Joeran Beel

Lennart Purucker, Lennart Schneider, Marie Anastacio et al.

Automated machine learning (AutoML) systems commonly ensemble models post hoc to improve predictive performance, typically via greedy ensemble selection (GES). However, we believe that GES may not always be optimal, as it performs a simple deterministic greedy search. In this work, we introduce two novel population-based ensemble selection methods, QO-ES and QDO-ES, and compare them to GES. While QO-ES optimises solely for predictive performance, QDO-ES also considers the diversity of ensembles within the population, maintaining a diverse set of well-performing ensembles during optimisation based on ideas of quality diversity optimisation. The methods are evaluated using 71 classification datasets from the AutoML benchmark, demonstrating that QO-ES and QDO-ES often outrank GES, albeit only statistically significant on validation data. Our results further suggest that diversity can be beneficial for post hoc ensembling but also increases the risk of overfitting.

Lennart Purucker, Joeran Beel

Many state-of-the-art automated machine learning (AutoML) systems use greedy ensemble selection (GES) by Caruana et al. (2004) to ensemble models found during model selection post hoc. Thereby, boosting predictive performance and likely following Auto-Sklearn 1's insight that alternatives, like stacking or gradient-free numerical optimization, overfit. Overfitting in Auto-Sklearn 1 is much more likely than in other AutoML systems because it uses only low-quality validation data for post hoc ensembling. Therefore, we were motivated to analyze whether Auto-Sklearn 1's insight holds true for systems with higher-quality validation data. Consequently, we compared the performance of covariance matrix adaptation evolution strategy (CMA-ES), state-of-the-art gradient-free numerical optimization, to GES on the 71 classification datasets from the AutoML benchmark for AutoGluon. We found that Auto-Sklearn's insight depends on the chosen metric. For the metric ROC AUC, CMA-ES overfits drastically and is outperformed by GES -- statistically significantly for multi-class classification. For the metric balanced accuracy, CMA-ES does not overfit and outperforms GES significantly. Motivated by the successful application of CMA-ES for balanced accuracy, we explored methods to stop CMA-ES from overfitting for ROC AUC. We propose a method to normalize the weights produced by CMA-ES, inspired by GES, that avoids overfitting for CMA-ES and makes CMA-ES perform better than or similar to GES for ROC AUC.

Lennart Purucker, Joeran Beel

Automated Machine Learning (AutoML) frameworks regularly use ensembles. Developers need to compare different ensemble techniques to select appropriate techniques for an AutoML framework from the many potential techniques. So far, the comparison of ensemble techniques is often computationally expensive, because many base models must be trained and evaluated one or multiple times. Therefore, we present Assembled-OpenML. Assembled-OpenML is a Python tool, which builds meta-datasets for ensembles using OpenML. A meta-dataset, called Metatask, consists of the data of an OpenML task, the task's dataset, and prediction data from model evaluations for the task. We can make the comparison of ensemble techniques computationally cheaper by using the predictions stored in a metatask instead of training and evaluating base models. To introduce Assembled-OpenML, we describe the first version of our tool. Moreover, we present an example of using Assembled-OpenML to compare a set of ensemble techniques. For this example comparison, we built a benchmark using Assembled-OpenML and implemented ensemble techniques expecting predictions instead of base models as input. In our example comparison, we gathered the prediction data of $1523$ base models for $31$ datasets. Obtaining the prediction data for all base models using Assembled-OpenML took ${\sim} 1$ hour in total. In comparison, obtaining the prediction data by training and evaluating just one base model on the most computationally expensive dataset took ${\sim} 37$ minutes.

Jannik Nitschke, Lukas Wegmeth, Joeran Beel

Ensemble methods are frequently used in recommender systems to improve accuracy by combining multiple models. Recent work reports sizable performance gains, but most studies still optimize primarily for accuracy and robustness rather than for energy efficiency. This paper measures accuracy energy trade offs of ensemble techniques relative to strong single models. We run 93 controlled experiments in two pipelines: 1. explicit rating prediction with Surprise (RMSE) and 2. implicit feedback ranking with LensKit (NDCG@10). We evaluate four datasets ranging from 100,000 to 7.8 million interactions (MovieLens 100K, MovieLens 1M, ModCloth, Anime). We compare four ensemble strategies (Average, Weighted, Stacking or Rank Fusion, Top Performers) against baselines and optimized single models. Whole system energy is measured with EMERS using a smart plug and converted to CO2 equivalents. Across settings, ensembles improve accuracy by 0.3% to 5.7% while increasing energy by 19% to 2,549%. On MovieLens 1M, a Top Performers ensemble improves RMSE by 0.96% at an 18.8% energy overhead over SVD++. On MovieLens 100K, an averaging ensemble improves NDCG@10 by 5.7% with 103% additional energy. On Anime, a Surprise Top Performers ensemble improves RMSE by 1.2% but consumes 2,005% more energy (0.21 vs. 0.01 Wh), increasing emissions from 2.6 to 53.8 mg CO2 equivalents, and LensKit ensembles fail due to memory limits. Overall, selective ensembles are more energy efficient than exhaustive averaging,

Dominika Tkaczyk, Rohit Gupta, Riccardo Cinti et al.

