Image aesthetics assessment (IAA) has been traditionally addressed as a supervised learning problem, where the goal is to accurately predict information related to user opinions, such as the mean opinion score, image ratings, or a binary quality label, usually crafted by using a mean score threshold to label images as highly or lowly aesthetic. Supervised approaches fail to take into account the subjectiveness of this problem, as the idea of aesthetic pleasantness varies among different people and different cultures, thus making the labels extremely noisy. However, the existence of worldwide photographic contests, exhibitions and masters implies that, to a reasonable degree, there is a broader consensus about the quality of very high-quality images and photographs. Furthermore, labelling image data for IAA is a difficult process, as a large amount of non-trivial aesthetic judgements are required for obtaining a large-scale IAA dataset. Therefore, in this work we analyse the potential of positive-unlabelled techniques for solving IAA. We propose techniques for building PU datasets from traditional IAA datasets and from available reference datasets of high-quality images, and test several well-known PU algorithms on these. Our results highlight the potential of PU approaches for IAA, as we obtain results close to the state-of-the-art with much smaller sets of labelled data: in experiments with only 5% of labelled in AVA, we reach accuracy levels only 0.03 points below NIMA, and we reach competent balanced accuracy levels in settings with a very limited amount of labelled data and with very simple models. • We study solving image aesthetic assessment as a positive unlabelled problem. • We make positive unlabelled datasets from known image aesthetic assessment datasets. • We get a balanced accuracy 0.03 points below NIMA with 5% of labelled images in AVA.

Aesthetic assessment evaluates the quality of a given image using subjective annotations, commonly user ratings, as a knowledge base. Rating complexity is usually relaxed in state-of-the-art works by employing a binary high/low quality label computed from the mean value of rating votes. Nevertheless, this approach introduces uncertainty to average-quality images, which may affect the performance of machine learning models trained from annotated data. In this work, we present a novel approach to aesthetic assessment based on redefining the rating-based groundtruths present in most datasets. Our intent is twofold: to reduce the rating uncertainty and to automatically group them into clusters reflecting high and low quality patterns, thus avoiding an arbitrary threshold like 5 in 1–10 ratings. The experimentation uses the well-known AVA dataset, which consists of more than 255,000 images, and we train several CNN models to test our new groundtruths against the baseline ones. The results show that our approach achieves significant performance gains, between 3% and 9% more balanced accuracy than the baseline groundtruths.

Aesthetic classification of images is an inherently subjective task. There does not exist a validated collection of images/photographs labeled as having good or bad quality from experts. Nowadays, the closest approximation to that is to use databases of photos where a group of users rate each image. Hence, there is not a unique good/bad label but a rating distribution given by users voting. Due to this peculiarity, it is not possible to state the problem of binary aesthetic supervised classification in such a direct mode as other Computer Vision tasks. Recent literature follows an approach where researchers utilize the average rates from the users for each image, and they establish an arbitrary threshold to determine their class or label. In this way, images above the threshold are considered of good quality, while images below the threshold are seen as bad quality. This paper analyzes current literature, and it reviews those attributes able to represent an image, differentiating into three families: specific, general and deep features. Among those which have been proved more competitive, we have selected a representative subset, being our main goal to establish a clear experimental framework. Finally, once features were selected, we have used them for the full AVA dataset. We have to remark that to perform validation we report not only accuracy values, which is not that informative in this case, but also, metrics able to evaluate classification power within imbalanced datasets. We have conducted a series of experiments so that distinct well-known classifiers are learned from data. Like that, this paper provides what we could consider valuable and valid baseline results for the given problem.

There exist a wide number of works in the literature related to new systems devoted to manage thermal control in buildings. Commonly, their evaluation is performed by using simulation of users and environmental conditions. Thus, in this work we choose a successful thermal-comfort system, formerly evaluated with simulations, and evaluate it by using data from project ASHRAE RP-884, which provides logs of real data coming from different buildings, in a wide variety of climates, and occupied by people with different thermal preferences. From these logs, we propose a pre-processing and evaluation methodology in order to achieve more realistic evaluations.

The Averaged One-Dependence Estimators (AODE) classifier is one of the most attractive semi-naive Bayesian classifiers and hence a good alternative to Naive Bayes (NB), as it obtains fairly low error rates maintaining under control the computational complexity. Unfortunately, as most of the methods designed within the framework of Bayesian networks, AODE is exclusively defined to deal with discrete variables. Several approaches to avoid the use of discretization pre-processing techniques have already been presented, all of them involving in lower or greater degree the assumption of (conditional) Gaussian distributions. In this paper, we propose the use of Mixture of Truncated Exponentials (MTEs), whose expressive power to accurately approximate the most commonly used distributions for hybrid networks has already been demonstrated. We perform experiments on the use of MTEs over a large group of datasets for the first time, and we analyze the importance of selecting a proper number of points when learning MTEs for NB and AODE, as we believe, it is decisive to provide accurate results.

When a Bayesian network (BN) is modified, for example adding or deleting a node, or changing the probability distributions, we usually will need a total recompilation of the model, despite feeling that a partial (re)compilation could have been enough. Especially when considering dynamic models, in which variables are added and removed very frequently, these recompilations are quite resource consuming. But even further, for the task of building a model, which is in many occasions an iterative process, there is a clear lack of flexibility. When we use the term Incremental Compilation or IC we refer to the possibility of modifying a network and avoiding a complete recompilation to obtain the new (and different) join tree (JT). The main point we intend to study in this work is JT-based inference in Bayesian networks. Apart from undertaking the triangulation problem itself, we have achieved a great improvement for the compilation in BNs. We do not develop a new architecture for BNs inference, but taking some already existing framework JT-based for probability propagation such as Hugin or Shenoy and Shafer, we have designed a method that can be successfully applied to get better performance, as the experimental evaluation will show.

AODE (Aggregating One-Dependence Estimators) is considered one of the most interesting representatives of the Bayesian classifiers, taking into account not only the low error rate it provides but also its efficiency. Until now, all the attributes in a dataset have had to be nominal to build an AODE classifier or they have had to be previously discretized. In this paper, we propose two different approaches in order to deal directly with numeric attributes. One of them uses conditional Gaussian networks to model a dataset exclusively with numeric attributes; and the other one keeps the superparent on each model discrete and uses univariate Gaussians to estimate the probabilities for the numeric attributes and multinomial distributions for the categorical ones, it also being able to model hybrid datasets. Both of them obtain competitive results compared to AODE, the latter in particular being a very attractive alternative to AODE in numeric datasets.

Triangulation of Bayesian networks (BNs) is an NP-hard problem and of great importance for the efficiency of propagation algorithms. Several approaches, most of them basically heuristic, have been proposed to search optimal solutions for this problem. Recently, Olesen and Madsen1 launched the possibility of applying the maximal prime subgraph decomposition (MPSD) to the problem of triangulation. The idea is to retriangulate separately each MPS, so that we can work on smaller graphs. In this article, we exploit this idea by using both greedy heuristic algorithms and stochastic ones [genetic algorithms (GAs)]. From an experimentation performed over 10 real complex networks, we study empirically the usefulness of applying this MPSD-based retriangulation. © 2003 Wiley Periodicals, Inc.