We characterized the structure of 343 bipartite ecological networks with wellknown interaction types (antagonistic versus mutualistic) and ecologies (e.g., pollination, herbivory, seed dispersal) using three different approaches that focus on different network scales (see Methods, Fig. 1; [25]). We first investigated the macroscale structure of the networks by using global metrics: connectance, nestedness, modularity, network size (i.e., the total number of species), and absolute difference in guild size (i.e., the number of species on each side of the network). We found differences in macro-scale structures between antagonistic and mutualistic networks; however, these were explained by specific ecologies rather than the overall interaction type. Specifically, nestedness was significantly higher in mutualistic networks (mean: 2.16 ± s.d. 0.71) than in antagonistic ones (1.94 ± 1.09; Wilcoxon-Mann-Whitney test: W = 8673, p-value < 0.001), whereas antagonistic networks were significantly more connected (0.23 ± 0.17) than mutualistic ones (0.15 ± 0.09; Wilcoxon-Mann-Whitney test: W = 17709, p-value < 0.001). The difference in terms of modularity was not significant (0.41± 0.15 for antagonistic networks and 0.43 ± 0.10 for mutualistic ones; Wilcoxon-Mann-Whitney test: W = 12559, p-value = 0.19) (Fig. 2A). In addition, we found that mutualistic networks were composed of more species compared to antagonistic ones (76.1± 92.3 for antagonistic networks and 82.7 ± 72.7 for mutualistic ones; Wilcoxon-Mann-Whitney test: W = 11124, p-value = 0.003) and that absolute differences between guild sizes were also more important in mutualistic networks (28.5 ± 46.7 for antagonistic networks and 35.4 ± 50.1 for mutualistic ones; WilcoxonMann-Whitney test: W = 10768, p-value < 0.001). When controlling for the different types of ecology (e.g., pollination, herbivory, seed dispersal...) using mixed models, antagonistic and mutualistic networks were not significantly different in terms of their nestedness (Wald test: χ2 = 0.060, p-value = 0.81), connectance (Wald test: χ2 = 0. 41, p132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 value = 0.52), modularity (Wald test: χ2 = 0.42, p-value = 0. 52), the total number of species (Wald test: χ2 = 0. 29, p -value = 0.59), and absolute difference in the number of species per guild (Wald test: χ2 = 0. 95, p-value = 0.33). Model comparisons confirmed that differences in terms of global metrics were better explained by the different types of ecology than

by the antagonistic/mutualistic type of the interactions (Supplementary Table 1). For instance, compared to other types of ecology, bacteriaphage networks were more connected but generally less nested, while herbivory networks were more nested (Supplementary Fig. 1, Supplementary Table 2 A-C).

failed at accurately classifying antagonistic versus mutualistic networks. Supervised artificial neural networks (ANN) performed much better. Using motif frequencies and ANN, we were able to properly classify around 82% of the empirical interaction networks. This percentage of successful classifications is higher than the success rate achieved by the only previous attempt we are aware of to classify networks with ANN based exclusively on their structure [20], and achieved without accounting for environmental conditions (as in 24). Our results suggest that motif frequencies, which depict patterns of species interactions at the mesoscale, are better at capturing structural differences between antagonistic and mutualistic networks compared to global metrics or the Laplacian spectral density, and that these differences are robust to different environmental conditions.

Our analyses confirm that global metrics, including the modular/nested dichotomy, are not sufficient to discriminate between mutualistic and antagonistic 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 network structures [20,21]. We did not find a significant difference in modularity values between mutualistic and antagonistic networks. We did find that empirical mutualistic networks tend to be more nested and less connected than antagonistic networks, as previously suggested by [3,12], but this dichotomy is not consistent enough (i.e., there are too many exceptions) to provide a robust classification criterion. There are several empirical examples of exceptions to the nested/modular dichotomy in the literature, such as a nested structure in host-parasite antagonistic networks [18]. Within our database, some pollination and mycorrhizal networks are known to be modular [9,39–43], while they assemble mainly mutualistic interactions.

