This official extension to ASReview LAB extends the software with tools for plotting and extracting the statistical results of several performance metrics. The extension is especially useful in combination with the simulation functionality of ASReview LAB.
ASReview Insights can be installed from PyPI:
pip install asreview-insightsAfter installation, check if the asreview-insights package is listed as an
extension. Use the following command:
asreview --helpIt should list the 'plot' subcommand and the 'metrics' subcommand.
The ASReview Insights extension is useful for measuring the performance of active learning models on collections of binary labeled text. The extension can be used after performing a simulation study that involves mimicking the screening process with a specific model. As it is already known which records are labeled relevant, the simulation can automatically reenact the screening process as if a screener were using active learning. The performance of one or multiple models can be measured by different metrics and the ASReview Insights extension can plot or compute the values for such metrics from ASReview project files. O'Mara-Eves et al. (2015) provides a comprehensive overview of different metrics used in the field of actrive learning. Below we describe the metrics available in the software.
The recall is the proportion of relevant records that have been found at a certain point during the screening phase. It is sometimes also called the proportion of Relevant Record Found (RRF) after screening an X% of the total records. For example, the RRF@10 is the recall (i.e., the proportion of the total number of relevant records) at screening 10% of the total number of records available in the dataset.
The confusion matrix consist of the True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN). Definitions are provided in the following table retrieved at a certain recall (r%).
| Definition | Calculation | |
|---|---|---|
| True Positives (TP) | The number of relevant records found at recall level | Relevant Records * r% |
| False Positives (FP) | The number of irrelevant records reviewed at recall level | Records Reviewed – TP |
| True Negatives (TN) | The number of irrelevant records correctly not reviewed at recall level | Irrelevant Records – FP |
| False Negatives (FN) | The number of relevant records not reviewed at recall level (missing relevant records) | Relevant Records – TP |
The Work Saved over Sampling (WSS) is a measure of "the work saved over and above the work saved by simple sampling for a given level of recall" (Cohen et al., 2006). It is defined as the proportion of records a screener does not have to screen compared to random reading after providing the prior knowledge used to train the first iteration of the model. The WSS is typically measured at a recall of .95 (WSS@95), reflecting the proportion of records saved by using active learning at the cost of failing to identify .05 of relevant publications.
Kusa et al. (2023) propose to normalize the WSS for class imbalance (denoted as the nWSS). Moreover, Kusa et al. showed that nWSS is equal to the True Negative Rate (TNR). The TNR is the proportion of irrelevant records that were correctly not reviewed at level of recall. The nWSS is useful to compare performance in terms of work saved across datasets and models while controlling for dataset class imbalance.
The following table provides a hypothetical dataset example:
| Dataset characteristics | Example value |
|---|---|
| Total records | 2000 |
| Records Reviewed | 1100 |
| Relevant Records | 100 |
| Irrelevant Records | 1900 |
| Class imbalance | 5% |
With this information, the following metrics can be calculated:
| Metric | Example value |
|---|---|
| TP | 95 |
| FP | 1100 – 95 = 1005 |
| TN | 1900 – 1005 = 895 |
| FN | 100 – 95 = 5 |
| TNR95% | 895 / 1900 = 0.47 |
A variation is the Extra Relevant records Found (ERF), which is the proportion of relevant records found after correcting for the number of relevant records found via random screening (assuming a uniform distribution of relevant records).
The following plot illustrates the differences between the metrics Recall (y-axis), WSS (blue line), and ERF (red line). The dataset contains 1.000 hypothetical records with labels. The stepped line on the diagonal is the naive labeling approach (screening randomly sorted records).
Both recall and WSS are sensitive to the position of the cutoff value and the distribution of the data. Moreover, the WSS makes assumptions about the acceptable recall level whereas this level might depend on the research question at hand. Therefore, Ferdinands et al. (2020) proposed two new metrics: (1) the Time to Discover a relevant record as the fraction of records needed to screen to detect this record (TD); and (2) the Average Time to Discover (ATD) as an indicator of how many records need to be screened on average to find all relevant records in the dataset. The TD metric enables you to pinpoint hard-to-find papers. The ATD, on the other hand, measures performance throughout the entire screening process, eliminating reliance on arbitrary cut-off values, and can be used to compare different models.
The ASReview Insights package extends ASReview LAB with two new subcommands
(see asreview --help): plot and metrics. The plots
and metrics are derived from an ASReview project file. The ASReview file
(extension .asreview) can be
exported
from ASReview LAB after a
simulation,
or it is generated from running a simulation via the command
line.
For example, an ASReview can be generated with:
asreview simulate benchmark:van_de_schoot_2017 -s sim_van_de_schoot_2017.asreview --init_seed 535To use the most basic options of the ASReview Insights extension, run
asreview plot recall YOUR_ASREVIEW_FILE.asreviewwhere recall is the type of the plot, or
asreview metrics sim_van_de_schoot_2017.asreviewMore options are described in the sections below. All options can be
obtained via asreview plot --help or asreview metrics --help.
