change datasets

TACM12K/README.md

@@ -1,48 +1,65 @@
 # TACM12K
 
-**Table-ACM12K (TACM12K)** is a relational table dataset reorganized from the ACM heterogeneous graph dataset[1]. It includes four tables: papers, authors, Paper-Author relationships, and citation relationships. The paper features include year, title, and abstract, while author features include name and affiliation. Additionally, we label the papers based on the conferences in which they were published.
+## Overview
 
-## Processed
+**Table-ACM12K (TACM12K)** is a relational table dataset reorganized from the ACM heterogeneous graph dataset [1][2]. It comprises four tables: papers, authors, writings relationships, and citations relationships. The "papers" table and the "authors" table contain several key attributes related to papers and authors, respectively. The "writings" table captures the authorship relationships between authors and papers, while the "citations" table represents the citation relationships between papers. The new version contains 12,499 papers, 17,431 authors, 37,055 writings, and 37,055 citations.
 
-First, we created the Author table by extracting the original authors' IDs and names from the file. We added the firm information to each author based on the 'AvsF' data, resulting in the Author table (17431 entries).
+## Data Processing
 
-Next, we worked on the Paper table, manually annotating the year information for the venues. Using the 'PvsV' matrix, we added the year attribute to the Paper entries. Due to mislabeling in the original data where the STOC conference was incorrectly marked as the COLT conference in the "PvsC" matrix, we corrected this in the "VvsC" matrix. We then recalculated the PvsV * VvsC to obtain the corrected PvsC matrix, ultimately assigning the conference attribute to each paper. The number of papers per conference is shown in Table 1.
+First, the "authors" table was constructed by extracting the original author IDs and names from the dataset. Subsequently, firm information for each author was incorporated using the "AvsF" matrix, leading to the complete formation of the "authors" table.
 
-The original dataset contained text information for each paper as a concatenation of title and abstract. To make this structure clearer, we utilized three methods to extract the title from the original text information:
-1. Use the scholarly library in Python to accurately query the last two sentences of the original text.
-2. Utilize the Tongyi Qianwen language model to extract the title from the original text.
-3. Match the abstract in the Aminer dataset [tj2008Aminer] to obtain the title.
+Next, the "papers" table was created by manually labeling the year information for each venue. Utilizing the "PvsV" matrix, the year attribute was successfully added to the entries in the "papers" table. During this process, a mislabeling issue was identified in the original data, where the STOC conference had been mistakenly labeled as COLT in the "PvsC" matrix. This error was corrected in the "VvsC" matrix, and a corrected "PvsC" matrix was generated by recalculating the product of the "PvsV" and "VvsC" matrices.
 
-We then compared the results of these three methods. If two results were the same, we considered it the title. If all three results differed, we manually verified and supplemented the title information. This process allowed us to separate the original text into a distinct title and abstract. Additionally, we supplemented some incomplete abstracts.
+The original dataset contained concatenated text for each paper, combining the title and abstract. To clarify this structure and differentiate the title from the abstract, three methods of extraction were applied:
 
-The Write table inherited the "PvsA" matrix from the original data. The Cite table inherited the "PvsP" matrix from the original data.
+1. The scholarly library [3] in Python was used to query the last two sentences of the original text accurately.
+2. The Tongyi Qianwen [4] large language model was employed to extract the title.
+3. The abstract from the Aminer dataset [5] was matched with the original text to identify the title.
 
-- 
+The results from these three methods were then compared. When two results matched, that version was accepted as the title. In instances where all three differed, the title was manually verified and supplemented.
 
-<center>Table 1: Each conference corresponds to its number of papers</center>
+Finally, the "writings" table was generated from the "PvsA" matrix in the original dataset, while the "citations" table was constructed using the "PvsP" matrix from the same source.
 
-| Conference | Number | Conference | Number |
-| :--------: | :----: | :--------: | :----: |
-|    KDD     |  1061  |    SOSP    |  332   |
-|   SIGMOD   |  1417  |    SPAA    |  473   |
-|    WWW     |  1653  |  SIGCOMM   |  648   |
-|   SIGIR    |  1572  |  MobiCOMM  |  322   |
-|    CIKM    |  1724  |    ICML    |  456   |
-|    SODA    |  662   |    COLT    |   83   |
-|    STOC    |  1519  |    VLDB    |  577   |
+## Dataset Composition
 
-## Content
+- **papers.csv**: This file contains 12,499 papers from 14 conferences across various years, with each row representing a unique paper identified by paper_id, along with its publication year, conference, title, and abstract.
 
