Semantics-aware Dataset Discovery from Data Lakes with Contextualized Column-based Representation Learning
- Python 3.7.10
- PyTorch 1.9.0+cu111
- Transformers 4.9.2
- NVIDIA Apex (fp16 training)
Install requirements:
pip install -r requirements
Datasets for table union search:
- Santos: https://zenodo.org/record/7758091
- TUS: https://github.com/RJMillerLab/table-union-search-benchmark
WDC web tables:
- See download instructions here for the 50M relational English tables.
Viznet: https://github.com/megagonlabs/sato/tree/master/table_data
The main entry point is run_pretrain.py. Example command:
CUDA_VISIBLE_DEVICES=0 python run_pretrain.py \
--task viznet \
--batch_size 64 \
--lr 5e-5 \
--lm roberta \
--n_epochs 3 \
--max_len 128 \
--size 10000 \
--projector 768 \
--save_model \
--augment_op drop_col \
--fp16 \
--sample_meth head \
--table_order column \
--run_id 0
Hyperparameters:
--task: the tasks that we current support include "santos", "santosLarge", "tus", "tusLarge", "large", "small", "viznet". The tasks "large" and "small" are for column matching and "viznet" is for column clustering.--batch_size,--lr,--n_epochs,--max_len: standard batch size, learning rate, number of training epochs, max sequence length--lm: the language model (we use roberta for all the experiments)--size: the maximum number of tables/columns used during pre-training--projector: the dimension of projector (768 by default, same in all the experiments)--save_model: if this flag is on, the model checkpoint will be saved to the directory specified in the--logdirflag, such as"results/viznet/model_drop_col_head_column_0.pt"--augment_op: augmentation operator for contrastive learning. It includes["drop_col", "sample_row", "sample_row_ordered", "shuffle_col", "drop_cell", "sample_cells", "replace_cells", "drop_head_cells", "drop_num_cells", "swap_cells", "drop_num_col", "drop_nan_col", "shuffle_row"]- Column-level:
drop_col(drops a random column),shuffle_col(shuffles columns),drop_num_col(drops random numeric columns),drop_nan_col(drops columns with mostly NaNs) - Row-level:
sample_row(sample rows),sample_row_ordered(sample rows but preserve order),shuffle_row(shuffles the order of rows) - Cell-level:
drop_cell(drops a random cell),sample_cells(sample cells),replace_cells(sample random cells and replace with first ordered cells),drop_head_cells(drop first quarter cells),drop_num_cells(drop a sample of numeric cells),swap_cells(swap two cells)
- Column-level:
--sample_meth: table pre-processing operator that preserves order and de-duplicates. It includes["head", "alphaHead", "random", "constant", "frequent", "tfidf_token", "tfidf_entity", "tfidf_row", "pmi"]- Row-level:
tfidf_row(takes the rows with highest average tfidf scores),pmi(get highest pmi of pairs of column with topic column) - Entity-level:
tfidf_entity(takes entities in a column with highest after tfidf scores over its tokens) - Token-level:
head(take first N tokens),alphaHead(take first N sorted tokens),random(randomly sample tokens),constant(take every Nth token),frequent(take most frequently-occurring tokens),tfidf_token(take tokens with highest tfidf scores)
- Row-level:
--fp16: half-precision training (always turn this on)--table_order: row or column order for pre-processing, "row" or "column"--single_column: if this flag is on, then it will run the single-column variant ignoring all the table context--mlflow_tag: use this flag to assign any additional tags for mlflow logging
Run extractVectors.py. Example command:
python extractVectors.py \
--benchmark santos \
--table_order column \
--run_id 0
Hyperparameters
--benchmark: the current benchmark for the experiment. Examples includesantos,santosLarge,tus,tusLarge,wdc--single_column: if this flag is on, then it will retrieve the single-column variant--run_id: the run_id of the job (I use 0 for experiments)--table_order: column-ordered or row-ordered (always usecolumn)--save_model: whether to save the vectors in a pickle file, which is then used in the online processing
- Linear & Bounds: Run
test_naive_search.py. Some scripts are intus_cmd.shandrun_tus_all.py(for slurm scheduling). Example command:
