Quantifying the Dynamics of Network Detachment Across Scientific, Technological, Commercial, and Pharmacological Domains Authors: Binglu Wang, Ching Jin, Chaoming Song, Johannes Bjelland, Brian Uzzi, Dashun Wang Published: 2026 ABSTRACT ABSTRACTDespite the vast literature on complex network dynamics, our understanding of the network detachment process remains limited. This understanding is essential for deeper insight into network evolution. Here, we analyze four large‐scale datasets that capture the structural patterns of detachment across scientific, technological, commercial, and pharmacological domains. Across these diverse domains, we uncover one simple pattern of preferential detachment in social systems, whereby the probability for individuals or organizations to disengage from a previously adopted practice correlates with the number of network neighbors who have already detached. To examine the structural consequences of this empirical pattern, we show that preferential detachment can fundamentally alter how networked systems disintegrate, inducing a phase transition in systems typically considered robust against detachment. We further derive an analytical framework to systematically understand the impact of preferential detachment on network dynamics, pinpointing the topological conditions under which it accelerates, decelerates, or leaves unchanged the breakdown of the network compared with random detachment. Together, these results demonstrate that the detachment dynamics follow simple yet reproducible patterns, deepening our quantitative understanding of detachment within networked social systems, with implications for the robustness and functioning of innovation communities.Keywords: complex network, computational social science, computer science, innovation, network dynamics, phase transition, preferential detachment, science of science Introduction Over the past several decades, research on complex networks has transformed our understanding of how interconnected systems form and expand across a diverse range of domains, from scientific collaboration and technological adoption to online information diffusion [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]. A fundamental property of many networked systems is that popularity or activity eventually decays over time [15, 16, 17, 18, 19]. While the growth of networks has been extensively studied, the counterpart of this process—how networks disconnect through the disengagement or detachment of nodes—has received relatively little attention, especially within social contexts [16, 20], leaving open important questions about the dynamics of network detachment. Existing research on network breakdown largely stems from studies of non‐social systems such as power grids, the Internet, financial infrastructures, and ecological food webs [2, 3, 21, 22]. In these domains, node removal is typically modeled as resulting from random failures, targeted attacks, or optimal percolation strategies, with a focus on structural robustness and systemic resilience [23, 24]. While these frameworks provide valuable insights, they abstract away from a defining feature of social systems: detachment is rarely random. Instead, real‐world social systems may exhibit critical network dynamics that arise from socially embedded decisions, shaped by peer behavior, structural interdependencies, and shifting incentives [25]. Such socially mediated detachment can propagate through networks, producing fragmentation or even systemic collapse, yet its structural regularities remain largely understudied. One key reason for this gap in social systems is the historical lack of large‐scale datasets that systematically capture the detachment dynamics of networks [16]. Disengagement decisions are rarely recorded with the same precision as adoption or engagement events, making it difficult to systematically analyze patterns of detachment. To address this limitation, we compile and analyze four large‐scale datasets spanning the scientific, technological, commercial, and pharmacological domains. The first dataset (D 1) captures the detachment dynamics of scientific fields as scientists shift their focus away from certain areas of research. Here we use the Microsoft Academic Graph data [26] to measure the fraction of publishing scientists in each field over time, allowing us to delineate the life cycle of 1,219 scientific fields from 1940 to 2015 and quantify the dynamics of scientists leaving a field (Section S1.1). Our second dataset (D 2) records the dynamics of inventors shifting across patenting domains, including 6.9 m patents granted by the United States Patent and Trademark Office from 1976 to 2015. Similar to D 1, here we trace the life cycles of 291 patenting domains, approximated by Cooperative Patent Classification (CPC) codes, and measure the proportion of inventors in each domain over time (Section S1.2). For D 1 and D 2, we also construct the collaboration network using co‐authorship and co‐inventor information, respectively [4]. Our third dataset (D 3) focuses on a commercial domain and traces the individual usage history of mobile handsets within a European country (704 types of handsets by 3.5 m users), allowing us to quantify the detachment dynamics of mobile handsets through changes in daily active users over a period of nine years (2006–2014). In addition, we collect anonymized phone call records among 1.7 m users within a three‐month period (October 1, 2013‐January 1, 2014) to approximate the social network among a subset of users for the network‐based analyses (Section S1.3). Lastly, we construct our fourth dataset (D 4) tracing the dynamics of pharmacological technologies and analyze the detachment of drug mechanisms, i.e., biological mechanisms of action (MOAs), as pharmaceutical organizations shift their focus away from certain MOAs in drug development. Here we use the Cortellis Drug Discovery Intelligence database, which contains the largest and most comprehensive drug development records in the world, covering 648 702 drugs over the past 25 years [27, 28]. To quantify the detachment dynamics of pharmacological technologies, we examine MOAs associated with each drug and measure the research and development (R&D) focus of pharmaceutical companies on these MOAs. In total, we trace 248 MOAs by 5133 pharmaceutical companies and measure the fraction of companies using each MOA in drug development. We also calculate the similarity in market focus among pharmaceutical companies to approximate the underlying proximity between two companies (Section S1.4). The item‐specific relational networks exhibit heterogeneous, fat‐tailed degree distributions across all four empirical domains (Section S2.1). Together, these four datasets offer us an unprecedented opportunity to examine the temporal life‐cycle dynamics and the structural pattern in which the detachment processes unfold. Results Figure 1A,D,G,J plots the popularity dynamics of different scientific fields, patenting domains, mobile handsets, and drug MOAs, respectively. All four domains exhibit a similar rise‐and‐fall pattern, demonstrating that both the adoption and detachment processes are important for understanding these life‐cycle dynamics. Here we focus on the declining phase of each item and measure the rate of detachment ν as a function of time (t). We find that, across scientific fields, patenting domains, mobile handsets, and drug MOAs, the average detachment probability (ν) systematically increases with time (t) (Figure 1B,E,H,K). Going beyond the aggregate estimates, we test if the observed trend holds for individual scientific fields, patenting domains, mobile handsets, or drug MOAs. In particular, for each item, we build panel data containing all of its users, fit this panel data using linear regression, and estimate the slope of the fit for each item, to approximate how its detachment rate changes as a function of time (Figure 1C,F,I,L). We find that the majority of scientific fields (98%), patenting domains (92%), mobile handsets (91%), and drug MOAs (84%) can be approximated by a positive slope, suggesting the detachment probability of different items appears to trend upward in a rather ubiquitous manner. Temporal dynamics of innovation detachment. (A,D,G,J) Popularity dynamics (NN∗) of scientific fields (A), patenting domains (D), mobile handsets (G), and drug MOAs (J). Here we calculate the proportion of scientists and inventors in each field, the ratio of active users for each mobile handset, and the fraction of pharmaceutical organizations focusing on each MOA. Each line represents a scientific field (A), patenting domain (D), handset type (G), or drug MOA (J). The color code corresponds to peak dates (when the population reaches its maximum), shifting from blue to red. (B,E,H,K) The average detachment probability ν≡ΔN−N as a function of time t for all scientific fields (B), patenting domains (E), mobile handsets (H), and drug MOAs (K), showing that the detachment rate, on average, increases with time. The shaded region represents a 95% confidence interval. (C,F,I,L) We estimate the relationship between the detachment probability and time for individual scientific fields (C), patenting domains (F), mobile handsets (I), and drug MOAs (L), and we plot the distributions of the estimated slope coefficients. To elucidate the relationship between detachment rates and underlying network structures, we further test the hypothesis that detachment processes follow Poissonian dynamics. This is a prevailing assumption in computational models, often derived from epidemiological frameworks [5, 29, 30, 31, 32, 33, 34, 35, 36, 37], and it has shown initial success in modeling substitutive systems [5]. This assumption reflects the idea that items follow a characteristic lifespan and that the probability of detachment is independent of the behavior of an individual's network neighbors. To test this implication, we control for potential temporal effects by restricting our measurements to a short time interval beginning around each item's half‐life (i.e., the stage at which its popularity has fallen to half of its peak), where detachment turnover is most active and a common life‐cycle stage is shared across items with different lifespans (Figure 2 and Section S2.2). We verify that our results are robust to this choice of time window (Section S2.3). For each item i, we consider its item‐specific relational network, namely the subnetwork of the underlying relational network formed by users associated with i, and divide users into two groups depending on whether detachment events have occurred among their network neighbors (ri > 0 vs. ri = 0) within