=Paper=
{{Paper
|id=Vol-3318/paper1
|storemode=property
|title=Estimating The MBTI Of Twitter Accounts With Graph Neural Networks Over Neo4j
|pdfUrl=https://ceur-ws.org/Vol-3318/paper1.pdf
|volume=Vol-3318
|authors=Georgios Drakopoulos,Eleanna Kafeza
|dblpUrl=https://dblp.org/rec/conf/cikm/DrakopoulosK22a
}}
==Estimating The MBTI Of Twitter Accounts With Graph Neural Networks Over Neo4j==
https://ceur-ws.org/Vol-3318/paper1.pdf
Estimating The MBTI Of Twitter Accounts With Graph
Neural Networks Over Neo4j
Georgios Drakopoulos1,∗ , Eleanna Kafeza2
1
Department of Informatics, Ionian University, Tsirigoti Sq. 7, Kerkyra 49100, Hellas
2
College of Interdisciplinary Studies, Zayed University, UAE
Abstract
Intelligent agents are indispensable and flexible autonomous tools for efficiently mining large graphs for heterogeneous
knowledge. Twitter is a prime case in point with structural and functional attributes such as original multimedia content
posting and retweeting revealing important affective information about accounts. Additionally, this can be facilitated by
including hashtag emotional polarity and reactions to political, social, or even historical events. Further insight can be gained
by moving one step forward from individual emotional reactions to integrated personality estimations such as the MBTI
taxonomy. An intelligent agent has been developed with a stochastic account visiting policy based on preferential attachment,
an optional evolving forgetting factor for penalizing vertices appearing too frequently, and the capability to yield an MBTI
estimate based on a graph neural network. The results indicate the superior performance of the proposed heuristic based on
evaluation criteria including community size distribution and hashtag coherency.
Keywords
Intelligent agents, random walks, graph neural networks, affective communities, emotional polarity, personality taxonomies,
MBTI, preferential attachment, Twitter, Neo4j, py2neo, PyTorch
1. Introduction sion making processes and marketing performance. This
is achieved as businesses are allowed through Twitter
Intelligent agents (IAs) are autonomous digital entities mining to gain invaluable insights on the dynamics and
with extensive capabilities for maintaining and ensuring collective behavior of their customer base or any other
the smooth operation of massive infrastructure, mostly online target group for that matter. In turn this yields
networks of various types. Depending mainly on their more accurate predictions of key factors such as future
technology and operating principles, an IA may have demand, reactions to new products, or brand loyalty to
extended command and control capabilities while requir- name only a few. Twitter analytics tailored for this task
ing only the initial programming in order to properly include community structure discovery algorithms, hash-
function, excluding of course any sort of necessary local tag flow analysis and information diffusion strategies,
input. Recently IA has advanced beyond a fixed set of digital influence computation, and link prediction tools.
rules to integrating various level of machine learning Such insight is obtained frequently from the computa-
(ML) capabilities, provided sufficient computing power is tionally challenging task of processing a diverse set of
available. The inclusion of neural networks architectures follow relationships, hashtags, or tweets. Such attributes
such as graph neural networks (GNNs) which take full ad- are either of structural nature in the sense that they are
vantage of the local network topology constitutes a major about the social graph itself or functional as they pertain
addition. Perhaps the most well-known representation of to the activity of the various entities, mostly the Twitter
such an agent in pop culture, albeit possessing far more accounts, which use said graph.
capabilities than those of its contemporary counterparts, Among the functional features the affective ones have
is that of the iconic agent Smith from The Matrix 1 . recently garnered considerable research attention since
Twitter mining analytics provide a significant oppor- emotions are the primary motivations behind human
tunity to various organizations to improve both deci- actions. To this end, attributes such as the emotional po-
larity of tweets and hashtags are considered as major in-
CIKM’22: 31st ACM International Conference on Information and dicators of how a Twitter account would react to various
Knowledge Management (companion volume), October 17–21, 2022, events and rely heavily on emotion models such as those
Atlanta, GA proposed by Plutchik or Ekman. However, personality
∗
Corresponding author.
taxonomies such as the Myers-Brigs taxonomy indicator
Envelope-Open c16drak@ionio.gr (G. Drakopoulos); eleana.kafeza@zu.ac.ae
(E. Kafeza) (MBTI) go beyond individual emotional responses and
Orcid 0000-0002-0975-1877 (G. Drakopoulos); 0000-0001-9565-2375 provide a more general framework for systematically
(E. Kafeza) evaluating sequences of account reactions as they take
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0). into consideration the higher cognitive functions driving
CEUR
CEUR Workshop Proceedings (CEUR-WS.org)
them. For instance, personalities with an extrovert pre-
http://ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
1
https://www.imdb.com/title/tt0133093
1
Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–10
disposition will be typically more vociferous compared ture extending as a result the digital awareness of the
to introvert ones for the same event. organizations deployed them. IAs often have to take de-
cisions [2, 3] which in turn rely on operational criteria
Table 1 based on aspects like anthropomorphism [4], maintain-
Notation Summary ing trust with human users [5], cognitive functionality
[6], connecting IAs with sensors [7], action explainabil-
Symbol Meaning First in
ity [8], and even the possible role of voice [9]. IAs can
△
= Equality by definition Eq. (1) be employed in many capacities like protecting critical
{𝑠1 , … , 𝑠𝑛 } Set with elements 𝑠1 , … , 𝑠𝑛 Eq. (2) industrial cyber-physical infrastructure [10], communi-
(𝑡1 , … , 𝑡𝑛 ) Tuple with elements 𝑡1 , … , 𝑡𝑛 Eq. (1) cating with humans through dynamic oral conversations
⟨𝑠𝑘 ⟩ Sequence with elements 𝑠𝑘 Algo. 1 [11], facilitating social interactions [12], recommending
|𝑆| Set, sequence, or tuple cardinality Eq. (4)
charging points for electric vehicles [13], modeling finan-
(𝑢, 𝑣; 𝑙) Edge from 𝑢 to 𝑣 with label 𝑙 Eq. (18)
Γ𝑖 (𝑣) Inbound neighborhood of 𝑣 Eq. (7)
cial markets [14], and even shaping fashion trends [15].
