This study site was located in the Qizimeishan National Nature Reserve (30.041565°N, 109.77305° E, 1,800–1,900 m; Supplementary Figure 1 ), Hubei Province, China, which is rich in biodiversity in monsoon-dominated continental East Asia (Man et al., 2008; Li et al., 2021). Hubei Qizimeishan National Nature Reserve is a forest ecosystem nature reserve and mainly protects mid-subtropical forest ecosystem and rare and endangered wild fauna and flora, which is on the IUCN Green List (https://iucngreenlist.org/explore/green-list-sites/). The annual average temperature and precipitation in the study site is 8.9°C and 1876.0 mm, respectively (Man et al., 2008). In July and August, the forest margin herb community is composed of many co-flowering herbaceous and shrub species, which provide an ideal system to conduct studies on plant–pollinator interactions. We collected data on plant–pollinator interactions, combining plots and transect sampling method (Chamberlain et al., 2012). Along a track from the Nature Reserve to the town Chunmuying, field works were repeated in six transects, which were separated from each other by more than 0.5 km. To characterize plant–flower visitor interactions in an unbiased manner, we established 50 permanent plots (3 × 3 m for grassland plot or 5 × 2 m for roadside plot) in 2017 for a long-term pollination observation at the community level. An alternative plot was re-established near the origin one when it was destroyed or when the distribution ranges changed among years. As a result, pollination observation was conducted in a total 9–12 plots for each transect in each year, which could include all the flowering species to the greatest extent possible. The total area observed in each year was 510 m² in 2017, 537 m² in 2018, and 458 m² in 2019, respectively. Finally, 66 insect-pollinated plant species from 28 families were identified and included in our pollination network ( Supplementary Table 1 ).

Field observations were carried out during July and August 2017, 2018, and 2019. To build the pollination networks, we conducted all surveys at 3-day intervals for each transect. Pollination observations were arranged in 15-min observation periods from 0900 h to 1700 h in sunny and cloudy days when pollinators were active. Three plots were surveyed simultaneously with two or three observers each plot, recording the number of pollinators foraging each plant species during an observation period for one plot. In addition to the relatively abundant species, we also took care to sample the rare co-flowering species in the plot, ensuring to record as possible more plant–pollinator interactions as we could (Albor et al., 2022). A visitor that encountered the anther or stigma of a flower was considered a pollinator. Morphological species were recorded in the field, and specimens were identified to the lowest taxonomical level possible by specialists later. Voucher specimens of insects and plants were kept in South-Central Minzu University, Wuhan, China. To reduce the sampling bias due to differences in foraging-time among pollinators, plots in each transect were observed as a given sequence every day, and, finally, observations for each plot included morning, midday, and afternoon-periods. We focused on the diurnal pollination network and thus did not record nocturnal pollinators in the field work because rare species were pollinated at night in this study. In total, 1,255 hours of observations were assigned in 3 years (152.5 hours on 12–21 August 2017; 571.25 hours on 10 July to 21 August 21 2018; and 531.25 hours on 7 July 7 to 25 August 25 2019).

We collected the following traits for plants species: flower density, floral size, floral shape, floral symmetry, and floral color. Flower density was defined as the floral abundance of a species divided by the area of plots ( Supplementary Table 1 ). We used flower density instead of flower abundance because it played a more important role in pollinator attraction especially when mass-flowering plants were distributed mostly patchily. Here, we counted the numbers of open attractive units (flowers or inflorescences) per plot during field pollination observation (see Supplementary Table 1 for details of attractive unit for each species; Lázaro et al., 2020). For Asteraceae species, a capitula is considered as an attractive unit except for Artemisia lavandulifolia and Eupatorium lindleyanum, in which a spike and a synflorescence (small capitula numerous in apical dense corymb) as the unit respectively ( Supplementary Table 1 ).

For 61 focal species, we conducted size measurement on at least 30 fully opened flowers/inflorescence (the same attractive floral units as that used to estimate flower density) from 10 randomly selected individuals in 2021 and 2022. For each species, we measured width, height, and tube length (if possible) of a flower unit, using a digital vernier caliper. As the simplest measure that can be compared among all species, the width/diameter of a flower/inflorescence was finally used to assess the effect of floral size on network metrics at the species level ( Supplementary Table 1 ; Lázaro et al., 2020). For Houttuynia cordata, the diameter of involucral bracts was measured instead of the width of inflorescences. For the other five species (A. lavandulifolia, spike width; Astilbe chinensis, panicle width; Cryptotaenia japonica, umbellule diameter; Heracleum hemsleyanum, umbellule diameter; Hydrangea strigose, cyme width), in which inflorescences act as the attractive units, inflorescence size was collected from Flora of China (http://www.iplant.cn/foc) because only the flower size was measured in the field.

