes a Load-Explore-Synchronize execution model that exploits spatial similarity and temporal monotonicity in data access patterns. The system sequentially loads temporal graph chunks into cache in temporal order. A fine-grained synchronization mechanism triggers all relevant tasks to explore only loaded data during search tree expansion. This approach enables task sharing of graph traversals for identical chunks while restricting fragmented memory accesses to cached data.

Results

Experimental evaluation demonstrates DTMiner achieves performance improvements ranging from 1.14× to 11.98× compared to existing state-of-the-art temporal motif mining solutions. The system effectively reduces data access overhead by regularizing irregular access patterns inherent in different motif matching tasks. The synchronization mechanism successfully constrains memory fragmentation to in-cache graph data, improving overall computational efficiency across diverse temporal graph datasets and motif types.

Implications

The data-centric approach identifies optimization opportunities in graph mining beyond algorithmic improvements alone. By reorganizing execution around data locality and access patterns, DTMiner demonstrates that system-level design can substantially improve performance for irregular workloads. This architecture may extend to other graph analysis tasks exhibiting similar characteristics of redundant traversals and irregular memory access patterns.

The observed properties of spatial similarity and temporal monotonicity in motif mining data accesses suggest broader applicability of the LES model. Future temporal graph analysis systems could adopt similar execution models to coordinate multiple concurrent tasks around shared data access patterns. The framework enables more efficient utilization of cache hierarchies in systems where multiple independent computation tasks operate over common data structures.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.