Appendix

1 Dataset Statistics and Modeling

1.1 CS Subfield Definition and Statistics

Table 1: GF graph dataset statistics for 12 CS subfields. This table details each subfield's name, number of authors, topics, papers, links, and corresponding venues and coarse-grained topics.

CS Subfield Name	#Authors	#Topic	#Papers	#Links	Representative venue	Representative coarse-grained topics
Natural Language Processing	13240	136	208601	235217	ACL, LREC	Text mining, Speech recognition
Visualization	7081	84	85714	93646	TVCG, VIS	Visualization, Visual analytics
Computer Security	10732	132	232216	211868	USENIX Security, CCS	Cryptography, Malware detection
Graphics	5318	79	79627	76524	TVCG, CGF	Computer graphics (images), Rendering
Computer Architecture	20214	218	617377	737024	ISCA, MICRO	Parallel computing, Distributed systems
Software & Programming Language	12557	132	283600	396670	PLDI, ICSE	Programming languages, Software engineering
Human-Computer Interaction	5182	105	153033	182585	CHI, CSCW	User interface, User experience
Networking	20574	282	1049489	1338620	SIGCOMM, INFOCOM	Network protocols, Distributed systems
Graphics & Visualization	11980	105	271418	262957	TVCG, CGF	Visualization, Computer graphics (images)
Database & Data Science	12758	129	225653	194952	SIGMOD, VLDB	Database management, Data mining
Theory & Algorithms	24738	278	772323	891339	STOC, SODA	Algorithm, Theoretical computer science
Artificial Intelligence at Large	54637	497	4836561	8420229	AAAI, IJCAI	Machine learning, Neural networks

As described in Section III.A, we selected 12 representative subfields from the CS domain, as shown in Table 1. The table lists two representative venues and coarse-grained topics for each subfield (raw data, the full list of venues and coarse-grained topic can be found in GitHub). We then matched the corresponding coarse-grained topic and venues in the MAG database and extracted papers associated with these matched entities, including their authors and citation information, to construct a domain-specific database. For the defined domains, we selected top authors with an h-index above a certain threshold (default is 5, adjustable per field) and built self-citation graphs for each scholar. After constructing the graphs, we utilize the GeneticFlow graph model to conduct node and edge profiling in order to construct the GF graph, thereby identifying core papers and their citation relationships.

1.2 GeneticFlow Graph Model

The pre-processed academic data is modeled by GF analytics [1] to compute a GF graph for each recipient/scholar/venue. The GF graph is designed to represent the evolution of a scholar/venue's scientific impact and contribution. Take the GF analytics of a scholar $s$ as an example. His/her full GF graph is defined by $G = {V, E}$, where $V$ denotes the set of $n$ paper nodes authored by $s$ and $E$ denotes the set of $m$ reversed self-citation edges among $V$. The reversal of citation links is a key design decision in the GeneticFlow graph model, shifting the focus from citation links to citation influence modeling, which emphasizes influence and potential content inheritance from one paper to another, capturing the chronological progression of research contributions. Each node $v_i$ is further associated with a timestamp $t$ (publication year by default), an ordered list of paper co-authors $A = {a_1, a_2, \ldots}$, and an extra set of paper attributes $\Phi$. Each edge $e_{ij} = (v_i, v_j)$ is derived from a citation from paper $v_j$ to $v_i$.

To represent the main component of a scholar's citation-based research evolution, the concept of core GF graph is proposed, which is a subgraph $G^*$ of $G$ that best represents the research evolution of scholar $s$. The core GF graph profiling problem is decomposed into two sub-problems: node profiling and edge profiling.

