Blaž Škrlj

Boshko Koloski, Senja Pollak, Roberto Navigli et al.

Large semantic knowledge bases are grounded in factual knowledge. However, recent approaches to dense text representations (i.e. embeddings) do not efficiently exploit these resources. Dense and robust representations of documents are essential for effectively solving downstream classification and retrieval tasks. This work demonstrates that injecting embedded information from knowledge bases can augment the performance of contemporary Large Language Model (LLM)-based representations for the task of text classification. Further, by considering automated machine learning (AutoML) with the fused representation space, we demonstrate it is possible to improve classification accuracy even if we use low-dimensional projections of the original representation space obtained via efficient matrix factorization. This result shows that significantly faster classifiers can be achieved with minimal or no loss in predictive performance, as demonstrated using five strong LLM baselines on six diverse real-life datasets. The code is freely available at \url{https://github.com/bkolosk1/bablfusion.git}.

Blaž Škrlj, Benjamin Ben-Shalom, Grega Gašperšič et al.

Field-aware Factorization Machines (FFMs) have emerged as a powerful model for click-through rate prediction, particularly excelling in capturing complex feature interactions. In this work, we present an in-depth analysis of our in-house, Rust-based Deep FFM implementation, and detail its deployment on a CPU-only, multi-data-center scale. We overview key optimizations devised for both training and inference, demonstrated by previously unpublished benchmark results in efficient model search and online training. Further, we detail an in-house weight quantization that resulted in more than an order of magnitude reduction in bandwidth footprint related to weight transfers across data-centres. We disclose the engine and associated techniques under an open-source license to contribute to the broader machine learning community. This paper showcases one of the first successful CPU-only deployments of Deep FFMs at such scale, marking a significant stride in practical, low-footprint click-through rate prediction methodologies.

Jan Kralj, Blaž Škrlj, Živa Ramšak et al.

Biological systems can be studied at multiple levels of information, including gene, protein, RNA and different interaction networks levels. The goal of this work is to explore how the fusion of systems' level information with temporal dynamics of gene expression can be used in combination with non-linear approximation power of deep neural networks to predict novel gene functions in a non-model organism potato \emph{Solanum tuberosum}. We propose DDeMON (Dynamic Deep learning from temporal Multiplex Ontology-annotated Networks), an approach for scalable, systems-level inference of function annotation using time-dependent multiscale biological information. The proposed method, which is capable of considering billions of potential links between the genes of interest, was applied on experimental gene expression data and the background knowledge network to reliably classify genes with unknown function into five different functional ontology categories, linked to the experimental data set. Predicted novel functions of genes were validated using extensive protein domain search approach.

Boshko Koloski, Blaž Škrlj, Marko Robnik-Šikonja et al.

The cross-lingual transfer is a promising technique to solve tasks in less-resourced languages. In this empirical study, we compare two fine-tuning approaches combined with zero-shot and full-shot learning approaches for large language models in a cross-lingual setting. As fine-tuning strategies, we compare parameter-efficient adapter methods with fine-tuning of all parameters. As cross-lingual transfer strategies, we compare the intermediate-training (\textit{IT}) that uses each language sequentially and cross-lingual validation (\textit{CLV}) that uses a target language already in the validation phase of fine-tuning. We assess the success of transfer and the extent of catastrophic forgetting in a source language due to cross-lingual transfer, i.e., how much previously acquired knowledge is lost when we learn new information in a different language. The results on two different classification problems, hate speech detection and product reviews, each containing datasets in several languages, show that the \textit{IT} cross-lingual strategy outperforms \textit{CLV} for the target language. Our findings indicate that, in the majority of cases, the \textit{CLV} strategy demonstrates superior retention of knowledge in the base language (English) compared to the \textit{IT} strategy, when evaluating catastrophic forgetting in multiple cross-lingual transfers.

Kristian Miok, Matej Klemen, Blaz Škrlj et al.

Explaining the predictions of neural models in clinical NLP remains a significant challenge, especially for complex tasks involving long, unstructured medical texts. While post-hoc methods like LIME and SHAP are widely used, they often fall short when applied to clinical narratives. In this paper, we identify core limitations of token-level and perturbation-based explanation techniques through targeted demonstra- tions on a hospital length-of-stay prediction task. Our findings reveal issues such as overemphasis on non-informative tokens, instability in at- tributions, and high-confidence predictions for incoherent input variants. These results underscore the need for explanation strategies that are clin- ically meaningful, semantically grounded, and robust to linguistic noise.

Guillermo Bernárdez, Lev Telyatnikov, Marco Montagna et al.