Floral shape was categorized into four types as follows: open dish/bowl, open tube, flag or gullet, and tube ( Supplementary Table 1 ). 1) Open dish/bowl: Species have exposed nectar and pollen, or pollen presented as pollinia. This group included species in Adoxaceae, Apiaceae, Apocynaceae, Araliaceae, Caprifoliaceae, Celastraceae, Gentianaceae, Geraniaceae, Hydrangeaceae, Hypericaceae, Epilobium, Primulaceae, Ranunculaceae, and Rosaceae. 2) Open tube: Species have a head of small ray and disc tubular flowers such as Asteraceae species or with densely flowered capitula or spike such as Dipsacus asper and Polygonum species. Stamens and pistils are exposed, and nectar are concealed at the base of narrow tubes. 3) Flag or gullet: Species have exposed stamens and pistils but concealed nectar at the bottom of narrow or wide tubes, including Fabaceae, Gentianaceae, Lamiaceae, Ranunculaceae (Aconitum hemsleyanum), Commelinaceae, Boraginaceae, and Primulaceae. 4) Tube: Species have deep narrow or wide corolla tube, hidden pollen, and concealed nectar such as in Campanulaceae, Scrophulariaceae, Gentianaceae, Asparagaceae, Balsaminaceae, Liliaceae, and Cucurbitaceae. In general, both pollen and nectar are easy access for open dish/bowl and open tube flowers. However, flag-, gullet-, and tube-shaped flowers are always mechanically strong.

Flower symmetry was categorized as zygomorphy or actinomorphy for each species. Asteraceae species have actinomorphic inflorescences (attractive unit) with zygomorphic flowers and thus were assigned as actinomorphy. Furthermore, we recorded floral color of each species in the field, which was recognized as five types: white/pink, yellow, yellow green, blue, or purple.

To obtain a global overview of plant–pollinator network structure, a quantitative visitation network was constructed from the interaction data pooled together sites and years, because insect abundance, diversity, and the plants that they use may vary among years (Ouvrard et al., 2018). Rather than on species identity, networks based on insect functional groups can reveal patterns in the functionality and sustainability of complex plant–pollinator communities when studied across gradients or replicates (Fontaine et al., 2006; Geslin et al., 2013; Koski et al., 2015). Therefore, we categorized insect visitors into 12 functional groups according to their body size and foraging behavior. That is: six Hymenoptera types (ants, ANT; bumblebees, BB; large solitary bees, LL; honeybees, HB; small bees, SB; wasps, WASP), two Diptera types (hoverflies, HF; other flies, FL), two Lepidoptera types (butterflies and moths except hawkmoths, BF; hawkmoths, HM), one Coleoptera type (beetles, BT), and other visitors (see Supplementary Table 3 for details of pollinator groups).The strength of each interaction was identified by the number of flower pollinators in a particular functional group that were observed visiting a focal plant species ( Supplementary Figure 2 ).

The following five network metrics at the species level were calculated from the plant–pollinator network. (1) Species strength is the sum of dependencies of each species, with higher value indicating more pollinator functional groups depending on it (see the works of Watts et al., 2016; Suárez-Mariño et al., 2022). (2) Weighted closeness centrality is calculated as the sum of the number of shortest distances between the species in question and all other species in the network, with all ties weighted as 1/(link weight/average link weight in the network) (Opsahl et al., 2010). Low closeness scores indicate specialization, and high closeness scores indicate nodes (pollinators) are more “central”, e.g., closer to all other species in the network. (3) Specialization d’ calculates how strongly a species deviates from a random sampling of interacting partners available, based on the observed interaction frequencies (Blüthgen et al., 2006). It ranges between 0 for extreme generalization and 1 for extreme specialization, respectively (Olesen et al., 2007). (4) Nestedness contribution estimates the individual contribution of each plant to the overall nested structure of the network (Saavedra et al., 2011; Suárez-Mariño et al., 2022). (5) Modularity contribution, namely, z value in the network, evaluates the individual contribution from each plant species to entire network modularity (Guimeraà et al., 2005; Suárez-Mariño et al., 2022). The first three metrics quantify plant specialization, and the last two metrics refer to its consequences on network structure. All network metrics were calculated by the “bipartite” package in R.

