Balance Strategies

There are several balance strategies that rebalance and reorder the training data. This is sometimes necessary, because the data is often very inbalanced: there are many more papers that should be excluded than included (otherwise, automation cannot help much anyway).

Parameters in the config file should be under the section [balance_param].

We have currently implemented the following balance strategies:

Full Sampling

See asreview.balance_strategies.full_sample()

This just uses all the data.


See asreview.balance_strategies.undersample()

This undersamples the data, leaving out excluded papers so that the included and excluded papers are in some particular ratio (closer to one). Configuration options are as follows:

# Shuffle the samples

# Set the ratio of included/excluded to 1

Triple Balance

See asreview.balance_strategies.triple_balance()

This divides the training data into three sets: included papers, excluded papers found with random sampling and papers found with max sampling. They are balanced according to forumla’s depending on the percentage of papers read in the dataset, the number of papers with random/max sampling etc. Works best for stochastic training algorithms.

# Shuffle the samples

# Start with a random/max weight ratio of 10 at 0% read papers.

# Decay to 1 with the following decay exponent:

# Decrease the included/excluded ratio with the power:

# Cap the included/excluded ratio at: