Disclosing the interactive mechanism behind scientists’ topic selection behavior from the perspective of the productivity and the impact

doi:10.1016/j.joi.2023.101409

Journal of Informetrics

Volume 17, Issue 2, May 2023, 101409

https://doi.org/10.1016/j.joi.2023.101409 Get rights and content

Highlights

•
This paper proposes two correlation metrics, by which we verify that the productivity and the impact is related to the evolution of scientists’ research interests. This study may help researchers deeply understand scientists’ topic selection process.
•
This paper proposes a novel Q seashore walk model, which effectively imitates the interaction process between scientists and the environment. Our model contributes to the theoretical advancement of applying the reinforcement learning theory to analyze scientists’ behavior.
•
This paper shows that proper rewards effectively stimulate the performance of scientists, but excessive rewards inhibit their performance. This study may provide theoretical evidence for policy intervention in scientific research.

Abstract

The productivity and the impact are two most recognized aspects to evaluate the research performance of scientists. Figuring out whether and how these two factors shape the evolution of scientists’ research interests may facilitate researchers to go deep into scientists’ topic selection behavior. In this paper, we employ Microsoft Academic Graph as our data source, and propose two correlation metrics, by which over 20,000 scientists’ publication sequence from the computer science field are analyzed. We confirm that the productivity and the impact are related to the evolution of scientists’ research interests, and scientists tend to select topics which help them produce the productivity and the impact. To further explore how these two factors affects topic selection behavior, we propose a novel Q seashore walk model based on the interactive mechanism hypothesis. Our analysis results based on the simulation data are consistent with those based on the empirical data, which confirms the validity of our model and reports the evidence for the interactive mechanism. Based on the simulation data, we also analyze the role of reward for scientists’ research performance, and find that “too much is as bad as too little”. This research may help researchers deeply understand the process of topic selection, and provide a theoretical basis for research and development policy formulation.

Keywords

Scientometrics

Topic selection behavior

Random walk

Reinforcement learning

1. Introduction

It is particularly crucial to quantitatively figure out what factors affect scientists’ topic selection behavior and how they do so, because this not only affects how scientists are trained, and funded (Foster et al., 2015; Jia et al., 2017; Sinatra et al., 2016; Wang et al., 2013), but also implicitly reflects the process of collective discovery of new knowledge and how collective decisions shape the science (Yu et al., 2021; Zeng et al., 2019).

Previous studies have found various internal factors (e.g. gender, research interest, etc.) and external factors (e.g. topic value, topic novelty, funding, etc.) affecting topic selection behavior (Bu et al., 2018; Fadhly et al., 2018; Keshavarz & Shekari, 2020; Laudel, 2006; Li et al., 2017; Wei et al., 2013). However, few researches quantitively study whether and how the productivity and the impact acquired by studying a topic shape the evolution of scientists’ research interest. The interaction process between scientists and the external environment behind topic selection behavior is still unclear.

Jia et al. (2017) recently simulated the process of evolution of research interests, and proposed the seashore walk model, in which one of the basic assumptions is that scientists execute a random walk in topic selection. However, scientists actually may modify their research strategy when they receive the incentive from the external environment. As shown in Fig. 1, the environment includes other scientists, reviewers, journals, and etc. Scientists interact with the environment to improve the productivity and the impact, which are considered as critical evaluation factors of research performance evaluation of scientists in some countries (e.g., China). For example, in the process of a paper from submission to publication, authors need to communicate with reviewers, and apart from self-citation, an author can only accumulate citations when other scientists cite his/her papers. There are other evaluation factors as well. Under the conflict defined as “the essential tension” (Kuhn, 1977) and “publish or perish” (Qiu, 2010), scientists may shift their research focus to chase their goals. This study aims to examine whether the productivity and the impact acquired by studying a topic is related to the evolution of scientists’ research interest in a data-driven way, and disclose and simulate the interactive mechanism behind topic selection behavior.

In this paper, we selected over 20,000 scientists in the computer science field as the analysis objects, and employed “field of study” generated from Microsoft Academic Graph (MAG) (Sinha et al., 2015; Tang et al., 2008; Zhang et al., 2019) to represent topics. We proposed two correlation indicators, by which we disclose that the productivity and the impact shape the correlation in the evolution of research interest. To further explore how the two factors affect scientists’ topic selection behavior, we also proposed a variant of the seashore walk model (Jia et al., 2017), Q seashore walk model based on the Q-learning algorithm (Watkins, 1989; Watkins & Dayan, 1992), to simulate the process of topic choice based on the interactive mechanism hypothesis. To test the validity of our model, we compared the metrics distribution generated by the empirical data and that generated by the simulation data, and analyzed the role of rewards in scientists’ research performance.

The current study has the following theoretical and practical implications. We present that the productivity and the impact shape the correlation in the evolution of research interest of scientists, and used a simulation model to explain how this happens. Therefore, this study may help researchers deeply understand scientists’ topic selection behavior, and provides a theoretical basis for formulating research and development policy. The spirit of our model provides an interesting perspective to examine scientists’ decision-making process, and contributes to the theoretical advancement of applying the reinforcement learning theory to analyze scientists’ behavior.

The rest of this paper is organized as follows. Section 2 reviews the related work on scientists’ topic selection behavior and reinforcement learning theory. The problem definition and methodology are clarified in Section 3. The data set is clearly introduced in Section 4. In Section 5, we provide the experimental setup and analysis of experimental results. In Section 6, we offer the contributions and limitations of this study. Finally, we summarize the paper in Section 7.

2. Background

2.1. The factors affecting topic selection behavior

How scientists select their research topics is an area of continuous focus, and previous studies have found a variety of internal and external factors affecting this behavior. In this paper, we simply define external factors as the objective environmental factors (e.g., topic nature, funding, etc.), and internal factors as the influencing factors closely related to the scientists themselves (e.g., gender, research interest, etc.).

Many studies focused on analyzing external factors. For example, Wallace and Ràfols (2018) conducted analysis of avian influenza research by qualitative interviews and quantitative bibliometric approaches, and confirmed that institutional pressures and funding pressures have a powerful influence on the topic choice of researchers. Laudel (2006) also found that physicists adapt their research content to obtain funding. Hoonlor et al. (2013) also revealed that funding may raise interest in the supported disciplines. After interviewing three Indonesian scientists, Fadhly et al. (2018) identified 11 ways (e.g. research trends, available data, etc.) to select research topics, and concluded that scientists from different disciplines have their own approaches to select topics. Wei et al. (2013) and Li et al. (2017) found that scientists from physics, mathematics, economics, and biomedicine tend to follow large and hot topics, and Chinese scientists prefer to follow large topics. Lakeh and Ghaffarzadegan (2017) analyzed the abstracts of over 200,000 HIV/AIDS scientific papers, and found that where scientists live affect their research topics. Buehling (2021) aimed to examine whether journal rankings may influence scientists’ topic selection process. After checking the Handelsblatt Ranking (HBR) for economists, they confirmed that there is no significant relationship between the two.

Some studies explored to analyze internal factors, and disclose the inherent characteristics of topic selection behavior. Duch et al. (2012) showed that gender difference may result in the inequality in academic resource allocation, which drives topic selection behavior. Jia et al. (2017) found that the evolution of research interests follows an exponential distribution, which is determined by three fundamental factors (i.e., heterogeneity, recency, and subject proximity). Bu et al. (2018) analyzed the scientific careers of scientists from the computer science field, and found that the collaborators of high-impact scientists tend to study diverse research topics. Keshavarz and Shekari (2020) analyzed data collected through a questionnaire survey of 391 postgraduate students, and revealed four factors affecting topic selection (i.e. personal issues, topic nature, information resources, and research operability). In addition, Zeng et al. (2019) analyzed the co-citing network of papers of 3420 scientists, and showed that scientists nowadays move more frequently among topics than those in the past. Yu et al. (2021) performed an analysis of publication records from over 14,000 scientists in physics, and disclosed that the change of research interest may help scientists create papers with increased impact, but it is not associated with productivity. Huang et al. (2022b) recently proposed five novel research strategies under exploration and exploitation, and uncovered the relationship between scientists’ research performance (i.e., productivity and impact) and their preference for these strategies. However, they did not further explore how different incentives and/or rewards drive topic selection behavior.

In short, various internal factors and external factors affect scientists’ research topic selection process. However, previous studies generally neglected to explore whether and how the productivity and the impact affect the evolution of scientists’ research interests. The interaction mechanism between scientists and the environment in the process of topic selection is still understudied. In this study, we aim to answer the above questions.