Γ𝑜 (𝑣) Outbound neighborhood of 𝑣 Eq. (7) In order for IAs to adapt to complex and nonstationary
prob {Ω} Probability of event Ω occurring Eq. (7) environments, ML capabilities have been recently added
𝑆1 ⧵ 𝑆 2 Asymmetric set difference Eq. (8) to them [16]. ML techniques cover a broad spectrum of
𝜑(⋅) Logistic function Eq. (12) options such as variational encoders [17], reinforcement
𝜓(⋅) Hyperbolic tangent function Eq. (14) learning [18], deep learning in various forms [19], and
⟨𝑔||𝑓⟩ KL divergence of 𝑔 from 𝑓 Eq. (20) cooperative learning [20]. Possible extensions for use
I𝑛 𝑛 × 𝑛 identity matrix Eq. (4) with IAs are tensor stack networks (TSNs) [21], GNNs
[22], adversarial neural networks (ANNs) [23], and self
The primary research objective of this conference pa- organizing maps (SOMs) [24].
per is the development of an IA determining its next Graph mining extracts nontrivial knowledge from
jump based on a strategy exploiting local structural in- linked data [25]. For example, approximating directed
formation as well as the MBTI profile of the neighboring graphs with undirected ones based on density criteria
vertices. The latter is achieved by a preprocessing stage [26]. Community discovery for large graphs has taken
where a GNN performs personality type prediction based many forms due not only to instance size [27] but also
on the standard MBTI profiles. Each given vertex has a because many and equally valid graph community def-
set of ground truth affective attributes which allow the initions exist [28, 29]. For instance, communities may
deployment of multiple psychological interfaces depend- well be build on trust [30], spatiotemporal patterns [31],
ing on the application and the type of IA querying the social behavior [32], multiple connectivity criteria [33],
vertex in question. Additionally, IAs take into considera- noiseless patterns [34], spatial behavior akin to that of
tion the MBTI personality of neighboring vertices prior geolocation services [35], privacy preserving constraints
to moving to one of them. The aforementioned factors [36], and simultaneous structural and affective criteria
differentiate significantly this conference paper from the [37]. Applications include trajectory planning for au-
vast majority of previous approaches. tonomous race vehicles [38], political [39] and commer-
The remainder of this conference paper is structured as cial [40] digital campaign designs, biomedical document
follows. In section 2 the recent scientific literature regard- recommendation based on a keyword-term-document
ing IAs, graph mining, and personality models are briefly tensor model [41], opinion mining [42], and assessing
reviewed. IA design is described in detail in section 3. the affective resilience of Twitter graphs [43, 44].
The results obtained by the action of the proposed IA on Personality indicators [45] go beyond emotion models
benchmark graphs are examined in section 4, whereas like the emotion wheel [46] or big five [47] since they
possible future research directions are given in section provide a framework for predicting reactions to a wider
5. Sets are denoted by capital letters. Bold capital letters array of stimuli [48], whereas the various emotion models
denote matrices, small bold vectors, and small scalars. only describe a single reaction. MBTI is used among
Acronyms are explained the first time they are encoun- others for designing digital campaigns on social media
tered in the text. In function definitions parameters are [49] perhaps in conjunction with other major attributes
separated of the arguments by a semicolon. Finally, in of the underlying domain [50]. Also MBTI taxonomy
table 1 the notation of this work is summarized. has been used to assess leadership traits [51]. Examples
include exploring political Twitter [52], predicting brand
loyalty [53], and analyzing the dynamics in social media
2. Related Work communities [54]. Tensor distance metrics [55] can be
used to cluster personalities in the MBTI space, especially
IA design relies on a number of approaches [1] as they when personalization is intended [56].
have to function on their own in large digital infrastruc-
2
Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–10
3. Intelligent Agent Design label in 𝐿 carries a different semantic value and denotes
relationships of varying strength. In the proposed ap-
3.1. Graph representation proach this is translated to different edges having priority
depending on their label, but this does not exclude the
Multilayer graphs in this work will be the algorithmic cor-
association of specific edge weights with different labels.
nerstone for representing and analyzing the interaction
between Twitter accounts and the IA as the latter will be
circulating along the edges of such a graph. Multilayer 3.2. MBTI personality system
graphs extend ordinary ones by including a set of dis- As mentioned earlier, MBTI is one of the most popular
tinct edge labels along with the requirement that an edge personality models stemming in large part from the the-
has exactly one and, thus, allowing as many as multiple ory of Jung and it aims at assessing the combination
edges between any two vertices as long as their labels are of individual cognitive functions. Specifically, in MBTI
distinct. As such, concepts such as vertex degrees and there are four independent axes, each corresponding to
paths have to be redefined to take into consideration this a function, with their two endpoints being personality
extension. As each label carries its own structural, func- attributes. This yields a total of sixteen basic personal-
tional, and semantic meaning, many underlying domains ity types known by the initials of their corresponding
can be better modeled. Definition 1 formally introduces attributes as shown in figure 1. Therein the respective
the class of labeled multilayer graphs. frequencies are also shown. Observe that in this case as
well holds the fundamental result of probability theory
Definition 1 (Labeled multilayer graphs). A labeled
stating that for every discrete distribution with 𝑛 data
multilayer graph 𝐺 expresses simultaneous connections
points there is at least one such point whose probability
between its vertices allowing the formulation of higher
is strictly more than 1/𝑛. In the context of MBTI this
order patterns and it is defined as the ordered triplet of (1):
implies there is at least one personality type which is
△
𝐺 = (𝑉 , 𝐸, 𝐿) (1) more common than the others.
The components of a multilayer 𝐺 and the role they
play in representing Twitter accounts and the associated ISTJ ISFJ INFJ INTJ
interactions and activities are the following: 12% 8% 4% 7%
• The set of vertices 𝑉 comprises of the entities the
relationships are built on. In the context of the ISTP ISFP INFP INTP
proposed methodology 𝑉 consists of anonymized 4% 3% 4% 4%
Twitter accounts.
