Fine-tuning pre-trained language models has become the state-of-the-art approach for Entity Resolution (ER), but this has created a divide between two dominant architectures: fast-but-less-accurate bi-encoders and accurate-but-slow cross-encoders. However, a concrete gap in prior ER benchmarking remains unresolved: existing studies often evaluate architectures in isolation or on limited datasets. It remains unclear which base models and architectures are best suited for the diverse range of real-world ER datasets, each with unique characteristics and performance bottlenecks. This paper bridges this gap through an extensive empirical evaluation. We systematically compare three popular pre-trained models (MiniLM, MPNet, and BGE) across three distinct architectural paradigms: a pre-trained bi-encoder, a fine-tuned bi-encoder, and a fine-tuned cross-encoder. We tested these combinations on eight diverse real-world and semi-synthetic datasets, analyzing their performance, training costs, and final resolution times. Our results reveal a clear accuracy-vs-efficiency trade-off, identifying the fine-tuned bi-encoder as the optimal balance between performance and practical resolution speed. More importantly, we demonstrate that fine-tuning is not a universal solution. Its effectiveness is highly contingent on the dataset: it provides substantial gains on specialized domains by fixing pre-existing performance gaps but is detrimental to performance on datasets where pre-trained models are already well-aligned. These findings provide a practical guide for practitioners on selecting the optimal model and architecture based on their specific data and application requirements.

The problem of data imbalance is defined as the uneven distribution of the training examples to the existing classes of a dataset. Among a wide variety of solutions, the oversampling techniques try to mitigate the problem by synthesizing artificial examples associated with the minority class. The huge success of Generative Adversarial Networks (GANs) rendered them an attractive choice for oversampling and numerous researchers proposed modifications of GANs for imbalanced datasets. Nevertheless, the existing models employ the entire minority class for sample generation, thus being vulnerable to outliers and noisy data instances. In addition, the majority of the relevant research concerns image classification tasks, leaving a large gap for research with tabular data. Finally, another powerful and popular generative model, the Variational Autoencoder (VAE) has been rather overlooked by the community in class imbalance solutions. In this paper we present SB-GAN and SB-VAE, two generative models that identify borderline and noisy samples before they are trained. In this manner SB-GAN and SB-VAE learn better class distributions that are not distorted by the existence of outliers. The experimental evaluation of SB-GAN and SB-VAE with 4 tabular datasets revealed a superior performance against 8 state-of-the-art oversampling techniques.

Short text clustering deals with the problem of grouping together semantically similar documents with small lengths. Nowadays, huge amounts of text data is being generated by numerous applications such as microblogs, messengers, and services that generate or aggregate entitled entities. This large volume of highly dimensional and sparse information may easily overwhelm the current serial approaches and render them inefficient, or even inapplicable. Although many traditional clustering algorithms have been successfully parallelized in the past, the parallelization of short text clustering algorithms is a rather overlooked problem. In this paper we introduce pVEPHC, a short text clustering method that can be executed in parallel in large computer clusters. The algorithm draws inspiration from VEPHC, a recent two-stage approach with decent performance in several diverse tasks. More specifically, in this work we employ the Apache Spark framework to design parallel implementations of both stages of VEPHC. During the first stage, pVEPHC generates an initial clustering by identifying and modelling common low-dimensional vector representations of the original documents. In the sequel, the initial clustering is improved in the second stage by applying cluster split and merge operations in a hierarchical fashion. We have attested our implementation on an experimental Spark cluster and we report an almost linear improvement in the execution times of the algorithm.

Short text clustering is a popular problem that focuses on the unsupervised grouping of similar short text documents, or entitled entities. Since the short texts are currently being utilized in a vast number of applications, the problem in question has been rendered increasingly significant in the past few years. The high cluster homogeneity and completeness are two among the most important goals of all data clustering algorithms. However, in the context of short texts, their fulfilment is particularly difficult, because this type of data is typically represented by sparse vectors that collectively comprise a very high dimensional space. In this article we introduce VEPHC, a two-stage clustering algorithm designed to confront the sparseness and high dimensionality traits of short texts. During the first stage (or else, the VEP part), the initial feature vectors are projected onto a lower dimensional space by constructing and scoring variable-sized combinations of features (that is, terms). In the second stage (or else, the HC part), VEPHC improves the homogeneity and completeness of the generated clusters through split and merge operations that are based on the similarities of all inter-cluster elements. The experimental evaluation of VEPHC on two real-world datasets demonstrates its superior performance over numerous state-of-the-art clustering algorithms in terms of F1 scores and Normalized Mutual Information.