2.2. Reinforcement learning theory

The fundamental goal of reinforcement learning is to find good policies, which can resolve sequential decisions problems for agents, by optimizing cumulative future rewards (Sutton & Barto, 2018). The Q- learning algorithm (Watkins, 1989; Watkins & Dayan, 1992) is one of the classical valued-based algorithms, which teaches the agents to learn how to act in the controlled Markovian environment, and the convergence of the Q-learning algorithm has been strictly proven. In recent years, deep learning algorithm and Q-learning algorithm have been effectively combined. Mnih et al. (2013) first combined the neural network with reinforcement learning and proposed the Deep Q-Network (DQN), which achieved excellent performance in seven Atari games. Van Hasselt et al. (2016) analyzed and tackled the DQN's overestimation of action value by presenting the Double Q-learning algorithm (DDQN), which achieves better performances in some Atari games. Wang et al. (2016) presented the dueling network, which separately estimates the state value function and the state-dependent action advantage function. Their model exceeds the state of the art in the Atari 2600 domain. Moreover, reinforcement learning has also been widely utilized in various fields. For example, AlphaGo Zero has achieved a superhuman performance in Go, and also defeated previous AlphaGo models (Silver et al., 2017). Duan et al. (2021) employed the Q-learning algorithm to solve the log anomaly detection task, and proposed the QLLog model. Ciranka et al. (2022) recently employed the reinforcement learning policy to imitate the human behavior in relational learning. They found that human learner used symmetric learning policy under full feedback contexts, but asymmetric learning policy under partial feedback contexts.

To the best of our knowledge, the interaction mechanism between scientists and the environment behind their topic selection behavior generally is relatively understudied, and therefore remains unclear. Inspired by Jia et al. (2017) and Ciranka et al. (2022)’s research, we proposed a novel Q seashore walk model based on the Q-learning algorithm, in which the process of topic selection is imitated. Our simulation model aims to help researchers better understand scientists’ topic selection behavior, and provide evidence for policy intervention in scientific research.

3. Methodology

3.1. Problem definition

The productivity and the impact are two widely employed aspects to evaluate the research performance of a scientist, and are generally quantified by number of publications and citation frequency (Huang et al., 2021, 2022a; Khosrowjerdi & Bornmann, 2021; Perianes-Rodriguez & Ruiz-Castillo, 2018; Sinatra et al., 2016; Zhu et al., 2021). Due to scientists’ pursuit of career advancement, the productivity and the impact may be seen as the incentive and/ or reward from external environment. Specifically, the authors should interact with the environment to publish papers and accumulate citations. For example, in the process of a paper from submission to publication, authors should consult with reviewers. Apart from self-citation, an author can only accumulate citations when other scientists cite his/her papers. Hence, there may exist the interaction mechanism between scientists and the environment, which drives scientists’ topic selection behavior, and topics that bring richer benefits may be more attractive to scientists.

This study aims to explore whether and how the productivity and the impact acquired by studying a topic shape scientists’ topic selection behavior in a data-driven way, and disclose and simulate the interactive mechanism composed of the productivity and the impact behind topic selection. In this paper, firstly, we propose TDC (Topic distribution correlation) and TFC (Topic frequency correlation) metrics to show that the productivity and the impact shape the correlation of topic distribution in scientists’ publication sequence. Second, we propose the Q seashore walk model to further verify and explain our results. Finally, we also conduct a series of robustness tests.

3.2. Topic distribution correlation metrics and topic frequency correlation metrics

To disclose the correlation among the evolution of scientists’ research interests from the perspective of the productivity and the impact, we proposed two correlation metrics. Before formally providing the formulas of the two metrics, we first introduce some definitions and notations used in this study.

For each scientist, $α$ , we sorted $α$ ’s publication sequence by the publication year of each paper (i.e., $d_{1, α}, d_{2, α}, \dots, d_{n, α}$ ), where the research content of each paper is represented by a list of topics (e.g., FoS in MAG dataset). We proposed two division approaches to split the publication sequence into adjacent windows. In the first division method, the publication sequence is divided by the time span, $Δ t$ , and papers published between $t$ and $t + Δ t$ are grouped into the same group. In the second division method, the publication sequence is divided by the number of publications, $Δ m$ , and consecutive $Δ m$ papers are grouped into the same group. To make full use of the publication sequence, we employed the sliding window to traverse the publication sequence from beginning to end. To clarify the two division methods, we provide a simple example, as shown in Fig. 2. $α$ has published $n$ papers from 2000 to 2020. $Δ t$ and $Δ m$ are set to 2 and 3, respectively. Thus, in the first division method, papers published in 2000 and 2001(i.e., $d_{1, α}, d_{2, α}$ ) constitute the first window, and $d_{2, α}, d_{3, α}, d_{4, α}$ constitute the second window, and so on. In the second division method, the first three papers (i.e., $d_{1, α}, d_{2, α}, d_{3, α}$ ) constitute the first window, and $d_{2, α}, d_{3, α}, d_{4, α}$ also are grouped into the second window, and so on.

Subsequently, we introduce the topic distribution correlation metric (TDC) and topic frequency correlation metrics (TFC). Specifically, topic vector (Jia et al., 2017) is employed in TDC and TFC. To evaluate the level of importance of topics from the perspective of the productivity and the impact, we expand the definition of topic vector. We respectively employed the frequency of each topic adopted by $α$ and the citation count acquired by the papers studying the topic authored by $α$ in each window ( $i \in I$ ) to count each topic. The topic vectors calculated by the usage frequency and citation frequency in $i$ are denoted as $g_{i, f}$ and $g_{i, c}$ , respectively, and are a K-dimensional vector. $K$ is the number of unique topics in $α$ ’s publication sequence. For example, as shown in the Fig. 3, in the first window, $g_{1, f}$ is $(2, 1, 1, 0, 0)$ , because “Java” is adopted twice in two papers, and “Machine learning” and “Computer vision” are respectively adopted once in one paper. $g_{1, c}$ is $(1, 1, 0, 0, 0)$ , because the papers adopting “Java” and “Machine learning” are cited only once in a citation window, $Δ t^{'}$ .

Because we focus on the difference of the importance ranking of topics in the topic vectors, the Spearman correlation coefficient is adopted. Specifically, in the first division method, we calculated the Spearman correlation coefficient between $g_{i, f}$ and $g_{i + Δ t, f}$ , $S p e a r m a n r (g_{i, f}, g_{i + Δ t, f})$ . The gap, $Δ t$ , avoids the overlapping of two paper sets from two windows. Similarly, we also calculated $S p e a r m a n r (g_{i, f}, g_{i + Δ m, f})$ in the second division method. The positive correlation value suggests that scientists prefer to continue to study the topic that previously generated more papers. However, the zero and negative values mean that productivity have nothing to do with the scientists’ topic selection behavior. Subsequently, we calculated $S p e a r m a n r (g_{i, c}, g_{i + Δ t, f})$ and $S p e a r m a n r (g_{i, c}, g_{i + Δ m, f})$ .The positive value suggests that scientists tend to choose the topic that previously had a greater impact, and the zero and negative values mean that the impact is a negligible factor for the scientists’ topic selection behavior. Notably, to consider the whole career life span of a scientist, the Spearman correlation coefficient between all pairs of windows is calculated, and these coefficients with significance level above a certain level, $p_{v}$ , are summed and divided by the total number of pairs of windows, ( $I (Δ t) - Δ t$ and $I (Δ m) - Δ m$ ). Finally, the formulas of $T D C (P r o d u c t i v i t y (I m p a c t), Δ t (Δ m))$ can be found in Eqs. (1)-(4).(1) $T D C (P r o d u c t i v i t y, Δ t) = \frac{1}{I (Δ t) - Δ t} \sum_{i}^{I (Δ t) - Δ t} I (p < p_{v}) S p e a r m a n r (g_{i, f}, g_{i + Δ t, f})$ (2) $T D C (I m p a c t, Δ t) = \frac{1}{I (Δ t) - Δ t} \sum_{i}^{I (Δ t) - Δ t} I (p < p_{v}) S p e a r m a n r (g_{i, c}, g_{i + Δ t, f})$ (3) $T D C (P r o d u c t i v i t y, Δ m) = \frac{1}{I (Δ m) - Δ m} \sum_{i}^{I (Δ m) - Δ m} I (p < p_{v}) S p e a r m a n r (g_{i, f}, g_{i + Δ m, f})$ (4) $T D C (I m p a c t, Δ m) = \frac{1}{I (Δ m) - Δ m} \sum_{i}^{I (Δ m) - Δ m} I (p < p_{v}) S p e a r m a n r (g_{i, c}, g_{i + Δ m, f})$

