Most network studies on biological interactions consider only a single interaction type. However, individual species are simultaneously positioned in various types of interactions. The ways in which different network types are merged and entangled, and the variations in network structures between different sympatric networks, require full elucidation. Incorporating interaction types and disentangling complex networks is crucial, because the integration of various network architectures has the potential to alter the stability and co-evolutionary dynamics of the whole network.
To reveal how different types of interaction networks are entangled, we focused on the interaction between birds and flowers of temperate plants in Japan, where flower-feeding birds are mainly generalist passerines, acting as pollinators and predators of flowers.
Using long-term monitoring data, we investigated the flower-feeding episodes of birds. We constructed the whole network (WN) between birds and plants, separating the network into mutualistic and antagonistic sub-networks (MS and AS, respectively). We investigated structural properties of the three quantified networks and species-level characteristics of the main bird species. For bird species, we evaluated dietary similarity, dietary specialization and shifts of feeding behaviour relative to plant traits.
Our results indicate that WN comprises entangled MS and AS, sharing considerable proportions of bird and plant assemblages. We observed distinctive differences in the network structural properties between the two sub-networks. In comparison with AS, MS had lower numbers of bird and plant species, showed lower specialization and modularity and exhibited higher nestedness. At the species level, the Japanese white-eye acted as pollinator, while the brown-eared bulbul acted as both pollinator and predator for large numbers of flowers, based on its behavioural plasticity. Overall, the pattern of avian feeding behaviour was influenced by flower size and plant origin. Birds showed nectarivory for plants with medium-sized flowers and exotic origins.
Our results highlight the complex patterns of interactions between birds and the flowers of plants in temperate regions. They also indicate that understanding the interaction type for each species pair and consideration of the behavioural plasticity of animal species are important for elucidating integrated network structures.
Mutualistic interactions between organisms play crucial roles in the maintenance of biodiversity and drive evolutionary dynamics in communities. The community-level architecture of mutualistic interactions has considerable implications for these two aspects and has been extensively studied. In the last decade, network approaches have been applied to the analysis of community-level mutualistic interactions, such as pollination and seed dispersal (Bascompte et al. 2003; Bascompte & Jordano 2007; Ings et al. 2009). This approach has advanced our understanding of mutualisms by quantifying the various structural characteristics and demonstrating their implications for community persistence and stability.
Currently, most network studies on biological interactions have considered only a single interaction type (i.e. mutualistic or antagonistic). In general, individual species are simultaneously positioned in various types of interactions (Melián et al. 2009; Fontaine et al. 2011). The interaction between plants and flower visitors is a key target of mutualistic network research. Within this interaction, the whole network can be a complex of mutualism and antagonism, because flower visitors can include predators or cheaters, along with legitimate pollinators (Strauss & Whittall 2006; Genini et al. 2010). Moreover, a single animal species can act as a mutualist and an antagonist for different plant species or even for the same plant species. Thus, it is difficult to draw an a priori dichotomy between mutualists and antagonists. Incorporating interaction types and disentangling complex networks is crucial, because the integration of various network architectures has the potential to alter the stability and co-evolutionary dynamics of the whole network (Thébault & Fontaine 2008; Fontaine et al. 2011; Mougi & Kondoh 2012). Nevertheless, Fontaine et al. (2011) emphasized that even descriptive data regarding actual merging sub-networks are scarce.
Interactions between flowers and flower-visiting birds have been studied as typical mutualistic systems, in relation to ecological consequences and co-evolutionary dynamics (Temeles & Kress 2003; Cronk & Ojeda 2008; Johnson & Nicolson 2008). Network structures of flowering plant and bird species are mainly investigated in the regions around the tropics, where avian nectar specialists (e.g. hummingbirds) are distributed, having evolved into highly specialized systems (Dalsgaard et al. 2009, 2011; but see Johnson & Nicolson 2008). However, consideration of interaction types has not been conducted on flower–bird interactions, despite the presence of avian flower predators and nectar robbers (Galetti 1993; Traveset, Willson & Sabag 2002). The evaluation of interaction types is critical, especially the network of temperate East Asia and Europe, where avian nectar specialists are absent. In these regions, assemblages of generalist passerines are the main avian flower visitors. Some of these passerines have been shown to act as effective pollinators of specific winter-flowering plants (Kunitake et al. 2004; Fang, Chen & Huang 2012) and as flower predators (Numazawa 1989; Corlett 2004). Therefore, complex interactions including mutualisms and antagonisms may evolve between assemblages of flowers and flower-feeding birds. This indicates the importance of integrating interaction type, when elucidating integrated network structures and evaluating their ecological and evolutionary implications.
