The study of user interactions with information retrieval systems has been examined from different perspectives, including direct observation and transaction logs. The growing availability of server-side transaction logs has provided researchers with a rich set of data that can objectively record the actions of system users, whether for online public access catalogs, bibliographic database systems, web search engines, or digital libraries. Studies of user interactions with public search engines have been carried out since the mid-1990s. Jansen (2006) has reported that stored data in transaction logs of Web search engines, intranets and websites can offer valuable insights about the information searching process of online searchers. A number of researchers (Jansen, Spink, and Saracevic 2000; Spink, Wolfram, Jansen, and Saracevic 2001; Jansen and Spink 2006; Zhang, Wolfram, and Wang 2009) have conducted transaction log query analysis of various websites. These studies have revealed that interactions with public search services such as search engines are generally brief with little browsing of results and modification of findings (Spink, Wolfram, Jansen, and Saracevic 2001; Jansen, Spink, and Koshman 2007; Markey 2007; Jansen, Booth, and Spink 2009). These studies have also focused primarily on query-related data and not on the full click-through data of transaction logs, i.e., actions related to all aspects of the information seeking process, including search results and related actions taken by users. The present research is prompted by the desire to determine if the findings regarding user interactions for public search services are similar for digital libraries (DLs) where, unlike general search engines, users will rely on specialized document collections, such as image collections, for specific purposes. Similarly, does the nature of the collection (e.g., images) also influence how users search?

The research questions addressed by the current study include: 1) What types of search actions do users engage in most frequently when searching image-based collections? 2) How do users engage in search actions during search sessions as reflected by an action transition network? 3) What are the frequently observed queries and terms in searching the studied imagebased collections?4) Are there differences in query patterns between queries submitted internal to the digital collections and those originating from outside the collections, such as search engines?

This study is novel in that it investigates users’ search actions, especially transitions in those actions, in an image-based digital library. In research questions 1 and 2, we intend to understand which types of user actions occur most or least frequently and identify frequent transition patterns between those actions during the search process in the context of an image-based digital library. By analyzing unobtrusive transition logs, we are able to quantitatively identify the most and least frequently applied user actions. Furthermore, we are able to identify the most common transition patterns in actions that enable us to better understand users’ engagement in search processes. The understanding of users’ search actions is important to model users’ search behavior in an image-based digital library, and to generate insights into more efficient and effective system design. Research questions 3 and 4 are prompted by the desire to determine if there are specific areas of the collections that are of particular interest to searchers and whether users are more likely to engage in more natural language-based searches or if users rely on metadata content from the digital collections. Given that the content of the digital collections is also externally accessible through search engines, a comparison of queries submitted internally through the digital collections interface and those external to the system may reveal different patterns arising from access to the digital collections search features that would not be available through a search engine interface.

To date, user interaction studies involving DLs have been carried out using direct search observation (Xie and Cool 2009; Joo and Xie 2012) and resource usage logs (Wolfram and Xie 2002). Transaction log analysis (TLA) studies of DLs also have been undertaken, but have focused on systems that are primarily text-based or mixed media (Jones, Cunningham, McNab, and Boddie 2000; Bollen and Luce 2002; Jamali, Nicholas and Huntington 2005). Khoo et al. (2008) discussed the use of web metrics to analyze various user actions in digital libraries including session length and page views. They pointed out the difficulty in identifying firm inferences of users’ intentions made from web metrics using transaction logs. In other words, web metrics record users’ behavior rather than the thought processes and intentions. The authors also noted the difficulty in distinguishing separate sessions from the same IP address.

Similarly, earlier studies of image or multimedia searching using public search engines focused largely on query-related activities surrounding image collections (Goodrum and Spink 2001; Choi and Rasmussen 2003; Jörgensen and Jörgensen 2005; Tjondronegoro, Spink, and Jansen 2009; Huurnink, Hollink, van den Heuvel, and de Rijke 2010) or click-through data for results assessment (Smith, Brien, and Ashman 2012). Effective searching for images presents challenges not found in text-based retrieval. The search process and actions undertaken over the course of a session (i.e., a series of actions undertaken to fulfill an information need) are not well understood. Moreover, studies of image retrieval on different information retrieval systems investigated users’ query patterns. Choi and Rasmussen (2003) found that users’ queries for images in digital archives focused on searches for names of events, actions or conditions, individuals and place names. Jansen (2008) identified users’ image query structure on the Web and provided attributes to classify Web image queries, while Pu (2008) examined the characteristics of users’ failed queries for images on the Web. The results revealed that users showed successful image retrieval on the Web with short queries while users’ failed searches for images were associated with longer queries. Finally, Choi, and Hsieh-Yee (2010) conducted a query analysis for images using an OPAC. The authors found that the Boolean operator “AND” was the most commonly used feature in the queries and that the most frequently occurring initial queries were composed of two terms.

Although previous studies of DL interactions have relied on user observation and screen capture software, the amount of data collected from this approach is more limited than the types of data that can be collected using click-through data automatically collected by Web servers. The present study sheds new light on user interactions with image- based DL collections by examining user session behavior on a larger scale than is feasible with direct observation. Each user engages in search behaviors that are reflected in recorded actions. When combined with other users’ data, a more complete picture emerges of overall user behavior that reveals patterns which may help to inform how digital libraries are designed, and particularly how to better facilitate browsing.