The introduction of flash SSDs has accelerated the performance of DBMSes. However, the intrinsic characteristics of flash motivated many researchers to investigate new efficient data structures. The emergence of 3DXPoint, a new non-volatile memory, sets new challenges: 3DXPoint features low latency and high IOPS even at small queue depths. However, the cost of 3DXPoint is 4 times higher than that of a flash-based device, rendering hybrid storage systems a good alternative. In this paper we pursue exploiting the efficiency of both 3DXPoint and flash-based devices introducing H-Grid, a variant of Grid-File for hybrid storage. H-Grid uses a flash SSD as main store and a small 3DXPoint device to persist the hottest data. The performance of the proposed index is experimentally evaluated, comparing it against GFFM, a flash efficient implementation of Grid File. The results show that H-Grid is faster than GFFM execution on a flash SSD, reducing the single point search time from 35% up to 43%.

Rank aggregation is a popular problem that combines different ranked lists from various sources (frequently called voters or judges), and generates a single aggregated list with improved ranking of its items. In this context, a portion of the existing methods attempt to address the problem by treating all voters equally. Nevertheless, several related works proved that the careful and effective assignment of different weights to each voter leads to enhanced performance. In this article, we introduce an unsupervised algorithm for learning the weights of the voters for a specific topic or query. The proposed method is based on the fact that if a voter has submitted numerous elements which have been placed in high positions in the aggregated list, then this voter should be treated as an expert, compared to the voters whose suggestions appear in lower places or do not appear at all. The algorithm iteratively computes the distance of each input list with the aggregated list and modifies the weights of the voters until all weights converge. The effectiveness of the proposed method is experimentally demonstrated by aggregating input lists from six TREC conferences.

During the past few years, the e-commerce platforms and marketplaces have enriched their services with new features to improve their user experience and increase their profitability. Such features include relevant products suggestion, personalized recommendations, query understanding algorithms and numerous others. To effectively implement all these features, a robust products categorization method is required. Due to its importance, the problem of the automatic products classification into a given taxonomy has attracted the attention of multiple researchers. In the current literature, we encounter a broad variety of solutions, ranging from supervised and deep learning algorithms, as well as convolutional and recurrent neural networks. In this paper we introduce a supervised learning method which performs morphological analysis of the product titles by extracting and processing a combination of words and n-grams. In the sequel, each of these tokens receives an importance score according to several criteria which reflect the strength of the correlation of the token with a category. Based on these importance scores, we also propose a dimensionality reduction technique to reduce the size of the feature space without sacrificing much of the performance of the algorithm. The experimental evaluation of our method was conducted by using a real-world dataset, comprised of approximately 320 thousand product titles, which we acquired by crawling a product comparison Web platform. The results of this evaluation indicate that our approach is highly accurate, since it achieves a remarkable classification accuracy of over 95%.

In the big data era, the efficient processing of large volumes of data has became a standard requirement for both organizations and enterprises. Since single workstations cannot sustain such tremendous workloads, MapReduce was introduced with the aim of providing a robust, easy, and fault-tolerant parallelization framework for the execution of applications on large clusters. One of the most representative examples of such applications is the machine learning algorithms which dominate the broad research area of data mining. Simultaneously, the recent advances in hardware technology led to the introduction of high-performing alternative devices for secondary storage, known as Solid State Drives (SSDs). In this paper we examine the perfor-mance of several parallel data mining algorithms on MapReduce clusters equipped with such modern hardware. More specifically, we investigate standard dataset preprocessing methods including vectorization and dimensionality reduction, and two supervised classifiers, Naive Bayes and Linear Regression. We compare the execution times of these algorithms on an experimental cluster equipped with both standard magnetic disks and SSDs, by employing two different datasets and by applying several different cluster configurations. Our experiments demonstrate that the usage of SSDs can accelerate the execution of machine learning methods by a margin which depends on the cluster setup and the nature of the applied algorithms.