We characterized the pattern of bird–flower interactions in temperate Japan, by incorporating interaction types and decomposing the whole bird–flower network (hereafter WN) into mutualistic and antagonistic sub-networks (hereafter MS and AS, respectively). We tested the hypothesis that these entangled sympatric networks differ in their structural properties, as suggested by a general cross-comparison of different network types (Bascompte et al. 2003; Thébault & Fontaine 2008). We based our analysis on data obtained from long-term monitoring by regional volunteer naturalists. First, we developed three types of bird–flower quantitative networks, that is, WN, MS and AS and elucidated how the two sub-networks merge and entangle within WN. We characterized the networks using various metrics regarding their structures. Secondly, we analysed the species–level characteristics of the main bird and plant assemblages. We assessed the variation in feeding behaviours to flowers among the main avian flower feeders, including the variation in dietary specialization. Finally, we constructed a general pattern of bird–flower interactions, from the perspective of both network and species levels by integrating our results.
Materials and methods
The study was undertaken in Kanagawa prefecture, Japan (35°13′N to 35°67′N, 138°92′E to 139°80′E; 2416 km2). The area around Yokohama has a temperate climate with a mean annual rainfall of 1622·5 mm and a mean annual temperature of 15·5 °C. The temperate forest constitutes 39·7% of this prefectural area, followed by urban areas (26·8%) and farmland (8·6%). For more detailed information, see Yoshikawa & Isagi (2012).
We analysed the flower-feeding records collected by volunteer members of the Kanagawa branch of the Wild Bird Society of Japan (Kanagawa Branch of The Wild Bird Society of Japan 2007). This ongoing monitoring began in 1977 with the main aim of monitoring the spatial distribution of bird species and their long-term changes, to reveal the ecology of the birds. These monitoring data are composed of field observations that members of the branch throughout the prefecture document during their individual birding visits, depending on their own interests. The branch previously offered members reporting cards or Microsoft Excel files to standardise information such as the name of the observer, date of observation, bird species, number of individuals, bird behaviour, location and site environment. However, a description of bird behaviour is not obligatory. The observers voluntarily filled out these cards or files and sent them to the society. The data base includes about 180 000 individual observations on 380 bird species from 1977 to 2006. There are currently 400 observers, with approximately 2200 across the 29-year period. In addition, specific regional observers, classed as ‘veteran naturalists’ (currently about 40 members) are obliged to report at least one episode on observed bird species per month within assigned regional blocks (one to seven blocks per observer, with 104 total blocks within the prefecture; see Yoshikawa & Isagi 2012) and are encouraged to describe avian behaviours with the maximum possible detail.
Selection of avian flower-feeding episodes and characterization of interaction type
We focused on the flower-feeding episodes of terrestrial bird species on animal-pollinated flowers. We examined all records regarding birds’ foraging behaviour, selecting flower-feeding episodes. We focused on: (i) foraging episodes to animal-pollinated flowers; (ii) flower-feeding episodes, where we could identify plant species; and (iii) flower-feeding episodes in lowland areas, that is, <500 m altitude. We focused on the interactions in the lowland areas, as considerable differences occur in bird and plant assemblages between the lowland and highland areas. The observation episodes were mainly from the lowland areas. The locations of the observed flower-feeding episodes were broadly distributed in the prefecture lowlands, reflecting a general interaction pattern in the region. For all of the selected episodes, we checked plant species names and nomenclatures, using the BG Plant Database (Yonekura & Kajita 2003). We checked plant family names according to the Angiosperm Phylogeny Family Database (http://ctfs.arnarb.harvard.edu/webatlas/apgnames) based on APG III. The focal region contained several species of cherry trees (genus Cerasus), including native species and cultivars (Flora-Kanagawa Association 2001). These closely related species are similar in their flower morphologies and sometimes hybridised and are often reported in the data set using the general name. In such cases, we categorised the species into two species, natural species and cultivated species (mainly Cerasus × yedoensis), according to the environment where they were observed. The study region contains urban areas, where some observations occurred in areas such as along roadsides and in gardens. Consequently, some exotic and ornamental plant species were present in the selected episodes. Therefore, the interaction pattern between birds and flowering plant species differs from native interactions in the region, but reflects current interactions.
