Scenario Generation With Transitive Rules for Counterfactual Event

Analysis

Aigerim Mussina

1 a

, Paulo Trigo

2 b

and Sanzhar Aubakirov

1 c

Department of Computer Science, Al-Farabi Kazakh National University, 71 al-Farabi Ave., Almaty, Kazakhstan

GuIAA, ISEL - Instituto Superior de Engenharia de Lisboa, Lisbon, Portugal

Keywords:

Event Detection, Association Rules, What-If Analysis.

Abstract:

Event detection on online social networks is one of the comprehensive approaches for analyzing people’s dis-

cussions. However, it is not enough to detect an event as people often look for ways to inﬂuence the course

of an event. Often, in the course of a discussion, the introduction of a new topic can shift the focus to another

subject and thus move from one event to another. The causal relationship between topics and events can be ex-

plored by extracting association rules among the topics covered in each event. The scenario generation based

on those causal relationships can support what-if (counterfactual) analysis and explain transitions between

events. In this paper our goal is to generate what-if scenarios among topics of detected events. The association

rule approach was chosen as a method for its human-readable output that can be transposed into a counter-

factual scenario. We propose methods for time-window constrained topic-based what-if scenario generation

founded on market-basket analysis.

1 INTRODUCTION

Nowadays people are often immersed in a continu-

ous stream of textual data generated from social net-

works. A large volume of data usually emerges (and

grows) from people’s discussions and around aggre-

gating concepts commonly designated as “events”.

Also, within en event, the people’s discussions unfold

around certain “topics”. There is thus an increase in

research effort in the processing of textual data origi-

nating from social networks, with the goal of automat-

ically detecting both the events and their topics. Re-

searchers are interested in detecting events-and-topics

as it appears to become a powerful method for the

follow-up of people’s discussions. However, just de-

tecting events-and-topics is not enough.

The follow-up of people’s discussions also in-

volves the challenge of trying to predict the cause-

and-effect relationship that results from introducing a

new topic into a discussion. People usually (and in-

tuitively) seek to understand not only the source-and-

ﬂow of a discussion but also the impact that their own

participation, via the introduction of a topic, may have

https://orcid.org/0000-0002-7043-0810

https://orcid.org/0000-0001-5850-615X

https://orcid.org/0000-0002-8416-527X

on that same ﬂow.

Therefore, there is a “topic space” where each per-

son searches for a subset of topics through the gene-

ration of different scenarios in order to decide how to

best intervene in the ﬂow of a discussion. In this con-

text, a scenario is constructed by combining different

topics that may have originated from the same event

(intra-event scenario) or from different events (inter-

event scenario).

The overall process starts with the event detection

process being applied in the context of an online so-

cial network (OSN). Therefore, an event is described

by a set of topics-of-interest (ToI) that, in turn, were

addressed by people, in the course of their OSN inter-

actions (discussions) during a given time-window.

We formally deﬁne an event as E(w, T ) =

, ...,t

}, where w is a time-window, T is a ToI

dictionary and the {t

, ...,t

} ⊆ T is a topic-set;

hence, each t

∈ E(w, T ) represents the topic, x, as

taken from T and addressed, over w, in the detected

event E. We point out that the event detection algo-

rithm resorts to a ToI dictionary in order to improve

the process accuracy (Mussina et al., 2022).

The counterfactual analysis was chosen as a base

approach to the search for “topic space”. It aims

to explore cause-and-effect relations by searching for

statements such as “if A occurred, then B is also

Mussina, A., Trigo, P. and Aubakirov, S.

Scenario Generation With Transitive Rules for Counterfactual Event Analysis.

