Special sessions

Non vector data are ubiquitous in contemporary data science applications. They take various forms, from simple categorical observations to dynamic graphs. In some cases, they also integrate a temporal component. Because of the geometry implied by their quantized nature, non vector data raise specific challenges in the context of abnormal behaviour detection. Consider for instance a time series which values are vertices of a graph: how can we extend the notion of change point detection in such a series? What about a time series of full graphs?

Numerous research directions have been explored in this context, including, among others the following ones:

• vector embedding
• dissimilarity based methods
• generative models
• kernel methods
• deep learning
• etc.

The focus of this special session is on anomaly, outlier and change point detection in non vector data. We welcome all types of non vector data: categorical data, texts, graphs and other (semi)structured data. For data without a temporal dimension, we are interested in novel methods to detect outliers and anomalies. For time series, we look for new change point detection methods. Applications papers are also welcome.

Machine Learning (ML) achievements enabled automatic extraction of actionable information from data in a wide range of decision-making scenarios (e.g. health care, cybersecurity, and education). ML models
are nowadays ubiquitous pushing even further the process of digitalization and datafication of the real and digital world producing more and more complex and interrelated data. This demands for improving both
ML technical aspects (e.g. design and automation) and human-related metrics (e.g. fairness, robustness, privacy, and explainability), with performance guarantees at both levels.

The aforementioned scenario posed three main challenges: (i) Learning from Complex Data (i.e. sequence, tree and graph data), (ii) Learning Trustworthily, and (iii) Learning Automatically with Guarantees. The scope of this special session is then to address one or more of these challenges with the final goal of Learning Trustworthily, Automatically, and with Guarantees from Complex Data.

Examples of methods and problems in these challenges are:

• efficient and effective models capable of directly learning from data natively structured or collected from interrelated heterogeneous sources (e.g. social and relational data, knowledge graphs), characterized by entities, attributes, and relationships, without relying on human skills to encode this complexity into a rich and expressive (vectorial) representation;
• design ML models from a human-centered perspective, making ML trustworthy by design, by removing human biases from the data (e.g. gender discrimination), increasing robustness (e.g. to adversarial data perturbation), preserving individuals’ privacy (e.g. protecting ML models from differential attacks), and increasing transparency (e.g. via ML models and output explanation);
• automatizing the ML design and deployment parts which are currently handcrafted by highly skilled and trained specialists. For this reason, ML is required to be empowered with self-tuning properties (e.g. architecture and hyperparameter automatic selection), understanding and guaranteeing the final performance (e.g. with worst case and statistical bounds) with respect to both technical and human relevant metrics.

The focus of this special session is to attract both solid contributions or preliminary results which show the potentiality and the limitations of new ideas, refinements, or contaminations between the different fields of machine learning and other fields of research in solving real world problems. Both theoretical and practical results are welcome to our special session.

In today’s world, important data is often distributed over different facilities or devices. E. g. in the context of the internet of things, smart devices, or the health care domain. In the latter one,
different hospitals and private doctors, will hold different information of the same patient and models can be derived in a federated learning scheme to avoid data transmissions and to ensure
privacy constraints.

Just collecting data from all facilities/devices as it is, is mostly forbidden due to data privacy reasons. One direction to address this issue is given by means of differential privacy concepts.

Another problem is, that in some scenarios, the data in some of the facilities are changing rapidly, and thus updated data would need to be sent to the model very frequently. Thus, minimizing communication effort is also a challenge. To address this issue, there currently exist different algorithms of the Federated Learning domain, like Federated Averaging (FedAvg), which address these problems.

This special session welcomes novel research and applications in the field of Federated Learning and beyond. Especially addressing the challenges of expensive communication, systems heterogeneity, statistical heterogeneity, and privacy concerns.

We encourage the submission of papers on novel Federated Learning methods by means of computational intelligence and machine learning approaches, including but not limited to:

• data analysis and pattern recognition approaches for federated environments
• preprocessing approaches in collaborative learning
• learning in heterogeneous systems
• effective communication for updating distributed models
• representation and modelling of distributed models
• approximation techniques for Federated Learning
• online and incremental learning (dimensionality reduction, classification, clustering and regression, outlier detection) with a particular design for federated environments
• Federated Transfer Learning algorithms
• Vertical Federated Learning algorithms
• Horizontal Federated Learning algorithms
• security and privacy preservation in federated environments
• differential privacy techniques
• model compression and adaptive model aggregation
• application of Deep Learning in the Federated Learning context
• particular interesting applications for Federated Learning e.g. in IoT, recommendation systems, medicine, sensor networks, text processing...

Graphs are very flexible data structures that can be used to represent, at the same time, entities and the relationships among them.Graphs can  describe networks of interacting elements, e.g. in social graphs or metabolomics, as well as data where topological variations influence the feature of interest, such as molecular compounds.  Data-driven processing and adaptive learning on structured data has a long-standing history but it has recently grown to become one of the most active research topics in the deep learning field.