-- paper.csv: It contains 12,499 papers of 14 conferences in different years, in which each row of data is the paper's unique identification paper_id, publication year, published conference, title of the paper, and abstract of the paper.
-- author.csv: It includes 17,431 authors, each with a unique identification of author_id, name, and firm or organization for which they work.
-- cite.csv: A total of 30789 paper citation relationships are included, two of which are listed as paper_id and paper_id_cited, that is, the former cited the latter.
-- paper_author.csv: It includes 37055 papers and author relationships, and each line is paper ID paper_id and author ID author_id. A paper may have multiple authors.
-- masks.pt: The training set (280), verification set(500) and test set(1000) are divided. The training set selected 20 papers for each category. While the verification set and the test set meet the natural distribution law as much as possible, the balance adjustment is made so that the number of papers in the least category and the number of papers in the most category are not more than four times.
-- paper_embeddings.npy: The paper information embedding was carried out using the "all-MiniLM-L6-v2" model.
-- author_embeddings.npy: The author information embedding was carried out using the "all-MiniLM-L6-v2" model.
+- **authors.csv**: This file includes 17,431 authors, each identified by a unique author_id, along with their name and the firm or organization they are associated with.
 
-## Refferences
+- **citations.csv**: This file captures 30,789 citation relationships, where each entry consists of paper_id and paper_id_cited, indicating that the former cites the latter.
 
-[1]. Wang X, Ji H, Shi C, et al. Heterogeneous graph attention network[C]//The world wide web conference. 2019: 2022-2032.
+- **writings.csv**: This file documents 37,055 paper-author relationships, with each entry comprising a paper_id and an author_id, allowing for multiple authors per paper.
 
-[2]. Tang J, Zhang J, Yao L, et al. Arnetminer: extraction and mining of academic social networks[C]//Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. 2008: 990-998.
+- **<span style="color: black;">masks.pt</span>**: This file divides the dataset into training (280), verification (500), and test (1,000) sets. The training set includes 20 papers from each category, while the verification and test sets are balanced to ensure that the number of papers in the least and most represented categories does not exceed a fourfold difference.
+
+- **paper_embeddings.npy**: This file contains embeddings for paper information generated using the "all-MiniLM-L6-v2" model.
+
+- **author_embeddings.npy**: This file holds embeddings for author information, also created with the "all-MiniLM-L6-v2" model.
+
+## References
+
+[1]. https://github.com/Jhy1993/HAN/tree/master/data/acm
+
+[2]. Wang X, Ji H, Shi C, et al. Heterogeneous graph attention network[C]//The world wide web conference. 2019: 2022-2032.
+
+[3]. Steven A. Cholewiak, Panos Ipeirotis, Victor Silva, and Arun Kannawadi. SCHOLARLY: Simple access to Google Scholar authors and citation using Python, 2021.
+
+[4]. https://tongyi.aliyun.com/
+
+[5]. Tang J, Zhang J, Yao L, et al. ArnetMiner: extraction and mining of academic social networks[C]//Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. 2008: 990-998.
+
+## Citing
+
+If you find this useful in your research, please cite our paper, thx:
+
+```
+@article{rllm2024,
+      title={rLLM: Relational Table Learning with LLMs}, 
+      author={Weichen Li and Xiaotong Huang and Jianwu Zheng and Zheng Wang and Chaokun Wang and Li Pan and Jianhua Li},
+      year={2024},
+      eprint={2407.20157},
+      archivePrefix={arXiv},
+      primaryClass={cs.AI},
+      url={https://arxiv.org/abs/2407.20157}, 
+}
+```

TLF2K/README.md

@@ -1,29 +1,64 @@
------------
-Description
------------
+# TLF2K
 
-Table-LastFM2K extends the original lastfm-2k dataset (https://grouplens.org/datasets/hetrec-2011/) with enriched artist data. As the original dataset was collected in 2011 and last.fm data is continuously updated, we have selected pre-2011 information for each artist to augment the data. The augmented data is available in the 'regression' folder as artists.csv. This new version comprises two tasks, each utilizing distinct sets of data.
+## Overview
 
-The basic task is to classify artists' genre label.
+**Table-LastFM2K (TLF2K)** extends the original Lastfm-2k dataset [1][2] with enriched artist data. It includes three tables: artists, user_artists, and user_friends. The "artists" table contains several key attributes related to artists, while the "user_artists" table captures the listening counts between specific users and artists. The "user_friends" table captures the friendships between specific users. This new version includes 9,047 artists, 80,009 user-artist relationships, and 12,717 user-friend relationships.
 