the observation window. We find that the distributions of detachment rates for the two groups show visibly different patterns (Figure 2A–D), with the distribution for the first group (ri > 0) shifting systematically to the right of the ri = 0 group. These findings suggest that the detachment rate tends to be higher when detachment events have occurred among an individual's network neighbors. We then test whether the detachment probability (ν i ) increases with the number of network neighbors who have also detached from the item (ri ). Figure 2E–H shows four case studies, where we measure the rate of detachment as a function of ri for one selected example in each dataset. Averaged curves across all items in each domain show similar upward trends, confirming that the representative examples are consistent with population‐level patterns (Section S2.4). Across these four cases, ν i shows a positive relationship with ri , indicating that the detachment probability appears to increase with the number of network neighbors who have detached. By contrast, in a null model of random detachment, ν i is independent of ri (Figure 2E–H, grey dashed line). To systematically test across different items in different domains (i.e., scientific fields, patenting domains, mobile handsets, and drug MOAs), we measure ν i and ri for each item in the four datasets, and we fit them with a linear ansatz, finding that a substantial fraction of scientific, technological, commercial, and pharmacological items exhibit a positive slope (73%, 70%, 79%, and 70%, respectively) (Figure 2I–L). This pattern also remains robust when exposure is normalized by degree, ri /ki , indicating that the result is not driven solely by the raw number of detached neighbors (Section S2.5). Detachment and the underlying social network. (A–D) Distributions of detachment probability plotted separately for people whose network neighbors have detached (orange) or not (grey). Here we control for temporal effects by taking a short time interval for our measurements (one year for the scientific fields, patenting domains, and drug MOAs, and 100 days for the mobile handsets). To maintain consistency of measurements across the four domains, we begin our measurements around the half‐life of each life cycle, i.e., when the item's popularity falls to half of its peak, which provides a comparable life‐cycle stage across items with different absolute lifespans; robustness checks using alternative reference points are reported in Section S2.3. We find the two distributions deviate from each other, suggesting that the underlying network further plays a role in detachment dynamics. (E–H) Detachment probability ν i as a function of ri for one case selected from each domain as an illustrative example (Peptide YY for scientific fields, A61K 38/00 for patenting domains, iPhone 3GS for handsets, and Drugs Targeting Calcium Channels (Voltage‐Gated) for drug MOAs). The four cases were selected as recognizable illustrative examples with sufficient observations across multiple values of ri for stable visualization. The dashed line indicates the null model, where detachment probability is independent of how many network neighbors have detached (ri ). (I–L) We estimate the relationship between the detachment probability ν i and ri for all scientific fields (I), patenting domains (J), mobile handsets (K), and drug MOAs (L), and we plot the distribution of the coefficients, finding that a vast majority of them can be approximated by a positive slope. Overall, these empirical findings reveal a systematic pattern that we refer to as preferential detachment: the probability that an individual disengages increases with the number of detachments among their network neighbors. This observation contrasts with the Poissonian assumption that underlies most prevailing computational frameworks for modeling network detachment dynamics. We use the term “preferential” in a statistical sense, rather than to imply conscious preference or a specific causal mechanism. By analogy to preferential attachment, it denotes a non‐uniform detachment probability with respect to a structural quantity, here local exposure to prior detachment among network neighbors. We also distinguish this usage from prior uses of “preferential detachment” in network growth and rewiring models, where the term often refers to degree‐dependent edge removal, link rewiring, node deletion, or the loss of structural connections [38, 39, 40, 41]. Here detachment is defined as item‐specific disengagement conditioned on neighbors’ detachment states. Most importantly, the scientific, technological, commercial, and pharmacological domains we studied differ substantially in their scope, scale, temporal resolution, and nature. Yet despite these clear differences, and the myriad factors that might affect the decisions of individuals and organizations to detach from an innovative pursuit [42, 43, 44, 45, 46, 47], the patterns of preferential detachment show remarkable consistencies across these rather diverse domains. The concept of preferential detachment challenges the Poissonian assumption that has traditionally underpinned computational models of network decline. At the same time, it connects to several important ideas in social science, from peer effects and network externalities [44, 48, 49, 50, 51, 