Node profiling focuses on identifying the set of core papers $V^* \subseteq V$ authored by a scholar. A paper is considered a core paper if it meets either of the following conditions: (a) the scholar is the first author, or (b) the scholar is the supervisor of the first author. The contribution of the $k$-th author $a_k$ to a paper is quantified as:

$$ p_{\text{cont}}(a_k) = \begin{cases} 1, & \text{if } k = 1, \\ p_{AA}(a_k, a_1, t), & \text{if } k \neq 1, \end{cases} $$

Here, $p_{AA}(a_k, a_1, t)$ denotes the probability of $a_k$ being the advisor of the first author $a_1$ at time $t$. In practice, contributions of papers exceeding a predefined threshold (default: 0.5) are considered core papers. To detect the advisor-advisee relationship, an unsupervised, human-interpretable algorithm is introduced, which estimates the advisor-advisee probability between $a_k$ and $a_l$ by:

$$ p_{adr}(a_k, a_l, t) = \frac{N_{a_k}(0, t) - N_{a_k, a_l}(0, t)}{N_{a_k, a_l}(0, t)} \\ p_{ade}(a_k, a_l, t) = \max_{\substack{t_0 \leq t \leq t_1 \ t_1 - t_0 \geq S_0 \ \text{numerator} \geq S_{\text{adr}}}} \frac{\sum_{t_0 \leq t \leq t_1} \hat{N}_{a_k, a_l}(t)}{\hat{N}_{a_l}(t_0, t_1)} \\ p_{AA}(a_k, a_l, t) = \min(1.0, p_{adr}(a_k, a_l, t)) \times \min(1.0, p_{ade}(a_k, a_l, t)) $$

where $N_{a_k}(0, t)$ is the number of papers published by $a_k$ until time $t$, and $N_{a_k, a_l}(0, t)$ is the number of papers co-authored by $a_k$ and $a_l$ until time $t$, $\hat{N}{a_k, a_l}(t)$ is the number of major papers of a co-authored by $a_k$ and $a_l$ at time $t$, and $\hat{N}{a_l}(t_0, t_1)$ is the number of major papers by $a_l$ in the period ([t_0, t_1]). By default, a major paper of $a_l$ means he/she is the top-3 author. The use of the $\min$ function in $p_{AA}$ ensures that the supervising rate $p_{adr}(a_k, a_l, t)$ does not exceed 1.

Edge profiling involves detecting the set of core citation edges $E^* \subseteq E$ that represent the evolution of the scholar's scientific contribution. The approach focuses on classifying self-citation links according to an established taxonomy and then extracting the class of extend-type citations. To achieve this goal, a supervised learning algorithm is applied to a labeled dataset to classify citation types. Twenty features are selected and input to an Extra-Tree model. The model achieves an F1 score of 0.646 with 10-fold cross-validation. The features used in classification include metadata of cited and citing papers, citation network features, temporal correlation measures, and lexical patterns extracted from the citation context and text.

The graph modeling process mentioned above has been evaluated in the real-world task of scientific award inference, as demonstrated in [1], where the GF model outperforms existing indicator-based and network-based profiling methods. For instance, using a GNN model for award inference in the NLP-ARC domain, the F1 score exceeded predictions based on author-level indicators by 19.8%.

1.3 Sensitivity Analysis on Topic Thresholds

Figure 1: Sensitivity analysis of the topic-assignment threshold applied to two datasets (VCG: a, c; AI: b, d). (a-b) Mean number of topics per paper versus threshold; (c-d) Density of the N-th highest similarity score versus threshold.

We sincerely thank you for valuable suggestions to strengthen replicability. Complexity and runtime analysis are addressed in our response to Comment 4. To investigate the sensitivity of the topic-assignment mechanism, we conducted an analysis on two datasets: "Visualization and Graphics" (VCG: 271,418 papers, 105 topics) and "Artificial Intelligence at Large" (AI: 4,836,561 papers, 497 topics). The topic-assignment threshold was set within the range of $\theta \in$ [0.3, 0.8], centered around the default setting of $\theta$ = 0.6. In our work, the assignment rule operates as follows: each paper is assigned to topics where similarity exceeds $\theta$, with the constraint that when more than 3 topics satisfy this condition, only the top-3 most similar are assigned, and when no topics exceed $\theta$, the single most similar topic is assigned.