Bibliographic reference parsers extract machine-readable metadata such as author names, title, journal, and year from bibliographic reference strings. To extract the metadata, the parsers apply heuristics or machine learning. However, no reference parser, and no algorithm, consistently gives the best results in every scenario. For instance, one tool may be best in extracting titles in ACM citation style, but only third best when APA is used. Another tool may be best in extracting English author names, while another one is best for noisy data (i.e. inconsistent citation styles). In this paper, which is an extended version of our recent RecSys poster, we address the problem of reference parsing from a recommender-systems and meta-learning perspective. We propose ParsRec, a meta-learning based recommender-system that recommends the potentially most effective parser for a given reference string. ParsRec recommends one out of 10 open-source parsers: Anystyle-Parser, Biblio, CERMINE, Citation, Citation-Parser, GROBID, ParsCit, PDFSSA4MET, Reference Tagger, and Science Parse. We evaluate ParsRec on 105k references from chemistry. We propose two approaches to meta-learning recommendations. The first approach learns the best parser for an entire reference string. The second approach learns the best parser for each metadata type in a reference string. The second approach achieved a 2.6% increase in F1 (0.909 vs. 0.886) over the best single parser (GROBID), reducing the false positive rate by 20.2% (0.075 vs. 0.094), and the false negative rate by 18.9% (0.107 vs. 0.132).

Joeran Beel, Andrew Collins, Akiko Aizawa

Recommender systems in academia are not widely available. This may be in part due to the difficulty and cost of developing and maintaining recommender systems. Many operators of academic products such as digital libraries and reference managers avoid this effort, although a recommender system could provide significant benefits to their users. In this paper, we introduce Mr. DLib's "Recommendations as-a-Service" (RaaS) API that allows operators of academic products to easily integrate a scientific recommender system into their products. Mr. DLib generates recommendations for research articles but in the future, recommendations may include call for papers, grants, etc. Operators of academic products can request recommendations from Mr. DLib and display these recommendations to their users. Mr. DLib can be integrated in just a few hours or days; creating an equivalent recommender system from scratch would require several months for an academic operator. Mr. DLib has been used by GESIS Sowiport and by the reference manager JabRef. Mr. DLib is open source and its goal is to facilitate the application of, and research on, scientific recommender systems. In this paper, we present the motivation for Mr. DLib, the architecture and details about the effectiveness. Mr. DLib has delivered 94m recommendations over a span of two years with an average click-through rate of 0.12%.

Mark Collier, Joeran Beel

Neural Turing Machines (NTMs) are an instance of Memory Augmented Neural Networks, a new class of recurrent neural networks which decouple computation from memory by introducing an external memory unit. NTMs have demonstrated superior performance over Long Short-Term Memory Cells in several sequence learning tasks. A number of open source implementations of NTMs exist but are unstable during training and/or fail to replicate the reported performance of NTMs. This paper presents the details of our successful implementation of a NTM. Our implementation learns to solve three sequential learning tasks from the original NTM paper. We find that the choice of memory contents initialization scheme is crucial in successfully implementing a NTM. Networks with memory contents initialized to small constant values converge on average 2 times faster than the next best memory contents initialization scheme.

Joeran Beel, Min-Yen Kan, Moritz Baumgart

A major step toward Artificial General Intelligence (AGI) and Super Intelligence is AI's ability to autonomously conduct research - what we term Artificial Research Intelligence (ARI). If machines could generate hypotheses, conduct experiments, and write research papers without human intervention, it would transform science. Sakana recently introduced the 'AI Scientist', claiming to conduct research autonomously, i.e. they imply to have achieved what we term Artificial Research Intelligence (ARI). The AI Scientist gained much attention, but a thorough independent evaluation has yet to be conducted. Our evaluation of the AI Scientist reveals critical shortcomings. The system's literature reviews produced poor novelty assessments, often misclassifying established concepts (e.g., micro-batching for stochastic gradient descent) as novel. It also struggles with experiment execution: 42% of experiments failed due to coding errors, while others produced flawed or misleading results. Code modifications were minimal, averaging 8% more characters per iteration, suggesting limited adaptability. Generated manuscripts were poorly substantiated, with a median of five citations, most outdated (only five of 34 from 2020 or later). Structural errors were frequent, including missing figures, repeated sections, and placeholder text like 'Conclusions Here'. Some papers contained hallucinated numerical results. Despite these flaws, the AI Scientist represents a leap forward in research automation. It generates full research manuscripts with minimal human input, challenging expectations of AI-driven science. Many reviewers might struggle to distinguish its work from human researchers. While its quality resembles a rushed undergraduate paper, its speed and cost efficiency are unprecedented, producing a full paper for USD 6 to 15 with 3.5 hours of human involvement, far outpacing traditional researchers.

