Bookdata-tools/README.md

This repository contains the code to import and integrate the book and rating data that we work with. It imports and integrates data from several sources in a single PostgreSQL database; import scripts are primarily in Python, with Rust code for high-throughput processing of raw data files.

If you use these scripts in any published reseaerch, cite our paper:

Michael D. Ekstrand, Mucun Tian, Mohammed R. Imran Kazi, Hoda Mehrpouyan, and Daniel Kluver. 2018. Exploring Author Gender in Book Rating and Recommendation. In Proceedings of the 12th ACM Conference on Recommender Systems (RecSys '18). ACM, pp. 242–250. DOI:10.1145/3240323.3240373. arXiv:1808.07586v1 [cs.IR].

Note: the limitations section of the paper contains important information about the limitations of the data these scripts compile. Do not use this data or tools without understanding those limitations. In particular, VIAF's gender information is incomplete and, in a number of cases, incorrect.

In addition, several of the data sets integrated by this project come from other sources with their own publications. If you use any of the rating data, cite the appropriate original source paper. For each data set below, we have provided a link to the page that describes the appropriate citation.

Requirements

• PostgreSQL 10 or later with orafce and pg_prewarm (from the PostgreSQL Contrib package) installed.
• Python 3.6 or later with the following packages:
  • psycopg2
  • numpy
  • tqdm
  • pandas
  • colorama
  • chromalog
  • natural
  • dvc
• The Rust compiler (available from Anaconda)
• 2TB disk space for the database
• 100GB disk space for data files

It is best if you do not store the data files on the same disk as your PostgreSQL database.

The environment.yml file defines an Anaconda environment that contains all the required packages except for the PostgreSQL server. It can be set up with:

conda create -f environment.yml

All scripts read database connection info from the standard PostgreSQL client environment variables:

• PGDATABASE
• PGHOST
• PGUSER
• PGPASSWORD

Alternatively, they will read from DB_URL.

We use Data Version Control (dvc) to script the import and wire its various parts together. A complete re-run, not including file download time, takes approximately 8 hours on our hardware (24-core 2GHz Xeon, 128GiB RAM, spinning disks).

Initializing and Configuring the Database

After creating your database, initialize the extensions (as the database superuser):

CREATE EXTENSION orafce;
CREATE EXTENSION pg_prewarm;
CREATE EXTENSION "uuid-ossp";

The default PostgreSQL performance configuration settings will probably not be very effective; we recommend turning on parallelism and increasing work memory, at a minimum.

Downloading Data Files

This imports the following data sets:

• Library of Congress MDSConnect Open MARC Records from https://www.loc.gov/cds/products/MDSConnect-books_all.html (auto-downloaded).
• LoC MDSConnect Name Authorities from https://www.loc.gov/cds/products/MDSConnect-name_authorities.html (auto-downloaded).
• Virtual Internet Authority File - get the MARC 21 XML data file from http://viaf.org/viaf/data/ (not auto-downloaded).
• OpenLibrary Dump - the editions, works, and authors dumps from https://openlibrary.org/developers/dumps (auto-downloaded).
• Amazon Ratings - the 'ratings only' data for Books from http://jmcauley.ucsd.edu/data/amazon/ and save it in data (auto-downloaded). If you use this data, cite the paper.
• BookCrossing - the BX-Book-Ratings CSV file from http://www2.informatik.uni-freiburg.de/~cziegler/BX/ (auto-downloaded). If you use this data, cite the paper.
• GoodReads - the GoodReads books, works, authors, and full interaction files from https://sites.google.com/eng.ucsd.edu/ucsdbookgraph/home (not auto-downloaded). If you use this data, cite the paper.

Several of these files can be auto-downloaded with the DVC scripts; others will need to be manually downloaded.

Running Everything

You can run the entire import process with:

dvc repro

Individual steps can be run with their corresponding .dvc files.

Layout

The import code consists of Python, Rust, and SQL code, wired together with DVC.

Python Scripts

Python scripts live under scripts, as a Python package. They should not be launched directly, but rather via run.py, which will make sure the environment is set up properly for them:

python run.py sql-script [options] script.sql

DVC Usage and Stage Files

In order to allow DVC to be aware of current database state, we use a little bit of an unconventional layout for many of our DVC scripts. Many steps have two .dvc files with associated outputs:

• step.dvc runs import stage step.
• step.transcript is (consistent) output from running step, recording the actions taken. It is registered with DVC as the output of step.dvc.
• step.status.dvc is an always-changed DVC stage that depends on step.transcript and produces step.status, to check the current status in the database of that import stage.
• step.status is an uncached output (so it isn't saved with DVC, and we also ignore it from Git) that is registered as the output of step.status.dvc. It contains a stable status dump from the database, to check whether step is actually in the database or has changed in a meaningful way.

Steps that depend on step then depend on step.status, not step.trasncript.

The reason for this somewhat bizarre layoutis that if we just wrote the output files, and the database was reloaded or corrupted, the DVC status-checking logic would not be ableto keep track of it. This double-file design allows us to make subsequent steps depend on the actual results of the import, not our memory of the import in the Git repository.

The file init.status is an initial check for database initialization, and forces the creation of the meta-structures used for tracking stage status. Everything touching the database should depend on it, directly or indirectly.

In-Database Status Tracking

Import steps are tracked in the stage_status table in the database. For completed stages, this can include a key (checksum, UUID, or other identifier) to identify a 'version' of the stage. Stages can also have dependencies, which are solely used for computing the status of a stage (all actual dependency relationships are handled by DVC):

• stage_deps tracks stage-to-stage dependencies, to say that one stage used another as input.
• stage_file tracks stage-to-file dependencies, to say that a stage used a file as input.

The source_file table tracks input file checksums.

Projects using the book database can also use stage_status to obtain data version information, to see if they are up-to-date.

Utility Code

The bookdata package contains Python utility code, and the src directory contains a number of utility modules for use in the Rust code. To the extent reasonable, we have tried to mirror design patterns and function names.

Design for Datasets

The general import philosophy is that we import the data into a PostgreSQL table in a raw form, only doing those conversions necessary to be able to access it with PostgreSQL commands and functions. We then extract information from this raw form into other tables and materialized views to enable relational queries.

Each data set's import process follows the following steps:

1. Initialize the database schema, with an SQL script under schemas/.
2. Import the raw data, controlled by DVC steps under import/. This may be multiple steps; for example, OpenLibrary has a separate import step for each file. Actual import is usually handled by Rust or Python code.
3. Index the data into relational views and tables. This is done by SQL scripts under index/.

Data integration then happens after the data sets are indexed (mostly).

Book Identifiers

Each data set comes with its own identification scheme for books:

• LCCN
• OpenLibrary key
• ASIN
• GoodReads book and work identifiers

We integrate these through two steps. First, we map ISBNs to numeric IDs with the isbn_id table. This table contains every ISBN (or ISBN-like thing, such as ASIN) we have seen and associates it with a unique identifier.

Second, we map ISBN IDs to clusters with the isbn_cluster table. A cluster is a collection of related ISBNs, such as the different editions of a work. They correspond to GoodReads or OpenLibrary 'works' (in fact, when a GoodReads or OpenLibrary work is available, it is used to generate the clusters).

This allows us to connect ratings to metadata with maximal link coverage, by pulling in metadata across the whole book cluster.