Towards Inferring Searcher Intent Eugene Agichtein Intelligent Information Access Lab (IRLab) • • • • Text and data mining Modeling information seeking behavior Web search and social media search Tools for medical informatics and public health Ablimit Aji (2nd year PhD) Qi Guo (3rd year Phd) In collaboration with: - Beth Buffalo (Neurology) - Charlie Clarke (Waterloo) - Ernie Garcia (Radiology) - Phil Wolff (Psychology) - Hongyuan Zha (GaTech) 1st year graduate students: Julia Kiseleva, Dmitry Lagun, Qiaoling Liu, Wang Yu Eugene Agichtein, Emory University, IR Lab 2 Online Behavior and Interactions Information sharing: blogs, forums, discussions Search logs: queries, clicks Client-side behavior: Gaze tracking, mouse movement, scrolling Eugene Agichtein, Emory University, IR Lab 3 Research Overview Discover Models of Behavior (machine learning/data mining) Intelligent search Information sharing Health Informatics Eugene Agichtein, Emory University, IR Lab Cognitive 4 Diagnostics 4 Main Application Areas • Search: ranking, evaluation, advertising, search interfaces, medical search (clinicians, patients) • Collaborative information sharing: searcher intent, success, expertise, content quality • Health informatics: self reporting of drug side effects, co-morbidity, outreach/education • Automatic cognitive diagnostics: stress, frustration, other impairments … Eugene Agichtein, Emory University, IR Lab 5 Talk Outline Overview of the Emory IR Lab Intent-centric Web Search Classifying intent of a query Contextualized search intent detection Eugene Agichtein, Emory University, IR Lab 6 Web Retrieval Architecture [from Baeza-Yates and Jones, WWW 2008 tutorial] Example centralized parallel architecture Web Crawlers Eugene Agichtein, Emory University, IR Lab Information Retrieval Process (User view) Source Selection Resource Query Formulation Query Search Ranked List Selection query reformulation, vocabulary learning, relevance feedback source reselection Eugene Agichtein, Emory University, IR Lab Documents Examination Documents Delivery 8 Some Key Challenges for Web Search • Query interpretation (infer intent) • Ranking (high dimensionality) • Evaluation (system improvement) • Result presentation (information visualization) Eugene Agichtein, Emory University, IR Lab 9 Intent is “Hidden State” Generating Actions Satisfied Unsatisfied Intent “States” • First (naïve) generative model of user actions: – Given a state (e.g., “Unsatisfied” with results) User generates actions such as query, click, browse Eugene Agichtein, Emory University, IR Lab 10 Problem Statement • Given: Sequence of user actions and background knowledge, Predict user intent and future actions – Will define intent classes, actions next • Example applications: – Predict document relevance (ranking, result presentation, summarization) – Predict next query (query suggestion, spelling correction) – Predict user satisfaction (market share) Eugene Agichtein, Emory University, IR Lab 11 Intent Classes (top level only) [from SIGIR 2008 Tutorial, Baeza-Yates and Jones] User intent taxonomy (Broder 2002) – Informational – want to learn about something (~40% / 65%) History nonya food – Navigational – want to go to that page (~25% / 15%) Singapore Airlines – Transactional – want to do something (web-mediated) (~35% / 20%) • Access a serviceDownloads • Shop – Gray areas Jakarta weather Kalimantan satellite images Nikon Finepix • Find a good hub Car rental • Exploratory search “see what’s there” Eugene Agichtein, Emory University, IR Lab Kuala Lumpur Search Actions • Keystrokes – query, scroll, CTRL-C, …) • GUI: All of these can be easily captured on SERP (javascript) – scrolling, button press, clicks • Mouse: – moving, scrolling, down/up, scroll • Browser: – new tab, close, back/forward Eugene Agichtein, Emory University, IR Lab 13 Problem 1: Detect Query Intent [Ashkan et al., ECIR 2009] • Query Intent Detection in multiple dimensions: Commercial, if assumed purpose of query is to make an immediate or future purchase Navigational, if assumed purpose of query is to locate a specific Website, informational otherwise • Clickthrough Calculation - Estimating the average ad clickthrough rate for each query type Eugene Agichtein, Emory University, IR Lab 14 Dataset Construction [Ashkan et al., ECIR 