1 / 35

AI on the Web, Part I

AI on the Web, Part I. CS592 Class Spring 2000. Part I: Personalized Prediction. NewsDude @ UCI [*]. [*] D.Billsus, M.Pazzani. A Hybrid User Model for News Story Classification . Proc. In UM99, Banff, Canada, June99.

holzer
Download Presentation

AI on the Web, Part I

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. AI on the Web, Part I CS592 Class Spring 2000

  2. Part I: Personalized Prediction

  3. NewsDude @ UCI[*] [*] D.Billsus, M.Pazzani. A Hybrid User Model for News Story Classification. Proc. In UM99, Banff, Canada, June99. • Intelligent agent compiles a daily news program for individual user (info retrieval) • Architecture: How it works? • Short-term vs. Long-term models for user modeling • Time-coded feedback to increase prediction accuracy

  4. NewsDude Architecture

  5. Learning Models • short-term model (NN) • news threads for ongoing recent events • long-term model (Naïve Bayes Classifier) • general news preferences • hybrid • Use short-term model • Use long-term model • Assign default score

  6. Time-coded feedback • Use the amount of time a user has listened to a story as implicit feedback • User’s direct binary feedback + Time-coded feedback = Fine-grained scale w/out extra burden on user (similar to Lieberman’s Letizia) • pl = proportion of a story user has heard • If story was rated as uninteresting: score = 0.3 * pl • If story was rated as interesting: score = 0.7 + 0.3*pl • If user asked for more information: score = 1.0

  7. NewsDude evaluation

  8. NewsDude: Strengths and Limitations • Tracks user’s changing interests in real-time without sacrificing general interests • Simple feedback but accurate prediction • Rate enough before personalizing • Not flexible: recalculate classifier if adding new keywords • Similar systems: GroupLen (U.Minnesota)

  9. Adaptive Web Sites(Mike Perkowitz and Oren Etzioni, IJCAI97) • Web sites that improve their organization and presentation by learning user access patterns.

  10. Issues in Adaptive Web Sites Design • Observation -- Discover User’s Interests • Special user’s interests • Group users’ interests • All users’ interests • Adjustment -- Adjust the original web-page design according to the observation • Customization (by link, by keyword reordering) • Optimization (by refining search results)

  11. Adaptive Web Sites: Two Examples • WebWatcher • Modify the original design by promotion, demotion and highlighting of links, and by linking web pages • Constructed by Carnegie Mellon University • http://www.cs.cmu.edu/~webwatcher/ • PageGather • Generate new web pages: Index Page Synthesis • Constructed by University of Washington, Seattle • http://www.cs.washington.edu/research/adaptive/

  12. WebWatcher(T. Joachims, D. Freitag, T. Mitchell, IJCAI97) • A software agent acts as a tour guide for web visitor; • Making suggestions on where to go next; • Learning from information provided when users enter and exit the web page, and also learns user’s access patterns. • Reorganizes web pages for user

  13. WebWatcher: architecture • A User • Requests • WebWatcher commands • Highlight advice • Replaced URLs • WebWatcher • World Wide Web • WebWatcher: a proxy agent between users and WWW

  14. Learning from Previous Tours • Users provide keywords of interest before the tour starts; • Those key words are added to the descriptions of every hyperlink this user follows; • Interests and hyperlink descriptions are represented by high-dimensional feature vectors. Their elements are calculated by using TF-IDF heuristic; • LinkQuality = Evaluation of probability that a user follow this hyperlink • estimated as the average similarity of k highest ranked keywords associated with the hyperlink.

  15. WebWatcher: TFIDF • WebWatcher uses the TFIDF with cosine similarity measure to calculate the current user’s similarity to hyperlink description • TFIDF calculates feature vector V as follows: • Vi = Freq(Word i) * [ log2(n) - log2(DocFreq(Word i)) ] • Freq(Word i) : the number of occurrences of Word i in this page • DocFreq(Word i) : the number of pages Word i appears • n is the total number of pages

  16. WebWatcher: Reinforcement Learning • Reinforcement learning: learn control strategies that select optimal actions • R(s): reward function at state s • Q(s, a): the goodness of action a in state s • Q(s,a) = R(s) +  * max { Q(s’, a’) | a’} • s is the current state, s’ is the next state through a, •  (0   < 1 ): a discount factor that determines how severely to discount the value of rewards received further into the future.