------------
-Artists file format
------------
-artists.csv follows the format outlined below:
-artistID,type,name,born,yearsActive,location,label,tag_list,biography,url
-
-* artistID: The identifier given in the original lastfm dataset.
-* type: Used to indicate whether the artist is an individual (such as a singer, DJ, or composer) or a group (band, musical ensemble), left blank if indeterminate.
-* name: The name of the artist.
-* born: Only present for single-type artists, indicating the artist's birthdate.
-* yearsActive: Only present for multiple-type artists, indicating the artist's years of activity.
-* location: The birthplace or location of establishment of the artist.
-* label: The artist's genre label, generated by ChatGPT along with manual identification, is one of the following: [country, electronic, hip-Hop, jazz, latin, pop, punk, reggae, rock, metal, soul].
-* tag_list: Added used to identify tags associated with the artist.
-* biography: The artist's personal biography, collected from the final version of information predating 2011. The information contains line breaks represented as "\n".
-* url: The artist's homepage on the last.fm website.
-
-### ---------------- The Following is lastfm-2k's original Readme ----------------
+## Data Processing
+
+First of all, since the original dataset was collected in 2011, only information on artists before 2011 from Last.fm was fetched to enrich the dataset.
+
+Next, based on Wikipedia's classification of popular music [4], each artist was categorized into one of eleven types of music genres. In this labeling process, ChatGPT was employed to assign labels to each musician based on their respective tag lists. Human intervention was required when ChatGPT exhibited insufficient confidence.
+
+Finally, artists not associated with the chosen eleven labels were removed, and the corresponding artists in the user_artists table were also eliminated. Additionally, the artist IDs were rearranged accordingly.
+
+## Dataset Composition
+
+- **artists.csv**: This file contains 9,047 artists, following the format outlined below: artistID, type, name, born, yearsActive, location, label, tagList, biography and url information.
+  * artistID: The identifier given in the original lastfm dataset.
+  * type: Used to indicate whether the artist is an individual (such as a singer, DJ, or composer) or a group (band, musical ensemble), left blank if indeterminate.
+  * name: The name of the artist.
+  * born: Only present for single-type artists, indicating the artist's birthdate.
+  * yearsActive: Only present for multiple-type artists, indicating the artist's years of activity.
+  * location: The birthplace or location of establishment of the artist.
+  * label: The artist's genre label ([Country, Electronic, Hip-Hop, Jazz, Latin, Pop, Punk, Reggae, Rock, Metal, Soul]), generated by ChatGPT along with manual identification.
+  * tagList: Retained to identify tags associated with the artist.
+  * biography: The artist's personal biography, collected from the final version of information predating 2011. The information contains line breaks represented as "\n".
+  * url: The artist's homepage on the last.fm website.
+- **user_artists.csv**: This file includes 80,009 user_artists relationships, each row of data represents the number of times a certain user has listened to a specific artist, composed of userID, artistID and weight.
+- **user_friends.csv**: This file contains 12,717 user_friends relationships, with each row representing a friendship between two users, composed of UserID and FriendID..
+- **<span style="color: black;">masks.pt</span>**: This file divides the dataset into training (220), verification (500), and test (1,000) sets. The training set includes 20 artists for each tag label, while the verification and test sets are structured to align as closely as possible with the natural distribution of artists.
+
+## References
+
+[1]. GroupLens. (2013, September 6). The 2nd International Workshop on Information Heterogeneity and Fusion in Recommender Systems (HetRec 2011) datasets. https://grouplens.org/datasets/hetrec-2011/
+
+[2]. Cantador, I., Brusilovsky, P., & Kuflik, T. (2011, October 23). In Proceedings of the fifth ACM conference on Recommender systems (pp. 387–388). Association for Computing Machinery. https://doi.org/10.1145/2043932.2044016
+
+[3]. Last.fm. from https://www.last.fm/
+
+[4]. Wikipedia.  Music genre. from https://en.wikipedia.org/wiki/Music_genre
+
+## Citing
+
+If you find this useful in your research, please cite our paper, thx:
+
+```
+@article{rllm2024,
+      title={rLLM: Relational Table Learning with LLMs}, 
+      author={Weichen Li and Xiaotong Huang and Jianwu Zheng and Zheng Wang and Chaokun Wang and Li Pan and Jianhua Li},
+      year={2024},
+      eprint={2407.20157},
+      archivePrefix={arXiv},
+      primaryClass={cs.AI},
+      url={https://arxiv.org/abs/2407.20157}, 
+}
+```
+
+
+---
+
+### ---------- The Following is Movielens-1M's Original README ----------
 