52] to the theory of creative destruction [53] to vacancy chains [47, 54]. Interestingly, preferential detachment mirrors preferential attachment [1, 55, 56, 57, 58], the well‐established mechanism governing network growth that generates scale‐free structures with heterogeneous degree distributions. If preferential attachment governs how networks grow, preferential detachment may help explain how they decline. This parallel raises a natural question: what structural consequences might follow from this empirically observed detachment pattern? To examine one such consequence, we study how preferential detachment reshapes network structure during periods of decline. We incorporate this empirical pattern into a minimal network model. Specifically, we approximate disengagement as node removal and formalize a detachment process on a network with N nodes. At each time step, node i is removed with probability: where ri denotes the number of neighbors of node i that have already been detached, α captures the strength of preferential detachment, and S denotes the set of surviving nodes. When α = 0, nodes are removed randomly with the same probability, and Equation (1) recovers the existing modeling framework of random detachment. As α increases, node removal becomes increasingly conditioned on local detachment in the network. Estimating an empirical analogue of the preferential‐detachment strength further shows positive item‐level α values across all four domains (Section S2.6), providing a direct empirical bridge between the observed exposure–detachment relationship and the model parameter. To analytically investigate the process, we adopt the heterogeneous mean‐field approximation (HMA) and derive a self‐consistent condition for the critical threshold (fc ) under preferential detachment. The critical point satisfies: where G0(x)≡∑kPkxk and G1(x)≡∑k1⟨k⟩(k+1)Pk+1xk are the standard generating functions of the degree distribution. The auxiliary function h(f), which captures the evolving local survival probability of edges during detachment, follows the differential equation: with initial condition h (f 0)= 1. Together, Equations (2) and (3) provide a self‐consistent percolation threshold under preferential detachment. When α = 0, preferential detachment reduces to random detachment and Equation (2) collapses to the canonical Molloy–Reed criterion. For scale‐free networks with γ < 3, the canonical criterion predicts no nontrivial phase‐transition point below complete removal when expressed as the removed‐node fraction, fcto1. The situation changes qualitatively under preferential detachment (α > 0). As detachment progresses, h(f) drops below 1, which makes the factor ∑kPk(k2−k)h(fc)k−2 converge and thereby yields a finite critical fc < 1 satisfying Equation (2). Thus, regardless of how weak the preference is, preferential detachment restores a genuine phase transition in scale‐free networks. In other words, even an infinitesimal amount of preferential detachment fundamentally changes the breakdown dynamics: a phase transition emerges that is absent in the random‐detachment case. To verify this theoretical prediction, we first simulate the detachment process on a specific scale‐free network (Pk∼k−γ, with γ = 2.5), and we study its robustness to random detachment (α = 0) and preferential detachment (α = 0.5). Specifically, we trace the size of the giant component (S) as a function of the fraction of nodes removed (f). We find that under random detachment (Figure 3A, grey line), the network shows a high degree of robustness [2, 29, 37], as its giant component remains connected until all nodes are removed. If, however, the system undergoes preferential detachment (Figure 3A, red line), a fundamentally different pattern emerges. First, conditional on removing the same number of nodes, the network breaks down much more quickly than under random detachment. More importantly, the network appears to undergo a critical process and become fragmented when the removed node fraction exceeds a percolation threshold (fc ). Preferential detachment across different network topologies. Simulations of detachment dynamics on networks with different degree distributions: a scale‐free network with a heavy‐tailed degree distribution (exponent γ = 2.5; A–C), a network with a Poisson degree distribution (D–F), and a light‐tailed network with a normal degree distribution (G–I), all with the same average degree. Simulations are performed on networks of size N = 106 and averaged over 20 realizations. (A,D,G) Size of the giant component S, measured as the fraction of remaining nodes, as a function of the fraction of nodes removed (f), under preferential detachment (α = 0.5), and under random detachment (α = 0). Network visualizations show each network when half of its nodes have detached (f = 0.5) and under preferential (α = 0.5) and random (α = 0) detachment. Here the nodes that have detached are shown in grey; among the remaining nodes, those in the giant connected component are shown in orange, and all others are shown in blue. (B,E,H) Critical fraction of node removal as a function of α for networks with heavy‐tailed (B), Poisson (E), and light‐tailed (H) degree distributions. For scale‐free networks (γ = 2.5), simulated results under preferential detachment are shown as red squares, with the corresponding