For each feeding episode, we identified the bird's feeding behaviour towards the flower, categorising it into three types: nectarivory, flower predation and unidentified. Behaviours such as sucking nectar and inserting the bill into the flower corolla were categorised as nectarivory. Nectarivory does not always mean successful pollen attachment to the birds, but can be a prerequisite for successful pollination. Behaviours resulting in whole or partial destruction of the flower, such as swallowing the whole flower, petal or bud, were categorised as flower predation. Nectar robbing (i.e. feeding on nectar by piercing the flower) can partially destroy flowers, but may result in pollen transport (Maloof & Inouye 2000). This behaviour was recorded infrequently and categorised as flower predation. In cases where ambiguous terminology was used in reporting avian behaviour, we classed as unidentified. When nectarivory and predation behaviours were reported in a single episode, we divided these occurrences into two episodes.
To validate if the monitoring data reflected the actual abundance of birds in the region, we compared them with line census data obtained by members of the Society branch independent of the focal monitoring. These line censuses were conducted monthly along 13 lowland routes (1·2–5·0 km) for 10 years (1999–2008), with a total of 1560 censuses. For each of the 78 terrestrial bird species, we compared the number of reported observations in the lowlands monitoring data with the number of censuses where the bird was recorded. We fitted a linear regression to the number of reported observations, with the number of censuses where the bird was recorded as the explanatory variable (without intercept; estimate = 2·623, P <0·0001, Appendix S1). However, we found a negative correlation between the explanatory variable and the model residual (r = −0·392, P <0·001, Appendix S1). This suggests that these monitoring data reflect the actual abundance of the birds in the region, but are biased towards rare bird species. This is probably because the observers reported rare birds for conservation purposes, for example, when the observer confirmed the presence of such birds. We believe this bias for rare birds could diminish after selecting the observations concerning avian feeding.
To determine if the observed flower-feeding episodes involved all the bird and plant species pairs identified in the three types of interactions (i.e. whole interaction, nectarivory and flower predation), we conducted a rarefaction analysis of the bird and plant species for each interaction. For each interaction (i.e. for WN, MS and AS), species accumulation curves were generated by re-sampling the episodes without replacement using EstimateS ver. 8.2.0 (Colwell 2006). This procedure was repeated 100 times, with species accumulation curves of the birds and plants obtained by linking the expected number of species (i.e. the mean number of species within a re-sampled, randomised set of episodes) with each episode number.
Structural properties of the three network types
From our flower-feeding episodes data, we reconstructed three types of bird–flower networks. WN was composed directly from the total number of flower-feeding episodes. We constructed MS and AS from the collected nectarivory and flower-predation episodes, respectively. The flower-feeding episodes categorised as unidentified were included in WN, but were excluded when constructing MS and AS.
Each network was bipartite, comprising two series of nodes (bird species and plant species), with links between the nodes (i.e. reported interactions between the species). The links were weighted by the number of interactions calculated as the total number of reported feeding episodes between the two species in the entire period. First, we measured the similarity of bird and plant assemblages between the two sub-networks to evaluate how much the sub-networks share species, thus, how much they were entangled. Similarity was measured using the Morisita–Horn index, which is resistant to under-sampling (Jost, Chao & Chazdon 2011). We characterized the network structure of the obtained bird–flower interactions using the following metrics: network-level specialization, nestedness and modularity (Dormann et al. 2009). We measured the network-level specialization using H2′, the standardied two-dimensional Shannon entropy (Blüthgen, Menzel & Blüthgen 2006). This metric, ranging from zero to one, increases when the network is specialized. The significances of the H2′ values were compared with those from the ‘r2dtable’ null model based on the Patefield algorithm (see Blüthgen et al. 2008). Nestedness is a measure of the departure from the systematic arrangement of species by niche width, interpreted as specialization asymmetry (Bascompte et al. 2003; Almeida-Neto et al. 2008). High nestedness means that specialists interact with species forming well-defined subsets, with which generalists in the network interact. Among several versions of this metric, we adopted the WNODF (weighted nestedness metric based on overlap and decreasing fill; Almeida-Neto & Ulrich 2011) to evaluate a quantified interaction matrix. This metric ranges from zero to 100 (no nestedness and perfect nestedness, respectively), calculated using the NODF program (Almeida-Neto & Ulrich 2011). Among the available null models (Ulrich, Almeida-Neto & Gotelli 2009; Ulrich et al. 2012; Ulrich & Gotelli 2013), we used those based on the ‘rc’ algorithm, which retains the sum of the rows and columns of the original interaction matrix. Modularity is the extent to which species in a network interact more within groups than among groups and is assumed to be associated with the spread speed of disturbance (Olesen et al. 2007). We measured the network modularity (M) using the NETCARTO program and an algorithm based on simulated annealing (Guimerà & Amaral 2005). Modularity was calculated for binary (i.e. 0/1) interaction matrices. We tested the significance of this value by comparing modularity with the values of 100 randomly generated networks, implemented in NETCARTO. This null model yielded models that retain the same degree of distribution and connectivity as the focal network.
