The Promotion Effect of LOESS Smoothing Technique in Short-term Traffic Volume Clustering

In short-term traffic volume clustering, one important issue is the representation of traffic profiles. This article focuses on how the LOESS smoothing technique enhances the clustering effect and what the best value of parameter span of LOESS is. This article used K-Means clustering algorithm and compared the clustering effect using raw data and smoothed data. The experiment result shows when the traffic profiles are slight smoothed, the clustering effect enhances from 39.15% to 66.48%. And the best range of parameter span is 0.2~0.4 to keep the balance of clustering effect and profiles details. This article verifies the promotion effect of LOESS smoothing technique in shortterm traffic volume clustering and gives advice on the best value of LOESS parameter. Introduction With the rapid development of information techniques such as Internet of things, big data, cloud computing, Intelligent Transport System (ITS) is making great progress not only in laboratories but also in real traffic management systems in China. Short term traffic analysis is the essential component of ITS to identify specific traffic patterns and make forecasting of short-term traffic flow. Short term traffic clustering, which can identify similar traffic flow of different sections of road on different day, is enlightening to find similar traffic patterns and support traffic management. Some researchers used clustering to support data driven forecasting methods such as K Nearest Neighbor regression and used locally estimated scatterplot smoothing (LOESS) to smooth the noise in traffic flow profiles. However, the promotion effect of LOESS to identify similar traffic profiles was rarely been focused on. This research aims to fill up this gap to find whether LOESS can enhance the clustering effect of short term traffic profiles and what the best parameter of LOESS is to keep the balance of maintaining the profiles details and clustering effect. Short term traffic profile is one type of time series. Aghabozorgi (2015) [1] has reviewed the four main elements of time series clustering, including representation method, similarity degree, prototype definition, clustering algorithm. This article uses Euclidean distance as similarity measurement, average value as prototype definition and K-Means as clustering algorithm. Raw profiles and LOESS smoothing profiles as two kinds of representation methods will be conducted in clustering experiments and the clustering effects will be compared. Methodology K-Means Clustering Methods K-Means is one of the most popular clustering algorithms based on partition. K-Means produces k clusters from n unlabeled objects to make sure that there is at least one object in each cluster [1]. The clustering prototype of K-Means is the average of the objects. The principle of K-Means is to minimize the sum of the distances (generally expressed in Euclidean distance) between all objects in the cluster and their cluster center (i.e. prototype). So this article chooses the result of between-cluster sum of squares by total squared distance as an index of clustering index. For the data in Euclidean International Conference on Modeling, Analysis, Simulation Technologies and Applications (MASTA 2019) Copyright © 2019, the Authors. Published by Atlantis Press. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/). Advances in Intelligent Systems Research, volume 168