EXTRA (EXplanaTion RAnking) Datasets

Papers

• Lei Li, Yongfeng Zhang, Li Chen. EXTRA: Explanation Ranking Datasets for Explainable Recommendation. SIGIR'21 Resource.
• Lei Li, Yongfeng Zhang, Li Chen. On the Relationship between Explanation and Recommendation: Learning to Rank Explanations for Improved Performance. ACM Transactions on Intelligent Systems and Technology (TIST), 2022.

Datasets to download

• Amazon Movies & TV
• TripAdvisor Hong Kong
• Yelp 2019

If you are interested in how to build models on the datasets, please refer to BPER.

Data format

• IDs.pickle can be loaded via the pickle package as a python list, where each record is a python dict in the form of

{'user': '7B555091EC0818119062CF726B9EF5FF',  # str
'item': '1068719',  # str
'rating': 5,  # int, not important to the ranking task
'time': '2017-05-06', # str in the format of YYYY-MM-DD, not available on TripAdvisor
'exp_idx': ['34', '85'],  # a list of str, they are the indices of explanations after sentence grouping via LSH
'oexp_idx': ['91', '15']}  # a list of str, they are the indices of original sentences, corresponding to senID in the following

• Open id2exp.json via a text editor, e.g., Sublime, if you are curious about what the explanation indices correspond to. Or you can load it via testing.py by updating the parameters (line 5-7).

---

• IDs.txt and id2exp.txt are compatible with IDs.pickle and id2exp.json. It would be easier to check the content with plain text files.
• Each line in IDs.txt is in the format of userID::itemID::rating::timestamp::expID:expID::senID:senID, where timestamp is not available on TripAdvisor, and expID/senID are separated by ":" when there are multiple explanations.
• Each line in id2exp.txt is in the format of expID::explanation sentence.
• You can load the two files via movielens_load.py by updating the paths (line 1-2).

---

• Folders named 1, 2, 3, 4 and 5 are data splits.
• Each folder contains train.index and test.index which indicate the indices of their records in the list of IDs.pickle/IDs.txt.
• train.index/test.index contain a line of numbers (indices), e.g., 5 8 9 10.

Creation steps

• Run the scripts in the following order:
• Modify format_amazon.py, including the path (line 6, 7), the keys (line 13, 14, 17, 19-23), and how you load and iterate over each review (line 10) if your data are stored in other data formats, such as JSON, Excel or TXT. Remove line 13-16, if your dataset has no summary or tip, which is meant to include as much textual data as possible. This script is not limited to Amazon datasets. You can modify it to apply to other datasets.
• Update the paths (line 6, 7) in process_sentence.py.
• Update the parameters (line 6-10) in group_sentence.py.
• Update the paths (line 5-9) in keep_valid.py.
• Update the paths (line 6-9) in movielens.py, if you want to process the data into the MovieLens format.

Friendly reminder

• Run the program on a machine with sufficient memory
• Creating the datasets may take some time (e.g., hours for Yelp)

Code dependencies

• Python 3.6
• NLTK
• Datasketch

Citations

@inproceedings{SIGIR21-EXTRA,
	title={EXTRA: Explanation Ranking Datasets for Explainable Recommendation},
	author={Li, Lei and Zhang, Yongfeng and Chen, Li},
	booktitle={SIGIR},
	year={2021}
}
@article{TIST22-BPER,
	title={On the Relationship between Explanation and Recommendation: Learning to Rank Explanations for Improved Performance},
	author={Li, Lei and Zhang, Yongfeng and Chen, Li},
	journal={ACM Transactions on Intelligent Systems and Technology (TIST)},
	year={2022}
}