The recall is an important metric to study the performance of active learning algorithms in the context of information retrieval. ASReview Insights offers a straightforward command line interface to plot a "recall curve". The recall curve is the recall at any moment in the active learning process.
To plot the recall curve, you need a ASReview file (extension .asreview). To
plot the recall, use this syntax (Replace YOUR_ASREVIEW_FILE.asreview by
your ASReview file name.):
asreview plot recall YOUR_ASREVIEW_FILE.asreviewThe following plot is the result of simulating the PTSD data via
the benchmark platform (command asreview simulate benchmark:van_de_schoot_2017 -s sim_van_de_schoot_2017.asreview).
On the vertical axis, you find the recall (i.e, the proportion of the relevant records) after every labeling decision. The horizontal axis shows the proportion of total number of records in the dataset. The steeper the recall curve, the higher the performance of active learning when comparted to random screening. The recall curve can also be used to estimate stopping criteria, see the discussions in #557 and #1115.
asreview plot recall YOUR_ASREVIEW_FILE.asreviewThe Work Saved over Sampling (WSS) metric is a useful metric to study the performance of active learning alorithms compared with a naive (random order) approach at a given level of recall. ASReview Insights offers a straightforward command line interface to plot the WSS at any level of recall.
To plot the WSS curve, you need a ASReview file (extension .asreview). To
plot the WSS, use this syntax (Replace YOUR_ASREVIEW_FILE.asreview by your
ASReview file name.):
asreview plot wss YOUR_ASREVIEW_FILE.asreviewThe following plot is the result of simulating the PTSD data via
the benchmark platform (command asreview simulate benchmark:van_de_schoot_2017 -s sim_van_de_schoot_2017.asreview).
On the vertical axis, you find the WSS after every labeling decision. The recall is displayed on the horizontal axis. As shown in the figure, the WSS is linearly related to the recall.
The Extra Relevant Records found is a derivative of the recall and presents the proportion of relevant records found after correcting for the number of relevant records found via random screening (assuming a uniform distribution of relevant records).
To plot the ERF curve, you need a ASReview file (extension .asreview). To
plot the ERF, use this syntax (Replace YOUR_ASREVIEW_FILE.asreview by your
ASReview file name.):
asreview plot erf YOUR_ASREVIEW_FILE.asreviewThe following plot is the result of simulating the PTSD data via
the benchmark platform (command asreview simulate benchmark:van_de_schoot_2017 -s sim_van_de_schoot_2017.asreview).
On the vertical axis, you find the ERF after every labeling decision. The horizontal axis shows the proportion of total number of records in the dataset. The steep increase of the ERF in the beginning of the process is related to the steep recall curve.
Optional arguments for the command line are --priors to include prior
knowledge, --x_absolute and --y_absolute to use absolute axes.
See asreview plot -h for all command line arguments.
It is possible to show the curves of multiple files in one plot. Use this
syntax (replace YOUR_ASREVIEW_FILE_1 and YOUR_ASREVIEW_FILE_2 by the
asreview_files that you want to include in the plot):
asreview plot recall YOUR_ASREVIEW_FILE_1.asreview YOUR_ASREVIEW_FILE_2.asreviewTo make use of the more advanced features, you can make use of the Python API. The advantage is that you can tweak every single element of the plot in the way you like. The following examples show how the Python API can be used. They make use of matplotlib extensively. See the Introduction to Matplotlib for examples on using the API.
The following example show how to plot the recall with the API and save the result. The plot is saved using the matplotlib API.
import matplotlib.pyplot as plt
from asreview import open_state
from asreviewcontrib.insights.plot import plot_recall
with open_state("example.asreview") as s:
fig, ax = plt.subplots()
plot_recall(ax, s)
fig.savefig("example.png")Other options are plot_wss and plot_erf.
It's straightforward to customize the plots if you are familiar with
matplotlib. The following example shows how to update the title of the plot.
import matplotlib.pyplot as plt
from asreview import open_state
from asreviewcontrib.insights.plot import plot_wss
with open_state("example.asreview") as s:
fig, ax = plt.subplots()
plot_wss(ax, s)
plt.title("WSS with custom title")
fig.savefig("example_custom_title.png")
It's possible to include prior knowledge in your plot. By default, prior knowledge is excluded from the plot.
import matplotlib.pyplot as plt
from asreview import open_state
from asreviewcontrib.insights.plot import plot_wss
with open_state("example.asreview") as s:
fig, ax = plt.subplots()
plot_wss(ax, s, priors=True)By default, all axes in ASReview Insights are relative. The API can be used to change this behavior. The arguments are identical for each plot function.
