Classification of webpages as Ephemeral or Evergreen
Download as PPTX, PDF1 like684 views
This document summarizes a project to classify web pages as either ephemeral (short-lived) or evergreen (long-lived) content. The goal is to build a classifier using models like Naive Bayes, logistic regression, SVM, and random forests. Data is scraped from websites and preprocessed using techniques like bag-of-words and TF-IDF. Initial results show the SVM and random forest models performing best, with accuracies of around 86-80% respectively. Further work involves verifying outliers, ensemble methods, and applications in recommendation systems, archival projects, and targeted advertising.
![Formula for calculating IDF
D is the set of training examples (documents)
|D| is the number of training examples
|{d ∈ D : t ∈ d}| is the number of documents where the
word t appears [6].](https://image.slidesharecdn.com/minorprojectpresentation-150215131340-conversion-gate01/85/Classification-of-webpages-as-Ephemeral-or-Evergreen-9-320.jpg)
![References
[1] Kaggle, “StumbleUpon Evergreen Classification Challenge”,
http://www.kaggle.com/c/stumbleupon
[2] T. Fawcett, “An Introduction to ROC Analysis”, Pattern Recognition Letters,
Issue 27, 2006, pp 861-874
[3] J. Ramos, “Using TF-IDF to Determine Word Relevance in Document Queries”,
ICML, 2003](https://image.slidesharecdn.com/minorprojectpresentation-150215131340-conversion-gate01/85/Classification-of-webpages-as-Ephemeral-or-Evergreen-25-320.jpg)
Classification of webpages as Ephemeral or Evergreen
- 1. Classifying Ephemeral vs. Long Lasting Content on the Web Akif Khan Yusufzai 11CSS07 Moonis Javed 11CSS40
- 2. Introduction ● Web classification is a very important machine learning problem with wide applicability in tasks such as news classification, content prioritization, focused crawling and sentiment analysis of web content. ● On the Web, classification of page content is essential to focused crawling, to the assisted development of web directories, to topic specific Web link analysis and to contextual advertising. ● Our GOAL: Study a specific instance of this broad and vital web classification problem and developing a successful prediction system.
- 3. Goal Of Project: ● Building a crawler to scrape data from different websites across the internet such as bloomberg.com, Blogs from blogger.com, Times of India news website. ● Building a classifier to categorize web pages as evergreen or ephemeral. ● Binary classes of classification: o Ephemeral (short lived) o Evergreen (Long lived)
- 4. Ephemeral Content ● Short Lived i.e it loses its relevance after a certain period of time. ● Based on current happenings or interests. ● Fades out after some time and its viewership or hit count is negligible. ● Examples o A news topic o A viral video
- 5. Long Lasting Content ● Content doesn't loses its relevance even after a very long time. ● Usually based on everlasting topics ● Example o A cooking recipe o Information about monument such as Taj Mahal, or about history.
- 6. Technical Details● Dataset from Kaggle’s ‘Stumbleupon Challenge’ ● Contains approximately 10,000 HTML documents for training and testing purpose. ● To improve our model we will scrape recent data from other websites. ● Fields o URL o Boilerplate text : contains title and body of html page. o Number of characters o Number of links o Number of words in url o User determined label (only in training) , etc.
- 7. Preprocessing ● Bag of Words Model: o Create a dictionary of all the words with their frequency of occurrence. o remove the high frequency words (filler words like the, is, etc.) and lowest frequency words as their presence doesn’t affect the prediction model. o remove the least frequency words which do not occur enough so as to be helpful for prediction.
- 8. Preprocessing Another approach… we use the Term frequency- Inverse document frequency (tf-idf) of each word. ● The TF-IDF is the product of the Term frequency ( indicating the number of times a word appears in a given document). ● Inverse document frequency: which measures how commonly the word appears across all documents.
- 9. Formula for calculating IDF D is the set of training examples (documents) |D| is the number of training examples |{d ∈ D : t ∈ d}| is the number of documents where the word t appears [6].
- 10. Classifier Models used 1. Naive Bayes Model 2. Logistic regression Analysis 3. Support Vector Machine (SVM) 4. Decision tree model - Random Forest
- 11. Naïve Bayes Model ● family of simple probabilistic classifiers ● based on applying Bayes' theorem ● strong (naive) independence assumptions between the features ● popular method for text categorization ● used with word frequencies as features
- 12. Logistic Regression ● used for predicting the outcome of a categorical dependent variable (i.e., a class label) based on one or more predictor variables (features) ● makes use of one or more predictor variables that may be either continuous or categorical data ● Binomial logistic regression used as final predictions are binary (0 - ephemeral 1 - long lasting)
- 13. Support Vector Machine (SVM) ● builds a model that assigns new examples into one category or the other, making it a non-probabilistic binary linear classifier ● used to analyze data and recognize patterns ● A set of features that describes one case (i.e., a row of predictor values) is called a vector. ● creates an optimal N-dimensional hyperplane which separates the data into two categories
- 14. Decision Tree Model - Random Forest ● ensemble learning method for classification (and regression). ● operate by constructing a multitude of decision trees at training time. ● outputting the class that is the mode of the classes output by individual trees. ● applies the general technique of bootstrap aggregating, or bagging.
- 15. Outliers
- 16. Results of Different Models SVM ● Linear SVM on Tf-Idf vectorized body 20 fold CV score : 86.8915% ● Linear SVM on Tf-Idf vectorized body after outlier removal 20 fold CV score : 87.2765%
- 17. Results of Different Models L
- 18. Results of Different Models Gaussian NB: ● 20 Fold CV Score 69.825% ● 20 Fold CV Score after Outlier Removal 70.379%
- 19. Results of Different Models Random Forest: ● 20 Fold CV Score 79.95% ● 20 Fold CV Score after Outlier Removal 80.1174%
- 20. Results of Different Models
- 21. Word Cloud of highest frequency words
- 22. Work To Be Done ● Verify Outliers using clustering methods like K-means Clustering, DBSCAN ● Build Ensemble Method by combining multiple models ● Apply AdaBoost to improve Ensemble Accuracy
- 23. Area of Application ● useful for recommenders attempting to classify different news stories based on type. ● used for archival projects to determine what web content merits inclusion. ● for content sites interested in capacity planning for hosting different pages based on expected longevity. ● used for putting up of advertisement, and the companies can bid more on ads displayed on this page as it will be visible for a longer time.
- 24. Future Scope ● Apply distributed or cloud computing to further improve accuracy and provide real time classification. ● Calculate lifetime prediction ( time for which it receives a fair amount of hits ) ● Sentiment analysis of long lasting web pages
- 25. References [1] Kaggle, “StumbleUpon Evergreen Classification Challenge”, http://www.kaggle.com/c/stumbleupon [2] T. Fawcett, “An Introduction to ROC Analysis”, Pattern Recognition Letters, Issue 27, 2006, pp 861-874 [3] J. Ramos, “Using TF-IDF to Determine Word Relevance in Document Queries”, ICML, 2003
- 26. Thank You Any Questions ?