Figure 1(a-b) demonstrates how the constraint bounds the mean topics per paper between 1 and 3. The curve rises asymptotically toward 3 as $\theta$ decreases to 0.3 and falls toward 1 as $\theta$ increases to 0.8. The steepest transitions occur near $\theta$=0.61 (VCG) and $\theta$=0.63 (AI), indicating heightened sensitivity where threshold adjustments most substantially impact topic assignments.

Figure 1(c-d) reveals the underlying similarity distributions, showing the N-th highest similarity peaks at progressively lower $\theta$ values: N=1 ($\theta$=0.67), N=2 ($\theta$=0.63), N=3 ($\theta$=0.61), and N=5 ($\theta$=0.58) in VCG. We observe clear separation of the N=1 curve from others, indicating strong discriminative power for primary topics. The broad N=5 distribution (peak 0.58) signifies weak semantic associations, while the N=3 peak (0.61) closely aligns with our sensitivity threshold selection.

Our selection of $\theta = 0.6$ is specifically designed to minimize the number of assigned topics while preserving high relevance. This optimal balance is supported by two key observations: (1) The steepest transitions in the average number of topics per paper occur around $\theta \approx 0.6$ (Figure 1a–b), indicating that further increases in the threshold yield diminishing reductions in topic count. (2) The N=3 similarity peak around $\theta \approx 0.6$ (Figure 1c-d) suggests this threshold captures substantial semantic relevance. Although our initial case study employed $\theta = 0.55$, which was calibrated on older MAG data, our current production system utilizes MAG+OpenAlex data, where $\theta = 0.6$ provides more favorable tradeoffs between variability and relevance. We test the difference between $\theta = 0.55$ and $\theta = 0.6$ on the production system, and only find minor difference on the case studied in this paper. Therefore, we do not make big change to the case study section here.

The interface preserves all top-3 similarity scores, enabling users to dynamically adjust $\theta$ via a slider and instantly observe impacts on the topical landscape. Decreasing $\theta$ enhances diversity (increasing three-topic assignments), while increasing $\theta$ prioritizes relevance (favoring the top highly relevant topic). This reflects in real time the sensitivity relationships quantified in our analysis.

2 Integrated Flow Hierarchical Layout Algorithm

In GeneticPrism visualization, we utilize the hierarchical layout algorithm to visualize the prism's per-topic GF sub-graph on each lateral face. However, it overlooks the interactions between the central and related topics. To address this, we propose the Integrated Flow Hierarchical Layout (IFHL) for the GeneticScroll visualization, considering not only the layout of the central graph but also integrating its influence on other topics through the flow map and streamgraph. In contrast to alternative designs of GeneticScroll, which directly draw influence edges after drawing the topic GF sub-graph, our approach integrates the layout of the central graph and influence edges together. As shown in Figure 2(a) to Figure 2(d), three steps are adopted to distinguish the central graph from the newly added influences and effectively prevent visual clutter.

Figure 2: Steps from the original GF sub-graph layout to the GeneticScroll view. (a) Original per-topic GF sub-graph; (b) GF sub-graph with influence edges; (c) GF sub-graph with influence edges using bundling; (d) GeneticScroll view: add flow map layout for influence edges and place them on the background layer, integrating with influx/efflux streamgraphs

2.1 Objective

The objective of IFHL is to lay out the context edges integrated with the original hierarchical layout. To support the need for overlapping topics, we conduct GF graph partitioning for each topic. We construct a corresponding GF sub-graph for each topic, defined as:

$$ G(T_k) = {V(T_k), E(T_k)} \\ V(T_k) = { n \mid n \in V, T_k \in \text{topic}(n) } \\ E(T_k) = { e_{ij} \mid e_{ij} \in E, v_i \in V(T_k) \wedge v_j \in V(T_k) } $$

In this formula, (e_{ij}) is an extension edge (a reverse citation edge), (E) is the set of all extension edges in the author((s))'s collection, (v_i) and (v_j) are the source and target of edge (e_{ij}). The function (\text{topic}) maps a paper to its corresponding set of topics (with a maximum size of 3).

