International Journal of Semantic Computing

The performance and efficiency of A* search algorithm heavily depends on the quality of the heuristic function. Therefore, designing an optimal heuristic function becomes the primary goal of developing a search algorithm for specific domains in artificial intelligence. However, it is difficult to design a well-constructed heuristic function without careful consideration and trial-and-error, especially for complex pathfinding problems. The complexity of a heuristic function increases and becomes unmanageable to design when an increasing number of parameters are involved. Existing approaches often avoid complex heuristic function design: they either trade-off the accuracy for faster computation or taking advantage of the parallelism for better performance. The objective of this paper is to reduce the difficulty of complex heuristic function design for A* search algorithm. We aim to design an algorithm that can be automatically optimized to achieve rapid search with high accuracy and low computational cost. In this paper, we present a novel design and optimization method for a Multi-Weighted-Heuristics function (MWH) named Evolutionary Heuristic A* search (EHA*) to: (1) minimize the effort on heuristic function design via Genetic Algorithm (GA), (2) optimize the performance of A* search and its variants including but not limited to WA* and MHA*, and (3) guarantee the completeness and optimality. EHA* algorithm enables high performance searches and significantly simplifies the processing of heuristic design. We apply EHA* to multiple grid-based pathfinding benchmarks to evaluate the performance. Our experiment result shows that EHA* (1) is capable of choosing an accurate heuristic function that provides an optimal solution, (2) can identify and eliminate inefficient heuristics, (3) is able to automatically design multi-heuristics function, and (4) minimizes both the time and space complexity.

The comprehensive set of neuronal connections of the human brain, which is known as the human connectomes, has provided valuable insight into neurological and neurodevelopmental disorders. Functional Magnetic Resonance Imaging (fMRI) has facilitated this research by capturing regionally specific brain activity. Resting state fMRI is used to extract the functional connectivity networks, which are edge-weighted complete graphs. In these complete functional connectivity networks, each node represents one brain region or Region of Interest (ROI), and each edge weight represents the strength of functional connectivity of the adjacent ROIs. In order to leverage existing graph mining methodologies, these complete graphs are often made sparse by applying thresholds on weights. This approach can result in loss of discriminative information while addressing the issue of biomarkers detection, i.e. finding discriminative ROIs and connections, given the data of healthy and disabled population. In this work, we demonstrate a novel framework for representing the complete functional connectivity networks in a threshold-free manner and identifying biomarkers by using feature selection algorithms. Additionally, to compute meaningful representations of the discriminative ROIs and connections, we apply tensor decomposition techniques. Experiments on a fMRI dataset of neurodevelopmental reading disabilities show the highly interpretable nature of our approach in finding the biomarkers of the diseases.

As a primary user group, Deaf or Hard of Hearing (D/HOH) audiences use Closed Captioning (CC) service to enjoy the TV programs with audio by reading text. However, the D/HOH communities are not completely satisfied with the quality of CC even though the government regulators entail certain rules in the CC quality factors. The measure of the CC quality is often interpreted as an accuracy on translation and regulators use the empirical models to assess. The need of a subjective quality scale comes from the gap in between current empirical assessment models and the audience perceived quality. It is possible to fill the gap by including the subjective assessment by D/HOH audiences. This research proposes a design of an automatic quality assessment system for CC which can predict the D/HOH audience subjective ratings. A simulated rater is implemented based on literature and the CC quality factor representative value extraction algorithm is developed. Three prediction models are trained with a set of CC quality values and corresponding rating scores, then they are compared to find the feasible prediction model.

