HSC-Bench

Service Composition

Task definition, QoS metrics, composition datasets, model families, and results for workflow-level service composition.

Task definition

Service composition converts complex user requirements into a service sequence, DAG, or executable workflow. Inputs include the user need, service library, functional constraints, input/output compatibility, and QoS constraints. The output should satisfy both functional requirements and non-functional optimization objectives.

Key distinction: composition is not just Top-K retrieval; it produces an executable workflow whose quality depends on service compatibility and aggregated QoS.

Composition datasets

Dataset	Domain	QoS	Requirement Type	Description
QWS	Web Service	RT, TP, Availability, Reliability	Simulated requirements	Classic QoS service composition dataset.
WS-Dream	Web Service	Response Time, Throughput	QoS prediction oriented	Can support recommendation, composition, and QoS prediction; functional semantics are weaker.
HSC	AI Model Service	Metadata + QoS	Real-inspired workflows	Provides AI service workflows from Hugging Face model services.
HSC+	AI Model Service	Response time, waiting time, reliability, successability	Generated realistic requirements and workflows	Core dataset for the benchmark.

Composition model library

Heuristic / Evolutionary Optimization

GA

Genetic algorithm baseline for QoS-aware service composition.

Code

Heuristic / Evolutionary Optimization

DAAGA

Evolutionary optimization method for multi-objective composition.

Code

Heuristic / Evolutionary Optimization

MWOA

Whale optimization variant for QoS composition search.

Code

Heuristic / Evolutionary Optimization

CSSA

Swarm intelligence baseline for service composition optimization.

Code

Heuristic / Evolutionary Optimization

SDFGA

Genetic algorithm variant for service dependency and QoS constraints.

Code

Heuristic / Evolutionary Optimization

BPSC

Population-based service composition optimization baseline.

Code

Heuristic / Evolutionary Optimization

PK-IDPSO

Particle swarm optimization method for QoS-aware composition.

Code

Learning-based Service Composition

GNNPN-SC

Graph neural network and pointer network based workflow generation model.

Code

Learning-based Service Composition

RL-based composition

Reinforcement learning formulation for sequential composition decisions.

Code

Learning-based Service Composition

Pointer-network based composition

Sequence generation baseline for executable service workflows.

Code

LLM / Agentic

LLM Planner

LLM-based planner for requirement-to-workflow generation.

Code

LLM / Agentic

Multi-agent pipeline

Agentic service generation pipeline from natural-language need to service plan.

Code

LLM / Agentic

Code completion agent

Agent that translates workflow plans into executable code artifacts.

Code

QoS and workflow metrics

Response Time ↓

Total execution latency or path-level response time, typically aggregated along the workflow path.

Cost ↓

Total service cost. If real cost is unavailable, the benchmark should define a simulated cost setting.

Throughput ↑

Workflow throughput, often determined by the bottleneck service.

Availability ↑

Probability that the composed service is available, commonly multiplied across component services.

Reliability ↑

Probability of successful execution for the workflow.

Utility ↑

Weighted normalized QoS score. The page should always document normalization and weights.

Composition results

Dataset Model Type

Model	Dataset	Type	Utility ↑	RT ↓	Cost ↓	Throughput ↑	Availability ↑	Reliability ↑	Code	Official	Unified Protocol
GNNPN-SC	HSC+	Learning-based	TBD	TBD	TBD	TBD	TBD	TBD	Link	Planned	Yes
SDFGA	HSC+	Optimization-based	TBD	TBD	TBD	TBD	TBD	TBD	Link	Planned	Yes
DAAGA	HSC+	Optimization-based	TBD	TBD	TBD	TBD	TBD	TBD	Link	Planned	Yes
GA	QWS	Optimization-based	TBD	TBD	TBD	TBD	TBD	TBD	Link	Yes	TBD
LLM Planner	HSC+	LLM-based	TBD	TBD	TBD	TBD	TBD	TBD	Link	Planned	Yes
Multi-agent pipeline	HSC+	LLM-based	TBD	TBD	TBD	TBD	TBD	TBD	Link	Planned	Yes