analytical prediction shown as a solid orange line. Results for targeted attacks on scale‐free networks (black dotted line) and random detachment (blue dotted line) are shown for comparison. We repeat the same analyses for networks with a Poisson distribution in (E) and a light‐tailed degree distribution in (H). (C,F,I) Relationship between the current degree kf and the number of detached neighbors r at different time snapshots (f = 0.1, 0.3, 0.5, 0.7, 0.9), obtained from Monte Carlo simulation under random detachment (circles) and preferential detachment (α = 0.5; squares) on heavy‐tailed (C), Poisson (F), and light‐tailed (I) networks. The dashed black line indicates the asymptotic behavior predicted by Equation (4) in the large‐r regime (see Section S4 for details). We further simulate the preferential detachment process with different α (Figure 3B), and simulation results show excellent agreement with our analytical calculations derived from Equation (2). To further illuminate the process, we visualize the network when half of the nodes are removed (f = 0.5) for both processes (subpanels in Figure 3A). While the two networks have the same number of nodes remaining, they show markedly different patterns. Whereas a giant component composed of high‐degree nodes persists under random detachment (α = 0), the giant component in the preferential detachment case (α = 0.5) only contains a small number of low‐degree nodes. This highlights a key difference between random and preferential detachment on scale‐free networks: compared with random detachment, hubs become disproportionately vulnerable under preferential detachment. These results align closely with the network science literature [2, 5, 16, 29, 37, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69], which has extensively examined the importance of hubs in maintaining the robustness of networks. When hubs are deliberately targeted, the network collapses early (Figure 3B, dotted line). This raises a natural question: is preferential detachment equivalent to a targeted hub‐attack mechanism? To answer this question, we further examine preferential detachment on two additional network topologies: Erdős–Rényi (ER) networks with a Poisson degree distribution and a Gaussian network with a normal degree distribution. If preferential detachment is merely a hub attack, a similar earlier onset of phase transition should be expected for both networks (Figure 3E,H, black dotted line) [2, 56, 65]. However, Equation (2) predicts different behavior. Specifically, for ER networks, G 0 (x) = G 1 (x) = e 〈k〉(x − 1) . Substituting this into Equation (2) yields the canonical Molloy–Reed criterion. In other words, preferential detachment follows indistinguishable dynamics from random detachment for the ER network. Interestingly, for Gaussian networks with degree distributions narrower than Poisson, preferential detachment induces an even later percolation transition than random detachment (Figure 3G), where the critical point is delayed as α increases (Figure 3H). Both predictions are further confirmed by simulation results (Figure 3D–I). These results indicate that preferential detachment is not a targeted hub‐attack mechanism. Instead, it has highly non‐trivial effects on different network topologies, highlighting a novel form of critical behavior and motivating a closer examination of the underlying mechanism. To further understand how the percolation threshold shifts for different network topologies, we examine the relationship between a node's number of detached neighbors (r) and its current degree (kf ) after a fraction f of nodes has detached. We first measure kf vs. r at different levels of f for the scale‐free network (Figure 3C), finding that the two quantities follow an asymptotically linear relationship, suggesting that nodes with larger r are also characterized by a higher remaining degree, mimicking the attack processes (see Section S4 for details). However, when we repeat the analyses on ER networks, we find that kf (r) remains constant for any given r (Figure 3F) under preferential detachment, as it does under random detachment. Because connections in ER networks are formed randomly, the number of a node's detached neighbors is statistically independent of its remaining degree during the detachment process. While nodes with higher r are more likely to be selected under preferential detachment, the corresponding kf is random for ER networks, explaining why preferential detachment on ER networks follows the same dynamics as random detachment. Mathematically, the asymptotic behavior of kf (r) as r→∞ can be directly calculated by our model: suggesting that this key relationship is governed by the original degree distribution of the network P(k) (see Section S4 for details). For the standard degree distribution considered here, whose tails are eventually decreasing in the large‐r regime, we have Pk(r+1)Pk(r)<1. Equation (4) therefore predicts that kf (r) is asymptotically bounded by a linear scaling. For example, it saturates to the linear scaling bound for scale‐free networks. However, for the Poisson degree P(k)∼λkk!, Equation (4) predicts kf∼λ, indicating kf does not change with r. Thus, the Poisson degree distribution serves as a key division here. If the degree distribution has a heavier tail than the Poisson distribution, such