import matplotlib.pyplot as plt
from asreview import open_state
from asreviewcontrib.insights.plot import plot_wss
with open_state("example.asreview") as s:
fig, ax = plt.subplots()
plot_wss(ax, s, x_absolute=True, y_absolute=True)
fig.savefig("example_absolute_axis.png")
If you have multiple curves in one plot, you can customize the legend:
import matplotlib.pyplot as plt
from asreview import open_state
from asreviewcontrib.insights.plot import plot_recall
fig, ax = plt.subplots()
with open_state("tests/asreview_files/sim_van_de_schoot_2017_1.asreview") as s1:
with open_state("tests/asreview_files/"
"sim_van_de_schoot_2017_logistic.asreview") as s2:
plot_recall(ax,
[s1, s2],
legend_values=["Naive Bayes", "Logistic"],
legend_kwargs={'loc': 'lower center'})
fig.savefig("docs/example_multiple_lines.png")
The metrics subcommand in ASReview Insights can be used to compute metrics
at given values. The easiest way to compute metrics for a ASReview project
file is with the following command on the command line:
asreview metrics sim_van_de_schoot_2017.asreview
which results in
"asreviewVersion": "1.0",
"apiVersion": "1.0",
"data": {
"items": [
{
"id": "recall",
"title": "Recall",
"value": [
[
0.1,
1.0
],
[
0.25,
1.0
],
[
0.5,
1.0
],
[
0.75,
1.0
],
[
0.9,
1.0
]
]
},
{
"id": "wss",
"title": "Work Saved over Sampling",
"value": [
[
0.95,
0.8913851624373686
]
]
},
{
"id": "erf",
"title": "Extra Relevant record Found",
"value": [
[
0.1,
0.9047619047619048
]
]
},
{
"id": "atd",
"title": "Average time to discovery",
"value": 101.71428571428571
},
{
"id": "td",
"title": "Time to discovery",
"value": [
[
3898,
22
],
[
284,
23
],
[
592,
25
],
...
[
2382,
184
],
[
5479,
224
],
[
3316,
575
]
]
},
{
"id": "tp",
"title": "True Positives",
"value": [
[
0.95,
39
],
[
1.0,
42
]
]
},
{
"id": "fp",
"title": "False Positives",
"value": [
[
0.95,
122
],
[
1.0,
517
]
]
},
{
"id": "tn",
"title": "True Negatives",
"value": [
[
0.95,
6023
],
[
1.0,
5628
]
]
},
{
"id": "fn",
"title": "False Negatives",
"value": [
[
0.95,
3
],
[
1.0,
0
]
]
},
{
"id": "tnr",
"title": "True Negative Rate (Specificity)",
"value": [
[
0.95,
0.980146
],
[
1.0,
0.915867
]
]
}
]
}
}
Each available item has two values. The first value is the value at which the metric is computed. In the plots above, this is the x-axis. The second value is the results of the metric. Some metrics are computed for multiple values.
| Metric | Description pos. 1 | Description pos. 2 | Default |
|---|---|---|---|
recall |
Labels | Recall | 0.1, 0.25, 0.5, 0.75, 0.9 |
wss |
Recall | Work Saved over Sampling at recall | 0.95 |
erf |
Labels | ERF | 0.10 |
atd |
Average time to discovery (in label actions) | - | - |
td |
Row number (starting at 0) | Number of records labeled | - |
cm |
Recall | Confusion matrix values at recall | 0.95, 1 |
It is possible to override the default values of asreview metrics. See
asreview metrics -h for more information or see the example below.
asreview metrics sim_van_de_schoot_2017.asreview --wss 0.9 0.95
{
"asreviewVersion": "1.0",
"apiVersion": "1.0",
"data": {
"items": [
{
"id": "recall",
"title": "Recall",
"value": [
[
0.1,
1.0
],
[
0.25,
1.0
],
[
0.5,
1.0
],
[
0.75,
1.0
],
[
0.9,
1.0
]
]
},
{
"id": "wss",
"title": "Work Saved over Sampling",
"value": [
[
0.9,
0.8474220139001132
],
[
0.95,
0.8913851624373686
]
]
},
{
"id": "erf",
"title": "Extra Relevant record Found",
"value": [
[
0.1,
0.9047619047619048
]
]
},
{
"id": "atd",
"title": "Average time to discovery",
"value": 101.71428571428571
},
{
"id": "td",
"title": "Time to discovery",
"value": [
[
3898,
22
],
[
284,
23
],
[
592,
25
],
...
[
2382,
184
],
[
5479,
224
],
[
3316,
575
]
]
}
]
}
}
Metrics can be saved to a file in the JSON format. Use the flag -o or
--output.
asreview metrics sim_van_de_schoot_2017.asreview -o my_file.json
Optional arguments for the command line are --priors to include prior
knowledge, --x_absolute and --y_absolute to use absolute axes.
See asreview metrics -h for all command line arguments.
Metrics are easily accesible with the ASReview Insights API.
Compute the recall after reading half of the dataset.
from asreview import open_state
from asreviewcontrib.insights.metrics import recall
with open_state("example.asreview") as s:
print(recall(s, 0.5))Other metrics are available like wss and erf.
It's possible to include prior knowledge to your metric. By default, prior knowledge is excluded from the metric.
from asreview import open_state
from asreviewcontrib.insights.metrics import recall
with open_state("example.asreview") as s:
print(recall(s, 0.5, priors=True))This extension is published under the MIT license.
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.