Download presentation
Presentation is loading. Please wait.
Published byJennifer Hart Modified over 9 years ago
1
Amazon review utility estimator
2
Overview Goal: To determine the “usefulness” of Amazon.com reviews Using Mallet classifiers Several custom features If accurate, this system could be applied beyond Amazon, including other product reviews or even Slashdot/Digg comments.
3
Reviews Used Amazon ECS: Collected large number of reviews over 4 categories: Textbooks, Digital Cameras, Music, DVD Textbooks: 24,419 reviews with over 5 votes Digital Cameras: 22,566 Music: 43,328 DVD: 132,208
4
Regression? All of the length features seem to have a trend when grouped in buckets DVD data Avg TotalAvg WordAvg Para 0-25%133.965.581.84 26-50%197.045.662.33 51-75%248.725.682.79 76-100%281.665.722.84
5
Regression R 2 ~.3 Rating # of words
6
Regression Rating Avg Sentence Length
7
Features Bag of words Average: length, sentence length, word length % of words that are stop words # of spelling errors # of paragraphs Pronouns, articles, Proper nouns etc. Punctuation History
8
Stuff We Learned Some good reviews are hard to find “e-toys has this for 19.99” rated helpful by 17/21 people. And some people are just stupid “and there you have it. That's the secret. ” 77%... “On DVD, I'll buy this NOW! Not on VHS...Jezus...” 78%... We attempted manually classifying ~100 reviews In 4 buckets around 30% accuracy In 2 buckets around 55%.... abstract.cs.washington.edu/~kylej1/quiz.php
9
Cont. Trade off between Precision and Recall: Many features increase precision but hurt recall The range of good reviews is very broad Word Count / Sentence Length / % stopwords have biggest impact Precision +5%, Recall -8% Diminishing returns..
10
Cont. Precision in the High 80s with the right combination of features Recall suffers, drops to between 40-50% Experimenting with multiple classifiers in series. To boost recall without destroying precision Similar to Boosting.
11
Future When should computer override customer rating? Amazon has huge # of “Labeled” data…but the labels are sometimes poor Review Quality is very subjective Weight based on # of total votes? ○ Some concerns with this Bias detection Positive or Negative impact?
12
End Questions?
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.