asreview / asreview-makita

Workflow generator for simulation studies using the command line interface of ASReview LAB
https://asreview.ai
MIT License
30 stars 17 forks source link
asreview data machine-learning python simulation systematic-literature-reviews systematic-reviews utrecht-university

ASReview Makita

PyPI version Downloads PyPI - License DOI

ASReviews' Makita (MAKe IT Automatic) is a workflow generator for simulation studies using the command line interface of ASReview LAB. Makita can be used to effortlessly generate the framework and code for your simulation study.

A simulation involves mimicking the screening process for a systematic review of a human in interaction with an Active learning model (i.e., a combination of a feature extractor, classifier, balancing method, and a query strategy). The simulation reenacts the screening process as if a researcher were using active learning. The performance of one or multiple model(s) can then be measured by performance metrics, such as the Work Saved over Sampling, recall at a given point in the screening process, or the average time to discover a relevant record.

Using Makita templates, different study structures can be generated to fit the needs of your very own study. If your study requires a unique template, you can create a new one and use it instead.

With ASReview LAB, you can simulate with the web interface, the Python API, or the Command Line Interface (CLI). Makita makes use of the CLI.

What Makita does:

What Makita does not do:

For a tutorial on using Makita we refer to the Exercise on Using the ASReview Simulation Mode.

Installation

Prerequisites:

Requires Python 3.8 or higher.

Install the Makita extension with pip:

pip install asreview-makita

For upgrading, use:

pip install --upgrade asreview-makita

After installing the extension, ASReview should automatically detect Makita. If installed correctly, asreview --help should list Makita as an option.

Getting started

You can create the framework and code for your own simulation study in 4 steps.

  1. Create an project folder on your computer.
  2. Create a subfolder named data and fill it using one or more datasets.
  3. Using your preferred command line tool, cd into the project folder.
  4. Create a simulation study from a template found in the list of templates via
asreview makita template NAME_OF_TEMPLATE

where NAME_OF_TEMPLATE is one of the templates (e.g. asreview makita template arfi).

Your simulation study is now properly set up and ready for use. To start the simulations, execute the following shell script in the project folder:

sh jobs.sh

Or on Windows:

jobs.bat

The jobs script runs all jobs in the project folder.

Platform support

By default, ASReview Makita renders job files for the platform of rendering. It is possible to render jobs for other platforms as well. Use the argument --platform with values "Windows", "Linux", or "Darwin" (MacOS) to change the output.

asreview makita template basic --platform Windows

By default, the job file depends on the platform. Windows users will see a jobs.bat file, while other users will see jobs.sh. You can overwrite this with

asreview makita template basic --job_file my_jobs_file.my_ext

Templates

The following table gives an overview of the available templates. See Getting started for instructions on usage.

Note: If no seed is set with the template command, the default seed is used. While this is important for the reproducibility of the results, it may lead to long-term bias. To avoid seed-related bias across different simulation studies, a seed should be for the prior records and models.

Basic template

command: basic

The basic template prepares a script for conducting a simulation study with one run using the default model settings, and two randomly chosen priors (one relevant and one irrelevant record).

optional arguments:

  -h, --help                                show this help message and exit
  --job_file JOB_FILE, -f JOB_FILE          The name of the file with jobs.                 Default jobs.bat for Windows, otherwise jobs.sh.
  -s DATA_FOLDER                            Dataset folder
  -o OUTPUT_FOLDER                          Output folder
  --init_seed INIT_SEED                     Seed of the priors.                             Seed is set to 535 by default.
  --model_seed MODEL_SEED                   Seed of the models.                             Seed is set to 165 by default.
  --template TEMPLATE                       Overwrite template with template file path.
  --platform PLATFORM                       Platform to run jobs: Windows, Darwin, Linux.   Default: the system of rendering templates.
  --n_runs N_RUNS                           Number of runs.                                 Default: 1.
  --skip_wordclouds                         Disables the generation of wordclouds.
  --overwrite                               Automatically accepts all overwrite requests.
  --classifier CLASSIFIER                   Classifier to use.                              Default: nb.
  --feature_extractor FEATURE_EXTRACTOR     Feature_extractor to use.                       Default: tfidf.
  --query_strategy QUERY_STRATEGY           Query strategy to use.                          Default: max.
  --balance_strategy BALANCE_STRATEGY       Balance strategy to use.                        Default: double.
  --instances_per_query INSTANCES_PER_QUERY Number of instances per query.                  Default: 1.
  --stop_if STOP_IF                         The number of label actions to simulate.        Default 'min' will stop simulating when all relevant records are found.