This paper describes the 2nd edition of the ICML Topological Deep Learning Challenge that was hosted within the ICML 2024 ELLIS Workshop on Geometry-grounded Representation Learning and Generative Modeling (GRaM). The challenge focused on the problem of representing data in different discrete topological domains in order to bridge the gap between Topological Deep Learning (TDL) and other types of structured datasets (e.g. point clouds, graphs). Specifically, participants were asked to design and implement topological liftings, i.e. mappings between different data structures and topological domains --like hypergraphs, or simplicial/cell/combinatorial complexes. The challenge received 52 submissions satisfying all the requirements. This paper introduces the main scope of the challenge, and summarizes the main results and findings.

Blaž Škrlj, Adi Schwartz, Jure Ferlež et al.

Hyperparameter optimization is the process of identifying the appropriate hyperparameter configuration of a given machine learning model with regard to a given learning task. For smaller data sets, an exhaustive search is possible; However, when the data size and model complexity increase, the number of configuration evaluations becomes the main computational bottleneck. A promising paradigm for tackling this type of problem is surrogate-based optimization. The main idea underlying this paradigm considers an incrementally updated model of the relation between the hyperparameter space and the output (target) space; the data for this model are obtained by evaluating the main learning engine, which is, for example, a factorization machine-based model. By learning to approximate the hyperparameter-target relation, the surrogate (machine learning) model can be used to score large amounts of hyperparameter configurations, exploring parts of the configuration space beyond the reach of direct machine learning engine evaluation. Commonly, a surrogate is selected prior to optimization initialization and remains the same during the search. We investigated whether dynamic switching of surrogates during the optimization itself is a sensible idea of practical relevance for selecting the most appropriate factorization machine-based models for large-scale online recommendation. We conducted benchmarks on data sets containing hundreds of millions of instances against established baselines such as Random Forest- and Gaussian process-based surrogates. The results indicate that surrogate switching can offer good performance while considering fewer learning engine evaluations.

Boshko Koloski, Nada Lavrač, Senja Pollak et al.

In the domain of semi-supervised learning, the current approaches insufficiently exploit the potential of considering inter-instance relationships among (un)labeled data. In this work, we address this limitation by providing an approach for inferring latent graphs that capture the intrinsic data relationships. By leveraging graph-based representations, our approach facilitates the seamless propagation of information throughout the graph, effectively incorporating global and local knowledge. Through evaluations on biomedical tabular datasets, we compare the capabilities of our approach to other contemporary methods. Our work demonstrates the significance of inter-instance relationship discovery as practical means for constructing robust latent graphs to enhance semi-supervised learning techniques. The experiments show that the proposed methodology outperforms contemporary state-of-the-art methods for (semi-)supervised learning on three biomedical datasets.

Sebastian Koralewski, Merwan Barlier, Yulia Stolin et al.

Online recommendation platforms increasingly use Large Language Models (LLMs) to extract structured features from ad creatives. While deploying a single-call LLM annotation agent yields significant Click-Through Rate (CTR) improvements in our live production environment, per-creative prompting is prohibitively expensive to scale. The redundant system instructions sent in every request account for 94% of billed input tokens. To break this cost bottleneck, we introduce PromptPack, a scalable, high-throughput LLM annotation agent. PromptPack achieves this scale via in-context batching, combining a shared system prompt, a strict XML structural envelope, and an output correction layer to ensure deterministic, pipeline-ready feature extraction across multiple creatives simultaneously. We evaluate PromptPack via an offline retrieval benchmark using a downstream logistic-regression ranker. To deeply profile the agent's behavior, we measure AUC and introduce Volume-Weighted Absolute Lift (VWAL), a novel metric capturing the signal quality of the generated features. Compared to our live, unbatched production baseline, PromptPack at batch size 20 cuts our LLM costs by 89% and accelerates throughput by 2.5x while fully preserving AUC.

Boshko Koloski, Nada Lavrač, Bojan Cestnik et al.

In an era marked by a rapid increase in scientific publications, researchers grapple with the challenge of keeping pace with field-specific advances. We present the `AHAM' methodology and a metric that guides the domain-specific \textbf{adapt}ation of the BERTopic topic modeling framework to improve scientific text analysis. By utilizing the LLaMa2 generative language model, we generate topic definitions via one-shot learning by crafting prompts with the \textbf{help} of domain experts to guide the LLM for literature mining by \textbf{asking} it to model the topic names. For inter-topic similarity evaluation, we leverage metrics from language generation and translation processes to assess lexical and semantic similarity of the generated topics. Our system aims to reduce both the ratio of outlier topics to the total number of topics and the similarity between topic definitions. The methodology has been assessed on a newly gathered corpus of scientific papers on literature-based discovery. Through rigorous evaluation by domain experts, AHAM has been validated as effective in uncovering intriguing and novel insights within broad research areas. We explore the impact of domain adaptation of sentence-transformers for the task of topic \textbf{model}ing using two datasets, each specialized to specific scientific domains within arXiv and medarxiv. We evaluate the impact of data size, the niche of adaptation, and the importance of domain adaptation. Our results suggest a strong interaction between domain adaptation and topic modeling precision in terms of outliers and topic definitions.