The set of context edges includes all edges that are connected directly to the topic $T_k$ and that are not within $T_k$. The in'' edges (influx), connect to the neighbor nodes upstream of the topic, while the out'' edges (efflux) connect to those downstream. To display the influx and efflux patterns, we place the in'' and out'' edges on the left and right sides, respectively. We introduce influx nodes $v_l(y)$ and efflux nodes $v_r(y)$, and $y$ represents the year associated with the influx/efflux node. These nodes act as proxies for external endpoints: When a context edge connects from an external node to a topic-internal node, the source is replaced by the corresponding influx node; when connecting from an internal node to external, the target is replaced by the efflux node. This replacement yields influence edges, defined as:

$$ E_I(T_k) = &{ e_{pj} \mid e_{ij} \in E, v_i \notin V(T_k) \wedge v_j \in V(T_k) \rightarrow v_p = v_l(y(v_i)) } \\ \cup &{ e_{ip} \mid e_{ij} \in E, v_i \in V(T_k) \wedge v_j \notin V(T_k) \rightarrow v_p = v_r(y(v_j)) } $$

Here, $V(T_k)$ is the set of central nodes, $E(T_k)$ is the set of central edges, $E_C(T_k)$ is the set of context edges, and $E_I(T_k)$ is the set of influence edges. The functions source and target map edges to their start and end points, respectively, and $\text{year}_l$ and $\text{year}r$ map nodes to the years of the influx/efflux nodes. To further abstract and reduce visual elements, we aggregate years by granularity. The influx nodes are mapped to the earliest year within the granularity, and the efflux nodes are mapped to the latest year (given by Equation below), ensuring a top-down extension. For example, for the year 2008, we create nodes $v_l(2008)$ and $v_r(2008)$. $v_l(2008)$ represents all nodes entering the central graph from 2008, and all influx influence edges $e{ij}$ with (\text{year}_l(v_k)=2008) originate from $v_l(2008)$. We compress context information using influx/efflux nodes into a single point, reducing visual elements.

$$ \text{year}_l(v_k) = \left\lfloor \frac{\text{year}(v_k)}{\text{grid}} \right\rfloor \cdot \text{grid} \\ \text{year}_r(v_k) = \left(\left\lfloor \frac{\text{year}(v_k)}{\text{grid}} \right\rfloor+1\right) \cdot \text{grid} - 1 $$

Table 2: Description of Nodes, Edges, and crossings.

Type	Type ID	Type	Description	Cost
Node	0	$v_c$	Central node
Node	1	$v_{cc}$	Dummy node between central nodes
Node	2	$v_p$	Proxy node, including $v_l, v_r$
Node	3	$v_{cp}$	Dummy node between influx/efflux node and central node
Edge	0	$e_c$	Edge between $v_c$, $v_{cc}$, a segment of central edges	$\alpha$
Edge	1	$e_p$	Edge concerning $v_p$ or $v_{cp}$, a segment of influence edges	weight($e_p$)
Crossing	0		All crossing	$\text{cost}(e_1) \cdot \text{cost}(e_2)$
Crossing	1		Weighted crossing	$\text{cost}(e_1) \cdot \text{cost}(e_2)$
Crossing	2	$(e_f, e_f)$	Internal crossing	$\text{cost}(e_1) \cdot \text{cost}(e_2)$
Crossing	3	$(e_p, e_p)$	External crossing	$\text{cost}(e_1) \cdot \text{cost}(e_2)$

2.2 Weighted Crossing

During the layout of the GF graph, we adhere to the rules that influx/efflux nodes are always positioned on the left and right sides, arranged by year. In detail, we enhanced the Hierarchical Layout provided by GraphViz (DOT) to lay out the whole graph with influence edges and influx/efflux nodes rather than the central graph. As shown in Table 2, we considered fine-grained node types during the layout process. Due to the addition of dummy nodes in the layout process, node types include central nodes $v_c$, influx/efflux nodes $v_p$, dummy nodes $v_{cc}$ between central nodes, and dummy nodes $v_{cp}$ between central and influx/efflux nodes. Edges are divided into segments of central edges $e_c$ and segments of influence edges $e_p$. During the node ordering step, we ensure that influx nodes $v_l$ and efflux nodes $v_r$ are positioned at the extreme left and right of the layout, respectively, as highlighted in Algorithm 1.

