This research starts by considering users of a digital library (DL) and aims at using the data extracted from logged les, including search strategies, and queries, to build e ective user-interaction pro les. and use them to guide designers and systems developers in the production of more usable, useful and e ective interaction pro les. It is important to stress that the proposed interaction pro les are built by extracting a number of features from the log les. Thus, they contain information about real searching experiences, including usage patterns, user familiarity with the system, and time intervals. This study is conducted in collaboration with RERO Doc digital library1. RERO Doc is the network of the libraries of Western Switzerland. Users from di erent parts of the world can search on di erent domains: Nursing, Economics, Computer Science and others. RERO Doc provides various document types such as books, articles, theses, periodicals,etc. Thus, the research questions are:
(1) What are the most suitable techniques to produce rich/realistic groups of data extracted from log les in order to build interaction pro les? (2) What are the main features to characterise interaction pro les? (3) What is the minimum size of data to produce robust groups to use for building interaction pro les?
Data preparing and processing: Interface analysis: the interface is inspected in order to understand di erent search options.
Preprocessing phase: we follow the framework of [2] as the following: Data loading: the dataset consists of 59 million records 20 GB collected over a six-month period.
Data cleaning: including users identi cations hidden, elimination of the erroneous, and corrupted records.
Data parsing: consists of sessions recognition and removing the nonhuman sessions e.g. Googlebot, SemanticScolarBot. The session "is a common unit of interaction that is used in search log analysis" [3]. Session recognition depends on the user interactions and on the features of the interface. This phase is crucial for identifying distinct classes of searching patterns [2, 3]. We identify a session by the combinations of user IP, time stamp, and user agent extracted from the log les. Also, the non-human requests are removed in this stage and so data is reduced to 9 GB.
Data coding: in this phase, the URL requests were analysed and divided into meaningful parts including user IP, time stamp, request, referrer, user agent, session IP. Researchers follow di erent strategies to analyse the URLs embodied on the log les. For example,[1] build a hierarchical taxonomy of the website, and [2] de ne a code schema of the all types of the interactions based on analysing and understanding the structure of the website. In this research we consider both strategies.
Features engineering: Based on the interface and log les analysis, the meaningful features are identi ed.
Mining user behaviour: The remaining sessions are further analysed and grouped based on the available variables in the records. We started the analysis of the rst million record which consists of 125000 session records. 72095 sessions were detected with only one record. The aim of this phase is to identify di erent usage patterns among user interactions and group them accordingly. Two main di erent grouping techniques were used: topic modelling and K-Means. For 6 topic model, the Coherence Score of the topic modelling is 0.35. For K-Means the estimated number of clusters is 6 with Silhouette Coe cient of 0.95. So far, six di erent usage patterns have been identi ed and interpreted qualitatively: (1) Single sessions or known-item, where searchers visit RERO for downloading documents without any interactions. (2) Complicated sessions, where the usage pattern is characterised by heavily interactions including submitting queries, browsing, and using di erent functions. (3) Light navigators who navigate the library for navigating without using di erent functions on the interface. (4) Advance navigators whose navigations are characterised by using di erent functions and many iterations. (5) Light browsers whose searching is simple, short without using di erent functions. (6) Advanced browsers, their interactions is long and including advance search functions.
In conclusion, log le analysis is an unobtrusive method to detect usage patterns of digital library. The aim of this research in progress is to build interactions pro les in the digital library context in order to gain more insights into users searching experiences. We plan more experiments to test and compare the e ectiveness of the techniques used to group data for preparing interaction pro les. Then, we will involve experts to assess the quality of these interaction pro les and how e ective these are in assisting designers and system developers in the production of more usable, useful and e ective tools to support searchers.