Conference PaperPDF Available


One classic problem definition in social network analysis is the study of diffusion in networks, which enables us to tackle problems like favoring the adoption of positive technologies. Most of the attention has been turned to how to maximize the number of influenced nodes, but this approach misses the fact that different scenarios imply different dif-fusion dynamics, only slightly related to maximizing the number of nodes involved. In this paper we measure three different dimensions of social prominence: the Width, i.e. the ratio of neighbors influenced by a node; the Depth, i.e. the degrees of separation from a node to the nodes perceiv-ing its prominence; and the Strength, i.e. the intensity of the prominence of a node. By defining a procedure to extract prominent users in complex networks, we detect associations between the three dimensions of social prominence and classical network statistics. We validate our results on a social network extracted from the Last.Fm music platform.
The Three Dimensions of Social Prominence
Diego Pennacchioli1, Giulio Rossetti1Luca Pappalardo1, Dino Pedreschi1,
Fosca Giannotti1, and Michele Coscia2
1KDDLab ISTI-CNR, Via G. Moruzzi 1, Pisa, Italy
2CID Harvard University, 79 JFK St, Cambridge, MA, US
Abstract. One classic problem definition in social network analysis is
the study of diffusion in networks, which enables us to tackle problems
like favoring the adoption of positive technologies. Most of the attention
has been turned to how to maximize the number of influenced nodes, but
this approach misses the fact that different scenarios imply different dif-
fusion dynamics, only slightly related to maximizing the number of nodes
involved. In this paper we measure three different dimensions of social
prominence: the Width, i.e. the ratio of neighbors influenced by a node;
the Depth, i.e. the degrees of separation from a node to the nodes perceiv-
ing its prominence; and the Strength, i.e. the intensity of the prominence
of a node. By defining a procedure to extract prominent users in complex
networks, we detect associations between the three dimensions of social
prominence and classical network statistics. We validate our results on a
social network extracted from the Last.Fm music platform.
1 Introduction
One classic problem in social network analysis is understanding diffusion effects
in networks. Modeling diffusion processes on complex networks enables us to
tackle problems like preventing epidemic outbreaks [6] or favoring the adop-
tion of new technologies or behaviors by designing an effective word-of-mouth
communication strategy. In our paper, we are focused on the social prominence
aspect of the diffusion problem in networks.
In the setting of favoring social influence, most of the attention of researchers
has been put on how to maximize the number of nodes subject to the spread-
ing process. This is done by choosing appropriate seeds in critical parts of the
network, such that their likelihood of being prominent users, i.e. nodes that are
active on an innovation before all the other nodes, is maximum, to possibly
achieve larger cascades. While larger cascades are obviously part of the over-
all aim, we argue that it is not the unique dimension of this problem. Three
other dimensions are relevant: the width, the depth and the strength of the so-
cial prominence of any given node in a network. The width of a node is being
prominent for its immediate neighbors; the depth is its ability to be the root of
long cascades; the strength is being the root of an intense activity.
Real-world scenarios focus on specific diffusion patterns requiring a multidi-
mensional understanding of the prominence mechanics at play, along the three
mentioned dimensions. Some examples are: (i) an analyst needs information from
the personal acquaintances of a subject, the important aspect is that many sub-
ject’s direct connections respond, ignoring people two steps away or more; (ii) a
person wants to find another person with a given object, the important aspect
is that some people are able to pass her message through a chain pointing to the
target; (iii) an artist wants to influence people in a social network to her art, the
important aspect is that some people are influenced above the threshold that
will make them aware of the art. In (i) we want a broad diffusion in the first de-
gree of separation. In (ii) we require a targeted diffusion similar to a Depth First
Search. In (iii) there is the need of a high-intensity diffusion. Different scenarios
may require any combination of the three.
In this paper, we make use of three measures to capture the characteristics
of these three scenarios: the Width, Depth and Strength of social prominence.
The Width measures the ratio of the neighbors of a node that follows the node’s
actions. The Depth measures how many degrees of separation there are between
a node and the other nodes that followed its actions. The Strength measures the
intensity of the action performed by some nodes after the leader.
We study what the relations are between these three measures to understand
if we are capturing three orthogonal dimensions of social prominence. We also
study the relations between the Width, Depth and Strength measures and differ-
ent node properties, with the aim of predicting the diffusion patterns of different
events, given the characteristics of the nodes that lead their diffusion.
To validate our concepts, we constructed a social network from the music
platform Last.Fm3, along with the data about how many times and when each
user listens to a song performed by a given artist. We detect who are the promi-
nent users for each artist, i.e. the users who start listening to an artist before
any of their neighbors. We calculate for each prominent user its Width, Depth
and Strength, along with its network statistics such as the degree and the be-
tweenness centrality, looking for associations between them. We then create a
case study to understand what are the different dynamics in the spread of artists
belonging to different music genres, by using the artists’ tags.
To sum up, the contributions of our paper are: (i) a proof that social diffusion
indeed follows at least these three dimensions, which are uncorrelated or anti-
correlated; (ii) the discovery of some significant associations between the three
dimensions of social prominence and some traditional network measures; (iii)
the ability to predict the patterns of diffusion of particular events by looking at
the characteristics of the leaders spreading them.
2 Related Work
In the last decade, there has been growing interest in the studies of diffusion pro-
cesses. Two phenomena are tightly linked to the concept of diffusion: the spread
of biological [6] or computer [17] viruses, and the spread of ideas and innovation
through social networks, the so-called “social contagion” [2], [8]. In both cases,
the patterns through which the spreading takes place are determined not just
by the properties of the pathogen/idea, but also by the network structures of
the population it is affecting.
Some models have been defined to understand the contagion dynamics: the
SIR [11], SIS and SIRS [16] models. The idea behind them is that each individual
transits between some stages in the life cycle of a disease: from Susceptible (S) to
Infected (I), and from Infected to either Recovered (R) or again Susceptible. The
availability of Big Data conveying information about human interactions and
movements encouraged the production of more accurate data-driven epidemic
models. For example, [6] takes into account the spatio-temporal dimension. In
[17], authors study the spreading patterns of a mobile virus outbreak.
Christakis and Fowler studied the role of social prominence in the spread
of obesity [4], smoking [5] and happiness [9]. Their results suggest that these
health conditions may exhibit some amount of “contagion” in a social sense:
although the dynamics of diffusion are different from the biological virus case,
they nonetheless can spread through the social network.
