An Integer Linear Programming Framework for Mining Constraints from Data
Authors: Tao Meng, Kai-Wei Chang
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We verify the proposed constraint mining algorithm in various synthetic and real-world applications and demonstrate that the proposed approach successfully identifies the feasible set at scale. In particular, we show that our approach can learn to solve 9x9 Sudoku puzzles and minimal spanning tree problems from examples without providing the underlying rules. Our algorithm can also integrate with a neural network model to learn the hierarchical label structure of a multi-label classification task. |
| Researcher Affiliation | Academia | 1Department of Computer Science, University of California, Los Angeles, USA. Correspondence to: Tao Meng <tmeng@cs.ucla.edu>, Kai-Wei Chang <kwchang@cs.ucla.edu>. |
| Pseudocode | No | The paper describes its methods in narrative text and does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | The source code and data are available at https://github.com/uclanlp/ILPLearning. |
| Open Datasets | Yes | We experiment on the dataset introduced in Wang et al. (2019). The dataset contains 9, 000 training and 1, 000 test samples, each of which has a unique solution. ... We conduct experiments on Im CLEF07A (Dimitrovski et al., 2011), which contains 10, 000 training samples and 1, 006 test samples. ... All the date and code can be found in https://github.com/Mt Some Three/ILPLearning. |
| Dataset Splits | Yes | We experiment on the dataset introduced in Wang et al. (2019). The dataset contains 9, 000 training and 1, 000 test samples, each of which has a unique solution. ... We generate a dataset with 7 nodes. The dataset contains 20, 000 training and 500 test data. ... We conduct experiments on Im CLEF07A (Dimitrovski et al., 2011), which contains 10, 000 training samples and 1, 006 test samples. ... to tune the neural network, we generate another 500 dev data points. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU models, or cloud instance types) used for running the experiments. It only mentions using 'Gurobi v8.1.1 (Gurobi Optimization, 2019) as the ILP solver'. |
| Software Dependencies | Yes | For all the experiments, we use the Gurobi v8.1.1 (Gurobi Optimization, 2019) as the ILP solver. |
| Experiment Setup | Yes | MST Experiments: ...3-layer feedforward neural network as the baseline model with Re LU activation. The hidden dimension is set to be 50. The input and output dimension is 21. We use the sigmoid function to regularize the output in ( 1, 1). We train the model for 300 epochs and we use the Adam optimizer(Kingma & Ba, 2015) to optimize the model. The learning rate is set to be 0.001. ... Hierarchical Multi-label Classification Experiments: ...We enumerate the number of layers in {1, 2, 3}, the number of training epochs in {1, 5, 50, 300}., and the learning rate in {0.001, 0.0003, 0.0001}. The hidden dimension is set to be 100. The input dimension is 80 and the output dimension is 96. In hidden layer we use Re LU as the activation and in output we use sigmoid function to regularize the output. |