In addition, z values (within-module degree) and c values (among-module degree) were also calculated. Moreover, linear models (LMs) detected no significant relationship between c values and z values in the plant–pollinator network (t = 0.601, P = 0.55). Thus, z value for each species can be considered as its modularity contribution in the network (Suárez-Mariño et al., 2022). Similar with the previous studies, weighted versions of z and c were calculated using species strength instead of species degree here (Watts et al., 2016). To objectively define thresholds, 100 null models for original networks were run quantiles 95 (q95) as critical c and z values were employed. At the same time, we also computed quantiles 50 (q50) of the c and z, respectively. The c and z values were all calculated using czvalues function of R based on the quantitative modularity (Q), which was estimated by the QuanBiMo to algorithm (Dormann et al., 2009; Dormann and Strauss, 2014).

For each plant species, a topological role in the network was then assigned on the basis of the shape of the c and z frequency distribution in the network (Lázaro et al., 2020). A network hub (z_i ≥ z _q95, c_i > c _q50) is highly linked to species within their own module and species of other modules, which is important for the connectivity among species both within its own module and within the network (Olesen et al., 2007). Whereas, a module hub (z_i ≥ z _q95, c_i ≤ c _q50) plays an important role in connecting species within its own module. A connector species (z_i < z _q95, c_i > c _q50) is crucial for among-module connectivity but plays an inferior role within its own module. Peripheral species (z_i < z _q95, c_i ≤ c _q50) have a few interactions inside their own module and rarely link to any other modules.

Pollinators use floral traits such as size, shape, symmetry, and color to locate and discriminate between different co-flowering species in the community (Chittka and Raine, 2006). Therefore, different functional pollinators may have innate preferences for certain shapes and colors (Bascompte et al., 2003; Fontaine et al., 2006; Geslin et al., 2013; Reverté et al., 2016). In line with other studies, we did find similar pollinator foraging preferences in the study community ( Figure 1 ; Supplementary 5 ). For example, long-mouthpart pollinators such as lepidoptera and bumblebees preferentially foraged on more mechanical flowers with tubular corollas in module I and module II, whereas short-mouthpart ones such as honeybees, solitary bees, and flies tend to forage on flowers with open-corolla and radial symmetry (Wignall et al., 2006; Geslin et al., 2013). Furthermore, our study also revealed color preferences of functional pollinators: bumblebees favor blue and purple, whereas flies favor white and yellow ( Figure 1 ). Moreover, honeybees and small bees also preferred the most mass-flowering ones with significant largest flowers ( Figure 1 , Supplementary Figure 5 ). To some extent, this finding is consistent with what pollination syndrome theory and pollination system described (Faegri and Van der Pijl, 1979; Carstensen et al., 2016; Dellinger at al., 2019). However, our results also indicate that pollinators made their foraging decisions due to mixed traits rather than a single trait, because plant species with a single similar trait did not attract similar pollinator assemblages (Reverté et al., 2016).

As expected, flower density reflects local resource abundance, which is determined by neutral ecological process rather than phylogeny history (Violle et al., 2007). He et al. (2019) defined density as an ecosystem trait to link functional traits such as other floral traits to macroecology. In contrast, floral morphological traits are often phylogenetically conserved but also community dependent (Blomberg et al., 2003; Chamberlain et al., 2014; Reginato and Michelangeli, 2016). For example, Chamberlain et al. (2014) found that flower symmetry (radial and bilateral) was most frequently phylogenetically conserved, whereas flower size was less (phylogenetic signal detected in 41% and 28% of the total trees, respectively). In most cases, however, previous studies indicated that phylogenetic signal for floral color was lacking (except in the works of Shrestha et al., 2014; Reverté et al., 2016). In our community, all of the floral morphological traits measured (size, shape, symmetry and color) showed to be phylogenetical dependent ( Supplementary Figure 3 ), meaning that related species had similar floral morph in the focal community.