Species-level characteristics of the main bird and plant species
We characterized the behaviours of the main flower-visiting bird species in relation to plant species. First, we quantified the dietary similarity between the main bird species within each sub-network (MS and AS) and across the sub-networks. We used the Morisita–Horn index to measure similarity in dietary composition. Secondly, we evaluated the dietary specialization of bird species in relation to plant species, using a rarefaction analysis (Herrera 2005) based on feeding episodes and a related species diversity index, Chao2. Plant species accumulation curves were obtained by re-sampling the episodes of each bird species without replacement using EstimateS version. 8.2.0 (Colwell 2006). This procedure was repeated 100 times, and plant species accumulation curves were obtained linking the expected number of plant species (i.e. the mean number of plant species within a re-sampled, randomised set of episodes) at each episode number. We calculated Chao2, estimating the total number of dietary species of a bird by incorporating the occurrence frequency of rare dietary species. Chao2 was calculated using EstimateS with the bias-corrected option. We checked Chao's estimated coefficients of variation (CVs) of incidence distribution. We followed Colwell's (2006) recommendations that if the CV were >0·5, we adopted the largest of the Chao2 calculated with the classic option and incidence-based coverage estimate (Magurran 1988) as the best estimate. The dietary specialization of each bird was evaluated for MS and AS.
Relationships between plant traits and avian feeding behaviours
For each plant species, we calculated the proportion of nectarivory to the total number of identified feeding behaviours by all bird species and by main bird species. To investigate the effects of plant traits on the avian feeding behaviours (i.e. nectarivory vs. flower predation), we analysed the data using generalized linear mixed models (GLMMs). We designated feeding behaviour towards a flower (pooled across all bird species) as the response variable, with plant flower size and flower size squared (considering a nonlinear effect, Crawley 2005), flower type (choripetalous/gamopetalous) and plant origin (native/non-native) as explanatory variables. We treated plant species as a random factor, using the binomial distribution and logit link function in the models. We constructed and checked all of the models with each possible combination of the four explanatory variables, adopting the model with the lowest Akaike information criterion (AIC) value as the best model. In addition, when large variations in feeding behaviour were detected within a bird species, we conducted the same GLMM analysis to investigate the effects of plant traits on the probability of nectarivory of the focal bird. Flower sizes were taken from the literature (Kitamura & Murata 1980; Satake et al. 1989 and Iwatsuki et al. 2001). We adopted petal length for choripetalous flowers and corolla length for tubular gamopetalous flowers as the flower size. For a few plant species, we were unable to obtain petal length and adopted flower radius as an alternative measure for flower size. Data regarding flower type and plant origin were extracted from the same literature sources.
