prune_recover: Performance evaluation of link predictors
In galanisl/LinkPrediction: Link prediction in complex networks

Description Usage Arguments Value Author(s) References Examples

Given a network of interest and any number of link prediction function names, evaluates the performance of these link predictors by pruning edges from the network and checking if the methods give high likelihood scores to the removed edges. Performance is assessed with four different metrics: Recall@k, AUPR, AUROC and Average Precision.

1 2	prune_recover(g, ..., probes = seq(0.1, 0.5, 0.1), epochs = 10, preserve_conn = FALSE, use_weights = FALSE)

`g`	igraph; The network of interest.
`...`	character; One or more function names able to produce a tibble with 3 columns, the first two being node IDs (candidate link) and the third their likelihood of interaction. This tibble should be sorted from most to less likely candidate links.
`probes`	numeric; A vector indicating the fraction of links to be pruned from `g` at random.
`epochs`	integer; Number of times each fraction of links is randomly removed.
`preserve_conn`	logical; Should network connectivity be preserved after edge pruning? This is important for some prediction techniques that can only be applied to connected topologies. To preserve connectivity, the network's Minimum Spanning Tree (MST) is computed and links are pruned around it.
`use_weights`	logical; If connectivity is to be preserved, this parameter indicates whether the MST should be computed taking edge weights into account.

Tibble with the following columns:

`method`	Link prediction method.
`epoch`	Epoch of link removal.
`frac_rem`	Fraction of links pruned at the specified epoch.
`links_rem`	Number of links pruned at the specified epoch.
`recall_at_k`	Fraction of top candidates links that are in the set of removed edges.
`aupr`	The area under the Precision-Recall curve.
`auroc`	The area under the Receiving Operating Characteristic curve.
`avg_prec`	The average Precision at the point where Recall reaches its maximum value of 1.

Gregorio Alanis-Lobato galanisl@uni-mainz.de

Lu, L. et al. (2015) Toward link predictability of complex networks. PNAS 112(8):2325-2330

1
2
3

# Assess the performance of three link predictors applied to the Zachary 
# Karate Club network
assessment <- prune_recover(g = karate_club, "lp_cn", "lp_aa", "lp_pa")