We found that structural differences between mutualistic and antagonistic networks are much clearer when looking at motif frequencies. Densely connected motifs (e.g. motifs 31 to 37) are significantly more frequent in antagonistic networks, while mutualistic networks are characterized by motifs with asymmetrical specialization (e.g., motifs 10, 19, 20). Motifs focus on precise structural patterns of interspecific interactions that can be missed when summarizing all the information in a single global value at the network macro-scale [25]. Why motifs performed better than the spectral density of the Laplacian graph is unclear, as this latter representation is supposed to integrate all scales. Using a constant bandwidth value to convert the Laplacian graph spectrum into a density may result in information loss when comparing networks of different sizes. Also, transforming the discrete Laplacian spectrum into a continuous density may be problematic for small networks with only a few eigenvalues.

Mutualistic and antagonistic networks cover a wide range of ecologies (e.g. pollination, parasitism, herbivory) and our analyses showed that the type of ecology matters in the supervised classification, a phenomenon referred to as data linkage: networks with different ecologies tend to have their own structural signatures beyond the mutualistic/antagonistic dichotomy, and the ANN classifier at least partially learns to identify each type of ecology rather than to discriminate mutualistic versus 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 antagonistic networks. In fact, after controlling for the type of ecology, the only significant difference between mutualistic and antagonistic networks was the frequency of a few motifs with asymmetrical specialization (e.g., motifs 17 and 44), suggesting that a significant part of the structural singularities comes from the type of ecology rather than the type of interaction (antagonistic versus mutualistic). These results on the importance of the type of ecology corroborate previous studies that found, for instance, significant structural differences between pollination and seed dispersal networks [44] or between mutualistic plant-fungus and plant-animal networks [45]. Interestingly, we found herbivory and pollination networks to have similar structural properties in both spectral density and motifs frequency, which may explain the low classification rate of herbivory networks. In terms of the ability to classify mutualistic and antagonistic networks, our results on ecology suggest that a better classification of networks of a particular type of ecology will be achieved by using networks from the same ecology during the training of the classifiers.

size and quality of the dataset used for training. We used one of the largest publicly available databases of empirical networks, but it contained only 343 networks in total. In addition, it had a strong unbalanced representation of the different types of ecology, with 80% of the mutualistic networks represented by pollination networks and 60% of the antagonistic networks represented by host-parasite networks. There was also a strong imbalance in the representation of interaction intimacy, which is known to impact network structure [5,17,23,46]: mutualistic networks were mainly low intimacy networks, except ant-plants (n=3) or mycorrhizal networks (n=9), and antagonistic networks were mainly high intimacy networks, except herbivory networks (n=23). As a consequence, only ~60% of the high-intimacy mutualistic networks and ~40% of the low-intimacy antagonistic networks were correctly classified. These results confirm that interaction intimacy may confer a particular structure to ecological networks [5,17]. Increasing the number of empirical networks, especially mutualistic networks 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 with high intimacy (e.g. symbiotic networks) and antagonistic networks with low intimacy (e.g. trophic networks) would improve further our ability to classify ecological networks by better representing the large structural heterogeneity of ecological networks [20]. Another way to improve the classification would be to consider interaction weights (the abundance of interacting individuals in a pair of interacting species). Such information can emphasize structural differences that are not contained in binary networks [47,48]. Unfortunately, the database of weighted empirical networks is even smaller than that of binary networks.

The efficiency of supervised classification approaches can also strongly depend on the type, architecture, and tuning of the classifier. For example, the overfitting we noticed in the ANN (which did not seem to negatively impact our results) could potentially be avoided by using other regularization technics, such as dropout [49]. We chose to use simple, few-layers neural networks implemented in R for their availability and ease of usage to the large community of R users, as well as because our dataset for training was rather small. However, several other machine learning tools exist and could be tested, although they may require larger training sets; graph neural networks in particular have been specifically designed to learn information from graphs such as ecological networks [50].

