Wasserstein Logistic Regression with Mixed Features
Authors: Aras Selvi, Mohammad Reza Belbasi, Martin Haugh, Wolfram Wiesemann
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that our method outperforms both the unregularized and the regularized logistic regression on categorical as well as mixed-feature benchmark instances. (Abstract) and We report numerical results in Section 4. (Intro) and Figure 1: Left: Estimates of β for the standard logistic regression... All results are reported as averages over 2,000 statistically independent runs. (Section 2.1) and Figure 2: Runtime comparison between our column-and-constraint scheme and a naïve solution of problem (4) as a monolithic exponential conic program. (Section 4.1). |
| Researcher Affiliation | Academia | Aras Selvi Mohammad Reza Belbasi Martin B. Haugh Wolfram Wiesemann Imperial College Business School, Imperial College London, United Kingdom {a.selvi19, r.belbasi21, m.haugh, ww}@imperial.ac.uk |
| Pseudocode | Yes | Algorithm 1 Column-and-Constraint Generation Scheme for Problem (4). and Algorithm 2 Identification of Most Violated Constraints in the Reduced Problem (4). (Section 3) |
| Open Source Code | Yes | All source codes and detailed results are available on Git Hub (https://github.com/selvi-aras/Wasserstein LR). (Section 4) |
| Open Datasets | Yes | on the 14 most popular UCI data sets that only contain categorical features having more than 30 rows [11] (varying licenses). (Section 4.2) |
| Dataset Splits | Yes | All results are reported as means over 100 random training set-test set splits (80%:20%). and The radius ϵ {...} as well as the Lasso penalty γ {...} are selected via 5-fold crossvalidation. (Section 4.2) |
| Hardware Specification | Yes | All algorithms were implemented in Julia [5] (MIT license) and executed on Intel Xeon 2.66GHz processors with 8GB memory in single-core mode. (Section 4) |
| Software Dependencies | Yes | All algorithms were implemented in Julia [5] (MIT license)... We use MOSEK 9.3 [26] (commercial) to solve all exponential conic programs through Ju MP [12] (MPL2 License). (Section 4) |
| Experiment Setup | Yes | The radius ϵ {0, 10 5, . . . , 10 4, . . . , 1} of the Wasserstein ball as well as the Lasso penalty γ {0, 1/2 10 5, . . . , 1/2 10 4, . . . , 1/2} are selected via 5-fold crossvalidation. We consider two variants of our DR logistic regression that employ a different output label weight (κ = 1 vs. κ = m) in the ground metric (cf. Definition 2). (Section 4.2) |