Create Graph

Created: 24-08-2018.Last Update: 18-07-2026.

The Scope allows you to create edges between the source nodes (Objects in the selection of the currently active Type) and target nodes (Objects referenced by the indicated path in the Scope).

Network

All reported network measures per node read the same way: a higher value means more, e.g. more connected, more central, more influential. A value of 0 means the node is part of the graph but has no relevance to the measure. A non-existent value (-) means there was nothing to calculate for the node. Every entry below states what the node values mean (Primary, and Secondary when the algorithm reports a second value) and which summary figures the algorithm adds to the network overview (Statistic).

The graph follows from configuration using the Scope. It is directed by default, or undirected simply when both directions are set. The graph is always one-mode: when nodes of one Type are connected because they share a node of another Type, that shared node acts as a collapsed in-between and does not become a node of the graph itself; the connection is projected between the two nodes as a single co-occurrence. Such a projected (resolved) edge counts once, being a single shared fact written both ways. A direct Type-to-Type edge (native) sums its two directions, each a relationship the data asserts in its own right, so a mutual edge counts stronger than a one-way one. An edge is unweighted, or weighted as a closeness or as a distance; how a weight is built and how each family of algorithms reads it is described in the Weight section below.

The currently available network algorithms are:

Algorithm	Description
Connections
Count	The number of edges running from a node to other nodes, or their total edge length on a weighted graph. The most direct measure of activity: it counts connections, not their importance. More info. Returns Primary: The number of edges of the node, or their total edge length when weighted. In an events network, the count simply ranks who took part in the most gatherings, before any question of how well-placed those gatherings were.
Shortest Path	The shortest path between one or a group of nodes (From) and any other node or a group of nodes (Target). More info. Options Centrality: weighs every node on these paths by betweenness centrality (absolute, relative, or normalised), showing which nodes the connection depends on. Tolerance: (detour/slack) widens the centrality to near-shortest (ε-shortest) paths: a node is included when the best path via it is at most the tolerance longer than the shortest path (in weights, or steps when unweighted), showing which nodes would carry the connection if the shortest paths were unavailable. The tolerance applies when a Target is set. Returns Primary: The node's distance from the From selection: the length of the shortest path to it, in weights or steps when unweighted; From nodes report 0, nodes on no path report nothing. Secondary: With the centrality enabled, how strongly the connection depends on the node: the number of best routes between From and Target that run through it, or its betweenness for the paths from the From selection to the whole network when no Target is set. In a correspondence network, the shortest path from a scholar to the court shows through whose hands a request most plausibly travelled; the tolerance adds who could have carried it had a connection failed. Literature Dijkstra, E. W. (1959). 'A note on two problems in connexion with graphs'. Numerische Mathematik 1, 269-271. Freeman, L. C. (1977). 'A set of measures of centrality based on betweenness'. Sociometry 40(1), 35-41. Brandes, U. (2001). 'A faster algorithm for betweenness centrality'. Journal of Mathematical Sociology 25(2), 163-177.
Disjoint Paths	The largest set of routes between a selection of nodes (From) and another selection (Target) that share no edge: each route is an independent chain of documented relationships, so the count states how robust the connection between the two selections is and how many relationships must fail before it breaks (Menger's edge connectivity). Where the shortest path finds the single best route, this finds how many ways the connection survives. The number of routes depends only on which edges exist; among all largest sets, the reported one has the least total distance, so the routes shown are the most plausible representatives. More info. Returns Primary: The number of disjoint routes that run through the node; the From and Target nodes carry their own routes. Secondary: The node's earliest step along a route, counted from the From side (a position, not a score). Statistic: Paths: the number of independent routes between the selections. Distance: the total distance of that set of routes. In a relationship network, disjoint paths show whether the contact between two milieus rests on a single go-between or on several independent chains of acquaintance. Literature Menger, K. (1927). 'Zur allgemeinen Kurventheorie'. Fundamenta Mathematicae 10, 96-115. Ford, L. R. & Fulkerson, D. R. (1956). 'Maximal flow through a network'. Canadian Journal of Mathematics 8, 399-404. Suurballe, J. W. & Tarjan, R. E. (1984). 'A quick method for finding shortest pairs of disjoint paths'. Networks 14(2), 325-336. Bhandari, R. (1999). Survivable Networks: Algorithms for Diverse Routing. Kluwer Academic Publishers.