as power‐law distributions, kf increases with r. Hence, in such networks, hubs are more likely to be removed under preferential detachment, leading to an earlier onset of the phase transition. By contrast, for networks with a narrower degree distribution than the Poisson distribution, such as normal distributions, there is a negative correlation between kf and r (Figure 3I), where once nodes with higher r are removed, preferential detachment actually targets nodes with lower current degrees, thus inducing an even slower percolation than random detachment. Table 1 summarizes the predictions of Equation (4) for several common networks. TABLE 1Summary of different network structures.NetworkDegree Distribution kf(r)|r→∞ scale‐free networks k −γ r ER networks λke−λk! λGaussian networks e−12(k−μσ)2 0 Summary of different network structures. Discussion Taken together, this paper examines the declining dynamics across four distinct domains and makes three primary contributions. First, it empirically documents a simple yet ubiquitous phenomenon of preferential detachment governing the dynamics of network detachment. Across scientific, technological, commercial, and pharmacological domains, the probability of detachment is consistently associated with local exposure to prior detachment among network neighbors. These findings provide the central empirical contribution of the paper and suggest that detachment in social and innovation systems is not merely an individual‐level or item‐level life‐cycle process; it is also shaped by local network structure. Second, building on this empirical regularity, we incorporate preferential detachment into current modeling frameworks and find that its presence creates novel yet complex dynamics in the breakdown of the overall ecosystem, generating a new form of structural collapse in networked systems that are thought to be robust against random detachments. Third, we derive an analytical framework to systematically understand the impact of preferential detachment on network breakdown, pinpointing specific conditions under which it accelerates, decelerates, or leaves unchanged the breakdown of the network compared with random detachment, depending on the network topology. Overall, these results provide a new quantitative basis for understanding network detachment in social systems, which has implications for the robustness and functioning of interconnected communities. While this paper synthesizes empirical evidence from four distinct domains to document the phenomenon of preferential detachment and its impact on the disintegration of networked communities, it is important to acknowledge a significant limitation: the study does not establish the origins of preferential detachment and its potential interplay with preferential attachment. For example, it is possible that the observed preferential detachment is an extension or byproduct of the preferential attachment mechanisms that govern the growth phase of the new communities. Although the findings of this paper remain valid irrespective of the specific origins of preferential detachment, further research into the connections between the growth and decline of networked communities could pave the way for an integrated framework for the whole life cycle of network dynamics, which would deepen our understanding of and ability to predict how networked communities evolve and dissolve. Further, analyzing data across four diverse domains, this paper mainly focuses on universal detachment patterns that appear consistent across domains. At the same time, there are substantial heterogeneities both within and across domains. For example, scientists often have considerable autonomy in selecting research fields, whereas inventors typically operate under firm‐driven managerial strategies. While both scientists and inventors act as originators of ideas, mobile handset users function primarily as consumers. In contrast, the drug‐development dataset reflects organizational (firm‐level) decisions, where companies shift their R&D focus across different mechanisms of action. The four domains also differ in the nature of their underlying relational networks. Scientific fields and patenting domains are based on collaboration networks, and mobile handsets are based on direct communication ties, whereas drug MOAs are based on similarity in market focus among pharmaceutical organizations. The persistence of preferential detachment across these different relational settings suggests a common exposure‐dependent pattern, while variation in the estimated strength of detachment across domains indicates that the degree of “stickiness” may depend on the type of actors, ties, and institutional context. Observing consistent detachment patterns across these contextually distinct domains highlights important underlying commonalities in how detachment unfolds and further suggests fruitful directions for future research. For example, which factors tend to be associated with a higher or lower α parameter? Beyond the universal patterns presented in this paper, insights into the heterogeneity in the detachment processes may further our understanding of the potential mechanisms behind the detachment dynamics. Our framework also connects to recent work on dynamical processes in complex networks, including studies of strategy