This study examines the difference in people’s perceptions of an embodied and a nonembodied conversational agent. It also discusses the potential use of embodied AI in the telehealth industry. A GPT assistant was created with the instruction to act as an empathetic doctor, and an embodied conversation agent system was designed for it to use in interactions. Research participants interacted with both embodied and nonembodied versions of the assistant and recorded their experience through a survey. Results showed that participants had higher standards for an embodied AI and desired a more human-like interaction. If these expectations are not met, their preference shifts in favor of nonembodied AI. Users would be comfortable using embodied AI in telehealth settings, however, a preference still exists for actual doctors. Measures would also have to be taken to ensure their medical information is kept secure.

Topic evolution is a recently introduced field of research as a substitute for a more traditional text-based topic evolution, allowing the tracking of more complex evolutionary events with the use of network structures. Network-based topic evolution showed that the neighborhood characteristics of newly introduced topics can be utilized to determine when a topic would emerge in a given domain. Predicting emerging topics requires a method for generating pseudo-neighbors of previously unseen topics as the neighborhood for an emerging topic is not known before its appearance. The authors proposed the Descendant-Aware Clustering algorithm to generate a set of neighborhood candidates for future emerging topics, surpassing existing algorithms in both performances and computation times. Optimizing the algorithm parameters enhances the performance even further. Significant performance improvements were observed when NSGA-III multi-objective algorithm was applied to over 100 research domains. A set of enhanced default values are introduced to the proposed algorithm removing the necessity for dataset-specific optimization, cementing the position of the Descendant-Aware Clustering as the best clustering algorithm for detecting ancestors of future emerging topics.

Topic evolution automatically tracks a set of concepts within a given dataset over time, assisting researchers to overview various research domains. Network-based topic evolution is one of the recent approaches incorporating relational models instead of traditional text-based models for allowing the detection of topic correlation events. Topics are represented with co-occurrence relationships instead of word vectors, connected over time through their positions in a network instead of their semantic similarities. This paper shows that the topics and their network representations share meaningfully similar semantics. The existence of such contextual relationships allows topics to be labeled without having enough direct textual appearances in the document collection. Forty fields-of-study keywords with 64,215 to 5.8 million related articles were selected from the Microsoft Academic Graph dataset containing more than 200 million publications. The semantics of topics within forty topic networks were found using three sets of word embeddings trained from a collection of 10.5 million Medline abstracts from the year 2000 to 2016, and word embeddings of the topics are compared against their network-based representations, which are their neighborhoods in the previous timeslot. Cosine similarities between topics and their neighbors consistently resulted in moderate correlations from the year 2001 to 2015, showing higher values for topics that were already present in the topic networks compared to newly emerging topics. The result suggested that labeling topics based on the network structure are possible without semantic analysis, which is necessary for predicting topic evolutions such as the topic emergence when future documents are unavailable.

Online portfolio selection, one of the major fundamental problems in finance, has been explored quite extensively in recent years by machine learning and artificial intelligence communities. Recent state-of-the-art methods have focused on Mean Reversion significantly and have demonstrated outstanding performance. Another version of the same phenomenon, Median Reversion, has also performed well and demonstrated its ability to be robust against noises and outliers. Another important characteristic is Momentum. In this paper, a Bayesian ensembling approach to extract knowledge from both Mean Reversion and Median Reversion simultaneously based on the momentum associated with each one is proposed for the online portfolio selection task. The proposed method demonstrates its effectiveness by outperforming current state-of-the-art algorithms on several datasets.

As the Age of Information has evolved over the last several decades, the demand for technology which stores, analyzes, and utilizes data has increased substantially. Countless industries such as the medical, the retail, and the aircraft rely on this technology to guide their decision making. In the present paper, we propose a hybrid machine learning algorithm consisting of Decision Trees and Neural Networks which can effectively and efficiently classify data of varying volume and variety. The structure of the hybrid algorithm consists of a decision tree where each node of the tree is a neural network trained to classify a specific category of the output using binary classification. The data with which we used to train and test the classification ability of our algorithm is the Federal Aviation Administration's (FAA's) Boeing 737 maintenance dataset which consists of 137,236 unique records each composed of 72 variables. We perform this by classifying the discrepancy, or cause, of the incident into whether or not the incident occurred during scheduled maintenance operations and then further classifying specific details relating to the incident. Our results indicate that our hybrid algorithm is able to effectively classify incidents with high accuracy and precision. Additionally the algorithm is able to identify the most significant inputs regarding a classification allowing for higher performance and greater optimization. This demonstrates the algorithm's applicability in real-world scenarios while also showcasing the benefits of combining decision trees and neural networks as opposed to using them individually.