3 Leader Detection
Each diffusion process has its starting points. Any idea, disease or trend is firstly
adopted by particular kinds of actors. Such actors are not like every other actor:
they have an increased sensibility and they are the first to perform an action in
a given social context. We call such actors prominent users, or leaders, because
they are able to anticipate how other actors will behave. Given a graph, several
interesting problems arise regarding how information spreads over its topology:
can we identify the leaders? Can we characterize them? What kind of knowledge
should we expect to extract from their analysis?
Our approach aims to detect leaders through the analysis of two correlated
entities: the topology of the social graph and the set of actions performed by the
actors (nodes). When discussing the roles of those entities, we refer respectively
to the following definitions:
Definition 1 (Social Graph). A social graph Gis composed by a set of actors
(nodes) Vconnected by their social relationships (edges) E. Each edge eEis
defined as a couple (u, v)with u, v Vand, where not otherwise specified, has
to be considered undirected. With Γ(u)we identify the neighbor set of a node u.
Definition 2 (Action). An action au,ψ = (w, t)defines the adoption by an
actor uV, at a certain time t, of a specific object ψwith a weight w∈ R. The
set of all the actions of nodes belonging to a social graph Gwill be identified by
A, while the object set will be called Ψ.
We identify with Gψ= (Vψ, Eψ), where VψVand EψE, the induced
subgraph on Grepresenting respectively the set of all the actors that have per-
formed an action on ψ, and the edges connecting them. We depict an example
24 7
5 63 8
a1,x =(1,0) a4,y =(2,1)
a2,x =(2,1) a7,y =(4,3)
a3,x =(1,2) a8,y =(5,2)
a4,x =(4,6) a6,y =(6,2)
a5,x =(1,4) a1,y =(2,3)
a6,x =(6,7) a2,y =(1,5)
5 63
5 63
1 1
2 4
Fig. 1: Toy Example. On the left the social graph Gand action set A, where
x, y Ψare the objects of the actions; in the center the induced subgraph for
the action x; on the right the diffusion tree for x. In red we highlighted the leader
(root) for the given tree.
of the social graph and the set of actions in Figure 1 (left), where the induced
subgraph for the object xis highlighted with a dashed line. In the Figure, a1,x
refers to the user 1 performing the action x; and a1,x = (1,0) means that user 1
performed xone time, starting at the timestep 0.
Given the nature of a diffusion process, we would expect that each leader
will be prominent among its neighbors, being the root of a cascade event that
follows some rigid temporal constraints. Our constraint is that a node uprecedes
a neighbor viff given tu,ψ au,ψ and tv,ψ av,ψ is verified that tv,ψ > tu,ψ and
tv,ψ tu,ψ δ. Here, δis a temporal resolution parameter that limits the cascade
effect: if tv,ψ tu,ψ > δ, we say that vexecuted action av,ψ independently from
u, as u’s prominence interval is over.
We transform each undirected subgraph Gψin a directed one imposing that
the source node of an edge must have performed its action before the target
node. After that, each edge (u, v) will be labeled with min(tu,ψ , tv,ψ ) to identify
when the diffusion started going from one node to the other. The directed version
of Gψrepresent all the possible diffusion paths that connect leaders with their
“tribes” (Figure 1 (center) an example for the object xΨ).
From now on, for a given object ψ, we will refer to the corresponding leader
set as Lψ: when no action is specified the set Lwill be used to describe the
union of all the Lψfor the graph G. To be defined a leader an actor should not
have any incoming edges in Gψ. This is because a prominent user cannot act
after another user (they are, in their surroundings, innovators), and is a direct
consequence to the adoption of a directed graph to express diffusion patterns.
Given this definition, for each directed connected component Cψ⊂ Gψmultiple
nodes can belong to Lψ.
Realistically, a leader may be influenced by exogenous events. This is not a
problem as we are not measuring a node’s influence, but a node’s prominence,
i.e. its propensity to act faster than others to any kind of exogenous and/or
endogenous influence. To study the path of diffusion given an action aand a
leader lwe use a minimum diffusion tree:
Definition 3 (Leader’s Minimum Diffusion Tree). Given an action aψ, a
directed connected component Cψand a leader l∈ Lψ, the minimum diffusion
tree Tl,ψ ⊂ Cψis the Minimum spanning tree (MST) having its root in land
built minimizing the temporal label assigned at the edges.
An example of minimum diffusion tree for the node 1 and object xis shown
in Figure 1 (right). For each object, the diffusion process on a given network is
independent. Moreover, given temporal dependencies on its adoption (expressed
through actions a∈ A), it is possible to identify the origin points of the
diffusion. The identified leaders will show different topological characteristic and
will be prominent in their surroundings in different ways: our aim is to classify
diffusion leaders characterizing some of their common traits.
4 Measures
To capture the three dimensions of social prominence we need three network
measures. We call these measures Width, the ratio of neighbors mirroring an
action after a node; Depth, how many degrees of separation are in between a
node and the most distant of the nodes mirroring its actions; and Strength, how
strongly nodes are mirroring a node’s action.
Given a leader, the Width aims to capture the direct impact of her actions
on her neighbors, i.e. the degree of importance that a leader has over her friends.
Definition 4 (Width). Let Gbe a social graph, ψΨan object and l∈ Lψ
Va leader: the function width :Lψ[0,1] is defined as:
width(l, ψ) = |{u|uΓ(l)∧ ∃au,ψ ∈ A}|
The value returned is the ratio of all the neighbors that, after the action of
the leader, have adopted the same object.
The Depth measure evaluates how much a leader can be prominent among
other prominent leaders, which can be prominent on other leaders and so on.
Definition 5 (Depth). Let Tl,ψ be a minimum diffusion tree for a leader l∈ Lψ
and a given object ψΨ: the function depth :Tl,ψ Ncomputes the length of
the maximal path from lto a node uTl,ψ . The function depthavg :Tl,ψ R
computes the average length of paths from lto any leaf of the tree.
The last proposed measure, the Strength, tries to capture quantitatively the
total weight of the adoption of an object after the leader’s action. A leader is
strongly prominent if the nodes among which she is prominent are very engaged
in adopting what she adopted. Direct prominence diminishes as new adopters be-
come more distant, in the network sense, from the original innovator. Therefore,
we decided to introduce a distance damping factor.
Definition 6 (Strength). Let Tl,ψ be a minimum diffusion tree for a leader
l∈ Lψand an object ψΨ;0< β < 1a damping factor: the function strength :
Tl,ψ ×(0,1) Ris defined as:
strength(Tl,ψ , β ) = X
βiL(Tl,ψ, i) (2)
where L:Tl,ψ ×NRis defined as:
L(Tl,ψ, i) = X
and represents the sum, over all the nodes uat distance ifrom l, of the ratio
between the weight of action ψand the total weight of all the actions taken.