Algorithm 1: Node Ordering with Constraint of Proxy Nodes

Input: Graph g
Output: Best order best
order = init_order(g)
best = order
for i = 0 to Max_iterations do
    wmedian(order, i) // Calculate the median positions for each layer and sort nodes based on these positions, alternating the direction of each iteration
    transpose(order) // Swap adjacent nodes to reduce edge crossings layer by layer
    for r = 1 to Max_rank do
        adjust_order(order[r], v_l(r), v_r(r))  // Adjust order[r] to place v_l(r) at the start and v_r(r) at the end, shifting other nodes accordingly
    end for
    if crossing(order) < crossing(best) then
        best = order
    end if
end for
return best

The optimization metric is changed from edge crossing to weighted crossing. The weight of edges in the middle is set to $\alpha$, and the weight of influence edges is proportional to edge width. In the crossing function, we use $\text{cost}(e)$ instead of the default "1" as the edge weight. The weighted crossing between $e_1$ and $e_2$ is defined as $\text{cost}(e_1) \cdot \text{cost}(e_2)$. We use total weighted crossing as the optimization objective, as highlighted in Algorithm 2.

Algorithm 2: Weighted Crossing Count Calculation

Input: Order order
Output: crossing count count
count = 0
Count = {} // Count of connections to node i in the next layer
for r = 1 to Max_rank do
    for v in order[r] do
        for e in ND_out(v) do
            for k = order(aghead(e)) to max(keys(Count)) do
            count += Count[k] × cost_weighted(e)
            end for
        Count[order(aghead(e))] += cost_weighted(e)
        end for
    end for
end for
return count

To evaluate the optimal weight $\alpha$ for the best layout, we used crossing counts as indicators, including internal crossing (between central edges) and external crossing (between influence edges). We tested four optimization objectives: all crossing, internal crossing, external crossing, and weighted crossing, defined as:

$$ \text{cost}_\text{all}(e) = 1 \\ \text{cost}_\text{internal}(e) = \begin{cases} 1 & \text{if } e \in E(T_k) \\ 0 & \text{if } e \in E_I(T_k) \end{cases} \\ \text{cost}_\text{external}(e) = \begin{cases} 0 & \text{if } e \in E(T_k) \\ 1 & \text{if } e \in E_I(T_k) \end{cases} \\ \text{cost}_\text{weighted}(e) = \begin{cases} \alpha & \text{if } e \in E(T_k) \\ \text{weight}(e) & \text{if } e \in E_I(T_k) \end{cases} $$

Figure 3: Internal and external crossings under different optimization objectives, in an example graph of 50 nodes, are used for determining the optimal weight ($\alpha$) using the elbow method.

We aimed to find an optimal objective where internal crossing was low (primary goal) while maintaining low external crossing. Using the elbow method, we determined the optimal $\alpha$. For example, in a graph with 50 nodes, Figure 3 shows internal and external crossings under different optimization objectives. Using "all crossing" as the baseline resulted in low external crossing but high internal crossing. "Internal crossing" had the lowest internal crossing but high external crossing, while "external crossing" had the opposite effect. Weighted crossing achieved a good balance; with $\alpha$ over 6, internal crossing is relatively low, and further increasing $\alpha$ provided no benefit while the external crossing arises extensively. Thus, we selected $\alpha = 6$ as the optimal weighted crossing for this example.

To generalize, we selected $\alpha$ incrementally, stopping when the reduction in internal crossing did not outweigh the increase in external crossing. This algorithm calculates an adaptive $\alpha$, resulting in an enhanced topic GF graph with detailed influence information and minimal deviation from the original layout.