Ardalan Arabzadeh, Tobias Vente, Joeran Beel

As recommender systems become increasingly prevalent, the environmental impact and energy efficiency of training large-scale models have come under scrutiny. This paper investigates the potential for energy-efficient algorithm performance by optimizing dataset sizes through downsampling techniques in the context of Green Recommender Systems. We conducted experiments on the MovieLens 100K, 1M, 10M, and Amazon Toys and Games datasets, analyzing the performance of various recommender algorithms under different portions of dataset size. Our results indicate that while more training data generally leads to higher algorithm performance, certain algorithms, such as FunkSVD and BiasedMF, particularly with unbalanced and sparse datasets like Amazon Toys and Games, maintain high-quality recommendations with up to a 50% reduction in training data, achieving nDCG@10 scores within approximately 13% of full dataset performance. These findings suggest that strategic dataset reduction can decrease computational and environmental costs without substantially compromising recommendation quality. This study advances sustainable and green recommender systems by providing insights for reducing energy consumption while maintaining effectiveness.

Moritz Baumgart, Lukas Wegmeth, Tobias Vente et al.

To combat the rising energy consumption of recommender systems we implement a novel alternative for k-fold cross validation. This alternative, named e-fold cross validation, aims to minimize the number of folds to achieve a reduction in power usage while keeping the reliability and robustness of the test results high. We tested our method on 5 recommender system algorithms across 6 datasets and compared it with 10-fold cross validation. On average e-fold cross validation only needed 41.5% of the energy that 10-fold cross validation would need, while it's results only differed by 1.81%. We conclude that e-fold cross validation is a promising approach that has the potential to be an energy efficient but still reliable alternative to k-fold cross validation.

Tobias Vente, Joeran Beel

Automated Machine Learning (AutoML) has greatly advanced applications of Machine Learning (ML) including model compression, machine translation, and computer vision. Recommender Systems (RecSys) can be seen as an application of ML. Yet, AutoML has found little attention in the RecSys community; nor has RecSys found notable attention in the AutoML community. Only few and relatively simple Automated Recommender Systems (AutoRecSys) libraries exist that adopt AutoML techniques. However, these libraries are based on student projects and do not offer the features and thorough development of AutoML libraries. We set out to determine how AutoML libraries perform in the scenario of an inexperienced user who wants to implement a recommender system. We compared the predictive performance of 60 AutoML, AutoRecSys, ML, and RecSys algorithms from 15 libraries, including a mean predictor baseline, on 14 explicit feedback RecSys datasets. To simulate the perspective of an inexperienced user, the algorithms were evaluated with default hyperparameters. We found that AutoML and AutoRecSys libraries performed best. AutoML libraries performed best for six of the 14 datasets (43%), but it was not always the same AutoML library performing best. The single-best library was the AutoRecSys library Auto-Surprise, which performed best on five datasets (36%). On three datasets (21%), AutoML libraries performed poorly, and RecSys libraries with default parameters performed best. Although, while obtaining 50% of all placements in the top five per dataset, RecSys algorithms fall behind AutoML on average. ML algorithms generally performed the worst.

Christopher Mahlich, Tobias Vente, Joeran Beel

This paper introduces e-fold cross-validation, an energy-efficient alternative to k-fold cross-validation. It dynamically adjusts the number of folds based on a stopping criterion. The criterion checks after each fold whether the standard deviation of the evaluated folds has consistently decreased or remained stable. Once met, the process stops early. We tested e-fold cross-validation on 15 datasets and 10 machine-learning algorithms. On average, it required 4 fewer folds than 10-fold cross-validation, reducing evaluation time, computational resources, and energy use by about 40%. Performance differences between e-fold and 10-fold cross-validation were less than 2% for larger datasets. More complex models showed even smaller discrepancies. In 96% of iterations, the results were within the confidence interval, confirming statistical significance. E-fold cross-validation offers a reliable and efficient alternative to k-fold, reducing computational costs while maintaining comparable accuracy.

Joeran Beel, Bela Gipp, Tobias Vente et al.

Recommender-systems research has accelerated model and evaluation advances, yet largely neglects automating the research process itself. We argue for a shift from narrow AutoRecSys tools -- focused on algorithm selection and hyper-parameter tuning -- to an Autonomous Recommender-Systems Research Lab (AutoRecLab) that integrates end-to-end automation: problem ideation, literature analysis, experimental design and execution, result interpretation, manuscript drafting, and provenance logging. Drawing on recent progress in automated science (e.g., multi-agent AI Scientist and AI Co-Scientist systems), we outline an agenda for the RecSys community: (1) build open AutoRecLab prototypes that combine LLM-driven ideation and reporting with automated experimentation; (2) establish benchmarks and competitions that evaluate agents on producing reproducible RecSys findings with minimal human input; (3) create review venues for transparently AI-generated submissions; (4) define standards for attribution and reproducibility via detailed research logs and metadata; and (5) foster interdisciplinary dialogue on ethics, governance, privacy, and fairness in autonomous research. Advancing this agenda can increase research throughput, surface non-obvious insights, and position RecSys to contribute to emerging Artificial Research Intelligence. We conclude with a call to organise a community retreat to coordinate next steps and co-author guidance for the responsible integration of automated research systems.