Given the example in Figure 1, what are the Width, Depth and Strength
values for the red node leader and the action x?
Width: from Figure 1(left) we see that Γ(1) = {2,4,7,8}, i.e. 4 nodes. Given
that Γx(1) = {u|uΓ(1) ∧ ∃au,x}={2,4}, we have width(1, x) = |Γx(1)|
|Γ(1)|= 0.5.
Depth: the leaves in Figure 1(rigth) are nodes 3, 4 and 6. Node 4 is a direct
neighbor of 1, while node 3 is two edges away. The longest chain is 1 2
56, therefore depth(T1,x) = 3. We can also calculate depthav g (T1,x), that is
the average path length in the tree from node 1 to all the leaves: 1+2+3
3= 2.
Strength: we need to use the number of times each node performed action x.
We also set our damping faction β= 0.5. At the first degree we have nodes 2 and
4, that performed action x2 and 4 times respectively; they also performed action
y1 and 2 times respectively: their contribution is then β0×(2
2+1 +4
4+2 ). Nodes
2 and 5 are at the second degree of separation as they never performed action y,
therefore they add: β1×(1 + 1). Finally, at the third degree of separation, node
6 adds β2×6
6+6 . Wrapping up, strength(T1,x ,0.5) = 2.458¯
5 Experiments
In this section we present our data extracted from the music social media
Last.Fm. We use the data to characterize the Width, Depth and Strength mea-
sures, by searching for associations with network topology measures. Finally, we
analyze the prominence of different users for different musical genres.
5.1 Data
Last.Fm is an online social network platform, where people can share their own
music tastes and discover new artists and genres basing on what they, or their
friends, like. Users send data about their own listenings. For each song, a user can
express her preferences and add tags (e.g. genre of the song). Lastly, a user can
add friends (undirected connections, the friendship request must be confirmed)
and search her neighbors w.r.t. musical tastes. A user can see, in her homepage,
her friends’ activities. The co-presence of these characteristics makes Last.Fm
the ideal platform on which test our method, as it contains everything we need:
social connections that can convey social prominence, a measure of intensity
proportional to the number of listening of an artist, rich metadata attached to
each song/artist and an intrinsic temporal dimension of users’ actions.
Using Last.Fm APIs4, we obtained a sample of the UK user graph, exploring
the network with a breadth-first approach, up until the fifth degree of separation
from our seeds. For each user, we retrieved: (a) her connections, and (b) for each
week in the time window from Jan-10 to Dec-11, the number of single listenings
of a given artist (e.g. in the week between April 11,2010 and April 18,2010 the
user 1234 has listened 66 songs from the artist Metallica).
Fig. 2: Log-binned distribution of the nodes’ degree. Log-binned distribution of
number of listeners per artist.
For each artist we have a list of tags, weighted with the number of users that
assigned the tag to the artist (e.g. Metallica has 4 tags: “metal” with counter
50670, “hard rock” with 23405, “punk” with 10500 and “adrenaline” with 670).
We split tags, associating the counter to each single word (in the last example:
(metal, 50670), (punk, 10500), (hard, 23405), (rock, 23405), (adrenaline, 670)),
then we filtered the words referring to a musical genre ((metal, 50670), (punk,
10500), (rock, 23405)). Finally, we assigned a musical genre to an artist iff the
survived tag with the greater counter had the relative rate 0.5 (in the example:
rmetal(M etallica) = 50670
50670+10500+23405 '0.6, so Metallica are definitely metal).
After the crawl and cleaning stages, we built our social graph G. In Geach
node is a user and each edge is generated using the user’s friends in the social
media platform. The total amount of nodes is 75,969, with 389,639 edges con-
necting them. In Figure 2 (left) we depicted the log-binned degree distribution of
G, along with the best fit. Each action in the data is one user listening to an artist
wtimes in week t. In Figure 2 (right) we depicted the log-binned distribution of
the number of listeners per artist, along with the best fit.
Since we are interested in leaders, we need to focus only on new artists that
were previously not existent. If an artist was in activity before our observation
time window, there is no way to know if a user has listened to it before, therefore
nullifying our leader detection strategy. For this reason, we focus only on artists
whose first listening is recorded six months after the beginning of our observation
period. Each artist belongs to a music genre (coded in its tag) and we want to
use this information in Section 5.3. We decided to focus on music genres with
sufficient popularity, namely: dance, electronic, folk, jazz, metal, pop, punk, rap
and rock. A genre’s popularity is determined by having at least 10 artists with
at least 100 listeners. To sum up, we focus on the artists who appear for the first
time after six months in our observation period, with at least 100 listeners and
belonging to one of the mentioned nine tags. The cardinality of our action set A
is 168,216 actions, while the object set Ψcontains a total of 402 artists.
In our experimental settings, we set our damping factor β= 0.5 for the
calculation of the Strength measure. We also set δ= 3, meaning that if a user
listened to a particular artist three weeks or more after its neighbor then we do
not consider her neighbor to be prominent for her for that action.5
5.2 Characterization of the Measures
For each leader, besides Width, Depth and Strength, we calculated also the
Degree (number of edges connected to the node), the Clustering coefficient (ratio
of triangles over the possible triads centered on the node), the Neighbor Degree
(average degree of the neighbors of the node), the Betweenness (share of the
shortest paths that pass through the node) and Closeness Centrality (inverse
average distance between the node and all the other nodes of the network).
Width Strength Degree Clustering Neigh Deg Bet Centr Clo Centr
AVG Depth -0.03 -0.23 -0.08 0.05 -0.08 -0.02 -0.13
Width - 0.01 -0.31 0.13 0.05 -0.07 -0.59
Strength - - 0.02 -0.02 0.03 0.00 0.04
Degree - - - -0.16 -0.02 0.77 0.56
Clustering - - - - -0.05 -0.06 -0.32
Neigh Deg - - - - - -0.00 0.39
Bet Centr - - - - - - 0.22
Table 1: Pearson correlation coefficient ρbetween Width, Depth, Strength and
other network statistics for our leaders.
In Table 1 we report the Pearson correlation coefficient ρbetween the network
measures. We highlighted the correlations whose p-value was significant or whose
absolute value was strong enough to draw some conclusions. For the significance
of p-values, the traditional choice is to set the threshold at p < 0.01. However,
given our number of observations, we decided to be more restrictive, setting our
threshold at p < 0.0005. We also consider a ρvalue significant if |ρ|>0.1.
