Simulate with Python API

The ASReview Python API provides advanced control over the ASReview software, allowing users to customize models, implement different sampling strategies, and more. This example demonstrates how to simulate a systematic review using the ASReview API and save the results in an ASReview project file.

import asreview as asr
from synergy_dataset import Dataset

Here, we use a dataset from the SYNERGY collection, accessed via the synergy-dataset package.

d = Dataset("Hall_2012").to_frame()
d.head()

	doi	title	abstract	label_included
openalex_id
https://openalex.org/W2131536587	https://doi.org/10.1109/indcon.2010.5712716	Computer vision based offset error computation...	The use of computer vision based approach has ...	0
https://openalex.org/W2557025555	https://doi.org/10.1109/induscon.2010.5740045	Design and development of a software for fault...	This paper presents an on-line fault diagnosis...	0
https://openalex.org/W2143148279	https://doi.org/10.1109/tpwrd.2005.848672	Analytical Approach to Internal Fault Simulati...	A new method for simulating faulted transforme...	0
https://openalex.org/W2111816457	https://doi.org/10.1109/icelmach.2008.4799852	Nonlinear equivalent circuit model of a tracti...	The paper presents the development of an equiv...	0
https://openalex.org/W3142547111	https://doi.org/10.1109/ipdps.2006.1639408	Fault tolerance with real-time Java	After having drawn up a state of the art on th...	0

Next, we import the required models for the simulation.

from asreview.models.balancers import Balanced
from asreview.models.classifiers import SVM
from asreview.models.feature_extractors import Tfidf
from asreview.models.queriers import Max, TopDown
from asreview.models.stoppers import IsFittable

We create a simulation workflow that begins with a top-down reading strategy until both a relevant and an irrelevant article are identified. Afterward, the simulation transitions to an active learning phase powered by an SVM classifier.

learners = [
    asr.ActiveLearningCycle(querier=TopDown(), stopper=IsFittable()),
    asr.ActiveLearningCycle(
        querier=Max(),
        classifier=SVM(C=3),
        balancer=Balanced(ratio=5),
        feature_extractor=Tfidf(),
    ),
]

sim = asr.Simulate(
    d,
    d["label_included"],
    learners,
)
sim.review()

Relevant records found: 100%|██████████| 104/104 [02:06<00:00,  1.22s/it]
Records labeled       :  65%|██████▍   | 5672/8793 [02:06<01:09, 44.75it/s]

Loss: 0.022
NDCG: 0.656

Finally, we review the simulation results to analyze the performance and outcomes of the systematic review process.

sim._results

	record_id	label	classifier	querier	balancer	feature_extractor	training_set	time	note	tags	user_id
0	0	0	None	top_down	None	None	0	1.745846e+09	None	None	None
1	1	0	None	top_down	None	None	1	1.745846e+09	None	None	None
2	2	0	None	top_down	None	None	2	1.745846e+09	None	None	None
3	3	0	None	top_down	None	None	3	1.745846e+09	None	None	None
4	4	0	None	top_down	None	None	4	1.745846e+09	None	None	None
...	...	...	...	...	...	...	...	...	...	...	...
5667	8389	0	svm	max	balanced	tfidf	5667	1.745846e+09	None	None	None
5668	1739	0	svm	max	balanced	tfidf	5668	1.745846e+09	None	None	None
5669	4807	0	svm	max	balanced	tfidf	5669	1.745846e+09	None	None	None
5670	5160	0	svm	max	balanced	tfidf	5670	1.745846e+09	None	None	None
5671	5647	1	svm	max	balanced	tfidf	5671	1.745846e+09	None	None	None

5672 rows × 11 columns