Ladislav Peska

Enhancing Recommender System with Linked Open Data

In this paper, we present an innovative method to use Linked Open Data LOD to improve content based recommender systems. We have selected the domain of secondhand bookshops, where recommending is extraordinary difficult because of high ratio of objects/users, lack of significant attributes and small number of the same items in stock. Those difficulties prevents us from successfully apply both collaborative and common content based recommenders. We have queried Czech language mutation of DBPedia in order to receive additional attributes of objects books to reveal nontrivial connections between them. Our approach is general and can be applied on other domains as well. Experiments show that enhancing recommender system with LOD can significantly improve its results in terms of object similarity computation and top-k objects recommendation. The main drawback hindering widespread of such systems is probably missing data about considerable portion of objects, which can however vary across domains and improve over time.

Recommending for Disloyal Customers with Low Consumption Rate

In this paper, we focus on small or medium-sized e-commerce portals. Due to high competition, users of these portals are not too loyal and e.g. refuse to register or provide any/enough explicit feedback. Furthermore, products such as tours, cars or furniture have very low average consumption rate preventing us from tracking unregistered user between two consecutive purchases. Recommending on such domains proves to be very challenging, yet interesting research task. For this task, we propose a model coupling various implicit feedbacks and object attributes in matrix factorization. We report on promising results of our initial off-line experiments on travel agency dataset. Our experiments corroborate benefits of using object attributes; however we are yet to decide about usefulness of some implicit feedback data.

Hybrid Recommending Exploiting Multiple DBPedia Language Editions

In this paper we describe approach of our SemWex1 group to the ESWC 2014 RecSys Challenge. Our method is based on using an adaptation of Content Boosted Matrix factorization [1], where objects are defined through their content-based features. Features were comprised of both direct DBPedia RDF triples and derived semantic information (with some WIE and NLP features). Total of seven DBPedia language editions were used to form the dataset. In the paper we will further describe our methods for semantic information creation, data filtration, algorithm details and settings as well as decisions made during the challenge and dead ends we explored.

UPComp - A PHP Component for Recommendation Based on User Behaviour

In this paper, we investigate the possibilities of interpreting user behaviour in order to learn his/her preferences. UP Comp, a PHP component enabling use of user preferences for recommendation, is described. UP Comp is a standalone component that can be integrated into any PHP web with only basic knowledge of PHP, HTML and SQL. The methods of user behaviour interpretation are evaluated on a real web shop with tourist trips using UP Comp.

Negative implicit feedback in e-commerce recommender systems

In this paper, we imagine the situation of a typical e-commerce portal employing personalized recommendation. Such website typically receives user feedback from their implicit behavior such as time on page, scrolling etc. The implicit feedback is generally understood as positive only, however we present several methods how to identify some of the implicit feedback as negative user preference, how to aggregate various feedback types together and how to recommend based on it. We have conducted several off-line experiments with real user data from travel agency website confirming that treating some implicit feedback as negative preference can significantly improve recommendation quality.

Using Implicit Preference Relations to Improve Recommender Systems

Our work is generally focused on making recommendations for small or medium-sized e-commerce portals, where we are facing scarcity of explicit feedback, low user loyalty, short visit durations or a low number of visited objects. In this paper, we present a novel approach to use a specific user behavior pattern as implicit feedback, forming binary relations between objects. Our hypothesis is that if a user selects a specific object from the list of displayed objects, it is an expression of his/her binary preference between the selected object and others that are visible, but ignored. We expand this relation with content-based similarity of objects. We define implicit preference relation (IPR) a partial ordering of objects based on similarity expansion of ignored-selected preference relation. We propose a merging algorithm utilizing the synergic effect of two approaches this IPR partial ordering and a list of recommended objects based on any/another algorithm. We report on a series of offline experiments with various recommending algorithms on two real-world e-commerce datasets. The merging algorithm could improve the ranked list of most of the evaluated algorithms in terms of nDCG. Furthermore, we also provide access to the relevant datasets and source codes for further research.

Classification of fMRI data using dynamic time warping based functional connectivity analysis

The synchronized spontaneous low frequency fluctuations of the BOLD signal, as captured by functional MRI measurements, is known to represent the functional connections of different brain areas. The aforementioned MRI measurements result in high-dimensional time series, the dimensions of which correspond to the activity of different brain regions. Recently we have shown that Dynamic Time Warping (DTW) distance can be used as a similarity measure between BOLD signals of brain regions as an alternative of the traditionally used correlation coefficient. We have characterized the new metrics stability in multiple measurements, and between subjects in homogenous groups. In this paper we investigated the DTW metrics sensitivity and demonstrated that DTW-based models outperform correlation-based models in resting-state fMRI data classification tasks. Additionally, we show that functional connectivity networks resulting from DTW-based models as compared to the correlation-based models are more stable and sensitive to differences between healthy subjects and patient groups.