Jarne Mathi Decker, Joeran Beel

The Algorithm Selection Problem for recommender systems-choosing the best algorithm for a given user or context-remains a significant challenge. Traditional meta-learning approaches often treat algorithms as categorical choices, ignoring their intrinsic properties. Recent work has shown that explicitly characterizing algorithms with features can improve model performance in other domains. Building on this, we propose a per-user meta-learning approach for recommender system selection that leverages both user meta-features and automatically extracted algorithm features from source code. Our preliminary results, averaged over six diverse datasets, show that augmenting a meta-learner with algorithm features improves its average NDCG@10 performance by 8.83% from 0.135 (user features only) to 0.147. This enhanced model outperforms the Single Best Algorithm baseline (0.131) and successfully closes 10.5% of the performance gap to a theoretical oracle selector. These findings show that even static source code metrics provide a valuable predictive signal, presenting a promising direction for building more robust and intelligent recommender systems.

Andrew Collins, Laura Tierney, Joeran Beel

Recommendation algorithms perform differently if the users, recommendation contexts, applications, and user interfaces vary even slightly. It is similarly observed in other fields, such as combinatorial problem solving, that algorithms perform differently for each instance presented. In those fields, meta-learning is successfully used to predict an optimal algorithm for each instance, to improve overall system performance. Per-instance algorithm selection has thus far been unsuccessful for recommender systems. In this paper we propose a per-instance meta-learner that clusters data instances and predicts the best algorithm for unseen instances according to cluster membership. We test our approach using 10 collaborative- and 4 content-based filtering algorithms, for varying clustering parameters, and find a significant improvement over the best performing base algorithm at alpha=0.053 (MAE: 0.7107 vs LightGBM 0.7214; t-test). We also explore the performances of our base algorithms on a ratings dataset and empirically show that the error of a perfect algorithm selector monotonically decreases for larger pools of algorithm. To the best of our knowledge, this is the first effective meta-learning technique for per-instance algorithm selection in recommender systems.

Oisín Carroll, Joeran Beel

Games such as go, chess and checkers have multiple equivalent game states, i.e. multiple board positions where symmetrical and opposite moves should be made. These equivalences are not exploited by current state of the art neural agents which instead must relearn similar information, thereby wasting computing time. Group equivariant CNNs in existing work create networks which can exploit symmetries to improve learning, however, they lack the expressiveness to correctly reflect the move embeddings necessary for games. We introduce Finite Group Neural Networks (FGNNs), a method for creating agents with an innate understanding of these board positions. FGNNs are shown to improve the performance of networks playing checkers (draughts), and can be easily adapted to other games and learning problems. Additionally, FGNNs can be created from existing network architectures. These include, for the first time, those with skip connections and arbitrary layer types. We demonstrate that an equivariant version of U-Net (FGNN-U-Net) outperforms the unmodified network in image segmentation.

Mohammed Al-Rawi, Joeran Beel

The fashion industry is looking forward to use artificial intelligence technologies to enhance their processes, services, and applications. Although the amount of fashion data currently in use is increasing, there is a large gap in data exchange between the fashion industry and the related AI companies, not to mention the different structure used for each fashion dataset. As a result, AI companies are relying on manually annotated fashion data to build different applications. Furthermore, as of this writing, the terminology, vocabulary and methods of data representation used to denote fashion items are still ambiguous and confusing. Hence, it is clear that the fashion industry and AI companies will benefit from a protocol that allows them to exchange and organise fashion information in a unified way. To achieve this goal we aim (1) to define a protocol called DDOIF that will allow interoperability of fashion data; (2) for DDOIF to contain diverse entities including extensive information on clothing and accessories attributes in the form of text and various media formats; and (3)To design and implement an API that includes, among other things, functions for importing and exporting a file built according to the DDOIF protocol that stores all information about a single item of clothing. To this end, we identified over 1000 class and subclass names used to name fashion items and use them to build the DDOIF dictionary. We make DDOIF publicly available to all interested users and developers and look forward to engaging more collaborators to improve and enrich it.

Rohan Anand, Joeran Beel

We introduce Auto-Surprise, an Automated Recommender System library. Auto-Surprise is an extension of the Surprise recommender system library and eases the algorithm selection and configuration process. Compared to out-of-the-box Surprise library, Auto-Surprise performs better when evaluated with MovieLens, Book Crossing and Jester Datasets. It may also result in the selection of an algorithm with significantly lower runtime. Compared to Surprise's grid search, Auto-Surprise performs equally well or slightly better in terms of RMSE, and is notably faster in finding the optimum hyperparameters.

Joeran Beel, Bryan Tyrell, Edward Bergman et al.