DOI: 10.5220/0011895000003393

In Proceedings of the 15th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2023) - Volume 3, pages 1047-1051

ISBN: 978-989-758-623-1; ISSN: 2184-433X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

1047

likely to occur” (Menzies and Beebee, 2001). Con-

sidering A and B as events and E deﬁnition (above),

we may rephrase such a statement as “if, within

a time-window, w, topics {t

, ...,t

} occurred,

then topics {t

, ...,t

} are also likely to oc-

cur in w”; where t

, here simply means that topic

x was addressed in event E. Following the idea

of what-if perspective, we suppose that A’s topic-

set, {t

, ...,t

}, includes “topic space” that have

been intervened such that discussion goes to B’s topic-

set, {t

, ...,t

In this paper, we follow the counterfactual per-

spective and propose methods for time-window con-

strained topic-based what-if scenario generation.

Generated scenarios have two types: intra-event and

inter-event. The inter-event scenario generation in-

cludes two approaches: a one-rule-based approach

and a two-rules-transitivity-based approach. We fol-

lowed a market-basket analysis approach and the pro-

posed methods resort to (unsupervised) association

rule extraction techniques when applied to detected

events (baskets) and the sets of corresponding topics

discussed therein (item-sets within baskets).

2 RELATED WORK

An approach to the event analysis process (that fol-

lows from event-detection) falls into the broader re-

search ﬁeld of “event association extration”. Usually

researchers resort to graph-based formulations, where

events are usually modelled as nodes (Shahaf et al.,

2015) and the value of edges between nodes (events)

are computed from the frequency of common words.

The “Connecting the dots” concept with graph neu-

ral networks outperformed the results of the state-of-

the-art methods (Wu et al., 2020). Analyzing event

associations one could extract cause-and-effect rela-

tionship.

Another recent research approach is the “counter-

factual event analysis” which is being achieved in the

healthcare ﬁeld where researchers are focused on the

prediction of the outcome of treatments (Zou et al.,

2022). Research aims not only to predict risks of

a treatment, but also to weight the cause-effect rela-

tion of alternative interventions (Prosperi et al., 2020).

Another approach is related to the “prescriptive anal-

ysis” that ﬁnds relevant applications in business man-

agement processes and decision making (Lepenioti

et al., 2020).

The main limitations of modern solutions in pre-

scriptive and counterfactual analysis are both the: a)

the generation of models that are complex to explain,

and b) focus on narrow areas of application. In this

work we apply a market-basket analysis, which pro-

vides, as a result, a human readable model based on

association rules. Such a model is used to generate

what-if counterfactual scenario. In this context, we

could not compare results with other researchers right

now.

3 COUNTERFACTUAL EVENT

ANALYSIS

The market-basket analysis approach is formalized

with the following main concepts: basket, item,

itemset, association rule, left-hand-side (LHS), right-

hand-side (RHS), support and conﬁdence. In a con-

text of event analysis, the detected event is consid-

ered as a basket and topic is considered as an item.

The itemset I = {t

, . . . , t

} is a, k size, set of items

where each t

is a topic (item). The association rule is

constructed from subitemsets and described as impli-

cation of the form LHS ⇒ RHS, where LHS, RHS ⊂ I

and LHS ∩ RHS =

0. The support and conﬁdence are

metrics that each association rule satisﬁes. The sup-

port is the joint probability, P(LHS, RHS), that items

in LHS and RHS occur together. The conﬁdence is a

conditional probability of the form P(RHS|LHS).

Deﬁnition 1. A what-if scenario is generated from

an itemset I

as an association rule, LHS ⇒ RHS, that

takes the form:

Base

∪W I

⇒ RHS

(1)

where Base

∪ W I

= LHS and Base

∩ W I

0, also LHS ⊂ I

and RHS ⊂ I

; each L, M and R

subscript is the size of the respective subitemset.

In this way the counterfactual scenario perspective

can be read as “if W I

occurs together with the Base

then RHS

is also likely to occur”.

Deﬁnition 2. An intra-event scenario is taken

from a what-if scenario where support(W I

) =

support(Base

) = support(RHS

)

In the intra-event scenario the equal support

means that subitemsets are likely to appear together

in each event. This comes from a small time-window

event detection where distinct events describe the

same real-world discussions.

