Recommending topics for self-descriptions in online user profiles

Werner Geyer, Casey Dugan, David R. Millen, Michael J. Muller, Jill FreyneRecSys 2008 · DOI

TL;DR

This paper tackles the challenge of low user engagement in online profile self-descriptions by developing a recommender system that suggests topics for users to create, rather than just content to consume. Their findings show that recommending topics significantly increases profile content creation and updates, with recommendations based on social network connections proving more effective than content-based ones.

What Problem Does This Paper Address?

Online social networking sites often allow users to create rich self-descriptions through freeform "About You" sections. However, this flexibility paradoxically leads to low participation and infrequent updates, with many users creating few entries and rarely modifying them. The authors identify this as a critical problem for impression formation and social connection. Traditional recommender systems focus on helping users consume content (e.g., recommending movies or books). This paper addresses the novel problem of recommending content for users to create, specifically "About You" topics, aiming to boost user engagement and profile richness.

Methods

The core of the paper is the development and evaluation of an "About You" recommender system. They propose three main algorithms for generating recommendations:

  • Content-based: This algorithm mines user-provided text (status messages, existing "About You" entries, shared content) and corporate directory information (job title, location) to find topics similar in content to what the user has already shared.
  • Network-based: This approach leverages the user's social network (friends list). It recommends topics that have been used by at least one of the user's friends. This is a novel application of social network information for content creation recommendations.
  • Excluded: This serves as a baseline, recommending topics that were not suggested by the other two algorithms, essentially random selections from the remaining pool.

The system was evaluated through two studies: a preliminary evaluation with 100 users and a larger controlled study with 2,000 users, comparing an experimental group receiving recommendations against a control group.

Datasets

The primary dataset used is from Beehive, IBM's internal enterprise social networking site. The authors collected data on user-created "About You" entries from 5,788 members, identifying 2,861 distinct "About You" candidates for recommendations. They also utilized corporate directory information. No new benchmarks were introduced, but the study uses real user data from an existing social network.

Key Insights and Ideas

  • Recommending Creation Boosts Engagement: The fundamental idea that recommending topics for creation can significantly increase user participation is validated. This shifts the recommender system paradigm from consumption to active contribution.
  • Social Network is King for Creation: Recommendations derived from a user's social network (what their friends are talking about) are significantly more effective for driving "About You" content creation than content-based approaches. This suggests that users are influenced by their peers' self-expression and are more likely to engage with topics that resonate within their social circle.
  • Profiles Become Less Stale: The recommender system not only increased the number of entries but also extended the time between a user's first and last entry, indicating that recommendations help maintain profile activity over time.
  • User Preference for Self-Representation: Users are highly selective about their self-descriptions, prioritizing topics that they feel "represent me" or reveal something meaningful about them. This highlights the importance of impression management in online profiles.
  • Algorithm Simplicity vs. Domain Complexity: While the algorithms used were relatively straightforward (e.g., binary overlap for content, simple presence for network), they yielded significant results, underscoring that the domain of self-description recommendation is more about understanding user intent and personality than complex matching.

Key Results

  • Increased Entries: The experimental group (receiving recommendations) added 663 new "About You" entries, compared to 169 in the control group. This represents a 22% participation rate in the experimental group versus 8.4% in the control group.
  • More Users Participating: The number of users adding "About You" entries for the first time increased by 7.8% in the experimental group, significantly more than the 3.5% in the control group.
  • Less Stale Profiles: The average time between a user's first and last "About You" entry increased from 11.22 days to 29.71 days in the experimental group post-experiment, indicating sustained engagement.
  • Algorithm Effectiveness: In the controlled study, the Network-based recommendation algorithm achieved the highest acceptance rate (27.6%), followed by Content-based (20.7%) and Excluded (15.8%). All differences were statistically significant.
  • Preliminary Evaluation: In the preliminary study, Network-based recommendations were liked most (44.1%), followed by Content-based (36.2%). Acceptance rates for creating entries were 13.2% for Network, 10.5% for Content, and 8.6% for Excluded.

Limitations and Future Work

  • Moderate Increase: While significant, the average increase in entries and user participation was described as "moderate." The authors suggest potential reasons include the effort required for content creation and the quality of recommendations.
  • Network Algorithm Coverage: The Network-based algorithm's coverage was limited to 48.4% of users who had friends with "About You" entries.
  • Duplicate Recommendations: The question reduction technique for identifying distinct candidates was not perfect, leading to some similar recommendations, particularly for the Content-based algorithm.
  • Future Work:
    • Hybrid Algorithms: Investigating hybrid algorithms that combine Content and Network-based approaches to improve coverage and effectiveness.
  • Newcomer Support: Developing strategies to recommend content and connections for new users who lack established social networks.
  • Improved Ranking: Enhancing the Network algorithm by considering factors like friend proximity, relationship strength, and topic popularity.
  • Understanding Popularity: Exploring how to present less popular but potentially valuable topics effectively.
  • Source Attribution: Investigating whether revealing the source of a recommendation (e.g., "recommended by a friend") would further improve acceptance rates.

Summary updated

BibTeX
@inproceedings{geyer2008recommending,
  author = {Werner Geyer and Casey Dugan and David R. Millen and Michael J. Muller and Jill Freyne},
  title = {{Recommending topics for self-descriptions in online user profiles}},
  booktitle = {RecSys},
  year = {2008},
  doi = {10.1145/1454008.1454019},
}

Loading related content...

Figure