• The set 𝐸 ⊆ 𝑉 × 𝑉 × 𝐿 is the set of labeled edges.
Therefore, not only does each edge have orien- ESTP ESFP ENFP ENTP
tation but also a label. In this work they denote 5% 5% 8% 5%
Twitter interaction.
• The label set 𝐿 depends directly on the semantics
of the underlying domain. In this particular case, ESTJ ESFJ ENFJ ENTJ
𝐿 contains the four elements of equation (2). 12% 8% 5% 6%
△
𝐿 = {𝑓 𝑜𝑙𝑙𝑜𝑤 , 𝑟𝑒𝑡𝑤𝑒𝑒𝑡 , 𝑚𝑒𝑛𝑡𝑖𝑜𝑛 , 𝑟𝑒𝑝𝑙𝑦} (2)
Each layer is formed by the edges of a single label, Figure 1: MBTI personalities.
therefore allowing in total |𝐿| layers. Each of them
represents different account interaction patterns.
The four axes representing basic cognitive functions
Definition 1 takes into consideration structural and or fundamental personality traits which determine the
functional graph characteristics. The former rely on com- MBTI taxonomy are the following:
binatorial properties, whereas the latter depend directly
• Extroversion vs Introversion: This axis deter-
on the Twitter activity. For the purposes of the analysis
mines how social an individual is. In turn this
done here edges have directions, capturing the one-way
determines the nature and amount of sensory in-
nature of Twitter interaction.
put an individual can cognitively process.
Each of the Twitter graph layers is formed by a specific
label and their endpoints. A given layer need not be sym- • Sensing vs iNtuition: This direction indicates the
metric and, in fact, such a layer is a special case. Each role tangible data play in a person’s cognitive pro-
3
Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–10
cess, namely whether they rely more on concrete The GNN performing the vertex classification works as
input or abstract terms. follows. It consists of 𝐿0 layers where each one acts like a
• Feeling vs Thinking: This variable denotes local spatial filter applied in parallel on the various graph
whether an individual tends to reason in order to segments similarly to a convolutional neural network
understand a situation or relies on accumulated (CNN). This is achieved through successive applications
experience in the form of hunches. of the same nonlinear mapping 𝜎(⋅), frequently reported
• Perceiving vs Judging: This factor finally shows as the activation function, on a linear transform of the
whether a person tends to place themselves inside graph adjacency matrix. In this equation P is the graph
a given situation during decision making process Laplacian matrix of (4), H𝑘 is the output of the previous
or they reason from a detached standpoint. layer, and W𝑘 is a trainable weight matrix.
The system described above captures a major part of M𝑘+1 = 𝜎(PH𝑘 W𝑘 ) (3)
human decision making process as it describes how and
which information is collected and furthermore how it is In (3) the graph Laplacian matrix P is defined as in (4),
processed. As stated earlier the personalities of the MBTI where A is the graph adjacency matrix and D is the graph
taxonomy are shown in figure 1 in such a way that ech neighborhood size matrix, namely the diagonal matrix
personality differs at exactly one trait from its neighbor- containing the total number of labeled edges whose end-
ing ones. This property in fact extends to personalities point is the respective vertex. Since Twitter graphs are
which are adjacent if the taxonomy map would wrap. directed, then there are two adjacency matrices, one for
Thus, personalities are encoded in a scheme similar to the inbound and one for the outbound edges. Moreover,
that of Gray coding allowing the easy grouping by factors since in this work there can be multiple edges between
similar to Karnaugh maps. the same vertex pair, both A and D contain the count of
In order to estimate the MBTI personality of each Twit- edges whose head or tail is the respective vertex. This
ter account in the dataset a GNN was used which oper- implies that also the diagonal of A should be bolstered
ated on the entire graph. Since each of the four variables with the maximum number of parallel edges, since each
of the taxonomy are independent, their estimation was vertex is strongly connected with itself. Instead of us-
recast as four separate instances of vertex classification. ing an one-hot encoding for each of personality type, a
The ground truth state vector contains the following vector with four entries was used as it is a more natural
attributes, which can be linked to how the basic person- representation for the MBTI taxonomy.
alities of the MBTI taxonomy manifest themselves. △
P = D−1/2 (A + |𝐿|I𝑛 )D−1/2 (4)
• Average number of connections.
• Average number of characters. As a result of the above, there are two CNNs, one
• Average number of nonwhite characters. operating on inbound and another on outbound neigh-
borhoods. The nonlinear activation function 𝜎(⋅) shown
• Fraction of alphabetical characters.
in (5) is applied elementwise to the matrix argument.
• Fraction of digits.
△
• Fraction of uppercase characters. 𝜎(𝑥) = ln (𝑒 𝛽0 𝑥 + 1) (5)
• Fraction of white spaces.
• Fraction of special characters. The synaptic weights of matrix W𝑘 are also individu-
• Average number of words. ally updated using the delta rule of equation (6). Specifi-
cally, each element of the weight matrix in the 𝑘-th layer
• Fraction of unique works.
receives a correction term of the form:
• Number of long words.
• Average word length. 𝛽0 𝑒 𝛽0 M𝑘+1 [𝑖,𝑗]
• Number of unique stopwords. ΔW𝑘 [𝑖, 𝑗] = 𝜂0 ( )M𝑘+1 [𝑖, 𝑗] (6)
1 + 𝑒 𝛽0 M𝑘+1 [𝑖,𝑗]
• Fraction of stopwords.
• Number of sentences. The learning parameter 𝜂0 decays with a cosine rate
• Number of long sentences (at least 10 words). whose frequency 𝜔0 depends on the training size 𝑝0 .
When both CNN run and the personality types are
• Average number of characters per sentence.
obtained, then in case there is a difference between the
• Average number of words per sentence.
two CNNs for the value a specific variable, it is decided
• Percentage of positive words. by the majority of the values of the respective variables
• Percentage of negative words. of the neighboring vertices ignoring direction. When
• Percentage of neutral words. this is not possible, then second order neighborhoods are
also included, again ignoring direction.