Deﬁnition 3. A one-rule-based inter-event scenario

is taken from a what-if scenario where:

• support(W I

) is the minimum from all subitem-

sets of size M,

• support(Base

) is the maximum from all

subitemsets of size L,

• support(RHS

) is the maximum from all

subitemsets of size R.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

1048

Events may have common topic-subsets but they

actually represent different real-world discussions. In

that case, we use the inter-event scenario that can

show the path from one event to another. The asso-

ciation rules extraction for the inter-event scenario is

focused on topics that appear in several events. A

threshold value, h, is used to control the number of

events with common topics.

Deﬁnition 4. A two-rules-transitivity-based inter-

event scenario is taken from two association rules of

the form:

Base

⇒ W I

W I

⇒ RHS

(2)

where M > L, M > R and Base

,W I

⊂ I

W I

, RHS

⊂ I

, I

and I

are different itemsets.

Analysis of two rules at once is another approach

of inter-event scenario generation. Here we try

to pass from Base

(left-hand-side) to RHS

(right-

hand-side) through W I

(transitive subitemset). The

key feature is the association rule’s conﬁdence anti-

monotone property. It states that conﬁdence is anti-

monotone with respect to the number of items on the

right-hand-side of the rule. i.e., an increase in right-

hand-side dimension decreases or maintains its conﬁ-

dence. In our case, if M > L in Base

⇒ W I

then the

probability that W I

will happen when Base

hap-

pened is low. However, if we also have M > R in

W I

⇒ RHS

then the probability that RHS

will

happen when W I

happened is high. In this inter-

event scenario we pass from the low conﬁdence rule

to the high conﬁdence rule.

4 RESULTS

An overall view of our work is depicted in ﬁgure

1. The left side of the ﬁgure represents our pre-

vious work on the event detection where we ex-

plored two input data sources – corpora from Twit-

ter “Events2012” and Topics (ToI) (McMinn et al.,

2013) – and implemented the SEDTWik (Morabia

et al., 2019) algorithm for event detection replac-

ing Wikipedia with ToI dictionaries (Mussina et al.,

2022). Since event detection method uses frequency

counting to determine the burst in the use of cer-

tain words, it was necessary to remove the noise and

reduce each word to the same form. Tweets were

cleaned of stop words and converted to their stemmed

form. The NLTK library was used in the process of

cleaning tweets from stop words and stemming (with

PorterStemmer) (NLTK, 2022). The detected events

were combined into one ﬁle for each ToI dictionary,

where each line corresponds to an event.

Figure 1: Overall view of the work.

The right side of the ﬁgure 1 represents this pa-

per’s current work. The association rules extraction

resorts to the market-basket approach being applied

to the detected events. Given that the approach may

identify a large number of rules, it was limited by

the number of rules and minimum support and con-

ﬁdence. Therefore, the support-and-conﬁdence space

was explored to extract groups with at most 10 rules.

From those extracted rules we generate the scenar-

ios. As it was mentioned above the scenarios could

be intra-event and inter-event. Below we describe the

scenarios’ generation details along with preliminary

results.

4.1 Intra-Event Associations

Since detected events have a small number of top-

ics and the number of words in a “topic space” is

enormous, we faced many rules being extracted from

one event. One such rule, where the support of each

subitemset is about 0.029, is presented in the table

1. From that rule, the following counterfactual sce-

nario was generated: “if ‘shrine’ occurs together with

the ‘nearbi’, ‘ﬂood’ then ‘sanctuari’, ‘evacu’ are also

likely to occur”. This scenario can be interpreted as

“if a ﬂood occurs near the shrine, then people from

the sanctuary should be evacuated”.

4.2 Inter-Event Associations

The ﬁrst tests showed that the size of topic-sets should

be increased to extract the inter-event association

rules. The subitemsets support was lower than 0.1,

Scenario Generation With Transitive Rules for Counterfactual Event Analysis

1049

Table 1: Association rules for one real-world event.

Base

W I

RHS

itemsets ‘nearbi’,

‘ﬂood’

‘shrine’ ‘sanctu-

ari’,

‘evacu’

support 0.0294 0.0294 0.0294

which means that the probability of topics appearing

together is tiny. The average topic-set size was 9,

which is not enough to search for common topics in a

certain number of events. For the inter-event scenario

generation, the topic-set expanded by tweets contain-

ing topics, which construct event clusters in event de-

tection. The added tweets were cleared from non-ToI

dictionary topics. As a result, the average topic-set

size increased to 453.