training data thanks to simulated data. We investigated this possibility with BipartiteEvol, an individual-based model that mimics the eco-evolutionary emergence of bipartite interaction networks [38]. However, ANN classifiers trained on BipartiteEvol simulations were not able to discriminate mutualistic and antagonistic empirical networks, suggesting that the model somehow fails at capturing important structural differences between antagonistic and mutualistic networks. [38] showed that the model produces realistic networks, which we confirm here, and is able to reproduce major global differences observed between mutualistic and antagonistic networks. However, we find that BipartiteEvol simulations tend to have higher 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 frequencies of complex motifs and a lower connectance compared to empirical networks. Consequently, their projection only covers a small proportion of the space occupied by empirical networks. One possible explanation is the continuum between mutualism and antagonism in many interactions found in natural communities [51]. For instance, it has been shown that parrot-plant networks can combine the two types of interactions [51]. Similarly, pollination or mycorrhizal networks often contain cheating species that use antagonistic strategies and induce structural changes within the interaction networks [52–54]. Interactions can also shift from mutualism to antagonism on short timescales, depending on some external conditions [55,56]. By contrast, BipartiteEvol models the emergence of a network in strict antagonistic or mutualistic communities. This calls for future efforts to improve the BipartiteEvol model. Simulations under other generative models (e.g., 57,58) or data augmentation methods could also be tested.

Ultimately, supervised classifying approaches could be particularly useful to predict the type of undetermined interactions. For instance, many plant-fungus interaction networks are assembled from high-throughput sequencing data, but we currently ignore the mutualistic versus the antagonistic type of many of these endophytic interactions [31,32]. Similarly, an increasing number of networks are built from co-occurrence data generated with high-throughput sequencings, such as the ones between microscopic planktonic species sampled by the Tara Oceans expeditions [34,35]. The ecology of the majority of species in these networks is currently undetermined [35,36,59]. However, a preliminary analysis of bipartite networks built from co-occurrence data in the Tara Oceans data revealed that these networks were structurally very different from the empirical and simulated networks we used here (results not shown). Classifying these networks would thus require using a different training dataset. A potential source of difference might come from species delineation in microbial lineages in such databases, which rely on the clustering into operational 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 taxonomic units (OTUs) of short marker gene sequences (e.g. the small subunit rRNA gene or the ITS marker). This difference in species delimitation in networks involving microbial groups compared to networks involving macro-organisms (e.g. pollination and herbivory) [19] may affect network structure. This may explain why the classification of mycorrhizal networks was not very accurate in our study. One approach to tackle this would be to incorporate more microbial networks with known interaction types in the training of the supervised classifier, which would be highly recommended anyway given the importance of the type of ecology in the classification accuracy. Another source of differences comes from the use of co-occurrence as a proxy for interaction, a topic that is currently intensively debated [60]. Classifiers would need to be trained with co-occurrence networks with known interaction types or simulated co-occurrence networks.

Conclusion

Mutualistic and antagonistic networks have structural differences that can be particularly well captured by motif frequencies. Antagonistic networks are enriched in densely connected motifs, while mutualistic networks are enriched in motifs with asymmetrical specialization. Besides this dichotomy, there are structural differences across networks with distinct ecologies. These structural differences can be used to classify bipartite interaction networks with supervised machine learning with pretty good accuracy. These performances could be further improved by increasing the number of empirical networks used for training, diversifying the type of ecology sampled, and improving currently available models to simulate networks. While more work is needed to predict interaction types for co-occurrence networks, our analyses provide promising results on the ability to predict interaction types from network structure.

on different network scales (Fig. 1).

First, we characterized the macro-scale structure of the networks using 5 global metrics: the total size (number of species in both guilds), the absolute difference in the number of species in the two guilds, the connectance, the nestedness, and the 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 modularity. We quantified nestedness using the NODFc index that corrects for network size (<NODFc= function from the R-package maxnodf, version 1.0.0 [66]). We quantified modularity using the function computeModules from the R-package bipartite (version 2.16) [67] with the <Beckett= method.

Second, we characterized the mesoscale structure of the networks by performing bipartite motif analyses that depict the patterns of interactions between subsets of species (called motifs; Fig. 1) [37]. For each network, we computed the frequencies of the 44 possible motifs involving between 2 and 6 species using the function mcount from the bmotif R-package (version 2.0.2) [68] with the normalization procedure <normalise_sum= (i.e. which computes the frequency of motifs among the total number of motifs).

