Improving Classification Accuracy Using Automatically Extracted Training Data Ariel Fuxman A. Kannan, A. Goldberg, R. Agrawal, P. Tsaparas, J. Shafer Search.

Improving Classification Accuracy Using Automatically Extracted Training Data

Ariel FuxmanA. Kannan, A. Goldberg, R. Agrawal,

P. Tsaparas, J. Shafer

Search LabsMicrosoft Research – Silicon Valley

Mountain View, CA

• For classification tasks, large amounts of training data can significantly improve accuracy

• How do we create large training sets?– Conventional methods of using human labelers

are expensive and do not scale

• Thesis: The Web can be used to automatically create labeled data

Web as a Source of Training Data

2

In this talk

• Validate the thesis on a task of practical importance: Retail intent identification in Web Search

• Present desirable properties of sources of labeled data

• Show how to extract labeled data from the sources

3

Importance of Retail Intent Queries

4

Just Behave: A Look at Searcher Behavior -Total U.S. MarketComScore Feb 2009Just Behave: A Look at Searcher Behavior -Total U.S. MarketComScore Feb 2009

Share of Searches(% of total search queries)

Share of Paid Clicks(% of queries leading to paid clicks)

Application of Retail Intent

5

• Provide enhanced user experience around Commerce Search

Retail intent identificationDefinition:

A query posed to a search engine has retail intent if most users who type the query have the intent to buy a tangible product

Examples :Queries with retail intent Queries without retail

intent

Zune 80 gb Medical insurance

Buy ipod Free ringtones

Digital camera lenses Digital camera history

Data Sources for Retail Intent• Sources

– Web sites of retailers (e.g., Amazon, Walmart, Buy.com)

• Training Data– Queries typed directly on search box of retailers

• Extraction from toolbar logs

7

URL in toolbar log

Desirable Properties of Web Data Sources

• Popularity– Sources should yield large amounts of data

• Orthogonality– Sources should provide training data about

different regions of the training space

• Separation– Sources should provide either positive or negative

examples of the target class, but not both

8

Popularity• Sources should yield large amounts of

data• For retail intent identification

–Web site traffic is a proxy for popularity –More traffic means more queries–Choose Web sites of retailers based on

publicly available traffic report (Hitwise)

9

Orthogonality

• For retail intent identification• Positive examples: top sites from “Departmental Stores” and

“Classified Ads” (Amazon and Craigslist)

• Negative examples: top site from “Reference” (Wikipedia)

10

• Sources should provide training data about different regions of the training space

Separation

• Training examples must unambiguously reflect the intended meaning of most users

– Example: there is a book called “World War I”, but the intent of the query is mostly non-commercial

• Can be enforced by removing groups of confusable queries from the sources

Method to Enforce Separation

• Create “groups” of positive queries• Compare the word frequency distribution of

each group against the negative class using Jensen-Shannon divergence

• Remove groups with low divergence

12

Groups for Retail Intent

• Extracting groups from the toolbar log

13

URL in toolbar log

Enforcing separation property• JS Divergence of Amazon and Craigslist with

respect to Wikipedia

See paper for experimental validation

Experiments

• Setup– Built multiple classifiers using manual and

automatically extracted labels in the training sets– Classification method: logistic regression, using

unigrams and bigrams as features– Test set: 5K queries randomly sampled from a

query log and labeled using Mechanical Turk

Automatic vs. Manual

16

Accuracy of extracted labels classifier on par with manual labels classifier

Combining Manual and Automatically Extracted

Marginally different from using only automatically extracted labels

Using Unlabeled Data

18

Performance of the automatic labels classifier is still on par with classifiers that start with manual labels and exploit unlabeled data using self-training

Conclusions

• By carefully choosing the data sources, we can extract valuable training data

• Using large amounts of automatically extracted training data, we can get classifiers that are on par with those trained with manual labels

• As future work, we would like to apply this experience to other classification tasks

19