Automated per-instance algorithm selection often outperforms single learners. Key to algorithm selection via meta-learning is often the (meta) features, which sometimes though do not provide enough information to train a meta-learner effectively. We propose a Siamese Neural Network architecture for automated algorithm selection that focuses more on 'alike performing' instances than meta-features. Our work includes a novel performance metric and method for selecting training samples. We introduce further the concept of 'Algorithm Performance Personas' that describe instances for which the single algorithms perform alike. The concept of 'alike performing algorithms' as ground truth for selecting training samples is novel and provides a huge potential as we believe. In this proposal, we outline our ideas in detail and provide the first evidence that our proposed metric is better suitable for training sample selection that standard performance metrics such as absolute errors.

Mark Grennan, Joeran Beel

Citation parsing, particularly with deep neural networks, suffers from a lack of training data as available datasets typically contain only a few thousand training instances. Manually labelling citation strings is very time-consuming, hence synthetically created training data could be a solution. However, as of now, it is unknown if synthetically created reference-strings are suitable to train machine learning algorithms for citation parsing. To find out, we train Grobid, which uses Conditional Random Fields, with a) human-labelled reference strings from 'real' bibliographies and b) synthetically created reference strings from the GIANT dataset. We find that both synthetic and organic reference strings are equally suited for training Grobid (F1 = 0.74). We additionally find that retraining Grobid has a notable impact on its performance, for both synthetic and real data (+30% in F1). Having as many types of labelled fields as possible during training also improves effectiveness, even if these fields are not available in the evaluation data (+13.5% F1). We conclude that synthetic data is suitable for training (deep) citation parsing models. We further suggest that in future evaluations of reference parsers both evaluation data similar and dissimilar to the training data should be used for more meaningful evaluations.

Andrew Collins, Joeran Beel

The effectiveness of recommender system algorithms varies in different real-world scenarios. It is difficult to choose a best algorithm for a scenario due to the quantity of algorithms available, and because of their varying performances. Furthermore, it is not possible to choose one single algorithm that will work optimally for all recommendation requests. We apply meta-learning to this problem of algorithm selection for scholarly article recommendation. We train a random forest, gradient boosting machine, and generalized linear model, to predict a best-algorithm from a pool of content similarity-based algorithms. We evaluate our approach on an offline dataset for scholarly article recommendation and attempt to predict the best algorithm per-instance. The best meta-learning model achieved an average increase in F1 of 88% when compared to the average F1 of all base-algorithms (F1; 0.0708 vs 0.0376) and was significantly able to correctly select each base-algorithm (Paired t-test; p < 0.1). The meta-learner had a 3% higher F1 when compared to the single-best base-algorithm (F1; 0.0739 vs 0.0717). We further perform an online evaluation of our approach, conducting an A/B test through our recommender-as-a-service platform Mr. DLib. We deliver 148K recommendations to users between January and March 2019. User engagement was significantly increased for recommendations generated using our meta-learning approach when compared to a random selection of algorithm (Click-through rate (CTR); 0.51% vs. 0.44%, Chi-Squared test; p < 0.1), however our approach did not produce a higher CTR than the best algorithm alone (CTR; MoreLikeThis (Title): 0.58%).

Conor O'Sullivan, Joeran Beel

In this study, machine learning models were constructed to predict whether judgments made by the European Court of Human Rights (ECHR) would lead to a violation of an Article in the Convention on Human Rights. The problem is framed as a binary classification task where a judgment can lead to a "violation" or "non-violation" of a particular Article. Using auto-sklearn, an automated algorithm selection package, models were constructed for 12 Articles in the Convention. To train these models, textual features were obtained from the ECHR Judgment documents using N-grams, word embeddings and paragraph embeddings. Additional documents, from the ECHR, were incorporated into the models through the creation of a word embedding (echr2vec) and a doc2vec model. The features obtained using the echr2vec embedding provided the highest cross-validation accuracy for 5 of the Articles. The overall test accuracy, across the 12 Articles, was 68.83%. As far as we could tell, this is the first estimate of the accuracy of such machine learning models using a realistic test set. This provides an important benchmark for future work. As a baseline, a simple heuristic of always predicting the most common outcome in the past was used. The heuristic achieved an overall test accuracy of 86.68% which is 29.7% higher than the models. Again, this was seemingly the first study that included such a heuristic with which to compare model results. The higher accuracy achieved by the heuristic highlights the importance of including such a baseline.

Nicholas Bonello, Joeran Beel, Seamus Lawless et al.

Fantasy Premier League (FPL) performance predictors tend to base their algorithms purely on historical statistical data. The main problems with this approach is that external factors such as injuries, managerial decisions and other tournament match statistics can never be factored into the final predictions. In this paper, we present a new method for predicting future player performances by automatically incorporating human feedback into our model. Through statistical data analysis such as previous performances, upcoming fixture difficulty ratings, betting market analysis, opinions of the general-public and experts alike via social media and web articles, we can improve our understanding of who is likely to perform well in upcoming matches. When tested on the English Premier League 2018/19 season, the model outperformed regular statistical predictors by over 300 points, an average of 11 points per week, ranking within the top 0.5% of players rank 30,000 out of over 6.5 million players.

