A Survey of Decomposition-Based Evolutionary Multi-Objective Optimization: Part II - A Data Science Perspective

Mingyu Huang, Ke Li

Abstract

In the era of "big literature," traditional systematic reviews struggle to remain comprehensive and objective. This paper presents LitLA (Literature Landscape Analysis), an end-to-end, AI-enhanced workflow designed to automate the discovery and mapping of massive scientific domains. Using the MOEA/D optimization algorithm as a case study, the authors demonstrate how LitLA can identify research themes, measure field maturation, and predict future trends such as the integration of Large Language Models (LLMs).

Introduction

We are currently living in the era of "big literature." In many scientific fields, the number of publications is growing exponentially, making traditional Systematic Literature Reviews (SLRs)—where researchers manually read and categorize hundreds of papers—nearly impossible to keep "COOT" (Comprehensive, Objective, Open, and Transparent). Manual reviews are prone to selection bias and often struggle to capture the structural "skeleton" of a field, such as citation flows and cross-border collaborations. To address this, the authors propose a data-driven pipeline that automates literature discovery, categorization, and trend forecasting.

The LitLA Workflow

The researchers developed LitLA, a multi-stage workflow that moves beyond simple keyword searches to provide a holistic view of a research landscape.

The LitLA workflow: an end-to-end pipeline from database search to future trend forecasting.

As shown in the figure above, the workflow consists of several key stages:

  1. Knowledge Graph Construction: Building a heterogeneous graph connecting papers, authors, institutions, venues, and keywords.
  2. LLM-Enhanced Metadata: Utilizing GPT-4o for "data wrangling," including extracting standardized keywords and classifying "citation intent" (e.g., distinguishing between background references and core methodological extensions).
  3. Advanced Topic Modeling: Employing BERTopic, which combines deep learning embeddings with UMAP and HDBSCAN to identify distinct research themes.
  4. Network Science: Applying the CD index to measure "disruptiveness" and Main Path Analysis to identify the influential backbone of research history.
  5. Link Prediction: Treating future trends as a network problem to predict which currently unconnected concepts are likely to converge.

Case Study: The MOEA/D Landscape

The primary case study focused on the MOEA/D (Multi-objective Evolutionary Algorithm based on Decomposition) research landscape. The dataset included 5,404 papers from the Web of Science (2007–2024) and citation contexts for over 4,600 statements from Semantic Scholar S2ORC. The resulting Knowledge Graph mapped 10,532 researchers, 432 venues, and nearly 79,000 keywords.

Key Insights and Results

The analysis revealed several critical trends regarding the maturation and structure of the field:

  • Field Maturation: While the volume of papers is growing at an annual rate of 40.5%, the "disruptiveness" of these publications is declining. The field has transitioned from a revolutionary phase to a consolidation phase focused on refining existing ideas.
  • Geographic and Social Dynamics: China has emerged as the dominant powerhouse in this sub-field, accounting for over half of all involved researchers. Despite the large number of authors, the community exhibits a "small world" effect, with 70% of authors in the largest component being within three hops of the most central researcher.
  • Topic and Citation Analysis: The study identified 83 topics, split between methodological enhancements and application-driven research (led by Energy and Scheduling).
Statistical overview of the MOEA/D research landscape, including growth trends (A, B), subject areas (C), citation intent (D), and top publication venues (E).

As illustrated in the figure above, the citation intent analysis (Panel D) shows that only 25.6% of papers citing MOEA/D actually propose an "Extension" of the algorithm. A plurality (38.9%) use it as a "Method" for other problems, while 23.4% cite it for "Background."

Predicting the AI Convergence

By analyzing keyword co-occurrence patterns in 2023, the LitLA model successfully predicted that Large Language Models (LLMs) would become a major focus in evolutionary optimization in 2024—a trend subsequently validated by a surge in new publications. Expert validation from 20 domain specialists confirmed the utility of the workflow, with 95% finding the topic modeling insightful.

Limitations and Future Work

The study's reliance on papers explicitly citing the original MOEA/D work may introduce selection bias. Additionally, the BERTopic model requires careful hyperparameter tuning and expert consultation to ensure thematic clusters remain human-interpretable. Future work aims to expand LitLA to other rapidly evolving fields like quantum computing and climate science, while integrating topic modeling more closely with community detection to link specific researcher "cliques" to thematic breakthroughs.

Summary updated

BibTeX
@misc{huang2024survey,
  author = {Mingyu Huang and Ke Li},
  title = {{A Survey of Decomposition-Based Evolutionary Multi-Objective Optimization: Part II - A Data Science Perspective}},
  year = {2024},
  eprint = {2404.14228},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.14228},
}
Figure