TL;DR
This paper surveys recommender systems research not by their technical approaches (content-based vs. collaborative), but by how they connect people. It argues that all recommender systems are fundamentally about building social connections, either explicitly through user modeling or implicitly by discovering relationships in data, and proposes a "connection-centric" framework for understanding them.
What Problem Does This Paper Address?
The authors observe that recommender systems research has traditionally been categorized by filtering techniques (content-based vs. collaborative). However, they argue this overlooks a crucial aspect: recommender systems operate within a social context and their ultimate goal is to connect people. This "connection-centric" viewpoint is under-emphasized in the literature. The paper aims to reframe the understanding of recommender systems by focusing on how they model users to foster these connections, whether directly or indirectly.
Methods
The paper proposes a "connection-centric" viewpoint as its primary methodological lens. It categorizes recommender systems based on how they model users to bring people together:
- Explicit User Modeling: Directly soliciting user preferences (e.g., ratings, profiles) to build connections.
- Implicit User Modeling: Inferring user interests from their behavior (e.g., clickstream data, usage logs) to discover existing social networks and connections.
The survey then explores various techniques within these categories, including:
- Content-based filtering: Briefly mentioned as a precursor but excluded from the main survey due to its focus on item features rather than user connections.
- Collaborative filtering: A major focus, with discussions on its evolution and various implementations.
- Social Network Analysis: Techniques for mining and exploiting social structures (e.g., link analysis, small-world networks) to facilitate recommendations.
- Hybrid Approaches: Combining explicit and implicit methods, or content-based and collaborative filtering.
Datasets
The paper discusses various types of data used in recommender systems research, but does not list specific datasets used for training or evaluation in a structured manner. Instead, it refers to general data sources such as:
- User ratings (e.g., for movies, books)
- Usenet news messages
- Bookmark folders
- E-mail logs
- Web documents and web logs
- Transaction data
- Movie rating datasets (e.g., EachMovie, MovieLens)
Key Insights and Ideas
- Recommendation is Social: The core idea is that all recommender systems, at their heart, are about connecting people. This social aspect is more fundamental than the filtering technique used.
- Two Paths to Connection: Recommender systems connect people either by explicitly asking users about their preferences (explicit modeling) or by observing their behavior to infer relationships (implicit modeling).
- Implicit Modeling's Power: Implicit modeling, by leveraging existing social structures and user behavior, offers a more natural and potentially richer way to discover connections and foster serendipity.
- Social Network Analysis is Key: Understanding and mining social networks (e.g., through link analysis, small-world properties) is crucial for building more effective and insightful recommender systems.
- Beyond Predictive Accuracy: Traditional evaluation metrics like precision and recall are insufficient because they ignore the social process and user experience of recommendation.
Key Results
The paper is a survey and conceptual framework, not an empirical study presenting quantitative results. Its key contributions are:
- A Connection-Centric Framework: A new way to categorize and understand recommender systems research based on their social connection-building mechanisms.
- Emphasis on Implicit Modeling: Highlighting the growing importance and potential of implicit user modeling and social network discovery for recommendation.
- Identification of Broadening Issues: Discussing critical aspects beyond core algorithms, including evaluation, targeting, privacy, and trust, from a social perspective.
Limitations and Future Work
Limitations:
- Focus on Connections: The paper acknowledges that by focusing on connections, it might de-emphasize certain aspects of content-based filtering, which it states do not find place in its survey.
- Broadening Issues are Cursory: The discussion on evaluation, targeting, privacy, and trust is brief and intended to provide pointers rather than exhaustive coverage.
Future Work:
- Distributed Recommendation Infrastructure: Developing a broader, distributed infrastructure for recommendation where users can manage their own models and share them.
- Formal Modeling and Design Methodologies: Creating unified methodologies for studying, designing, and evaluating recommender systems, especially to address issues like sparsity and privacy.
- HCI Approach: Designing less intrusive and more salient interfaces for user interaction with recommender systems, including usability evaluations.
- Information Appliances: Leveraging ubiquitous computing devices and the "voice web" to deliver "recommendations on demand."
- Mediating Explicit and Implicit Approaches: The future of recommender systems lies in effectively combining user modeling with the exploitation of combinatorial social structures.
Summary updated