Link Prediction	Scores every node on how plausible an undocumented connection to a selection of nodes (From) is, based on the connections they share. Only nodes without a documented connection to the From selection are scored, making this a detector for connections the sources never state directly. On a weighted graph the edge strengths take over from the counts (Lü & Zhou), so a shared connection over two strong edges weighs more than one over two weak edges. More info Options Measure: sets how shared connections are weighed: Shared Connections (Common Neighbours): as a plain count. Relative Shared Connections (Jaccard): relative to all connections of both nodes. Selective Shared Connections (Adamic-Adar): discounting shared connections softly the more connections they have themselves. Exclusive Shared Connections (Resource Allocation): discounting shared connections strongly by spreading their evidence evenly over all their connections, so only sparsely connected in-between nodes carry real weight. Mode: the plausibility scores as they are (absolute), or rescaled between 0 and 1 so the most plausible node is 1 (normalised). Returns Primary: The node's best plausibility score against the From selection; nodes with a documented connection score 0. Secondary: The number of shared connections behind that score. Statistic: Candidates: the number of nodes scoring above 0. Best: the highest score found. In a correspondence network, link prediction flags two scholars who never appear in the same source yet share most of their correspondents: an acquaintance the letters imply but do not state. Literature Liben-Nowell, D. & Kleinberg, J. (2007). 'The link-prediction problem for social networks'. Journal of the American Society for Information Science and Technology 58(7), 1019-1031. Adamic, L. A. & Adar, E. (2003). 'Friends and neighbors on the Web'. Social Networks 25(3), 211-230. Zhou, T., Lü, L. & Zhang, Y.-C. (2009). 'Predicting missing links via local information'. The European Physical Journal B 71, 623-630. Lü, L. & Zhou, T. (2010). 'Link prediction in weighted networks: The role of weak ties'. EPL (Europhysics Letters) 89(1), 18001.
Katz Prediction	Scores nodes on the paths of every length between the node and the From selection, each path discounted the longer it is (attenuation). Where the other Link Prediction measures need directly shared connections, this prediction can surface nodes whose plausibility rests on deeper structure alone: shown by a high score with few or no shared connections. On a weighted graph a path multiplies the edge strengths along it, so a longer route over strong edges can outscore a short one over weak edges. More info Options Iterations: the maximum path length taken into account. Attenuation: how strongly longer paths are discounted; lower values keep the prediction local, higher values let distant structure count. Mode: the plausibility scores as they are (absolute), or rescaled between 0 and 1 so the most plausible node is 1 (normalised). Returns Primary: The node's best plausibility score against the From selection; nodes with a documented connection score 0. Secondary: The number of directly shared connections behind that best pair, which can be 0 when the plausibility rests on longer paths alone. Statistic: Candidates: the number of nodes scoring above 0. Best: the highest score found. In a book network, Katz prediction can surface a printer and an author who share no acquaintances but sit in tightly interlinked circles, a connection worth checking the sources for. Literature Katz, L. (1953). 'A new status index derived from sociometric analysis'. Psychometrika 18(1), 39-43. Liben-Nowell, D. & Kleinberg, J. (2007). 'The link-prediction problem for social networks'. Journal of the American Society for Information Science and Technology 58(7), 1019-1031.
Centrality - Influence
Degree Centrality	The summed strength of a node's edges, per direction: how much of the network's documented activity attaches to the node directly. Unweighted every edge is 1, so the values are the plain number of incoming and outgoing edges; on a weighted graph every edge counts by its strength, so a node with a few strong connections can outrank one with many weak ones. More info. Options Mode: the summed strengths as they are (absolute), or the incoming and outgoing sums each rescaled between 0 and 1 so the highest node is 1 (normalised). Returns Primary: The summed strength of the edges running to the node (in-degree). Secondary: The summed strength of the edges running from the node (out-degree). In a correspondence network, the in-degree ranks who received letters from the most correspondents and the out-degree who wrote the most widely; weighted, a sustained exchange counts for more than a single letter. Literature Freeman, L. C. (1979). 'Centrality in social networks: Conceptual clarification'. Social Networks 1(3), 215-239. Opsahl, T., Agneessens, F. & Skvoretz, J. (2010). 'Node centrality in weighted networks: Generalizing degree and shortest paths'. Social Networks 32(3), 245-251.
Eigenvector Centrality	The influence of a node based on the influence of the nodes linking to it: a few well-placed connections can outweigh many marginal ones. On a directed graph only incoming edges carry influence, so a node nothing points to scores 0; use Katz centrality when such nodes should still rank. More info. Options Iterations: how long the scores are passed around before they are read. Returns Primary: The node's influence, scaled so the most influential node is 1. Statistic: Iterations: the number run. Change: how much the scores still moved in the last iteration; near 0 means they have settled, larger means more iterations would still move them. In a patronage network, a courtier with three well-placed patrons outranks a merchant with thirty ordinary contacts. Literature Bonacich, P. (1972). 'Factoring and weighting approaches to status scores and clique identification'. Journal of Mathematical Sociology 2(1), 113-120. Bonacich, P. (1987). 'Power and centrality: A family of measures'. American Journal of Sociology 92(5), 1170-1182.
Katz Centrality	Eigenvector centrality in which every node keeps a base score, so nodes without incoming paths still rank; the attenuation factor discounts influence over longer paths. More info. Options Iterations: the maximum path length influence can travel. Attenuation: how strongly influence fades with every step; lower values keep the measure close to counting direct connections, higher values let the whole network weigh in. Returns Primary: The node's influence, scaled so the most influential node is 1. Statistic: Iterations: the number run. Change: how much the scores still moved in the last iteration; near 0 means they have settled. In a correspondence network, Katz centrality still ranks the peripheral figures: a copyist reached through a single chain of intermediaries keeps a score where eigenvector centrality would give none. Literature Katz, L. (1953). 'A new status index derived from sociometric analysis'. Psychometrika 18(1), 39-43.
PageRank	The importance of a node based on the number and quality of the edges leading to it: each node divides its own rank over its outgoing edges, so being linked by a selective, important node counts most. More info. Options Iterations: how often the rank is passed on. Damping: the share of rank that follows the edges; the remainder is the base score every node keeps, which stops rank from pooling in closed loops. Returns Primary: The node's rank, on the scale of the original paper: a node nothing points to keeps the base score of 1 minus the damping (0.15 by default), and the scores average around 1 over the network. Statistic: Iterations: the number run. Change: how much the scores still moved in the last iteration; near 0 means they have settled. In a citation or dedication network, PageRank ranks a work named by a few much-named works above one named often by works nobody names. Literature Brin, S. & Page, L. (1998). 'The anatomy of a large-scale hypertextual Web search engine'. Computer Networks and ISDN Systems 30(1-7), 107-117.
Hyperlink-Induced Topic Search (HITS)	Two scores per node: authority (linked to by good hubs) and hub (linking to good authorities). Useful in directed networks, where pointing and being pointed at mean different things. More info. Options Iterations: how often the two scores feed each other. Returns Primary: The node's authority score, scaled so the highest is 1. Secondary: The node's hub score, scaled so the highest is 1. Statistic: Iterations: the number run. Change: how much the two scores still moved in the last iteration; near 0 means they have settled. In a correspondence network, HITS separates the great senders from the great receivers: the secretary who writes to every authority is a hub, the authority every hub writes to is an authority. Literature Kleinberg, J. M. (1999). 'Authoritative sources in a hyperlinked environment'. Journal of the ACM 46(5), 604-632.
Centrality - Distance
Closeness Centrality	How near a node is to the whole network: closeness weighs every node by its distances to all other nodes. More info. Options Mode: 1 divided by the summed distances, in which an unreachable node counts as a distance longer than any real path (absolute); or rescaled so 1 means directly connected to every other node, comparable across networks, where nodes in small disconnected parts rank low (normalised). Returns Primary: How near the node is to the whole network: higher means nearer. In a travel or correspondence network, closeness centrality identifies from whom news could reach everyone in the fewest steps. Literature Bavelas, A. (1950). 'Communication patterns in task-oriented groups'. Journal of the Acoustical Society of America 22(6), 725-730. Sabidussi, G. (1966). 'The centrality index of a graph'. Psychometrika 31(4), 581-603. Wasserman, S. & Faust, K. (1994). Social Network Analysis: Methods and Applications. Cambridge University Press.
Closeness Eccentricity	How near a node is to the farthest node it can reach. Eccentricity identifies the centre (highest values) and periphery of each connected part, but does not weigh disconnection: use closeness centrality to compare across parts. More info. Returns Primary: 1 divided by the largest distance the node can reach. In a travel network, eccentricity finds the natural centre: the place from which no destination is truly remote. Literature Hage, P. & Harary, F. (1995). 'Eccentricity and centrality in networks'. Social Networks 17(1), 57-63.