The Depth measure is associated with low Closeness Centrality. This means
that a deep prominence is associated to nodes at the margin of the network.
It is expected that nodes with high Closeness Centrality have also low Depth:
being central, they cannot generate long chains of diffusion. The eccentricity of
all the nodes of the network ranges from 6 to 10, meaning that some leaders
cannot have a Depth larger than 5. To make a fair comparison, we recalculate
the Depth value capping it at 5, meaning that any Depth value larger than 5 is
manually reduced to 5. Then, we recalculate the correlation ρbetween the Depth
capped to 5 and the Closeness Centrality obtaining as result ρ=0.1366, with
p < 0.0005. We can conclude that central nodes are not associated with deep
spread of their prominence in a social network.
5To assure experiment repeatability, we made our cleaned dataset and our code avail-
able at the page
For the Width measure, the anti-correlation with the Degree is not meaning-
ful, as the Degree is in the denominator of Definition 4. However, we observe a
positive association with Clustering, i.e. nodes could be prominent in a tightly
connected community; and a negative association with Closeness Centrality, i.e.
central nodes could not spread a wide influence. Both associations could be ex-
plained with the negative correlation with Degree. Therefore, for both measures
we run a partial correlation, controlling for the Degree. In practice, we calcu-
late the correlation between Width and Clustering (or Closeness Centrality) by
keeping the Degree constant. Results are in Table 2: even if significant according
to the p-value, the relationship between Width and Clustering is very weak and
deserves further investigation. On the other hand, it is confirmed that central
nodes are also associated with low Width, regardless their degree.
Clustering Clo Centr
Partial ρ0.087216 -0.536861
p-value 1.57 ×1014 0
Table 2: Partial correlation and p-value of Clustering and Closeness Centrality
with Width, controlling for Degree values.
From Table 1, we see that the Strength measure is not correlated with tradi-
tional network statistics. As a consequence, hubs associated with low Depth and
low Width, do not have necessarily high Strength, making their prominence in
a network questionable. Moreover, Strength appears to be negatively associated
with Depth, suggesting a trade-off between how deeply a node can be prominent
in a network and how strong this prominence is on the involved nodes.
The anti-correlation between the Strength and the Depth may be due to β:
from Definition 6 βdecreases nodes’ contributions at each degree of separation
(i.e. at increasing Depths). As a consequence, nodes farther from the leader
contribute less to its Strength, i.e. the highest the Depth the smallest are the
contributions to the Strength. We recalculated the Strength values by setting
β= 1, therefore ignoring any damping factor and nullifying this effect. We
obtained as result ρ=0.4168 and a significant p-value, therefore concluding
that βis not causing the anti-correlation between Depth and Strength.
To sum up, we summarize the associations as follows: (i) central nodes are
not necessarily prominent in a social network (low Width and Depth), a result
that confirms [3] and [1]; (ii) longer cascades (higher Depths) are associated with
a lower degree of engagement (lower Strengths), a phenomenon possibly related
to the role played by “weak ties”; (iii) be prominent among neighbors is probably
easier if the node is in a tightly connected community, but more evidences have
to be brought to reject the role played by the node’s degree.
5.3 Case Study
Here, we present a case study based on Last.Fm data. Our aim is to use our
Leader extraction technique and the proposed Width, Depth and Strength mea-
sures to characterize the spread of musical genres among the users of the service.
Cluster size dance ele folk jazz met pop punk rap rock
0 1822 1.25 1.13 1.54 1.37 1.50 0.76 1.31 1.13 1.10
1 136 1.28 1.55 1.28 2.35 0.78 0.73 0.64 1.35 0.70
2 664 0.59 0.87 0.98 0.48 0.95 0.97 1.50 1.20 1.19
3 482 1.26 1.16 1.09 1.12 0.91 0.80 2.48 1.24 0.89
4 973 1.14 1.20 1.15 1.41 0.80 0.91 0.66 0.97 0.97
5 512 1.29 0.96 0.95 1.09 1.10 0.97 0.33 1.06 1.01
6 682 0.89 0.79 0.61 0.64 1.13 1.08 1.07 1.08 1.01
7 124 0.75 1.45 0.35 0.64 0 1.09 0 1.02 0.62
8 524 0.93 1.01 1.12 0.91 1.15 1.07 0.43 0.95 0.87
9 937 0.40 0.46 0.19 0.23 0.45 1.56 0.13 0.37 1.06
10 232 0.72 0.57 0.27 0.99 0.38 1.44 0.38 0.46 1.00
11 612 0.74 0.94 0.71 0.40 0.70 1.27 0.07 0.68 0.83
(a) (b)
Fig. 3: (a) The RCA scores of the presence of each tag in each cluster; (b) The
centroids of our clusters.
We recall that, as described in Section 5.1, the object set Ψis composed by 402
artists, each one having a tag corresponding to her main music genre.
For each couple leader land object ψ, we calculate Depth, Width and
Strength values; we compute the size of the Leader’s Minimum Diffusion Tree
(|Tl,ψ|); and we group together the objects with the same tag. To characterize
the typical values of Width, Depth and Strength for each tag we cannot use the
average or the median. This is because Strength and Width values are skewed,
and it is the combination of the three measures that really characterizes the
leaders. We cluster leaders using as features their Width, Depth and Strength
values. We used the Self-Organizing Map (SOM) method [13] because: (i) SOM
does not require to set the number of clusters k; (ii) k-means outperforms SOM
only if the number of resulting clusters is very small (less than 7) [14], but our
study of the best kto be used in k-means with the Sum of Squared Errors (SSE)
methodology resulted in a optimal number of clusters falling in a range between
9 and 13 (in fact, SOM returned 12 clusters); and (iii) SOM performs better if
the data points are contained in a warped space [12], which is our case.
In Table 3(a), we report a presence score for each tag in each cluster. There
are larger and smaller clusters and some tags attract more listeners than others.
To report just the share of leaders with a given tag in a given cluster is not
meaningful. We correct the ratio with the expected number of leaders with the
given tag in the cluster, a measure known as Revealed Comparative Advantage:
RCA(i, j) = freqi,j
freqi,.f req,j
freq,,where iis a tag, jis a cluster, f reqi,j is the number
of leaders who spread an artist tagged with tag ithat is present in cluster j. For
each cluster we highlighted the tag with the highest unexpected presence.
