-
Notifications
You must be signed in to change notification settings - Fork 0
Research papers
Mikhail Koltsov edited this page Nov 2, 2016
·
4 revisions
-
Subjectivity Classification using Machine Learning Techniques for Mining Feature-Opinion Pairs from Web Opinion Sources Subjectivity is used as a feature in one spam review research paper.
-
Survey of review spam detection using machine learning techniques
Interesting ideas:
- we can use "co-training" (semi-supervised learning): make a dataset of labeled data + big chunk of unlabeled data. Train two classifiers with distinct features (e.g. review-oriented and user-oriented) on labeled data. Make them predict unlabeled data. Look at samples that are most certainly (judging by probability of both classifiers) paid/unpaid. Add such samples to labeled dataset. Rinse and repeat;
- there were only three classifiers that proved to be performing: SVM, Naive Bayes, Logistic Regression. Other classifiers, ensembles and ideas (like Bagging and Boosting) were not researched (or maybe they proved to be weak?);
- synthetic review spam is different from "real-world" review spam;
- there are two main feature types: review-centric and reviewer-centric.
Has a comparison of previous work related to review spam classification: 3. Learning to Identify Review Spam