Betweenness Centrality	The number of times a node lies on the shortest paths between other nodes: bridges and gatekeepers, important for connecting others rather than being connected. More info. Options Mode: the plain path count (absolute), the share of all node pairs the node could sit between, comparable across networks (relative), or rescaled between 0 and 1 within this result (normalised). Returns Primary: The number of shortest paths between other nodes that run through the node. In a book-trade network, betweenness centrality finds the agent through whom two otherwise separate circles of printers and buyers had to pass. Literature Freeman, L. C. (1977). 'A set of measures of centrality based on betweenness'. Sociometry 40(1), 35-41. Brandes, U. (2001). 'A faster algorithm for betweenness centrality'. Journal of Mathematical Sociology 25(2), 163-177.
Structure & Cohesion
Clustering Coefficient	The degree to which the neighbours of a node are also connected to each other. Both ends inform: a well-connected node with a low coefficient brokers between groups that do not know each other. On a weighted graph a closed triangle counts by the strength of the node's edges into it (Barrat et al.), so a cluster held together by strong edges weighs more. More info. Returns Primary: The node's coefficient: 0 means none of its neighbours are connected to each other, 1 means they form a complete cluster. Statistic: Average: the mean coefficient over the analysed nodes, the overall clumping of the network. In a relationship network, the clustering coefficient distinguishes a salon, where everyone knows everyone, from a patronage web that only holds together through its patron. Literature Watts, D. J. & Strogatz, S. H. (1998). 'Collective dynamics of "small-world" networks'. Nature 393, 440-442. Barrat, A., Barthélemy, M., Pastor-Satorras, R. & Vespignani, A. (2004). 'The architecture of complex weighted networks'. Proceedings of the National Academy of Sciences (PNAS) 101(11), 3747-3752.
k-Core	The deepest level k at which a node still belongs to a group in which every member has at least k edges: high core numbers sit in the dense core of the network, low ones on the periphery. On a weighted graph the level is a summed edge strength instead of a count (the s-core), so a member held in by a few strong edges ranks with one held in by many weak ones. More info. Returns Primary: The node's core number, a strength level when weighted. Secondary: The node's degree, or its summed edge strength when weighted. Statistic: Degeneracy: the deepest core level the network holds. In an events network, the k-core peels the incidental attendees away layer by layer until the standing core of the scene remains. Literature Seidman, S. B. (1983). 'Network structure and minimum degree'. Social Networks 5(3), 269-287. Batagelj, V. & Zaveršnik, M. (2003). 'An O(m) algorithm for cores decomposition of networks'. arXiv:cs/0310049. Eidsaa, M. & Almaas, E. (2013). 'S-core network decomposition: A generalization of k-core analysis to weighted networks'. Physical Review E 88, 062819.
Articulation Points & Bridges	The nodes and edges that hold the network together: an articulation point is a node whose removal breaks a connected part into pieces, a bridge is an edge that does the same. Direction and weights play no role: this is about whether the structure holds, not how strongly. More info. Returns Primary: The number of extra pieces the node's removal would create; 0 means the network stays whole without it. Secondary: The number of bridges the node sits on. Statistic: Points: the number of articulation points in the network. Bridges: the number of bridges. A single documented marriage holding two family networks together is an articulation point: a warning that everything said about their connection rests on a single relationship in the sources. Literature Hopcroft, J. & Tarjan, R. (1973). 'Algorithm 447: Efficient algorithms for graph manipulation'. Communications of the ACM 16(6), 372-378.
Connected Components	Which island of the network a node belongs to, to read before any distance or centrality is compared across islands. The weak component ignores the direction of the edges; the strong component requires every member to reach every other following the arrows, so the strong components are the circulation loops inside the weak islands, and the two coincide when every edge is mutual. More info. Returns Primary: The node's weak component. The components are numbered by size, 1 being the largest; like the community identifiers these numbers identify, they do not score. Secondary: The node's strong component, numbered the same way. Connected components tell whether the sources describe one connected world or several separate ones; in a correspondence network the strong components show between whom letters actually travelled both ways. Literature Tarjan, R. (1972). 'Depth-first search and linear graph algorithms'. SIAM Journal on Computing 1(2), 146-160.
Brokerage	Whether a node's contacts are connected around it or only through it (Burt's structural holes). Effective size counts the contacts discounted by their redundancy: a node whose contacts all know each other keeps an effective size near 1, a node whose contacts are strangers to one another keeps nearly its full number of contacts and is the broker. A node without edges has no ego network and reports nothing. More info. Returns Primary: The node's effective size: higher means more brokerage. Secondary: Burt's constraint, at its comparable literature scale; constraint reads inverted, low marks the broker. Statistic: Average Constraint: the mean constraint over the analysed nodes that hold at least one edge. In a trade network, the bookseller whose buyers and suppliers do not know one another scores a high effective size: every deal runs through the bookseller, and the sources on either side may never mention the other. Literature Burt, R. S. (1992). Structural Holes: The Social Structure of Competition. Harvard University Press. Borgatti, S. P. (1997). 'Structural holes: Unpacking Burt's redundancy measures'. Connections 20(1), 35-38.
