Source-linked AI summary
ComLittee: Literature Discovery with Personal Elected Author Committees
Hyeonsu B. Kang, Nouran Soliman, Matt Latzke, Joseph Chee Chang, Jonathan Bragg
TL;DR
Literature discovery tools often center papers, while author-highlighting systems offer limited ways to capture users’ evolving knowledge of authors. ComLittee lets users curate authors and uses recommender and authorship-network signals to find and explain relevant authors and papers. In a within-subjects study, it improved reported discovery outcomes over a paper-centric baseline.
Problem
Many discovery systems center papers and do not capture or use users’ evolving knowledge of authors, limiting support for author-centric exploration.
Source: §1 · p. 1Method
ComLittee lets users curate seed papers and an author committee, then combines paper-recommender and authorship-network signals to recommend and explain relevant authors and papers.
Source: §1 · p. 2Results
In a within-subjects study (N=16), ComLittee outperformed a strong paper-centric baseline in author and paper discovery efficiency, author novelty, and paper interestingness.
Source: §1 · p. 2Takeaways & Limitations
The study demonstrates the feasibility and value of using curated authors to enrich relevance signals for further author and paper recommendations.
Source: §8 · p. 15Retrospective post-task ratings may invite confirmation bias, and short lab studies may leave time-pressure effects and shallow engagement unobserved.
Source: §7.1 · p. 12
Abstract
from arXiv · showhide
In order to help scholars understand and follow a research topic, significant research has been devoted to creating systems that help scholars discover relevant papers and authors. Recent approaches have shown the usefulness of highlighting relevant authors while scholars engage in paper discovery. However, these systems do not capture and utilize users' evolving knowledge of authors. We reflect on the design space and introduce ComLittee, a literature discovery system that supports author-centric exploration. In contrast to paper-centric interaction in prior systems, ComLittee's author-centric interaction supports curation of research threads from individual authors, finding new authors and papers with combined signals from a paper recommender and the curated authors' authorship graphs, and understanding them in the context of those signals. In a within-subjects experiment that compares to an author-highlighting approach, we demonstrate how ComLittee leads to a higher efficiency, quality, and novelty in author discovery that also improves paper discovery.
1 INTRODUCTION
Literature discovery systems commonly center papers, although authors and their publication networks can reveal related research threads. ComLittee makes authors a basis for discovery by letting users curate authors and combine that committee with recommendation and authorship-network signals.
1 INTRODUCTION
- Paper-centric tools often omit authorship-graph signals, even though authors’ related papers and networks can help reveal research threads, additional authors, and relevant papers.Systems that only highlight authors on surfaced papers may miss threads because papers are prioritized for predicted relevance, not author coverage.Source: §1 · p. 1
- Author-highlighting systems offer useful context, but limited author interaction and no way to save known-relevant authors constrain discovery beyond the surfaced papers.Without saved authors, recommendations rely on predicted relevance and may be harder for users to use as discovery mechanisms.Source: §1 · pp. 1–2
- ComLittee lets users save seed papers, curate a personal author committee, and combine paper-recommender scores with publication-network relations to update recommendations and explain relevance.Interactive filters help users find papers contributing to the relevance between authors.Source: §1 · p. 2
- In a within-subjects study (N=16), ComLittee showed gains over a strong paper-centric baseline in author and paper discovery efficiency, author novelty, and paper interestingness.The study also reports implications for future author-augmented literature discovery designs.Source: §1 · p. 2
2 RELATED WORK
The related work contrasts paper-centered discovery with author recommendation and interactive approaches, then draws on prior work about explanations to inform ComLittee’s design. Together, these strands frame the system’s author-centric exploration.