Boshko Koloski, Andrei Margeloiu, Xiangjian Jiang et al.

Tabular deep-learning methods require embedding numerical and categorical input features into high-dimensional spaces before processing them. Existing methods deal with this heterogeneous nature of tabular data by employing separate type-specific encoding approaches. This limits the cross-table transfer potential and the exploitation of pre-trained knowledge. We propose a novel approach that first transforms tabular data into text, and then leverages pre-trained representations from LLMs to encode this data, resulting in a plug-and-play solution to improv ing deep-learning tabular methods. We demonstrate that our approach improves accuracy over competitive models, such as MLP, ResNet and FT-Transformer, by validating on seven classification datasets.

Blaž Škrlj, Boshko Koloski, Senja Pollak et al.

Integrating structured knowledge from Knowledge Graphs (KGs) into Large Language Models (LLMs) enhances factual grounding and reasoning capabilities. This survey paper systematically examines the synergy between KGs and LLMs, categorizing existing approaches into two main groups: KG-enhanced LLMs, which improve reasoning, reduce hallucinations, and enable complex question answering; and LLM-augmented KGs, which facilitate KG construction, completion, and querying. Through comprehensive analysis, we identify critical gaps and highlight the mutual benefits of structured knowledge integration. Compared to existing surveys, our study uniquely emphasizes scalability, computational efficiency, and data quality. Finally, we propose future research directions, including neuro-symbolic integration, dynamic KG updating, data reliability, and ethical considerations, paving the way for intelligent systems capable of managing more complex real-world knowledge tasks.

Kristian Miok, Blaz Škrlj, Daniela Zaharie et al.

Clinical language models often struggle to provide trustworthy predictions and explanations when applied to lengthy, unstructured electronic health records (EHRs). This work introduces TT-XAI, a lightweight and effective framework that improves both classification performance and interpretability through domain-aware keyword distillation and reasoning with large language models (LLMs). First, we demonstrate that distilling raw discharge notes into concise keyword representations significantly enhances BERT classifier performance and improves local explanation fidelity via a focused variant of LIME. Second, we generate chain-of-thought clinical explanations using keyword-guided prompts to steer LLMs, producing more concise and clinically relevant reasoning. We evaluate explanation quality using deletion-based fidelity metrics, self-assessment via LLaMA-3 scoring, and a blinded human study with domain experts. All evaluation modalities consistently favor the keyword-augmented method, confirming that distillation enhances both machine and human interpretability. TT-XAI offers a scalable pathway toward trustworthy, auditable AI in clinical decision support.

Blaž Škrlj, Benoît Guilleminot, Andraž Tori

Large language models (LLMs) and their associated agent-based frameworks have significantly advanced automated information extraction, a critical component of modern recommender systems. While these multitask frameworks are widely used in code generation, their application in data-centric research is still largely untapped. This paper presents Agent0, an LLM-driven, agent-based system designed to automate information extraction and feature construction from raw, unstructured text. Categorical features are crucial for large-scale recommender systems but are often expensive to acquire. Agent0 coordinates a group of interacting LLM agents to automatically identify the most valuable text aspects for subsequent tasks (such as models or AutoML pipelines). Beyond its feature engineering capabilities, Agent0 also offers an automated prompt-engineering tuning method that utilizes dynamic feedback loops from an oracle. Our findings demonstrate that this closed-loop methodology is both practical and effective for automated feature discovery, which is recognized as one of the most challenging phases in current recommender system development.

Boshko Koloski, Senja Pollak, Roberto Navigli et al.

Building on the success of Large Language Models (LLMs), LLM-based representations have dominated the document representation landscape, achieving great performance on the document embedding benchmarks. However, the high-dimensional, computationally expensive embeddings from LLMs tend to be either too generic or inefficient for domain-specific applications. To address these limitations, we introduce FuDoBa a Bayesian optimisation-based method that integrates LLM-based embeddings with domain-specific structured knowledge, sourced both locally and from external repositories like WikiData. This fusion produces low-dimensional, task-relevant representations while reducing training complexity and yielding interpretable early-fusion weights for enhanced classification performance. We demonstrate the effectiveness of our approach on six datasets in two domains, showing that when paired with robust AutoML-based classifiers, our proposed representation learning approach performs on par with, or surpasses, those produced solely by the proprietary LLM-based embedding baselines.