Community Detection
Every community analysis divides the network into communities: groups of nodes that belong together by the mode's own standard, stated per entry. All modes read the network undirected except Infomap, which follows the arrows. Options Runs & Seed: a run beyond the first restarts the analysis with the nodes visited in a different order and keeps the best result by the mode's own quality; the seed makes the drawn orders, and so the whole analysis, repeat exactly. At 0 runs the mode picks its own number: 10 for Label Propagation, which must draw at random, 1 for the others, which take the best step every time. With more than one run the runs that are not kept are not lost: the modes that keep a best division also report how much the runs agreed on it, as each node's consensus (secondary value) and a figure over the whole network. Significance: a quality score on its own says little, because these measures score any network, a random one included. The setting compares the found quality against the same analysis on that many randomised networks in which every node keeps its number of edges but loses everything else; the resulting statistics report the mean and spread of the randomised qualities and the z-score and p of the found one, and a large z means the communities are real structure, not what the edge counts alone would produce. Resolution: a higher resolution gives fewer, larger communities; a lower resolution more, smaller ones. 1 runs the measure as the literature defines it. Size: how the size of a node's community is reported as the secondary value: absolute, relative, normalised, or not at all. Returns Primary: The community the node belongs to, numbered 1 up to the number of communities; the numbers are to be used as identifiers, not scores. Secondary: The size of the node's community, as set by the Size option; with more than one run Modularity, Leiden, Leiden CPM and Label Propagation replace it with the node's consensus, how consistently the members of its community stay with it across the runs (1 when they never part, a node between two circles less). Statistical Inference and Infomap replace it as stated in their entries. Statistic: Communities: the number found, reported by every mode. Consensus: with more than one run, the mean consensus over the nodes, how much the runs agreed on the division as a whole. The further figures are the mode's own and stated per entry. Literature Maslov, S. & Sneppen, K. (2002). 'Specificity and stability in topology of protein networks'. Science 296(5569), 910-913. Lancichinetti, A. & Fortunato, S. (2012). 'Consensus clustering in complex networks'. Scientific Reports 2, 336.
Modularity	Divides the network into communities in which edges are denser than expected by chance, with the classic Louvain search. Louvain can return communities that internally fall apart into disconnected pieces; it is kept for continuity and comparison, and Leiden is the recommended mode for new analyses. More info. Returns Statistic: Modularity: the achieved modularity, both as Newman defines it (roughly 0.3-0.7 signals clear community structure) and at the chosen resolution. In a correspondence network, modularity divides the writers into circles that write mostly among themselves, the epistolary equivalent of schools and milieus. Literature Newman, M. E. J. & Girvan, M. (2004). 'Finding and evaluating community structure in networks'. Physical Review E 69, 026113. Blondel, V. D., Guillaume, J.-L., Lambiotte, R. & Lefebvre, E. (2008). 'Fast unfolding of communities in large networks'. Journal of Statistical Mechanics: Theory and Experiment 2008(10), P10008. Fortunato, S. & Barthélemy, M. (2007). 'Resolution limit in community detection'. Proceedings of the National Academy of Sciences (PNAS) 104(1), 36-41.
Leiden Modularity	Modularity with the improved Leiden search: every found community is guaranteed to be internally connected, and the search is rerun from its own result until the division stops improving, which is where the method's quality guarantees live. On the same network Leiden finds an equal or better division than Louvain, so prefer it unless comparability with an older Modularity analysis is the point. More info. Returns Statistic: Modularity: the achieved modularity, both as Newman defines it (roughly 0.3-0.7 signals clear community structure) and at the chosen resolution. In a correspondence network, Leiden draws the same circles as modularity but guarantees each circle actually holds together, so no reported community quietly consists of two unrelated groups. Literature Traag, V. A., Waltman, L. & van Eck, N. J. (2019). 'From Louvain to Leiden: guaranteeing well-connected communities'. Scientific Reports 9, 5233.