The problem of matching product titles is of particular interest for both users and marketers. The former, frequently search the Web with the aim of comparing prices and characteristics, or obtaining and aggregating information provided by other users. The latter, often require wide knowledge of competitive policies, prices and features to organize a promotional campaign about a group of products. To address this interesting problem, recent studies have attempted to enrich the product titles by exploiting Web search engines. More specifically, these methods suggest that for each product title a query should be submitted. After the results have been collected, the most important words which appear in the results are identified and appended in the titles. In the sequel, each word is assigned an importance score and finally, a similarity measure is applied to identify if two or more titles refer to the same product. Nonetheless, these methods have multiple problems including scalability, slow retrieval of the required additional search results, and lack of flexibility. In this paper, we present a different approach which addresses all these issues and is based on the morphological analysis of the titles of the products. In particular, our method operates in two phases. In the first phase, we compute the combinations of the words of the titles and we record several statistics such as word proximity and frequency values. In the second phase, we use this information to assign a score to each combination. The highest scoring combination is then declared as label of the cluster which contains each product. The experimental evaluation of the algorithm, in a real world dataset, demonstrated that compared to three popular string similarity metrics, our approach achieves up to 36% better matching performance and at least 13 times faster execution.

In smart cities residential homes are fully equipped with information networking and computing technologies and are connected to the power grid via intelligent meters. Connectivity of meters allows formation of groups of residents, which are physically close, and as a result individual consumptions can be aggregated into a shared consumption. In this paper an approach of unfolding shared consumption and making inferences about resident personal usage is presented. The proposed approach tackles the problem of unfolding as a multiobjective problem in which a set of residential profiles is fitted to the measured consumption. A solution to the multiobjective problem is sought by using the Non-dominated Sorting Genetic Algorithm-II (NSGA-II) that utilizes the Pareto optimality theory to identify an optimal solution. The approach is applied to a set electricity consumption signals for making inferences about the personal energy usage of residential participants in the shared consumption pattern.

In the recent years, there is a growing interest in exploiting the advances of mobile and pervasive computing to museum environments. A mobile museum guide, named iMuse Mobile Tour is presented in this paper. The guide utilizes UHF RFID technology to provide context aware information services. It comprises predefined and self-defined tours as well as interactive games to stimulate learning. A group support service is introduced, which enables group visitors to exploit guidance services on their private mobile phones. The service provides access to information retrieved by museum's RFID enabled devices.

Social networking has become one of the most important trends on the Web, leading to the development of several social applications such as blogs. Blogs are locations on the Web where individuals are provided with the ability to express their opinion, experience, and knowledge about a product, an event, or any other subject. The tremendous popularity of these services has rendered the problem of identifying the most influential bloggers significant, since its solution can lead to numerous major benefits for commerce, advertising, and searching. The current works on this topic either ignore temporal aspects or they fail to gracefully incorporate recency, productivity, and influence at the same time. This paper investigates the issue of identifying bloggers who are both productive and influential by introducing the blogger's productivity index and blogger's influence index. The proposed metrics are evaluated against the state-of-the-art influential blogger identification methods by employing data collected from a real-world community blog site. The obtained results confirm that the new methods are able to identify significant patterns in the bloggers' behavior.

