In this paper, we focus on the popularity effect of the scientific collaboration process that popular authors have an advantage in making more publications. Standard network analysis has been used to analyze the scientific collaboration network. However, the standard network has limitations in explaining the scientific output by binary co-authorship relationships since papers have various numbers of authors. We propose a leading author model to understand the popularity effect mechanism while avoiding the use of the standard network structure. The estimation algorithm is presented to analyze the size of the popularity effect. Moreover, we can find influential authors through the estimated genius levels of authors by considering the popularity effect. We apply the proposed model to the real scientific collaboration data, and the results show positive popularity effects in all the collaborative systems. Furthermore, finding influential authors considering the genius level are discussed.
Jan 1, 2022
Sharing economy platforms for accommodation sharing offer a more flexible supply-demand model compared to traditional hotels that own a fixed number of rooms. However, this flexibility can cause uncertainty and become a challenge if not managed dynamically. An important task for such platforms is forecasting future daily occupancy in a certain area, with sufficient lead time so that they can reach out to new hosts and secure more rooms in time for peak demand. We developed such a forecasting solution for AsiaYo, the largest Chinese language online accommodation sharing platform. We evaluate and compare various forecasting algorithms, including statistical and machine learning methods, using two years of data from AsiaYo on occupancy in different cities. The empirical results show that the occupancy is highly dependent on the weekday, city, and holidays. We show the strengths and weaknesses of different methods in terms of required accuracy level, computation time, and flexibility.
Nov 1, 2019
With the dramatically increasing amount of data available owing to fast growing financial activities, data about customers has been fast accumulating and becoming one of the most valuable asset for financial holding companies (FHC). These financial activities have created a huge amount of customer behavioral data, which is regarded as important information for direct marketing. We utilized powerful deep learning algorithms combined with oversampling and feature extraction techniques, applied to behavioral financial data, for direct marketing. This paper describes an implementation of prediction models combined with practical business and analytic goals. Real-world data used in model training and testing were collected from a Portuguese bank marketing campaign. Our model achieved a lift index of 91.76%, which is comparable to the random forest (92.03%) and better than näıve Bayes (80.89%) and logistic regression (89.72%). We find that increasing the amount of training data available improves performance of the deep learning algorithm. Therefore, we believe that our model will potentially outperform the random forest with sufficiently large samples. With this close integration of business analytics and deep learning, our model can be thought of as a prototype for direct marketing and can be finally implemented in FHC systems to deal with large amounts of data.
Jun 1, 2018