 hetrec2011-lastfm-2k
 ====================

TLF2K/artists.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bfafc159dca1bcf439f18ae73572a71d43eaf2ec3f04a0a2cd0f47d55443c902
-size 25485819
+oid sha256:7fe4cf4d0c6212193a660793225eccaf4b14055283d7d9191df932dfe0f6fb12
+size 25416582

TML1M/README.md

@@ -1,67 +1,75 @@
-DESCRIPTION
-================================================================================
+# TML1M
 
-Table-Movielens1M extends the original Movielens-1M dataset (https://grouplens.org/datasets/movielens/1m/) with enriched movie data. This new version contains 6040 users, 3883 movies, and 1000209 ratings. Leveraging each MovielensID's corresponding movie page on movielens.org, we further fetch some "time-invariant" metadata for movies, including: Director, Cast, Runtime, Languages, Certificate, Plot, and Url.
+## Overview
 
-We have also prepared embeddings for movies with all-MiniLM-L6-v2 model, see embeddings.npy in zipfile.
+**Table-Movielens1M (TML1M)** is a relational table dataset extends the original Movielens-1M dataset [1][2] with enriched movie data. It includes three tables: users, movies and ratings. The "users" and the "movies" contain several key attributes related to users and movies. The "ratings" captures the relationships between specific users and movies. This new version contains 6,040 users, 3,883 movies, and 1,000,209 ratings. 
 
-The basic task is to classify users' age phase.
+## Data Processing
 
-ENHANCED MOVIE DATA FORMAT
-================================================================================
+The processing focused exclusively on the movies, extracting "time-invariant" metadata from each film. This metadata includes details such as the director, cast, running time, language, certification, plot, and URL, obtained by using the Movielens ID to access the corresponding movie page on movielens.org [3]. Afterward, the movie IDs were reordered to be consecutive for improved usability.
 
-movies.csv follows the format outlined below:
-MovielensID,Title,Year,Genre,Director,Cast,Runtime,Languages,Certificate,Plot,Url
-
-- MovielensID: This is the movie's id on the movielens website, inherited from the original Movielens-1M dataset.
-- Title: The title of the movie, sourced from the movielens website.
-- Year: The release year of the movie, sourced from the movielens website.
-- Genre: Movie genres, inherited from the original Movielens-1M dataset.
-- Director: The director's name of the movie, sourced from the movielens website.
-- Cast: The main cast of the movie, sourced from the movielens website.
-- Runtime: The duration of the movie, sourced from the movielens website.
-- Languages: Official language versions of the movie, sourced from the movielens website.
-- Certificate: Movie certificate information, sourced from the movielens website.
-- Plot: A brief summary of the movie's main plot, sourced from the movielens website.
-- Url: The movie's URL on the movielens website.
-
-SOME SPECIAL NOTES
-================================================================================
 