Dominika Tkaczyk, Andrew Collins, Joeran Beel

Information about the contributions of individual authors to scientific publications is important for assessing authors' achievements. Some biomedical publications have a short section that describes authors' roles and contributions. It is usually written in natural language and hence author contributions cannot be trivially extracted in a machine-readable format. In this paper, we present 1) A statistical analysis of roles in author contributions sections, and 2) NaïveRole, a novel approach to extract structured authors' roles from author contribution sections. For the first part, we used co-clustering techniques, as well as Open Information Extraction, to semi-automatically discover the popular roles within a corpus of 2,000 contributions sections from PubMed Central. The discovered roles were used to automatically build a training set for NaïveRole, our role extractor approach, based on Naïve Bayes. NaïveRole extracts roles with a micro-averaged precision of 0.68, recall of 0.48 and F1 of 0.57. It is, to the best of our knowledge, the first attempt to automatically extract author roles from research papers. This paper is an extended version of a previous poster published at JCDL 2018.

Mark Collier, Joeran Beel

Memory-augmented neural networks (MANNs) have been shown to outperform other recurrent neural network architectures on a series of artificial sequence learning tasks, yet they have had limited application to real-world tasks. We evaluate direct application of Neural Turing Machines (NTM) and Differentiable Neural Computers (DNC) to machine translation. We further propose and evaluate two models which extend the attentional encoder-decoder with capabilities inspired by memory augmented neural networks. We evaluate our proposed models on IWSLT Vietnamese to English and ACL Romanian to English datasets. Our proposed models and the memory augmented neural networks perform similarly to the attentional encoder-decoder on the Vietnamese to English translation task while have a 0.3-1.9 lower BLEU score for the Romanian to English task. Interestingly, our analysis shows that despite being equipped with additional flexibility and being randomly initialized memory augmented neural networks learn an algorithm for machine translation almost identical to the attentional encoder-decoder.

Andrew Collins, Joeran Beel

Many recommendation algorithms are available to digital library recommender system operators. The effectiveness of algorithms is largely unreported by way of online evaluation. We compare a standard term-based recommendation approach to two promising approaches for related-article recommendation in digital libraries: document embeddings, and keyphrases. We evaluate the consistency of their performance across multiple scenarios. Through our recommender-as-a-service Mr. DLib, we delivered 33.5M recommendations to users of Sowiport and Jabref over the course of 19 months, from March 2017 to October 2018. The effectiveness of the algorithms differs significantly between Sowiport and Jabref (Wilcoxon rank-sum test; p < 0.05). There is a ~400% difference in effectiveness between the best and worst algorithm in both scenarios separately. The best performing algorithm in Sowiport (terms) is the worst performing in Jabref. The best performing algorithm in Jabref (keyphrases) is 70% worse in Sowiport, than Sowiport`s best algorithm (click-through rate; 0.1% terms, 0.03% keyphrases).

Mark Collier, Joeran Beel

Syllabuses for curriculum learning have been developed on an ad-hoc, per task basis and little is known about the relative performance of different syllabuses. We identify a number of syllabuses used in the literature. We compare the identified syllabuses based on their effect on the speed of learning and generalization ability of a LSTM network on three sequential learning tasks. We find that the choice of syllabus has limited effect on the generalization ability of a trained network. In terms of speed of learning our results demonstrate that the best syllabus is task dependent but that a recently proposed automated curriculum learning approach - Predictive Gain, performs very competitively against all identified hand-crafted syllabuses. The best performing hand-crafted syllabus which we term Look Back and Forward combines a syllabus which steps through tasks in the order of their difficulty with a uniform distribution over all tasks. Our experimental results provide an empirical basis for the choice of syllabus on a new problem that could benefit from curriculum learning. Additionally, insights derived from our results shed light on how to successfully design new syllabuses.

Dominika Tkaczyk, Paraic Sheridan, Joeran Beel

Bibliographic reference parsers extract metadata (e.g. author names, title, year) from bibliographic reference strings. No reference parser consistently gives the best results in every scenario. For instance, one tool may be best in extracting titles, and another tool in extracting author names. In this paper, we address the problem of reference parsing from a recommender-systems perspective. We propose ParsRec, a meta-learning approach that recommends the potentially best parser(s) for a given reference string. We evaluate ParsRec on 105k references from chemistry. We propose two approaches to meta-learning recommendations. The first approach learns the best parser for an entire reference string. The second approach learns the best parser for each field of a reference string. The second approach achieved a 2.6% increase in F1 (0.909 vs. 0.886, p < 0.001) over the best single parser (GROBID), reducing the false positive rate by 20.2% (0.075 vs. 0.094), and the false negative rate by 18.9% (0.107 vs. 0.132).

Joeran Beel, Andrew Collins, Oliver Kopp et al.

