[Under Review] This repository contains the code for the paper submitted for KR 2024.
We suggest to create a conda environment:
conda create -n mtlearn python=3.8 numpy
conda activate mtlearnIn this paper, we have tried the other two different kinds of Datalog Rule Learner, whose information are provided as follows:
| Baselines | Code |
|---|---|
| AMIE+ (Luis Galárraga et al., 2015) | https://github.com/dig-team/amie |
| Popper(A Cropper et al., 2020) | https://github.com/logic-and-learning-lab/Popper |
The three benchmarks (icews14, icews18 and icews0515) for the extrapolation experiments could be downloaded from here; and the three benchmarks (icews14I and icews0515I) for the interpolation experiments could be downloaded from here; and tLUBM and weather benchmarks are contructed following the steps described in here, in which the 46 manually contructed rules are provided in the current repo (/programs/tLUBM.txt). For the iTemporal(1-5), we provide their configurations, so that the rules and the datasets could be generated by loading the configuration files into the iTemporal Web Interface. All downloaded datasets should be put in the data folder.
Before running, the user should preprocess datasets. First, we need to preprocess the data and make them suitable as the input for the Datalog rule learner. Assume you have temporal benchmark stored in data/extrapolation/icews14, which contains train.txt, valid.txt and test.txt, each of file comes as a set of quaduple，e.g., anne visit US 2019, and the separator can be blank or tab, then we run the following command to handle the data
python data_converter.py --data_dir data/extrapolation/icews14 --window_size 8where the two required arguments 'data_dir' and 'window_size' represent the data path and the window size we want to split. After this, we can see three folders~(train, valid and test) are created under 'data/YAGO', which contain the processed results.
Next, we can use the Datalog rule learner to learn the static rules, and we can run the following command
python generate_rules.py --data_dir data/extrapolation/icews14After this, we can see a new folder 'Datalog', which obtain many files containing 'Datalog' rules together with their confidence scores, and the number of files equals to the number of windows.
Next, we need to conver the Datalog rules into DatalogMTL, and we can run the following command
python rule_converter.py -data_dir data/extrapolation/icews14After this, we can see a new folder 'DatalogMTL', which obtain many files containing 'DatalogMTL' rules together with their confidence scores, and the number of files equals to the number of windows.
Finally, we use the specified scoring strategy to generate a final set of DatalogMTL rules.
python scoring.py --data_dir data/extrapolation/icews14 --strategy maximumwhere --strategy could be any of the following choices:
- maximum
- weighted maximum
- average
- weighted average
- meteor
In particular, if you use the 'meteor' scoring strategy, you can specify another argument --beta (the default value is 0.5).
Now, we are ready to evaluate
python evaluate.py --data_dir data/extrapolation/icews14 --rulefile_path data/extrapolation/icews14/DatalogMTL.txt --window_size 8 where 'datafile_path' denotes the path of the evaluated temporal dataset file, the 'rulefile_path' represents the path of our generated DatalogMTL rule file.
In particular, if you want to use 'RQ' evaluation metric, you can add the two additional arguments: '-RQ' and '--goldrules_path ', where the argument 'goldrules_path' denotes the path to the benchmark DatalogMTL rule.