cs.SIFeb 25, 2026
SaveBridging Through Absence: How Comeback Researchers Bridge Knowledge Gaps Through Structural Re-emergence
Organizations: University College Cork, Cork, Ireland · Shanghai Jiao Tong University, Shanghai, China · Indian Institute of Information Technology, Guwahati, India · Techno India University, West Bengal, Kolkata, India
Abstract
Understanding the role of researchers who return to academia after prolonged inactivity, termed "comeback researchers", is crucial for developing inclusive models of scientific careers. This study investigates the structural and semantic behaviors of comeback researchers, focusing on their role in cross-disciplinary knowledge transfer and network reintegration. Using the AMiner citation dataset, we analyze 113,637 early-career researchers and identify 1,425 comeback cases based on a three-year-or-longer publication gap followed by renewed activity. We find that comeback researchers cite 126% more distinct communities and exhibit 7.6% higher bridging scores compared to dropouts. They also demonstrate 74% higher gap entropy, reflecting more irregular yet strategically impactful publication trajectories. Predictive models trained on these bridging- and entropy-based features achieve a 97% ROC-AUC, far outperforming the 54% ROC-AUC of baseline models using traditional metrics like publication count and h-index. Finally, we substantiate these results via a multi-lens validation. These findings highlight the unique contributions of comeback researchers and offer data-driven tools for their early identification and institutional support.
Figures & tables
Figure 1: Figure shows the publication-count heatmaps (log-bins) for (Left) Dropout and (Right) Comeback authors. The heatmaps are overlaid with scaled probability curves, where the empirical probability is shown as a solid line and the model fit as a dashed line. The x-axis represents the year index, and the y-axis represents the logarithmic bin of publication counts.
Figure 2: Figure shows the grouped bar chart showing the percentage share of different venue types for Comeback versus Dropout authors. The y-axis shows the percentage (%), and the x-axis categorizes the venue types: Other, Conference, Journal, Workshop, and Symposium.
Figure 3: The figure represents global geographic distribution of comeback-publication percentages by country. A zoomed inset focuses on Europe. The color intensity or value on the map corresponds to the percentage of comeback publications for each country.
| Existing Metric | Reference / Origin | Use in This Study |
| Publication Count ( ) | Standard bibliometric | Baseline for author productivity (volume control) |
| Citation Count ( ) | Standard bibliometric | Baseline for impact (volume control) |
| h-index ( ) | Hirsch et al. Hirsch (2005) | Career performance baseline; used in baseline-only models |
| Cross-Community Citation Rate (XCC) | Cunningham (2025) Cunningham and Greene (2025) | Paper-weighted share of citations outside the paper’s community; reported in Fig. 7 |
| Proposed Metric | Reference / Origin | Use in This Study |
| Bridging Score ( ) | This work | Edge-weighted fraction of an author’s outgoing citations that cross communities; main structural indicator (Figs. 4 , 5 , 11 , 8 , 9 ) |
Table 1: Summary of bibliographic metrics used in this study, categorized as Existing and Proposed. For each metric, the table lists its name, reference or origin, and its specific use within this study.
Figure 4: Figure represents distribution of Distinct Communities (log-scaled, left) and Bridging Score (right) across author categories (Comeback, Dropout, Active). The y-axis for the left plot is the log-scaled count of distinct communities. The y-axis for the right plot is the Bridging Score value.
Figure 5: Figure shows the bootstrapped distributions of the mean difference (Comeback – Dropout) for (left) Bridging Score and (right) Distinct Authored-Community Count (ACC). The x-axis shows the mean difference value, and the y-axis shows the frequency from bootstrapping. The vertical dashed lines indicate the observed mean differences.
Figure 6: Figure represents the distribution of Gap Entropy for Comeback vs. Dropout authors. The y-axis represents the density or frequency, and the x-axis represents the Gap Entropy value.
Figure 7: Figure shows the boxplot of Cross-Community Citation Rate (XCC) for Comeback vs. Dropout authors. The y-axis shows the Cross-Community Citation Rate.
Figure 8: Figure shows the moving window comparison of the average Bridging Score for Comeback vs. Dropout authors across 5, 10, and 15-year intervals. The x-axis likely represents the midpoint of the time window, and the y-axis represents the average Bridging Score.
