Social Lens: Personalization Around User Defined Collections for Filtering Enterprise Message Streams

Elizabeth Daly, Michael Müller, Liang Gou, David R. Millen

Abstract

Social media has led to a data explosion and has begun to play an ever increasing role as a valuable source of information and a mechanism for information discovery. The wealth of data highlights the need for methods to filter and sort information in order to allow users to discover useful information. Most traditional solutions focus on the user, either the user's social network, or a form of personalization based on collaborative filtering or predictive user modeling. This paper presents a novel algorithm to view information through a lens based on a user defined collection while excluding the attributes of the user from the analysis. As a result, the lens is transparent, tunable and sharable amongst users and, additionally allows both a reduction in information overload while discovering new related content.

Introduction

In the modern enterprise, employees are increasingly overwhelmed by "message streams"—a constant flow of status updates, blog posts, wiki edits, and file shares. Traditional approaches to managing this information overload typically rely on ego-centric filtering (personalization based on an individual's social network or past behavior) or global discovery (highlighting what is popular across the entire organization).

However, ego-centric models suffer from several limitations. They face the "cold start" problem, requiring significant historical data to be effective. They can create "filter bubbles," limiting exposure to new information. Most importantly, they are inherently non-sharable; two colleagues working on the same project cannot easily share a "personalized" view because that view is tied to their unique identities. This paper introduces Social Lens, an algorithm that shifts the focus from the user to a topic-defined collection of entities.

The Social Lens Algorithm

The core innovation of Social Lens is building a filter around a user-defined "seed" collection rather than the user themselves. By excluding the user’s own metadata, the filter becomes a transparent and sharable "lens."

The Social Lens algorithm architecture, showing the flow from initial seed entities to weighted updates.

The process begins with a user selecting a few "seed" entities, such as specific colleagues or project wikis. The algorithm then performs a graph traversal across three interconnected networks: Person-to-Person, Entity-to-Entity (e.g., shared tags), and Person-to-Entity (e.g., authorship). Through a two-phase weighting process, the system identifies candidate entities and scores them using frequency, Friend-of-Friend similarity, and content similarity (TF-IDF). Finally, pseudo-relevance feedback re-weights the original seeds to ensure the most relevant components drive the final stream of updates.

Evaluation and Results

The system was evaluated within a complex enterprise social software environment at IBM. Researchers compared Social Lens against two baselines: "Top Updates" (ego-centric) and "Discover" (global/random).

Comparison of user ratings for interestingness across three streams: Social Lens, Top Updates, and Discover.

The results demonstrated that Social Lens significantly outperformed the baselines in terms of content relevance and "interestingness." As shown in the comparison of streams, Social Lens updates received the highest percentage of "interesting" ratings (approximately 42%), while the baseline streams were rated as "uninteresting" over 53% of the time.

Furthermore, the algorithm proved highly effective for discovery. Approximately 62% of the objects surfaced by the lens were authored by individuals outside the user's immediate social network, yet users consistently rated these objects as highly relevant. This suggests that Social Lens successfully breaks the "filter bubble" inherent in traditional social filtering.

Discussion and Conclusion

Social Lens aligns with the theory of "working spheres," acknowledging that employees operate in multiple distinct contexts throughout a workday. By allowing users to manually tune the importance of people versus objects through a simple UI, the system provides a level of transparency and control absent in "black-box" recommendation engines.

While the initial study was small, the findings highlight the potential for non-ego-centric personalization in the workplace. Future work will focus on improving the content snippets provided in updates and exploring how these lenses can evolve temporally as project priorities shift. Ultimately, Social Lens offers a path toward more collaborative and discoverable information management in the enterprise.

Summary updated

Abstract

Social media has led to a data explosion and has begun to play an ever increasing role as a valuable source of information and a mechanism for information discovery. The wealth of data highlights the need for methods to filter and sort information in order to allow users to discover useful information. Most traditional solutions focus on the user, either the user's social network, or a form of personalization based on collaborative filtering or predictive user modeling. This paper presents a novel algorithm to view information through a lens based on a user defined collection while excluding the attributes of the user from the analysis. As a result, the lens is transparent, tunable and sharable amongst users and, additionally allows both a reduction in information overload while discovering new related content.
BibTeX
@inproceedings{daly2021social,
  author = {Elizabeth Daly and Michael Müller and Liang Gou and David R. Millen},
  title = {{Social Lens: Personalization Around User Defined Collections for Filtering Enterprise Message Streams}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/icwsm.v5i1.14123},
}
Figure