Leiden CPM	The Leiden search on the Constant Potts Model (CPM), which holds every community against the same fixed standard: an edge density, rather than modularity's comparison with what chance would put inside it. This makes it resolution-limit-free: whether a small tight circle is recognised does not depend on how large the rest of the network is, where modularity merges small circles into their surroundings once the network grows. The resolution is read against the density of the whole network, so 1 asks of a community the density the network itself carries. More info. Returns Statistic: CPM: the achieved quality. Modularity: of the same division, for comparison. Absolute Resolution: the density threshold the resolution amounted to, which is the resolution as the literature and other software state it. In an events network spanning decades, the Constant Potts Model keeps recognising a small persistent circle of five regulars that modularity would swallow into a larger period cluster. Literature Traag, V. A., Van Dooren, P. & Nesterov, Y. (2011). 'Narrow scope for resolution-limit-free community detection'. Physical Review E 84, 016114. Traag, V. A., Waltman, L. & van Eck, N. J. (2019). 'From Louvain to Leiden: guaranteeing well-connected communities'. Scientific Reports 9, 5233.
Statistical Inference	Finds the division that best explains the network's edges, and only reports communities the evidence supports: where modularity always produces a division, statistical inference will not see communities in randomness. It weighs its own evidence, so it needs no randomised comparison networks. The model counts edges, so it reads the network as simple: weights are ignored (the statistics say so) and a mutual pair counts once. More info. Options Samples: sweeps of a sampler around the best division, replacing the community size with the membership certainty per node. Returns Secondary: The size of the node's community; with Samples set, its membership certainty instead: the share of the sampled divisions that agree on its community, so a node between two circles shows as genuinely uncertain rather than silently assigned. Statistic: Description Length: how compactly the division explains the network (in nats). Evidence: what the communities save over describing the network as a single community; at or below 0 the model found no communities worth reporting. With Samples set: the sweeps run, the mean certainty over the nodes, and the acceptance rate of the sampler's merge-split moves. Where every other mode will happily divide a network of chance co-occurrences into "circles", statistical inference reports one community, telling you the sources give no evidence of any grouping at all. Literature Zhang, L. & Peixoto, T. P. (2020). 'Statistical inference of assortative community structures'. Physical Review Research 2, 043271. Peixoto, T. P. (2014). 'Efficient Monte Carlo and greedy heuristic for the inference of stochastic block models'. Physical Review E 89, 012804. Peixoto, T. P. (2019). 'Bayesian stochastic blockmodeling'. In: P. Doreian, V. Batagelj & A. Ferligoj (eds.), Advances in Network Clustering and Blockmodeling, 289-332. Wiley.
Label Propagation	Every node repeatedly adopts the community most common among its neighbours until the labels settle: a node belongs where most of its connections belong. The method draws at random by design, so it always runs several times (the Runs setting) and keeps the division with the highest modularity. Remains practical when the network grows too large for the other methods. More info. Returns Secondary: As it always runs several times, Label Propagation reports the node's consensus across those runs by default (see the shared Returns above), in place of the community size. Statistic: Modularity: of the kept (best) division. On a network of hundreds of thousands of records, label propagation still returns readable circles in the time the other modes need for a fraction of the data. Literature Raghavan, U. N., Albert, R. & Kumara, S. (2007). 'Near linear time algorithm to detect community structures in large-scale networks'. Physical Review E 76, 036106.
Infomap	Communities as the modules of a flow: Infomap follows a random walker over the network and groups the nodes where the walker lingers, judged by how well the modules compress a description of its movements (the map equation). It reads the direction of the edges, so it suits networks in which something actually moves along the arrows; on a fully mutual network it reads the flow as undirected. More info. Options Teleportation: the share of steps the walker jumps to a random node instead of following an arrow, needed to keep a directed flow from pooling in dead ends; 0.15 is the rate of the original paper, and it applies to directed flow only. Returns Secondary: The node's visit rate instead of the community size: the share of the walker's steps spent at the node, its standing within the flow. Statistic: Codelength: the description length of the walker's movements, in bits per step. Saving: what the modules save over having none, this mode's evidence; a saving of 0 means no flow structure worth reporting. Flow: whether the walker followed the arrows or read the network undirected. Teleportation: the rate used, on directed flow. In a travel or citation network, Infomap groups the places or works among which the traffic actually circulates, where the density based modes only see who is connected to whom. Literature Rosvall, M. & Bergstrom, C. T. (2008). 'Maps of random walks on complex networks reveal community structure'. Proceedings of the National Academy of Sciences (PNAS) 105(4), 1118-1123.