2009] • Microsoft adCenter Search query log ~100M search impressions ~8M ad clicks associated with the impressions • Seed: 1700 queries labeled by three researchers – Examine query, search result page (SRP) • MTurk: 3000 new queries + 1000 Seed queries – 40 batches of 100 queries, each with 25 Seed, 75 MT – If agreement < 60% reject, redo; if >75% bonus • Results after resolution: – 42% Commercial; 55% Navigational Eugene Agichtein, Emory University, IR Lab 15 Amazon Mechanical Turk Service Eugene Agichtein, Emory University, IR Lab 16 Use Support Vector Machine (SVM) Classifier • SVMs maximize the margin around the separating hyperplane. • A.k.a. large margin classifiers • The decision function is fully specified by a subset of training samples, the support vectors. • Quadratic programming problem • Seen by many as most successful current text classification method Eugene Agichtein, Emory University, IR Lab Support vectors Maximize margin 17 Features for Classification Category Query Specific Content Clickthrough Feature [Ashkan et al., ECIR 2009] Description Query length Number of characters in the query string Query segments Number of words in the query string URL-element Whether the query string contains any URL element, such as .com, .org Organic domain Total number of domains listed among the organic results of which the query string is a substring SERP Frequency of keywords extracted from the first search result page Host # Number of different target ad hosts clicked as results of the query Click per host Total number of ad clicks recorded for the query divided by Host # Top host significance Number of times a click happens on the most frequent target host as a result of the query, divided by click per host Decrease level for top two hosts Number of times a click happens on the most frequent target host divided by the number of times the second most frequent target host receives a click Average substring # Number of target hosts of which the query is a substring divided by total number of different hosts clicked for the query Substring ratio Total number of clicks on target hosts of which the query is a substring divided by total number of ad clicks for the query Deliberation time The average time between entering a query and an ad click for that query Eugene Agichtein, Emory University, IR Lab 18 Intent Classification: Results [Ashkan et al., ECIR 2009] Setting Query + SERP + Clickthrough Classifier Precision Recall Accuracy Commercial 0.90 0.89 0.83 0.94 90% 0.85 0.86 0.80 0.90 85.5% Classifier Precision Recall Accuracy Navigational 0.86 0.81 0.87 0.80 84.5% 0.83 0.79 0.84 0.81 83.7% Noncommercial Commercial Query + SERP Setting Query + SERP + Clickthrough Noncommercial Informational Navigational Query + SERP Informational Eugene Agichtein, Emory University, IR Lab 19 Clickthrough for Varying Intent [Ashkan et al., ECIR 2009] Eugene Agichtein, Emory University, IR Lab 20 Talk Outline Overview of the Emory IR Lab Intent-centric Web Search Classifying intent of a query Contextualized search intent detection Eugene Agichtein, Emory University, IR Lab 21 How Do We Know “True” User Intent? Adapted from [Daniel M. Russell, 2007] • Ask the user (surveys, field studies, pop-ups) – Does not scale, users get annoyed • Observe user actions and guess – Intent usually obvious to humans but not always • Detect signals from user’s brain (fMRI, EEEG) and attempt to interpret neuron activity Eugene Agichtein, Emory University, IR Lab 22 “Eyes are a Window to the Soul” • Eye tracking gives information Camera about search interests: – Eye position – Pupil diameter – Seekads and fixations Reading Visual Search Eugene Agichtein, Emory University, IR Lab 23 “An Eye Tracker on Every Table” • And “nuclear reactor in every back yard”… Unlikely. • Eye tracking equipment is bulky and expensive • Can we infer gaze position from observable actions? • Exploratory study from Google (Rodden et al.) says maybe: mouse position is sometimes related to eye position Eugene Agichtein, Emory University, IR Lab 24 Relationship Between Mouse and Gaze Position [K. Rodden, X. Fu, A. Aula, and I. Spiro, Eye-mouse coordination patterns on web search results pages, Extended Abstracts of ACM CHI 2008] • Searchers might use the mouse to focus reading attention, bookmark promising