4
Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–10
3.3. Jump strategy vertex 𝑢 is obtained if both the numerator and the de-
nominator of equation (7) are divided by |Γ𝑖 (𝑣)|.
The IA proposed here moves along the edges of the graph
following a heuristic mechanism for community struc- △ |Γ𝑖 (𝑠)|
ture discovery which eventually approximates under a 𝑔(𝑣) = ∑ (8)
|Γ
𝑠∈Γ (𝑢)⧵𝑣 𝑖
(𝑣)|
set of mild assumptions a homogeneous Markov chain 𝑜
steady state distribution. IA starts from an arbitrary ver- If the middle form of equation (7) is used, then |Γ𝑖 (𝑠)|
tex and visits other vertices with probability proportional can be efficiently approximated under mild conditions
to their inbound degree locally normalized. Progressively by large set cardinality estimators [57]. Alternatively,
this constructs a long sequence ⟨𝑠𝑘 ⟩ containing the ver- the right hand side form of (7) is an inspiration for ap-
tices visited by the IA, which can then be mined by any proximating 𝑔(𝑣) in (8) by 𝑔(𝑣),
̂ when now the ratio of
agglomerative clustering algorithm in order for the final the estimators is used or, less frequently, an estimation
community list to be derived. Longer sequences lead to of their ratio since ratio statistics are in general difficult
more reliable discovery as they contain a higher vari- to be derived. In any case, this needs to be done only
ety of subsequences. Thus there are more indications once for static graphs. In that case the rightmost form of
whether certain vertices belong together. In particular if equation (7) can be approximated as in equation (9):
triangles local are successfully identified, then the more
likely a community is to be properly discovered. Addi- 1 1
= = 1 − 𝜑(ln 𝑔(𝑣))
̂ (9)
tional patterns assisting in community discovery are: 1 + 𝑔(𝑣)
̂ 1 + 𝑒 ln 𝑔(𝑣)
̂
• Bridges: Losing a bridge always implies connec- In order to derive a probability distribution from the
tivity loss. Thus, two communities may be con- scores obtained by (7), the softmax map is used in this
nected with at least one bridge. work. Specifically, the transition probability from 𝑢 to an
• Articulations: They are bridge endpoints. As outbound neighbor 𝑣 can be computed as:
such, they are critical for connectivity and belong
to different communities. exp (1 − 𝜑(ln 𝑔(𝑣)))
̂
prob {𝑢 → 𝑣} = , 𝑠 ∈ Γ𝑜 (𝑢)
• Subsequences: Frequent subsequences are in- ∑𝑠 exp (1 − 𝜑(ln 𝑔(𝑠)))
̂
dicators that certain vertices should be grouped (10)
together, especially for shorter ones. Alternatively, given the probabilistic approximation
• Hubs and authorities: Both are central in com- (9), any decision rule determining the next vertex IA
munities. As such, they should be grouped with moves to which relies on the ratio of the transition prob-
the vertices they appear with. abilities from 𝑢 to two distinct outbound neighbors 𝑣 and
𝑤 can use the equivalent computation of equation (11):
In the limit the IA random walk approximates the
steady state distribution of a homogeneous Markov chain prob {𝑢 → 𝑣} 1 − 𝜑(ln 𝑔(𝑣))
̂
= (11)
as the vertex selection mechanism is solely determined by prob {𝑢 → 𝑤} 1 − 𝜑(ln 𝑔(𝑤))
̂
the local connectivity patterns between the current and
the outbound neighboring vertices. This also eliminates In (9) 𝜑(⋅) is the logistic function defined in equation
the effect of the starting vertex. (12). It is always positive and also it is the derivative of
In order to determine the next vertex to be visited the softplus function commonly used as the nonlinear
the IA makes a probabilistic decision based on a mecha- activation function in neural networks [58].
nism reminiscent of preferential attachment. Recall that △ 1 𝑒𝑥
according to the latter the probability of moving from 𝜑(𝑥) = −𝑥 = 𝑥 , 𝑥∈ℝ (12)
1+𝑒 𝑒 +1
vertex 𝑢 to an outbound neighbor 𝑣 is shown in equation
(7). Thus, each candidate vertex 𝑣 is selected with a prob- An important property of the logistic function which
ability proportional to its locally normalized inbound has been used in (9) is shown in (13). This property
degree. The normalization constant is the sum of the suggests a balance or a conservation law between the
in-degrees of every outbound neighbor of 𝑢. logistic function and its symmetric with respect to its
argument and the horizontal axis 𝑥 = 1/2.
|Γ𝑖 (𝑣)| 1
prob {𝑢 → 𝑣} ∝ = (7) 1 𝑒 −𝑥
∑ |Γ𝑖 (𝑠)| 1 + 𝑔(𝑣) 𝜑(−𝑥) = = = 1 − 𝜑(𝑥) (13)
𝑠∈Γ𝑜 (𝑢) 1 + 𝑒𝑥 1 + 𝑒 −𝑥
Moreover, notice that 𝜑(⋅) is a rescaled and shifted
In (7) the positive quantity 𝑔(⋅) which is a function
version of the hyperbolic tangent 𝜓(⋅) of (14). The latter is
of the vertex 𝑣 and of the union of the inbound neigh-
the Bayesian estimator of a bipolar ±1 bit in the presence
borhoods of the outbound neighborhood of the current
5
Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–10
of additive Gaussian white noise [59]. Thus the logistic the selection of parameter 𝜀0 there will be sufficient visits
function is the Bayesian estimator for a 0/1 bit. to obtain important information regarding the role of
each vertex. Because of the decision rule (18), vertices
△ △ 𝑒 𝑥 − 𝑒 −𝑥 which are the endpoints of important edges will be visited
𝜓(𝑥) = tanh (𝑥) = (14)
𝑒 𝑥 + 𝑒 −𝑥 more frequently, whereas vertices peripheral in large
The absolute derivative value of the logistic function communities, with low degree, or difficult to reach will
is bound as shown in equation (15). This means that any be progressively neglected up to a point.
approximation errors of 𝑠 ̂ are not magnified by 𝜑(⋅).