Researchers often rely on the byline order in a publication to estimate relative contributions made by its authors, an assumption on which existing author contribution measures are based. This byline-based approach is, however, incompatible with the alphabetical author ordering, a practice still employed by many research fields. Manually requesting authors to state their contributions can overcome the limitation of the existing methods. Such approaches, however, require resource-intensive data acquisition and preprocessing, rendering them ungeneralizable to existing bodies of bibliographic records. The present paper proposed a possibility of order-independent automatic author contribution measure by focusing on distinguishing the main contributors from the rest of the authors using machine learning algorithms, bypassing the limitation of both the byline-based numerical author contribution methods and ungeneralizable manual approaches. The experiment validated the proposed approach by successfully classifying both the primary and the corresponding authors shown as the first and the last author without utilizing byline orders. The Random Forest classifier showed the best performances, successfully classifying the first author, the last author, and both with the accuracy of 0.90, 0.89, and 0.76 respectively.

Understanding the current research topics and their histories allow researchers to focus their capabilities on the current research trends. The field of topic evolution helps the understanding by automatically model and detect the set of shared research fields in the academic papers as topics. The authors propose a novel topic evolution method for identifying and predicting the emergence of new topics under the assumption that neighborhoods of new topics in the future have distinguishable structural features. Eight journals were selected from the Microsoft Academic Graph dataset, each representing topics networks with varying size, history, and research domains. Both retrospective classification and prospective prediction showed promising performance with classifications above 0.89 for six journals and coefficients of determination exceeding 0.95 for five journals. The result showed both the retrospective identification and the prospective prediction can be done, validating the assumption that topic evolution events can be predicted with a network-based approach.

A foundational step in the study of any animal is the establishment of an accurate behavioral model. Building a model that is capable of defining and predicting an animal's behavior is critical to advancing ethological theory and research, however many animal models fail to be sufficiently thorough or often do not exist at all. Great pools of data are available for improving these models through recorded video of animals posted on video hosting sites throughout the internet, however these sources are largely left unused due to their sheer quantity being too much for researchers to manually observe and annotate. This paper proposes a method for efficiently converting video of animals at any length into models capable of making accurate behavioral prediction. This predictive model is developed through a data processing pipeline merging an ensemble meta-algorithm for behavior classification with a long short-term memory network for temporal pattern recognition and prediction. The application of this pipeline produced results with a higher degree of predictive accuracy compared to more traditional autoregressive techniques. These findings suggest the method has significant potential as a tool for efficiently developing new models and findings in the study of animal behavior. The method's performance in prediction also suggests that it may have further application in building models for the prediction of events or failures in inorganic, semi-stochastic processes such as video monitored mechanical systems and equipment.

Journey by aircraft is the only option for long distance transportation and also one of the frequently used modes of transportation of passengers. As a result, safety of passengers and efficiency of the aircraft depend on maintaining efficient running conditions. Although many safety standards are followed in the design of the aircraft and thus there are fewer accidents, it is necessary to perform a thorough analysis to avoid risks that may occur during flight time. In the present work, we propose a maintenance strategy, Failing And Not Falling (F&!F), based on the Federal Aviation Administration (FAA) data in the United States. We work with the dataset of Boeing 737. The data consists of 72 features with 137,236 records which describe an aircraft accident or incident. These features are used to predict whether an incident will be identified during aircraft maintenance or during aircraft operation and what specific type of incident will occur. The prediction method is based on the integration of a decision tree and a unique neural network at each node of the decision tree. The results obtained using different architectures show how deep the neural networks should be, how to identify the relevant features, and the success of combining decision trees and neural networks. Moreover the neural networks and the decision tree approach also successfully identified the important features of maintenance. This method can be used for the maintenance of any data in multiple domains.