results, or not at all. • Behavior varies with task difficulty and user expertise Eugene Agichtein, Emory University, IR Lab 25 Assume “Transitivity” Holds • Given: – Gaze position ==> user intent and Mouse movement ==> gaze position Mouse movement ==> user intent Restate problem: Given user actions, infer current user’s intent, focusing on Individual User’s actions Eugene Agichtein, Emory University, IR Lab 26 From Query Type to Search Intent • “obama” navigational informational • Other examples: – Query bookmarks (refinding): ~40% of queries (J. Teevan et al., SIGIR 2007) – Research vs. Immediate Purchase • Incorrect to classify the query into a single intent Classify user goals for each query instance Eugene Agichtein, Emory University, IR Lab 27 Dataset Creation: EMU Train Prediction Models HTTP Server Usage Data HTTP Log Data Mining & Management • Firefox + LibX plugin • Track whitelisted sites e.g., Emory, Google, Yahoo search… • All SERP events logged (asynchronous http requests) •150 public use machines, ~5,000 opted-in users Eugene Agichtein, Emory University, IR Lab 28 EMU: Querying Behavior Data Eugene Agichtein, Emory Univesity, IR Lab 29 Playback Example Eugene Agichtein, Emory University, IR Lab 30 Problem 1: Search Intent Classification • Infer “personalized” intent {NAV, INFO, TRANSACT} for each search instance using EMU instrumentation – “obama” instance 1 NAV, but instance 2 INFO • Focus: – Contribution of client/GUI events (Mouse movements) Eugene Agichtein, Emory University, IR Lab 31 Navigational query: “facebook” Eugene Agichtein, Emory University, IR Lab 32 Informational query: “spanish wine” Eugene Agichtein, Emory University, IR Lab 33 Transactional query: “integrator” Eugene Agichtein, Emory University, IR Lab 34 Mouse Features: Simple • First representation: – Trajectory length – Horizontal range – Vertical range Eugene Agichtein, Emory University, IR Lab Horizontal range Trajectory length Vertical range 35 Mouse Features: Full • Second representation: – 5 segments: initial, early, middle, late, and end – Each segment: speed, acceleration, rotation, slope, etc. 1 2 3 4 5 Eugene Agichtein, Emory University, IR Lab 36 Learning to recover single search intent • Represent full client-side interactions with each SERP page as feature vectors • Apply standard machine learning classification methods Feature Vectors (training, test) Manual labels CSIP (using WEKA SVM, Decision Tree etc.) Eugene Agichtein, Emory University, IR Lab Predictions for test instances 37 Experimental Setup • Dataset: – Gathered from mid-January 2008 until mid-March 2008 from the public-used machines in Emory University libraries. – Consist of ~1500 initial query instances/search sessions – Randomly sample 300 initial query instances • Behavioral pattern for follow-up queries might be different Eugene Agichtein, Emory University, IR Lab 38 Creating “Truth” Labels • Use our best guess based on clues: – Query terms – Next URL (eg. clicked result) – How user behaves before click/exit Eugene Agichtein, Emory University, IR Lab 39 Intent Statistics in Labeled Sample Eugene Agichtein, Emory University, IR Lab 40 Results: Classifying Search Intent CSIP > CF >> CS > S Eugene Agichtein, Emory University, IR Lab 41 Results II: {Info/Transact} vs. Nav All improved. Still, CSIP > CF >> CS > S Eugene Agichtein, Emory University, IR Lab 42 Salient features (by Info Gain) Eugene Agichtein, Emory University, IR Lab 43 Case Studies Summary • CSIP can help identify: – Relatively rare navigational queries (re-finding queries or queries for obscure websites) – Informational queries that resemble navigational queries (coincides with a name of a website) Eugene Agichtein, Emory University, IR Lab 44 Outline • Overview of research at the Emory IR Lab • Dimensions of (commercial) search intent • Classifying intent of a query • Contextualized search intent detection Eugene Agichtein, Emory University, IR Lab 45 Informational vs. Transactional: Research vs. Purchase Intent • 10 Users (grad students and staff) asked to – 1. Search for a best deal on an item they want to purchase immediately (Purchase intent) – 2. Research a product they want to purchase eventually (Research intent) • Eye tracking and browser instrumentation performed in parallel • EyeTech sysstems TM3 (integrated) – At