Unlike TDC, which measures the relative importance of different topics, TFC aims to gage the level of involvement of a single topic. Specifically, for $k$ th topic, we count its usage frequency sequence and citation frequency sequence in $I$ windows, denoted as $h_{1, f}^{k}, h_{2, f}^{k} \dots h_{I, f}^{k}$ , and $h_{1, c}^{k}, h_{2, c}^{k} \dots h_{I, c}^{k}$ . For example, in Fig. 3, $h_{1 : I, f}^{J a v a}$ is $(2, 1, . . ., 2)$ . Then, in the first division method, we calculate the Pearson correlation coefficient between $h_{1 : I - Δ t, f}^{k}$ and $h_{1 + Δ t : I, f}^{k}$ , $P e a r s o n r (h_{1 : I (Δ t) - Δ t, f}^{k}, h_{1 + Δ t : I (Δ t), f}^{k})$ . Notably, the gap, $Δ t$ , avoids the overlapping of two paper sets from two adjacent windows. Similarly, we calculated $P e a r s o n r (h_{1 : I (Δ m) - Δ m, f}^{k}, h_{1 + Δ m : I (Δ m), f}^{k})$ in the second division method. The positive value suggests that scientists continuously select the topic that previously generated more productivity. We also calculate $P e a r s o n r (h_{1 : I (Δ t) - Δ t, c}^{k}, h_{1 + Δ t : I (Δ t), f}^{k})$ and $P e a r s o n r (h_{1 : I (Δ m) - Δ m, c}^{k}, h_{1 + Δ m : I (Δ m), f}^{k})$ . The positive value means that scientists continuously study the topic that previously generated more impact. To consider all topics adopted by a scientist, for each topic, the Pearson correlation coefficient is calculated, and these coefficients with significance level above a certain level, $p_{v}$ , are summed and divided by the total number of topics, $K$ . The formulas of $T F C (P r o d u c t i v i t y (i m p a c t), Δ t (Δ m))$ can be found in Eqs. (5)–(8). The notations used in TDC and TFC and its definitions can be found in Table 1.(5) $T F C (P r o d u c t i v i t y, Δ t) = \frac{1}{K} \sum_{k}^{K} I (p < p_{v}) P e a r s o n r (h_{1 : I (Δ t) - Δ t, f}^{k}, h_{1 + Δ t : I (Δ t), f}^{k})$ (6) $T F C (I m p a c t, Δ t) = \frac{1}{K} \sum_{k}^{K} I (p < p_{v}) P e a r s o n r (h_{1 : I (Δ t) - Δ t, c}^{k}, h_{1 + Δ t : I (Δ t), f}^{k})$ (7) $T F C (P r o d u c t i v i t y, Δ m) = \frac{1}{K} \sum_{k}^{K} I (p < p_{v}) P e a r s o n r (h_{1 : I (Δ m) - Δ m, f}^{k}, h_{1 + Δ m : I (Δ m), f}^{k})$ (8) $T F C (I m p a c t, Δ m) = \frac{1}{K} \sum_{k}^{K} I (p < p_{v}) P e a r s o n r (h_{1 : I (Δ m) - Δ m, c}^{k}, h_{1 + Δ m : I (Δ m), f}^{k})$

Table 1. Notations in the TDC and TFC metrics.

Notation	Definition
$Δ t$	The length of time span required in the first division method
$Δ m$	The number of publications required in the second division method
$I (Δ t)$	The total number of windows in the first division method
$I (Δ m)$	The total number of windows in the second division method
$Δ t^{'}$	The length of citation window
$i$	The index of $i$ th window
$g_{i, f}$	Topic vector calculated based on the productivity in $i$
$g_{i, c}$	Topic vector calculated based on the impact in $i$
$h_{i, f}$	Topic frequency calculated based on the productivity in $i$
$h_{i, c}$	Topic frequency calculated based on the impact in $i$
$h_{1 : I, f}$	$h_{1, f}, h_{2, f} \dots h_{I, f}$
$h_{1 : I, c}$	$h_{1, c}, h_{2, c} \dots h_{I, c}$
$p_{v}$	Significant level
$k$	The index of $k$ th topic
$K$	The total number of unique topics in a publication sequence

3.3. Q seashore walk model

In this subsection, we aim to explore how the productivity and the impact affect the scientists’ topic selection behavior. We employed a simulation model to imitate topic selection behavior based on the interactive mechanism hypothesis, in which the productivity and the impact are regarded as rewards.

Inspired by Isaac Newton's retrospection that his scientific career had been like “a boy playing on the seashore… finding… a prettier shell than ordinary” (Jia et al., 2017). Mandelbrote (2001) proposed the seashore walk model, which models the evolution of a scientist's research interests. However, the seashore walk model simply assumes that scientists perform an unbiased random walk on the 1-D lattice, which means that they follow a random exploration policy. However, we argued that scientists might chase goals, and shift their research focus according to the benefits gained by their previous strategies. Therefore, we proposed the Q seashore walk model, which aim to further figure out the role of the productivity and the impact in the process of selecting topics.

Before introducing our model, we simply review the seashore walk model. A scientist, $α$ , starts from a random position, and performs an unbiased random walk ( $p (a_{l e f t}) = p (a_{r i g h t}) = 0.5$ ) on a seashore, which is a 1-D lattice with piles of shells located on some sites, as shown in the Fig. 4. Different-colored shapes represent shells with different combinations of topics. The walker picks a shell at one site, corresponding to publishing a paper. The probability that shells exist on a site is $p$ . The number of shells at the site, $q$ , follows a power law distribution, $P (q) \sim {(1 + q)}^{- λ}; q \leq q_{m a x}$ , if there are any shells. $q_{m a x}$ is the maximum number of shells in a site. The total number of steps of a walker, $S$ , followed a truncated log-normal distribution, $l o g N o r m a l (S_{μ}, S_{σ})$ . The total length of the seashore is $L$ sites, and the length of each topic tool is $L_{j}$ sites. Each topic pool has three distinct topics, and two neighboring topic pools (e.g., $L_{j}$ and $L_{j + 1}$ ) vary by one topic. The topics of papers from $L_{j}$ are the random combination of topics from the topic tool.

In the Q seashore walk model, instead of assuming an random strategy, we employed a classical and effective reinforcement learning algorithm, the Q-learning algorithm (Watkins, 1989; Watkins & Dayan, 1992), to quantify the decision-making policy adopted by $α$ . Specifically, our model is established according to the following ideas: $α$ tends to stay at the topic pool that brings them rich rewards, and depart from the topic pool with poor benefits. In short, if $α$ picks one shell, $α$ would prefer to stay at the current position, otherwise $α$ try to leave. Specifically, the scientist is the agent, $α$ , and the seashore is the environment, $ε$ . The action space, $A$ , is composed of moving one step left or right, denoted as $a_{l e f t}$ and $a_{r i g h t}$ . The state space, $S$ , is composed of the left side of the topic pool and the right side of the topic pool, denoted as $s_{l e f t}$ and $s_{r i g h t}$ . The rewards gained by $α$ depend on three factors (i.e., action, state, and shell), as shown in Table 2. For example, as shown in Fig. 5, when $α$ at state, $s_{l e f t}$ , takes the action $a_{l e f t}$ and then picks a shell, he will get a negative reward, $- r_{1}$ , which suggests that leaving the topic pool is a wrong decision. Moreover, to consider the productive differences among topics, we divide the topic pool into high-yield topic pool and general topic pool with the probability, $p_{h i g h}$ and $1 - p_{h i g h}$ . The difference between the two is that the former has a larger $q_{m a x} (h i g h) = b q_{m a x}$ . To model the impact of simulated papers, we employed a log-normal distribution, $P (C) \sim l o g N o r m a l (C_{μ}, C_{σ})$ to imitate the citation distribution of simulated papers.

Table 2. Rewards in the Q seashore walk model.

