TastePaths: Enabling Deeper Exploration and Understanding of Personal Preferences in Recommender Systems

Document Title

TastePaths: Enabling Deeper Exploration and Understanding of Personal Preferences in Recommender Systems

Document Information

• Topic Area: Recommender Systems and Personalized User Exploration
• Keywords: Recommender Systems, Music Exploration, Interactive Search, Visualization, User Preferences
• Conference Information: IUI '22, March 22–25, 2022, Helsinki, Finland

Research Background and Problem

Background

Recommender systems play a critical role in daily information consumption, yet their traditional linear content presentation often traps users in "filter bubbles," limiting their ability to explore new domains and deepen their understanding of personal preferences. Particularly in the context of music recommendation, this limitation hinders users from discovering their interests more profoundly.

Problems and Motivation

1. Existing recommender systems overly focus on short-term consumption satisfaction rather than aiding users in achieving "self-actualization."
2. Users struggle to effectively learn and expand their interests when exploring unfamiliar domains using recommender systems.
3. Recommender systems lack interactive tools to support user control over the exploration process and understanding of recommendation algorithms.

Research Questions

1. Role of personalization in music exploration: How can personalized data help users delve deeper into and better understand their preferred music styles?
2. Optimization of exploration strategies: If the linear constraints of recommender systems are removed, what exploration strategies will users adopt? How can these strategies be better supported?
3. Learning user preferences: How does learning about music preferences help users improve their ability to express interests in a recommendation environment?

Solution

Methods and Innovations

1. TastePaths Tool: An interactive web tool that visualizes related artists within music genres using a graph-based approach.
2. Design Goals:
   • D1: Anchoring Artists: Start exploration with artists familiar to the user, supporting both personalized and non-personalized versions.
   • D2: Genre Space Overview: Provide an overall view of the music genre space through graph-based grouping, marking representative labels for subgenres.
   • D3: In-depth Exploration: Quickly present and facilitate detailed research on specific artists' musical information.

Implementation Steps and Key Technologies

1. Generating Music Graphs: Use knowledge graph data from music streaming services to construct a highly relevant graph of 150 nodes centered around three anchoring artists.
2. Clustering and Label Assignment: Apply the Louvain algorithm for graph clustering and use TF-IDF to select representative subgenre labels for each cluster.
3. Navigation Optimization: Highlight green paths from anchoring artists to key nodes in subgenres, providing guidance for user exploration.

Experimental Design

1. Conduct a comparative study with 16 participants to examine the performance of personalized and non-personalized versions of TastePaths in music exploration.
2. Collect data through questionnaires, operation logs, and interviews, combining quantitative and qualitative methods to evaluate exploration outcomes and user preferences under both conditions.

Research Findings

Experimental Results

1. Superiority of Personalization:

   • The personalized version of TastePaths helped users more effectively find music matching their personal interests.
   • Users reported higher satisfaction with the exploration experience using the personalized graph and collected significantly more songs compared to the non-personalized version.

2. Exploration Strategies:

   • Users discovered the most meaningful music by navigating genre space boundaries or intersections between clusters.
   • Various nonlinear exploration strategies were adopted, such as starting from anchoring artists or gradually traversing graph subsets.

3. Desire for Greater Control:

   • Users expressed a desire to dynamically expand or prune the graph structure to highlight areas relevant to their personal interests.
   • There was a demand for personalized navigation paths, such as real-time adjustments to guidance paths based on immediate feedback.

4. Enhanced Understanding Models:

   • TastePaths helped users comprehend how the recommender system organizes music.
   • Users formed clearer mental maps of genres and their components, learning how to express their interests more effectively.

Limitations and Future Directions

1. The shape and density of the graph structure may influence user exploration behavior; future research could investigate optimal configurations.
2. Information richness is insufficient: enhancing the display of subgenre characteristics, historical context, and artist backgrounds is necessary.
3. The test sample was limited to users open to exploring new music; future studies should expand to a broader user base.

Research Impact and Insights

1. Provides innovative interaction methods to guide users toward "self-actualization" in recommender systems, expanding interest boundaries.
2. Enhances user control and feedback mechanisms, potentially improving recommendation algorithms and user satisfaction.
3. Supports responsible recommendation design principles, such as introducing exploration loops to clarify exploration outcomes and a sense of goal completion.

Conclusion

The TastePaths project not only enhances users' understanding of their musical interests but also demonstrates the potential of interactive recommender systems in self-discovery and interest exploration. This offers significant directions for designing more proactive and user-centered recommendation tools in the future.