Properties of the three network types
We obtained 784 terrestrial bird-to-flower feeding episodes, where we could identify the plant species, from 203 observers. Large parts of these episodes occurred during the winter–spring seasons (ca. 96% of the total episodes occurred between November and April.) In WN, we recorded 24 terrestrial bird species and 60 animal-pollinated plant species, including three exotic bird species and 30 non-native (exotic and/or ornamental) plant species (Fig. 1a). The number of episodes involving non-native species was 385, representing 49% of all episodes. MS was composed of seven bird species and 24 plant species, with 463 feeding episodes (Fig. 1b). AS was composed of 21 bird species and 41 plant species, with 249 episodes (Fig. 1c). Five bird species and 15 plant species occurred in MS and AS. Among the 784 episodes, four episodes occurred as a combination of nectarivory and flower predation (i.e. belonging to MS and AS). The Morisita–Horn indices of species composition similarity between MS and AS were larger for the plant assemblage (0·753) compared with the bird assemblage (0·464). The rarefaction analysis (Fig. 2) showed that the number of recorded species did not align completely with increasing episode numbers in the three networks, especially for plant species in AS. This indicates that the obtained data did not record every species pair of bird and plant interactions.
The values of network-level specialization (H2′), nestedness (WNODF) and modularity (M) differed between the three network types (Fig. 3). For all of the networks, the H2′ value was significantly larger (110–221%) compared with the null model (z test, P <0·0001 for all the networks); AS showed the highest value (0·37), while MS had the lowest value (0·10). For nestedness, MS had the highest value (49·37), followed by WN (23·26) and AS (18·75). However, these values showed no significant differences compared with the null models for any of the three networks (P =0·28, 0·37 and 0·28 for WN, MS and AS, respectively). Although non-significant, MS was more nested than the null model, whereas AS and WN were less nested than the null model. AS had the highest value (0·43) of modularity (M), followed by WN (0·42) and MS (0·30). However, for all of the networks, the modularity did not differ significantly from those based on the null models (P =0·69, 0·53 and 0·84 for WN, MS and AS, respectively).
Species-level characteristics of the main avian flower consumers and plant species
In WN, the bird species with the most feeding episodes was the Japanese white-eye (Zosterops japonicus), followed by the brown-eared bulbul (Microscelis amaurotis), the bullfinch (Pyrrhula pyrrhula) and the Eurasian tree sparrow (Passer montanus) (Table 1, Fig. 1). In MS, the white-eye and the bulbul accounted for > 95% of all feeding episodes. In AS, the bulbul accounted for the highest proportion (ca. 44%) of feeding episodes, followed by the bullfinch, hawfinch and sparrow. In MS, there was high dietary similarity (0·86) between the two main bird species, the white-eye and the bulbul (Table 2). Conversely, dietary similarity for AS was low between the bulbul and the bullfinch or sparrow, but high between the bullfinch and the sparrow. The white-eye and the sparrow showed very similar dietary compositions in MS and AS, whereas the bulbul showed lower similarity for dietary composition between the two sub-networks (Table 2). Analysis based on rarefaction (Fig. 4) revealed variations in dietary breadth between the bird species and between mutualistic and antagonistic relationships. The most diverse dietary breadth occurred in the AS for the bulbul (Chao2 = 54·6), followed by the white-eye in the MS (Chao2 = 27·6) and the bulbul in the MS (Chao2 = 13·0). Thus, the bulbul used a higher diversity of flowers for predatory feeding compared with nectarivory feeding. In comparison with the bulbul, the bullfinch and the sparrow showed relatively limited dietary breadths in the AS (Chao2 = 6·98 and 6·28, respectively).
Table 1. Basic information on the four main flower-feeding bird species. Numbers in parentheses indicate the number of episodes for non-native species, or the numbers of non-native plant species consumed, respectively. Bill lengths of the birds (means ± standard deviations) were measured using museum specimens
No. of episodes
No. of consumed plant species
No. of episodes
No. of consumed plant species
No. of episodes
No. of consumed plant species
14·7 ± 0·7 (n =9)
35·2 ± 2·0 (n =4)
13·8 ± 1·1 (n =4)
Eurasian house sparrow
15·2 ± 1·2 (n =3)
Table 2. Similarities (Morisita–Horn index) between dietary plant composition of the main bird species. Values on the left and right of the virgules indicate similarity between bird pairs within the mutualistic sub-network (MS) and antagonistic sub-network (AS), respectively. Values in square brackets indicate similarity within the same bird species across the two sub-networks (MS and AS). The bullfinch belongs to AS, therefore some values are missing from the table
We observed a large variation in the feeding behaviour of birds (Fig. 5a). We found considerable variation in the proportion of nectarivory feeding, between and within the main flower-feeding bird species (Fig. 5b–e). In our GLMM analysis (Table 3), the best model for proportions of nectarivory by all bird species included flower size and flower size squared and plant origin as an explanatory variable. The best model for plant species used by the bulbul was flower size and flower size squared. The bulbul used various sized flowers, showing marked predatory feeding behaviour for plant species with small and large flowers (Fig. 5c). The white-eye showed almost total nectarivory feeding behaviour for the flower assemblage, except for one plant with large flowers (Fig. 5b). The bullfinch and the house sparrow showed flower-predation behaviours for almost all plant species (Fig. 5d,e).