𝜏0 = |𝑉| log𝜀0 |𝑉| (19)
𝜕𝜑(𝑥) −𝑒 −𝑥
| | = |− | = |𝜑(𝑥)𝜑(−𝑥)| ≤ 1 (15)
𝜕𝑥 (1 + 𝑒 −𝑥 )2 Algorithm 1 IA operational framework
Extending (7) in order to include multiple labels gives Require: Maximum number of hops 𝜏0
(18). Therein the transition probability from 𝑢 to 𝑣 takes Ensure: Obtain vertex sequence ⟨𝑠𝑘 ⟩
into account the parallel edges wherever they exist with 1: place IA in a random vertex
additional sums raging over 𝐿. Alternatively, this can be 2: repeat
replaced by the cardinality of the set of edges of the form 3: for all outbound vertices 𝑣 do
(𝑢, 𝑣; 𝑙) for the various possible labels 𝑙 ∈ 𝐿. Additionally, 4: compute 𝑎𝑢,𝑣 from (16)
let locations of the MBTI personalities of 𝑢 and 𝑣 in figure 5: end for
1 be (𝑖𝑢 , 𝑗𝑢 ) and (𝑖𝑣 , 𝑗𝑣 ). Then the factor 𝑎𝑢,𝑣 (16) which 6: if forgetting factor is enabled then
depends on these locations is computed as follows: 7: compute 𝜆𝑢,𝑣 from (17)
8: end if
2 2 9: compute destination 𝑠 from (18)
(𝑖 − 𝑖 ) + (𝑗𝑢 − 𝑗𝑣 )
𝑎𝑢,𝑣 = exp (− 𝑢 𝑣
△
) (16) 10: move to 𝑠 and place 𝑠 in the vertex sequence
2
11: until 𝜏0 is reached
A last but optional modification to (7) is the addition of 12: return sequence ⟨𝑠𝑘 ⟩
a forgetting factor 𝜆𝑢,𝑣 which penalizes outbound neigh-
bors which are selected too frequently in favor of others. For static or slowly evolving graphs the jumps of IA
The rationale for taking into consideration past selec- can be cached such that 𝑔(⋅) can be efficiently approxi-
tions is that the IA should be able to escape very dense mated without a new aggregation of edges in the vicinity
segments of a particular community in order to explore of a vertex. Exploiting this locality may considerably
alternative paths inside the community or even move to accelerate the computations of IA.
other communities. The particular form of the forgetting
factor used here is shown in equation (17). Therein 𝑞𝑢,𝑣 is
the number of times 𝑣 has been selected as a destination 4. Results
for the IA from 𝑢 the last 𝛾0 times.
4.1. Implementation
△ 𝑞𝑢,𝑣
𝜆𝑢,𝑣 = 1 − (17) In figure 2 the components of the proposed system are
1 + 𝛾0
shown. Therein can be seen that the IA moves on a
With these observations, the original decision rule multilayer graph stored in a standalone Neo4j instance
of equation (7) is now modified to take into considera- along with the activity on it such as tweets and mentions.
tion the fact that the IA moves along a multilayer graph Another major element is the GNN which has been ex-
where vertices have their own personality according to ecuted as a preprocessing step and has given to each
the MBTI taxonomy giving equation (18). Moreover, the vertex its MBTI personality. The third component is the
forgetting factor is also present, but its use is optional. IA itself, which is lightweght as it only has to implement
the decision rule of (18) based on local input.
𝜆𝑢,𝑣 𝑎𝑢,𝑣 |{(𝑢, 𝑣; 𝑙)}| The operational parameters of the GNN and the IA are
prob {𝑢 → 𝑣} ∝ (18)
∑𝑣∈Γ𝑜 (𝑢) ∑𝑠∈Γ𝑖 (𝑣) |{(𝑠, 𝑣; 𝑙)}| shown in table 2 and they pertain to various equations
presented earlier in the text.
The maximum number of jumps 𝜏0 the IA is allowed The dataset used in this work has been collected by
to make is determined by (19), which is a mechanism the Twitter crawler used among others in [26]. The data
for eventually terminating the IA route. The rationale therein pertains to three different graphs constructed
behind that limit is that the IA must visit each vertex from topic sampling using a main hashtag. Specifically,
more than once, but not too many times. Depending on the three Twitter graphs are the following:
6
Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–10
Neo4j
LGNN
agent
+follow MBTI estimations
->retweet
@reply
@mention
#hashtags evaluation
jump decisions
MBTI estimations
Figure 2: System architecture.
Table 2 To form the CNN ground truth from the collected Twit-
GNN and IA Parameters ter dataset two different ways were used:
Parameter Value • By picking the MBTI type an account has posted
Number of layers 𝐿0 7 on their Twitter bio. This is the preferred way
Activation function 𝜎(⋅) Eq. (5) and it was used whenever it was possible.
Scaling 𝛽0 3/2 • By locating a reference to an MBTI personality
Frequency 𝜔0 2𝜋/(1 + 𝑝0 ) up to two words apart from the words I, am, feel,
Training size 𝑝0 3000 myself, me, and being.
Number of jumps 𝜏0 Eq. (19) Accounts with self-reported MBTI type were not in-
Factor 𝜀0 3/2 cluded in the vertex classification by the CNN.
