Knowledge-Based Navigation of Complex Information Spaces

AAAI 1996

TL;DR

This paper introduces "FINDME" systems, a novel approach to navigating large, complex information spaces by providing "assisted browsing." Instead of relying solely on rigid keyword queries, FINDME systems actively monitor user interactions and offer intelligent suggestions, helping users discover relevant information and understand the underlying structure of the data. This is particularly useful in domains where users have fuzzy goals or struggle to articulate precise search criteria.

What Problem Does This Paper Address?

The explosion of online information has created a significant challenge: how do users effectively find what they need in vast, multi-dimensional information spaces? Traditional keyword-based retrieval requires users to precisely define their queries, which is often difficult when users have vague ideas or are exploring unfamiliar domains. Browsing offers an alternative, but without structure, it can lead to users getting lost. The authors tackle this by developing "assisted browsing" systems that combine standard interfaces with intelligent guidance, aiming to make information discovery more intuitive and efficient, even for users with imprecise goals.

Methods

The core of the FINDME approach is "assisted browsing," which involves monitoring user progress and providing relevant assistance. This is achieved through several key mechanisms:

  • Initial Query/Zooming Stage: Users begin by stating a starting point within the information space.
  • Assisted Browsing Phase: Users traverse the space, and the system provides support.
  • Trade-off Explanation: The system helps users understand inherent trade-offs within a domain (e.g., the relationship between car horsepower and fuel efficiency).
  • Similarity-Based Retrieval: Users can find items similar to those they are currently viewing, but with a nuanced approach that considers user-specific feature weighting based on their goals.
  • Tweaking: Users can refine their search by critiquing an item and articulating new criteria based on that critique (e.g., "I like this apartment, but I want something cheaper").

FINDME systems are designed to be flexible, allowing users to redirect their search at any time without getting trapped. The paper highlights three specific FINDME systems: CAR NAVIGATOR (for new cars), VIDEO NAVIGATOR (for movie rentals), and RENTME (for apartment rentals).

Datasets

The paper does not explicitly list or describe datasets used for training or evaluation in a formal sense. Instead, it refers to the underlying data sources for each FINDME system:

  • CAR NAVIGATOR: Implicitly uses data from car specifications and ratings, presented in a magazine-like interface.
  • VIDEO NAVIGATOR: Utilizes a database of 7500 movies from a popular video reference work (Wiener, 1993).
  • RENTME: Operates on a database of classified ads for rental apartments in a metropolitan area. The system includes a natural language processing step to extract features from these terse ad texts.

Key Insights and Ideas

  • Assisted Browsing as a Paradigm: The paper argues that for complex domains with fuzzy user goals, assisted browsing is a more effective paradigm than pure retrieval or unguided browsing.
  • User-Centric Feature Weighting: Users don't always value features equally. The system's ability to infer or allow users to express the relative importance of features (e.g., engine size for towing vs. general driving) is crucial for effective similarity-based retrieval.
  • "Tweaking" as a Powerful Interaction Model: Allowing users to critique an item and then request variations ("cheaper," "nicer," "bigger") provides a natural and intuitive way to refine searches, especially when users struggle to articulate explicit criteria.
  • Knowledge-Based Heuristics are Key: The effectiveness of FINDME systems relies on incorporating domain-specific knowledge to understand trade-offs, relationships between features, and effective search strategies. This moves beyond simple keyword matching.
  • Flexibility is Paramount: Users should never feel trapped. The ability to easily redirect or restart a search is essential for a positive user experience in complex information spaces.

Key Results

The paper focuses on the conceptual framework and the qualitative demonstration of the FINDME systems rather than presenting extensive quantitative results. The key findings are demonstrated through examples:

  • CAR NAVIGATOR: Showcases how users can manipulate car features and receive explanations of trade-offs (e.g., high horsepower vs. fuel efficiency). Buttons like "sportier," "roomier," etc., allow for multi-variable adjustments.
  • VIDEO NAVIGATOR: Illustrates how "clerk" agents can suggest related movies based on genre, actors, or directors, mimicking the advice of knowledgeable friends.
  • RENTME: Demonstrates the "tweaking" mechanism with concrete examples of how users can refine apartment searches by requesting "cheaper" or "nicer" apartments, leading to the system exploring similar neighborhoods or adjusting price brackets. The system's natural language processing can distinguish nuanced phrases like "No dogs" vs. "Dogs welcome."

Limitations and Future Work

  • Knowledge Representation: The authors acknowledge that in RENTME, neighborhood and apartment qualities are currently represented by simplified numerical weights. They suggest that a more sophisticated approach is needed to better approximate human evaluation, especially for understanding baseline expectations associated with different rent magnitudes.
  • Domain Specificity: While the FINDME framework is general, the effectiveness of the knowledge-based retrieval strategies is tied to the specific domain. Elaborating the knowledge base for each domain is an ongoing effort.
  • User Interface Design: The paper notes that while direct manipulation of features was initially appealing, users often lacked the patience for it. This suggests that interface design needs to balance fine-grained control with intuitive, higher-level operations.
  • Future Work: The authors imply that further development would involve more sophisticated knowledge representation, richer natural language understanding, and potentially more advanced user modeling to better anticipate user needs and preferences.

Summary updated

BibTeX
@inproceedings{1996knowledge,
  title = {{Knowledge-Based Navigation of Complex Information Spaces}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
}
Figure