Table 3. Results of the generalized linear mixed models (GLMMs) for the probability of nectarivory for all bird species and for bulbuls. The binomial distribution and logit link function were used in the analysis. Feeding behaviours (nectarivory vs. flower predation) are designated as response variables and plant traits as explanatory variables. Plant species was designated as a random factor. The estimated coefficients in the best models (based on AIC) are shown
Our study is the first quantitative evaluation of a bird–flower network examining the variation in interaction types. Our results indicate that the whole bird–flower network comprises entangled mutualistic and antagonistic sub-networks, sharing considerable proportions of bird and plant assemblages. This was shown in the relatively high species composition similarity between the sub-networks. The structures of the two sub-networks differed considerably. First, compared with AS, MS had a significantly smaller network size, that is, the number of species occurring within a network but a higher number of feeding episodes (Fig. 1). This indicates that the frequency of interactions per bird–plant pair was larger than for AS. Thus, we propose that the inter-relationship between bird and plant species was stronger in MS than in AS, although further data relating to the cost/benefit ratio for each interaction type are required to demonstrate this hypothesis. Secondly, the two sub-networks differed in their structural properties (Fig. 3). Nestedness was higher in MS compared with AS, making the network more robust to perturbation or extinction of species (Bascompte et al. 2003; Memmott, Waser & Price 2004; Ulrich, Almeida-Neto & Gotelli 2009). However, we found no significant nested structures for the three networks. The reason is unclear, but their small network sizes may have restricted detection of nested structures. Currently, evaluations of nestedness for quantitative networks are rare requiring more data. Meanwhile, when applying NODF metrics, which are for binary (i.e. 0/1) matrices, to our networks, MS did not significantly show nestedness (P =0·805), whereas WN and AS showed anti-nested patterns (P =0·014 and P =0·005, respectively). The network-level specialization H2′ value was higher for AS compared with MS (Fig. 3a). This low degree of specialization in MS is similar to Dalsgaard et al. (2011) data for hummingbird–flower mutualistic networks in temperate North America. In contrast, the high H2′ value for AS was significantly higher compared with temperate mutualistic interactions (Dalsgaard et al. 2011). High network specialization in AS derived from the relative specialization of predators, except from the bulbul, and the little dietary similarity between bulbul and others (Fig. 4). We detected no significant modularity in any of the networks (Fig. 3c). This finding may be attributed to the small size of focal networks (Olesen et al. 2007).
Overall, the structural features of MS and AS are consistent with previously recorded patterns during the cross-comparison of different network types from various regions (Bascompte et al. 2003; Thébault & Fontaine 2010). Our findings indicate that these tendencies hold true, even for entangled sympatric sub-networks sharing species assemblages. However, the present AS has a generalist dominant predator (the bulbul, Fig. 4), although the high specialization observed in antagonistic networks is attributed to predator dietary specialization (Fontaine, Thébault & Dajoz 2009; Genini et al. 2010; Fontaine et al. 2011). Most studies examining bipartite antagonistic networks have focused on the networks existing between herbivorous insects and plants, or between parasites and host animals (Fontaine, Thébault & Dajoz 2009; Graham et al. 2009; Thébault & Fontaine 2010; Cagnolo, Salvo & Valladares 2011). Nevertheless, interaction patterns may differ according to focal taxonomic positions and require careful assessment.