Decision criterion Eq. (18)
Forgetting factor 𝜆𝑢,𝑣 Eq. (17)
Window size 𝛾0 4 4.2. Evaluation
The Kullback-Leibler divergence between the distribu-
tion of the MBTI types returned by CNN for each graph
• #Julia: This topic is about the Julia programming compared to the global reference distribution shown in
language. At the sampling time the next Julia de- figure 1 is shown in table 4. Recall that the divergence
veloper conference was about to begin and, more- between a distribution 𝑔 and a reference one 𝑓 when they
over, earlier that year a major language update are both discrete is computed by equation (20).
was released. Therefore, there was high interest
at the time for the particular topic with mostly △ 𝑔𝑘
positive or neutral feelings. ⟨𝑔||𝑓⟩ = ∑ 𝑔𝑘 log ( ) (20)
𝑘
𝑓𝑘
• #Windows11: At about the same time a major
news update about the upcoming Windows 11 The low divergence between the three empirical distribu-
was made to the public. This generated mostly tions and the reference one mean that the three graphs
positive sentiment, but also some negative ones are representative in terms of MBTI personalities.
concerning the removal of some of the expected Hashtag coherency is a functional way to assess the
features and the question about obsolete hard- community structural uniformity. The rationale behind
ware support still lingering among users. this approach is that a successful partitioning will re-
• #BlackList: Just before the beginning of the final sult in communities of accounts with similar interests
season of this the major hit it was announced as expressed by hashtags. Specifically, if the Tanimoto
that the main protagonist would not join the cast. coefficient 𝜌0 between the hashtag sets of two accounts
Furthermore this was aggravated by leaked plots is used as a distance metric, then the minimum 𝑑𝑚 and
where her character is removed in a way deemed the average 𝑑𝑎 intercluster distances can be used figures
anticlimactic by fans. These stirred considerable of merit. The results can be seen in table 5.
controversy among them. Table 5 should be read as follows. Three cases were
tested. The first was the decision rule of (18) without the
7
Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–10
Table 3
Twitter Social Graph Properties (from [26]).
Property #Julia #Win11 #BlackList Tweets #Julia #Win11 #BlackList
Vertices 143019 152231 122535 Polarity% (pos/neg) 45.11/2.67 27.25/29.13 45.67/52.77
Edges 9232117 8536771 8425224 Length (avg/std) 167.33/45.12 145.17/37.83 154.86/41.84
Avg i-deg 66.21 61.89 72.43 Distinct hashtags 1182 1263 1314
Avg o-deg 71.36 63.18 76.08 Hashtags (avg/std) 5.13/0.89 8.42/1.17 7.18/1.01
Triangles 2458114 2282375 2946268 Replies (avg/std) 14.22/5.17 11.22/3.76 19.46/6.22
Squares 1034216 100736 117874 Mentions (avg/std) 17.63/4.38 13.38/3.29 15.49/5.34
Diameter 17 21 16 Density (linear/log) 64.55/1.35 56.08/1.38 68.76/1.36
Table 4 edge frequencies, which in turn heavily rely on a local
MBTI Type Distribution Divergence. decision rule for selecting the destination vertex. More-
over, destination vertices with similar MBTI profile types
#Julia #Win11 #BlackList
are given priority, whereas vertices frequently selected
1.4413 1.2417 1.2215 may be optionally penalized in order to allow the IA to
escape from especially dense segments of the Twitter
graph. Experimental results indicate that both factors
Table 5
Minimum and average intercluster distances. increase the partitioning quality in terms of increased
minimum and average intercluster distance.
Metric #Julia #Win11 #BlackList Concerning future research directions, the most imme-
𝑑𝑚 0.7514 0.7816 0.7932 diate one is to apply the proposed approach on more and
𝑑𝑎 0.7833 0.7724 0.7832 larger benchmark Twitter graphs. Moreover, IAs may
work in parallel in cooperative or adversarial modes.
Metric +MBTI +MBTI +MBTI
𝑑𝑚 0.8532 0.8717 0.8365
𝑑𝑎 0.8433 0.8876 0.8666 Acknowledgments
Metric +factor +factor +factor
This conference paper is part of Project 451, a long term
𝑑𝑚 1 0.9615 0.9725 research initiative with a primary objective of develop-
𝑑𝑎 0.9918 1 0.9811 ing novel, scalable, numerically stable, and interpretable
higher order analytics.
Moreover, this work is part of Project Vega.
MBTI similarity factor 𝑎𝑢,𝑣 and the forgetting factor 𝜆𝑢,𝑣 .
The second was the same rule with only 𝑎𝑢,𝑣 and the third
was with both 𝑎𝑢,𝑣 and 𝜆𝑢,𝑣 . For all three cases 𝑑𝑚 and 𝑑𝑎 References
were collected and each was normalized with respect to
[1] R. Calegari, G. Ciatto, V. Mascardi, A. Omicini,
its respective maximum. This allows the percentage of
Logic-based technologies for multi-agent systems:
change between jump strategies to be shown. In light of
A systematic literature review, Autonomous Agents
this, enabling both factors result to better communities,
and Multi-Agent Systems 35 (2021) 1–67.
while excluding them both yields the worst ones.
[2] D. Ha, Y. Tang, Collective intelligence for deep
Finally, the role of forgetting factor 𝜆𝑢,𝑣 it was positive
learning: A survey of recent developments, Collec-
as it can be seen from the above metrics. This can be
tive Intelligence 1 (2022).
attributed to the fact that, although at first iy may look
[3] I. M. Enholm, E. Papagiannidis, P. Mikalef,
counter-intuitive, it has a linear and not an exponential
J. Krogstie, Artificial intelligence and business
decay, so it forces the IA to choose less likely outbound
value: A literature review, Information Systems
neighbors but not too often.
Frontiers (2021) 1–26.
[4] S. Moussawi, M. Koufaris, R. Benbunan-Fich, How
5. Future Work perceptions of intelligence and anthropomorphism
affect adoption of personal intelligent agents, Elec-
This conference paper focuses on the development of an tronic Markets 31 (2021) 343–364.
intelligent agent (IA) which performs random walks on [5] T. Kim, H. Song, How should intelligent agents
Twitter multilayer graphs with the purpose of estimating apologize to restore trust? Interaction effects be-
8
Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–10
tween anthropomorphism and apology attribution M. Dunion, E. Fosong, S. Garcin, S. Guo, B. Gyevnar,
on trust repair, Telematics and Informatics 61 T. McInroe, et al., Deep reinforcement learning for
(2021). multi-agent interaction, AI Communications (2022)
[6] Z. Nagoev, I. Pshenokova, O. Nagoeva, Z. Sundukov, 1–12.
