Revisiting of 'Revisiting the Stop-and-Stare Algorithms for Influence Maximization'

SSA/DSSA were introduced in SIGMOD’16 as the first algorithms that can provide rigorous \(1-1/e-\epsilon \) guarantee with fewer samples than the worst-case sample complexity \(O(nk \frac{\log n}{\epsilon ^2 OPT_k})\). They are order of magnitude faster than the existing methods. The original SIGMOD’16 paper, however, contains errors, and the new fixes for SSA/DSSA, referred to as SSA-fix and D-SSA-fix, have been published in the extended version of the paper [11]. In this paper, we affirm the correctness on accuracy and efficiency of SSA-fix/D-SSA-fix algorithms. Specifically, we refuse the misclaims on ‘important gaps’ in the proof of D-SSA-fix’s efficiency raised by Huang et al. [5] published in VLDB in May 2017. We also replicate the experiments to dispute the experimental discrepancies shown in [5]. Our experiment results indicate that implementation/modification details and data pre-processing attribute for most discrepancies in running-time. (We requested the modified code from VLDB’17 [5] last year but have not received the code from the authors. We also sent them the explanation for the gaps they misclaimed for the D-SSA-fix’s efficiency proof but have not received their concrete feedback.)