2.3 Citation Influence Flow Map

Algorithm 3: Flow Map Adjustment Algorithm

Input: Edge bundles contextEdges, layout of influence edges layout
Output: Adjusted flow map layout

pointTree ← buildPointTree(layout)
order ← topologicalSort(pointTree)

for each point p in reverse(order) do
    totalWidth ← sum of widths of all paths at p
    parentEdge ← findParentEdge(pointTree)
    parentEdge.width ← totalWidth
    normal ← computeNormal(parentEdge)
    sortedPaths ← sortPaths(p)  // Sort paths at p using weighted angles: 0.5·angle1 + 0.3·angle2 + 0.2·angle3
    point ← getInitialPosition(p, normal)

    for each path in sortedPaths do
        adjustStartPoint(path, point, normal)
        point ← updatePosition()  // For next path
    end for
end for

return adjusted layout

The flow map layout ensures that the edge widths increase progressively at intersection points, akin to a Sankey diagram. The primary goal is to prevent visual clutter and edge crossings for graphs with influence edges (Figure 2(b)), resulting in a clear and comprehensible visualization of complex network flows. To achieve the Influx/Efflux flow map, we first bundle the influence edges (Figure 2(c)), calculating their intersection points. Through modeling these intersection points, we adjust the layout by tweaking the positions, thicknesses, and order of the split edges, thereby achieving the effect of a Sankey graph (Figure 2(d)).

We handle edge bundling by merging adjacent dummy nodes into a single node. However, direct merging could confuse central edges with influence edges. Therefore, central edges and influence edges are treated separately. Not bundling central edges made the pattern more evident, while context information was less critical. Therefore, we achieved edge bundling by merging only influence edges ($v_{cc}$ or $v_{cp}$) to simplify visual elements and highlight the topic GF graph.

To adapt the flow map layout based on edge bundling results, a series of specific methodological steps are followed. The edge bundling process outputs control points defining Bézier curves. Notably, the control points at the extremities of these curves may coincide at shared locations, thereby defining branching points referred to as intersections. To render the flow map and generate a Sankey diagram utilizing these edge bundling control points, it becomes necessary to model the intersections accurately.

As shown in Algorithm 3, this is achieved by first constructing an intersection tree from the set of pre-arranged curves. Subsequently, a topological sort is performed on this tree to establish a hierarchical ordering of the intersections. This sort ensures a progressive increase in edge widths downstream from parent to child nodes, facilitating an organized flow representation. Finally, initiating the adjustment process from the downstream terminus of the topological sort, the curves are progressively refined moving upstream towards the source points.

At each intersection, the total width of all curves is calculated by summing the widths of all paths. The normal direction at the intersection is then determined using the curve of the parent intersection. Curves at each child intersection are sorted using weighted angles, calculated based on the tangent lines formed by the curve's closest three points to the intersection, with the closest point having higher weights because these Bézier curves are influenced more by the points closest to the intersection. Finally, the starting points of curves at each child intersection are adjusted based on the normal direction and the sorted order.

The flow map effectively visualizes complex network interactions by combining edge bundling and layout adjustment. Progressive edge widths and clear separation of line flows are achieved, avoiding visual clutter and crossings. This approach realizes a control-point-based Sankey diagram, providing an efficient method for visualizing intricate networks and their interactions.

2.4 Complexity and Efficiency Analysis

Figure 4: Scalability Analysis: Relationship between Graph Size and Runtime (log scale). (a) node count versus average time cost; (b) node count versus edge count, colored by runtime.

Our system employs a client-side computational architecture using WebAssembly to execute the core DOT layout algorithm directly within the user's web browser. This design significantly offloads server processing and enhances overall system scalability. Our server infrastructure (Django serving static resources) underwent rigorous stress testing, which confirmed its stable support for concurrent access by more than 1,000 users. Consequently, our service performance is primarily constrained by network bandwidth rather than server computation resources.