Learning algorithm for an intelligent decision mak- [20] I.-J. Liu, U. Jain, R. A. Yeh, A. Schwing, Coopera-
ing system based on multi-agent neurocognitive ar- tive exploration for multi-agent deep reinforcement
chitectures, Cognitive Systems Research 66 (2021) learning, in: ICML, PMLR, 2021, pp. 6826–6836.
82–88. [21] G. Drakopoulos, E. Kafeza, P. Mylonas, L. Iliadis,
[7] A. Gupta, S. Savarese, S. Ganguli, L. Fei-Fei, Embod- Transform-based graph topology similarity met-
ied intelligence via learning and evolution, Nature rics, NCAA 33 (2021) 16363–16375. doi:1 0 . 1 0 0 7 /
communications 12 (2021) 1–12. s00521- 021- 06235- 9.
[8] G. Vilone, L. Longo, Notions of explainability and [22] X. Zhou, W. Liang, W. Li, K. Yan, S. Shimizu, I. Kevin,
evaluation approaches for explainable artificial in- K. Wang, Hierarchical adversarial attacks against
telligence, Information Fusion 76 (2021) 89–106. graph neural network based IoT network intrusion
[9] K. Seaborn, N. P. Miyake, P. Pennefather, M. Otake- detection system, IEEE Internet of Things Journal
Matsuura, Voice in human–agent interaction: A (2021).
survey, CSUR 54 (2021) 1–43. [23] F. E. Fernandes Jr, G. G. Yen, Pruning of generative
[10] Z. Lv, D. Chen, R. Lou, A. Alazab, Artificial intelli- adversarial neural networks for medical imaging
gence for securing industrial-based cyber–physical diagnostics with evolution strategy, Information
systems, Future generation computer systems 117 Sciences 558 (2021) 91–102.
(2021) 291–298. [24] G. Drakopoulos, I. Giannoukou, S. Sioutas, P. My-
[11] E. C. Ling, I. Tussyadiah, A. Tuomi, J. Stienmetz, lonas, Self organizing maps for cultural con-
A. Ioannou, Factors influencing users’ adoption and tent delivery, NCAA 31 (2022). doi:1 0 . 1 0 0 7 /
use of conversational agents: A systematic review, s00521- 022- 07376- 1.
Psychology & Marketing 38 (2021) 1031–1051. [25] H. Motoda, K. Saito, Critical link identification
[12] S. K. Lee, P. Kavya, S. C. Lasser, Social interactions based on bridge detection for network with uncer-
and relationships with an intelligent virtual agent, tain connectivity, in: ISMIS, Springer, 2018.
International Journal of Human-Computer Studies [26] G. Drakopoulos, E. Kafeza, P. Mylonas, S. Sioutas,
150 (2021). Approximate high dimensional graph mining with
[13] W. Zhang, H. Liu, F. Wang, T. Xu, H. Xin, D. Dou, matrix polar factorization: A Twitter application,
H. Xiong, Intelligent electric vehicle charging rec- in: IEEE Big Data, IEEE, 2021, pp. 4441–4449. doi:1 0 .
ommendation based on multi-agent reinforcement 1109/BigData52589.2021.9671926.
learning, in: The Web Conference, ACM, 2021, pp. [27] G. Mei, X. Wu, Y. Wang, M. Hu, J.-A. Lu, G. Chen,
1856–1867. Compressive-sensing-based structure identification
[14] J. Lussange, I. Lazarevich, S. Bourgeois-Gironde, for multilayer networks, IEEE transactions on cy-
S. Palminteri, B. Gutkin, Modelling stock markets bernetics 48 (2017) 754–764.
by multi-agent reinforcement learning, Computa- [28] X. Su, S. Xue, F. Liu, J. Wu, J. Yang, C. Zhou, W. Hu,
tional Economics 57 (2021) 113–147. C. Paris, S. Nepal, D. Jin, et al., A comprehensive
[15] O. Allal-Chérif, V. Simón-Moya, A. C. C. Ballester, survey on community detection with deep learn-
Intelligent purchasing: How artificial intelligence ing, IEEE Transactions on Neural Networks and
can redefine the purchasing function, Journal of Learning Systems (2022).
Business Research 124 (2021) 69–76. [29] G. Rossetti, R. Cazabet, Community discovery in
[16] S. Carta, A. Ferreira, A. S. Podda, D. R. Recupero, dynamic networks: A survey, CSUR 51 (2018) 1–37.
A. Sanna, Multi-DQN: An ensemble of deep Q- [30] G. Drakopoulos, E. Kafeza, P. Mylonas,
learning agents for stock market forecasting, Expert H. Al Katheeri, Building trusted startup teams from
systems with applications 164 (2021). LinkedIn attributes: A higher order probabilistic
[17] B. Wu, S. Nair, R. Martin-Martin, L. Fei-Fei, C. Finn, analysis, in: ICTAI, IEEE, 2020, pp. 867–874.
Greedy hierarchical variational autoencoders for doi:1 0 . 1 1 0 9 / I C T A I 5 0 0 4 0 . 2 0 2 0 . 0 0 1 3 6 .
large-scale video prediction, in: CVPR, IEEE/CVF, [31] S. van den Beukel, S. H. Goos, J. Treur, An adaptive
2021, pp. 2318–2328. temporal-causal network model for social networks
[18] M. L. Olson, R. Khanna, L. Neal, F. Li, W.-K. Wong, based on the homophily and more-becomes-more
Counterfactual state explanations for reinforce- principle, Neurocomputing 338 (2019) 361–371.
ment learning agents via generative deep learning, [32] P. Bindu, P. S. Thilagam, D. Ahuja, Discovering
Artificial Intelligence 295 (2021). suspicious behavior in multilayer social networks,
[19] I. H. Ahmed, C. Brewitt, I. Carlucho, F. Christianos, Computers in Human Behavior 73 (2017) 568–582.
9
Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–10
[33] M. Abdolali, N. Gillis, M. Rahmati, Scalable and [47] Q. Agbaria, A. A. Mokh, Coping with stress during
robust sparse subspace clustering using randomized the coronavirus outbreak: The contribution of big
clustering and multilayer graphs, Signal Processing five personality traits and social support, Interna-
163 (2019) 166–180. tional Journal of Mental Health and Addiction 20
[34] G. Drakopoulos, E. Kafeza, P. Mylonas, S. Sioutas, (2022) 1854–1872.