2 RELATED WORK
- The review covers paper-centric systems, expertise finding, interactive machine learning, and relevance explanations as inputs to research questions and system design.It presents these areas as the review’s organizing topics.Source: §2 · p. 2
2.1 Paper-centric Literature Discovery Systems
- Prior paper-centric systems support searching, triaging, organizing, and reading papers, while ComLittee uses authors as the mode of interaction and discovery while retaining paper exploration.The reviewed systems include query expansion, paper triage, thread organization, and collection overviews.Source: §2.1 · p. 3
2.2 Expert-finding Systems
- Expert-finding work includes co-authorship prediction, conference-based triadic closure, and faceted author recommendation [43]; studies also identify authors as valuable in literature discovery.Recent systems use authorship graphs to augment papers with author highlights [20].Source: §2.2 · p. 3
2.3 Interactive Machine Learning
- Regroup iteratively updates priors from feedback on group members; prior work also motivates studying how users discover novel authors as human and model relevance change over time.A cited study found recall-oriented objectives improved user perception and adoption willingness over precision-oriented objectives at equal performance.Source: §2.3 · pp. 3–4
2.4 Relevance Explanation
- Prior explanation designs use lenses and transparent recommendation sources, or connect recommendations to familiar papers and authors [20].ComLittee applies lens-style explanations at author and paper levels and connects recommended authors to authors users have saved.Source: §2.4 · p. 4
3 DESIGN SPACE FOR AUTHOR-AUGMENTED LITERATURE DISCOVERY
The design space organizes author-augmented discovery around seven axes, from user control and explanation targets to relevance sources and distance. ComLittee occupies choices including curated authors, multiple relevance signals, and indirect relations, while omitting direct author relations for its topic-focused task.
3 DESIGN SPACE FOR AUTHOR-AUGMENTED LITERATURE DISCOVERY
- The seven axes span user controllability, explained entity, update timing, presentation, purpose, relevance source, and relevance distance.The space is intended to help future work explore new axis values, axes, and combinations.Source: §3 · p. 4
Who
- ComLittee lets users manually add or remove indirect reference authors in an elected committee, rather than relying only on inferred authors.The design contrasts user steering with systems that minimize effort by inferring reference authors from publication and interaction logs.Source: §3 · p. 4
What; When
- ComLittee explains both papers and authors using paper-recommender scores and citation or coauthor relations; it updates recommendations sequentially rather than in batches.The text notes that low-latency sequential updates may trade off some accuracy, while the effect of accumulated feedback divergence remains open.Source: §3 · p. 4
Where; Why; How: Relevance Source
- ComLittee pushes authors as top-level recommendations and uses in-situ highlights for explanations, not author recommendations; it combines recommender, graph, and saved-committee signals.Its relevance sources include paper relevance, coauthor and cited-author relations, user interaction history, and saved committee authors.Source: §3 · pp. 4–5
How: Relevance Distance
- ComLittee supports indirect author relations but not direct relations to the user, because the experimental task concerns entities related to a topic rather than the user generally.The design space distinguishes direct from indirect reference-author relations.Source: §3 · p. 5
4 SYSTEM DESCRIPTION
ComLittee starts from topic papers and potentially familiar authors, then uses paper-relevance signals and authorship connections to recommend authors and papers. Its interface lets users curate authors, inspect recommendation explanations, and explore relevant publications.
4 SYSTEM DESCRIPTION
- ComLittee initializes from topic papers and potentially familiar authors, then helps users discover other relevant authors and their papers through connections to familiar committee authors.The system assumes users can find at least five relevant papers, though a familiar author committee is optional.Source: §4 · p. 5, §4.2 · p. 6
4.1 Example User Scenario
- The example workflow moves from importing topic papers and saving familiar authors to reviewing their relevant papers and exploring unfamiliar authors linked through citations.As the author committee grows, the example describes the system continuing to recommend authors and improve explanations and paper rankings.Source: §4.1 · pp. 5–6
4.2 Author-Centric Recommendation Strategies
- Authors of saved papers are ranked by frequency, while authors of the 100 most topically relevant recent papers also receive votes.The recent-paper search covers publications from the last 180 days and can use saved or downvoted papers as its basis.Source: §4.2.1 · p. 6, §4.2.2 · p. 6
- Coauthorship expansion ranks authors from relevant papers by saved authors, while citation expansion ranks authors through references cited by saved authors.Citation votes are assigned per saved author, self-citations are excluded, and up to 100 highly voted references are sampled.Source: §4.2.3 · p. 6, §4.2.4 · p. 6
4.2.5 Batch generation.
- ComLittee interleaves recommendations from four ranked lists in batches of eight and updates recommendation explanations as users interact.Explanation filters connect author recommendations to relationships and relevant-paper counts; paper explanations show predicted relevance and frequent citing authors.Source: §4.2.5 · pp. 6–7, §4.3 · p. 7
4.4 Design Iterations
- Three pilot rounds and feedback from a senior UI designer informed iterations aimed at improving clarity and usability.The iterations addressed confusion points and usability, including loading-latency indicators.Source: §4.4 · p. 7
4.5 Final System Interface
- The interface combines saved-item panels, author publication histograms, relevance filters, paper feedback, author details, and expandable publication lists.Users can save or downvote papers, and saved authors are highlighted in paper recommendations.Source: §4.5 · p. 7
4.6 Baseline
- The baseline centers paper recommendations, with author hover details and FeedLens highlighting; both systems provide author-detail modals, while ComLittee adds more relevance explanations.The baseline’s author highlights update as the user encounters papers and the shared relevance model scores them.Source: §4.6 · pp. 7–8
5 EXPERIMENTAL DESIGN
The study compared ComLittee with a paper-centric author-highlighting baseline in timed, within-subjects discovery tasks. It measured discovery behavior and ratings, then analyzed quantitative and qualitative data against four research questions.