Empty Cell	$s_{l e f t}$		$s_{r i g h t}$
Any shells	$a_{l e f t}$	$a_{r i g h t}$	$a_{l e f t}$	$a_{r i g h t}$
Yes	$- r_{1}$	$r_{1}$	$r_{1}$	$- r_{1}$
No	$r_{2}$	$- r_{2}$	$- r_{2}$	$r_{2}$

Finally, we introduced the definition of Q value, $Q (s, a)$ , which gauges the long-term benefits of taking an action at a state, and which can be estimated iteratively by Bellman equation, as shown in Eq. (9). $α$ takes an action by $a r g m a x_{a} Q (s, a)$ with the probability, $1 - p_{e x p l o r e}$ , and executes an unbiased random walker with the probability, $p_{e x p l o r e}$ . For example, as shown in the upper of Fig. 5, due to $Q (s_{l e f t}, a_{l e f t}) > Q (s_{l e f t}, a_{r i g h t})$ , the agent takes $a_{l e f t}$ . When $Q (s, a_{l e f t}) = Q (s, a_{r i g h t})$ , we used the random strategy. Thus, the Q seashore walk model degenerates into the seashore walk model, when the random strategy is always adopted. In our model, there are only four combinations of the action and state, and therefore $Q (s, a)$ only takes four values. In each $L_{j}$ , the $Q_{j} (s, a)$ , are employed to manage the policy of $α$ , andare initialized with zero values. We also introduced three rules to update $Q_{j} (s, a)$ . (1). when $α$ does not leave $L_{j}$ , and $s$ remains unchanged (e.g., $s_{t} = s_{l e f t}, s_{t + 1} = s_{l e f t}$ ), Eq. (9) is utilized to update $Q_{j} (s, a)$ . (2). when $α$ does not leave $L_{j}$ , but $s$ changes (e.g., $s_{t} = s_{l e f t}, s_{t + 1} = s_{r i g h t}$ ), we update $Q_{j} (s_{t}, a_{l e f t})$ and $Q_{j} (s_{t}, a_{r i g h t})$ to zero values. This is due to the fact that $α$ has been at the middle of the topic pool, and both actions seem to be good decisions. (3). when $α$ leaves $L_{j}$ (e.g., $α$ arrives at $L_{j + 1}$ ), Eq. (9) is utilized to update $Q_{j} (s_{t}, a_{t})$ , Eq. (10) is employed to update $Q_{j + 1} (s_{t + 1}, a_{t})$ , and then $Q_{j} (s_{t}, a_{l e f t})$ and $Q_{j} (s_{t}, a_{r i g h t})$ are updated to zero value. $1 - a_{t}$ indicates the opposite direction of $a_{t}$ . Actually, Eq. (10) indicates the intensity of $α$ 's desire to leave $L_{j}$ . The notations used in the Q seashore walk model can be found in Table 3.(9) $Q_{j} (s_{t}, a_{t}) = r + γ Q_{j} (s_{t}, a_{t})$ (10) $Q_{j + 1} (s_{t + 1}, a_{t}) = Q_{j} (s_{t}, a_{t}) - Q_{j} (s_{t}, 1 - a_{t})$

Table 3. Notations in the Q seashore walk model.

Notation	Definition
$α$	The agent (walker, scientist)
$A$	Action space ( ${a_{l e f t}, a_{r i g h t}}$ )
$S$	State space ( ${s_{l e f t}, s_{r i g h t}}$ )
$ε$	Environment (seashore)
$Q (s, a)$	The long-term benefits of taking an action at a state
$r_{1}$	The reward when a shell has been found
$r_{2}$	The reward when no shell is found
$γ$	The discount parameter in the Bellman equation
$p$	The probability that a site contains any shells
$q$	The number of shells at the site ( $P (q) \sim {(1 + q)}^{- λ}$ )
$q_{m a x}$	The maximum number of shells in a site
$p_{h i g h}$	The probability that the topic pool is high-yield
$b$	$q_{m a x} (h i g h) = b q_{m a x}$
$S$	The total number of steps of a walker ( $P (S) \sim l o g n o r m a l (S_{μ}, S_{σ})$ )
C	Citation distribution of simulated papers ( $P (C) \sim l o g N o r m a l (C_{μ}, C_{σ})$ )
$S_{m a x}$	The maximum number of steps for a walker
$p_{e x p l o r e}$	The probability that the walker adopts an exploration action
$L_{j}$	The $j t h$ topic pool (The length of a topic pool)
$L$	The length of all topic pools ( $\sum_{j} L_{j}$ )
$N_{1}$	The number of 1-D lattices
$N_{2}$	The number of scientists sampled in each 1-D lattice

3.4. Data collection and preprocessing

In this paper, we utilized Microsoft Academic Graph (MAG) data from Open Academic Graph (OAG) 2.1 (Sinha et al., 2015; Tang et al., 2008; Zhang et al., 2019) as our data source. MAG comprise over 200 million papers, and is one of the most widely used data sets in bibliometrics (De Domenico et al., 2016; Jin et al., 2021). In MAG, there are totally 228,251 “field of study” (FoS), which are nested within 293 sub-disciplines, which are nested within 19 disciplines (Shen et al., 2018). Each paper collected from MAG has been associated with a list of FoS (i.e. topics).

Since we are familiar with the computer science field (CS), we extracted papers from this field by the FoS keyword, “computer science”, as our data set (hereafter, $M A G_{C S}$ ). There are 23,471,718 papers, and 29,073,597 authors who have published at least one paper in $M A G_{C S}$ .To analyze the role of the productivity and the impact in the evolution of scientists’ research interests, we considered scientists who have published at least 100 papers in the CS field. This is because scientists who have published fewer papers cannot provide sufficient data for our analysis, and introduce randomness into the results. Finally, 21,239 scientists constitute our analysis object. Our dataset will be described in detail in the following “Descriptive statistics” subsection.

4. Experiments and results

4.1. Descriptive statistics

The distribution characteristics of $M A G_{C S}$ is shown in Fig. 6. The annual number of publications in the CS field experienced exponential growth from 1950 to 2019, as shown in Fig. 6(a). The number of new authors in the CS field per year also experienced substantial growth, as shown in Fig. 6(b). However, the majority of authors (77.09%) only published one paper, and the proportion of authors with less than four papers is over 90%. As we mentioned previously, only 21,239 authors with over 100 papers ( $t h r e s h o l d_{1}$ ) are selected as our analysis object. Fig. 6(c) presents the annual number of new topics, which increased rapidly from 1950 to 2010 and then slowed down from 2011 to 2020, which means that the CS field has gradually become mature, but it is still developing. Notably, the lower number of papers, authors, and topics in 2020 is due to the truncation of data collection. The number of publications in each topic can be found in Fig. 6(d). The only topic with over 3 million papers is “artificial intelligence”, and the majority of topics have less than 100,000 papers. Actually, many topics have only been selected a few times during the scientific career of $α$ . Hence, topics adopted by $α$ less than 10 times ( $t h r e s h o l d_{2}$ ) are filtered out, when we calculated the TDC and TFC. In the “Robustness tests” subsection, we also repeated our experiments with different value of $t h r e s h o l d_{1}$ and $t h r e s h o l d_{2}$ .

4.2. Correlation analysis based on TDC and TFC metrics

We employed TDC and TFC metrics to show that the productivity and the impact shape the correlation in the evolution of research interest. To analyze the magnitude of the correlation coefficient, we built a control group by shuffling scientists’ publication sequence. Specifically, we randomly sort the order of a scientist's papers on the premise that annual number of papers in the reshuffled sequence is the same as that in the original sequence. In the following experiments, $Δ t$ and $Δ m$ in two division methods are set to 1 and 8, respectively. $p_{v}$ is set to 0.1. TDC and TFC distributions of over 20,000 scientists based on the empirical data and the randomized data are shown in Figs. 7 and 8.

As shown in Fig. 7(a) and (b), $T D C (P r o d u c t i v i t y, Δ t (Δ m))$ distribution based on the empirical data (the red bar) are bell-shaped distributions, and their mean values are greater than zero (0.135 and 0.166), which means that scientists prefer to continuously study topics that bring more productivity. We also found that $T D C (P r o d u c t i v i t y, Δ t (Δ m))$ distribution based on the shuffled data (the gray bar) obviously shift to the negative half of the x-axis, and their mean values are smaller than those based on the empirical data (0.067 and 0.111). Moreover, as shown in Fig. 7(c) and (d), $T F C (P r o d u c t i v i t y, Δ t (Δ m))$ distribution based on the empirical data also are bell-shaped distributions with mean values greater than zero (0.133 and 0.105), which also means that scientists prefer to study the topics that help them produce papers. Not surprisingly, $T F C$ distribution based on the shuffled data also obviously shifts to the negative half of the x-axis, and their mean values are 0.078 and −0.080. Notably, in Fig. 7, TDC and TFC distributions based on the empirical data and those based on the randomized data are significantly different under the KS test with $p v a l u e < 1 e^{- 3}$ .

Subsequently, we used the citation count acquired by a paper in following 3 years to count the topic vector. As shown in Fig. 8(a) and (b), we found that $T D C (I m p a c t, Δ t (Δ m))$ distribution based on empirical data presents the bell-shaped distribution, and are at the right side of $T D C (I m p a c t, Δ t (Δ m))$ distribution based on shuffled data. As shown in Fig. 8(c) and (d), we get the similar results on $T F C (I m p a c t, Δ t (Δ m))$ . In addition, $TDC and TFC$ distribution based on empirical data and its control group are also significantly different under the KS test with $p v a l u e < 1 e^{- 3}$ . Thus, scientists tend to continuously study the topic that produce the greater impact.

