The predictive power of online chatter

1
The predictive power of online chatter

• D. Gruhl, R. Guha, R. Kumar, J. Novak, A. Tomkins

Presented by R. Paiu
2
Motivation Methodology

• Goal
• Demonstrate that online blog postings can be used
  to predict spikes in sales ranks
• Methodology
• Blog postings ?? Amazon books sales
• Create hand-crafted queries and use them for
  predictions
• Use queries to discover blog posts discussing a
  given product
• Plot sales ranks and of blog posts as two time
  series
• Implement and test automated query generation
• Develop automated prediction algorithms

3
Data Sources

• IBM WebFountain Project
• Contains hundreds of thousands of postings
• Data sources include 300,000 blogs and 200,000
  articles
• Approximately 200,000 new blog postings stored
  daily
• Amazon.com Sales Rank Data
• Collected rank data for 120 days (July October
  2004) for all books in top 300
• In total, 2,430 books were included in the data
  set

4
Correlation of Blog Posts and Sales Rank

• Steps
• Detecting spikes in sales ranks
• Locating blog mentions about corresponding book
• Plotting and correlating time series

5
Detecting Spikes in Sales Ranks (step 1)

• Minimal rank m must occur more than 2 weeks from
  the start and the end of the considered period
  (120 days)
• One week before and after m occurs, the rank
  value must be greater than max(m50, 1.5m)
• 50 books (out of 2430) contain a spike

6
Locating Blog Mentions about Books (Step 2)

• Retrieve blog mentions about the 50 books
• Hand-crafted queries produced by users
• Iteratively re-fined until results seem to match
  the desired book

7
Plotting and Correlating Time Series (Step 3)

• Plot the sales rank of the product and the number
  of corresponding postings vs. time
• y-axis is scaled so that values are 01
• Best lag k where cross-correlation is maximum
• Best lag
• Leading, if negative
• Trailing, if positive
• Results
• Sales rank spikes potentially
• correlated with blog activity
• Spikes in sales rank may occur
• despite insignificant blog activity
• Causes include discount pricing,
• bulk buying of books, etc.
• Highly correlated spikes in sales ranks and
• blog mentions were obtained for 10 books

8
Queries Automatic Generation

• Generation algorithm
• Based on the names of the authors of the books
• 1900 US Census used for estimating the names
  ambiguity
• Experimental results
• 182 automated queries
• generated
• 45 of queries have
• cross-correlation gt 0.5
• Cross-correlations gt 0.5
• correspond to strong
• correlations between blog
• mentions and sales rank
• 35 of those queries have
• best lag lt 2 weeks
• Blog posts are more likely
• leading rank spikes than
• trailing them

9
Sales Rank Prediction

• Given the time series representing sales data up
  to a point t, does the addition of blog mention
  data for the same period improve the prediction
  of what will happen to the sales rank?
• Outcomes
• MOTION VOLATILITY Predicting whether
  tomorrows sales rank for a particular book will
  be higher or lower appears to be hard
• SPIKES Analysis of blog data up to a point t
  allows to effectively predict when there will be
  a future spike in the sales rank, without
  recourse to information from the future and even
  without recourse to history of sales ranks.

10
Predicting Motion Volatility

• Analysis of various predictor algorithms
• Volatility prediction predict if tomorrows
  sales rank would differ from todays by more than
  a certain threshold value.

11
Predicting Spikes 1

• Original data is processed as follows (thruthed
  data)
• The point of the minimal sales rank is located, m
• A threshold is set,
• Spikes region is taken to be the maximum
  interval containing the point of minimal sales
  rank and no point of sales rank greater than
• Steps
• A product and a time t are fixed
• The predictor is given as input the number of
  blog postings for this product, for all days up
  to and including t
• The predictor outputs a bit indicating whether it
  believes a spike in sales rank will occur in the
  near future
• Results of the predictor are evaluated against
  the truthed data

12
Predicting Spikes 2

• Goals
• Find spikes that appear to be biggest ever
• Find spikes that significantly exceed historical
  averages
• Find spikes that rise relatively quickly

13
Predicting Spikes 3

• Evaluation
• Four categories
• Leading a spikes occurs after t, but within 2
  weeks
• Trailing a spike occurred within the past 2
  weeks
• Inside a spike is currently occurring
• Incorrect a spike does not occur within 2 weeks
  of t
• 39 predictions made
• 2/3 of them (26) were leading or trailing
• No indication in the paper regarding accuracy of
  predictions!

14
Conclusions Future Work

• Conclusions
• Rank motion and volatility prediction is
  difficult
• The volume of blog postings related to a product
  can be used to predict sales rank spikes
• Automatic query generation for detecting blog
  post mentions of products produces good results
• Future work
• Develop better understanding of when blogs are
  useful for prediction
• Create new tool features and new techniques for
  automated query generation and prediction
• Expand the area of prediction to other product
  sales ranks, electoral voting, and public opinion
  on important decisions
• Study and analyze proposed Hidden Markov Model
  relating discussion involvement of the blogger to
  sales prediction lag