dependency networks for collaborative filtering and data visualization uai-2000 발표 : 황규백
TRANSCRIPT
Dependency Networks for Collaborative Dependency Networks for Collaborative Filtering and Data VisualizationFiltering and Data Visualization
UAI-2000
발표 : 황규백
AbstractAbstract
Dependency networks An alternative for the Bayesian network A (cyclic) directed graph
Basic properties of dependency networks Dependency networks for collaborative filtering Dependency networks for data visualization
IntroductionIntroduction
A dependency network A collection of regression/classification models among variables
combined using Gibbs sampling Disadvantages
• Not useful for encoding a causal relationships
Advantages• Quite useful for encoding a predictive relationships
Representation of Joint DistributionRepresentation of Joint Distribution
In Bayesian networks
In dependency networks Via ordered Gibbs sampler
• Initialize each variable randomly.
• Resample each Xi according to
Theorem 1:• An ordered Gibbs sampler applied to a dependency network for X, w
here each Xi is discrete and each local distribution p(xi|pai) is positive, has a unique stationary distribution for X.
n
i iixpp1
)|()( pax
)|(),...,,,...,|( 111 iiniii xpxxxxxp pa
Conditional DistributionConditional Distribution
Gibbs sampling is used. Not so disadvantageous
Learning Not representing the causal relationships Each local distribution can be learned without regard to acyclicity
constraints.
Consistency and inconsistency Inconsistent dependency networks
• All conditional distributions are not obtainable from a single joint distribution p(x).
Theorem 2:• If a dependency network for X is consistent with a positive distributio
n p(x), then the stationary distribution defined in Theorem 1 is equal to p(x).
Other Properties of Dependency Other Properties of Dependency NetworksNetworks
Markov networks and dependency networks Theorem 3:
• The set of positive distributions consistent with a dependency network structure is equal to the set of positive distributions defined by a Markov network structure with the same adjacencies.
Defining the same distributions, however, representational forms are different.
• Potentials vs. Conditional probabilities
Minimality of the dependency network For every node Xi, and for every parent pai
j, Xi is not independent of pai
j given the remaining parents of Xi.
Theorem 4:• A minimal consistent dependency network for a positive distribution
p(x) must be bi-directional.
Learning Dependency NetworksLearning Dependency Networks
Each local distribution for Xi is simply a regression/classification model for xi with X \ {xi} as inputs. Generalized linear models, neural networks, support-vector machi
nes, …
In this paper, the decision tree was used. A simple hill-climbing approach with a Bayesian score
Collaborative FilteringCollaborative Filtering
Preferences prediction Implicit/explicit voting Binary/non-binary preferences
Bayesian network approach
In a dependency network
)0\|1( ii XXp X
)|()0\|1( iiii XpXXp paX
Datasets for Collaborative FilteringDatasets for Collaborative Filtering
MS.COM(Webpages), Nielsen(TV show), MSNBC(Stories in the site)
Evaluation Criteria and Experimental Evaluation Criteria and Experimental ProcedureProcedure
Accuracy of the list given by a predictive model
Average accuracy of a model A case in the test set
• <input set | measurement set> (randomly partitioned)
• <0, 1, 1, 0, 1, 0, | 1, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1>
ak
kk
kp
kp/
1
2)(
)()list(cfaccuracy
N
iK
k
k ik
i kp
kp
N 11
0)(
)(100)list(cfaccuracy
Results on AccuracyResults on Accuracy
Higher score indicates better performance.
Results on the Prediction TimeResults on the Prediction Time
Number of predictions per second
Results on Computational ResourcesResults on Computational Resources
Computational resources for model learning
Data VisualizationData Visualization
Predictive relationships (not causal) Bayesian networks often interfere with the visualization of such rel
ationships. Dependent or independent
Example DNViewer Media Metrix data
DNViewerDNViewer
A dependency network for Media Metrix data
DNViewer for Local DistributionDNViewer for Local Distribution
Local probability distribution
Summary and Future WorkSummary and Future Work
The dependency network defines a joint distribution for variables. is easy to learn from data. is useful for collaborative filtering and data visualization. is for conditionals.
The Bayesian network is for joint probability distribution.