6.3. Recommendations on Choosing the Appropriate Algorithm
In terms of the average computational time from Table 1, FastNCA is faster than the other four algorithms. However, FastNCA is not recommended herein, since it shows a high degree of inconsistency and inaccuracy. Moreover, even though NCA performs very well in both the noise and noise + outliers cases, the run time of NCA is hundreds and thousands of times slower than the other four algorithms using the small-dimensional synthetic data. It can be inferred that NCA is more computationally inefficient for reconstructing large-dimensional TRNs. In the case where the accuracy of the connectivity matrix is the first priority, PosNCA is recommended due to the fact that PosNCA has a high degree of accuracy in estimating A, especially in the scenario where the SNR is high. NINCA and ROBNCA can be selected as the general methods to solve the TRN inference problem, since they are consistent and accurate in both the noise and noise + outliers cases. Moreover, the run time of NINCA and ROBNCA is also comparable to FastNCA. Between these two algorithms, ROBNCA is more preferable if the existence of outliers is known a priori.