Detecting Buzz from Time-Sequenced Document Streams

1
Detecting Buzz from Time-Sequenced Document
Streams

• Jeonghee Yi
• IBM Almaden Research Center
• EEE05

2
Introduction

• Goal detecting emerging and changing interests
  that appear in document streams (e.g. email, news
  articles, and blogs).
• We detecting buzzwords in the documents over
  time.
• Buzzwords are terms that occur with strong
  momentum for a relatively short period of time.
• Our approach for buzz detection is based on the
  notion of burst of activities proposed by
  Kleinberg.

3
Related Works

• J. Kleinberg. Bursty and hierarchical structure
  in streams
• burst event model
• Yahoo! Buzz Index
• counts the percentage of the total number of
  people searching for a specific query term
• not take into account the duration of the buzz
• R. Kumar, J. Novak, P. Raghavan, and A. Tomkins.
  On the bursty evolution of blogspace
• evolution of community in blogspace
• D. Gruhl, R. Guha, D. Liben-Nowell, and A.
  Tomkins. Information diffusion through blogspace
• information and topic propagation using blogspace
  as an example domain

4
Kleinbergs Burst Analysis

• Modeling the stream using an infinite-state
  automaton
• Burst appear ? state transition
• There is a cost associated with any state
  transition.
• Given an event stream the method finds a low cost
  state sequence that is likely to generate that
  stream.

5
Two-State Automaton

• Given a set of n1 messages D (d0, d1, , dn),
  message inter-arrival gaps X (x1, x2, , xn).
• Two states, q0 (low) and q1 (high).
• For each state
• Exponential density function f(x) ae-ax, a gt 0.
• The probability that the gap exceeds x is equal
  to e-ax.
• a is the rate of message arrivals.
• The expected value of the gap in this model is
  e-1.
• The automaton changes states with probability p ?
  (0, 1), remaining in its current state with
  probability 1-p.

6
Finding Optimal State Sequences

• Q (qi1, qi2, , qin), where qi1, qi2, , qin ?
  q0, q1
• Each Q induces a density function fQ(x1, x2, ,
  xn) ?t 1..n fit(xt)
• b number of state transitions in QPrQ
• PrQX
• Maxmizing PrQX? minimizing lnPrQX
• Finding Q that minimizes the cost function

favor small b
favor Q that conform X well
7
Burst Analysis Examples
8
Buzzword Detection

• Not all bursty events by the Kleinbergs model
  can be considered as buzz.
• Take into account the relative duration and the
  mass of the bursts
• Ii(w) high-interval of w, Ii-(w)
  low-interval of w
• mass of Ii(w), , doc containing w
  arrived during Ii(w).
• arrival rate of Ii(w),
• momentum of Ii(w),
• span ratio of Ii(w),
• concentration of Ii(w),
• w qualifies to be a buzzword for time period
  Ii(w) if

9
Buzzword Word Detection Algorithm

• 1. For each term w in the stream, compute the
  optimal state sequence Q(w) by Kleinbergs model.
• 2. For each high interval of the state sequence
  Q(w), Ii(w), compute the degree of concentration
• 3. w is a buzzword if the following holds

10
Dataset

• 1M blog pages crawled by WebFountain Crawlers.
• Data preprocessing
• Template Removal
• Page Segmentation
• Detagging HTML Tags
• Topic Page Detection
• Winnow classifier
• 29K blog postings with political contents in the
  first quarter of 2004.

11
Empirical Results (1/2)
12
Empirical Results (2/2)

• Ralph Nader
• Feb. 20 Ralph Nader would enter the 2004
  presidential race
• Feb. 22 formal announcement
• John Kerry
• Many bursts got low score of span ratio, but they
  associated with his victory of Iowa Caucus, and
  his victory on Super Tuesday.
• Our algorithm was able to detect buzz on the
  individual events.
• John Kerry can be considered as a buzzword of
  much longer span.

buzz
buzz
13
Conclusion

• We presented a formal model of buzz and proposed
  an algorithm to detect them from a text document
  stream.
• We proposed a formal method of computing the
  degree of concentration.
• Our algorithm is experimentally verified on blog
  postings and the results show that the method is
  highly promising in detecting buzz.