Blaž Škrlj, Yonatan Karni, Grega Gašperšič et al.

The Deep and Cross architecture (DCNv2) is a robust production baseline and is integral to numerous real-life recommender systems. Its inherent efficiency and ability to model interactions often result in models that are both simpler and highly competitive compared to more computationally demanding alternatives, such as Deep FFMs. In this work, we introduce three significant algorithmic improvements to the DCNv2 architecture, detailing their formulation and behavior at scale. The enhanced architecture we refer to as DCN^2 is actively used in a live recommender system, processing over 0.5 billion predictions per second across diverse use cases where it out-performed DCNv2, both offline and online (ab tests). These improvements effectively address key limitations observed in the DCNv2, including information loss in Cross layers, implicit management of collisions through learnable lookup-level weights, and explicit modeling of pairwise similarities with a custom layer that emulates FFMs' behavior. The superior performance of DCN^2 is also demonstrated on four publicly available benchmark data sets.

Boshko Koloski, Nada Lavrač, Blaž Škrlj

Construction of neural network architectures suitable for learning from both continuous and discrete tabular data is a challenging research endeavor. Contemporary high-dimensional tabular data sets are often characterized by a relatively small instance count, requiring data-efficient learning. We propose HorNets (Horn Networks), a neural network architecture with state-of-the-art performance on synthetic and real-life data sets from scarce-data tabular domains. HorNets are based on a clipped polynomial-like activation function, extended by a custom discrete-continuous routing mechanism that decides which part of the neural network to optimize based on the input's cardinality. By explicitly modeling parts of the feature combination space or combining whole space in a linear attention-like manner, HorNets dynamically decide which mode of operation is the most suitable for a given piece of data with no explicit supervision. This architecture is one of the few approaches that reliably retrieves logical clauses (including noisy XNOR) and achieves state-of-the-art classification performance on 14 real-life biomedical high-dimensional data sets. HorNets are made freely available under a permissive license alongside a synthetic generator of categorical benchmarks.

Miha Malenšek, Blaž Škrlj, Blaž Mramor et al.

Synthetic datasets are important for evaluating and testing machine learning models. When evaluating real-life recommender systems, high-dimensional categorical (and sparse) datasets are often considered. Unfortunately, there are not many solutions that would allow generation of artificial datasets with such characteristics. For that purpose, we developed a novel framework for generating synthetic datasets that are diverse and statistically coherent. Our framework allows for creation of datasets with controlled attributes, enabling iterative modifications to fit specific experimental needs, such as introducing complex feature interactions, feature cardinality, or specific distributions. We demonstrate the framework's utility through use cases such as benchmarking probabilistic counting algorithms, detecting algorithmic bias, and simulating AutoML searches. Unlike existing methods that either focus narrowly on specific dataset structures, or prioritize (private) data synthesis through real data, our approach provides a modular means to quickly generating completely synthetic datasets we can tailor to diverse experimental requirements. Our results show that the framework effectively isolates model behavior in unique situations and highlights its potential for significant advancements in the evaluation and development of recommender systems. The readily-available framework is available as a free open Python package to facilitate research with minimal friction.

Blaž Škrlj, Nir Ki-Tov, Lee Edelist et al.

Real-world production systems often grapple with maintaining data quality in large-scale, dynamic streams. We introduce Drifter, an efficient and lightweight system for online feature monitoring and verification in recommendation use cases. Drifter addresses limitations of existing methods by delivering agile, responsive, and adaptable data quality monitoring, enabling real-time root cause analysis, drift detection and insights into problematic production events. Integrating state-of-the-art online feature ranking for sparse data and anomaly detection ideas, Drifter is highly scalable and resource-efficient, requiring only two threads and less than a gigabyte of RAM per production deployments that handle millions of instances per minute. Evaluation on real-world data sets demonstrates Drifter's effectiveness in alerting and mitigating data quality issues, substantially improving reliability and performance of real-time live recommender systems.

Blaž Škrlj, Blaž Mramor

The design of modern recommender systems relies on understanding which parts of the feature space are relevant for solving a given recommendation task. However, real-world data sets in this domain are often characterized by their large size, sparsity, and noise, making it challenging to identify meaningful signals. Feature ranking represents an efficient branch of algorithms that can help address these challenges by identifying the most informative features and facilitating the automated search for more compact and better-performing models (AutoML). We introduce OutRank, a system for versatile feature ranking and data quality-related anomaly detection. OutRank was built with categorical data in mind, utilizing a variant of mutual information that is normalized with regard to the noise produced by features of the same cardinality. We further extend the similarity measure by incorporating information on feature similarity and combined relevance. The proposed approach's feasibility is demonstrated by speeding up the state-of-the-art AutoML system on a synthetic data set with no performance loss. Furthermore, we considered a real-life click-through-rate prediction data set where it outperformed strong baselines such as random forest-based approaches. The proposed approach enables exploration of up to 300% larger feature spaces compared to AutoML-only approaches, enabling faster search for better models on off-the-shelf hardware.