evolution on dynamic networks and the evolution of prosocial behavior in multilayer populations [70, 71, 72]. These studies show how network architecture, temporal variation, and multilayer coupling can shape microscopic behavioral dynamics. Our work is complementary to this literature. Rather than examining how strategies, behaviors, or ties co‐evolve on a fixed, multilayer, or time‐varying network, we focus on item‐specific detachment: how actors disengage from a focal scientific field, patenting domain, handset model, or drug mechanism of action as local detachment accumulates around them. This distinction is also important when relating our framework to temporal rich‐club structure. In much of the temporal‐network literature, an edge represents a social, communication, or interaction tie, and a hub refers to an actor who is highly connected in the broader temporal network. In our setting, by contrast, the empirical analyses are conducted on item‐specific relational networks: subnetworks of the underlying relational network formed by actors associated with a focal item. Therefore, a hub in our analysis is not necessarily a socially central actor in the general network, but an actor with many neighbors within the item‐specific participation network. Likewise, detachment does not imply disappearance from the broader network; it means leaving a focal field, domain, product, or mechanism of action. Temporal rich‐club stability and preferential detachment therefore capture different levels of network dynamics. Future work could further examine this relationship by stratifying detachment dynamics by degree class, k‐core level, and rich‐club membership, and by extending the present framework to explicitly multilayer or time‐varying network settings. Overall, this paper demonstrates that the dynamics of detachment across scientific, technological, commercial, and pharmacological domains follow simple yet reproducible patterns, commonly governed by preferential detachment. The phenomenon of preferential detachment fundamentally alters the breakdown of networked ecosystems, with implications for the robustness and functioning of the underlying systems. Curiously, preferential detachment resembles behavior observed across a wide range of settings, including organizational strategy [43, 46], culture [45], financial markets [44], and healthcare [47], suggesting that the concept of preferential detachment may have broad relevance that goes beyond these examples. Author Contributions B.W., C.J., C.S., B.U., and D.W. designed the research; B.W. and C.J. collected data and conducted empirical and analytical analyses with help from D.W., C.S., J.B., and B.U.; all authors discussed and interpreted results; B.W., C.J., and D.W. wrote the manuscript; all authors edited the manuscript. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Supporting File: advs78029‐sup‐0001‐SuppMat.pdf. Acknowledgements We thank all members of the Kellogg Center for Science of Science & Innovation (CSSI) and the Northwestern Institute on Complex Systems (NICO) for helpful discussions. D.W. is supported by the Air Force Office of Scientific Research under award numbers FA9550‐17‐1‐0089 and FA9550‐19‐1‐0354, National Science Foundation grant SBE 1829344, the Alfred P. Sloan Foundation G‐2019‐12485, and Peter G. Peterson Foundation 21048. C.J. acknowledges support from UK Research and Innovation/ESRC through the Metascience Research Grants opportunity OPP569 [grant number UKRI1090]. The funders had no role in study design, data collection and analysis, the decision to publish, or the preparation of the manuscript. Data Availability Statement The scientific publication data used in this work are from Microsoft Academic Graph. The raw data were publicly available at https://www.microsoft.com/en‐us/research/project/microsoft‐academic‐graph. The patent data used in this work are from PatentsView, which is publicly available at https://patentsview.org/download/data‐download‐tables. The mobile phone data are not publicly available because they contain commercially sensitive information, but they are available from the corresponding author upon reasonable request. This work also uses drug data sourced from Cortellis, and researchers who wish to access these raw data should contact Cortellis directly. Data and code necessary for reproducing the main results are available at https://github.com/BW‐KH/preferential‐detachment. References Associated Data Supplementary Materials Supporting File: advs78029‐sup‐0001‐SuppMat.pdf. Data Availability Statement The scientific publication data used in this work are from Microsoft Academic Graph. The raw data were publicly available at https://www.microsoft.com/en‐us/research/project/microsoft‐academic‐graph. The patent data used in this work are from PatentsView, which is publicly available at https://patentsview.org/download/data‐download‐tables. The mobile phone data are not publicly available because they contain commercially sensitive information, but they are available from the corresponding author upon reasonable request. This work also uses drug data sourced from Cortellis, and researchers who wish to access these raw data should contact Cortellis directly. Data and code necessary for reproducing the main results are available at https://github.com/BW‐KH/preferential‐detachment.