Finesse: Fine-Grained Feature Locality based Fast Resemblance Detection...
summary
Proposes replacing costly N-transform super-features by splitting each chunk into fixed-size subchunks, computing Rabin fingerprints per subchunk and grouping those into super-features to exploit fine-grained feature locality (about 87% of corresponding subchunks share features). Implements a deduplication+delta-compression prototype and evaluates on six datasets, reporting similarity computation speedups of 3.2×~3.5× and system throughput gains of 41%–85% while keeping compression ratios comparable.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 2:18Data Encoding
- 2:50Resemblance Detection
- 4:13Traditional Superficial Approach
- 5:12FineGrained Feature locality
- 6:10FineGrained Future locality
- 7:21Faster resemblance detection
- 7:52Two steps feature
- 8:43Grouping principle
- 9:51Computation overhead
- 11:05TCE
- 12:44Computing Speed
- 14:39Future Work
- 15:16Conclusion
- 16:15Question
related talks
22mLoopDelta: Embedding Locality-aware Opportunistic Delta Compression in Inline...
26mUsing Random Projections to Make Sense of High-Dimensional ...
46mML Use Cases at Think Big Analytics
24mEfficient Similar-Object Search Algorithms for Terabytes of Data
09mLogCrisp: Fast Aggregated Analysis on Large-scale Compressed Logs by Enabling...
39mBuilding Highly Scalable Facial Recognition Pipelines
41mIntegrating Functions via Distance Sensitive Hashing
29mLabeling Tools are Great, but What About Quality Checks?