Boshko Koloski, Senja Pollak, Blaž Škrlj et al.

Keyword extraction is the task of retrieving words that are essential to the content of a given document. Researchers proposed various approaches to tackle this problem. At the top-most level, approaches are divided into ones that require training - supervised and ones that do not - unsupervised. In this study, we are interested in settings, where for a language under investigation, no training data is available. More specifically, we explore whether pretrained multilingual language models can be employed for zero-shot cross-lingual keyword extraction on low-resource languages with limited or no available labeled training data and whether they outperform state-of-the-art unsupervised keyword extractors. The comparison is conducted on six news article datasets covering two high-resource languages, English and Russian, and four low-resource languages, Croatian, Estonian, Latvian, and Slovenian. We find that the pretrained models fine-tuned on a multilingual corpus covering languages that do not appear in the test set (i.e. in a zero-shot setting), consistently outscore unsupervised models in all six languages.

Urh Primožič, Blaž Škrlj, Sašo Džeroski et al.

The need for learning from unlabeled data is increasing in contemporary machine learning. Methods for unsupervised feature ranking, which identify the most important features in such data are thus gaining attention, and so are their applications in studying high throughput biological experiments or user bases for recommender systems. We propose FRANe (Feature Ranking via Attribute Networks), an unsupervised algorithm capable of finding key features in given unlabeled data set. FRANe is based on ideas from network reconstruction and network analysis. FRANe performs better than state-of-the-art competitors, as we empirically demonstrate on a large collection of benchmarks. Moreover, we provide the time complexity analysis of FRANe further demonstrating its scalability. Finally, FRANe offers as the result the interpretable relational structures used to derive the feature importances.

Sebastian Mežnar, Matej Bevec, Nada Lavrač et al.

The increasing amounts of semantic resources offer valuable storage of human knowledge; however, the probability of wrong entries increases with the increased size. The development of approaches that identify potentially spurious parts of a given knowledge base is thus becoming an increasingly important area of interest. In this work, we present a systematic evaluation of whether structure-only link analysis methods can already offer a scalable means to detecting possible anomalies, as well as potentially interesting novel relation candidates. Evaluating thirteen methods on eight different semantic resources, including Gene Ontology, Food Ontology, Marine Ontology and similar, we demonstrated that structure-only link analysis could offer scalable anomaly detection for a subset of the data sets. Further, we demonstrated that by considering symbolic node embedding, explanations of the predictions (links) could be obtained, making this branch of methods potentially more valuable than the black-box only ones. To our knowledge, this is currently one of the most extensive systematic studies of the applicability of different types of link analysis methods across semantic resources from different domains.

Boshko Koloski, Timen Stepišnik-Perdih, Marko Robnik-Šikonja et al.

Increasing amounts of freely available data both in textual and relational form offers exploration of richer document representations, potentially improving the model performance and robustness. An emerging problem in the modern era is fake news detection -- many easily available pieces of information are not necessarily factually correct, and can lead to wrong conclusions or are used for manipulation. In this work we explore how different document representations, ranging from simple symbolic bag-of-words, to contextual, neural language model-based ones can be used for efficient fake news identification. One of the key contributions is a set of novel document representation learning methods based solely on knowledge graphs, i.e. extensive collections of (grounded) subject-predicate-object triplets. We demonstrate that knowledge graph-based representations already achieve competitive performance to conventionally accepted representation learners. Furthermore, when combined with existing, contextual representations, knowledge graph-based document representations can achieve state-of-the-art performance. To our knowledge this is the first larger-scale evaluation of how knowledge graph-based representations can be systematically incorporated into the process of fake news classification.

Blaž Škrlj, Marko Jukič, Nika Eržen et al.

The COVID-19 pandemic triggered a wave of novel scientific literature that is impossible to inspect and study in a reasonable time frame manually. Current machine learning methods offer to project such body of literature into the vector space, where similar documents are located close to each other, offering an insightful exploration of scientific papers and other knowledge sources associated with COVID-19. However, to start searching, such texts need to be appropriately annotated, which is seldom the case due to the lack of human resources. In our system, the current body of COVID-19-related literature is annotated using unsupervised keyphrase extraction, facilitating the initial queries to the latent space containing the learned document embeddings (low-dimensional representations). The solution is accessible through a web server capable of interactive search, term ranking, and exploration of potentially interesting literature. We demonstrate the usefulness of the approach via case studies from the medicinal chemistry domain.

