15h30 - 15h55
Text analytics and topic extraction from employee feedbacks
In this talk, we present a methodology to extract the main topics in a corpus composed of over ten thousand textual employee feedbacks, from a single company. We developed an analytics tool and used DBSCAN to find semantic clusters of feedbacks automatically.
15h55 - 16h20
Predicting engagement and sentiment in commercial B2C Facebook pages using textual information
We propose an alternative to manual tagging of social media data by exploring the use of two training metrics. The first is derived from platform users feedback metrics unique to Facebook. The second is the aggregate emotional polarity expressed by platform users through their comments on a status. In both cases, a model using the company status is proposed that predicts the reaction of the users.
16h20 - 16h45
Railway demand forecasting using machine learning approaches
In railway industries, demand forecasting is the estimation of the number of passengers aiming to travel by train with a specific itinerary. Railway uses this information to satisfy their demands and maximize the revenues. In this study, we present detailed analyses of applications of various machine learning algorithms and preprocessing techniques to predict the future bookings in railway industries in two different aggregation levels. As a result, stacked generalization method combined with proper preprocessing techniques outperformed other approaches at both levels. We successfully achieved 11% Mean Absolute Percentage Error for level-1 aggregation and 18% Weighted Absolute Percentage Error for level-2.
16h45 - 17h10
VRP routing with driver experience considerations: automating the process of driver to route assignment according to previous deliveries
This study introduces a method for representing the spatial and the temporal knowledge of drivers. The objective is to take into account their knowledge for the planning of delivery routes in vehicle routing problems. Assigning drivers to routes in areas they are familiar with will potentially lead in improvements in terms of accuracy, execution speed and overall transportation costs. Computer experiments from real data show the soundness of our approach.