The centroids of the SOM are depicted in Figure 3(b): Depth on the x-
axis, Strength on the y-axis and the Width as the color (Strength and Width
are in log scale). We can identify the clusters characterized by the highest and
lowest Strength (9 and 4 respectively); by the highest and lowest Depth (2 and
9 respectively); and by the highest and lowest Width (11 and 1 respectively).
There are also clusters with relatively high combinations of two measures: cluster
10 with high Strength and Width or cluster 5 with high Depth and Width.
From Table 3(a) we obtain a description of what values of Width, Depth and
Strength are generally associated with each tag. For space constraints, we report
only a handful of them for the clusters with extreme values. Jazz dominates
clusters 1 (with the lowest Width) and 4 (with the lowest Strength): this fact
suggests that jazz is a genre for which it is not easy to be prominent.
Cluster 9, with the lowest Depth but the highest Strength, is dominated by
pop (that dominates also clusters 10 and 11, both with high Strength but low
Depth). As a result, we can conclude that prominent leaders for pop artists are
embedded in groups of users very engaged with the new artist. On the other
hand, it is unlikely that these users will be prominent among their friends too.
Finally, cluster 2 with the highest density has a large majority of punk lead-
ers. From this evidence, we can conclude that punk is a genre that can achieve
long cascades, exactly the opposite of the pop genre.
We move on to the topological characteristics of the leaders per tag. A caveat:
a leader is not bounded to be leader just for one object ψ, but she is free to
be prominent in many ψ. Thus, one leader can be counted in more than one
tag. To help understand the magnitude of the issue, we depicted in Figure 4 the
number of leaders influencing their neighbors for a given amount of actions (left)
and for a given amount of tags (right). The y axis is logarithmic. The typical
leader influences one neighbor for one artist. However, some leaders express their
leadership for 8 objects and 4 tags.
# Leaders
# Objects
1 2 3 4 5 6
# Leaders
# Tags
Fig. 4: Distribution of number of objects (left) and of tags (right) per leader.
In Figure 5 we depict the log-binned distributions, for the leaders of each
tag, of four of the topological measures studied in Section 5.2: Degree, Closeness
Centrality, Clustering and Neighbor Degree. We omit Betweenness Centrality for
its very high correlation with Degree. Overall, there is no significant distinction
between the tags in the distributions of the topological features.
The most noticeable information is carried by the Degree distributions (Fig-
ure 5, top left). Each distribution appears very different from the overall degree
distribution (Figure 2 (left)). There are fewer leaders with low Degree than ex-
pected, therefore it appears that a high Degree increases the probability of being
a leader. On the other hand, we know that central hubs have on average lower
Depth and Width. As a consequence, it appears that the best leader candidates
are the nodes with an average degree, and from Figure 5 (top left) we see that
each tag has many leaders with a Degree between 10 and 100.
1.5 x 10-1 2 x 10-1 3 x 10-1 4 x 10-1
10-3 10-2 10-1 100
p(Neighbor Degree)
Neighbor Degree
Electronic Folk
Jazz Metal
Pop Punk
Rap Rock
Fig. 5: Distribution of leaders’ Degree (top left), Closeness Centrality (top right),
Clustering (bottom left) and Neighbor Degree (bottom right) per tag.
Using our leaders’ Minimum Diffusion Trees, we extract some patterns that
help us obtaining a complementary point of view over the leader prominence
for different music genres. We mine a graph dataset composed by all diffusion
trees Tl,ψ with the VF2 algorithm [7]. Suppose we are interested in counting how
frequent is the following star pattern: a leader influences three of its neighbors
in the diffusion trees of pop artists. In our data, we have 5,043 diffusion trees
for pop artists, and 581 have at least four nodes. Since the VF2 algorithm found
the star pattern in 186 of these graphs, we say that it appears in 3.69% of the
trees, or in 32.01% of the trees that have enough nodes to contain it.
In Table 3 we report the results of mining three patterns of four nodes: i) the
star-like pattern described above; ii) a chain where each node is prominent for (at
least) one neighbor; iii) a split where the leader is prominent for a node, which
itself is prominent for two other neighbors. Two values are associated to each
pattern and tag pair: the relative overall frequency, and the relative frequency
considering only the trees with at least four nodes (in parentheses).
There is no necessary relation between the patterns and Width, Depth and
Strength measures: a low Depth does not imply the absence of the chain pattern,
nor does a high Width imply a high presence of the star pattern. However, the
combination of the two measures may provide some insights. For instance, we
saw in Table 3(a) that jazz leaders are concentrated in the lowest Width cluster.
However, many jazz leaders who affect at least three nodes tend to be prominent
in their neighbors, much more than in any other genre (7.25% of all leaders,
62.5% of leaders who are prominent for at least three other nodes). Therefore,
Pattern dance electronic folk jazz metal pop punk rap rock
Table 3: Presence of different diffusion patterns per tag.
jazz leaders have low prominence among their friends, however they are likely
to have at least three neighbors for which they are prominent.
The chain pattern is more commonly found in pop leaders than in folk ones,
even though the clusters of their leaders described in Table 3(a) would suggest the
opposite. It seems that pop leaders are not likely to be prominent for nodes any
further than the third degree of separation, while folk leaders tend to generate
longer cascade chains. Also in this case, punk leaders are commonly found in
correspondence with chain patterns, just as Table 3(a) suggested.
Although pop leaders show a much greater Strength value than metal ones
(by confronting in Table 3(a) their presence in high Strength clusters like 9 or
10 and low Strength clusters like 8 and 0), the split pattern tends to be more
frequent in the metal genre (6.90% against 4.73% of the trees). This phenomenon
suggests us that metal leaders tend to be prominent for nodes strongly devoted to
metal, inducing them to spread the music to their neighbors. Pop leaders, on the
other hand, affect more neighbors with higher Width and Strength, presumably
flooding their ego networks with the songs they like.
6 Conclusion
In this paper, we presented a study of the propagation of behaviors in a social
network. Instead of just studying cascade effects and the maximization of influ-
ence by a given starting seed, we decided to analyze three different dimensions:
the prominence of a leader on how many neighbors, on how distant nodes and
on how engaged nodes. We characterized each of these concepts with a different
measure: Width, Depth and Strength. We applied our leader detection algorithm
to a real world network. Our results show that: (i) central hubs are usually inca-
pable of having a strong effect in influencing the behavior of the entire network;
(ii) there is a trade-off between how long the cascade chains are and how engaged
each element of the chain is; (iii) to achieve maximum engagement it is better
to target leaders in tightly connected communities, although for this last point
we do not have conclusive evidence. We also included a case study in which we
show how artists in different musical genres are spread through the network.
