README.md

metadata

license: mit
task_categories:
  - feature-extraction
tags:
  - biology

SequenceDistanceDataset (SDD)

A Cross-Domain Benchmark for Sequence Similarity Analysis

Overview

A comprehensive benchmarking dataset for evaluating distance metrics in two domains:

• 🧬 Biological Sequences (Proteins from UniProt/UniRef)
• 🗺️ Movement Trajectories (GPS data from 3 cities)

Designed to support research in similarity search, metric learning, and cross-domain analysis.

---

Key Features

✅ Precomputed Distance Matrices

• Eliminates computation overhead for direct benchmarking
• Includes both training and test sets

✅ Curated Data Splits

• Standardized train/query/base splits for retrieval tasks
• Reproducible evaluation protocols

✅ Diverse Metrics

Domain	Metrics
Proteins	Edit Distance (ED), Needleman-Wunsch (NW)
Trajectories	DTW, Hausdorff, Fréchet, EDR, EDwP

✅ Real-World Scale

• Contains large mobility datasets (e.g., Porto taxi trajectories)
• Represents diverse geographical contexts (Beijing, Chengdu, Portugal)

---

Dataset Structure

Protein Domain

Protein/
├── {UniProt, UniRef}/
│   ├── train_list.txt
│   ├── query_list.txt
│   ├── base_list.txt
│   └── {ED, NW}/  
│       ├── {train/test}_distance_matrix_result

Trajectory Domain

Trajectory/
├── {Geolife, Porto, Chengdu, TrajCL_Porto}/
│   ├── train/
│   ├── query/
│   ├── base/
│   └── {DTW, Haus, DFD, EDR, EDwP}/  
│       ├── {train/test}_distance_matrix_result

Use Cases

🔬 ​​Bioinformatics​​

• Compare protein alignment algorithms
• Study sequence homology detection

🚕 ​​Urban Computing​​

• Evaluate trajectory similarity methods
• Analyze mobility pattern variations

🤖 ​​Machine Learning​​

• Train distance metric learning models
• Benchmark neural embedding methods