Bias/Variance Analysis for Network Data Jennifer Neville and David - PowerPoint PPT Presentation

Bias/Variance Analysis for Network Data Jennifer Neville and David Jensen Knowledge Discovery Laboratory Knowledge Discovery Laboratory University of Massachusetts Amherst University of Massachusetts Amherst

Collective inference + + + + + + + + + – – – + + + – – – − Apply models to collectively infer – – – + + + class labels throughout network + + + + + + – – – – – – – – – − Exploit autocorrelation to + + + – – – – – – – – – improve model performance + + + – – – – – – – – – – – – – – + + + – – – – – – – − Collective SRL models + + + – – – – – – − Probabilistic relational models + + + – – – + + + (e.g., RBNs, RDNs, RMNs) + + + + + + – – – – – – − Probabilistic logic models – – – – – – – – – (e.g., BLPs, MLNs) − Adhoc collective models + + + + + + + + + – – – (e.g., pRNs, LBC) + + + – – – – – – + + + 2/13

Comparing collective models Latent group models Relational dependency networks 3/13

Comparing collective models Latent group models Relational dependency networks Why do RDNs perform poorly when few instances are labeled in test set? 3/13

Understanding RDN performance − Hypothesis − High autocorrelation → features selection chooses class label rather than observed attributes − Few labeled test set instances → identifiability problem − Gibbs sampling → increased variance − How to evaluate hypothesis? − Variance is due to collective inference procedure − Need an analysis framework that can differentiate model errors due to learning and inference 4/13

Bias/variance analysis − Conventional bias/variance analysis − Decomposes errors due to learning alone − Assumes no variation due to inference − Relational bias/variance analysis − Collective inference introduces new source of error − SRL models exhibit different types of errors − Network characteristics affect performance 5/13

Conventional bias/variance framework M 1 M 2 Model predictions Test Set Training M 3 Set Samples Models 6/13

Conventional bias/variance framework M 1 variance M 2 Model predictions _ Y* bias Y − Expected Expected error error per per instance instance Test Set Training − Decompose Decompose into into model model bias/variance bias/variance M 3 Set Samples Models 6/13

Bias/variance framework for relational data – – + M 1 + + + – – + + – – – – – – + – – – + + – – M 2 + – – – + – – + – – – – – + – – – – – + + – – + – + Model predictions + + – – – – – Fully labeled Test Set – – – Training – – M 3 – – Set – – – + Samples Models 7/13

Bias/variance framework for relational data learning bias – – + M 1 + + + – – learning variance + + – – – – – – + – – – + + – – M 2 + – – – + – – + – – – – – + – – – – – + + _ – – + – + Model predictions + + Y* Y L – – – – – Fully labeled Measure learning bias Test Set − Measure bias and and variance variance with with full full labeling labeling – – – Training – – M 3 – – Set – – – + Samples Models 7/13

Bias/variance framework for relational data – – – – – – – – – – – – + M 1 + + + – – – – – – – – – + + – – – – – – + – – – + + M 2 + – – – + – – + – – – + – – + + – – Model predictions + – + + – – – + – – – – – – – – – – – – – – – Training – – M 3 – – Set – – – + – – – – – – – Samples Models – – – Test Set Inference Runs 8/13

Bias/variance framework for relational data – – – total – – bias – – – – – – – + M 1 + + + – – – – – total – – variance – – + + – – – – – – + – – – + + M 2 + – – – + – – + – – – + – – + + _ – – Model predictions + – + + – – – + Y* Y – – – – – – – – – Measure total bias − Measure bias and and variance variance – – – – – – Training – – M 3 − Expectation over training Expectation over training and test sets test sets – – Set – – – + – – – – – – – Samples Models – – – Test Set Inference Runs 8/13

Bias/variance framework for relational data – – – learning total inference – – bias bias bias – – – – – – – + M 1 + + + – – – – – total – – variance – – + + – – – – – – + – – – + + M 2 + – – – + – – + – – – + – – + + _ _ _ – – Model predictions + – + + – – – + Y* Y* Y L Y Y – – – – – – – – – Measure total bias − Measure − Measure Measure learning bias bias and bias and and variance and variance variance variance with with full full labeling labeling – – – – – – Training – – M 3 − Expectation over training Expectation over training and test sets test sets − Measure Measure total bias bias and and variance variance – – Set – – – − Expectation over training Expectation over training and test sets test sets + – – – – – Difference: inference bias − Difference: bias and and variance variance – – Samples Models – – – Test Set Inference Runs 8/13

Synthetic data experiments − Vary group size, linkage, autocorrelation − Compare LGMs, RDNs, RMNs − Preliminary findings − LGMs: high learning bias when algorithm cannot identify underlying group structure − RDNs: high inference variance when little information seeding inference process − RMNs: high inference bias when network is densely connected or tightly clustered 9/13

Feature selection increases RDN inference variance 10/13

Feature selection increases RDN inference variance Inference Variance 10/13

Modified inference decreases variance 11/13

Improved performance on real data 12/13

Conclusions − Framework can be used to explain mechanisms behind SRL model performance − Improves understanding of model behavior − Suggests algorithmic modifications to increase performance − Future work − Extend framework (e.g., loss functions, joint estimation) − Investigate interaction effects between learning and inference errors − Real data experiments to evaluate design choices 13/13

Further information: jneville@cs.umass.edu kdl.cs.umass.edu 14/13