The University of Wisconsin-Milwaukee libraries house a digital library consisting of approximately 30 digitized visual collections. The collections comprise more than 54,000 photographic images and maps. In particular, the selected digital library has specialized collections of regional archives that address Milwaukee history, Milwaukee neighborhoods, and University of Wisconsin-Milwaukee-related archives. In addition, the digital library houses collections of the American Geographical Society Library, so there are a large number of maps or geographical items in the collections. The selected digital library has been developed using CONTENTdm software (http://www.contentdm.org/), which is a content management tool for archiving, managing, and servicing digitized items. CONTENTdm basically provides a search function, including basic search, advanced search, and faceted search, and supports different viewing options. Also, the UWM digital collections provide well organized metadata based on the Dublin Core metadata standard. Item view pages present images or videos with metadata fields to help users evaluate items. The metadata fields provide different attributes of a digitized item, including title, notes, date of photograph, photographer, subject terms, regional information, item size, rights, and other information. The primary entry point to the DL consists of several HTML pages that permit users to browse the collections or to submit queries to search the collection metadata. A screenshot taken from the DL appears in Figure 1.

Most user actions were carried out as PHP requests. PHP is a scripting language for dynamic and interactive Web pages. The presented analysis is based on one month of data collected in August of 2012. The transaction log files consist of all requests made by users and system actions taken. Each transaction log entry consists of an encrypted Internet Protocol (IP) address, time, date, HTTP command with PHP actions, and originating Uniform Resource Locator (URL). In addition, transaction records include users’ query input and selection of topic categories through the HTTP “Post” data transition method. Data were first cleaned to remove system-initiated action requests that were not user-related. Unique page requests from outside the DL were also removed. These were assumed to represent direct links to specific image files within the collections and were not representative of search sessions.

One challenge for TLA in environments that do not require a login is the identification of session boundaries, i.e., when a search session begins and ends. IP addresses are usually used to identify actions associated with a given session. This is also an issue in public DLs as noted by Khoo et al. (2008). For the present study, IP addresses were used as one indicator to delineate sessions. Actions taken on different days for the same IP address were also used to delimit sessions. The challenge with differentiating session boundaries arises when actions are associated with the same IP address on the same day, but in reality represent different sessions. It is possible that IP addresses may be shared by different computers, so that actions associated with one IP address may represent multiple, concurrent sessions. We have assumed that consecutive actions associated with the same IP address represent the same session. However, a period of inactivity between actions for the same IP address may also indicate a different session. How does one determine a cutoff point for periods of inaction that delineate different sessions for the same IP address? The cutoff determination for adjacent sessions with the same IP address should not be arbitrary. We employ a novel method for estimating the cutoff time between adjacent sessions for the same IP address. The cutoff time between actions served as the session boundary for actions associated with the same IP address on the same day. To identify the cutoff point for inactivity, we analyzed the distribution of time intervals in combination with the occurrence of those time intervals. To easily interpret the result, the data were transformed using logarithms. The distribution of intervals between actions exhibited a “V” shaped curvilinear pattern. We assumed that the inflection point implies the time interval that would be long enough to show the termination of a session. Before the inflection point, two adjacent actions were considered to occur in the same session. As the distribution is non-linear, Kernel regression was applied using MATLAB to find the inflection point. Kernel regression is a non-parametric estimation method to predict the conditional expectation of a random variable (Shawe-Taylor and Cristianini 2004). Unlike parametric regression, kernel regression includes a smoothing function as an estimation method, which consists of the kernel and bandwidth:

In this study, an Epanechnikov kernel was selected as a weight function (Li and Racine 2007), while the bandwidth was calculated based on Silverman’s rule (Silverman 1986). The kernel regression result revealed the time interval cutoff between adjacent sessions was: 10^2.4 = 251.19 seconds. This is reflected in Figure 2.

As a further limiter to sessions, because system content is accessible with direct links from outside sources, only sessions beginning with the home page or one of its subordinate PHP pages were included in the analysis.

The resulting cleaned log file was analyzed using descriptive statistics for the different types of available actions. A network analysis of adjacent actions undertaken within sessions was conducted using UCINET 6 (Borgatti, Everett, and Freeman 2002) to reveal the centrality and relationships among different actions. Since users’ search processes can be represented as transitions in search actions, we decided to focus on the changes of users’ actions. A network analysis is a compelling tool to analyze and visualize those transitions. In this study, we viewed each type of action as a node connected by a directed edge to the adjacent action in the same session. The network analysis summarizes the entire transitions of user actions during search processes, and highlights the most frequent transition patterns of search actions in searching the digital library. Even though the network analysis does not show sequential search action patterns in individual sessions, it enables us to understand the overall users’ search patterns comprehensively. By employing network analysis, we were able to identify which user actions play a central role in the search process and what types of transitions occur most frequently.

An analysis of queries submitted during the data collection period was also conducted. There were two kinds of search query strings: those originating from the digital collection site (internal) and those from outside sources such as search engines (external). The identification of queries in the raw transaction log file was challenging. Internally submitted queries were identified by locating transaction log lines that contained the feature “CISOBOX1” as a field name for the primary search function, while the external searches were identified with “q=” for the query field. In order to compare the difference between internal searches and the external searches, all the transaction lines with the string “CISOBOX1=” in the referral field and all lines with the string of “q=” in the referral field were extracted into two files.

For both files, the entries were sorted in alphabetic order of the encrypted IP addresses. From the sorted data set, all lines with the repeated same referral fields from the same IP addresses were eliminated. The repeated lines were generated because often one displayed page has multiple components such as images, icons, and CSS files. In this way, all the unique queries were identified.

To analyze the records at the query and term level, only the query strings were extracted from the URL encodings at each referral field. The extracted query strings were further cleaned by removing all the non-alphabetic characters (digits and special characters) to make counting simpler. All the remaining alphabetic characters were transformed into lower case for more accurate word and query counting.

The transaction log analysis revealed 9758 sessions with an average of 10.86 page views or actions per session (median 2; min. 1; max. 446; s.d. 33.60) with a frequency distribution that was highly skewed (skewness 6.53). A summary of the total actions and page requests extracted from the transaction log file appears in Table 1.