Network Statistics

Every analysis closes with the same overview of the network it ran on, next to the statistics of the algorithm itself (stated in the entries above). The distance based figures (diameter, radius, average path length) count unweighted steps following the direction of the edges.

Statistic	Description
Average Degree	The average number of edges per node.
Average Weighted Degree	The average summed edge weight per node, as the weights were provided.
Density	The share of all possible edges that exist: 1 would mean every node is connected to every other.
Reciprocity	The share of edges that also run in the reverse direction: 1 means every edge is mutual and the network is effectively undirected.
Diameter	The longest shortest path in the network: how many steps the two most separated (still connected) nodes are apart.
Radius	The steps needed from the best-placed node to its farthest reachable node; the centre of the network sits at this depth.
Average Path Length	The average number of steps between all pairs of connected nodes: how far apart the network is on average.
Connected Components	The number of islands the network falls into when the direction of the edges is ignored.
Largest Component	The number of nodes in the largest island: close to all nodes means one whole with stray fragments, far below it a real split.
Strongly Connected Components	The number of groups within which every node can reach every other following the arrows.
Largest Strong Component	The number of nodes in the largest such group; when every strong component is a single node, the network holds no cycle at all.

Similarity

With the support of vectors, it is possible to embed vectorised data that results from a machine learning model (e.g. large language models, computer vision, natural language processing) and use metrics for evaluating their similarity. The currently available similarity algorithms are:

Algorithm	Description
Vector distance	The distance between source and target vectors, so vectorised data from a machine learning (ML/AI) model can be compared: the nearer two vectors, the more alike the data behind them. A distance reads the reverse of the network measures above: a smaller value is a nearer, more similar pair. Options Approach: sets what the distance is measured between: Connected: use the Scope to build a network between the vectors and measure only along its edges, so the distance reflects similarity where the network already connects the vectors; the target vector has to be last in the Scope selection. Weighting applies in this approach, read as strength. Disconnected: disregard the network and measure between the selected source and target vectors directly. Calculate: the distance measure: Euclidean, the straight-line distance between the vectors, or Cosine, the angle between them regardless of their magnitude. Mode: the summed distances as they are (absolute), each as its share of the summed distance over all nodes (relative), or rescaled between 0 and 1 within this result (normalised). Returns Primary: The node's summed distance to the target vectors, each weighted by its edge's strength when a weighted mode is set; a smaller value is nearer, more similar. In a corpus of letters embedded by a language model, vector distance ranks which letters read most alike, surfacing near-duplicates and shared sources the catalogue never linked. Literature Salton, G., Wong, A. & Yang, C. S. (1975). 'A vector space model for automatic indexing'. Communications of the ACM 18(11), 613-620. Mikolov, T., Sutskever, I., Chen, K., Corrado, G. & Dean, J. (2013). 'Distributed representations of words and phrases and their compositionality'. Advances in Neural Information Processing Systems 26, 3111-3119.

Algorithm

Description

Vector distance

The distance between source and target vectors, so vectorised data from a machine learning (ML/AI) model can be compared: the nearer two vectors, the more alike the data behind them. A distance reads the reverse of the network measures above: a smaller value is a nearer, more similar pair.

