Personalized I nform ation Delivery on Objectives of this Talk the - PowerPoint PPT Presentation

Personalized I nform ation Delivery on Objectives of this Talk the Static and Mobile W eb � Traditional IR vs. mobile IR � Information Push as the default information access model � Estimating user interests via search engine clickthroughs Dik Lun Lee Departm ent of Com puter Science and Engineering Hong Kong University of Science and Technology Nov 2 , 2 0 0 9 1 Profiling User Interests in Search 2 Engine Web Search vs. Mobile Search Proactiveness: While you are shopping… � � Do you want your mobile devices to be loaded with Simple mobile search model Increasingly context aware useful coupons, store information and sales items? � Shrink the desktop/ web search onto a mobile device � Voice I/ O, auto-completion (Google Suggest), query � suggestion, aiming at reducing the user I/ O effort What about a bookstore offering a discount on a � book that you browsed on Amazon yesterday? Vertical search services to cater for common mobile search � Route, restaurant, directory search � � What about the time for the next bus that you take Yahoo Go!, Google Mobile every day? � Proactive model � � Up-to-date and relevant information are pushed to mobile … … device, replacing explicit requests by local browsing � Make possible by large local storage and high bandwidth � Require profiling user interests and context awareness � Best-effort suggestions Profiling User Interests in Search 3 Profiling User Interests in Search 4 Engine Engine

User Profiling: Online vs Mobile Location-Based Search Query Match Content & Rank Keywords Keywords Content/ Keyword driven Profile driven Content Query Keywords Keywords Match & Rank Location Time & Location Names User Profile Content Query Documents Web Web Keywords Keywords Match & Rank Repository Location Location Names Names Location User Match & Keywords Location Rank What does the user really want? Profiling User Interests in Search 5 Profiling User Interests in Search 6 Engine Engine User Profiling as a Universal Requirement User Profiling: Online vs Mobile � � Web/ desktop search, mobile search, pro-active or passive, Comprehensive profiling knowing the user interest is very important � Online tracking: search and web browsing � More relevant search results � Predictive of future events and needs � Suggest relevant queries � Display related information � Mobile tracking � Predictive of local interests (both temporal and spatial) and action items � Question: how to collect, derive, represent, utilize and � Location semantics: semantic location modeling refine Profiling User Interests in Search 7 Profiling User Interests in Search 8 Engine Engine

User Profiling – An Example User Profiling – Concept Extraction Planning ( 1 w eek to 1 m onth) Engaging ( a few days) Search Brow se Search Brow se Widm 2009 Query Widm 2009 Airport -> Hotel Widm ‘09 homepage Date, venue Concept space -Hotel name View and Search Result -Registration page -Address browse ( Snippets) -Workshop page Content -Reservation No. Program -Widm ‘09 page Relevant Other hotels Hotels stayed Concepts Location Before: -Hotel Names -Hilton -Websites Clicked Pages -Hyatt -Phone numbers Content Hotel homepages Names -Peninsula -Current prices Refined Phones -Old prices Concepts Availability Location -Etc… Flights, etc. profiles Engagement User profile Profiling User Interests in Search 9 Profiling User Interests in Search 10 Engine Engine Clickthrough Data Inferring User Preferences (Joachims) � Assumption: Users read the results from top to bottom, click on Doc Clicked Search results relevant results and skip non-relevant results √ d 1 Apple Computer � E.g., the user clicked # 1, # 4 and # 8, we can Result list: d 2 Apple – Quicktime infer that # 1, # 4 and # 8 are relevant while 1. Apple Store √ d 3 Apple – Fruit # 2, # 3, # 5, # 6 and # 7 are non-relevant 2. Apple - QuickTime √ d 4 Apple - Mac � It cannot infer if # 9 and # 10 are relevant or 3. Apple - Fruit History of Apple Computer not since it is not sure if the user has d 5 4. Apple .Mac √ examined the items below the last click d 6 Apple Mac News 5. www.applehistory.com � Instead of a relevant vs non-relevant decision, 6. Adam Country Nursery d 7 Apple tree 7. Apple cookbook the following user preferences can be inferred: √ d 8 Apple – Support 8. Apple Support √ � # 1 over # 2, # 3, # 5, # 6 and # 7 d 9 AppleInsider � # 4 over # 2, # 3, # 5, # 6 and # 7 9. … … � # 8 over # 2, # 3, # 5, # 6 and # 7 10.… … � Preference mining: Given the clickthrough data, what is � no further preference can be concluded the user interested in? Profiling User Interests in Search 11 Profiling User Interests in Search 12 Engine Engine