4.3. Robustness tests

To test the robustness of our results on empirical data, we respectively filtered out the high-frequency topics, changed the values for $Δ t$ and $Δ m$ , changed the values for $t h r e s h o l d_{1}$ and $t h r e s h o l d_{2}$ , and did not use the sliding window method, and repeated the above experiments.

Specifically, (1). considering that some topics are adopted frequently by a scientist and the control group may keep the order of these high-frequency topics, we simply filter out the topics studied by more than 60% of the total number of papers published by a scientist, and repeated the experiments. (2). We change the values for $Δ t$ and $Δ m$ ( $Δ t = 2$ and $Δ m = 12$ ), and repeat our experiments. (3). We change the values for $t h r e s h o l d_{1} = 60$ , employ authors who have published from 60 to 100 papers as our analysis objects, and repeat our experiments. (4). We change the values for $t h r e s h o l d_{2} = 5$ , allow more low frequency topics, and repeat our experiments. (5). We do not use the sliding window method in two division methods, and repeat our experiments. For example, in the second division method, when $Δ m = 3$ , we have $(d_{1}, d_{2}, d_{3}), (d_{4}, d_{5}, d_{6}), (d_{n - 2}, d_{n - 1}, d_{n})$ , instead of $(d_{1}, d_{2}, d_{3}), (d_{2}, d_{3}, d_{4}), (d_{n - 2}, d_{n - 1}, d_{n})$ .

In the above robustness tests, our results remain unchanged. Specifically, the difference between the average value of TDC (TFC) distribution based on the empirical data and that based on the randomized data is always positive, which means that the random distribution shifts to the negative half of the x-axis. The results are listed in Table 4. Moreover, in all the aforementioned experimental results, TDC (TFC) kernel density based on empirical data and its control group are significantly different under the KS test with $p v a l u e < 1 e^{- 3}$ .

Table 4. $μ (e m p i r i c a l d a t a) - μ (r a n d o m i z e d d a t a)$ in robustness tests.

Robustness tests	(1)	(2)	(3)	(4)	(5)
$T D C (P r o d u c t i v i t y, Δ t)$	0.072^⁎⁎⁎	0.034^⁎⁎⁎	0.037^⁎⁎⁎	0.086^⁎⁎⁎	0.069^⁎⁎⁎
$T D C (P r o d u c t i v i t y, Δ m)$	0.058^⁎⁎⁎	0.047^⁎⁎⁎	0.018^⁎⁎⁎	0.067^⁎⁎⁎	0.053^⁎⁎⁎
$T D C (I m p a c t, Δ t)$	0.069^⁎⁎⁎	0.055^⁎⁎⁎	0.051^⁎⁎⁎	0.074^⁎⁎⁎	0.060^⁎⁎⁎
$T D C (I m p a c t, Δ m)$	0.068^⁎⁎⁎	0.078^⁎⁎⁎	0.045^⁎⁎⁎	0.076^⁎⁎⁎	0.072^⁎⁎⁎
$T F C (P r o d u c t i v i t y, Δ t)$	0.057^⁎⁎⁎	0.041^⁎⁎⁎	0.038^⁎⁎⁎	0.061^⁎⁎⁎	0.045^⁎⁎⁎
$T F C (P r o d u c t i v i t y, Δ m)$	0.186^⁎⁎⁎	0.203^⁎⁎⁎	0.168^⁎⁎⁎	0.142^⁎⁎⁎	0.131^⁎⁎⁎
$T F C (I m p a c t, Δ t)$	0.048^⁎⁎⁎	0.052^⁎⁎⁎	0.032^⁎⁎⁎	0.052^⁎⁎⁎	0.032^⁎⁎⁎
$T F C (I m p a c t, Δ m)$	0.114^⁎⁎⁎	0.127^⁎⁎⁎	0.094^⁎⁎⁎	0.096^⁎⁎⁎	0.078^⁎⁎⁎

Note: *** indicates pvalue< 0.001.

4.4. Modeling the process of the topic selection by the Q seashore walk model

In this subsection, we employed the Q seashore walk model to imitate the process of selecting topics. The simulation experiment aims to further explore how the productivity and the impact affect the evolution of scientists’ research interests, and help us deeply comprehend the role of rewards for scientists in scientific research.

Specifically, after filtering out topics adopted by a scientist less than 10 times in his publication sequence, each paper from 21,239 scientists covers an average of 3.21 topics. Thus, the number of unique topics in each topic pool is set to three. To reproduce the real TDC and TFC distribution, we set different parameter settings of the Q seashore model, as shown in Table 5. First, to reproduce the real TDC distribution, we adopted the set of parameters in the first column of Table 5. There are totally 3 million pieces of simulated publication sequence, in which 8339 walkers have more than 100 shells. Since there is no publication year in the simulated data, we only calculated the $T D C (P r o d u c t i v i t y (I m p a c t), Δ m)$ in the simulation 1 data. $Δ m$ is also set to 8. As shown in Fig. 9, the real TDC distribution and the simulated one are almost overlapping, and have the similar mean and standard deviation. Second, to reproduce the real TFC distribution, we adopted the set of parameters in the second column of Table 5. There are totally 15 million pieces of simulated publication sequence, in which only 4880 walkers have more than 100 shells. We calculated the $T F C (P r o d u c t i v i t y (I m p a c t), Δ m)$ in the simulation 2 data. As shown in Fig. 10, the real TFC distribution and the simulated one are broadly similar, and have the similar mean and standard deviation. Notably, the difference between the kernel density of the real distribution and that of the simulated distribution is not significant under the KS test with $p v a l u e > 5 e^{- 2}$ , and the $p v a l u e$ can be found in Figs. 9 and 10. Considering that simulating scientists’ behavior is a complex and challenging task, we argue that the Q seashore walk model has achieved satisfactory simulation results. Thus, our model effectively imitates the scientists’ topic selection process based on the interactive mechanism. The interactive mechanism is a reasonable explanation of how the productivity and the impact affect the scientists’ topic selection behavior. Specifically, the scientist may perceive the productivity and the impact as a kind of rewards, and then adjust their current topic selection policy for their goals, as shown in Fig. 1.

Table 5. Parameter settings in the Q seashore walk model.

Notations	Simulation 1	Simulation 2	Simulation 3
$r_{1}$	5.0	5.0	[−10,25]
$r_{2}$	1.0	1.0	1.0
$γ$	1.0	1.0	1.0
$p$	0.15	0.035	0.035
$q$	$P (q) \sim {(1 + q)}^{- 1}$	$P (q) \sim {(1 + q)}^{- 1}$	$P (q) \sim {(1 + q)}^{- 1}$
$q_{m a x}$	10	10	10
$p_{h i g h}$	0.15	0.15	0.0
$b$	10	10	10
$S$	$l o g n o r m a l (0.5, 2.5)$	$l o g n o r m a l (0.5, 2.5)$	$l o g n o r m a l (0.5, 2.5)$
C	$l o g n o r m a l (0.25, 1)$	$l o g n o r m a l (0.25, 1)$	$l o g n o r m a l (0.25, 1)$
$S_{m a x}$	5000	5000	20,000
$p_{e x p l o r e}$	0.1	0.0	0.0
$L_{j}$	20	20	50
$L$	3000	3000	5000
$N_{1}$	300	300	150
$N_{2}$	10,000	50,000	50,000

To figure out how the magnitude of rewards affect scientists’ topic selection behavior and then affect their research performance, we repeated our experiments under the set of parameters in the third column of Table 5, in which the value of $r_{1}$ ranges from −10 to 25 with the interval equal to 1. As shown in Fig. 11(a) and (b), the x-axis indicates $r_{1}$ , and the y-axis represents the average accumulative number of papers and the average accumulative number of citations in all simulated scientists. With the increase of $r_{1}$ , the average number of papers per simulated scientist rises first and then decreases (red dot), and the trend can be perfectly fitted by a quadratic function (black line). Since there exist a linear correlation between the cumulative number of papers of a scientist and the cumulative number of citations of the scientist, Fig. 11(b) shows the same trend as Fig. 11(a). To be specific, when the $r_{1}$ gradually becomes positive, the average number of papers (citations) increase, which is in line with our common sense that proper rewards promote the scientists’ research performance. However, when $r_{1}$ is too large, rewards have a negative effect on the productivity and the impact of scientists, which is also line with our common sense that excessive rewards make researchers become satisfied with the status quo. We summarize this phenomenon as “too much is as bad as too little”. The optimal value of $r_{1}$ is about 7, which means there exists a theoretical optimal reward to promote scientists’ research performance by coordinating the interactive mechanism. Moreover, in Fig. 11(c), the y-axis represents the average unique topics adopted by the simulated scientists. We found that gray points can be fitted well by a linear function with a negative slope, −0.0014 ( $p v a l u e < 1 e^{- 3}$ ), which means that excessive rewards also discourage scientists from exploring new topics. However, exploration is one of the key factors in scientific innovation (Foster et al., 2015), and therefore excessive rewards may undermine the chances of originality. Hence, we provide theoretical evidence that a proper rewards mechanism may help the government and institutions adjust the research strategy of scientists, so as to better promote scientific development. Notably, because the majority of simulated scientists do not have any shells, and therefore the average number of papers in Fig. 11(a) is less than 1.