-1. Please note that due to the age of the Movielens-1M dataset, some MovielensID values no longer correspond to actual pages, meaning there is a discrepancy between MovielensID and the actual page ID on the movielens website.
-
-For movies facing this issue, we have extracted information from their actual pages. However, to minimize changes, we have not modified their MovielensID. Here is a list of such movies with their MovielensID and real ID:
-
-   * 3589: Original ID changed to 54605
-   * 557: Original ID changed to 8739
-   * 578: Original ID changed to 6531
-   * 669: Original ID changed to 6918
-   * 863: Original ID changed to 193303
-   * 978: Original ID changed to 4970
-   * 1108: Original ID changed to 6677
-   * 1205: Original ID changed to 4006
-   * 1294: Original ID changed to 5060
-   * 1362: Original ID changed to 4424
-   * 1494: Original ID changed to 6425
-   * 1868: Original ID changed to 6474
-   * 2645: Original ID changed to 5649
-   * 3027: Original ID changed to 203829
-   * 3366: Original ID changed to 6808
-   * 3416: Original ID changed to 6460
-   * 3482: Original ID changed to 72267
-   * 3532: Original ID changed to 5560
-   * 3842: Original ID changed to 5717
-   * 3935: Original ID changed to 6910
-
-1. Additionally, there are some other issues:
-	
-   * Movie with MovielensID 2228 in movielens is missing, so we use information from IMDB.
-   * Movies with MovielensID 1741 and 1758 are duplicates of 1795 and 2563 respectively.
-
-2. Users and Ratings file are followed original format.
-
-
----------------- The Following is Movielens-1m's original Readme ----------------
+In addition to these adjustments, several discrepancies were identified:
+- Movie with Movielens ID 2228 (MovieID 2160) is missing from Movielens, and data from IMDB was used as a substitute.
+- Movie with Movielens IDs 1741 (MovieID 1691) and 1758 (MovieID 1706) are duplicates of 1795 (MovieID 1736) and 2563 (MovieID 2495), respectively.
+
+## Dataset Composition
+
+- **users.csv**: This file contains 6040 users, each user has UserID, Gender, Age, Occupation and Zip-code. Occupations are shown in numerical code form. Age is chosen from the following ranges:
+	*  1:  "Under 18"
+	* 18:  "18-24"
+	* 25:  "25-34"
+	* 35:  "35-44"
+	* 45:  "45-49"
+	* 50:  "50-55"
+	* 56:  "56+"
+
+- **movies.csv**: This file includes 3883 movies with MovieID, Title, Year, Genre, Director, Cast, Runtime, Languages, Certificate, Plot and Url information.
+- 
+    * MovieID: The unique identifier for each movie, organized in sequential order.
+    * Title: The title of the movie, sourced from the Movielens website.
+    * Year: The release year of the movie, obtained from the Movielens website.
+    * Genre: The genres of the movie, derived from the original Movielens-1M dataset, separated by vertical bars (“|”).
+    * Director: The name(s) of the director(s) of the movie, sourced from the Movielens website, listed in commas if there is more than one.
+    * Cast: The primary cast of the movie, obtained from the Movielens website, separated by commas.
+    * Runtime: The duration of the movie, sourced from the Movielens website.
+    * Languages: The official language versions of the movie, obtained from the Movielens website, separated by commas if there are multiple.
+    * Certificate: The movie certification information, sourced from the Movielens website.
+    * Plot: A brief summary of the movie's main plot, sourced from the Movielens website.
+    * URL: The URL of the movie on the Movielens website.
+
+
+- **ratings.csv**: This file contains 1,000,209 ratings, each row of data represents a specific user's evaluation of a particular movie, composed of UserID, MovieID, Rating and Timestamp.
+- **<span style="color: black;">masks.pt</span>**: This file divides the dataset into training (140), verification (500), and test (1,000) sets. The training set includes 20 users from each age group, while the verification and test sets are structured to align as closely as possible with the natural distribution of users.
+- **embeddings.npy**: This file contains embeddings for movie information generated using the “all-MiniLM-L6-v2” model.
+
+## References
+[1]. GroupLens. (2015). MovieLens 1M dataset. Retrieved from https://grouplens.org/datasets/movielens/1m/
+
+[2]. Harper, F. M., & Konstan, J. A. (2015). The MovieLens datasets: History and context. ACM Transactions on Interactive Intelligent Systems (TiiS), 5(4), Article 19. https://doi.org/10.1145/2827872
+
+[3]. GroupLens. (2013). MovieLens. from https://movielens.org/
+
+## Citing
+
+If you find this useful in your research, please cite our paper, thx:
+```
+@article{rllm2024,
+      title={rLLM: Relational Table Learning with LLMs}, 
+      author={Weichen Li and Xiaotong Huang and Jianwu Zheng and Zheng Wang and Chaokun Wang and Li Pan and Jianhua Li},
+      year={2024},
+      eprint={2407.20157},
+      archivePrefix={arXiv},
+      primaryClass={cs.AI},
+      url={https://arxiv.org/abs/2407.20157}, 
+}
+```
+
+
+---
+
+### ---------- The Following is Movielens-1M's Original README ----------
+
 SUMMARY
 ================================================================================