From Page Preference to Concept Preference Page j Page i Page j Page i fruit computer fruit computer < q juice iPod iPod juice Now we know concepts are used to profile a user’s iPhone farm iPhone farm interests How to know if a concept is content or location [computer, iPod, iPhone] < q [fruit, juice, farm] related? Feature vector / User profile a i computer iPod iPhone fruit juice farm … weight 1 1 1 -1 -1 -1 0 Profiling User Interests in Search 13 Profiling User Interests in Search 14 Engine Engine Example: Location Query Example: Location Query Location concepts Location concepts � � A query can be described by A query can be described by Daytona Beach Cambodia the concepts it retrieves the concepts it retrieves Huntington Beach Indian Ocean Long Beach Indonesia Malaysia Myrtie Beach Q= beach Q= Southeast Asia Thailand Palm Beach Singapore Venice Beach camp Vietnam biking Content Content hotel language concepts concepts resort people restaurant relief effort vacation travel Profiling User Interests in Search 15 Profiling User Interests in Search 16 Engine Engine

Concept Extraction Concept Ontology � Content concepts are organized into hierarchy � The longest sequence of words appear in > n snippets. � Similarity(x,y) = > x and y coexist in the same snippets m � Snippets are considered by the search engine as the most times important document segment relevant to a query � Parent-Child(x,y) = > x coexists with many concepts, � Identify longest meaningful phrases in the snippets including y but not vice versa Search Brow se Query Concept space Content Relevant Search Result Concepts ( Snippets) Location Profiling User Interests in Search 17 Profiling User Interests in Search 18 Engine Engine Location Ontology User Behaviors � User behaviors are described by the concepts they clicked � Content feature vector | | Location feature vector Retrieved Pages Concept space Content Relevant Clicks Concepts Location Content Content feature vector Clicked Concepts � Prebuilt location hierarchy Location Location feature vector � A concept that matches a node is a location concept User profile Profiling User Interests in Search 19 Profiling User Interests in Search 20 Engine Engine

Measuring Content and Location Richness � How much content and location is a query associated to? � A concept is location oriented if it is associated with a large number of different locations � Is a concept either 100% content or 100% location? A concept is content oriented if it is associated with a large number of different concepts Hong Kong ⇒ ~ 100% location � A concept may be both content and location oriented with Programming ⇒ ~ 100% content different degrees of richness Java ⇒ half-half ??? HKUST ⇒ 80-20 ??? � Content entropy: What about ` ` Books’’, ` ` Physics’’, … ? � Location entropy: Profiling User Interests in Search 21 Profiling User Interests in Search 22 Engine Engine Measuring Content and Location Interests Query Classes � Four combinations of content and location entropies: � � Clicked content entropy: low/ low, high/ low, low/ high and high/ high � Explicit, content, location, and ambiguous queries � Note: Beijing is not entirely location-oriented and Manchester � is rich in content as well !!! Clicked location entropy: � Given a concept, is a user interested in the content and/ or the location aspects of the query? Consider ` ` Java’’, ` ` apple’’, etc. � Did the user click on a large number of various locations? � Did the user click on a large number of various concepts? Profiling User Interests in Search 23 Profiling User Interests in Search 24 Engine Engine