5.1 Objective & Research Questions
- The evaluation asked whether ComLittee improved author discovery efficiency and quality, supported discovery of unfamiliar relevant authors, and affected paper discovery and effort.Efficiency was saved items per fixed time, quality was mean relevance or interestingness ratings, and novelty compared unfamiliar-relevant with known-relevant authors.Source: §5.1 · p. 8
5.2 Participants
- Sixteen active researchers participated, including eight women; their fields included HCI, NLP, information retrieval, neuroscience, and oncology.Participants were 28.3 years old on average (SD 4.32).Source: §5.2 · p. 8
5.3.1 Structure.
- In a within-subjects study, participants explored two self-selected topics, with systems randomly assigned to topics and presentation order counterbalanced.Each topic folder began with two participant-curated seed papers; the remote study lasted 1.5 hours.Source: §5.3.1 · p. 8, Figure 4 · p. 9
5.3.2 Tutorials.
- Each condition began with a system tutorial followed by a 15-minute task, during which participants saved relevant and interesting items and downvoted irrelevant papers.Participants were encouraged to save at least five authors and papers and to downvote at least three papers early to calibrate recommendations.Source: §5.3.2 · p. 8, §5.3.3 · p. 8
5.3.4 Post-task Ratings and Surveys.
- Post-task measures included familiarity, relevance, and interestingness ratings for sampled saved items, plus surveys on workload, performance, compatibility, learning, and discovery effectiveness.Item relevance and interestingness used seven-point Likert scales; survey workload used an adapted NASA-TLX scale.Source: §5.3.4 · p. 8
5.3.5 Data Collection.
- The researchers logged timestamped interactions and analyzed the research questions with paired tests, repeated-measures analyses, ANOVA, and qualitative coding of transcripts.Analyses included author novelty, paper discovery measures, recommendation relevance scores, and steering behavior.Source: §5.3.5 · pp. 8–9, §5.3.6 · p. 9
6 FINDINGS
ComLittee users saved more authors and papers than baseline users, with particularly strong discovery of unfamiliar authors. They also reported lower workload, greater discovery helpfulness, and higher interest in saved papers, while qualitative accounts described author-centered exploration as a way to find connections and earlier work.
6.1.1 Users saved more authors in ComLittee and found saved fa-
- ComLittee users saved more authors overall (10.6 vs. 6.6) and unfamiliar authors (4.6 vs. 1.7), with no significant familiar-author count difference.Source: §6.1.1 · p. 9, Figure 5 · p. 10
6.1.2 Users found ComLittee features helpful for discovery.
- Users rated ComLittee more helpful for author and paper discovery and reported lower workload, greater workflow compatibility, and higher adoption plausibility.Source: §6.1.2 · p. 10
6.2 RQ2. ComLittee users saved familiar authors to scaffold subsequent discovery of unfamiliar authors
- Both conditions began with about twice as many familiar as unfamiliar saved authors, but ComLittee users reached familiar–unfamiliar parity in the second half.Source: Figure 6 · p. 10, §6.2 · p. 10
6.3.1 Efficiency.
- Users saved more papers in ComLittee (25.5 vs. 19.4) and downvoted fewer (11.2 vs. 23.2) than in baseline.Source: §6.3.1 · p. 11
6.3.2 Quality.
- Saved papers had higher average interestingness in ComLittee (6.1 vs. 5.6), with a significant difference for unfamiliar papers but not familiar papers.Source: §6.3.2 · p. 11
6.4.1 Users saved multiple papers from each discovered author at once; 6.4.2 ComLittee users expended significantly less effort during pa-
- Author-page visits yielded over three relevant papers at once; ComLittee users gave less negative feedback per saved-paper count overall, though more per author page than baseline users.Source: §6.4.1 · p. 11, §6.4.2 · p. 11
