This position paper proposes a platform for the creation of digital twins for social networks as semantic digital twins for people. These are mainly aimed at simulating human behavior from a cognitive point of view. The proposal relies on a semantic data infrastructure aimed at analytical purposes, which is directly fed with real data from social networks. Summarized data and data generation methods are then combined to produce new data streams according to the analyst requirements. All these data are stored in a dynamic knowledge graph, which plays a central role in the design of the digital twins. First experiments will be conducted on two scenarios where semantic data is already available, namely: Tourism and Fashion.
Digital Twins (DT) [
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
DTs are starting to be general-purpose tools, but the adoption of AI and DT
is hardly visible these days. In fact, Gartner market research predicted that by
2022 [
In this position paper, we present our approach for developing a platform that enables the de nition of DTs from cognitive and social network data as semantic digital twins for people. The main aim of the intended DTs is to simulate human behavior from a cognitive/social point of view. More speci cally, these DTs will rely on AI cognitive models which will be derived from social networks data. As a result, we will be able to simulate di erent situations and analyze the e ectiveness of decisions taken (what-if analysis). The proposed platform will be applied to two verticals, namely: Tourism and Fashions. 2
In this section recent work related to the development of DTs and Cognitive DTs
are presented, including DTs for social media. The proposal of this paper will
follow this last trend, developing a Social and Cognitive DT. For developing the
new cognitive and social digital twins it is necessary to capture and analyse data
from di erent platforms and sources which might be heterogeneous in syntax,
schema, or semantics, making data integration di cult. Therefore, the new DT
will make use of the dynamic SLOD-BI platform [
The paper [
In [
[
Finally, on the context of social media, [
The main motivation behind SLOD-BI (Sentiment Linked Open Data for
Business Intelligence) was to build a data infrastructure aimed at sharing extracted
sentiment data from social networks [
The main drawback of SLOD-BI is that it focused on generating static
datasets like other LOD projects. However, social networks are extremely
dynamic which are not well suited for LOD nor BI tools. Instead, social data
must be regarded as a continuous stream where dimensions continuously change.
Thus, we proposed Dynamic SLOD-BI [
In this paper, the main goal is adapting this infrastructure in order to build
digital twins of social network streams. More speci cally, we aim at re-using
semantic and summarized data from SLOD-BI to simulate new data streams
coping with some speci c constraints and parameters.
Generative models are machine learning methods that estimate the join
distribution of target and training data. This learned distribution can be used to
generate new data similar to the data the models were trained on. The current
methods have matured to the point in which they are able to generate very high
quality data, text and images. They have produced many practical applications,
including highly visible ones such as the generation of photo-realistic images,
and they have also been used for non-image data such as time series [
Current techniques are almost universally based on Deep Learning. Neural
approaches for the development of generative models including Deep Belief
Networks Boltzmann Machines, Variational Autoencoders and transformers, which
are used for text generation [
The main challenge in this project is how to adapt existing techniques to the speci c needs of a social network DT. 3
The knowledge graph (KG) includes all existing vocabularies for SLOD-BI as well
the new required vocabularies for designing DTs. Parameters and probabilistic
distributions will be taken from the SLOD-BI infrastructure, since they regard
the elements analysts targeted to. Gathered streamed data in SLOD-BI will serve
as a basis for estimating all required distributions that will guide the generation
of the new data streams. The KG must be extended in order to regard DT
generative methods and how they are linked to SLOD-BI concepts. This will
allow designers to choose the most appropriate generative methods depending
on the parameter settings. We will adopt a similar approach to that of BigOWL
[
This component aims at designing and implementing a speci c DT for a speci c scenario. The output of this component is a data stream simulating a real one but conditioned to a series of parameters and constraints. This component consists of three main modules, namely: (1) parameter setting, (2) a data generator composer, and (3) a data validator.
Parameter setting consists in de ning the shape of the distributions we aim at for each of the entities involved in the DT. For example, we can de ne the particular distribution of user pro les we want in the data stream, biasing towards journalists or professionals.
The generator composed will select the most appropriate methods to generate the intended data stream according to the knowledge expressed in the SLODDT subgraph. This part involves traditional distribution generators, multimedia content generators (e.g., text and images) as well as time series generators.
Probability distributions are taken by default from SLOD-BI data but it can be changed in the parameter setting module.
Due to the complexity of a social network stream, designing a DT requires combining many di erent generative methods to simulate how users, topics, events, posts, and so on occur in that data stream. SLOD-BI patterns can help in de ning the order in which di erent entities are generated and how they condition the other data to be generated. In Fig. 3 we show an example of a composed data generator following the boilerplate notation. Parameters are represented with and can be tuples of complex parameters. For example, post is composed of the parameters for text, image, hashtag and mentions parameters. In our approach, we will regard both traditional probabilistic generation models and current state-of-the-art generators based on deep learning. 3.3
As the data streams are randomly generated, we need to check that they do not contain inconsistencies, impossible values and incoherent contents. For this reason, we propose to apply consistency rules expressed in OWL2-RL to validate the generated data. Inconsistent data will be removed or replaced till the generated data stream becomes consistent and coherent. Examples of these constraints are: the limit values for user and post metrics, a user can only give a like or retweet once a post, and a user cannot interact to its own posts. 3.4
There is a great variety of analytical tasks associated with social networks, for example: bot/spam detection, community discovery, user pro ling, event detection, identifying in uencers and checking data quality. Analytical tools aim at visualizing and detecting anomalies in data whereas predictive analytics are aimed at automatically classifying, predicting and recognising entities from data streams. Predictive analytics mainly rely on data-driven machine learning methods, which usually require many labeled examples. For both kinds of tools, the generation of simulated data is crucial for evaluating them in new scenarios before they are seen in real data streams. Dynamic SLOD-BI provides some of these tools which could be tested on the DTs outputs.
A rst scenario we want to address is that of Tourism. The intended DTs are
mainly aimed at simulating the human behavior from a cognitive perspective.
These DTs should be anthropomorphic representations of people who interact
with tourist facilities and express their feelings about them. The internal
structure of DTs will be designed by taking into account both cognitive and social
aspects which must be also present in the KG. Currently, some cognitive data
as well as image generation methods have been tested in this domain [
A second scenario is that of tracking fashion trends in social media. In this
scenario we want to recreate the behaviour of coolhunters and the possible
reaction of followers, for example to predict the stock of a new season after a new
advertising campaign. The main idea is to develop a DT able to recreate new
situations where image colors and text contents can be tuned according to some
unseen trend. A preliminary work in this direction was presented in [
In this paper we propose a new paradigm for a semantic-driven de nition of DTs for social networks. This proposal takes pro t from the summarized data gathered from a Business Intelligence data infrastructure to set the parameters of a DT following the analyst requirements.
Semantic Web technology plays a relevant role in this approach since the SLOD-BI data and DT parameters and constraints are expressed according to the provided vocabularies and ontologies. Moreover, the approach relies on a dynamic KG representation since social data are continuously changing. Data generation algorithms are also expressed in the KG and linked to the concepts and parameters that best suit them. In this way, the de nition and implementation of a DT is fully driven by the KG.
We plan to apply this proposal to some verticals already explored by the authors within the SLOD-BI project (e.g., automotive and medicine), as well new ones like Tourism and Fashion that would greatly bene t from social networks DTs. Preliminary results are expected soon for those domains where a good volume of data have been already gathered.
This project has been funded by the Ministry of Economy and Commerce with project contract TIN2016-88835-RET and by the Universitat Jaume I with project contract UJI-B2020-15.
Berlanga et al.