1. Acclimatizing Taxonomic Semantics for Hierarchical Content Categorization --- Lei Tang, Jianping Zhang and Huan Liu

2. Taxonomies and Hierarchical Models Web pages can be organized as a tree- structured taxonomy (Yahoo!, Google directory) Parental control: Web filters to block children ’ s access to undesirable web sites.  Parents want accurate content categorization of different granularity  Service providers appreciate the decision path how a blocking/non-blocking is made for fine tuning. Hierarchical Model: Exploit the taxonomy for classification strategy or loss function

3. Quality of Taxonomy Most hierarchical models use a predefined taxonomy, typically semantically sound. A librarian is often employed to construct the semantic taxonomy. Is semantically-sound taxonomy always good?  Subjectivity can result in different taxonomies  Semantics change for specific data

4. A Motivating Example Hurricane Federal Emergency Management Agency Geography Politics Normally During Katrina

5. A “Bayesian” View Stagnant nature of predefined Taxonomy (Prior Knowledge) Dynamic change of Semantics reflected in Data Data-Driven Taxonomy Inconsistent

6. “Start from Scratch” - Clustering Throw away the predefined taxonomy information, clustering based on labeled data. Two categories: divisive or hierarchical Usually require human experts to specify some parameters like the maximum height of a tree, the number of nodes in each branch, etc. Difficult to specify parameters without looking at the data

7. Optimal Hierarchy Optimal hierarchy: How to estimate the likelihood? Hierarchical model’s performance and the likelihood are positively related. Use hierarchical models’ performance statistics on validation set to gauge the likelihood. Brute-force approach to enumerate all taxonomies is infeasible.

8. Constrained Optimal Hierarchy Predefined taxonomy can help. Assumption: the optimal hierarchy is near the neighborhood of predefined taxonomy H 0 Constrained optimal hierarchy H ’ for H 0 : H ’ results from a series of elementary operations to adjust H 0 until no likelihood increase is observed.

9. Elementary Operations 1 234 56 (H1)(H1) 4 2 356 1 (H3)(H3) 7 2 34 56 1 (H4)(H4) (H2)(H2) 234 5 6 1 ‘Promote’ ‘Merge’ ‘Demote’ (All the leaf nodes remain unchanged)

10. Search in Hierarchy Space Given a predefined taxonomy, find its best constrained optimal hierarchy. Search in the hierarchy space. H0H0 H 12 H 11 H 32 H 33 H 31 H 03 H 02 H 13 H 21 H 22 H 23 H 24 H 01 H 04

11. Finding Best COH Greedy Search Follow the track with largest likelihood increase at each step to search for the best hierarchy. H0H0 H 12 H 11 H 32 H 33 H 31 H 03 H 02 H 13 H 21 H 22 H 23 H 24 H 01 H 04

12. Framework (a wrapper approach) Given: H 0, Training Data T, Validation Data V 1. Generate neighbor hierarchies for H 0, 2. For each neighbor hierarchy, train hierarchical classification models on T 3. Evaluate hierarchical classifiers on V. 4. Pick the best neighbor hierarchy as H 0 5. Repeat step 1 until no improvement

13. Hierarchy Neighbors Elementary operations can be applied to any nodes in the tree. Neighbors of a hierarchy could be huge. Most operations are repeated for evaluation. 2 1 32’ 1 3 H1H2

14. Finding Neighbors Check nodes one by one rather than all the nodes at the same time in each search step. ‘Merge’ and ‘Demote’ only consider the node most similar to the current one. Nodes at higher levels affects more for classification. Top-down traversal: Generate neighbors by performing all possible elementary operations to the shallowest node first.

15. Further consideration 2 types of top-down traversal: 1. ‘Promote’ operation only to generate neighbors 2. ‘Demote’ and ‘Merge’ operations only to generate neighbors Repeat 2-traversals procedure until no improvement. Root Geography Hurricane Politics If a node is inproperly placed under a parent, we need to ‘promote’ it first.

16. Experiment Setting 10-fold cross validation Naïve Bayes Classifier (Multinomial) Use information gain to select features Due to the scarcity of documents in each class, we use training data to validate the likelihood of a hierarchy.

17. Data Sets Data: Soc and Kids Human labeled web pages with a predefined taxonomy SocKids Classes69244 Nodes83299 Height45 Instances524815795 Vocabulary3400348115

18. Results on Soc

19. Results on Kids

20. Over-fitting? As we optimize the hierarchy just based on training data, it’s possible to over-fit the data.

21. Robust Method Instead of multiple traversals(iterations), just do 2- traversals once.

22. Conclusions Semantically sound taxonomy does not necessarily lead to intended good classification performance. Given a predefined taxonomy, we can accustom it to a data-driven taxonomy for more accurate classification Taxonomy generated by our method outperforms human-constructed taxonomy and the taxonomy generated “starting from scratch”.

23. Future work An initial work to combine “noisy” prior knowledge and data. How to implement an efficient filter model that can find a good taxonomy by exploiting the predefined taxonomy? Feature selection could alleviate the difference between taxonomies. How to use the taxonomy information for feature selection?