Figure 9: The figure represents Kaplan-Meier survival curves showing the probability of academic persistence over time, stratified by a binary Bridging Score group (High vs. Low). The x-axis represents time (e.g., years), and the y-axis represents the survival probability.
| Paper | Author | ||
| Attributes | Attribute Definition | Attributes | Attribute Definition |
| index | Paper ID | index | Author ID |
| Paper Title | Title of the publication | Author | Author name |
| Authors | List of authors (separated by ‘;’) | Affiliation | Current affiliation |
| Published Year | Year of publication | Papers | Total paper count |
| Publication Venue | Conference or journal name | Citations | Total citation count |
Table 2: Data attributes extracted and processed from AMiner.
| Raw Dataset Summary | Value |
| Total Papers Processed | 2,092,356 |
| Total Authors Processed | 1,712,433 |
| Valid Authors with Parsed Details | 1,132,637 |
| Parsed Author-Paper Pairs | 4.7 million |
| Reference Edges Extracted | 20 million |
Table 3: Summary statistics of the raw parsed AMiner dataset.
| Curated Researcher Statistics | Value |
| Number of Comeback Authors | 1,425 |
| Average Year Gap (Comeback) | 6.2 years |
| Number of Dropout Authors | 11,351 |
| Average Active Years (Dropouts) | 4.7 years |
| Publication Span Considered | 1980—2014 |
Table 4: Basic statistics of comeback and dropout researchers used in this study.
Figure 10: The figure represents the schematic overview of the methodological workflow. The diagram shows the pipeline from data preprocessing and author segmentation to network analysis, metric computation, and final statistical and predictive evaluation.
| Metric | CB mean | DO mean | Mean diff [95% CI] | (CB–DO), | Effects (d/ ) |
| Bridging | 0.608 | 0.565 | 0.044 [0.035, 0.052] | 10.08, | 0.131 / 0.016 |
| ACC (communities) | 74.78 | 33.08 | 41.70 [38.88, 44.57] | 29.34, | 0.402 / 0.616 |
Table 5: Statistical comparison of the Bridging Score (B) and Authored-Community Count (ACC) between Comeback (CB) and Dropout (DO) cohorts within the observation window. The table reports group means, mean differences with 95% confidence intervals, Welch’s t-test results (t-statistic and p-value), and effect sizes (Cohen’s / Cliff’s )
Figure 11: The figure represents heatmap of the average Bridging Score across publication years, segmented by author category (Comeback, Dropout, Active). The x-axis represents the year, and the y-axis represents the author category. The color intensity corresponds to the average Bridging Score value.
Figure 12: Figure represents the sunburst visualization of the topic mix for each author cohort. Each ring aggregates the top subject terms by frequency within the observation window, with segment size proportional to term frequency.
| Model | Accuracy | Precision | Recall | F1 | ROC–AUC |
| Random Forest | 0.927 | 0.974 | 0.888 | 0.929 | 0.972 |
| XGBoost | 0.911 | 0.971 | 0.858 | 0.911 | 0.965 |
| LightGBM | 0.906 | 0.959 | 0.862 | 0.908 | 0.963 |
| KNN (k=5) | 0.900 | 0.950 | 0.859 | 0.902 | 0.943 |
| Gradient Boosting | 0.871 | 0.943 | 0.809 | 0.871 | 0.927 |
| SVM (RBF) | 0.735 | 0.802 | 0.672 | 0.731 | 0.792 |
Table 6: The Table shows the out-of-sample performance metrics for various classification models predicting comeback status, using the bridging and entropy feature set (ACC, B, XCC, ). The metrics reported are Accuracy, Precision, Recall, F1-score, and ROC-AUC, obtained via stratified cross-validation.
Figure 13: Figure shows the receiver Operating Characteristic (ROC) curves for various machine learning models predicting comeback status. The x-axis represents the False Positive Rate, and the y-axis represents the True Positive Rate. The diagonal dashed line represents a random classifier.
Figure 14: The figure represents the SHAP summary plot showing the impact of the top features on the model’s output. Each point represents a researcher instance. The x-axis is the SHAP value (impact on model output), and the y-axis lists the features. The color represents the feature value from low (blue) to high (red).
| Metric / Model | CB vs. DO (or Score) | Primary test / stat | Where shown |
| A. CB–DO Metric contrasts | |||
| Distinct authored communities (ACC) | CB: 74.78 (mean); DO: 33.08 (mean); CI excludes 0 | Bootstrap mean diff (95% CI ) | Figs. 4 , 5 |
| Bridging score ( ) | CB: 0.608 (mean); DO: 0.565 (mean); CI excludes 0 | Bootstrap mean diff (95% CI ) | Figs. 4 , 5 |
| Cross-community citation (XCC) | CB higher than DO | , | Fig. 7 |
| Gap entropy ( ) | CB: 2.65 (mean); DO: 1.52 (mean) | , | Fig. 6 |
| Kaplan-Meier survival (shows persistence) | CB shows higher survival | log-rank , | Fig. 9 |
Table 7: Metrics & validation scores summarizing Comeback (CB)–Dropout (DO) contrasts and predictive performance, with statistical tests and figure references.