Partial Substitute tong-wang@uiowa.edu Poster #67 A black-box - PowerPoint PPT Presentation

Tong Wang Gaining Free or Low-Cost University of Iowa Transparency with Interpretable Tippie College of Business Partial Substitute tong-wang@uiowa.edu

Poster #67 A black-box model + High predictive performance - non-interpretable An interpretable model A hybrid of both? + interpretable - lower predictive performance

Poster #67 A black-box model + High predictive performance - non-interpretable An interpretable model A hybrid of both? + interpretable - lower predictive performance A key observation: there might exist a subspace where a black-box is overkill and a simple interpretable model can perform just as well as the black-box

Poster #67 A black-box model An interpretable model A hybrid of both? + High predictive performance An effective trade-off + interpretable A key observation: there might exist a subspace where a black-box is overkill and a simple interpretable model can perform just as well as the black-box The proposed solution: to substitute the black-box model with an interpretable model, where there is no or low- cost of predictive performance

Poster #67 A black-box model An interpretable model A hybrid of both? + High predictive performance An effective trade-off + interpretable A key observation: there might exist a subspace where a black-box is overkill and a simple interpretable model can perform just as well as the black-box The proposed solution: to substitute the black-box model with an interpretable model, where there is no or low- cost of predictive performance Predicted by a black-box model + - Predicted by an interpretable model Predicted by an interpretable model

Poster #67 A A hybrid rid pr predi dict ctive mod odel Define transparency of model: 𝐸 𝑗 𝐸

Poster #67 A hybrid A rid Learning Objective pr predi dict ctive • Predictive performance mod odel • Interpretability of 𝑔 𝑗 Define transparency of model: 𝐸 𝑗 • Transparency 𝐸

Poster #67 A hybrid A rid Learning Objective pr predi dict ctive • Predictive performance mod odel • Interpretability of 𝑔 𝑗 Define transparency of model: 𝐸 𝑗 • Transparency 𝐸 A Hybrid Rule Set Poster #67

Poster #67 Model Training 𝑜 { 𝑦 𝑗 } 𝑗=1 training data 𝑜 { (𝑦 𝑗 , 𝑧 𝑗 )} 𝑗=1 Any pre-trained Training and black box Predicted labels algorithm 𝑜 classifier { ො 𝑧 𝑐𝑗 } 𝑗=1 Stochastic Local Search 𝑜 { ො 𝑧 𝑐𝑗 } 𝑗=1 Input of the based algorithm (see the training algorithm paper for more details)

Poster #67 Evaluation: An efficient frontier that characterizes the trade-off between transparency and accuracy

Poster #67 Performance on Juvenile dataset accuracy transparency

Thank you! Poster #67 in Pacific Ballroom