5. Discussion

Under the conflict defined as “the essential tension” (Kuhn, 1977) and “publish or perish” (Qiu, 2010), the productivity and the impact may be seen as the intermediate goals for scientists, to pursue the career advancement. In this paper, we explore to disclose and model the interactive mechanism composed of the productivity and the impact. In the following paper, we discuss the theoretical and practical implications and limitations of this study.

This paper has the following theoretical implications. First, we propose two correlation metrics to disclose the correlation in term of topic distribution from scientists' publication sequence, which verifies that the productivity and the impact is related to the evolution of scientists’ research interests. Specifically, some scientists tend to continuously select topics which help them produce more productivity and impact. Second, we propose a novel Q seashore walk model. The simulated publication sequences generated by our model also present the correlation consistent with that in the empirical data. Therefore, our model effectively imitates the process of topic selection, and further confirms the above conclusions. More importantly, our model shows that the productivity and the impact may be seem as a reward, and the interactive mechanism drives scientist's topic selection behavior. In addition, the spirit of our model provides an interesting perspective to examine scientists’ decision-making process, and contributes to the theoretical advancement of applying the reinforcement learning theory to analyze scientists’ behavior. Third, based on the simulated data, we also prove that proper rewards effectively stimulate the performance of scientists, but excessive rewards inhibit their performance. This research may help researchers deeply understand the process of topic selection, and provide a theoretical basis for research and development policy formulation.

This study has the following practical implications. According to our simulation results, we report evidence that the research performance of scientists may be affected by the reward, and follows the trend of increasing first and then decreasing. Hence, government departments and institutions may adjust their research and development policy, and help scientists to build a benign interactive mechanism, so as to better promote the scientific development. Taking the revolutionary scientific research evaluation mechanism in China in recent years for example, Academic Representative Work System is advocated and gradually implemented, which effectively prevents scientists from blindly pursuing the number of publications rather than papers’ quality.

However, there are still some limitations in this study. First, there still exists other many factors affecting scientists’ topic selection behavior, which may or may not affect the interactive mechanism. Hence, in future studies, other factors need to be further analyzed. Second, our model confirms that there exists an optimal reward value to maximize the research performance of scientists. However, due to the fact that rewards from the real world are complex, the theoretical optimal reward is not equivalent to the real one. Thus, a sophisticated reward mechanism should be proposed, by which our model may provide more detailed guidance for research and development policy formulation. Furthermore, we only analyzed scientists from the CS field. Actually, scientists from different fields and from scientific strata may attach different importance to the same type of reward, or completely focus on a different type of reward, and therefore constitute a completely different interactive mechanism. Identifying what kinds of rewards scientists pursue can help government departments, academic publishing companies, and policy makers put forward practical proposals to promote the development of science.

6. Conclusion

In this study, we proposed two correlation metrics (TDC and TFC) to analyze 20,000 scientists’ publication sequence from the computer science field. We disclose that the productivity and the impact is related to the evolution of scientists’ research interests. Indeed, career advancement needs a steady stream of publications (Jia et al., 2017), and therefore scientists tend to continuously select topics which help them produce more productivity and impact. Notably, it should be emphasized that the observed results are statistical, not definitive, and individual scientists might exhibit a diversity of behaviors. To further figure out how the productivity and the impact drive the evolution of scientists’ research interests, we proposed a novel Q seashore walk model, which reported the evidence that the interactive mechanism is a reasonable explanation of how the productivity and the impact affect the scientists’ topic selection behavior. We also analyzed the role of reward for scientists’ research performance, and proved that proper rewards stimulate the performance of scientists, but excessive rewards inhibit their performance. We summarize this phenomenon as “too much is as bad as too little”. Thus, this study may help researchers figure out the motivations behind scientists’ decision-making process, deeply understand their topic selection behavior, and provide theoretical evidences for policy intervention in scientific research.

Declaration of Competing Interest

The authors declare no competing interests.

Acknowledgments

This work was supported by the Youth Science Foundation of the National Natural Science Foundation of China (grant no. 72004168).