Third, we used a multi-scale approach stemming from graph theory that integrates the whole structural information of a network, from micro- to macro-scales: the spectral density of the Laplacian graph [69]. We computed the normalized Laplacian graph of each network [69], defined as Lnorm = D-1/2 L D-1/2, where D is the degree matrix (a diagonal matrix counting the number of interactions for each species) and L is the Laplacian graph, defined as the difference between the adjacency matrix (a square binary matrix whose elements represent the interactions between species of the network) and the degree matrix (Supplementary Fig. 2A). The normalized Laplacian graph is a symmetric and positive-definite matrix whose eigenvalues are constrained between 0 and 2 [70], which makes them comparable between networks of different sizes, and symmetrical around 1. There are as many eigenvalues equal to 0 (or 2) as there are connected components (i.e. perfect modules of species that are not connected with any species from other modules), and we can thus expect a modular network to have more eigenvalues close to 0 (and 2) than a nested network. To be able to efficiently compare Laplacian graphs across networks of different sizes, we converted the eigenvalues into a spectral density using a Gaussian convolution [29](Supplementary Fig. 2A). To do so, we adapted the spectR function from the RPANDA R-package 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 (version 1.9) [71] for bipartite networks and used a bandwidth of 0.025 for the Gaussian convolution. We tested the influence of the bandwidth value (from 0.001 to 0.091) on the resulting spectral densities. We characterized each network by extracting its spectral density values at 100 points uniformly distributed between 0 and 1.

In sum, we characterized the structure of each network with 149 variables: 5 global metrics, 44 motif frequencies, and 100 values extracted from its Laplacian spectral density. For each metric of network structure, we used a separate Wilcoxon-MannWhitney test to assess whether antagonistic versus mutualistic networks were significantly different according to that metric. In addition, we accounted for the different types of ecology by performing mixed-effects models (function lmer from the R-package lme4 (version 1.1-27)) with the type of interaction (i.e. mutualistic or antagonistic) as a fixed effect and the type of ecology as a random effect (Type II Wald test). We also investigated which metric significantly differed between each type of ecology by performing an ANOVA for each metric (Fisher test). Finally, we analyzed whether the differences in structural metrics were better explained by the type of ecology or the type of interaction by comparing the Akaike information criterion (AIC) of the different models (see Supplementary Table 1).

Unsupervised classification

component analyses (PCA), with networks characterized either in terms of their global metrics, motif frequencies, or Laplacian spectral densities. PCAs were performed with the functions PCA and fviz_pca_ind from the FactoMineR and factoextra R-packages (versions 2.4 and 1.0.7 respectively) [72].

classification on the two main principal components projections (see Fig. 3, Supplementary Tables 3). K-means were performed with the function kmeans with 2 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 clusters. We investigated whether the two clusters tend to separate mutualistic from antagonistic networks, and networks with different types of ecology, by performing chi-squared tests on the K-means classifications. To quantify the ability to separate antagonistic and mutualistic networks, we used a measure of classification accuracy computed as the F-score with the function F1_score from the R-package MLmetrics. The F-score is the harmonic mean of the precision (e.g. the frequency of true mutualistic among predicted mutualistic networks) and the recall (e.g. the frequency of predicted mutualistic among true mutualistic networks).

Supervised classification

Next, we used supervised classifiers to classify antagonistic versus mutualistic networks based on their structure. As input for the supervised classifier, for each interaction network, we used a set of structural variables, either its global metrics, its motif frequencies, its Laplacian spectral density, or any combination of these sets of variables. As an output, the supervised classifier provides an interaction type: <mutualistic= or <antagonistic=. Supervised classifiers are based on supervised learning: the classifiers have first to be trained with a random subset of the interaction networks (the <training set=) in order to optimize its internal parameters; once trained, it can classify the other networks (the <test set=) from the input variables. The percentage of networks from the test set that are accurately classified by the classifier indicates the performance of the approach.