We introduce the first 'living lab' for scholarly recommender systems. This lab allows recommender-system researchers to conduct online evaluations of their novel algorithms for scholarly recommendations, i.e., recommendations for research papers, citations, conferences, research grants, etc. Recommendations are delivered through the living lab's API to platforms such as reference management software and digital libraries. The living lab is built on top of the recommender-system as-a-service Mr. DLib. Current partners are the reference management software JabRef and the CORE research team. We present the architecture of Mr. DLib's living lab as well as usage statistics on the first sixteen months of operating it. During this time, 1,826,643 recommendations were delivered with an average click-through rate of 0.21%.

Joeran Beel, Barry Smyth, Andrew Collins

The main contribution of this paper is to introduce and describe a new recommender-systems dataset (RARD II). It is based on data from Mr. DLib, a recommender-system as-a-service in the digital library and reference-management-software domain. As such, RARD II complements datasets from other domains such as books, movies, and music. The dataset encompasses 94m recommendations, delivered in the two years from September 2016 to September 2018. The dataset covers an item-space of 24m unique items. RARD II provides a range of rich recommendation data, beyond conventional ratings. For example, in addition to the usual (implicit) ratings matrices, RARD II includes the original recommendation logs, which provide a unique insight into many aspects of the algorithms that generated the recommendations. The logs enable researchers to conduct various analyses about a real-world recommender system. This includes the evaluation of meta-learning approaches for predicting algorithm performance. In this paper, we summarise the key features of this dataset release, describe how it was generated and discuss some of its unique features. Compared to its predecessor RARD, RARD II contains 64% more recommendations, 187% more features (algorithms, parameters, and statistics), 50% more clicks, 140% more documents, and one additional service partner (JabRef).

Andrew Collins, Dominika Tkaczyk, Joeran Beel

The effectiveness of recommendation algorithms is typically assessed with evaluation metrics such as root mean square error, F1, or click through rates, calculated over entire datasets. The best algorithm is typically chosen based on these overall metrics. However, there is no single-best algorithm for all users, items, and contexts. Choosing a single algorithm based on overall evaluation results is not optimal. In this paper, we propose a meta-learning-based approach to recommendation, which aims to select the best algorithm for each user-item pair. We evaluate our approach using the MovieLens 100K and 1M datasets. Our approach (RMSE, 100K: 0.973; 1M: 0.908) did not outperform the single-best algorithm, SVD++ (RMSE, 100K: 0.942; 1M: 0.887). We also develop a distinction between meta-learners that operate per-instance (micro-level), per-data subset (mid-level), and per-dataset (global level). Our evaluation shows that a hypothetically perfect micro-level meta-learner would improve RMSE by 25.5% for the MovieLens 100K and 1M datasets, compared to the overall-best algorithms used.

Andrew Collins, Dominika Tkaczyk, Akiko Aizawa et al.

"Position bias" describes the tendency of users to interact with items on top of a list with higher probability than with items at a lower position in the list, regardless of the items' actual relevance. In the domain of recommender systems, particularly recommender systems in digital libraries, position bias has received little attention. We conduct a study in a real-world recommender system that delivered ten million related-article recommendations to the users of the digital library Sowiport, and the reference manager JabRef. Recommendations were randomly chosen to be shuffled or non-shuffled, and we compared click-through rate (CTR) for each rank of the recommendations. According to our analysis, the CTR for the highest rank in the case of Sowiport is 53% higher than expected in a hypothetical non-biased situation (0.189% vs. 0.123%). Similarly, in the case of Jabref the highest rank received a CTR of 1.276%, which is 87% higher than expected (0.683%). A chi-squared test confirms the strong relationship between the rank of the recommendation shown to the user and whether the user decided to click it (p < 0.01 for both Jabref and Sowiport). Our study confirms the findings from other domains, that recommendations in the top positions are more often clicked, regardless of their actual relevance.

Joeran Beel

In this position paper, we question the current practice of calculating evaluation metrics for recommender systems as single numbers (e.g. precision p=.28 or mean absolute error MAE = 1.21). We argue that single numbers express only average effectiveness over a usually rather long period (e.g. a year or even longer), which provides only a vague and static view of the data. We propose that recommender-system researchers should instead calculate metrics for time-series such as weeks or months, and plot the results in e.g. a line chart. This way, results show how algorithms' effectiveness develops over time, and hence the results allow drawing more meaningful conclusions about how an algorithm will perform in the future. In this paper, we explain our reasoning, provide an example to illustrate our reasoning and present suggestions for what the community should do next.

Joeran Beel, Zeljko Carevic, Johann Schaible et al.