python test_naive_search.py \
--encoder cl \
--benchmark santos \
--augment_op drop_col \
--sample_meth tfidf_entity \
--matching linear \
--table_order column \
--run_id 0 \
--K 10 \
--threshold 0.7
Hyperparameters
--encoder: choice of encoder. Options include "cl" (this is for both full Starmie and singleCol baseline), "sato", "sherlock"--benchmark: choice of benchmark for data lake. Options include "santos", "santosLarge", "tus", "tusLarge", "wdc"--augment_op: choice of augmentation operator--sample_meth: choice of sampling method--matching: "linear" matching (full) or "bounds". If you would like to run "greedy", add the function call to the code--table_order: "column" or "row" (just use column)--run_id: always 0--single_column: when set to True, run the single column baseline--K: what you would like to set K to in top-K results--threshold: the similarity threshold
FOR ERROR ANALYSIS: bucket (bucket number between 0 and 5), analysis (either "col" for number of columns, "row" for number of rows,numeric" for percentage of numerical columns
FOR SCALABILITY EXPERIMENTS: scal (what fraction of data lake do we want to get the metrics scores for – 0.2,0.4,0.6,0.8,1.0)
- LSH: Run test_lsh.py (example script: lsh_cmd.sh). Example command:
python test_lsh.py \
--encoder cl \
--benchmark santosLarge \
--run_id 0 \
--num_func 8 \
--num_table 100 \
--K 60 \
--scal 1.0
Hyperparameters:
--encoder: choice of encoder. Options include "cl" (this is for both full Starmie and singleCol baseline), "sato", "sherlock"--benchmark: choice of benchmark for data lake. Options include "santos", "santosLarge", "tus", "tusLarge", "wdc"--run_id: always 0--single_column: when set to True, run the single column baseline--num_func: number of hash functions (always use 8 for ‘cl’ encoder)--num_table: number of tables (always use 100 for ‘cl’ encoder)--K: what you would like to set K to in top-K results
FOR SCALABILITY EXPERIMENTS: scal (what fraction of data lake do we want to get the metrics scores for – 0.2,0.4,0.6,0.8,1.0)
- HNSW: Run test_hnsw_search.py (example script: hnsw_cmd.sh). Example command:
python test_hnsw_search.py \
--encoder cl \
--benchmark santosLarge \
--run_id 0 \
--K 60 \
--scal 1.0
Hyperparameters:
--encoder: choice of encoder. Options include "cl" (this is for both full Starmie and singleCol baseline), "sato", "sherlock"--benchmark: choice of benchmark for data lake. Options include "santos", "santosLarge", "tus", "tusLarge", "wdc"--run_id: always 0--single_column: when set to True, run the single column baseline--K: what you would like to set K to in top-K results
FOR SCALABILITY EXPERIMENTS: scal (what fraction of data lake do we want to get the metrics scores for – 0.2,0.4,0.6,0.8,1.0)
Run discovery.py. We assume:
- A model checkpoint in
results/viznet/model_drop_col_head_column_0.pt - The viznet dataset in
data/viznet/
Run the script by
python discovery.py
The code will print out the MSE for NoJoin, contrastiving learning (CL), Jaccard, and Overlap. The joined tables will be output to pickled files named none_joined_tables.pkl, cl_joined_tables.pkl, jaccard_joined_tables.pkl, and overlap_joined_tables.pkl.
See Line 273 and Line 128 of the file sdd/pretrain.py.
To run column clustering, you can run a sequence of commands (remember to check the file paths):
CUDA_VISIBLE_DEVICES=7 python run_pretrain.py \
--task viznet \
--batch_size 64 \
--lr 5e-5 \
--lm roberta \
--n_epochs 3 \
--max_len 128 \
--size 10000 \
--projector 768 \
--save_model \
--augment_op drop_col \
--fp16 \
--sample_meth head \
--table_order column \
--run_id 0
Copy the clustering results:
cp *.pkl data/viznet/multi_column
Each run will pre-train the models on 10k viznet tables and cluster all the columns. The clustering results will be stored at data/viznet/multi_column/clusters.pkl and data/viznet/single_column/.
To view the clusters, you can use the jupyter notebook in notebook/offline.ipynb. Running the last cell should print out some clusters like
artist ---- 1. I Don't Give A ...; 2. I'm The Kinda; 3. I U She; 4. Kick It [featuring Iggy Pop]; 5.