Many future developments are possible. The limited prominence that central
hubs have on the overall network may be studied in conjunction with the prob-
lem of network controllability [15]. Alternative leader detection techniques, such
as the ones presented in [10], can be confronted with our proposed algorithm.
Finally, a deeper analysis of the properties of the Width, Depth and Strength
measures can be performed, using additional techniques and exploiting data from
other social media services like Twitter and Facebook.
Acknowledgments. The authors want to thank Prof. Otto Koppius for his
presentation about the prominence network measures and useful discussions.
This work has been partially supported by the European Commission under the
FET-Open Project n. FP7-ICT-270833, DATA SIM.
1. Michele Berlingerio, Michele Coscia, and Fosca Giannotti. Mining the temporal
dimension of the information propagation. In IDA, pages 237–248, 2009.
2. Ronald S Burt. Social contagion and innovation: Cohesion versus structural equiv-
alence. American Journal of Sociology, 92(6):1287–1335, 1987.
3. M. Cha, H. Haddadi, F. Benevenuto, and K.P. Gummadi. Measuring user influence
in twitter: The million follower fallacy. In ICWSM, 2010.
4. Nicholas A Christakis and James H Fowler. The spread of obesity in a large social
network over 32 years. New England Journal of Medicine, 357(4):370–379, 2007.
5. Nicholas A Christakis and James H Fowler. The collective dynamics of smoking in
a large social network. New England Jou. of Medicine, 358(21):2249–2258, 2008.
6. Vittoria Colizza, Alain Barrat, Marc Barthelemy, Alain-Jacques Valleron, and
Alessandro Vespignani. Modeling the worldwide spread of pandemic influenza:
Baseline case and containment interventions. PLoS Medicine, 4(1):e13, 2007.
7. L. P. Cordella, P. Foggia, C. Sansone, and M. Vento. A (sub)graph isomorphism
algorithm for matching large graphs. Pattern Analysis and Machine Intelligence,
IEEE Transactions on, 26(10):1367–1372, 2004.
8. Michele Coscia. Competition and success in the meme pool: a case study on ICWSM, 2013.
9. J H Fowler and N A Christakis. Dynamic spread of happiness in a large social
network: longitudinal analysis over 20 years in the framingham heart study. Bmj
Clinical Research Ed., 337(2):a2338–a2338, 2008.
10. Amit Goyal, Francesco Bonchi, and Laks V. S. Lakshmanan. Discovering leaders
from community actions. In CIKM, pages 499–508, 2008.
11. W O Kermack and A G McKendrick. A contribution to the mathematical theory
of epidemics. The Royal Society of London Series A, 115(772):700–721, 1927.
12. M. Y. Kiang and A. Kumar. A comparative analysis of an extended som network
and k-means analysis. Int. J. Know.-Based Intell. Eng. Syst., 8(1):9–15, 2004.
13. Teuvo Kohonen. The self-organizing map. IEEE, 78:1464–1480, 1990.
14. U.A. Kumar and Y. Dhamija. A comparative analysis of som neural network with
k-means clustering algorithm. Proceedings of IEEE International Conference on
Management of Innovation and Technology, pages 55–59, 2004.
15. Yang-Yu Liu, Jean-Jacques Slotine, and Albert-Laszlo Barabasi. Controllability of
complex networks. Nature, 473(7346):167–173, May 2011.
16. Romualdo Pastor-Satorras and Alessandro Vespignani. Epidemic spreading in
scale-free networks. Physical review letters, 86(14):3200–3203, 2001.
17. P. Wang, M. C. Gonz´alez, C. A. Hidalgo, and A-L. Barab´asi. Understanding the
spreading patterns of mobile phone viruses. Science, 324(5930):1071–1076, 2009.
... The results hold as long as the assumptions and the dynamics of the models are an accurate approximation of reality. We provided evidence to motivate the bipolar model's assumptions, but there could still be factors unaccounted for, such as the role of originality [36] or of spreaders' effort [37] in making content go viral. Second, many aspects of the model were fixed and should be investigated. ...
Full-text available
Many people view news on social media, yet the production of news items online has come under fire because of the common spreading of misinformation. Social media platforms police their content in various ways. Primarily they rely on crowdsourced ‘flags’: users signal to the platform that a specific news item might be misleading and, if they raise enough of them, the item will be fact-checked. However, real-world data show that the most flagged news sources are also the most popular and—supposedly—reliable ones. In this paper, we show that this phenomenon can be explained by the unreasonable assumptions that current content policing strategies make about how the online social media environment is shaped. The most realistic assumption is that confirmation bias will prevent a user from flagging a news item if they share the same political bias as the news source producing it. We show, via agent-based simulations, that a model reproducing our current understanding of the social media environment will necessarily result in the most neutral and accurate sources receiving most flags.
... The emergence for dynamic networks analysis theoretical grounds has been highlighted in the book "temporal networks" [16] where the curators, Holme and Saramaki, proposed an ensemble of works covering different dynamic network analysis methodologies. As a first step, several works have tried to transpose known problems on static networks to temporal networks: Temporal motifs mining [21], Diffusion [23,28], Link prediction [31,39], community discovery [29,30,32], are only a few examples. ...
Conference Paper
Nowadays the analysis of dynamics of and on networks represents a hot topic in the Social Network Analysis playground. To support students, teachers, developers and researchers we introduced a novel framework, named NDlib, an environment designed to describe diffusion simulations. NDlib is designed to be a multi-level ecosystem that can be fruitfully used by different user segments. Upon NDlib, we designed a simulation server that allows remote execution of experiments as well as an online visualization tool that abstracts its programmatic interface and makes available the simulation platform to non-technicians.
... The emergence for dynamic networks analysis theoretical grounds has been highlighted in the book "temporal networks" [16] where the curators, Holme and Saramaki, proposed an ensemble of works covering different dynamic network analysis methodologies. As a first step, several works have tried to transpose known problems on static networks to temporal networks: Temporal motifs mining [21], Diffusion [23,28], Link prediction [31,39], community discovery [29,30,32], are only a few examples. ...
Full-text available
Nowadays the analysis of dynamics of and on networks represents a hot topic in the social network analysis playground. To support students, teachers, developers and researchers, in this work we introduce a novel framework, namely NDlib, an environment designed to describe diffusion simulations. NDlib is designed to be a multi-level ecosystem that can be fruitfully used by different user segments. For this reason, upon NDlib, we designed a simulation server that allows remote execution of experiments as well as an online visualization tool that abstracts its programmatic interface and makes available the simulation platform to non-technicians.