Blaž Škrlj, Matej Petkovič

Contemporary natural language processing (NLP) revolves around learning from latent document representations, generated either implicitly by neural language models or explicitly by methods such as doc2vec or similar. One of the key properties of the obtained representations is their dimension. Whilst the commonly adopted dimensions of 256 and 768 offer sufficient performance on many tasks, it is many times unclear whether the default dimension is the most suitable choice for the subsequent downstream learning tasks. Furthermore, representation dimensions are seldom subject to hyperparameter tuning due to computational constraints. The purpose of this paper is to demonstrate that a surprisingly simple and efficient recursive compression procedure can be sufficient to both significantly compress the initial representation, but also potentially improve its performance when considering the task of text classification. Having smaller and less noisy representations is the desired property during deployment, as orders of magnitude smaller models can significantly reduce the computational overload and with it the deployment costs. We propose CoRe, a straightforward, representation learner-agnostic framework suitable for representation compression. The CoRe's performance is showcased and studied on a collection of 17 real-life corpora from biomedical, news, social media, and literary domains. We explored CoRe's behavior when considering contextual and non-contextual document representations, different compression levels, and 9 different compression algorithms. Current results based on more than 100,000 compression experiments indicate that recursive Singular Value Decomposition offers a very good trade-off between the compression efficiency and performance, making CoRe useful in many existing, representation-dependent NLP pipelines.

Timen Stepišnik Perdih, Senja Pollak, Blaž \v{Skrlj}

Ontologies are increasingly used for machine reasoning over the last few years. They can provide explanations of concepts or be used for concept classification if there exists a mapping from the desired labels to the relevant ontology. Another advantage of using ontologies is that they do not need a learning process, meaning that we do not need the train data or time before using them. This paper presents a practical use of an ontology for a classification problem from the financial domain. It first transforms a given ontology to a graph and proceeds with generalization with the aim to find common semantic descriptions of the input sets of financial concepts. We present a solution to the shared task on Learning Semantic Similarities for the Financial Domain (FinSim-2 task). The task is to design a system that can automatically classify concepts from the Financial domain into the most relevant hypernym concept in an external ontology - the Financial Industry Business Ontology. We propose a method that maps given concepts to the mentioned ontology and performs a graph search for the most relevant hypernyms. We also employ a word vectorization method and a machine learning classifier to supplement the method with a ranked list of labels for each concept.

Sebastian Mežnar, Nada Lavrač, Blaž Škrlj

Understanding how information propagates in real-life complex networks yields a better understanding of dynamic processes such as misinformation or epidemic spreading. The recently introduced branch of machine learning methods for learning node representations offers many novel applications, one of them being the task of spreading prediction addressed in this paper. We explore the utility of the state-of-the-art node representation learners when used to assess the effects of spreading from a given node, estimated via extensive simulations. Further, as many real-life networks are topologically similar, we systematically investigate whether the learned models generalize to previously unseen networks, showing that in some cases very good model transfer can be obtained. This work is one of the first to explore transferability of the learned representations for the task of node regression; we show there exist pairs of networks with similar structure between which the trained models can be transferred (zero-shot), and demonstrate their competitive performance. To our knowledge, this is one of the first attempts to evaluate the utility of zero-shot transfer for the task of node regression.

Boshko Koloski, Senja Pollak, Blaž Škrlj et al.

Keyword extraction is the task of identifying words (or multi-word expressions) that best describe a given document and serve in news portals to link articles of similar topics. In this work we develop and evaluate our methods on four novel data sets covering less represented, morphologically-rich languages in European news media industry (Croatian, Estonian, Latvian and Russian). First, we perform evaluation of two supervised neural transformer-based methods (TNT-KID and BERT+BiLSTM CRF) and compare them to a baseline TF-IDF based unsupervised approach. Next, we show that by combining the keywords retrieved by both neural transformer based methods and extending the final set of keywords with an unsupervised TF-IDF based technique, we can drastically improve the recall of the system, making it appropriate to be used as a recommendation system in the media house environment.

Blaž Škrlj, Sašo Džeroski, Nada Lavrač et al.