Blogs have recently become one of the most favored services on the Web. Many users maintain a blog and write posts to express their opinion, experience and knowledge about a product, an event and every subject of general or specific interest. More users visit blogs to read these posts and comment them. This “participatory journalism” of blogs has such an impact upon the masses that Keller and Berry argued that through blogging “one American in tens tells the other nine how to vote, where to eat and what to buy” [9]. Therefore, a significant issue is how to identify such influential bloggers. This problem is very new and the relevant literature lacks sophisticated solutions, but most importantly these solutions have not taken into account temporal aspects for identifying influential bloggers, even though the time is the most critical aspect of the Blogosphere. This article investigates the issue of identifying influential bloggers by proposing two easily computed blogger ranking methods, which incorporate temporal aspects of the blogging activity. Each method is based on a specific metric to score the blogger’s posts. The first metric, termed MEIBI, takes into consideration the number of the blog post’s inlinks and its comments, along with the publication date of the post. The second metric, MEIBIX, is used to score a blog post according to the number and age of the blog post’s inlinks and its comments. These methods are evaluated against the state-of-the-art influential blogger identification method utilizing data collected from a real-world community blog site. The obtained results attest that the new methods are able to better identify significant temporal patterns in the blogging behaviour

Abstract Designing fair and unbiased metrics to measure the “level of excellence” of a scientist is a very significant task because they recently also have been taken into account when deciding faculty promotions, when allocating funds, and so on. Despite criticism that such scientometric evaluators are confronted with, they do have their merits, and efforts should be spent to arm them with robustness and resistance to manipulation. This article aims at initiating the study of the coterminal citations—their existence and implications—and presents them as a generalization of self‐citations and of co‐citation; it also shows how they can be used to capture any manipulation attempts against scientometric indicators, and finally presents a new index, the f index, that takes into account the coterminal citations. The utility of the new index is validated using the academic production of a number of esteemed computer scientists. The results confirm that the new index can discriminate those individuals whose work penetrates many scientific communities.

Since its introduction in 1984, R-tree has been proven to be one of the most practical and well-behaved data structures for accommodating dynamic massive sets of geometric objects and conducting a very diverse set of queries on such datasets in real-world applications. This success has led to a variety of versions, each one trying to tune the performance parameters of the original proposal. Among them, the most prominent one is R*-tree, which employs a number of carefully designed heuristics and is widely ccepted as achieving the best performance in most cases. However, in the presence of actively changing datasets, R*-tree still does not avoid performance tuning with forced reinsertion, i.e. a process that performs a kind of local rebuilding. The latter fact has motivated the investigation of the adaptation of a known dynamization technique, based on carefully triggered local rebuildings, for converting static or semi-dynamic, main memory data structures to dynamic ones onto R*-trees. In this paper, we present LR-trees, a new efficient scheme for dynamic manipulation of large datasets, which combines the search performance of the bulk-loaded R-trees with the updated performance of R*-trees. Experimental results provide evidence on the latter statement and illustrate the superiority of the proposed method.

We present simple algorithms for three problems belonging to the class of intersection query problems. The first algorithm deals with the static rectangle enclosure problem and can easily be extended to d dimensions, the second algorithm copes with the generalized c-oriented polygon intersection searching problem in two dimensions, while the third solves the static 2-dimensional dominance searching problem with respect to a set of obstacles. All algorithms are simple, are based on persistence and improve previous bounds. Also, as a corollary of the first algorithm, we present a result for the static d-dimensional range searching problem.

Let Q 1 be a set of'red' geometric objects and Q 2 a set of 'blue' ones. The objects in Q 1 and Q 2 are of arbitrarily large description size but each one is the union of simpler constant size components. We consider the problem of reporting all intersections between objects of Q 1 and objects of Q 2 in a time that depends on the size of the output. This is equivalent to painting the constant size objects in Q 1 and Q 2 and reporting colour intersections between Q 1 and Q 2 . We present a technique that yields simple output-sensitive algorithms for many kinds of geometric objects. We also show that, using the same technique, it is possible to report all intersecting pairs in a set of objects in an output-sensitive manner.

The retrieval capabilities of the signature file access method have become very attractive for data processing applications dealing with both formatted and unformatted data. Multimedia databases should also take advantage of this method, provided convenient storage structures for organizing signature files are available. In this paper a high performance dynamic signature file organization is proposed, combining the latest developments in storage structures (dynamic partitioning schemes) with the classical multilevel organizations in an efficient and effective way. We call the method the parametric weighted filter (PWF) and the performance analysis shows its superiority over previous methods. Finally, refinements are proposed that can further improve the scheme and that are worth further investigation.