Process mining analytics for Industry 4.0 with [48] J. Juárez, C. Santos, C. A. Brizuela, A comprehensive
graph signal processing, in: WEBIST, SCITEPRESS, review and a taxonomy proposal of team formation
2021, pp. 553–560. doi:1 0 . 5 2 2 0 / 0 0 1 0 7 1 8 3 0 0 0 0 3 0 5 8 . problems, CSUR 54 (2021) 1–33.
[35] Y. Xie, M. Gong, S. Wang, B. Yu, Community discov- [49] N. Cerkez, B. Vrdoljak, S. Skansi, A method for
ery in networks with deep sparse filtering, Pattern MBTI classification based on impact of class com-
Recognition 81 (2018) 50–59. ponents, IEEE Access 9 (2021) 146550–146567.
[36] X. Zheng, Z. Cai, G. Luo, L. Tian, X. Bai, Privacy- [50] K. A. Nisha, U. Kulsum, S. Rahman, M. Hossain,
preserved community discovery in online social P. Chakraborty, T. Choudhury, et al., A compar-
networks, Future Generation Computer Systems ative analysis of machine learning approaches in
93 (2019) 1002–1009. personality prediction using MBTI, in: Computa-
[37] D. Drakopoulos, K. C. Giotopoulos, I. Giannoukou, tional Intelligence in Pattern Recognition, Springer,
S. Sioutas, Unsupervised discovery of semantically 2022, pp. 13–23.
aware communities with tensor Kruskal decompo- [51] P. E. Roush, L. Atwater, Using MBTI to understand
sition: A case study in Twitter, in: SMAP, IEEE, transformational leadership and self-perception ac-
2020. doi:1 0 . 1 1 0 9 / S M A P 4 9 5 2 8 . 2 0 2 0 . 9 2 4 8 4 6 9 . curacy, Military Psychology 4 (1992) 17–34.
[38] T. Stahl, A. Wischnewski, J. Betz, M. Lienkamp, [52] A. Khatua, A. Khatua, E. Cambria, Predicting politi-
Multilayer graph-based trajectory planning for race cal sentiments of voters from twitter in multi-party
vehicles in dynamic scenarios, in: ITSC, IEEE, 2019, contexts, Applied Soft Computing 97 (2020).
pp. 3149–3154. [53] L. Lalicic, A. Huertas, A. Moreno, M. Jabreel, Emo-
[39] V. K. Brändle, C. Galpin, H.-J. Trenz, Brexit as ‘pol- tional brand communication on Facebook and Twit-
itics of division’: Social media campaigning after ter: Are DMOs successful?, Journal of Destination
the referendum, Social Movement Studies 21 (2022) Marketing & Management 16 (2020).
234–253. [54] E. J. Choong, K. D. Varathan, Predicting judging-
[40] L. Dolega, F. Rowe, E. Branagan, Going digital? The perceiving of Myers-Briggs Type Indicator (MBTI)
impact of social media marketing on retail website in online social forum, PeerJ 9 (2021).
traffic, orders and sales, Journal of Retailing and [55] G. Drakopoulos, I. Giannoukou, P. Mylonas,
Consumer Services 60 (2021). S. Sioutas, On tensor distances for self organiz-
[41] G. Drakopoulos, A. Kanavos, I. Karydis, S. Sioutas, ing maps: Clustering cognitive tasks, in: DEXA,
A. G. Vrahatis, Tensor-based semantically-aware volume 12392, Springer, 2020, pp. 195–210. doi:1 0 .
topic clustering of biomedical documents, Compu- 1007/978- 3- 030- 59051- 2_13.
tation 5 (2017). doi:1 0 . 3 3 9 0 / c o m p u t a t i o n 5 0 3 0 0 3 4 . [56] D. Fernau, S. Hillmann, N. Feldhus, T. Polzehl,
[42] J. Serrano-Guerrero, F. P. Romero, J. A. Olivas, Towards automated dialog personalization using
Fuzzy logic applied to opinion mining: A review, MBTI personality indicators, Proc. Interspeech 2022
KBS 222 (2021). (2022) 1968–1972.
[43] G. Drakopoulos, E. Kafeza, P. Mylonas, S. Sioutas, [57] G. Drakopoulos, S. Kontopoulos, C. Makris, Even-
A graph neural network for fuzzy Twitter graphs, tually consistent cardinality estimation with appli-
in: G. Cong, M. Ramanath (Eds.), CIKM companion cations in biodata mining, in: SAC, ACM, 2016.
volume, volume 3052, CEUR-WS.org, 2021. doi:1 0 . 1 1 4 5 / 2 8 5 1 6 1 3 . 2 8 5 1 8 8 7 .
[44] G. Drakopoulos, I. Giannoukou, P. Mylonas, [58] Y. Chen, Y. Mai, J. Xiao, L. Zhang, Improving the
S. Sioutas, A graph neural network for assessing antinoise ability of DNNs via a bio-inspired noise
the affective coherence of Twitter graphs, in: IEEE adaptive activation function rand softplus, Neural
Big Data, IEEE, 2020, pp. 3618–3627. doi:1 0 . 1 1 0 9 / computation 31 (2019) 1215–1233.
BigData50022.2020.9378492. [59] S. Cheng, H. Zhan, H. Yao, H. Fan, Y. Liu, Large-
[45] S. Štajner, S. Yenikent, Why is MBTI personality scale many-objective particle swarm optimizer with
detection from texts a difficult task?, in: European fast convergence based on alpha-stable mutation
Chapter of the Association for Computational Lin- and logistic function, Applied Soft Computing
guistics: Main Volume, 2021, pp. 3580–3589. (2020) 106947.
[46] X. Zeng, Q. Chen, X. Fu, J. Zuo, Emotion wheel
attention-based emotion distribution learning, IEEE
Access 9 (2021) 153360–153370.
10