Konstantinos Pliakos

Making sense of big data in health research: Towards an EU action plan

Medicine and healthcare are undergoing profound changes. Whole-genome sequencing and high-resolution imaging technologies are key drivers of this rapid and crucial transformation. Technological innovation combined with automation and miniaturization has triggered an explosion in data production that will soon reach exabyte proportions. How are we going to deal with this exponential increase in data production? The potential of “big data” for improving health is enormous but, at the same time, we face a wide range of challenges to overcome urgently. Europe is very proud of its cultural diversity; however, exploitation of the data made available through advances in genomic medicine, imaging, and a wide range of mobile health applications or connected devices is hampered by numerous historical, technical, legal, and political barriers. European health systems and databases are diverse and fragmented. There is a lack of harmonization of data formats, processing, analysis, and data transfer, which leads to incompatibilities and lost opportunities. Legal frameworks for data sharing are evolving. Clinicians, researchers, and citizens need improved methods, tools, and training to generate, analyze, and query data effectively. Addressing these barriers will contribute to creating the European Single Market for health, which will improve health and healthcare for all Europeans.

SIMULTANEOUS IMAGE TAGGING AND GEO-LOCATION PREDICTION WITHIN HYPERGRAPH RANKING FRAMEWORK

The development of social media has led to a burst of interest in image-related metadata information, such as tags and geo-tags. Tags are semantic keywords that are assigned to an image. Image tagging enables the users of social media sharing platforms to annotate images, facilitating image search and content description. Despite the volume of related research, issues such as accuracy or efficiency still remain open problems. Here, a novel method for simultaneous image tagging and geo-location prediction is proposed that is based on hypergraph learning. The method is further improved by enforcing group sparsity constraints. It fully exploits various types of information, such as social, image-related metadata, or similarities based on visual attributes. Experiments on a dataset crawled from Flickr demonstrate F1 at 10 top ranked tags equal to 0.558 for image tagging and cumulative geotagging prediction rate at 3 top ranks equal to 83%.

PLSA driven image annotation, classification, and tourism recommendation

A burst of interest in image annotation and recommendation has been witnessed. Despite the huge effort made by the scientific community in the aforementioned research areas, accuracy or efficiency still remain open problems. Here, efficient methods for image annotation, visual image content classification as well as touristic place of interest (POI) recommendation are developed within the same framework. In particular, semantic image annotation and touristic POI recommendation harness the geo-information associated to images. Both semantic image annotation and visual image content classification resort to Probabilistic Latent Semantic Analysis (PLSA). Several tourist destinations, strongly related to the query image, are recommended, using hypergraph ranking. Experimental results were conducted on a large image dataset of Greek sites, demonstrating the potential of the proposed methods. Semantic image annotation by means of PLSA has achieved an average precision of 90% at 10% recall. The average accuracy of content-based image classification is 80%. An average precision of 90% is measured at 1% recall for tourism recommendation.

Weight estimation in hypergraph learning

The unremitting rising popularity of social media has led to an exponential increase in web activity as manifested by the vast volume of uploaded images. This boundless volume of image data has triggered the interest in image tagging. Here, an efficient hypergraph weight estimation scheme is proposed that improves the accuracy of image tagging, using hypergraph learning. The proposed method models high-order relations between hypergraph vertices (i.e., users, user social groups, tags, geo-tags, and images) by hyperedges. The information captured by the hyperedges is efficiently distilled by estimating the hyperedge weights. Experiments conducted on a dataset crawled from Flickr demonstrate the effectiveness of the proposed approach. Specifically, an average precision of 91% at 26% recall has been achieved for image tagging.

Personalized and geo-referenced image recommendation using unified hypergraph learning and group sparsity optimization

The rapid development of social media has led to a surge of interest in multimedia recommendation. Several recommender systems have been developed, but achieving a satisfactory efficiency or accuracy still remains an open problem. In this paper, a novel multi-reference image recommendation system is proposed based on a unified hypergraph. Relevant images from a large pool are recommended to a reference user or a reference geo-location. In addition to that, the hypergraph ranking problem is enhanced by enforcing group sparsity constraints. By adjusting the different weights associated to the object groups, we control each object group effect in the recommendation process. Experiments on a dataset crawled from Flickr demonstrate the merits of the proposed method.