Feature ranking has been widely adopted in machine learning applications such as high-throughput biology and social sciences. The approaches of the popular Relief family of algorithms assign importances to features by iteratively accounting for nearest relevant and irrelevant instances. Despite their high utility, these algorithms can be computationally expensive and not-well suited for high-dimensional sparse input spaces. In contrast, recent embedding-based methods learn compact, low-dimensional representations, potentially facilitating down-stream learning capabilities of conventional learners. This paper explores how the Relief branch of algorithms can be adapted to benefit from (Riemannian) manifold-based embeddings of instance and target spaces, where a given embedding's dimensionality is intrinsic to the dimensionality of the considered data set. The developed ReliefE algorithm is faster and can result in better feature rankings, as shown by our evaluation on 20 real-life data sets for multi-class and multi-label classification tasks. The utility of ReliefE for high-dimensional data sets is ensured by its implementation that utilizes sparse matrix algebraic operations. Finally, the relation of ReliefE to other ranking algorithms is studied via the Fuzzy Jaccard Index.

Sebastian Mežnar, Blaž Škrlj

The competition "Predicting Generalization in Deep Learning (PGDL)" aims to provide a platform for rigorous study of generalization of deep learning models and offer insight into the progress of understanding and explaining these models. This report presents the solution that was submitted by the user \emph{smeznar} which achieved the eight place in the competition. In the proposed approach, we create simple metrics and find their best combination with automatic testing on the provided dataset, exploring how combinations of various properties of the input neural network architectures can be used for the prediction of their generalization.

Matej Petković, Dragi Kocev, Blaž Škrlj et al.

In this work, we propose two novel (groups of) methods for unsupervised feature ranking and selection. The first group includes feature ranking scores (Genie3 score, RandomForest score) that are computed from ensembles of predictive clustering trees. The second method is URelief, the unsupervised extension of the Relief family of feature ranking algorithms. Using 26 benchmark data sets and 5 baselines, we show that both the Genie3 score (computed from the ensemble of extra trees) and the URelief method outperform the existing methods and that Genie3 performs best overall, in terms of predictive power of the top-ranked features. Additionally, we analyze the influence of the hyper-parameters of the proposed methods on their performance, and show that for the Genie3 score the highest quality is achieved by the most efficient parameter configuration. Finally, we propose a way of discovering the location of the features in the ranking, which are the most relevant in reality.

Sebastian Mežnar, Nada Lavrač, Blaž Škrlj

Learning from complex real-life networks is a lively research area, with recent advances in learning information-rich, low-dimensional network node representations. However, state-of-the-art methods are not necessarily interpretable and are therefore not fully applicable to sensitive settings in biomedical or user profiling tasks, where explicit bias detection is highly relevant. The proposed SNoRe (Symbolic Node Representations) algorithm is capable of learning symbolic, human-understandable representations of individual network nodes, based on the similarity of neighborhood hashes which serve as features. SNoRe's interpretable features are suitable for direct explanation of individual predictions, which we demonstrate by coupling it with the widely used instance explanation tool SHAP to obtain nomograms representing the relevance of individual features for a given classification. To our knowledge, this is one of the first such attempts in a structural node embedding setting. In the experimental evaluation on eleven real-life datasets, SNoRe proved to be competitive to strong baselines, such as variational graph autoencoders, node2vec and LINE. The vectorized implementation of SNoRe scales to large networks, making it suitable for contemporary network learning and analysis tasks.

Matej Petković, Blaž Škrlj, Dragi Kocev et al.

We propose Fuzzy Jaccard Index (FUJI) -- a scale-invariant score for assessment of the similarity between two ranked/ordered lists. FUJI improves upon the Jaccard index by incorporating a membership function which takes into account the particular ranks, thus producing both more stable and more accurate similarity estimates. We provide theoretical insights into the properties of the FUJI score as well as propose an efficient algorithm for computing it. We also present empirical evidence of its performance on different synthetic scenarios. Finally, we demonstrate its utility in a typical machine learning setting -- comparing feature ranking lists relevant to a given machine learning task. In real-life, and in particular high-dimensional domains, where only a small percentage of the whole feature space might be relevant, a robust and confident feature ranking leads to interpretable findings as well as efficient computation and good predictive performance. In such cases, FUJI correctly distinguishes between existing feature ranking approaches, while being more robust and efficient than the benchmark similarity scores.

Kristian Miok, Blaz Skrlj, Daniela Zaharie et al.