Although the floral shape, symmetry, and color played important roles in flower choices of functional pollinators, they did not affect either network metrics measured ( Supplementary Figure 4 ). Beyond that, both phylogenetically conserved floral size and phylogenetically independent flower density had significant influences on plant specialization and thus plant–pollinator network structure ( Figure 2 ). However, some differences were also detected between PSEM and SEM. To some extent, therefore, the results implied different impacts of phylogenetically conserved and independent floral traits on plant functional specialization and thus its consequences on network structure. Furthermore, unexpected positive relations were revealed between complementary specialization d’ and the other two indices measured as species strength and weighted closeness, which means a species may both be central in the network (a generalist) but also be a specialist ranked by d’. The implications of these results will be discussed in detail in the next sections.

In addition to the regular pollinators, more rewarding plants (e.g., more flower abundance and high local flower density) may increase occasional visitors, enhancing pollinator sharing with less abundant plants (Makino et al., 2007; Carvalheiro et al., 2014). As a result, these species tended to have a greater influence on the pollination of co-flowering plant species, resulting in more central and decreased functional specialization (Lázaro et al., 2020). In consistent with the hypothesis, the SEM confirmed that phylogenetic independent flower density was directly positively related with weighted closeness in our network ( Figure 2B ). The PSEM revealed that the floral size had positive influences on both species strength and weighted closeness ( Figure 2A ). Similar relationships were found in that of Lázaro et al. (2020), which suggested a direct positive related with the linkage level of plant species and an indirect positive relation with closeness centrality in a diverse dune marshland. However, it is important to note that our result indicated the effect of floral size after controlling for the effects of phylogenetic relatedness by PGLS. This means that both phylogeny and floral size determined species’ position in the network. Closely related plant species had similar floral size, and those with larger size can attract more pollinator functional groups and also occupy central positions in the plant–pollinator network. This result is expected because floral displays that vary in shape, size, color, height, and scent can act as attraction signals for flower visitors (Campbell et al., 2012; Gibson et al., 2012; Junker et al., 2013, Carvalheiro et al., 2014). Furthermore, we did not find out any correlations between other phylogenetic conserved floral traits (shape, symmetry, and color) and plant functional specialization and its consequences on network structure ( Supplementary Figure 4 ). However, these morphological traits may be related to pollinator preference and thus lead to the modularity structure which will be discussed in the later section.

Finally, different metrics were used in the network analysis that might be correlated among themselves (Watts et al., 2016; Lázaro et al., 2020). A species with higher complementary specialization d’ means that it has more specialized plant–pollinator interactions and low pollinator sharing in a plant–pollinator network, and vice versa. Therefore, there is always a negative correlation between complementary specialization d’ and other specialization metrics such as species degree, species strength, and closeness centrality (Lázaro et al., 2020; Suárez-Mariño et al., 2022). However, some studies also revealed an opposite pattern, an unexpected positive relationship between complementary specialization d’ and weighted closeness (Pocock et al., 2011; Trøjelsgaard et al., 2019; Gaiarsa and Bascompte, 2022). It seems like a “paradox” because low closeness scores indicate specialization and high closeness scores more central (e.g., closer to all other species in the network). Whereas, our results also revealed such a paradox ( Figure 2 ). Species with larger floral size and flower density were quantified as moderately specialized by complementary specialization d’, although they were the most centralized participants in the networks and were considered as high generalization when quantifying specialization with the other two indices.

In our study system, species with larger flower and patchily mass-flowering (high flower density) occupied central positions in the network, especially the three network hubs (C. monocephalum in module II, H. strigose in module III, and E. annuusm in module III), which had significantly larger floral size ( Supplementary Figure 5 ). Larger flowers can enhance floral attraction to pollinator visitation (Willmer, 2011), and thus, all functional pollinator groups were found foraging on them. However, most interactions actually occurred only with one or several focal functional groups ( Supplementary Figure 2 ). As a result, the complementary specialization d’ ranked them as high specialism, which refer exclusively to the interaction frequencies relative to the availability of the partners and completely ignore the actual number of partners (Blüthgen et al., 2006; Watts et al., 2016; Pardo-De la Hoz et al., 2022). As a contrast, rare species with smaller flowers were more likely “opportunists” and visited by the most common flower visitor(s) in the focal community (Koski et al., 2015). In line with other studies, therefore, our study also convinces that specialization indices convey different concepts of specialization and hence quantify different aspects (Watts et al., 2016). Regardless, just as Watts et al. (2016) suggested, it requires careful consideration when defining a specialist.