Operate
artist ---- 1. Spoken Intro; 2. The Court; 3. Maze; 4. Girl Talk; 5. A La Mode
artist ---- 1. Street Fighting Man; 2. Gimme Shelter; 3. (I Can't Get No) Satisfaction; 4. The
Last Time; 5. Jumpin' Jack Flash
…
---------------------------------
type ---- Emerson Elementary School; Banneker Elementary School; Silver City Elementary
School; New Stanley Elementary School; Frances Willard Elementary School
type ---- Choctawhatchee Senior High School; Fort Walton Beach High School; Ami Kids
Emerald Coast; Gulf Coast Christian School; Adolescent Substance Abuse
city ---- Chilton; Stoughton
…
---------------------------------
description ---- Fri Sep 11,2015 3:30 PM (CST); Fri Sep 11,2015 6:00 PM (CST); Sat Sep
12,2015 10:00 AM (CST); Sat Sep 12,2015 12:00 PM (CST); Sat Sep 12,2015 5:30 PM (CST)
day ---- Sept. 1; Sept. 7; Sept. 22; Sept. 29; Oct. 5
description ---- Fri Sep 11,2015 3:30 PM (CST); Fri Sep 11,2015 6:00 PM (CST); Sat Sep
12,2015 10:00 AM (CST); Sat Sep 12,2015 12:00 PM (CST); Sat Sep 12,2015 5:30 PM (CST)
...
address ---- 1721 Papillon St, North Port FL; 4113 Wabasso Ave, North Port FL; 3681
Wayward Ave, North Port FL; 1118 N Salford Blvd, North Port FL; 2057 Bendix Ter, North
Port FL
address ---- 5 Brand Rd, Toms River NJ; 40 12th St, Toms River NJ; 75 Sea Breeze Rd,
Toms River NJ; 98 Oak Tree Ln, Toms River NJ; 67 16th St, Toms River NJ
address ---- 652 Martha St, Montgomery AL; 3184 Lexington Rd, Montgomery AL; 120 S
Lewis St, Montgomery AL; 1812 W 2nd St #OP, Montgomery AL; 3582 Southview Ave,
Montgomery AL
---------------------------------
If you are using the code in this repo, please cite the following in your work:
@article{DBLP:journals/pvldb/FanWLZM23,
author = {Grace Fan and
Jin Wang and
Yuliang Li and
Dan Zhang and
Ren{\'{e}}e J. Miller},
title = {Semantics-aware Dataset Discovery from Data Lakes with Contextualized
Column-based Representation Learning},
journal = {Proc. {VLDB} Endow.},
volume = {16},
number = {7},
pages = {1726--1739},
year = {2023}
}
Embedded in, or bundled with, this product are open source software (OSS) components, datasets and other third party components identified below. The license terms respectively governing the datasets and third-party components continue to govern those portions, and you agree to those license terms, which, when applicable, specifically limit any distribution. You may receive a copy of, distribute and/or modify any open source code for the OSS component under the terms of their respective licenses. In the event of conflicts between Megagon Labs, Inc. Recruit Co., Ltd., license conditions and the Open Source Software license conditions, the Open Source Software conditions shall prevail with respect to the Open Source Software portions of the software. You agree not to, and are not permitted to, distribute actual datasets used with the OSS components listed below. You agree and are limited to distribute only links to datasets from known sources by listing them in the datasets overview table below. You are permitted to distribute derived datasets of data sets from known sources by including links to original dataset source in the datasets overview table below. You agree that any right to modify datasets originating from parties other than Megagon Labs, Inc. are governed by the respective third party’s license conditions. All OSS components and datasets are distributed WITHOUT ANY WARRANTY, without even implied warranty such as for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE, and without any liability to or claim against any Megagon Labs, Inc. entity other than as explicitly documented in this README document. You agree to cease using any part of the provided materials if you do not agree with the terms or the lack of any warranty herein. While Megagon Labs, Inc., makes commercially reasonable efforts to ensure that citations in this document are complete and accurate, errors may occur. If you see any error or omission, please help us improve this document by sending information to [email protected].
If you have any questions regarding the code and the paper, please directly contact Grace Fan ([email protected]).