6.4.3 Human-AI alignment on relevance was improved in ComLittee.
- With the relevance model held constant, saved-paper relevance scores were higher in ComLittee than baseline in the final quarter (0.18 vs. 0.10).Source: §6.4.3 · p. 11
6.4.4 ComLittee users felt authors represented contextualized
- Users described finding unfamiliar-yet-relevant authors, connecting authors, and discovering earlier or less familiar work by known authors.Source: §6.4.4 · p. 11
7 DISCUSSION AND FUTURE WORK
The discussion reports early growth in saved authors’ connections to existing committee members, alongside longer recommendation latency and challenges steering discovery. It identifies study-design constraints and future directions for diversifying recommendations, adapting to changing feedback, and supporting exploration over time.
7.1 Study Design Limitations
- Post-task ratings may be affected by confirmation bias, short lab sessions may prompt shallow engagement, and longer-term effects require a prolonged field deployment.Participants chose their topics; they were instructed to select topics at a similar level of abstraction, and topics were randomly assigned to conditions.Source: §7.1 · p. 12
7.2.1 Latency of recommendations.
- Recommendation requests averaged 12.9s in ComLittee versus 4.1s in the baseline, and the authors identify optimization as important for scaling and longer-term use.The longer latency occurred despite observed discovery efficiency; the authors suggest pre-computed summary embeddings as one optimization avenue.Source: §7.2.1 · p. 12
7.2.2 Combatting early cold-start phenomena.
- 19% to 87% of saved authors had a relation to a saved author from the first to second quarter, while the utility of relation strength became marginal over time.The relation percentage plateaued after the second quarter; relation strength refers to the number of papers connecting a recommended author and a saved author.Source: §7.2.2 · p. 13, Figure 9 · p. 13
7.2.3 Getting stuck in a particular ‘school of thought’ vs. steering the system to diversify discovery in later stages.
- Users valued author relations but reported difficulty steering recommendations beyond close-knit groups, and some questioned whether early recommendations favor established authors over emerging researchers.One participant blocked well-known authors and then received less familiar names; users also described risks of echo chambers and difficulty leaving interconnected groups.Source: §7.2.3 · p. 13
7.3 Beyond triadic closure on citation networks
- Familiar authors can serve as navigational springboards to unfamiliar authors, but users later raised concerns about branching out of closely related groups.The authors present triadic closure on authorship graphs as feasible for this discovery path and identify surfacing effective navigational author nodes as future work.Source: §7.3 · p. 14
7.4 Finding non-interacting bodies of literature
- Citation-network methods may make work outside frequently co-cited bodies of literature less discoverable, motivating complementary cross-domain retrieval approaches.The discussion points to analogical relations between papers and approaches that explicitly use differences between knowledge domains.Source: §7.4 · p. 14
7.5 Diverse intentions of user feedback
- Users’ relevance intentions may shift between broader and more focused results, so future systems could support feedback with different degrees of change and reusable subsets.Suggested affordances include saving intermediate results and subselecting feedback to experiment with model behavior.Source: §7.5 · p. 14
7.6 Flow of time
- Users leveraged familiar authors early and sought novel authors later, while accumulated feedback could make later steering laborious.The discussion suggests filtering or prioritizing by attributes, reviewing feedback history, and allowing users to retire feedback that is no longer relevant.Source: §7.6 · p. 14
8 CONCLUSION
ComLittee supports author-centric literature exploration by letting users curate relevant authors and use those authors to generate relevance signals for further recommendations. In a controlled study, it was feasible and valuable, with more relevant papers and authors discovered in a given time, higher author novelty ratings, and more interesting paper ratings.
8 CONCLUSION
- ComLittee uses curated authors to compute relevance signals for other authors and their work, then uses those signals to recommend authors and contextualize recommendations.The conclusion reports that, in a controlled study, participants discovered more relevant papers and authors in a given time, rated authors as more novel, and rated papers as more interesting.Source: §8 · p. 15
A HOW AUTHOR RECOMMENDATIONS WERE SORTED IN A BATCH FOR PRESENTATION
The appendix describes a relevance ratio for ordering authors within recommendation batches, designed to make relevance signals easier to compare across authors.