Recently, research in machine learning has become more reliant on data-driven approaches. However, understanding the general theory behind optimal neural network architecture is, arguably, just as important. With the proliferation of deep learning and neural networks, finding optimal neural network architecture is vital for both accuracy and performance. Recently, extensive research on neural network architecture has been performed [3],[4]. Additionally, while there has been plenty of research on hidden layer neural network architecture [2], activation functions are often not considered. In a network, an activation function defines the output of a neuron and introduces non-linearities into the neural network, enabling it to be a universal function approximator [12]. In terms of activation functions, one significant paper is Krizhevsky's seminole work on ImageNet classification and the creation of the ReLU activation function [1]. In the paper, Krizhevsky outlines the construction of an image recognition model using the Rectified Linear Unit activation function (ReLU) for the ImageNET LSVRC-2010 competition which outperformed the state-of-the-art image recognition systems at the time [1]. Since then, ReLU has increased in popularity. In their 2018 conference paper, Bircanoğlu and Arica, with the assistance of 231 distinct training procedures, found ReLU to be the best general activation function [12]. In addition to comparisons of activation functions, Nwankpa, Ijomah, Gachagan, and Marshall conducted a meta analysis of the field of research centered around activation functions and found ReLU to be the most popular activation function choice [5]. In terms of optimizers, gradient descent has historically been the most popular loss optimization algorithm, but with Kingma and Ba's 2014 paper [8], Adam: A Method for Stochastic Optimization, Adam optimizer is slowly becoming the industry standard [11]. In their paper, Kingma and Ba cleverly combine momentum descent, RMSprop, and Adagrad optimization into one algorithm, Adam (or adaptive moment estimation) [8]. In addition to Adam, there are plenty of other optimizers to choose from, including gradient descent, RMSprop [9], Adagrad [10], and Adadelta [7]. Recently, many breakthroughs have been made in terms of neural network performance, improved GPU performance and adaptation to deep learning tasks has created massive efficiency increases for the whole field of machine learning. Furthermore, as machine learning becomes increasingly optimized, the importance of efficiency improvements will continue to rise. Thus, understanding the optimal activation function and optimizer choice for a neural network is relevant. The goal of this paper is to make comparisons between activation functions, optimizers, and, more generally, entire neural network architectures, through measured error in a training environment. In this paper, we examine the performance of a wide variety of neural network configurations on randomly generated polynomial data sets of fixed degree. To do this, we compare various neural network activation functions and optimizers while controlling for hidden layer configurations and degree of the underlying polynomial dataset. Curiously, we find that the Sigmoid activation function is more accurate than ReLU and Tanh for regression tasks on low-featured polynomial data. We also reach the same conclusion regarding Stochastic Gradient Descent (SGD) in comparison to the Adam optimization function and Root Mean Square Propagation (RMSprop). Additionally, we observe that SGD is more efficient in the short term for finding local minimums than Adam or RMSprop; however, after sufficiently many epochs, performance differences between the optimizers vanished.

The large volume of scientific publications is likely to have hidden knowledge that can be used for suggesting new research topics. We propose an automatic method that is helpful for generating research hypotheses in the field of physics using the massive number of physics journal publications. We convert the text data of titles and abstract sections in publications to a bipartite graph, extracting words of physical matter composed of chemical elements and extracting related keywords in the articles. The proposed method predicts the formation of new links between matter and keyword nodes based on collaborative filtering and matter popularity. The formation of links represents research hypotheses, as it suggests the new possible relationships between physical matter and keywords for physical properties or phenomena. The suggested method has better performance than existing methods for link prediction in the entire bipartite graph and the subgraph that contains only a specific keyword, such as `antiferromagnetism' or `superconductivity.' Our suggested concept for generating research hypotheses can be easily extended to various other research topics or industrial topics using patent literature.

The exponential rise in the volume of publications and the prevalence of multidisciplinary practice in scientific domains has made it increasingly difficult to keep track of changes in research trends. In this paper, we propose a framework for determining persistent and emerging research topics in scientific literature. The topics were represented as non-overlapping communities of keywords in a dynamic cooccurrence network derived from 21 million articles in PubMed that were published from 1980 to 2016. We detected a set of communities for each snapshot of the network and traced their instances in consecutive periods using a similarity threshold. Our approach provides a retrospective analysis of changes in research topics: their formation, growth, shrinkage, survival, merging, splitting, and dissolution. We also show that a feature set comprising of 43 temporal and structural attributes from these keyword communities can be used to predict their evolution. In particular, we found that the frequency of co-occurrences and the appearance of new keywords within the community are highly predictive of its persistence or dissolution in the next five years.

Video is an increasingly important method of information-sharing on the Web. Services such as YouTube, Vimeo, and Liveleak are platforms that support uploading User-Generated Content. Users tend to seek related information during or after watching an informative video by finding and reading comments on Web services. However, existing services only support sorting by recentness (newest) or rating (LIKES score), as opposed to related information. We suggest a novel method to find informative comments by considering original content and its relevance. We conducted a qualitative study of participants watching informative videos and analyzed how users interacted with the comments, the feature preferences, and the criteria for evaluation. We developed methods to measure informativeness priority, the user-provided level of information, classify intention of information posted, and cluster duplicate themes. Analysis of 1,861 TED talk videos and 380,619 comments show suggested methods can find more informative comments compared to existing methods (LIKES).