References

Bu et al., 2018
Y. Bu, Y. Ding, J. Xu, X. Liang, G. Gao, Y. Zhao
Understanding success through the diversity of collaborators and the milestone of career
Journal of the Association for Information Science and Technology, 69 (2018), pp. 87-97
CrossRef View in Scopus Google Scholar
Buehling, 2021
K. Buehling
Changing research topic trends as an effect of publication rankings–The case of German economists and the Handelsblatt Ranking
Journal of Informetrics, 15 (2021), Article 101199
View PDF View article View in Scopus Google Scholar
Ciranka et al., 2022
S. Ciranka, J. Linde-Domingo, I. Padezhki, C. Wicharz, C.M. Wu, B. Spitzer
Asymmetric reinforcement learning facilitates human inference of transitive relations
Nature Human Behaviour, 6 (2022), pp. 555-564
CrossRef View in Scopus Google Scholar
De Domenico et al., 2016
M. De Domenico, E. Omodei, A. Arenas
Quantifying the diaspora of knowledge in the last century
Applied Network Science, 1 (2016), pp. 1-13
CrossRef Google Scholar
Duan et al., 2021
X. Duan, S. Ying, W. Yuan, H. Cheng, X. Yin
QLLog: A log anomaly detection method based on Q-learning algorithm
Information Processing & Management, 58 (2021), Article 102540
View PDF View article View in Scopus Google Scholar
Duch et al., 2012
J. Duch, X.H.T. Zeng, M. Sales-Pardo, F. Radicchi, S. Otis, T.K. Woodruff
The possible role of resource requirements and academic career-choice risk on gender differences in publication rate and impact
PLOS One, 7 (2012), p. e51332
CrossRef View in Scopus Google Scholar
Fadhly et al., 2018
F.Z. Fadhly, E. Emzir, N. Lustyantie
Exploring cognitive process of research topic selection in academic writing
English Review: Journal of English Education, 7 (2018), pp. 157-166
CrossRef Google Scholar
Foster et al., 2015
J.G. Foster, A. Rzhetsky, J.A. Evans
Tradition and innovation in scientists’ research strategies
American Sociological Review, 80 (2015), pp. 875-908
CrossRef View in Scopus Google Scholar
Hoonlor et al., 2013
A. Hoonlor, B.K. Szymanski, M.J. Zaki
Trends in computer science research
Communications of the ACM, 56 (2013), pp. 74-83
CrossRef View in Scopus Google Scholar
Huang et al., 2022a
S. Huang, Y. Huang, Y. Bu, W. Lu, J. Qian, D. Wang
Fine-grained citation count prediction via a transformer-based model with among-attention mechanism
Information Processing & Management, 59 (2022), Article 102799
View PDF View article View in Scopus Google Scholar
Huang et al., 2022b
S. Huang, W. Lu, Y. Bu, Y. Huang
Revisiting the exploration-exploitation behavior of scholars’ research topic selection: Evidence from a large-scale bibliographic database
Information Processing & Management, 59 (2022), Article 103110
View PDF View article View in Scopus Google Scholar
Huang et al., 2021
S. Huang, J. Qian, Y. Huang, W. Lu, Y. Bu, J. Yang, et al.
Disclosing the relationship between citation structure and future impact of a publication
Journal of the Association for Information Science and Technology, 73 (7) (2021), pp. 1025-1042
View in Scopus Google Scholar
Jia et al., 2017
T. Jia, D. Wang, B.K. Szymanski
Quantifying patterns of research-interest evolution
Nature Human Behaviour, 1 (2017), pp. 1-7
Google Scholar
Jin et al., 2021
C. Jin, Y. Ma, B. Uzzi
Scientific prizes and the extraordinary growth of scientific topics
Nature Communications, 12 (2021), pp. 1-11
CrossRef Google Scholar
Keshavarz and Shekari, 2020
H. Keshavarz, M.R. Shekari
Factors affecting topic selection for theses and dissertations in library and information science: A national scale study
Library & Information Science Research, 42 (2020), Article 101052
View PDF View article View in Scopus Google Scholar
Khosrowjerdi and Bornmann, 2021
M. Khosrowjerdi, L. Bornmann
Is culture related to strong science? An empirical investigation
Journal of Informetrics, 15 (2021), Article 101160
View PDF View article View in Scopus Google Scholar
Kuhn, 1977
T.S. Kuhn
The essential tension: Selected studies in scientific tradition and change
University Of Chicago Press, Chicago, Illinois (1977)
Google Scholar
Lakeh and Ghaffarzadegan, 2017
A.B. Lakeh, N. Ghaffarzadegan
Global trends and regional variations in studies of HIV/AIDS
Scientific Reports, 7 (2017), pp. 1-8
Google Scholar
Laudel, 2006
G. Laudel
The art of getting funded: How scientists adapt to their funding conditions
Science and Public Policy, 33 (2006), pp. 489-504
CrossRef View in Scopus Google Scholar
Li et al., 2017
M. Li, L. Yang, H. Zhang, Z. Shen, C. Wu, J. Wu
Do mathematicians, economists and biomedical scientists trace large topics more strongly than physicists?
Journal of Informetrics, 11 (2017), pp. 598-607
View PDF View article View in Scopus Google Scholar
Mandelbrote, 2001
S. Mandelbrote
Footprints of the lion
Cambridge Univ. Library (2001)
Google Scholar
Mnih et al., 2013
V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, M.A. Riedmiller
Playing Atari with Deep Reinforcement Learning (2013)
Google Scholar
ArXiv, abs/1312.5602.
Perianes-Rodriguez and Ruiz-Castillo, 2018
A. Perianes-Rodriguez, J. Ruiz-Castillo
The impact of classification systems in the evaluation of the research performance of the Leiden Ranking universities
Journal of the Association for Information Science and Technology, 69 (2018), pp. 1046-1053
CrossRef View in Scopus Google Scholar
Qiu, 2010
J. Qiu
Publish or perish in China: The pressure to rack up publications in high-impact journals could encourage misconduct, some say
Nature, 463 (2010), pp. 142-144
Google Scholar
Shen, Ma, & Wang, 2018
Z. Shen, H. Ma, K. Wang
A Web-scale system for scientific knowledge exploration. Annual Meeting of the Association for Computational Linguistics (2018)
Google Scholar
Silver et al., 2017
D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, et al.
Mastering the game of go without human knowledge
Nature, 550 (2017), pp. 354-359
CrossRef View in Scopus Google Scholar
Sinatra et al., 2016
R. Sinatra, D. Wang, P. Deville, C. Song, A.L. Barabási
Quantifying the evolution of individual scientific impact
Science (New York, N.Y.), 354 (2016), p. aaf5239
View in Scopus Google Scholar
Sinha et al., 2015
A. Sinha, Z. Shen, Y. Song, H. Ma, D. Eide, B.J. Hsu, et al.
An overview of microsoft academic service (mas) and applications
Proceedings of the 24th international conference on world wide web (2015), pp. 243-246
CrossRef View in Scopus Google Scholar
Sutton and Barto, 2018
R.S. Sutton, A.G. Barto
Reinforcement learning: An introduction
MIT Press (2018)
Google Scholar
Tang et al., 2008
J. Tang, J. Zhang, L. Yao, J. Li, L. Zhang, Z. Su
Arnetminer: Extraction and mining of academic social networks
Proceedings of the 14th ACM SIGKDD international conference on knowledge discovery and data mining (2008), pp. 990-998
CrossRef View in Scopus Google Scholar
Van Hasselt et al., 2016
H. Van Hasselt, A. Guez, D. Silver
Deep reinforcement learning with double q-learning
Proceedings of the AAAI conference on artificial intelligence (2016)
Google Scholar
Wallace and Ràfols, 2018
M.L. Wallace, I. Ràfols
Institutional shaping of research priorities: A case study on avian influenza
Research Policy, 47 (2018), pp. 1975-1989
View PDF View article View in Scopus Google Scholar
Wang et al., 2013
D. Wang, C. Song, A.L. Barabási
Quantifying long-term scientific impact
Science (New York, N.Y.), 342 (2013), pp. 127-132
CrossRef Google Scholar
Wang et al., 2016
Z. Wang, T. Schaul, M. Hessel, H. Hasselt, M. Lanctot, N. Freitas
Dueling network architectures for deep reinforcement learning
Proceedings of the PMLR International Conference on Machine Learning (2016), pp. 1995-2003
View in Scopus Google Scholar
Watkins and Dayan, 1992
C.J. Watkins, P. Dayan
Q-learning
Machine learning, 8 (1992), pp. 279-292
Google Scholar
Watkins, 1989
C.J.C.H. Watkins
Learning from delayed rewards
Robotics & Autonomous Systems, 15 (4) (1989), pp. 233-235
View in Scopus Google Scholar
Wei et al., 2013
T. Wei, M. Li, C. Wu, X.Y. Yan, Y. Fan, Z. Di, et al.
Do scientists trace hot topics?
Scientific Reports, 3 (2013), pp. 1-5
View in Scopus Google Scholar
Yu et al., 2021
X. Yu, B.K. Szymanski, T. Jia
Become a better you: Correlation between the change of research direction and the change of scientific performance
Journal of Informetrics, 15 (2021), Article 101193
View PDF View article View in Scopus Google Scholar
Zeng et al., 2019
A. Zeng, Z. Shen, J. Zhou, Y. Fan, Z. Di, Y. Wang, et al.
Increasing trend of scientists to switch between topics
Nature Communications, 10 (2019), pp. 1-11
Google Scholar
Zhang et al., 2019
F. Zhang, X. Liu, J. Tang, Y. Dong, P. Yao, J. Zhang, et al.
Oag: Toward linking large-scale heterogeneous entity graphs
Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining (2019), pp. 2585-2595
CrossRef Google Scholar
Zhu, Liu and Yang, 2021
N. Zhu, C. Liu, Z. Yang
Team size, research variety, and research performance: do coauthors’ coauthors matter?
Journal of Informetrics, 15 (2021), Article 101205
Google Scholar

Cited by (1)

An author credit allocation method with improved distinguishability and robustness
2023, Journal of Data and Information Science

View Abstract

[1] Bu et al., 2018
Y. Bu, Y. Ding, J. Xu, X. Liang, G. Gao, Y. Zhao
Understanding success through the diversity of collaborators and the milestone of career
Journal of the Association for Information Science and Technology, 69 (2018), pp. 87-97
CrossRef View in Scopus Google Scholar

[2] Buehling, 2021
K. Buehling
Changing research topic trends as an effect of publication rankings–The case of German economists and the Handelsblatt Ranking
Journal of Informetrics, 15 (2021), Article 101199
View PDF View article View in Scopus Google Scholar

[3] Ciranka et al., 2022
S. Ciranka, J. Linde-Domingo, I. Padezhki, C. Wicharz, C.M. Wu, B. Spitzer
Asymmetric reinforcement learning facilitates human inference of transitive relations
Nature Human Behaviour, 6 (2022), pp. 555-564
CrossRef View in Scopus Google Scholar

[4] De Domenico et al., 2016
M. De Domenico, E. Omodei, A. Arenas
Quantifying the diaspora of knowledge in the last century
Applied Network Science, 1 (2016), pp. 1-13
CrossRef Google Scholar

[5] Duan et al., 2021
X. Duan, S. Ying, W. Yuan, H. Cheng, X. Yin
QLLog: A log anomaly detection method based on Q-learning algorithm
Information Processing & Management, 58 (2021), Article 102540
View PDF View article View in Scopus Google Scholar

[6] Duch et al., 2012
J. Duch, X.H.T. Zeng, M. Sales-Pardo, F. Radicchi, S. Otis, T.K. Woodruff
The possible role of resource requirements and academic career-choice risk on gender differences in publication rate and impact
PLOS One, 7 (2012), p. e51332
CrossRef View in Scopus Google Scholar

[7] Fadhly et al., 2018
F.Z. Fadhly, E. Emzir, N. Lustyantie
Exploring cognitive process of research topic selection in academic writing
English Review: Journal of English Education, 7 (2018), pp. 157-166
CrossRef Google Scholar

[8] Foster et al., 2015
J.G. Foster, A. Rzhetsky, J.A. Evans
Tradition and innovation in scientists’ research strategies
American Sociological Review, 80 (2015), pp. 875-908
CrossRef View in Scopus Google Scholar

[9] Hoonlor et al., 2013
A. Hoonlor, B.K. Szymanski, M.J. Zaki
Trends in computer science research
Communications of the ACM, 56 (2013), pp. 74-83
CrossRef View in Scopus Google Scholar

