Abstract Recommendation (recommender) systems (RS) have played a significant role in both research and industry in recent years. In the area of academia, there is a need to help researchers discover the most appropriate and relevant scientific information through recommendations. Nevertheless, we argue that there is a major gap between academic state-of-the-art RS and real-world problems. In this paper, we present a novel multi-staged RS based on clustering, graph modeling and deep learning that manages to run on a full dataset (scientific digital library) in the magnitude of millions users and items (papers). We run several tests (experiments/evaluation) as a means to find the best approach regarding the tuning of our system; so, we present and compare three versions of our RS regarding recall and NDCG metrics. The results show that a multi-staged RS that utilizes a variety of techniques and algorithms is able to face real-world problems and large academic datasets. In this way, we suggest a way to close or minimize the gap between research and industry value RS.

The k Nearest-Neighbor (k-NN) query is a common spatial query that appears in several big data applications. Typically, GPU devices have much larger numbers of processing cores than CPUs and faster device memory than main memory accessed by CPUs, thus, providing higher computing power. We propose and implement a new GPU-based partitioning algorithm for the k-NN query, using the CUDA runtime API. Due to partitioning, this algorithm avoids calculating distances for the whole dataset. Using synthetic and real datasets, we present an extensive experimental performance comparison against six existing algorithms. These algorithms are based on calculating distances for the whole in-memory dataset. This comparison shows that the new algorithm excels in all the conducted experiments and outperforms these six algorithms.

Spatial access methods have been extensively studied in the literature, during last decades. Access methods were designed for efficient processing of demanding queries and extensive comparisons between such methods have been presented. However, choosing the best values for the parameters that affect the performance of a spatial access method when such a method is expected to be utilized within a specific workload / context of operations has not been studied, so far. In this paper, we present the design and implementation of a framework to evaluate and optimize a spatial index. The (very popular) family of R-trees is chosen as the index of focus, though the same process can be applied for other (spatial, or not) indexes or combinations of them. We elaborate on the antagonizing aspects in the design of an R-tree, present the design and implementation of a benchmarking framework and develop a performance model for this index that incorporates benchmarking results. Next, we develop an optimization framework that uses this model to provide an optimized set up (node occupancies and node splitting method) for a specific use context (dataset type, tree size, number of queries and type of queries). We also present experimental results of an indicative use of the developed benchmarking framework and optimizer for a limited range of use contexts.

In modern database applications the similarity or dissimilarity of complex objects is examined by performing distance-based queries (DBQs) on data of high dimensionality. The R-tree and its variations are commonly cited multidimensional access methods that can beused for answering such queries. Although the related algorithms work well for low-dimensional data spaces, their performance degrades as the number of dimensions increases (dimensionality curse). In order to obtain acceptable response time in high-dimensional data spaces, algorithms that obtain approximate solutions can be used. Approximation techniques, like N-consider (based on the tree structure), α-allowance and ε-approximate (based on distance), or Time-consider (based on time) can be applied in branch-and-bound algorithms for DBQs inorder to control the trade-off between cost and accuracy of the result. In this paper, we improve previous approximate DBQ algorithms by applying a combination of the approximation techniques in the same query algorithm (hybrid approximation scheme). We investigate the performance of these improvements for one of the most representative DBQs (the K-closest pairs query, K-CPQ) in high-dimensional data spaces, as well as the influence of the algorithmic parameters on the control of the trade-off between the response time and the accuracy of the result. The outcome of the experimental evaluation, using synthetic and real datasets, is the derivation of the outperforming DBQ approximate algorithm for large high-dimensional point datasets.

The time split B-tree (TSBT) is modified for indexing a database of evolving binary images. This is accomplished by embedding ideas from linear region quadtrees that make the TSBT able to support spatio-temporal query processing. To improve query performance, additional pointers are added to the leaf-nodes of the TSBT. The resulting access method is called time split linear quadtree (TSLQ). Algorithms for processing five spatio-temporal queries have been adapted to the new structure. Such queries appear in multimedia systems, or geographical information systems (GIS), when searched by content. The TSLQ was implemented and results of extensive experiments on query time performance are presented, indicating that the proposed algorithmic approaches outbalance respective straightforward algorithms. The region data sets used in the experiments were real images of meteorological satellite views and synthetic raster images.

This paper addresses the problem of finding the K closest pairs between two spatial data sets, where each set is stored in a structure belonging in the R-tree family. Five different algorithms (four recursive and one iterative) are presented for solving this problem. The case of 1 closest pair is treated as a special case. An extensive study, based on experiments performed with synthetic as well as with real point data sets, is presented. A wide range of values for the basic parameters affecting the performance of the algorithms, especially the effect of overlap between the two data sets, is explored. Moreover, an algorithmic as well as an experimental comparison with existing incremental algorithms addressing the same problem is presented. In most settings, the new algorithms proposed clearly outperform the existing ones.

In this paper, indexing in spatio-temporal databases by using the technique of overlapping is investigated. Overlapping has been previously applied in various access methods to combine consecutive structure instances into a single structure, without storing identical sub-structures. In this way, space is saved without sacrificing time performance. A new access method, overlapping linear quadtrees is introduced. This structure is able to store consecutive historical raster images, a database of evolving images. Moreover, it can be used to support query processing in such a database. Five such spatio-temporal queries along with the respective algorithms that take advantage of the properties of the new structure are introduced. The new access method was implemented and extensive experimental studies for space efficiency and query processing performance were conducted. A number of results of these experiments are presented. As far as space is concerned, these results indicate that, in the case of similar consecutive images, considerable storage is saved in comparison to independent linear quadtrees. In the case of query processing, the results indicate that the proposed algorithmic approaches outperform the respective straightforward algorithms, in most cases. The region data sets used in experiments were real images of meteorological satellite views and synthetic random images with specified aggregation.

The S-tree is a dynamic height-balanced tree similar in structure to B+trees. S-trees store fixed length bit-strings, which are called signatures. Signatures are used for indexing textbases, relational, object oriented and extensible databases as well as in data mining. In this article, methods of designing multi-disk B-trees are adapted to S-trees and new methods of parallelizing S-trees are developed. The resulting structures aim at achieving performance gain by accessing two or more disks simultaneously. In addition, two different searching techniques that exploit parallel disk accessing are devised. Performance results of experiments based on the new structures and searching techniques are also presented and discussed.

Traditionalrandom models for spatial two-dimensional data proposed in literature show their limits in generating in a satisfactory way instances of regions having a desired aggregation level.This is because none of them is really oriented to this aim.Rather, they are thought to model the behaviour of the constituting elements of the spatial data (so loosing sight of the context), or, alternatively, to model particular data structure for their representation, underestimating the fact that there is in general no semantic link between a region data and its representation.This means from one hand, the impossibility to produce meaningful the oretical results on time and space average performances of different data structures used to represent spatial regions, and, on the other hand, in an applicative context, the difficulty to generate instances of spatial regions having a statistical behaviour close to that of real data.To overcome this trouble, we introduce in our paper a new random model that provides the possibility to generate spatial regions having a desired aggregation.