1 / 14

Progress Report

Progress Report. ekker. Problem Definition. In cases such as object recognition , we can not include all possible objects for training . So transfer learning could deal with this kind of problem.

charis
Download Presentation

Progress Report

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Progress Report ekker

  2. Problem Definition • In cases such as object recognition , we can not include all possible objects for training . So transfer learning could deal with this kind of problem. • Here we divide the complete transfer learning into two steps: node(link) classification ,transfer to other domain.

  3. Related Solution • Graph labeling • SNARE : A Link Analytic System for Graph Labeling and Risk Detection,MaryMcGlohon et al. KDD 2009. • Markov Logic Network • Markov logic network ,Matthew Richarson,PedroDomingos,Machine Learning 2006

  4. Overview of Graph Labeling

  5. Overview of Graph Labeling Given: 1.A graph G=(V,E),V is the entities, E is the interactions between them. 2.Binary Class label X. 3. A set of flags based on node attributes G=(V,E) Output: A mapping between each node and its class label. Information about this node is inferred from its neighbors.

  6. Overview of Graph Labeling Information about this node is inferred from its neighbors. G=(V,E) node i potential Message to node i Vi edge potential from I to j Vj Upon convergence , belief scores are determined by :

  7. Overview of Markov Logic Network • Using the first-order logic to capture the relation(attributes) of data . • Using the entities(constant in predicate) and formulas build up the MLN network. • Learn the weight of each formula . • Using MLN to inference the query probability.

  8. Overview of Markov Logic Network Weights 3. 1. Constants Two constants: Anna (A) and Bob (B) 4. Using MLN to inference query , such as P(Smokes(A)=>Cancer(A)|MLN) 2. Friends(A,B) Friends(A,A) Smokes(A) Smokes(B) Friends(B,B) Cancer(A) Cancer(B) Friends(B,A)

  9. Ideas • But for MLN using the weight and first order to capture the characteristic of data. • Could we extend the graph labeling method with more generality. • In real data , the relation between nodes is not only one type and the node type is node only binary ,too. => How to do graph labeling on heterogeneous network.

  10. Recommendation over a heterogeneous Social Network • Recommendation over a heterogeneous Social Network,JinZhang,Jie,Tang, et al. , WAIM08 • This papers goal is to investigated the recommendation system on a general heterogeneous Web social network. • Browsing : do recommendation s when a person is browsing one object • Search : do recommendation of different types of object when a person searches for one type of object by query.

  11. Approach • Global importance estimation. • Similar to PageRank. • Concerned with a homogenous graph.

  12. Pair-wise learning Algorithm Build up a transition graph of the homogenous graph.

  13. Pair-wise learning Algorithm • Build up a transition matrix between each pair of two types of nodes. • For example, in the previous figure , we may have 13 transition matrixes. • Then it can using the transition probability and the transition matrix to compute the score. • But for compute the score we need to compute the transition probability.

  14. Pair-wise learning Algorithm • To learn the transition probability λ. • Using the training data A ={(i,j)} the selected pair of object of the same type which important score of i larger than j. • Try to make the importance score in random walk algorithm as in training data.

More Related