ARFI template

command: arfi

The ARFI template (All relevant, fixed irrelevant) prepares a script for running a simulation study in such a way that for every relevant record 1 run will be executed with 10 randomly chosen irrelevant records which are kept constant over runs. When multiple datasets are available the template orders the tasks in the job file per dataset.

optional arguments:

  -h, --help                                show this help message and exit
  --job_file JOB_FILE, -f JOB_FILE          The name of the file with jobs.                 Default jobs.bat for Windows, otherwise jobs.sh.
  -s DATA_FOLDER                            Dataset folder
  -o OUTPUT_FOLDER                          Output folder
  --init_seed INIT_SEED                     Seed of the priors.                             Seed is set to 535 by default.
  --model_seed MODEL_SEED                   Seed of the models.                             Seed is set to 165 by default.
  --template TEMPLATE                       Overwrite template with template file path.
  --platform PLATFORM                       Platform to run jobs: Windows, Darwin, Linux.   Default: the system of rendering templates.
  --n_priors N_PRIORS                       Number of priors.                               Default: 10.
  --skip_wordclouds                         Disables the generation of wordclouds.
  --overwrite                               Automatically accepts all overwrite requests.
  --classifier CLASSIFIER                   Classifier to use.                              Default: nb.
  --feature_extractor FEATURE_EXTRACTOR     Feature_extractor to use.                       Default: tfidf.
  --query_strategy QUERY_STRATEGY           Query strategy to use.                          Default: max.
  --balance_strategy BALANCE_STRATEGY       Balance strategy to use.                        Default: double.
  --instances_per_query INSTANCES_PER_QUERY Number of instances per query.                  Default: 1.
  --stop_if STOP_IF                         The number of label actions to simulate.        Default 'min' will stop simulating when all relevant records are found.

Multimodel template

command: multimodel

The multiple model template prepares a script for running a simulation study comparing multiple models for one dataset and a fixed set of priors (one relevant and one irrelevant record; identical across models).

optional arguments:

  -h, --help                                show this help message and exit
  --job_file JOB_FILE, -f JOB_FILE          The name of the file with jobs.                 Default jobs.bat for Windows, otherwise jobs.sh.
  -s DATA_FOLDER                            Dataset folder
  -o OUTPUT_FOLDER                          Output folder
  --init_seed INIT_SEED                     Seed of the priors.                             Seed is set to 535 by default.
  --model_seed MODEL_SEED                   Seed of the models.                             Seed is set to 165 by default.
  --template TEMPLATE                       Overwrite template with template file path.
  --platform PLATFORM                       Platform to run jobs: Windows, Darwin, Linux.   Default: the system of rendering templates.
  --n_runs N_RUNS                           Number of runs.                                 Default: 1.
  --skip_wordclouds                         Disables the generation of wordclouds.
  --overwrite                               Automatically accepts all overwrite requests.
  --instances_per_query INSTANCES_PER_QUERY Number of instances per query.                  Default: 1.
  --stop_if STOP_IF                         The number of label actions to simulate.        Default 'min' will stop simulating when all relevant records are found.
  --classifiers CLASSIFIERS                 Classifiers to use                              Default: ['logistic', 'nb', 'rf', 'svm']
  --feature_extractors FEATURE_EXTRACTOR    Feature extractors to use                       Default: ['doc2vec', 'sbert', 'tfidf']
  --query_strategies QUERY_STRATEGY         Query strategies to use                         Default: ['max']
  --balance_strategies BALANCE_STRATEGY     Balance strategies to use                       Default: ['double']
  --impossible_models IMPOSSIBLE_MODELS     Model combinations to exclude                   Default: ['nb,doc2vec', 'nb,sbert']

Example usage

If you want to specify certain combinations of classifiers and feature extractors that should and should not be used, you can use the --classifiers, --feature_extractors, --query_strategies, --balance_strategies and --impossible_models option. For instance, if you want to exclude the combinations of nb with doc2vec and logistic with tfidf, use the following command:

asreview makita template multimodel --classifiers logistic nb --feature_extractors tfidf doc2vec --query_strategies max max_random max_uncertainty cluster --impossible_models nb,doc2vec logistic,tfidf

Prior template

command: prior

The prior template evaluates how a set of custom prior knowledge might affect simulation performance. It processes two types of data in the data folder: labeled dataset(s) to be simulated and labeled dataset(s) to be used as prior knowledge. The filename(s) of the dataset(s) containing the custom prior knowledge should use the naming prefix prior_[dataset_name].

The template runs two simulations: the first simulation uses all records from the prior_ dataset(s) as prior knowledge, and the second uses a 1+1 randomly chosen set of prior knowledge from the non-prior knowledge dataset as a minimal training set. Both runs simulate performance on the combined non-prior dataset(s).