Hate speech is an important problem in the management of user-generated content. To remove offensive content or ban misbehaving users, content moderators need reliable hate speech detectors. Recently, deep neural networks based on the transformer architecture, such as the (multilingual) BERT model, achieve superior performance in many natural language classification tasks, including hate speech detection. So far, these methods have not been able to quantify their output in terms of reliability. We propose a Bayesian method using Monte Carlo dropout within the attention layers of the transformer models to provide well-calibrated reliability estimates. We evaluate and visualize the results of the proposed approach on hate speech detection problems in several languages. Additionally, we test if affective dimensions can enhance the information extracted by the BERT model in hate speech classification. Our experiments show that Monte Carlo dropout provides a viable mechanism for reliability estimation in transformer networks. Used within the BERT model, it ofers state-of-the-art classification performance and can detect less trusted predictions. Also, it was observed that affective dimensions extracted using sentic computing methods can provide insights toward interpretation of emotions involved in hate speech. Our approach not only improves the classification performance of the state-of-the-art multilingual BERT model but the computed reliability scores also significantly reduce the workload in an inspection of ofending cases and reannotation campaigns. The provided visualization helps to understand the borderline outcomes.

Blaž Škrlj, Nika Eržen, Shane Sheehan et al.

Neural language models are becoming the prevailing methodology for the tasks of query answering, text classification, disambiguation, completion and translation. Commonly comprised of hundreds of millions of parameters, these neural network models offer state-of-the-art performance at the cost of interpretability; humans are no longer capable of tracing and understanding how decisions are being made. The attention mechanism, introduced initially for the task of translation, has been successfully adopted for other language-related tasks. We propose AttViz, an online toolkit for exploration of self-attention---real values associated with individual text tokens. We show how existing deep learning pipelines can produce outputs suitable for AttViz, offering novel visualizations of the attention heads and their aggregations with minimal effort, online. We show on examples of news segments how the proposed system can be used to inspect and potentially better understand what a model has learned (or emphasized).

Matej Martinc, Blaž Škrlj, Senja Pollak

With growing amounts of available textual data, development of algorithms capable of automatic analysis, categorization and summarization of these data has become a necessity. In this research we present a novel algorithm for keyword identification, i.e., an extraction of one or multi-word phrases representing key aspects of a given document, called Transformer-based Neural Tagger for Keyword IDentification (TNT-KID). By adapting the transformer architecture for a specific task at hand and leveraging language model pretraining on a domain specific corpus, the model is capable of overcoming deficiencies of both supervised and unsupervised state-of-the-art approaches to keyword extraction by offering competitive and robust performance on a variety of different datasets while requiring only a fraction of manually labeled data required by the best performing systems. This study also offers thorough error analysis with valuable insights into the inner workings of the model and an ablation study measuring the influence of specific components of the keyword identification workflow on the overall performance.

Blaž Škrlj, Sašo Džeroski, Nada Lavrač et al.

Black-box neural network models are widely used in industry and science, yet are hard to understand and interpret. Recently, the attention mechanism was introduced, offering insights into the inner workings of neural language models. This paper explores the use of attention-based neural networks mechanism for estimating feature importance, as means for explaining the models learned from propositional (tabular) data. Feature importance estimates, assessed by the proposed Self-Attention Network (SAN) architecture, are compared with the established ReliefF, Mutual Information and Random Forest-based estimates, which are widely used in practice for model interpretation. For the first time we conduct scale-free comparisons of feature importance estimates across algorithms on ten real and synthetic data sets to study the similarities and differences of the resulting feature importance estimates, showing that SANs identify similar high-ranked features as the other methods. We demonstrate that SANs identify feature interactions which in some cases yield better predictive performance than the baselines, suggesting that attention extends beyond interactions of just a few key features and detects larger feature subsets relevant for the considered learning task.

Blaž Škrlj, Andraž Repar, Senja Pollak

Keyword extraction is used for summarizing the content of a document and supports efficient document retrieval, and is as such an indispensable part of modern text-based systems. We explore how load centrality, a graph-theoretic measure applied to graphs derived from a given text can be used to efficiently identify and rank keywords. Introducing meta vertices (aggregates of existing vertices) and systematic redundancy filters, the proposed method performs on par with state-of-the-art for the keyword extraction task on 14 diverse datasets. The proposed method is unsupervised, interpretable and can also be used for document visualization.

Blaž Škrlj, Matej Martinc, Jan Kralj et al.

The use of background knowledge is largely unexploited in text classification tasks. This paper explores word taxonomies as means for constructing new semantic features, which may improve the performance and robustness of the learned classifiers. We propose tax2vec, a parallel algorithm for constructing taxonomy-based features, and demonstrate its use on six short text classification problems: prediction of gender, personality type, age, news topics, drug side effects and drug effectiveness. The constructed semantic features, in combination with fast linear classifiers, tested against strong baselines such as hierarchical attention neural networks, achieves comparable classification results on short text documents. The algorithm's performance is also tested in a few-shot learning setting, indicating that the inclusion of semantic features can improve the performance in data-scarce situations. The tax2vec capability to extract corpus-specific semantic keywords is also demonstrated. Finally, we investigate the semantic space of potential features, where we observe a similarity with the well known Zipf's law.