Citation Context

...cting an in-depth study of behavioral differences based on query type. ‘Why’ People Search for Images. The general goals and tasks that motivate search provide the context for search behavior. Broder =-=[3]-=- proposed a taxonomy of intent for web search, which was adapted to image search by Lux et al. [13]. No previous work has attempted to link these classes of image search intent to specific query types...

Citation Context

...mising sources of additional implicit feedback. (3) We show that CTR/HTR is heavily dependent on query type, as is dwell time and referral page CTR, which has important applications in click modeling =-=[11]-=-; the huge variation in CTR, HTR, dwell time, and referral CTR based on query type shows that, to effectively interpret implicit user feedback as a relevance signal, query type is essential. Similarly...

Citation Context

...t returned page, and ends the session with no further action. Click/Hover-Through Rates. We now measure clickthrough rate (CTR), which has been shown to have a strong correlation with image relevance =-=[4, 19]-=-, as well as hoverthrough rate (HTR). For each (image,query,position) triple, where position is the ranking of the image in the result list, we calculate CTR/HTR, the number of clicks/hovers on the im...

Citation Context

...ching for people, locations, etc., and on whether the search was about unique instances or non-unique instances, which is closely related to the facet-based Shatford-Panofsky categorization framework =-=[18]-=-. This framework was also used by Armitage and Enser [2] to classify image queries in offline multimedia archives. The above studies, and other similar studies [6, 8, 17, 20], agree in their main find...

Citation Context

...images of the same type, this can be used to improve relevance. Also, the facet-based classification results suggest what types of image classifiers the community should work on (e.g. the who facet). =-=(2)-=- We emphasize 3 alternative sources of implicit relevance feedback from search logs: hovering over images, dwell time in the preview page, and click-through from the image preview to the referral webs...

Citation Context

...es of query reformulation used and how this relates to query type. Query reformulation is normally defined as any modification that a user makes to the initial query in hope of finding better results =-=[7]-=-. We adopt the approach of Jansen [9] and distinguish the following types of reformulation between consecutive queries: (i) adding terms, where one or more terms are added to the original query, (ii) ...

Citation Context

...d-Panofsky categorization framework [18]. This framework was also used by Armitage and Enser [2] to classify image queries in offline multimedia archives. The above studies, and other similar studies =-=[6, 8, 17, 20]-=-, agree in their main findings that web image search is dominated by people queries or queries within the Arts & Entertainment category. They either explicitly focus on the most popular queries or do ...

Citation Context

...ation With today’s high volume traffic on web search engines, one of the most common approaches to gain insights into user behavior is the analysis of query logs. Previous studies on web image search =-=[1, 5, 8, 17, 20]-=- focused on characterizing overall characteristics based on aggregated search logs. These findings help us to gain an overall picture of how users search for images on the web, but do not capture vari...

Citation Context

...ation With today’s high volume traffic on web search engines, one of the most common approaches to gain insights into user behavior is the analysis of query logs. Previous studies on web image search =-=[1, 5, 8, 17, 20]-=- focused on characterizing overall characteristics based on aggregated search logs. These findings help us to gain an overall picture of how users search for images on the web, but do not capture vari...

Citation Context

...he web, but do not capture variation amongst different types of image queries; it is likely that behavior varies across query types, since user behavior in search is heavily dependent on task or goal =-=[12]-=-. Understanding such differences, and how they relate to user intent, is the main focus of this work. For instance, we posit that searches for ‘Britney Spears’ would, in general, exhibit different sea...

Citation Context

...ation With today’s high volume traffic on web search engines, one of the most common approaches to gain insights into user behavior is the analysis of query logs. Previous studies on web image search =-=[1, 5, 8, 17, 20]-=- focused on characterizing overall characteristics based on aggregated search logs. These findings help us to gain an overall picture of how users search for images on the web, but do not capture vari...

Citation Context

...g that image searches lead to more clicks and deeper exploration of the search results (search depth), and conclude that image search tends to be more ‘exploratory’ and requires greater interactivity =-=[1, 10]-=-. Other work focus on search behavior on specialized image sharing platforms like Flickr [14]. Andre et al. [1] argue that, while image search tends to be more exploratory than text search, image sear...

Citation Context

...ation With today’s high volume traffic on web search engines, one of the most common approaches to gain insights into user behavior is the analysis of query logs. Previous studies on web image search =-=[1, 5, 8, 17, 20]-=- focused on characterizing overall characteristics based on aggregated search logs. These findings help us to gain an overall picture of how users search for images on the web, but do not capture vari...

Citation Context

...w this relates to query type. Query reformulation is normally defined as any modification that a user makes to the initial query in hope of finding better results [7]. We adopt the approach of Jansen =-=[9]-=- and distinguish the following types of reformulation between consecutive queries: (i) adding terms, where one or more terms are added to the original query, (ii) deleting terms, where one or more ter...

Citation Context

...ation With today’s high volume traffic on web search engines, one of the most common approaches to gain insights into user behavior is the analysis of query logs. Previous studies on web image search =-=[1, 5, 8, 17, 20]-=- focused on characterizing overall characteristics based on aggregated search logs. These findings help us to gain an overall picture of how users search for images on the web, but do not capture vari...

Citation Context

...earch behavior differ depending on query type? Human interactions with computers never take place in a vacuum, but always occur in response to some user need or goal, and image search is no exception =-=[13]-=-. As with any information service, to provide the best possible service it is crucial to understand the underlying goals behind user behavior, motivating our final research question: RQ 3. Can we asso...

Citation Context

...t returned page, and ends the session with no further action. Click/Hover-Through Rates. We now measure clickthrough rate (CTR), which has been shown to have a strong correlation with image relevance =-=[4, 19]-=-, as well as hoverthrough rate (HTR). For each (image,query,position) triple, where position is the ranking of the image in the result list, we calculate CTR/HTR, the number of clicks/hovers on the im...

Citation Context

...t-based categorizations. ‘How’ People Search for Images. Following on from the work of Goodrum and Spink [5], researchers started working on characterizing image search behavior on web search engines =-=[1, 5, 6, 8, 15, 17, 20]-=-. These studies characterize the general behavior of users on image search platforms based on aggregated search log data, measuring features like session length, the number of result pages viewed, the...

Citation Context

...nce there is no standard subjectbased taxonomy for image queries, we chose the IPTC subject code taxonomy4, which is often used for classifying online content, and has previously been used for images =-=[21]-=-. The taxonomy includes 17 top level subject-based nodes (e.g. Arts, Culture & Entertainment, Lifestyle & Leisure). Through the process of manual annotation, we identified an additional root category,...