Regarding the algorithm's intrinsic complexity and scalability in the client side, our routing layout leverages the core Graphviz DOT (sugiyama-style) engine (implementation can be found at GitHub). While adhering to the theoretical complexity framework of DOT, which uses heuristics (like barycenter methods) to approximate complexity ranging from $O(|V||E|)$ to $O(|V|^3)$ depending on graph structure and iteration number, we introduced specific optimizations for our the scholarly context. The optimizations include incorporating context node constraints during the reordering phase, utilizing weighted crossing minimization instead of counting crossing numbers, and merging context dummy nodes during optimization steps. An additaional flow adjustment step in our implementation adds a complexity of $O(|E|)$. Thus, the overall layout complexity is dominated by the DOT step, falling within the $O(|V||E|)$ to $O(|V|^3)$ theoretically. Crucially, because our approach intelligently decomposes the large GF graph into single-topic GF subgraphs prior to layout, the typical subgraph size encountered is manageable. Our data indicates a 95th percentile node count of 134.7, along with a corresponding 95th percentile runtime of 177.4 ms. This indicates that interactive rendering (ten subgraph per scholar, approximately 1.77s in total) is generally feasible for the majority of individual scholars' GF graph visualizations on standard hardware.

To evaluate real-world user experience, we measured the performance of the WebAssembly-packaged ''viz.js'' module executing within the Chrome browser. Using the GF graph of the prominent AI researcher Yoshua Bengio (972 nodes, 2811 edges) as a representative large-scale scholar graph, five consecutive tests yielded an average end-to-end interactive delay of approximately 1.77 seconds, well within the acceptable threshold for perceived interactivity ($<$ 2 seconds). Crucially, peak memory consumption in the client side during this interaction remained a manageable 94 MB. These results affirm that the implementation effectively supports typical research scenarios involving graphs for individual scholars (below the 1000 node upper bound) while maintaining a fluid user experience.

In summary, our theoretical complexity analysis, combined with empirical runtime benchmarks in both native and deployed WebAssembly environments, provides robust evidence that our layout strategy offers good efficiency and scalability for the target scales of scholor's GF graphs. Achieving responsiveness on par with established systems like VisualBib [2], the client-side architecture effectively ensures real-time interaction at the individual user level even during high concurrent access, enabling the desired interactive exploration of scholarly graphs.

3 User Evaluation Questionnaire

This questionnaire is designed to gather your feedback on GeneticPrism, an advanced visualization system that helps explore the academic impact evolution of individual scholars. GeneticPrism is the 2.0 version of GeneticFlow, which serves as the baseline system for comparison in this study.

Your responses will help us evaluate the effectiveness and usability of these systems. By comparing GeneticPrism (v2) with GeneticFlow (v1), we aim to refine and improve the system's design and functionality. Your detailed feedback is incredibly valuable to us!

Please take your time to explore the systems and reflect on your experience. The evaluation includes 7 key questions, each including a rating for two systems (on a scale of five) followed by an open-ended question. It takes about 60 minutes to complete. For more context, refer to this user guide and case study, featuring 3 practical use cases to showcase the systems in action.

Q1. How useful did you find GeneticPrism and GeneticFlow in providing an overview of a scholar's scientific impact over time and across different topics? (shown in Figure 5)

Figure 5: Temporal and topical overview of a scholar's scientific impact (Q1). On the first hand, most users will need an overall understanding of a scholar's research impact, including the research contribution across topics and over time. The design should allow a macroscopic view with a time dimension that can juxtapose and compare the scholar's impact evolution on multiple topics. e.g. Andrew Y. Ng, left: GeneticPrism, right: GeneticFlow.

Open-ended question: In your opinion, how did the 3D prism metaphor in GeneticPrism help or hinder your understanding of a scholar's scientific evolution? How did GeneticFlow compare in this aspect?

Q2. How effectively did GeneticPrism and GeneticFlow help you understand the evolution of a single research topic of a scholar over time? (shown in Figure 6)

Figure 6: Detailed analysis of a scholar's impact evolution on a single research topic (Q2). For any specific topic, e.g., the one that the user is currently working on or the administrator is overseeing, s/he will need to drill down to details to complete their job. The low-level tasks include but are not limited to identifying key papers, understanding research threads/clusters, and predicting. e.g. Andrew Y. Ng (deep neural DNN topic), left: GeneticScroll of deep neural (detailed view of GeneticPrism), right: GeneticFlow with topic highlighted.