Recommender-system datasets are used for recommender-system evaluations, training machine-learning algorithms, and exploring user behavior. While there are many datasets for recommender systems in the domains of movies, books, and music, there are rather few datasets from research-paper recommender systems. In this paper, we introduce RARD, the Related-Article Recommendation Dataset, from the digital library Sowiport and the recommendation-as-a-service provider Mr. DLib. The dataset contains information about 57.4 million recommendations that were displayed to the users of Sowiport. Information includes details on which recommendation approaches were used (e.g. content-based filtering, stereotype, most popular), what types of features were used in content based filtering (simple terms vs. keyphrases), where the features were extracted from (title or abstract), and the time when recommendations were delivered and clicked. In addition, the dataset contains an implicit item-item rating matrix that was created based on the recommendation click logs. RARD enables researchers to train machine learning algorithms for research-paper recommendations, perform offline evaluations, and do research on data from Mr. DLib's recommender system, without implementing a recommender system themselves. In the field of scientific recommender systems, our dataset is unique. To the best of our knowledge, there is no dataset with more (implicit) ratings available, and that many variations of recommendation algorithms. The dataset is available at http://data.mr-dlib.org, and published under the Creative Commons Attribution 3.0 Unported (CC-BY) license.

Felix Beierle, Akiko Aizawa, Joeran Beel

We investigate the problem of choice overload - the difficulty of making a decision when faced with many options - when displaying related-article recommendations in digital libraries. So far, research regarding to how many items should be displayed has mostly been done in the fields of media recommendations and search engines. We analyze the number of recommendations in current digital libraries. When browsing fullscreen with a laptop or desktop PC, all display a fixed number of recommendations. 72% display three, four, or five recommendations, none display more than ten. We provide results from an empirical evaluation conducted with GESIS' digital library Sowiport, with recommendations delivered by recommendations-as-a-service provider Mr. DLib. We use click-through rate as a measure of recommendation effectiveness based on 3.4 million delivered recommendations. Our results show lower click-through rates for higher numbers of recommendations and twice as many clicked recommendations when displaying ten instead of one related-articles. Our results indicate that users might quickly feel overloaded by choice.

Joeran Beel, Siddharth Dinesh

Research on recommender systems is a challenging task, as is building and operating such systems. Major challenges include non-reproducible research results, dealing with noisy data, and answering many questions such as how many recommendations to display, how often, and, of course, how to generate recommendations most effectively. In the past six years, we built three research-article recommender systems for digital libraries and reference managers, and conducted research on these systems. In this paper, we share some experiences we made during that time. Among others, we discuss the required skills to build recommender systems, and why the literature provides little help in identifying promising recommendation approaches. We explain the challenge in creating a randomization engine to run A/B tests, and how low data quality impacts the calculation of bibliometrics. We further discuss why several of our experiments delivered disappointing results, and provide statistics on how many researchers showed interest in our recommendation dataset.

Joeran Beel

While user-modeling and recommender systems successfully utilize items like emails, news, and movies, they widely neglect mind-maps as a source for user modeling. We consider this a serious shortcoming since we assume user modeling based on mind maps to be equally effective as user modeling based on other items. Hence, millions of mind-mapping users could benefit from user-modeling applications such as recommender systems. The objective of this doctoral thesis is to develop an effective user-modeling approach based on mind maps. To achieve this objective, we integrate a recommender system in our mind-mapping and reference-management software Docear. The recommender system builds user models based on the mind maps, and recommends research papers based on the user models. As part of our research, we identify several variables relating to mind-map-based user modeling, and evaluate the variables' impact on user-modeling effectiveness with an offline evaluation, a user study, and an online evaluation based on 430,893 recommendations displayed to 4,700 users. We find, among others, that the number of analyzed nodes, modification time, visibility of nodes, relations between nodes, and number of children and siblings of a node affect the effectiveness of user modeling. When all variables are combined in a favorable way, this novel approach achieves click-through rates of 7.20%, which is nearly twice as effective as the best baseline. In addition, we show that user modeling based on mind maps performs about as well as user modeling based on other items, namely the research articles users downloaded or cited. Our findings let us to conclude that user modeling based on mind maps is a promising research field, and that developers of mind-mapping applications should integrate recommender systems into their applications. Such systems could create additional value for millions of mind-mapping users.

Joeran Beel, Akiko Aizawa, Corinna Breitinger et al.

Only few digital libraries and reference managers offer recommender systems, although such systems could assist users facing information overload. In this paper, we introduce Mr. DLib's recommendations-as-a-service, which allows third parties to easily integrate a recommender system into their products. We explain the recommender approaches implemented in Mr. DLib (content-based filtering among others), and present details on 57 million recommendations, which Mr. DLib delivered to its partner GESIS Sowiport. Finally, we outline our plans for future development, including integration into JabRef, establishing a living lab, and providing personalized recommendations.

Stefan Langer, Joeran Beel

For the past few years, we used Apache Lucene as recommendation frame-work in our scholarly-literature recommender system of the reference-management software Docear. In this paper, we share three lessons learned from our work with Lucene. First, recommendations with relevance scores below 0.025 tend to have significantly lower click-through rates than recommendations with relevance scores above 0.025. Second, by picking ten recommendations randomly from Lucene's top50 search results, click-through rate decreased by 15%, compared to recommending the top10 results. Third, the number of returned search results tend to predict how high click-through rates will be: when Lucene returns less than 1,000 search results, click-through rates tend to be around half as high as if 1,000+ results are returned.