Plants offering more resources are likely to be visited by more pollinators and thus more likely to influence another indirectly via shared pollinators that, in turn, may result in plant–pollinator network modular structure (Kunin, 1997; Cartar, 2009; Bartomeus et al., 2013; Carvalheiro et al., 2014). Several studies have convinced that floral abundance had direct or indirect effects on species’ closeness in their study communities (Sazima et al., 2010; Watts et al., 2016; Lázaro et al., 2020) and that species strength showed a strong direct or indirect association with the modular structure of plant–pollinator networks (Watts et al., 2016; Suárez-Mariño et al., 2022). Similarly, our results also revealed flower density can indirectly positively influence the modularity contribution of each plant species via its direct positive effects on species strength and/or weighted closeness ( Figure 2 ). Floral size was important in the three network metrics at the species level measured except the nestedness contribution. Not only the indirect influences but also a direct effect of floral size on modularity contribution was detected in the PSME model ( Figure 2 ). Floral size was significant positively correlated with z value ( Figure 2 ). In our community, species with larger flowers had larger values of z or interacted more within their modules. In contrast, a reverse relationship was found across Canadian plant–pollinator communities (Chamberlain et al., 2014). This inconsistency suggests inconstant effects of floral size of a particular species on its contribution to network modularity, which may be pollination network context dependent.

In a pollination network context, pollination syndromes and their corresponding functional group of pollinators could contribute to modularity structure because interactions within pollination systems principally occur inside modules (Carstensen et al., 2016). For example, co-flowering species can filter pollinators via floral traits such as size, shape, symmetry, and color, because they may act as barriers to certain pollinators and thus drive pollinator preferences (Stang et al., 2006; Stang et al., 2009; Wei et al., 2021). Our results did reveal that different functional groups had different foraging preferences. For the six most frequent pollinators, butterflies and bumblebees with longer mouthparts mostly prefer to visit flag, gullet, or tubular bule/purple flowers that are more mechanical but did not distinguish floral symmetry, whereas others (small bees, honeybees, hoverflies, and other flies) showed preferences on white and yellow flowers that are radial and open access for pollinators ( Figure 1 , Supplementary Figure 5 ). Moreover, honeybees and small bees were also mostly frequently found to visit on larger flowers with high local density than the others ( Supplementary Figure 5 ). Therefore, module organization in our network is partly caused by convergent trait sets including floral shape, symmetry, color, and size. Such a co-evolutionary unit may describe the relationship between interacting species and give insights into the dynamics of ecological communities (Olesen et al., 2007).

For the whole network in a diverse community, it represents the community-wide interactions where niche partitioning in pollinator use and asymmetric facilitation may confer fitness advantage of rarer species (Wei et al., 2021). For the sub-network within module, however, it more likely describes how common and rare species (low local flower density) may interact with each other when pollinator niches overlap. Then, facilitative interactions among plants via pollinator sharing may favor rare plant species (Moeller, 2004; Wei et al., 2021). Furthermore, previous studies revealed that only occasional visitors might increase with local flower density (Makino et al., 2007). Rare species may thus benefit from both pollinator attraction and traits if it was growing with abundant hub species grouped into different modules. For example, rare species with smaller attractive unit were valued significantly lower specialization d’ in the focal community, partly indicating that they were more likely to be foraged by the most common pollinators attracted by common species (Koski et al., 2015). Hence, sub-networks may help us to understand the mechanisms behind which evolutionary and ecological factors determined plant–pollinator interactions and their changes across space and time (Carstensen et al., 2016).

Both nestedness and modularity are thought to provide benefits for ecological communities (Fortuna et al., 2010). Inconsistent with network modularity, nestedness contribution was determined neither by floral traits nor species-level specialization metrics ( Figure 2 ), although the study network exhibited both a significant nested structure and a significant modular structure. No significant relationship between nestedness contribution and modularity contribution was detected, either. This finding indicated other unmeasured traits such as phenological factors, to some degree, accounting for the nested structure of plant–pollinator network in the study community. For instance, Suárez-Mariño et al. (2022) showed that an increase in flowering overlap led to a higher degree of plant generalization and, in turn, with consequences for network nestedness. Furthermore, pollinators may play a more active role in the definition of interaction identity than plants because of pollinator mobility (Bascompte and Jordano, 2007). As a result, related animal species are more likely to share host plants than related plant species are to share pollinator visitors (Reverté et al., 2016). In agreement with the findings of Rezende et al. (2007b), our results suggested pollinators might be of even greater importance than plants on nestedness structure in our network. For example, more generalized pollinator species were found likely responsible for higher nestedness (Chesshire et al., 2021).