We first investigated the performance of linear classifiers using logit and Lasso regressions. In Lasso regression, a constraint on the estimated model coefficients can make the regression with a high number of variables more reliable. Logit regressions were computed with the glm function, whereas the Lasso regressions were performed with the function cv.glmnet from the glmnet R-package (version 4.1-2) [73]. We 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 included 80% of the networks in the training set and replicated the classifications 50 times (with different training sets).

Second, we used artificial neural networks (ANN) implemented in the neuralnet function from the R-package neuralnet (version 1.44.2) [74]. We optimized their configuration (number of layers and number of neurons in each layer) according to the input variables. We selected a 3 layers network, constituted of the input and output layers and an intermediate layer with 5, 20, and 60 neurons for global metrics, motif frequencies, and spectral density inputs respectively. Preliminary analyses showed that these choices provided a good compromise between classification accuracy and overfitting (results not shown). We also selected a non-linear logistic transformation, and the maximal number of steps was fixed at 108. In addition, we found that training the ANN (i.e. optimizing neuron weights to reduce classification bias) with 80% of the interaction networks and testing it on the remaining 20% provided a good trade-off between the size of the test set and the accuracy of the classification.

We independently trained and tested the classifiers on empirical networks using the different sets of structural input variables. For each configuration, the ANN trainings were replicated 50 times with different training and test sets, and we reported the average percentages of correct classifications among antagonistic and mutualistic networks and their standard deviations. By comparing the performance of each classifier (percentage of correct classification and classification accuracy), we were therefore able to select the best set of input structural variables and the best classification approach to discriminate antagonistic and mutualistic networks.

Moreover, because the training sets were limited in size (we only had a total of 343 empirical networks), we introduced a measure of the classification confidence for each empirical network. We selected the most efficient classification approach (ANN on motif frequencies; see Results) and we trained and tested it for 10,000 independent training and test sets of empirical networks: each empirical network was therefore classified on average 2,000 times with different training sets. We then computed the 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 percentage of classifications that fell in each category, which gives an idea of classification confidence: we considered that a given network was classified as mutualistic (resp. antagonistic) with <high confidence= if it was classified as mutualistic (resp. antagonistic) in more than 75% of the classifications. Conversely, a given network was classified as mutualistic (resp. antagonistic) with <low confidence= if it was classified as mutualistic (resp. antagonistic) in more than 50% of the classifications but in less than 75% of them.

Next, we verified that our results were not impacted by the over-representation of mutualistic versus antagonistic networks in the training set (more than 60% of the networks were mutualistic). To do so, we trained the classifier with the same number of antagonistic and mutualistic networks (from 5 to 95 networks for each type of interaction) and compared the classification accuracy with the one obtained with the same-size training sets composed of randomly chosen networks. In each condition, we replicated the classification 50 times with different training sets.

Finally, we investigated if the classifiers were strictly learning to separate antagonistic from mutualistic networks or if they also learned the potential structural singularities of each type of ecology from our database. To do so, we removed one by one each type of ecology from the training test (e.g. pollination networks), we trained the classifiers on 50 independent training sets constituted by 80% of the remaining networks, and we (i) tested the classification on the 20% remaining networks (e.g. are other networks still as well classified as when pollination networks were not used in the training?) and (ii) tested the classification on the networks from the removed type of ecology (e.g. are pollination networks classified among mutualistic networks when no pollination network have been used in the training?). If the classifiers learn to separate antagonistic from

mutualistic networks regardless of ecology, the classifications of tests (i) and (ii) should perform as well as the original classification. If they learn to classify ecologies rather than the type of interaction, then test (ii) should perform very badly, and test (i) may be less accurate. Lastly, we also verified whether 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 our results were impacted by the over or under-representation of some types of ecology in the training set (e.g. the pollination networks represent ~50% of the complete database). To do so, we trained the classifier with the same number of networks per type of ecology (from 5 to 20 networks for each type of ecology; if a type of ecology had fewer networks than the number per type of ecology, all networks were included in the training set) and compared the classification accuracy with the one obtained with same-size training sets composed of randomly chosen networks (see Supplementary Fig. 8B). If the classifiers learn to separate antagonistic from mutualistic networks regardless of ecology, then an over- or under-representation of some ecologies should not impact the results.

