Andrej Košir

Using affective parameters in a content-based recommender system for images

There is an increasing amount of multimedia content available to end users. Recommender systems help these end users by selecting a small but relevant subset of items for each user based on her/his preferences. This paper investigates the influence of affective metadata (metadata that describe the user’s emotions) on the performance of a content-based recommender (CBR) system for images. The underlying assumption is that affective parameters are more closely related to the user’s experience than generic metadata (e.g. genre) and are thus more suitable for separating the relevant items from the non-relevant. We propose a novel affective modeling approach based on users’ emotive responses. We performed a user-interaction session and compared the performance of the recommender system with affective versus generic metadata. The results of the statistical analysis showed that the proposed affective parameters yield a significant improvement in the performance of the recommender system.

Unsupervised quantification of whisking and head movement in freely moving rodents

The rodent whisker system has become the leading experimental paradigm for the study of active sensing. Thanks to more sophisticated behavioral paradigms, progressively better neurophysiological methods, and improved video hardware/software, there is now the prospect of defining the precise connection between the sensory apparatus and brain activity in awake, exploring animals. Achieving this ambitious goal requires quantitative, objective characterization of head and whisker kinematics. This study presents the methodology and potential uses of a new automated motion analysis routine. The program provides full quantification of head orientation and translation, as well as the angle, frequency, amplitude, and bilateral symmetry of whisking. The system operates without any need for manual tracing by the user. Quantitative comparison to whisker detection by expert humans indicates that the programs correct detection rate is at >95% even on animals with all whiskers intact. Particular attention has been paid to obtaining reliable performance under nonoptimal lighting or video conditions and at frame rates as low as 100. Variation of the zoom across time is compensated for without user intervention. The program adapts automatically to the size and shape of different species. The outcome of our testing indicates that the program can be a valuable tool in quantifying rodent sensorimotor behavior.

Predicting and Detecting the Relevant Contextual Information in a Movie-Recommender System

Context-aware recommender system (CARS) is a highly researched and implemented way of providing a personalized service that helps users to find their desired content. One of the remaining issues is how to decide which contextual information to acquire and how to incorporate it into CARS. While the relevant contextual information will improve the recommendations, the irrelevant contextual information could have a negative impact on the recommendation accuracy. By testing the independence between the contextual variable on the users’ ratings for items, we can detect its relevanceandimpactonthefeedbackfortheitemconsumedinthatspecificcontext.Inthisarticle,we propose several new theoretical concepts that should help deciding which information to use, as well as a methodology for detecting which contextual information contributes to explaining the variance in the ratings, based on statistical testing. The experiment was conducted on the real movie dataset that contains 12 different pieces of contextual information. We used two statistical tests with power analysis for the detection, and three contextualized matrix-factorization algorithms with slightly different reasoning for the prediction of ratings. The results showed a significant difference in the prediction of ratings in the context that was detected as relevant by our method, and the one that was detected as irrelevant, pointing to the importance of the power analysis and the benefits of the proposed method in the case of a small dataset.

Personal Content Recommender Based on a Hierarchical User Model for the Selection of TV Programmes

In this paper we present our approach to user modeling for a personalized selection of multimedia content tested on a corpus of TV programmes. The idea of this approach is to classify content (TV programmes) based on the calculation of similarities between the description of content and the user model for each description attribute. Calculated similarities are then combined into a classification decision using the Support Vector Machines. The basis for the calculation of similarities is a hierarchical structure of the user model, overlaid upon a taxonomy of TV programme genres. Preliminary results show that it works well with a varying quality of content descriptions including incomplete genre classification and arbitrary number of description attributes. The evaluation of the system performance was based on content described using the TV-Anytime standard, but the approach can be adapted for search of other types of content with multi-attribute descriptions.

The music perception abilities of prelingually deaf children with cochlear implants

OBJECTIVE To investigate the music perception abilities of prelingually deaf children with cochlear implants, in comparison to a group of normal-hearing children, and to consider the factors that contribute to music perception. METHODS The music perception abilities of 39 prelingually deaf children with unilateral cochlear implants were compared to the abilities of 39 normal hearing children. To assess the music listening abilities, the MuSIC perception test was adopted. The influence of the childs age, age at implantation, device experience and type of sound-processing strategy on the music perception were evaluated. The effects of auditory performance, nonverbal intellectual abilities, as well as the childs additional musical education on music perception were also considered. RESULTS Children with cochlear implants and normal hearing children performed significantly differently with respect to rhythm discrimination (55% vs. 82%, p<0.001), instrument identification (57% vs. 88%, p<0.001) and emotion rating (p=0.022). However we found no significant difference in terms of melody discrimination and dissonance rating between the two groups. There was a positive correlation between auditory performance and melody discrimination (r=0.27; p=0.031), between auditory performance and instrument identification (r=0.20; p=0.059) and between the childs grade (mark) in school music classes and melody discrimination (r=0.34; p=0.030). In children with cochlear implant only, the music perception ability assessed by the emotion rating test was negatively correlated to the childs age (r(S)=-0.38; p=0.001), age at implantation (r(S)=-0.34; p=0.032), and device experience (r(S)=-0.38; p=0.019). The childs grade in school music classes showed a positive correlation to music perception abilities assessed by rhythm discrimination test (r(S)=0.46; p<0.001), melody discrimination test (r(S)=0.28; p=0.018), and instrument identification test (r(S)=0.23; p=0.05). CONCLUSIONS As expected, there was a marked difference in the music perception abilities of prelingually deaf children with cochlear implants in comparison to the group of normal hearing children, but not for all the tests of music perception. Additional multi-centre studies, including a larger number of participants and a broader spectrum of music subtests, considering as many as possible of the factors that may contribute to music perception, seem reasonable.