Towards Recommender Systems for Police Photo Lineup

Photo lineups play a significant role in the eyewitness identification process. This method is used to provide evidence in the prosecution and subsequent conviction of suspects. Unfortunately, there are many cases where lineups have led to the conviction of an innocent suspect. One of the key factors affecting the incorrect identification of a suspect is the lack of lineup fairness, i.e. that the suspect differs significantly from all other candidates. Although the process of assembling fair lineup is both highly important and time-consuming, only a handful of tools are available to simplify the task. In this paper, we describe our work towards using recommender systems for the photo lineup assembling task. We propose and evaluate two complementary methods for item-based recommendation: one based on the visual descriptors of the deep neural network, the other based on the content-based attributes of persons. The initial evaluation made by forensic technicians shows that although results favored visual descriptors over attribute-based similarity, both approaches are functional and highly diverse in terms of recommended objects. Thus, future work should involve incorporating both approaches in a single prediction method, preference learning based on the feedback from forensic technicians and recommendation of assembled lineups instead of single candidates.

Drug-target interaction prediction: A Bayesian ranking approach

BACKGROUND AND OBJECTIVE In silico prediction of drug-target interactions (DTI) could provide valuable information and speed-up the process of drug repositioning - finding novel usage for existing drugs. In our work, we focus on machine learning algorithms supporting drug-centric repositioning approach, which aims to find novel usage for existing or abandoned drugs. We aim at proposing a per-drug ranking-based method, which reflects the needs of drug-centric repositioning research better than conventional drug-target prediction approaches. METHODS We propose Bayesian Ranking Prediction of Drug-Target Interactions (BRDTI). The method is based on Bayesian Personalized Ranking matrix factorization (BPR) which has been shown to be an excellent approach for various preference learning tasks, however, it has not been used for DTI prediction previously. In order to successfully deal with DTI challenges, we extended BPR by proposing: (i) the incorporation of target bias, (ii) a technique to handle new drugs and (iii) content alignment to take structural similarities of drugs and targets into account. RESULTS Evaluation on five benchmark datasets shows that BRDTI outperforms several state-of-the-art approaches in terms of per-drug nDCG and AUC. BRDTI results w.r.t. nDCG are 0.929, 0.953, 0.948, 0.897 and 0.690 for G-Protein Coupled Receptors (GPCR), Ion Channels (IC), Nuclear Receptors (NR), Enzymes (E) and Kinase (K) datasets respectively. Additionally, BRDTI significantly outperformed other methods (BLM-NII, WNN-GIP, NetLapRLS and CMF) w.r.t. nDCG in 17 out of 20 cases. Furthermore, BRDTI was also shown to be able to predict novel drug-target interactions not contained in the original datasets. The average recall at top-10 predicted targets for each drug was 0.762, 0.560, 1.000 and 0.404 for GPCR, IC, NR, and E datasets respectively. CONCLUSIONS Based on the evaluation, we can conclude that BRDTI is an appropriate choice for researchers looking for an in silico DTI prediction technique to be used in drug-centric repositioning scenarios. BRDTI Software and supplementary materials are available online at www.ksi.mff.cuni.cz/∼peska/BRDTI.

Drug–target interaction prediction with Bipartite Local Models and hubness-aware regression

Abstract Computational prediction of drug–target interactions is an essential task with various applications in the pharmaceutical industry, such as adverse effect prediction or drug repositioning. Recently, expert systems based on machine learning have been applied to drug–target interaction prediction. Although hubness-aware machine learning techniques are among the most promising approaches, their potential to enhance drug–target interaction prediction methods has not been exploited yet. In this paper, we extend the Bipartite Local Model (BLM), one of the most prominent interaction prediction methods. In particular, we use BLM with a hubness-aware regression technique, ECkNN. We represent drugs and targets in the similarity space with rich set of features (i.e., chemical, genomic and interaction features), and build a projection-based ensemble of BLMs. In order to assist reproducibility of our work as well as comparison to published results, we perform experiments on widely used publicly available drug–target interaction datasets. The results show that our approach outperforms state-of-the-art drug–target prediction techniques. Additionally, we demonstrate the feasibility of predictions from the point of view of applications.