Simulations of bipartite interaction networks

The accuracy of supervised classifiers often depends on the size of the available training set (i.e., data for which the classification is known), which in the case of bipartite interaction networks is rather small (e.g. we were able to obtain 343 networks here). A possible approach to increase the training set is to use simulated data. We explored this possibility with the recently-developed BipartiteEvol model [38], an ecoevolutionary model that simulates the evolution of interacting individuals within two clades. We simulated 1,200 weighted networks using a wide range of antagonistic and mutualistic parameters (see [75], for details). We noticed that connectance values for these networks were on average lower than those of the empirical networks; to approach those of the empirical networks, we retained 70% randomly sampled individual interactions in the simulated networks, which had a direct effect of deleting rare, specialized species and increasing network connectance. We also carried out analyses with the original networks (no sampling), which did not qualitatively affect the results (results not shown). Retaining networks that had at least 10 species in each guild, we obtained a total of 1,175 binary networks, including 705 mutualistic and 470 antagonistic networks, hereafter referred to as BipartiteEvol networks. 653 654 655 656 657 658 659 660 661 662

We performed unsupervised classification using PCA and K-means. We also performed supervised classification: we independently trained and tested the supervised classifiers (Logit or Lasso regressions, or ANN) on BipartiteEvol networks using the different sets of structural input variables. We then tested the classifier on the empirical networks. As the ANN trained on BipartiteEvol simulations performed poorly to classify empirical networks, we investigated differences between networks simulated with BipartiteEvol and empirical networks by comparing the distributions of their global metrics (in terms of network size, connectance, nestedness, and modularity), their motifs frequencies, and the spectral densities of their Laplacian graph. 664 665 666 667 668

Writing – original draft: Benoît Pichon, Hélène Morlon, and Benoît Perez-Lamarque

connectance, nestedness, modularity)

S p e c t r a l d e n s i t y 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 network structure. (A) Connectance, nestedness (NODFc), and modularity (Newman9s index) of antagonistic and mutualistic empirical networks. Stars indicate the statistical significance of the difference between antagonistic and mutualistic networks, as measured by a Wilcoxon-Mann-Whitney test (*** = p-value < 0.01, ns = non-significant). (B) Densities of eigenvalues from the Laplacian spectral density that differ significantly between antagonistic and mutualistic networks (Wilcoxon-Mann-Whitney test: p-value<0.01). The discontinuities in the x-axis are due to the filtering of eigenvalues densities that differ significantly between antagonistic and mutualistic networks. (C) Frequencies of the motifs that differ significantly between antagonistic and mutualistic networks (Wilcoxon-Mann-Whitney test: p-value<0.01). Each motif is illustrated at the top of each boxplot (following the denomination of Simmons et al. 2018).

Mutualistic networks are shown in green and antagonistic ones in yellow. Boxplots present the median surrounded by the first and third quartiles, and whiskers extend to the extreme values but no further than 1.5 of the interquartile range. tend to discriminate mutualistic versus antagonistic networks at the mesoscale. Projection of the 343 empirical networks on the two principal components (PC1 and PC2) obtained using principal coordinate analysis (PCA) on (A) global metrics (i.e. nestedness, modularity, connectance, and network size), (B) spectral density of the Laplacian graph and (C) motif frequencies. The percentage of explained variance is indicated on each axis. Each associated table corresponds to the contingency table of unsupervised K-means classification based on the global metrics (A), motifs frequencies (B), and spectral density of the Laplacian graph (C). Chi-squared test associated with the K-means classification is indicated.

Unsupervised classification

Cluster 1 Cluster 2 0.63 χ² = 15.572, p-value<0.001 0.55 796 802 805 810 813 Unsupervised classification

Cluster 1 Cluster 2 F-scores Unsupervised classification

Cluster 1 Cluster 2 χ² = 55.093, df = 1, p-value <0.001 χ²= 23.607, p-value < 0.001 37 179 0.81 0.48 0.51 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 10. 11. 12. 13. 14.