Options

Approach: sets what the distance is measured between:
- Connected: use the Scope to build a network between the vectors and measure only along its edges, so the distance reflects similarity where the network already connects the vectors; the target vector has to be last in the Scope selection. Weighting applies in this approach, read as strength.
- Disconnected: disregard the network and measure between the selected source and target vectors directly.
Calculate: the distance measure: Euclidean, the straight-line distance between the vectors, or Cosine, the angle between them regardless of their magnitude.
Mode: the summed distances as they are (absolute), each as its share of the summed distance over all nodes (relative), or rescaled between 0 and 1 within this result (normalised).

Returns

Primary: The node's summed distance to the target vectors, each weighted by its edge's strength when a weighted mode is set; a smaller value is nearer, more similar.

In a corpus of letters embedded by a language model, vector distance ranks which letters read most alike, surfacing near-duplicates and shared sources the catalogue never linked.

Literature

Salton, G., Wong, A. & Yang, C. S. (1975). 'A vector space model for automatic indexing'. Communications of the ACM 18(11), 613-620.
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. & Dean, J. (2013). 'Distributed representations of words and phrases and their compositionality'. Advances in Neural Information Processing Systems 26, 3111-3119.

Export

Export the computed graph for use in external applications. The graph provides:

Nodes with their name, nodegoat URI, Type, and weight.
Edges with their direction, weight, and time (using temporality (T) configured in the Scope).

The currently available export format is GEXF (Graph Exchange XML Format).

Weight

When applicable to the algorithm you can indicate whether you want to make use of the graph's weight. A weight is the total edge length between two nodes and can be based on:

Duplicate edges becoming the one edge.
Weighting collapsed nodes and source/target nodes using the Conditions.
Combination of both.

A collapsed node shared by two nodes projects a single connection between them (see above); each such shared collapsed node adds 1 + (w_a − 1) + (w_b − 1) to the edge's weight (one for the shared node, plus each side's weight above the default of 1) so unweighted it is simply the number of shared nodes.

Weight	Description
To closeness	A higher weight is a nearer, stronger connection: edges add to a shorter total edge length.
To distance	A higher weight is a farther, weaker connection: edges add to a longer total edge length.

The algorithms read a weight in one of two ways, since what a weight should count for depends on the algorithm's purpose:

Reading Description

As length Sum the weights of the edges along a path, so a near connection is short and a far one long: unweighted an edge is 1 (a step), a closeness is reversed to (w_max + 1) − w against the highest weight in the graph, a distance counts as is. The best routes run over the shortest edges. Shortest Path, Disjoint Paths, Closeness, Eccentricity, Betweenness, and Count.

As strength Weigh every edge on its own, so a near connection is strong and a far one weak: unweighted an edge is 1, a closeness counts as is, a distance is reversed to 1 / w (and not against the highest weight: one far outlier would then shift every strength). A strength depends on the edge alone, so an edge added elsewhere moves nothing. Community detection, the influence centralities (Degree Centrality included), Clustering Coefficient, k-Core, Brokerage, the prediction algorithms, and Vector Distance.

Reading	Description
As length	Sum the weights of the edges along a path, so a near connection is short and a far one long: unweighted an edge is 1 (a step), a closeness is reversed to `(w_max + 1) − w` against the highest weight in the graph, a distance counts as is. The best routes run over the shortest edges. Shortest Path, Disjoint Paths, Closeness, Eccentricity, Betweenness, and Count.
As strength	Weigh every edge on its own, so a near connection is strong and a far one weak: unweighted an edge is 1, a closeness counts as is, a distance is reversed to `1 / w` (and not against the highest weight: one far outlier would then shift every strength). A strength depends on the edge alone, so an edge added elsewhere moves nothing. Community detection, the influence centralities (Degree Centrality included), Clustering Coefficient, k-Core, Brokerage, the prediction algorithms, and Vector Distance.

Create Graph

Network

Returns

Options

Returns

Returns

Options

Returns

Options

Returns

Options

Returns

Options

Returns

Options

Returns

Options

Returns

Options

Returns

Options

Returns

Returns

Options

Returns

Returns

Returns

Returns

Returns

Returns

Options

Returns

Returns

Returns

Returns

Options

Returns

Returns

Options

Returns

Network Statistics

Similarity

Options

Returns

Export

Weight

Documentation