  17. Reinforcement Learning in WebWatcher • States correspond to Web pages • Actions correspond to Hyperlinks • R keyword (s): the TFIDF value of the keyword for page s • Q keyword (s,a) will be learned as the sum of discounted TFIDF value of keyword over the optimal tour beginning with a. • For every word w, WebWatcher uses a separate reward function R w(s) and learns a distinct Q w(s,a).

  18. WebWatcher: An Example of state space • 0.9 • 0.81 • 1 • 0.73 • S • 0.9 • R=1 • 0.73 • 0.81 • Initially, R =0 except destination • R = 1 at destination web page •  = 0.9

  19. PageGather(Mike Perkowitz and Oren Etzioni, IJCAI97, 99) • Index Page Synthesis • Instead of modifying the original web page design, PageGather create new index pages that contain collections of links related but currently unlinked pages. • Based on cluster mining to find collections of related pages.

  20. Cluster Mining: co-occurrence frequencies • For each pair of pages P1 and P2, compute: • Pr(P1|P2) the probability of visiting P1 if P2 is visited • Co-occurrence frequency between P1and P2 is the minimum of Pr(P1|P2) and Pr(P2|P1) • Co-occurrence frequency is zero if these two pages are already linked. • Compute a Similarity matrix • Apply a threshold and set low similarities to zero

  21. PageGather Algorithm • Process the access log into visit data. • Compute the co-occurrence frequencies between pages and create a similarity matrix. • Create the graph corresponding to the matrix, and find cliques (or connected components) in the graph. • For each cluster found, create a web page consisting of links to the documents in the cluster.

  22. Next Web Document Prediction • Papers by Albrecht, Zukerman and Nicholson • “Predicting User’s Requests on the WWW”, UM99 • “Pre-sending Documents on the WWW”, IJCAI99 • Theme: use Markov Models to predict the next document requested, and pre-send it

  23. Prediction Models • Prediction models are of the form • P(DR1, TR1 | previous requests) • Assumptions • distribution of the time for requesting a document is independent of the actual document • the next document depends only on the previous document • the time of the next request depends only on the time of the last request

  24. Prediction Models (Cont...) • From these assumptions, we can derive • P(DR1, TR1 | previous requests) = P(DR1 | previous documents) x P(TR1|TR) • Need to estimate the value of each of the two terms in the above equation

  25. Request Time Prediction

  26. Document Prediction • Four models are used for prediction • Time Markov Model • Space Markov Model • Second-order Time Markov Model • Linked Space-Time Markov Model • Graphical representations are used to represent each document prediction model

  27. Document Prediction (Cont…) • If a document Di is request after an event Ei-1, then there is an arc between them • For the Time Markov Model, Ei-1 is the last document reuqest (Di-1) • For the Space Markov Model, Ei-1 is the referring document of Di

  28. Document Prediction (Cont…) • For the Second-order Time Markov Model, Ei-1 is a tuple which contains the last two documents requested • For the Linked Space-Time Model, Ei-1 is a tuple that contains the last document requested and its referer

  29. Document Prediction (Cont…) • Each arc from event Ei-1 to Di has an associated weight w(Ei-1,Di) which is the frequency of an event-document pair across all training sessions • The probability of the request is then

  30. Document Prediction (Cont…)

  31. Hybrid Prediction Models • MaxHybrid Model • Consults all the Markov prediction models and selects the one with the highest probability in its most likely prediction • OrderedHybrid Model • Orders the Markov models according to their performance: Linked, Second-order, Time, and Space. Selects the first one that can make a prediction

  32. Hybrid Prediction Models(Cont…) • SpaceLinkedHybrid Model • If the maximum prediction made by the Space Markov Model is > 0.77, then use its prediction. Otherwise, use those of the Linked Markov Model

  33. Results • The experimental data is 50 days of server log in the form of {client,referer,requestedDoc,time,size} • Prediction modesl were assessed in terms of the probability with which they predict the actual next request

  34. Results (Cont…)

  35. Pre-sending documents • IJCAI 99 Paper (same data set) • including two costs • cost of waiting for a documents (cost-per-second) • cost of transmitting a document (cost-per-byte) • Calculate the expected benefit using document probabilities: • Expected-Benefit = Expected-Wait-Reduction - Expected-Total-Cost • Result: Pre-sending with an 8-hr cache best!

More Related