FIG. 5 illustrates the average performance of each condition of the experimental results in numbers. For each condition (i.e., in a table cell), the three numbers indicate (1) the average metric of the baseline, (2) the average metric of the proposed method, and (3) the improvement of the proposed method over the five runs (on removing different numbers of links from the original dataset). The highest performance and improvement are highlighted in bold for each metric in each dataset. The performance metrics (R-Precision or Area Under Curve—Precision Recall (AUC PR)) are computed in each run with the input network built by removing a certain percentage of the links.

From the results, the proposed biclique oriented methods enhance their baselines in all the conditions with different levels of improvement on both R-Precision and AUC PR. Some of the performance gain is substantial, where the maximum improvement appears with the preferential attachment algorithm for the unweighted Atlantic Storm dataset (0.564 for R-Precision and 0.557 for AUC PR). Thus, through the implementations of the algorithms as described in FIGS. 4(a) and 4(b), improvement to related art algorithms can be achieved and related art link prediction algorithms can be enhanced to more accurately detect missing links.