Running this template creates a generated_data folder. This folder contains two datasets; dataset_with_priors.csv and dataset_without_priors.csv. The simulations specified in the generated jobs file will use these datasets for their simulations.

optional arguments:

  -h, --help                                show this help message and exit
  --job_file JOB_FILE, -f JOB_FILE          The name of the file with jobs.                 Default jobs.bat for Windows, otherwise jobs.sh.
  -s DATA_FOLDER                            Dataset folder
  -o OUTPUT_FOLDER                          Output folder
  --init_seed INIT_SEED                     Seed of the priors.                             Seed is set to 535 by default.
  --model_seed MODEL_SEED                   Seed of the models.                             Seed is set to 165 by default.
  --template TEMPLATE                       Overwrite template with template file path.
  --platform PLATFORM                       Platform to run jobs: Windows, Darwin, Linux.   Default: the system of rendering templates.
  --n_runs N_RUNS                           Number of runs.                                 Default: 1.
  --skip_wordclouds                         Disables the generation of wordclouds.
  --overwrite                               Automatically accepts all overwrite requests.
  --classifier CLASSIFIER                   Classifier to use.                              Default: nb.
  --feature_extractor FEATURE_EXTRACTOR     Feature_extractor to use.                       Default: tfidf.
  --query_strategy QUERY_STRATEGY           Query strategy to use.                          Default: max.
  --balance_strategy BALANCE_STRATEGY       Balance strategy to use.                        Default: double.
  --instances_per_query INSTANCES_PER_QUERY Number of instances per query.                  Default: 1.
  --stop_if STOP_IF                         The number of label actions to simulate.        Default 'min' will stop simulating when all relevant records are found.

Example usage

Put at least 2 datasets in the data folder. One starting with the prior_ prefix, and one without this prefix.

note: priors_ will also work.

asreview makita template prior --classifier logistic --feature_extractor tfidf

Advanced usage

Create and use custom templates

It is possible to overwrite the internal templates. This can be useful for simulation studies with different needs.

  1. Select an existing template that looks similar to your needs. For example, you want to run ARFI with a different model, then you pick the ARFI template.
  2. Download the template you selected in step 1 from the Github repository. Template files have the following structure template_*.txt.template. For the ARFI example, the template is template_arfi.txt.template.
  3. Save the downloaded template somewhere on your computer. The template is a so-called "Jinja" template. The template consists of ASReview command line commands combined with jinja syntax. The Jinja syntax is very intuitive. See this Cheatsheet.
  4. Edit the Jinja template to your needs.
  5. Run the custom template with the command line option --template PATH_TO_MY_TEMPLATE.txt.template. For the ARFI example, this would be asreview makita template arfi --template PATH_TO_MY_TEMPLATE.txt.template. Please keep in mind that you follow the usual steps for running a template.
  6. A jobs.sh file should be in the your folder.

Please contribute your templates back to the project by making a Pull Request. Then, we can integrate it in the core of the makita package.

Add and use scripts

Makita can add scripts to your repository. The scripts are mainly pre- and postprocessing scripts. These scripts are not (yet) available in any existing ASReview software. Therefore, they can be added manually with asreview makita add-script NAME_OF_SCRIPT.

For example, the results from ASReview datatools are merged via the script merge_descriptives.py (or merge_metrics.py for ASReview insights), using:

  1. Collect statistics (with template)
  2. Run asreview makita add-script merge_descriptives.py
  3. Run python scripts/merge_descriptives.py

Use -s (source) and -o (output) to tweak paths.

Some scripts are added automatically to the folder, as they are part of the template. For example, the get_plot.py script is added to the generated folder when using any template, as it is used to generate the plots.

Still, get_plot.py can be used on its own, as it is a standalone script. To use it, use -s (source) and -o (output) to tweak paths.

Adding a legend to the plot can be done with the -l or --show_legend flag, with the labels clustered on any of the following: 'filename', 'model', 'query_strategy', 'balance_strategy', 'feature_extraction', 'n_instances', 'stop_if', 'n_prior_included', 'n_prior_excluded', 'model_param', 'query_param', 'feature_param', 'balance_param'

Available scripts

The following scripts are available:

Time to Discovery Tables

The 'merge_tds.py' script creates a table of the time to discovery (TD) values for each dataset, with each row corresponding to each record ID of the relevant records in a dataset, and the columns correspond to each simulation run (e.g, for the Multimodel template each column corresponds to a simualtion run with each active learning model). Additionally, the tables includes the average-record-TD values (the average of the TD values for a record across multiple simulation runs), and the average-simulation-TD values (the average of the TD values across all records for a single simulation run).

Run Makita via Docker

To run Makita template with Docker use the following command:

docker run -v $PWD:/app ghcr.io/asreview/asreview makita <YOUR COMMAND>

License

This extension is published under the MIT license.

Contact

This extension is part of the ASReview project (asreview.ai). It is maintained by the maintainers of ASReview LAB. See ASReview LAB for contact information and more resources.