[10] Huang et al., 2022a
S. Huang, Y. Huang, Y. Bu, W. Lu, J. Qian, D. Wang
Fine-grained citation count prediction via a transformer-based model with among-attention mechanism
Information Processing & Management, 59 (2022), Article 102799
View PDF View article View in Scopus Google Scholar

[11] Huang et al., 2022b
S. Huang, W. Lu, Y. Bu, Y. Huang
Revisiting the exploration-exploitation behavior of scholars’ research topic selection: Evidence from a large-scale bibliographic database
Information Processing & Management, 59 (2022), Article 103110
View PDF View article View in Scopus Google Scholar

[12] Huang et al., 2021
S. Huang, J. Qian, Y. Huang, W. Lu, Y. Bu, J. Yang, et al.
Disclosing the relationship between citation structure and future impact of a publication
Journal of the Association for Information Science and Technology, 73 (7) (2021), pp. 1025-1042
View in Scopus Google Scholar

[13] Jia et al., 2017
T. Jia, D. Wang, B.K. Szymanski
Quantifying patterns of research-interest evolution
Nature Human Behaviour, 1 (2017), pp. 1-7
Google Scholar

[14] Jin et al., 2021
C. Jin, Y. Ma, B. Uzzi
Scientific prizes and the extraordinary growth of scientific topics
Nature Communications, 12 (2021), pp. 1-11
CrossRef Google Scholar

[15] Keshavarz and Shekari, 2020
H. Keshavarz, M.R. Shekari
Factors affecting topic selection for theses and dissertations in library and information science: A national scale study
Library & Information Science Research, 42 (2020), Article 101052
View PDF View article View in Scopus Google Scholar

[16] Khosrowjerdi and Bornmann, 2021
M. Khosrowjerdi, L. Bornmann
Is culture related to strong science? An empirical investigation
Journal of Informetrics, 15 (2021), Article 101160
View PDF View article View in Scopus Google Scholar

[17] Kuhn, 1977
T.S. Kuhn
The essential tension: Selected studies in scientific tradition and change
University Of Chicago Press, Chicago, Illinois (1977)
Google Scholar

[18] Lakeh and Ghaffarzadegan, 2017
A.B. Lakeh, N. Ghaffarzadegan
Global trends and regional variations in studies of HIV/AIDS
Scientific Reports, 7 (2017), pp. 1-8
Google Scholar

[19] Laudel, 2006
G. Laudel
The art of getting funded: How scientists adapt to their funding conditions
Science and Public Policy, 33 (2006), pp. 489-504
CrossRef View in Scopus Google Scholar

[20] Li et al., 2017
M. Li, L. Yang, H. Zhang, Z. Shen, C. Wu, J. Wu
Do mathematicians, economists and biomedical scientists trace large topics more strongly than physicists?
Journal of Informetrics, 11 (2017), pp. 598-607
View PDF View article View in Scopus Google Scholar

[21] Mandelbrote, 2001
S. Mandelbrote
Footprints of the lion
Cambridge Univ. Library (2001)
Google Scholar

[22] Mnih et al., 2013
V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, M.A. Riedmiller
Playing Atari with Deep Reinforcement Learning (2013)
Google Scholar

[23] ArXiv, abs/1312.5602.

[24] Perianes-Rodriguez and Ruiz-Castillo, 2018
A. Perianes-Rodriguez, J. Ruiz-Castillo
The impact of classification systems in the evaluation of the research performance of the Leiden Ranking universities
Journal of the Association for Information Science and Technology, 69 (2018), pp. 1046-1053
CrossRef View in Scopus Google Scholar

[25] Qiu, 2010
J. Qiu
Publish or perish in China: The pressure to rack up publications in high-impact journals could encourage misconduct, some say
Nature, 463 (2010), pp. 142-144
Google Scholar

[26] Shen, Ma, & Wang, 2018
Z. Shen, H. Ma, K. Wang
A Web-scale system for scientific knowledge exploration. Annual Meeting of the Association for Computational Linguistics (2018)
Google Scholar

[27] Silver et al., 2017
D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, et al.
Mastering the game of go without human knowledge
Nature, 550 (2017), pp. 354-359
CrossRef View in Scopus Google Scholar

[28] Sinatra et al., 2016
R. Sinatra, D. Wang, P. Deville, C. Song, A.L. Barabási
Quantifying the evolution of individual scientific impact
Science (New York, N.Y.), 354 (2016), p. aaf5239
View in Scopus Google Scholar

[29] Sinha et al., 2015
A. Sinha, Z. Shen, Y. Song, H. Ma, D. Eide, B.J. Hsu, et al.
An overview of microsoft academic service (mas) and applications
Proceedings of the 24th international conference on world wide web (2015), pp. 243-246
CrossRef View in Scopus Google Scholar

[30] Sutton and Barto, 2018
R.S. Sutton, A.G. Barto
Reinforcement learning: An introduction
MIT Press (2018)
Google Scholar

[31] Tang et al., 2008
J. Tang, J. Zhang, L. Yao, J. Li, L. Zhang, Z. Su
Arnetminer: Extraction and mining of academic social networks
Proceedings of the 14th ACM SIGKDD international conference on knowledge discovery and data mining (2008), pp. 990-998
CrossRef View in Scopus Google Scholar

[32] Van Hasselt et al., 2016
H. Van Hasselt, A. Guez, D. Silver
Deep reinforcement learning with double q-learning
Proceedings of the AAAI conference on artificial intelligence (2016)
Google Scholar

[33] Wallace and Ràfols, 2018
M.L. Wallace, I. Ràfols
Institutional shaping of research priorities: A case study on avian influenza
Research Policy, 47 (2018), pp. 1975-1989
View PDF View article View in Scopus Google Scholar

[34] Wang et al., 2013
D. Wang, C. Song, A.L. Barabási
Quantifying long-term scientific impact
Science (New York, N.Y.), 342 (2013), pp. 127-132
CrossRef Google Scholar

[35] Wang et al., 2016
Z. Wang, T. Schaul, M. Hessel, H. Hasselt, M. Lanctot, N. Freitas
Dueling network architectures for deep reinforcement learning
Proceedings of the PMLR International Conference on Machine Learning (2016), pp. 1995-2003
View in Scopus Google Scholar

[36] Watkins and Dayan, 1992
C.J. Watkins, P. Dayan
Q-learning
Machine learning, 8 (1992), pp. 279-292
Google Scholar

[37] Watkins, 1989
C.J.C.H. Watkins
Learning from delayed rewards
Robotics & Autonomous Systems, 15 (4) (1989), pp. 233-235
View in Scopus Google Scholar

[38] Wei et al., 2013
T. Wei, M. Li, C. Wu, X.Y. Yan, Y. Fan, Z. Di, et al.
Do scientists trace hot topics?
Scientific Reports, 3 (2013), pp. 1-5
View in Scopus Google Scholar

[39] Yu et al., 2021
X. Yu, B.K. Szymanski, T. Jia
Become a better you: Correlation between the change of research direction and the change of scientific performance
Journal of Informetrics, 15 (2021), Article 101193
View PDF View article View in Scopus Google Scholar

[40] Zeng et al., 2019
A. Zeng, Z. Shen, J. Zhou, Y. Fan, Z. Di, Y. Wang, et al.
Increasing trend of scientists to switch between topics
Nature Communications, 10 (2019), pp. 1-11
Google Scholar

[41] Zhang et al., 2019
F. Zhang, X. Liu, J. Tang, Y. Dong, P. Yao, J. Zhang, et al.
Oag: Toward linking large-scale heterogeneous entity graphs
Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining (2019), pp. 2585-2595
CrossRef Google Scholar

[42] Zhu, Liu and Yang, 2021
N. Zhu, C. Liu, Z. Yang
Team size, research variety, and research performance: do coauthors’ coauthors matter?
Journal of Informetrics, 15 (2021), Article 101205
Google Scholar

Outline

Cited by (1)

Figures (11)

Tables (5)

Journal of Informetrics

Highlights

Abstract

Keywords

1. Introduction

2. Background

2.1. The factors affecting topic selection behavior

2.2. Reinforcement learning theory

3. Methodology

3.1. Problem definition

3.2. Topic distribution correlation metrics and topic frequency correlation metrics

3.3. Q seashore walk model

3.4. Data collection and preprocessing

4. Experiments and results

4.1. Descriptive statistics

4.2. Correlation analysis based on TDC and TFC metrics

4.3. Robustness tests

4.4. Modeling the process of the topic selection by the Q seashore walk model

5. Discussion

6. Conclusion

Declaration of Competing Interest

Acknowledgments

References

Cited by (1)

An author credit allocation method with improved distinguishability and robustness