... In particular, as Forecasting success via early adoptions analysis: A data-driven study shown from the current work, three weakness of such model are often highlighted [16,17]: (i) it is not justified empirically or analytically; (ii) it assumes that the adoption trend is normally distributed; (iii) the threshold is univocally defined for all the possible products. In recent years problem of identify innovators has lived a second youth: diffusion of Memes [18], social prominence analysis [19], discovery of innovation communities [20] as well as identification of Trend-Setter [21][22][23][24] are only a few of the network studies which exploit [25] or model [26] techniques aimed at identifying key users that are able to influence the diffusion of a product, content or idea over a social network. Although being related Hit-Savvy and Trend-Setter carry different semantics: where a Hit-Savvy is a passive actor that operates on the basis of its personal taste, a Trend-Setter is often defined as an active one that with his own choice aims to influence his peers. ...
Full-text available
Innovations are continuously launched over markets, such as new products over the retail market or new artists over the music scene. Some innovations become a success; others don’t. Forecasting which innovations will succeed at the beginning of their lifecycle is hard. In this paper, we provide a data-driven, large-scale account of the existence of a special niche among early adopters, individuals that consistently tend to adopt successful innovations before they reach success: we will call them Hit-Savvy. Hit-Savvy can be discovered in very different markets and retain over time their ability to anticipate the success of innovations. As our second contribution, we devise a predictive analytical process, exploiting Hit-Savvy as signals, which achieves high accuracy in the early-stage prediction of successful innovations, far beyond the reach of state-of-the-art time series forecasting models. Indeed, our findings and predictive model can be fruitfully used to support marketing strategies and product placement.
... The real world network is a sample of 70 k users of the on-line platform (network statistics are provided in [10]). ...
Conference Paper
Full-text available
Ideas, information, viruses: all of them, with their mechanisms, can spread over the complex social tissues described by our interpersonal relations. Classical spreading models can agnostically from the object of which they simulate the diffusion, thus considering spreading of virus, ideas and innovations alike. Indeed, such simplification makes easier to define a standard set of tools that can be applied to heterogeneous contexts; however, it can also lead to biased, partial, simulation outcomes. In this work we discuss the concepts of active and passive diffusion: moving from analysis of a well-known passive model, the Threshold one, we introduce two novel approaches whose aim is to provide active and mixed schemas applicable in the context of innovations/ideas diffusion simulation. Our data-driven analysis shows how, in such context, the adoption of exclusively passive/active models leads to conflicting results, thus highlighting the need of mixed approaches.
... So, in many practical situations, it is not an optimization problem that we should solve when social networks are considered (i.e., applying the best inputs to the smallest set of driver nodes in order to reach the optimal final state), rather it is a problem of finding heuristics for maximizing the welfare. A possibility is to perturb the system evolution for modifying the basin of attraction (i.e., modify the dynamics so that the system that was attracted towards a certain space of final states becomes attracted towards a different one) [22,23]. The property of structural controllability is also not sufficient as a requirement, because often there are practical limitations to the type of perturbations that could be injected into the system through driver nodes and also limitations to the accessibility of driver nodes. ...
Full-text available
This work is aimed at studying realistic social control strategies for social networks based on the introduction of random information into the state of selected driver agents. Deliberately exposing selected agents to random information is a technique already experimented in recommender systems or search engines, and represents one of the few options for influencing the behavior of a social context that could be accepted as ethical, could be fully disclosed to members, and does not involve the use of force or of deception. Our research is based on a model of knowledge diffusion applied to a time-varying adaptive network, and considers two well-known strategies for influencing social contexts. One is the selection of few influencers for manipulating their actions in order to drive the whole network to a certain behavior; the other, instead, drives the network behavior acting on the state of a large subset of ordinary, scarcely influencing users. The two approaches have been studied in terms of network and diffusion effects. The network effect is analyzed through the changes induced on network average degree and clustering coefficient, while the diffusion effect is based on two ad-hoc metrics defined to measure the degree of knowledge diffusion and skill level, as well as the polarization of agent interests. The results, obtained through simulations on synthetic networks, show a rich dynamics and strong effects on the communication structure and on the distribution of knowledge and skills, supporting our hypothesis that the strategic use of random information could represent a realistic approach to social network controllability, and that with both strategies, in principle, the control effect could be remarkable.
Evaluating the public data from person‐to‐person communication destinations through the social network could create invigorating outcomes and bits of knowledge on the general assessment of practically any product, administration, or conduct. One of the best and precise public notion pointers is through information mining from social networks, as numerous clients seem to state their viewpoints on the social networks. The innovation in the Internet technologies figured out how to expand action in contributing to a blog, labeling, posting, and online informal communication. Therefore, individuals are beginning to develop keen on mining these immense information assets to evaluate the viewpoints. The Social Network Analysis (SNA) is the way toward researching social designs using graph hypothesis and networks. It integrates an assortment of procedures for analyzing the design of informal organizations, in addition with the hypotheses that target describing the hidden elements and the patterns in this framework. It is an intrinsically integrative field, which initially emerged from the sectors of graph hypothesis, statistics, and sociopsychology. This chapter will cover the hypothesis of SNA, with a short prologue to graph hypothesis and data spread. Then discuss the role of Python in SNA, followed up by building and suggesting informal communities from genuine Pandas and text‐based data sets.
As Nietzsche once wrote “Without music, life would be a mistake” (Twilight of the Idols, 1889.). The music we listen to reflects our personality, our way to approach life. In order to enforce self-awareness, we devised a Personal Listening Data Model that allows for capturing individual music preferences and patterns of music consumption. We applied our model to 30k users of Last.Fm for which we collected both friendship ties and multiple listening. Starting from such rich data we performed an analysis whose final aim was twofold: (i) capture, and characterize, the individual dimension of music consumption in order to identify clusters of like-minded Last.Fm users; (ii) analyze if, and how, such clusters relate to the social structure expressed by the users in the service. Do there exist individuals having similar Personal Listening Data Models? If so, are they directly connected in the social graph or belong to the same community?.
