Understanding user intent on the web through interaction mining $ Loredana Caruccio n , Vincenzo Deufemia, Giuseppe Polese Department of Computer Science, University of Salerno, 84084 Fisciano (SA), Italy article info Available online 30 October 2015 Keywords: User intent understanding HCI features Web search User behavior mining Query classification abstract Predicting the goals of internet users can be extremely useful in e-commerce, online entertainment, and many other internet-based applications. One of the crucial steps to achieve this is to classify internet queries based on available features, such as contextual information, keywords and their semantic relationships. Beyond these methods, in this paper we propose to mine user interaction activities to predict the intent of the user during a navigation session. However, since in practice it is necessary to use a suitable mix of all such methods, it is important to exploit all the mentioned features in order to properly classify users based on their common intents. To this end, we have performed several experiments aiming to empirically derive a suitable classifier based on the men- tioned features. & 2015 Elsevier Ltd. All rights reserved. 1. Introduction During an Internet navigation session the user per- forms several actions that can provide hints on his/her future activities. Being able to capture and interpret the hidden goals behind such actions can provide organiza- tions with a competitive advantage. For instance, e- commerce organizations might predict user needs, and advertise the products that users will most likely buy, also through the mining of previous customers purchase stra- tegies [1]. Thus, multimedia catalogues, web and infor- mation retrieval systems need to embed search engines capable of capturing user intent, which is the focus of user intention understanding (UIU) research area [2]. Many approaches for user intent understanding are based on the analysis of search behaviors [3–5], such as clicked URLs [6] and submitted queries. Most of them aim to capture semantic correlations among search behaviors of the same user, in order to let search engines produce customized results for each individual user. Other studies analyzed user interactions with Search Engine Result Pages (SERPs) to infer their intent [7–10]. However, by limiting the analysis to results contained in a SERP, such methods ignore many important interactions and contents visited from such results. For this reason, some approaches to user behavior analysis focus on user interactions with web pages to infer clues on their interest and satisfaction with respect to the visited contents [11–13]. Following this trend, in this paper we define a new model for UIU analyzing both interactions with SERP results and those on the visited web pages. The interaction features considered in the proposed model are local page level statistics, that is, they are fine-grained and refer to portions rather than the whole web pages. This provides the basis for a more promising prediction of the user intent, since several experiments with eye-trackers revealed that users analyze web pages by sections, over- looking those of low interest [14]. Contents lists available at ScienceDirect journal homepage: www.elsevier.com/locate/jvlc Journal of Visual Languages and Computing http://dx.doi.org/10.1016/j.jvlc.2015.10.022 1045-926X/& 2015 Elsevier Ltd. All rights reserved. ☆ This paper has been recommended for acceptance by Henry Duh. n Corresponding author. E-mail addresses: lcaruccio@unisa.it (L. Caruccio), deufemia@unisa.it (V. Deufemia), gpolese@unisa.it (G. Polese). Journal of Visual Languages and Computing 31 (2015) 230–236