In the context of an inter-event scenario genera-

tion, we use topics that appear in several events. If the

number of events containing a topic is greater than the

threshold value, h, then this topic will be included in

the itemset, I. We intuitively set the threshold value

as h = 10.

4.2.1 One-Rule-Based Approach

This subsection provides two examples. The ﬁrst ex-

ample is from the dictionary “Armed Conﬂicts and

Attacks”, and the second is from “Arts, Culture and

Entertainment”. For the ﬁrst experiments, the follow-

ing sizes of subitemsets of the what-if scenario were

intuitively chosen: L = 2, M = 1, and R = 2. The

extracted association rules corresponding to the inter-

event scenario deﬁnition are presented in the table 2.

Table 2: Association rules support - example 1.

Base

W I

RHS

itemsets ‘news’,

‘kill’

‘car’ ‘least’,

‘bomb’

support 0.6712 0.4521 0.4246

Table 3: Association rules support - example 2.

Base

W I

RHS

itemsets ‘news’,

‘kill’

‘car’ ‘least’,

‘bomb’

support 0.6712 0.4521 0.4246

It is necessary to identify events from the ex-

tracted rules in order to better understand the results.

We found which detected events contain the W I

subitemset. Detected events are listed in the table 4.

More detailed and readable descriptions of the real-

world situations are shown in the table 5. Descriptions

of the events taken from the open-source “Wikipedia

Current Events Portal” (Wikipedia, 2022).

Table 4: Detected events topics.

Detected Event A Detected Event B

Ex. 1 ’beirut’, ’car’,

’eight’, ’central’,

’explod’

’kill’, ’bomb’,

’syria’, ’turkey’,

’car’, ’attack’,

’suicid’, ’central’,

’wound’

Ex. 2 ’pope’, ’coptic’,

’egypt’, ’chris-

tian’, ’bishop’,

’chosen’, ’select’

’egypt’, ’chris-

tian’, ’chosen’,

’blindfold’,

’crystal’, ’copt’

Table 5: Real-world events description.

Real-world Event

Ex. 1 “A car bomb ex-

plodes at Sassine

Square in the

Lebanese capital

of Beirut, killing

at least eight

people ...”

“A car bomb

detonates in

Semdinli, Turkey,

killing 1 and

injuring 12...”;

“A suicide car

bomber detonates

a bomb in the

Hama province

of Syria ...”

Ex. 2 “Bishop Richard

Williamson is ex-

pelled from the

Society of Saint

Pius X (SSPX)

...”

“A shortlist of

successors to

the Coptic Pope

is drawn up; a

blindfolded child

is then expected

to pick from a list

of three. (BBC)”

From the obtained results, we could derive the fol-

lowing scenarios:

• “if ’car’ occurs together with the ’news’, ’kill’

then ’bomb’, ’least’ are also likely to occur”. If

there is a news about murder and additional infor-

mation as a car occurred, then we can suggest that

a bomb in a car exploded.

• “if ’choos’ occurs together with the ’pope’,

’egypt’ then ’christian’, ’coptic’ are also likely

to occur”. Even though events in real-world are

not strictly connected, it is interesting that Event

A was about exclusion of bishop, Event B was

about including new members in the Coptic Pope.

Events connected by topic ’choos’. This topic is a

stemmed version of word choose.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

1050

4.2.2 Two-Rules-Transitivity-Based Approach

In this subsection we present an example of a two-

rules-transitivity-based inter-event scenario genera-

tion. The example is presented in the table 6.

Table 6: Two-rules-transitivity-based approach example.