Deep learning has a hierarchical network structure to represent multiple features of input data. The adaptive structural learning method of Deep Belief Network (DBN) can reach the high classification capability while searching the optimal network structure during the training. The method can find the optimal number of hidden neurons for given input data in a Restricted Boltzmann Machine (RBM) by neuron generation–annihilation algorithm, and generate a new hidden layer in DBN by the extension of the algorithm. In this paper, the proposed adaptive structural learning of DBN (Adaptive DBN) was applied to the comprehensive medical examination data for cancer prediction. The developed prediction system showed higher classification accuracy for test data (99.5% for the lung cancer and 94.3% for the stomach cancer) than the several learning methods such as traditional RBM, DBN, Non-Linear Support Vector Machine (SVM), and Convolutional Neural Network (CNN). Moreover, the explicit knowledge that makes the inference process of the trained DBN is required in deep learning. The binary patterns of activated neurons for given input in RBM and the hierarchical structure of DBN can represent the relation between input and output signals. These binary patterns were classified by C4.5 for knowledge extraction. Although the extracted knowledge showed slightly lower classification accuracy than the trained DBN network, it was able to improve inference speed by about 1/40. We report that the extracted IF-THEN rules from the trained DBN for medical examination data showed some interesting features related to initial condition of cancer.

Nowadays, Twitter has become one of the fastest-growing microblogging services; consequently, analyzing this rich and continuously user-generated content can reveal unprecedentedly valuable knowledge. In this paper, we propose a novel two-stage system to detect and track events from tweets by integrating a Latent Dirichlet Allocation (LDA)-based approach and an efficient density–contour-based spatio-temporal clustering approach. In the proposed system, we first divide the geotagged tweet stream into temporal time windows; next, events are identified as topics in tweets using an LDA-based topic discovery step; then, each tweet is assigned an event label; next, a density–contour-based spatio-temporal clustering approach is employed to identify spatio-temporal event clusters. In our approach, topic continuity is established by calculating KL-divergences between topics and spatio-temporal continuity is established by a family of newly formulated spatial cluster distance functions. Moreover, the proposed density–contour clustering approach considers two types of densities: “absolute” density and “relative” density to identify event clusters where either there is a high density of event tweets or there is a high percentage of event tweets. We evaluate our approach using real-world data collected from Twitter, and the experimental results show that the proposed system can not only detect and track events effectively but also discover interesting patterns from geotagged tweets.

Twitter is one of the most popular social media platforms used by millions of users daily to post their opinions and emotions. Consequently, Twitter tweets have become a valuable knowledge source for emotion analysis. In this paper, we present a new framework, K2, for tweet emotion mapping and emotion change analysis. It introduces a novel, generic spatio-temporal data analysis and storytelling framework that can be used to understand the emotional evolution of a specific section of population. The input for our framework is the location and time of where and when the tweets were posted and an emotion assessment score in the range $[-1,+1]$, with $+1$ representing a very high positive emotion and $-1$ representing a very high negative emotion. Our framework first segments the input dataset into a number of batches with each batch representing a specific time interval. This time interval can be a week, a month or a day. By generalizing existing kernel density estimation techniques in the next step, we transform each batch into a continuous function that takes positive and negative values. We have used contouring algorithms to find the contiguous regions with highly positive and highly negative emotions belonging to each member of the batch. Finally, we apply a generic, change analysis framework that monitors how positive and negative emotion regions evolve over time. In particular, using this framework, unary and binary change predicate are defined and matched against the identified spatial clusters, and change relationships will then be recorded, for those spatial clusters for which a match occurs. We also propose animation techniques to facilitate spatio-temporal data storytelling based on the obtained spatio-temporal data analysis results. We demo our approach using tweets collected in the state of New York in the month of June 2014.

News videos are valuable multimedia information on real-world events. However, due to the incremental nature of the contents, a sequence of news videos on a related news topic could be redundant and lengthy. Thus, a number of methods have been proposed for their summarization. However, there is a problem that most of these methods do not consider the consistency between the auditory and visual contents. This becomes a problem in the case of news videos, since both contents do not always come from the same source. Considering this, in this paper, we propose a method for summarizing a sequence of news videos considering the consistency of auditory and visual contents. The proposed method first selects key-sentences from the auditory contents (Closed Caption) of each news story in the sequence, and next selects a shot in the news story whose “Visual Concepts” detected from the visual contents are the most consistent with the selected key-sentence. In the end, the audio segment corresponding to each key-sentence is synthesized with the selected shot, and then these clips are concatenated into a summarized video. Results from subjective experiments on summarized videos on several news topics show the effectiveness of the proposed method.