reasonable resolution, samples reliably at ~12-15 Hz Eugene Agichtein, Emory University, IR Lab 46 Research Intent Eugene Agichtein, Emory University, IR Lab 47 Purchase Intent Eugene Agichtein, Emory University, IR Lab 48 Relationship between behavior and intent? • Search intent is contextualized within a search session • Implication 1: model session-level state • Implication 2: improve detection based on clientside interactions Eugene Agichtein, Emory University, IR Lab 49 Contextualized Intent Inference • • • • SERP text Mouse trajectory, hovering/dynamics Scrolling Clicks Eugene Agichtein, Emory University, IR Lab 50 Model: Linear Chain CRF Eugene Agichtein, Emory University, IR Lab 51 Conditional Random Fields (CRFs) [from Lafferty, McCallum, Pereira 2001] From HMMs to MEMMs to CRFs s s1 , s2 ,...sn HMM o o1 , o2 ,...on St St+1 ... |o| P( s , o ) P( st | st 1 ) P(ot | st ) t 1 |o | MEMM St-1 P ( s | o ) P ( st | st 1 , ot ) t 1 j f j ( st , st 1 ) j 1 exp t 1 Z st 1 ,ot k g k ( st , ot ) k j f j ( st , st 1 ) |o | j 1 P( s | o ) exp Z o t 1 k g k ( st , ot ) k Ot-1 Ot St-1 Ot+1 St ... St+1 ... |o | CRF Eugene Agichtein, Emory University, IR Lab Ot-1 Ot St-1 Ot-1 Ot+1 St Ot ... St+1 ... Ot+1 ... 52 Problem 2: Search Ad Receptiveness • Hypothesis: the right time to serve any search ads: when searcher is receptive to seeing ads • Receptiveness ≈ some search intent – Commercial? (navigational or informational) – Non-commercial? – “Background” interest Eugene Agichtein, Emory University, IR Lab 53 Predict Future Ad Clicks Within Session Eugene Agichtein, Emory University, IR Lab 54 Dataset: 440 Emory College Students Eugene Agichtein, Emory University, IR Lab 55 Results: Ad Click Prediction • 200%+ precision improvement (within mission) Eugene Agichtein, Emory University, IR Lab 56 Varying Model Structure Eugene Agichtein, Emory University, IR Lab 57 Feature Analysis Eugene Agichtein, Emory University, IR Lab 58 Error Analysis: Mouse Noise Eugene Agichtein, Emory University, IR Lab 59 Within-mission intent change/frustration/digression Eugene Agichtein, Emory University, IR Lab 60 Current and Future Work • Unsupervised intent clustering • User vs. task • Personalized behavior models • Long-term interests/effects • User mental state (frustration, satisfaction, …) Eugene Agichtein, Emory University, IR Lab 61 Challenges • Separate context from intent (e.g., smart phones) • User variability: individual differences, tasks • Scale of data: representation, compression • Privacy: client-side data similar to other PII – Can be abused and must be protected • Obtaining realistic user data: see above – EMU toolbar tracking since 2007 in Emory Libraries (biased) Eugene Agichtein, Emory University, IR Lab 62 Other Application Areas • Search: ranking, evaluation, advertising, search interfaces, medical search (clinicians, patients) • Collaborative information sharing: searcher intent, success, expertise, content quality • Health informatics: self reporting of drug side effects, co-morbidity, outreach/education • Automatic cognitive diagnostics: stress, frustration, other impairments…. Eugene Agichtein, Emory University, IR Lab 63 Summary: From Behavior to State of Mind • Approach: – Machine learning methods for detecting searcher intent – Calibrated and augmented with lab studies • Foundational contributions: – Methods to mine and integrate wide range of interactions – Data-driven discovery of user state-of-mind • Impact: – Intelligent, intuitive search and information sharing Eugene Agichtein, Emory University, IR Lab 64 Main References • Classifying and Characterizing Query Intent, Azin Ashkan, Charles L. A. Clarke, Eugene Agichtein, Qi Guo, In ECIR 2009. • Qi Guo and Eugene Agichtein, Exploring Client-Side Instrumentation for Personalized Search Intent Inference: Preliminary Experiments, Proc. of AAAI 2008 Workshop on Intelligent Techniques for Web Personalization (ITWP 2008) • Qi Guo, Eugene Agichtein, Azin Ashkan and Charles L. A. Clarke: In the Mood to Click? Inferring Searcher Advertising Receptiveness, in Proc. of WI 2009 • Other papers here: http://www.mathcs.emory.edu/~eugene/publications.html Eugene Agichtein, Emory University, IR Lab 65 Thank you! • Yandex (for hosting my visit) Supported by: Eugene Agichtein, Emory University, IR Lab 66