Full-text available
Controlling complex networks is of paramount importance in science and engineering. Despite the recent development of structural controllability theory, we continue to lack a framework to control undirected complex networks, especially given link weights. Here we introduce an exact controllability paradigm based on the maximum multiplicity to identify the minimum set of driver nodes required to achieve full control of networks with arbitrary structures and link-weight distributions. The framework reproduces the structural controllability of directed networks characterized by structural matrices. We explore the controllability of a large number of real and model networks, finding that dense networks with identical weights are difficult to be controlled. An efficient and accurate tool is offered to assess the controllability of large sparse and dense networks. The exact controllability framework enables a comprehensive understanding of the impact of network properties on controllability, a fundamental problem towards our ultimate control of complex systems.
The prevalence of obesity has increased substantially over the past 30 years. We performed a quantitative analysis of the nature and extent of the person-to-person spread of obesity as a possible factor contributing to the obesity epidemic. We evaluated a densely interconnected social network of 12,067 people assessed repeatedly from 1971 to 2003 as part of the Framingham Heart Study. The body-mass index was available for all subjects. We used longitudinal statistical models to examine whether weight gain in one person was associated with weight gain in his or her friends, siblings, spouse, and neighbors. Discernible clusters of obese persons (body-mass index [the weight in kilograms divided by the square of the height in meters], > or =30) were present in the network at all time points, and the clusters extended to three degrees of separation. These clusters did not appear to be solely attributable to the selective formation of social ties among obese persons. A person's chances of becoming obese increased by 57% (95% confidence interval [CI], 6 to 123) if he or she had a friend who became obese in a given interval. Among pairs of adult siblings, if one sibling became obese, the chance that the other would become obese increased by 40% (95% CI, 21 to 60). If one spouse became obese, the likelihood that the other spouse would become obese increased by 37% (95% CI, 7 to 73). These effects were not seen among neighbors in the immediate geographic location. Persons of the same sex had relatively greater influence on each other than those of the opposite sex. The spread of smoking cessation did not account for the spread of obesity in the network. Network phenomena appear to be relevant to the biologic and behavioral trait of obesity, and obesity appears to spread through social ties. These findings have implications for clinical and public health interventions.
The self-organized map, an architecture suggested for artificial neural networks, is explained by presenting simulation experiments and practical applications. The self-organizing map has the property of effectively creating spatially organized internal representations of various features of input signals and their abstractions. One result of this is that the self-organization process can discover semantic relationships in sentences. Brain maps, semantic maps, and early work on competitive learning are reviewed. The self-organizing map algorithm (an algorithm which order responses spatially) is reviewed, focusing on best matching cell selection and adaptation of the weight vectors. Suggestions for applying the self-organizing map algorithm, demonstrations of the ordering process, and an example of hierarchical clustering of data are presented. Fine tuning the map by learning vector quantization is addressed. The use of self-organized maps in practical speech recognition and a simulation experiment on semantic mapping are discussed
The advent of social media has provided data and insights about how people relate to information and culture. While information is composed by bits and its fundamental building bricks are relatively well understood, the same cannot be said for culture. The fundamental cultural unit has been defined as a "meme". Memes are defined in literature as specific fundamental cultural traits, that are floating in their environment together. Just like genes carried by bodies, memes are carried by cultural manifestations like songs, buildings or pictures. Memes are studied in their competition for being successfully passed from one generation of minds to another, in different ways. In this paper we choose an empirical approach to the study of memes. We downloaded data about memes from a well-known website hosting hundreds of different memes and thousands of their implementations. From this data, we empirically describe the behavior of these memes. We statistically describe meme occurrences in our dataset and we delineate their fundamental traits, along with those traits that make them more or less apt to be successful.
The Self-Organizing Map (SOM) network, a variation of neural computing networks, is a categorization network developed by Kohonen. The main function of SOM networks is to map the input data from an n-dimensional space to a lower dimensional plot while maintaining the original topological relations. In this research, we apply an extended SOM network that includes a grouping function to further cluster input data based on the relationships derived from a lower dimensional SOM map, to market segmentation problems. A computer program for implementing the extended SOM networks has been developed and it was first compared with K-means analysis in an experimental design using simulated data sets with known cluster solutions. Test results indicate that the extended SOM networks perform better when the data are skewed. We then further test the performance of the method with a real-world data set from a widely referenced machine-learning case. We believe the findings from this research can be applied to other problem domains as well.
Two classes of network models are used to reanalyze a sociological classic often cited as evidence of social contagion in the diffusion of technological innovation: Medical Innovation. Debate between the cohesion and structural equivalence models poses the following question for study: Did the physicians resolve the uncertainty of adopting the new drug through conversations with colleagues (cohesion) or through their perception of the action proper for an occupant of their position in the social structure of colleagues (structural equivalence)? The alternative models are defined, compared, and tested. Four conclusions are drawn: (a) Contagion was not the dominant factor driving tetracyclene's diffusion. Where there is evidence of contagion, there is evidence of personal preferences at work.
The self-organizing map (SOM) is an automatic data-analysis method. It is widely applied to clustering problems and data exploration in industry, finance, natural sciences, and linguistics. The most extensive applications, exemplified in this paper, can be found in the management of massive textual databases and in bioinformatics. The SOM is related to the classical vector quantization (VQ), which is used extensively in digital signal processing and transmission. Like in VQ, the SOM represents a distribution of input data items using a finite set of models. In the SOM, however, these models are automatically associated with the nodes of a regular (usually two-dimensional) grid in an orderly fashion such that more similar models become automatically associated with nodes that are adjacent in the grid, whereas less similar models are situated farther away from each other in the grid. This organization, a kind of similarity diagram of the models, makes it possible to obtain an insight into the topographic relationships of data, especially of high-dimensional data items. If the data items belong to certain predetermined classes, the models (and the nodes) can be calibrated according to these classes. An unknown input item is then classified according to that node, the model of which is most similar with it in some metric used in the construction of the SOM. A new finding introduced in this paper is that an input item can even more accurately be represented by a linear mixture of a few best-matching models. This becomes possible by a least-squares fitting procedure where the coefficients in the linear mixture of models are constrained to nonnegative values.
Conference Paper
Cluster analysis, a set of tools for building groups from multivariate data objects is extensively applied in many fields. One of the most widely used classical approaches of clustering is K-means algorithm. Kohonen's Self Organizing map is a neural network clustering methodology that maps an n-dimensional input data to a lower dimensional output map. In this study, we have compared K-means algorithm with Self Organizing map on a real life data with known cluster solutions. The performance of these algorithms is examined with respect to changes in the number of clusters and number of observations. Misclassification rates and point biserial correlation are used to compare performance of both the methods.
An overview of the self-organizing map algorithm, on which the papers in this issue are based, is presented in this article.