A HOW AUTHOR RECOMMENDATIONS WERE SORTED IN A BATCH FOR PRESENTATION
- The relevance ratio counts unique papers across an author’s relevance tags and normalizes that count by the author’s publication total.The normalization is intended to make comparisons fairer when authors differ in total publication counts.Source: A HOW AUTHOR RECOMMENDATIONS WERE… · p. 16
- After pilot feedback found interleaving recommendation sources confusing, the authors replaced that ordering with a relevance-based ratio.Participants noted that recommendations with several relevance explanations could appear lower than recommendations with none.Source: A HOW AUTHOR RECOMMENDATIONS WERE… · p. 16
B PSEUDOCODE OF AUTHOR RECOMMENDATION
The appendix points readers to pseudocode for author recommendation in Algorithm 1.
B PSEUDOCODE OF AUTHOR RECOMMENDATION
- Algorithm 1 presents pseudocode for author recommendation.The passage identifies the algorithm but does not describe its steps.Source: B PSEUDOCODE OF AUTHOR RECOMMENDATION · p. 17
C USERS’ ACTIONS REPRESENTED AUTHENTIC RELEVANCE AND INTERESTINGNESS
Users’ save actions appear to reflect authentic interest and relevance, including when they save authors.
C USERS’ ACTIONS REPRESENTED AUTHENTIC RELEVANCE AND INTERESTINGNESS
- After removing authors rated 4 or lower on relevance, users saved significantly more authors in ComLittee (M=9.6, SD=2.63) than Baseline (M=6.7, SD=2.75; tpaired(29.94)=3.71, p=0.002).Source: C USERS’ ACTIONS REPRESENTED AUTHENTIC… · p. 17
- Users rated about 13 of 15 randomly sampled saved papers as at least 4 on the relevance scale in both conditions.The authors interpret this as evidence that saving papers reflected authentic relevance judgments, not merely recommender steering.Source: C USERS’ ACTIONS REPRESENTED AUTHENTIC… · p. 17
D DISTRIBUTION OF POST-TASK RATINGS ON SAVED PAPERS AND AUTHORS, BY FAMILIARITY
The appendix presents distributions of post-task paper and author ratings, including ratings grouped by familiarity type.
D DISTRIBUTION OF POST-TASK RATINGS ON SAVED PAPERS AND AUTHORS, BY FAMILIARITY
- The distributions of counts on paper ratings and author ratings are shown in Figures 10 and 11, respectively; the passage says nothing about panel distinctions or numerical distribution values.Source: D DISTRIBUTION OF POST-TASK RATINGS ON… · p. 17
E DESCRIPTION OF DESIGN ITERATIONS
The design iterations adjusted recommendation explanations and presentation to better match pilot users’ intent and preferences while balancing information density and decision support.
E DESCRIPTION OF DESIGN ITERATIONS
- Pilot users expected “cited by” explanations to surface papers by saved authors that cited recommended authors, so self-citations were removed from author and paper citation explanations.The removal was intended to align explanations with user intent and avoid self-promotional citations.Source: E DESCRIPTION OF DESIGN ITERATIONS · p. 17
- To prioritize unseen publications while retaining familiar papers for judging recommendations, familiar papers with user feedback were collapsed into an expandable stack.The stack was ordered from most to least recently interacted with and placed at the top of the author’s publications list.Source: E DESCRIPTION OF DESIGN ITERATIONS · p. 17
- Because pilot users found predicted relevant-paper counts useful but potentially misleading about recommendation rank, the count was made less prominent.The count was not the main determinant of recommendation order.Source: E DESCRIPTION OF DESIGN ITERATIONS · p. 17
F ADDITIONAL SURVEY RESULTS
Additional survey results favored ComLittee on workload and discovery-helpfulness measures, while several workflow and feature comparisons showed no significant difference.
F ADDITIONAL SURVEY RESULTS
- The appendix describes additional questionnaire items and says two-sided paired-sample t-tests were used to compute p-values between conditions.It directs readers to Section 6.1.2 for discussion of the results.Source: F ADDITIONAL SURVEY RESULTS · p. 17
- The excerpt reports two scores, 3.9 (SD=1.77) and 4.9 (SD=1.59), with p=.02 for the item; it does not identify which condition received either score.Source: F ADDITIONAL SURVEY RESULTS · p. 20
- ComLittee responses were significantly more favorable on all author-discovery helpfulness questions, and overall workload was significantly lower than Baseline.Adoption plausibility was higher, but overall technology-acceptance responses did not differ significantly.Source: Table 2 · p. 20
- ComLittee responses were significantly more favorable on all paper-discovery helpfulness questions; common-feature and system-specific-feature ratings showed no significant advantage.When aggregating each user’s most-favored feature, ComLittee had a significantly higher average.Source: Table 3 · p. 21