This special session solicits contributions on the general topic of deep learning for graphs including wide themes such as graph representation learning, graph generation, interpretability of deep graph networks, learning for graphs as a means to integrate symbolic-subsymbolic information. We welcome works focusing on methodological advances, theoretical works, and also impacting applications of deep learning for graphs. Topics that are of interest to this session include, but are not limited to:

• Deep learning and representation learning for graphs
• Learning with network data
• Learning with knowledge graphs and symbolic-subsymbolic integration
• Graph generation
• Graph coarsening and pooling
• Deep graph networks and combinatorial algorithms/problems
• Interpretable models for graphs
• Randomized neural networks for relational data
• Relational deep learning
• Applications of learning for graphs: e.g. Natural language processing, machine vision, materials science, chemoinformatics, computational biology, social networks, federated learning, recommendation systems.

Machine learning models are currently dominated by neural networks and, in particular, by deep variants of those networks. Frequently, these models achieve promising results. However, usually deep networks act as black-box, as do many other machine learning models. Further, due to powerful tools, the learning process, which is often a gradient descent approach, is hidden for the developer as well as for the applicant of the model. Therefore, the possibilities to assess the network are mainly performance evaluations. However, this is problematic for many application for example in medicine, engineering and economy/finance applications, which require a transparent decision or prediction process.

Recently there has been considerable effort to develop interpretable models in machine learning and approaches to explain the decision/prediction processes to the user.

The aim of this special session is to make these new approaches and models more highly visible to the community. We invite papers highlighting different aspects of interpretable models and explaining decision support processes and inference models involving artificial intelligence. The session covers a broad range of this topic, ranging from theoretical considerations and new machine learning models to machine learning applications requiring or benefitting from interpretability and explainability.

Topics include, but are not limited to:

• Machine learning models with inherent interpretability
• Methods to explain existing models
• Model verification
• Visualization and visual inspection of the operation of machine learning models
• Confidence and trustworthiness in AI
• Prediction confidence and quantification of uncertainty
• Trade-off between interpretability and performance
• Model transparency in safety critical applications

We welcome both, new theoretical developments as well as practical applications.

Machine learning is increasingly applied to the study of Internet-delivered social media such as
Facebook, Twitter, YouTube, and Instagram. In the last several years these sites have been employed to
influence people with regard to political or social ideologies involving specific and targeted messages. In
the United States and elsewhere, hate groups advocating racist or antisemitic action often employ these
platforms to achieve influence.

This special session will attract scholars who are employing their knowledge of machine learning to
more efficiently and effectively process Internet content, including text, images, video, or combinations
thereof. Study of these phenomena may help to better understand everything from political protest to
advertising of products and services. Much of the work in “Measuring and Analyzing Online Social
Communication” undertaken currently is manual in nature. Machine learning tools for that work are
immature, however, they can advance the state of the art in the development of tools to automate the
study of social communications.

We encourage submission of papers on novel machine learning applied to Social Communication, they
include but are not limited to:

• Computer vision (including feature detection algorithms)
• Natural language approaches
• Online marketing
• Hate speech
• Text mining
• Political influence
• Propaganda
• Graph embeddings
• Censorship
• Summarization
• Document Classification
• Sentiment Analysis

The last few decades have seen a faster development of digital systems for observing the
mobility of people and goods. Various sensing systems - such as radio communication, Wi-Fi,
Bluetooth, validation of smart cards, mobile phone, and road traffic monitoring systems - have
enabled researchers and practitioners to acquire large amounts of data, which generally refer
to individual and collective trajectories. The mobility data can be further enriched with side
information, such as text corpora from social media, survey data, and weather information.
These massive data, temporally and spatially structured, can benefit from advanced machine
learning and data mining methods, providing decision aid tools, and contributing to the
development of safer, cleaner, and more efficient transportation systems. They can also help to
implement new mobility services for the user.

This special session will be an opportunity to present recent advances in mobility data analysis.
This field refers to descriptive methods, generally based on unsupervised learning, highlighting
the main uses, routines, mobility patterns, and provides a basis for anomaly detection. The
second field of interest in mobility data processing concerns predictive analytics, which
performs in a supervised learning framework to predict quantities of interest such as transport
demand or operating conditions. The implementation of these methods faces challenges related
to the incompleteness, heterogeneity, and strong temporal and spatial correlation within the
data, and their high dimensionality or volume. This special session is therefore concerned with
applications in the following fields of Intelligent Transport Systems and Mobility analysis:

• public transportation, sharing mobility systems, traffic analysis,
• mobility systems with low ecological impact,
• urban planning/land use,
• interactive visualization of urban mobility data,

taking advantage of original methodologies from the following fields:

• clustering, co-clustering and dimensionality reduction,
• spatiotemporal mobility pattern mining and interpretation,
• descriptive analysis of time series and trajectories,
• deep neural networks for supervised and unsupervised learning,
• anomaly detection,
• modeling and forecasting of time series,
• scalable methods for processing massive data.