In plant reproduction, flower feeding by birds has different ecological consequences among plant species (Fig. 5a), because of functional variation between and within the bird species (Fig. 5b–e). The frequency of nectarivorous feeding differed significantly between plant species, varying according to plant origin and flower size (Table 3, Fig. 5a). For non-native plants and those with medium-sized flowers, bird feeding mainly resulted in pollination rather than predation (Table 3, Fig. 5a). Currently, while several plant species can be bird pollinated, the contribution of birds to pollination has been experimentally confirmed for relatively few plant species in East Asia (Camellia japonica, Kunitake et al. 2004; Eriobotrya japonica, Fang, Chen & Huang 2012; Stachyurus praecox, Kaoru Fujita, personal communication). The white-eye rarely acted as a predator of dietary plant species (Fig. 5b), indicating that this species acts as an effective pollinator for a wide range of flowers (Kunitake et al. 2004; Abe et al. 2011; Fang, Chen & Huang 2012; but see Ueda 1999). Conversely, the bulbul showed plasticity in feeding behaviour according to flower size (Fig. 5c), largely determining the overall association between the nectarivory rate and flower size (Fig. 5a). However, flower size may be determined by a plant's phylogenetic position. Of the 43 plant species that appeared in WN, we found that a larger part (c. 78%) of the variation in flower size was derived from the order level compared with the family (c. 6%) or genus level (c. 1%; nested anova, d.f. = 19, P =0·03 for plant order). Therefore, together with the net effect of flower size, the phylogenetic effect may be important for avian feeding patterns. The bullfinch and the house sparrow acted mainly as predators for a limited number of plant species (Figs 4 and 5d, e), possibly affecting the reproduction of those plants (e.g. cherry trees, Numazawa 1989; Mikami & Mikami 2012). The white-eye and the bulbul are common resident species in temperate Japan, distributed across various habitats from mountainous forests to urban areas, making their pollination role ubiquitous (e.g. Abe & Hasegawa 2007). Conversely, feeding adaptability by the bulbul (Fig. 5c) may constrain flower–bird mutualisms, and together with the white-eye's feeding pattern, may act as a selective pressure for flower traits such as flower size (Strauss & Whittall 2006).
Non-native species in the networks
In our study, non-native plant species represented a considerable proportion of dietary flowers, suggesting that non-native species contribute substantially to the diet of wintering passerines. Our results suggest that the birds tend to feed more on non-native nectar plants compared with native plants (Table 3). The reason for this association is unknown, but one possibility is that cultivated plants have more attractive floral traits such as rich dense nectar. Furthermore, the possibility exists that some non-native plants originated in the tropics, therefore, they have evolved with avian pollinators. Possible negative interaction effects between birds and non-native plant species are reduced pollination of native plant species via competition for avian pollinators and cross-pollination between related native and cultivated species (e.g. between natural and cultivated cherries Cerasus, Tsuruta et al. 2012). Plant mutualisms with birds may facilitate the invasion of exotic plant species (e.g. via seed dispersal; Gosper, Stansbury & Vivian-Smith 2005; Heleno et al. 2013). However, we expected that the contribution of nectarivorous bird species to plant invasion would be small, because the focal non-native plants did not include highly invasive species (Muranaka et al. 2005).
We demonstrated the existence of complex structures of bird–flower interaction networks in a temperate region. Functional variation between and within bird species resulted in entangled MS and AS structures. The results suggest that network research requires careful assessment of the interaction type for each species pair, without an a priori dichotomy of mutualist and antagonist animals. The behavioural plasticity of animals can play a crucial role in the merging and entanglement of different types of networks and should be considered for the elucidation of integrated networks. Future studies will require the integration of various types of biological networks comprised of a wide variety of taxa to elucidate whole community behaviours and co-evolutionary dynamics.
We are indebted to the members of the Kanagawa Branch of the Wild Bird Society of Japan, for their tremendous efforts in collecting data. We thank the editorial committee for the list of birds in Kanagawa prefecture and the late Tetsuichi Hamaguchi, for their generosity in allowing our research. We thank Takashi Ishii for preparing related data. We are grateful to the Yamashina Institute for Ornithology, Forestry and Forest Products Research Institute, Museum of Nature and Human Activities, Hyogo, for allowing us to measure the bill sizes of their bird specimens. We thank Shota Sakaguchi and Masato Ohtani for suggestions on plant phylogenetic analysis. We thank Kaoru Fujita for sharing information from her research and Manabu Kajita for his constant support throughout our study. Finally, we thank Anna Traveset and an anonymous referee for their valuable comments on an earlier version of our manuscript.