FeedLens: Polymorphic Lenses for Personalizing Exploratory Search over Knowledge Graphs

Document Title

FeedLens: Polymorphic Lenses for Personalizing Exploratory Search over Knowledge Graphs

Document Information

• Domain: Exploratory search and personalized recommendation based on knowledge graphs
• Keywords: Knowledge graph, exploratory search, recommendation systems, interaction techniques, polymorphic lenses, user study

Research Background and Problem

• Problems and Challenges:

  • With the widespread use of knowledge graphs (KG) in information retrieval, their vast scale and openness make exploratory search complex and cognitively demanding for users.
  • Existing knowledge graph systems often rely heavily on recommendation models tailored to a single entity type, making it difficult to extend them for cross-entity search and navigation.
  • In academic literature search, users not only need to find relevant articles but also wish to efficiently discover related authors, institutions, and conferences.

• Significance:

  • For complex domains (e.g., scientific literature or e-commerce), existing exploratory search tools fail to sufficiently simplify user navigation and enhance knowledge discovery efficiency.
  • Personalized preference models have been widely adopted but have not been fully utilized for navigation and ranking across entities, limiting their potential in knowledge graphs.

• Motivation and Related Work:

  • Many studies have shown that creating "lenses" based on user preferences (e.g., filtering search results by keywords) can improve the relevance of search results.
  • However, most existing research is limited to constructing lenses for a single entity type, lacking exploration into cross-type generalization.

Solution

• Core Approach:

  • Propose a novel technique called "polymorphic lenses":
    • Reapply existing personalized preference models designed for a single entity type (e.g., papers) to other entity types and relationships (e.g., authors, institutions, conferences) in knowledge graphs.
    • Generalize user preference models to score, rank, and summarize multiple types of entities based on preferences for foundational entities.

• Innovations:

  • Simplicity and Generality: Reuse existing user preference models without reconstruction; ensure applicability to various entities and collections in complex knowledge graphs.
  • Dynamic Association: Extend customized filtering to new exploration endpoints across the entire knowledge graph.
  • System Implementation: Realized in the Semantic Scholar (S2) academic literature navigation system, with design optimized based on experimental feedback.

• Implementation Steps and Key Techniques:

  1. Define Polymorphic Lenses: Use specific aggregation functions (e.g., count-based scoring) to infer relevance scores for other entity types based on preference models for foundational entities.
  2. Interface Integration: Design new front-end features in Semantic Scholar to intuitively display lens-based scores (e.g., heatmaps and ranking features for authors, papers, and conferences).
  3. Backend Implementation:
     • Use Support Vector Machines (SVM) and SPECTER embedding models to construct preference scores for papers.
     • Propose clustering strategies (e.g., abstract embedding strategies) to optimize large-scale data processing efficiency for knowledge graphs.

Research Outcomes

• Main Contributions:

  • Proposed and validated the "polymorphic lenses" technique, enabling interconnection among multiple entity types in knowledge graphs and supporting efficient user navigation.
  • Developed a concrete system, FeedLens, and demonstrated its practicality in academic literature search.

• Experimental Results:

  • A comparative study with 15 participants revealed that FeedLens outperformed the baseline system, Semantic Scholar, in several aspects:
    • Efficiency: Users explored more content (system interactions increased by approximately 4x).
    • Relevance: Selected authors' related papers were of higher relevance (improvement of approximately 2x).
    • User Experience: Significantly reduced cognitive load (NASA-TLX scores decreased), and system usability scores improved (SUS scores increased from 77 to 84).
    • Users gave near-perfect ratings (average >4/5) for some important new features, such as author recommendations.

• Limitations and Future Directions:

  • The current study is validated only in the academic literature domain; future work could extend to other domains (e.g., e-commerce, online streaming recommendations).
  • Data-driven blind spots: Overemphasis on small-sample random recommendations for authors may affect overall exploration quality, requiring further optimization in evaluation.
  • Limited sample size in user studies; broader data collection is needed to verify generalizability.

Through this mechanism, FeedLens effectively enhances the efficiency and experience of exploratory search, exploring a new paradigm for applying personalized recommendations in knowledge graphs.