# Large scale link based latent Dirichlet allocation for web document classification

**Authors:** István Bíró, Jácint Szabó  
**arXiv:** [1006.4953](https://arxiv.org/abs/1006.4953)  
**Date:** 2010-06-25  
**Summary page:** https://www.summarizepaper.com/en/paper/1006.4953v1/

## Key points

- Authors explore the applicability of latent Dirichlet allocation (LDA) for classifying large collections of web documents
- Introduce a novel influence model that considers the linkage between documents, allowing topics to propagate along links
- Develop LDA-specific boosting of Gibbs samplers, resulting in significant speedup in experiments
- Boosted LDA model can be used for classification as dimensionality reduction and yields link weights for processing the web graph
- Deploy LDA link weights in stacked graphical learning as an example
- Achieve a 4% improvement over plain LDA with BayesNet and an 18% improvement over tf-idf with SVM in terms of AUC of classification
- Gibbs sampling strategies result in about 5-10 times speedup without significant decreases in accuracy measured by likelihood and AUC of classification
- Overall, the paper presents a comprehensive exploration of using LDA for classifying large web document collections, with improved performance and efficiency compared to traditional approaches such as tf-idf with SVM.

## Summary

In the paper titled "Large scale link based latent Dirichlet allocation for web document classification," authors István Bíró and Jácint Szabó explore the applicability of latent Dirichlet allocation (LDA) for classifying large collections of web documents. The authors introduce a novel influence model that takes into account the linkage between documents, providing a fully generative model of the document content. In this setup, topics propagate along links, allowing linked documents to directly influence the words in the linking document. The authors also develop LDA-specific boosting of Gibbs samplers, resulting in a significant speedup in their experiments. This boosted LDA model can be applied for classification as dimensionality reduction, similar to latent semantic indexing. Additionally, the model yields link weights that can be used in algorithms to process the web graph. As an example, the authors deploy LDA link weights in stacked graphical learning. To evaluate their approach, the authors use Weka's BayesNet classifier and compare it with plain LDA using BayesNet and tf-idf with SVM. They achieve a 4% improvement over plain LDA with BayesNet and an 18% improvement over tf-idf with SVM in terms of the AUC of classification. Furthermore, their Gibbs sampling strategies result in about 5-10 times speedup without significant decreases in accuracy measured by likelihood and AUC of classification. Overall, this paper presents a comprehensive exploration of using LDA for classifying large web document collections. The proposed influence model and boosted Gibbs samplers contribute to improved performance and efficiency compared to traditional approaches such as tf-idf with SVM.