The LDOS-PerAff-1 corpus of facial-expression video clips with affective, personality and user-interaction metadata

We present the LDOS-PerAff-1 Corpus that bridges the affective computing and recommender system research areas, which makes it unique. The corpus is composed of video clips of subjects’ affective responses to visual stimuli. These affective responses are annotated in the continuous valence-arousal-dominance space. Furthermore, the subjects are annotated with their personality information using the five-factor personality model. We also provide the explicit ratings that the users gave to the images used for the visual stimuli. In the paper we present the results of four experiments conducted with the corpus; an affective content-based recommender system, a personality-based collaborative filtering recommender system, an emotion-detection algorithm and a qualitative study of the latent factors.

The impact of weak ground truth and facial expressiveness on affect detection accuracy from time-continuous videos of facial expressions

In this paper we address two issues concerning real-world time-continuous emotion detection from videos of users’ faces: (i) the impact of weak ground truth on the emotion detection accuracy and (ii) the impact of the users’ facial expressiveness on the emotion detection accuracy. We implemented an appearance-based emotion detection algorithm that uses Gabor features and a k nearest neighbors classifier. We tested the performance of this algorithm on two datasets with different ground truth strengths (a firm ground truth dataset and a weak ground truth dataset). Then we split the dataset into three subsets reflecting different levels of users’ facial expressiveness (low, mid and high) and performed separate emotion detection.

The hazards of DAPI photoconversion: effects of dye, mounting media and fixative, and how to minimize the problem

Immunocytochemistry is a powerful tool for detection and visualization of specific molecules in living or fixed cells, their localization and their relative abundance. One of the most commonly used fluorescent DNA dyes in immunocytochemistry applications is 4′,6-diamidino-2-phenylindole dihydrochloride, known as DAPI. DAPI binds strongly to DNA and is used extensively for visualizing cell nuclei. It is excited by UV light and emits characteristic blue fluorescence. Here, we report a phenomenon based on an apparent photoconversion of DAPI that results in detection of a DAPI signal using a standard filter set for detection of green emission due to blue excitation. When a sample stained with DAPI only was first imaged with the green filter set (FITC/GFP), only a weak cytoplasmic autofluorescence was observed. Next, we imaged the sample with a DAPI filter set, obtaining a strong nuclear DAPI signal as expected. Upon reimaging the same samples with a FITC/GFP filter set, robust nuclear fluorescence was observed. We conclude that excitation with UV results in a photoconversion of DAPI that leads to detection of DAPI due to excitation and emission in the FITC/GFP channel. This phenomenon can affect data interpretation and lead to false-positive results when used together with fluorochrome-labeled nuclear proteins detected with blue excitation and green emission. In order to avoid misinterpretations, extra precaution should be taken to prepare staining solutions with low DAPI concentration and DAPI (UV excitation) images should be acquired after all other higher wavelength images. Of various DNA dyes tested, Hoechst 33342 exhibited the lowest photoconversion while that for DAPI and Hoechst 33258 was much stronger. Different fixation methods did not substantially affect the strength of photoconversion. We also suggest avoiding the use of mounting medium with high glycerol concentrations since glycerol showed the strongest impact on photoconversion. This photoconversion effect cannot be avoided even when using narrow bandpass filter sets.

Radio Environment Maps: The Survey of Construction Methods

Radio environment maps (REMs) and geolocation database represent an important source of information for the operation of cognitive radio networks, replacing or complementing spectrum sensing information. This paper provides a survey of methods for constructing the radio frequency layer of radio environment map (RF-REM) using distributed measurements of the signal levels at a given frequency in space and time. The signal level measurements can be obtained from fixed or mobile devices capable of sensing radio environment and sending this information to the REM. The signal measurements are complemented with information already stored in different REM content layers. The combined information is applied for estimation of the RF-REM layer. The RF-REM construction methods are compared, and their advantages and disadvantages with respect to the spatial distribution of signal measurements and computational complexity is given. This survey also indicates possible directions of further research in indirect RF-REM construction methods. It emphasizes that accurate RF-REM construction methods should in the best case support operation with random and clustered signal measurements, their operation should not be affected by measurements outliers, and it must estimate signal levels comparably on all RF-REM locations with moderate computational effort.

Emotion-Aware Recommender Systems – A Framework and a Case Study

Recent work has shown an increase of accuracy in recommender systems that use emotive labels. In this paper we propose a framework for emotion-aware recommender systems and present a survey of the results in such recommender systems. We present a consumption-chain-based framework and we compare three labeling methods within a recommender system for images: (i) generic labeling, (ii) explicit affective labeling and (iii) implicit affective labeling.