LHS RHS

Rule 1 ‘least’, ‘syria’ ‘news’, ‘car’,

‘kill’, ‘bomb’

Rule 2 ‘news’, ‘car’,

‘kill’, ‘bomb’

‘soldier’,

‘sever’

Rule 1, from the table 6, describes an event with

description: “A car bomb explodes at Sassine Square

in the Lebanese capital of Beirut, killing at least eight

people and wounding up to 78 others. (BBC)”. Rule 2

describes an event with description: “Syrian civil war:

A Jordanian soldier dies during a gunﬁght between

Jordanian troops and Islamic militants attempting to

cross the border into Syria. (CTV News)”.

From that rules, the following counterfactual sce-

nario was generated: “if ‘news’, ‘car’, ‘kill’, ‘bomb’

occurs together with the ‘least’, ‘syria’ then ‘soldier’,

‘sever’ are also likely to occur”.

5 CONCLUSION

In this paper, we proposed methods for time-window

constrained topic-based what-if scenario generation,

in the counterfactual perspective, founded on market-

basket analysis and association rules extraction. Def-

initions of counterfactual scenarios, both intra-event

and inter-event, are given. Preliminary results illus-

trate the extraction of coherent causal effects and re-

quire more analysis and controlled experiments. Fu-

ture work will apply the methods to events detected

using other ToI dictionaries. We will also include

evaluating the proposed methods in the context of the

usefulness of association rules and scenarios.

ACKNOWLEDGEMENTS

This work was supported by the grant of the Min-

istry of Science and Higher Education of the Republic

of Kazakhstan, project BR10965311 “Development

of intelligent information and telecommunication sys-

tems for urban infrastructure: transport, ecology, en-

ergy, and data-analytics in the Smart City concept”.

REFERENCES

Lepenioti, K., Bousdekis, A., Apostolou, D., and Mentzas,

G. (2020). Prescriptive analytics: Literature review

and research challenges. International Journal of In-

formation Management, 50:57–70.

McMinn, A. J., Moshfeghi, Y., and Jose, J. M. (2013).

Building a large-scale corpus for evaluating event de-

tection on twitter. In Proceedings of the 22nd ACM in-

ternational conference on Information & Knowledge

Management, pages 409–418.

Menzies, P. and Beebee, H. (2001). Counterfactual theories

of causation.

Morabia, K., Murthy, N. L. B., Malapati, A., and Samant,

S. (2019). Sedtwik: segmentation-based event detec-

tion from tweets using wikipedia. In Proceedings of

the 2019 Conference of the North American Chapter

of the Association for Computational Linguistics: Stu-

dent Research Workshop, pages 77–85.

Mussina, A., Aubakirov, S., and Trigo, P. (2022).

Parametrized event analysis from social networks.

Scientiﬁc Journal of Astana IT University, 10(10).

NLTK (2022). NLTK :: sample usage for stem. https:

//www.nltk.org/howto/stem.html. [Online; accessed

09-November-2022].

Prosperi, M., Guo, Y., Sperrin, M., Koopman, J. S., Min,

J. S., He, X., Rich, S., Wang, M., Buchan, I. E., and

Bian, J. (2020). Causal inference and counterfactual

prediction in machine learning for actionable health-

care. Nature Machine Intelligence, 2(7):369–375.

Shahaf, D., Guestrin, C., Horvitz, E., and Leskovec, J.

(2015). Information cartography. Communications of

the ACM, 58(11):62–73.

Wikipedia (2022). Wikipedia Current Events Portal. https:

//en.wikipedia.org/wiki/Portal:Current\ events/. [On-

line; accessed 15-December-2022].

Wu, Z., Pan, S., Long, G., Jiang, J., Chang, X., and Zhang,

C. (2020). Connecting the dots: Multivariate time se-

ries forecasting with graph neural networks. In Pro-

ceedings of the 26th ACM SIGKDD international con-

ference on knowledge discovery & data mining, pages

753–763.

Zou, H., Li, B., Han, J., Chen, S., Ding, X., and Cui,

P. (2022). Counterfactual prediction for outcome-

oriented treatments. In International Conference on

Machine Learning, pages 27693–27706. PMLR.

Scenario Generation With Transitive Rules for Counterfactual Event Analysis

1051