Network representation with clustering tree features

Representing and inferring interaction networks is a challenging and long-standing problem. Modern technological advances have led to a great increase in both volume and complexity of generated network data. The size of networks such as drug protein interaction networks or gene regulatory networks is constantly growing and multiple sources of information are exploited to extract features describing the nodes in such networks. Modern information systems need therefore methods that are able to mine these networks and exploit the available features. Here, a novel data mining framework for network representation and mining is proposed. It is based on decision tree learning and ensembles of trees. The proposed scheme introduces an efficient network data representation, capable of addressing different data types, tackling as well data volume and complexity. The learning process follows the inductive setup and it can be performed in both a supervised or unsupervised manner. Experiments were conducted on six biomedical network datasets. The experimental evaluation demonstrates the merits of the proposed approach, confirming its efficiency.

Adaptive algorithms for hypergraph learning

Social media sharing platforms enable image content as well as context information (e.g., user friendships, geo-tags assigned to images) to be jointly analyzed in order to achieve accurate image annotation or successful image recommendation. The context information is expressed frequently in terms of high-order relations, such as the relations among users, tags, and images. Hypergraphs can model the aforementioned high-order relations between their vertices (i.e., users, user social groups, tags, geo-tags, and images) by hyper-edges, whose influence can be assessed by properly estimating their weights. Here, an efficient adaptive hypergraph weight estimation is proposed for image tagging. In particular, both equality and inequality constraints enforced during hypergraph learning are taken into account and an efficient adaptation step selection using the Armijo rule is proposed. Experiments conducted on a dataset demonstrate the superior performance of the proposed approach compared to the state-of-the-art.

Feature Induction and Network Mining with Clustering Tree Ensembles

The volume of data generated and collected using modern technologies grows exponentially. This vast amount of data often follows a complex structure, and the problem of efficiently mining and analyzing such data is crucial for the performance of various machine learning tasks. Here, a novel data mining framework for unsupervised learning tasks is proposed based on decision tree learning and ensembles of trees. The proposed approach introduces an informative feature representation and is able to handle data diversity and complexity. Moreover, a new scheme is proposed based on the aforementioned approach for mining interaction data. These data are often modeled as homogeneous or heterogeneous networks and they are present in various fields, such as social media, recommender systems, and bioinformatics. The learning process is performed in an unsupervised manner, following also the inductive setup. The experimental evaluation confirms the effectiveness of the proposed approach.

Personalized music tagging using ranking on hypergraphs

Social tagging enables users of social media sharing platforms to annotate multimedia items by employing arbitrary keywords (i.e., tags), which describe better the multimedia content. Several applications, such as personalized multimedia recommendation or music genre classification, to name a few, benefit from tagging. Clearly, tagging aims at bridging the semantic gap between human concepts and content retrieval exploiting low-level features extracted from the multimedia. Here, the problem of personalized tag recommendation is addressed in a “query and ranking” manner on hypergraphs. This way, the relationships between the different object types, such as user friendships, user groups, music tracks and tags are captured and tags are recommended for a certain track to a user. Ranking on hypergraphs is studied by enforcing either ℓ2 norm regularization or group sparsity. Experiments on a dataset collected from Last.fm demonstrate a promising tag recommendation accuracy.

Simultaneous Image Clustering, Classification and Annotation for Tourism Recommendation

The exponential increase in the amount of data uploaded to the web has led to a surge of interest in multimedia recommendation and annotation. Due to the vast volume of data, efficient algorithms for recommendation and annotation are needed. Here, a novel two-step approach is proposed, which annotates an image received as input and recommends several tourist destinations strongly related to the image. It is based on probabilistic latent semantic analysis and hypergraph ranking enhanced with the visual attributes of the images. The proposed method is tested on a dataset of 30000 images bearing text information (e.g., title, tags) collected from Flickr. The experimental results are very promising, as they achieve a top rank precision of 80% for tourism recommendation.