Open-ended question: How did the GeneticScroll visualization (in GeneticPrism) help clarify the evolution of a single topic? Was GeneticFlow as effective? What could be improved?

Q3. How effectively did GeneticPrism and GeneticFlow help you understand the interaction and influence patterns between multiple research topics of a scholar? (shown in Figure 7)

Figure 7: Discover influence and interaction patterns among multiple research topics of a scholar (Q3). For topic planning and research outlook, users will need to figure out the causal relationship among multiple research topics of a scholar. This can be achieved through the visual analysis of citation influence patterns among papers belonging to these topics. e.g. Pat Hanrahan (influence from topic hardware parallel to topic visualization), left: GeneticScroll of hardware parallel (detailed view of GeneticPrism), right: GeneticFlow with topic highlighted.

Open-ended question: How did the GeneticScroll and chord diagram in GeneticPrism help in visualizing the connections between different topics? Was GeneticFlow able to achieve a similar level of clarity?

Q4. How useful were GeneticPrism and GeneticFlow in helping you identify key interdisciplinary papers that had significant influence across topics or fields? (shown in Figure 8)

Figure 8: Identifying key interdisciplinary papers and their influence patterns (Q4). During topic transitions of a scholar, there usually is not a clear-cut pattern between old and new topics. It is crucial for our users to identify the key papers that conduct interdisciplinary research on these transitional topics. The context of this research can help explain the reasoning behind the process of scientific evolution. e.g. Pat Hanrahan (3D graphics topic), left: key interdisciplinary papers in GeneticScroll of 3D graphics (detailed view of GeneticPrism), right: same interdisciplinary papers in GeneticFlow.

Open-ended question: Did the six-degree impact metaphor in GeneticPrism help you identify interdisciplinary works more effectively? How did this compare to GeneticFlow in terms of revealing cross-disciplinary research?

Q5. How effective were the interactive features (e.g., navigating between topic facets, switching between prism and scroll, inspecting nodes and links) in GeneticPrism compared to GeneticFlow?

The 3D prism supports zoom and pan and can be rotated upon mouse drag to focus on topics more interesting to the current user. Another animation mode is also supported in that the prism will rotate around its center at a constant rate, presenting a dynamic overview for better analysis. Clicking on topic tag will drill down to the GeneticScroll visualization of the corresponding topic. For GeneticScroll, hovering an influx/efflux streamgraph layer highlights the part of the flow map representing the influx/efflux flows from/to the corresponding topic.

Open-ended question: Were there any specific interaction features in GeneticPrism that you found particularly engaging or frustrating? How does it compare to GeneticFlow?

Q6. How visually appealing and enjoyable did you find the GeneticPrism and GeneticFlow visualizations, including their layout and design?

Open-ended question: What aspects of the visual design in GeneticPrism stood out to you in comparison to GeneticFlow? Were there any parts that you found more enjoyable or easier to interpret? Were there any aspects of the visualization in GeneticPrism that made it easier or harder to understand the data?

Q7. How effectively did GeneticPrism and GeneticFlow model the research topics of each paper?

In GeneticFlow, each paper is assigned one topic, which can sometimes oversimplify the research if a paper covers multiple themes. On the other hand, GeneticPrism assigns multiple topics to each paper, selecting the top 3 topics based on cosine similarity and a relevance threshold. This allows for a more nuanced and accurate representation of a paper's research, especially when dealing with interdisciplinary or multifaceted works.

Open-ended question: Did the topic modelling method in GeneticPrism (which assigns multiple topics to a paper) more accurately reflect the research content compared to GeneticFlow (which assigns a single topic)? Did the multi-topic approach improve the representation of interdisciplinary or multi-faceted research?

References

[1] Y. Luo, L. Shi, M. Xu et al. Impact-oriented contextual scholar profiling using self-citation graphs. In KDD’ 23, pp. 4572–4583, 2023.

[2] A. Dattolo and M. Corbatto. Visualbib: A novel web app for supporting researchers in the creation, visualization and sharing of bibliographies. Knowledge-Based Systems, 182:104860, 2019.