1. Deep Web Mining and Learning for Advanced Local Search
CS8803 Advisor Prof Liu
Yu Liu, Dan Hou
Zhigang Hua, Xin Sun
Yanbing Yu

2. Competitors Yahoo! Local Yelp CitySearch Google Local Yellow Page
How to beat them?

3. Research Background Deep Web Crawling Sentimental Learning
Sentimental Ranking Model
Geo-credit Ranking Model
Social Network for Businesses

4. Show Time! Local Biz Space

5. Architecture Database JDBC Query-based Crawler HTML Parser
Sentimental Learner
Super
Local-Search
Apache
Server

6. Tools Open source social network platform Elgg, OpenSocial LAMP Server
Linux+Apache+Mysql+PHP
Google Map API,
eg, Geocode,

7. Crawling Dynamic Pages

8. Crawling Dynamic Pages

9. Parsing Dynamic Pages

10. Database Design

11. Sentimental Learning

12. Sentimental Learning

13. Sentimental Learning Can we use ONE score
to show how good/ bad the store is?

14. Sentimental Learning Objective Dataset
To identify positive and negative opinions of a store
Dataset
Reviews represented by bag-of-terms
Normalized TF-IDF feature (normalized)
Two ways of sentiment representation
Simply average the scores
but “what you think good might be bad for me”
Manual labeling
1 to 5 (“least satisfied” to “most satisfied”)
consensus based
time-accuracy tradeoff

15. Dimension Reduction High dimensionality Dimension Reduction
6857 tokens
Memory limitation
Possibly under-fitting
Dimension Reduction
PCA (Principle Component Analysis)
an orthogonal linear transformation
transforms the data to a new coordinate system
retains the characteristics of the data set that contribute most to its variance
Get the most important features without losing generality

16. Principle Component Analysis
Original Dimension: 6857
Covariance Reserved: 95%
Different Granularity
Manual Labeling:
Score Averaging:
Downsampling Rate
40%
20% 4%
Result Dimension
231
135 41
Downsampling Rate
40%
20% 4%
Result Dimension
344
211 37

17. Sentimental learning Features used for sentimental learning:
Vector Space Model (reviews/comments)
Some keywords related to sentiments:
Positive: good, happy, wonderful, excellent, awesome, great, ok, nice, etc
Negative: bad, sad, ugly, outdated, shabby, stupid, wrong, awful, etc
Most words unrelated to sentiments:
e.g. buy, take, go, iPod, apple, comment, etc…
Causing noise for sentimental learning!!

18. What we do?
How to learn sentiments from a large set of features with lots of noise?
Vector Space Model: MXN (Entity-Term, e.g. 6,000X20,000)
Dimensionality reduction (PCA)
Using supervised learning for sentimental learning
Human labeling vs. Average rating
An online entity always includes many reviews with each review containing a rating
Average Rating is an alternative labeling for the entity
Manual labeling:
1 (least satisfactory) – 5 (most satisfactory)
Three persons do labeling, most-vote-adopted

19. Manual labeling vs. Average rating
Machine learning
Around 300 entities from local search, 6800 features after stop words removing and stemming
Using different SVM kernels
Avoiding overfit
Leave-one-out estimation
Nonlinearity of features
Polynomial kernel achieves best performance
Manual labeling
Training more precise
Labeling more consistent
Rate averaging
Training less precise
Rating more random
E.g. average(5, 5, 1) = 3

20. What we learned? Dimensionality reduction is necessary
Term Vector Space Model (VSM) is huge in nature
Human labeling is necessary
Sentimental learning involved subjective judge instead of objective judge.
Human rating is very random because it is not consistent across different people
More labeling data is needed
Other methods to be used:
Unsupervised learning (clustering)
Gaussian Mixture Model (an alternative to learn sentiments, while it is difficult to know the # of hidden sentiments)

21. How to use learned sentiments?
Sentimental learning can be used to improve ranking of local search
Because sentimental value represents an important metrics to evaluate the rank of an entity
Local search is influenced by the sentiment
Sentimental ranking model (SRM):
SentiRank = a*ContentSim + (1-a)*SentiValue
Empirically setting the parameter as “0.5”.
Similar to PageRank
PageRank = b*ContentSim + (1-b)*PageImportance

22. Geocoding Geocoding of Addresses
For example , the geo-center of store AA National Auto Parts
Is located at 3410 Washington St, Phoenix, AZ,85009
Using Geocode, we can get the exact latitude and longtitude
( , )
Haversine Formula of Great-circle distance:
Distance between two pairs of coordinates on sphere
= (3959 * acos( cos( radians(33.448) ) * cos( radians( lat ) ) * cos( radians( lng ) - radians(-122) ) + sin( radians( ) ) * sin(radians( lat ) ) ) )

23. Geo-Sentimental Ranking Model (GSRM)
Three Measurements
Content Similarity term-frequency
Sentimental Value sentimental learning
Geo-distance Google Map API
GSRM Ranking model

24. Example

25. Thank You !
QA time