Abstract
Massive amounts of information are generated in various social media and spread across multi-social networks through individual forwarding and sharing, which greatly enhance the speed and scope of transmission, but also bring great challenges to the control and governance of misinformation. The characteristics of the spread of misinformation across multi-social networks are considered, this article investigates the novel problem of misinformation influence minimization by entity protection on multi-social networks, and systematically tackling this problem. We analyse the hardness and the approximation property of the problem. We construct a multi-social networks coupled method and devise a pruning and filtering rule. We develop a two-stage discrete gradient descent (TD-D) algorithm to solve NP-Hard problems. We also construct a two-stage greedy (TG) algorithm with the approximate guarantee to verify the algorithm we developed. Finally, the effectiveness of our proposed methods is analysed in synthetic and real multi-network datasets (contains up to 202K nodes and 2.5M edges). The results show that the ability of the TD-D and TG algorithms to suppress the spread of misinformation is basically the same, but the running time of the TG algorithm is much higher than (far more than 10 times) that of the TD-D algorithm.
Similar content being viewed by others
References
N PM (2018) Rumors of child-kidnapping gangs and other WhatsApp hoaxes are getting people killed in India. Available from: https://www.latimes.com/world/asia/la-fg-india-whatsapp-2018-story.html
Zhang Y, Su Y, Weigang L, Liu H (2018) Rumor and authoritative information propagation model considering super spreading in complex social networks. Phys A: Stat Mech Appl 506:395–411
Hosni AIE, Li K (2020) Minimizing the influence of rumors during breaking news events in online social networks. Knowl Based Syst 193:105452. https://doi.org/10.1016/j.knosys.2019.105452
Zhao J, Yang LX, Zhong X, Yang X, Wu Y, Tang YY (2019) Minimizing the impact of a rumor via isolation and conversion. Phys A: Stat Mech Appl 526:120867. https://doi.org/10.1016/j.physa.2019.04.103
Fang Q, Chen X, Nong Q, Zhang Z, Cao Y, Feng Y et al (2020) General rumor blocking: an efficient random algorithm with martingale approach. Theor Comput Sci 803:82–93. https://doi.org/10.1016/j.tcs.2019.05.044
Li W, Bai Q, Liang L, Yang Y, Hu Y, Zhang M (2021) Social influence minimization based on context-aware multiple influences diffusion model. Knowl Based Syst 227:107233. https://doi.org/10.1016/j.knosys.2021.107233
Yang D, Liao X, Shen H, Cheng X, Chen G (2018) Dynamic node immunization for restraint of harmful information diffusion in social networks. Phys A: Stat Mech Appl 503:640–649. https://doi.org/10.1016/j.physa.2018.02.128
Ju W, Chen L, Li B, Chen Y, Sun X (2021) Node deletion-based algorithm for blocking maximizing on negative influence from uncertain sources. Knowl Based Syst 231:107451. https://doi.org/10.1016/j.knosys.2021.107451
Wu P, Pan L (2017) Scalable influence blocking maximization in social networks under competitive independent cascade models. Comput Netw 123:38–50. https://doi.org/10.1016/j.comnet.2017.05.004
Ni P, ZHu J, Wang G (2020) Disinformation diffusion activity minimization by edge blocking in online social networks. Chinese journal of management science (10):1–10
Ding L, Hu P, Guan ZH, Li T (2021) An efficient hybrid control strategy for restraining rumor spreading. IEEE Trans Syst Man Cybern Syst 51(11):6779–6791. https://doi.org/10.1109/TSMC.2019.2963418
Korkmaz G, Kuhlman CJ, Ravi SS, Vega-Redondo F (2018) Spreading of social contagions without key players. World Wide Web 21(5):1187–1221. https://doi.org/10.1007/s11280-017-0500-y
Yang L, Li Z, Giua A (2020) Containment of rumor spread in complex social networks. Inf Sci 506:113–130. https://doi.org/10.1016/j.ins.2019.07.055
Domingos P, Richardson M (2001) Mining the network value of customers. In: Proceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. KDD ’01. New York, NY, USA: Association for Computing Machinery. pp 57–66
Kempe D, Kleinberg J, Tardos E (2003) Maximizing the spread of influence through a social network. In: Proceedings of the Ninth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. KDD ’03. New York, NY, USA: Association for Computing Machinery. pp 137–146
Litou I, Kalogeraki V, Katakis I, Gunopulos D (2017) Efficient and timely misinformation blocking under varying cost constraints. Online Social Networks and Media 2:19–31. https://doi.org/10.1016/j.osnem.2017.07.001
Liu W, Chen L, Chen X, Chen B (2020) An algorithm for influence maximization in competitive social networks with unwanted users. Appl Intell 50(2):417–437. https://doi.org/10.1007/s10489-019-01506-4
Pham CV, Phu QV, Hoang HX, Pei J, Thai MT (2019) Minimum budget for misinformation blocking in online social networks. J Comb Optim 38(4):1101–1127. https://doi.org/10.1007/s10878-019-00439-5
Zhang P, Bao Z, Niu Y, Zhang Y, Mo S, Geng F et al (2019) Proactive rumor control in online networks. World Wide Web 07(22):1799–1818
Yan R, Li Y, Wu W, Li D, Wang Y (2019) Rumor blocking through online link deletion on social networks. ACM Trans Knowl Discov Data 13(2). https://doi.org/10.1145/3301302
Ghoshal AK, Das N, Das S (2019) Misinformation containment in OSNs leveraging community structure. In: 2019 IEEE 10th International Conference on Awareness Science and Technology (iCAST). pp 1–6
Lv J, Yang B, Yang Z, Zhang W (2019) A community-based algorithm for influence blocking maximization in social networks. Cluster Computing 05:22. https://doi.org/10.1007/s10586-017-1390-6
Yang D, Chow TWS, Zhong L, Zhang Q (2018) The competitive information spreading over multiplex social networks. Phys A: Stat Mech Appl 503:981–990
Hosni AIE, Li K, Ahmad S (2020) Minimizing rumor influence in multiplex online social networks based on human individual and social behaviors. Inf Sci 512:1458–1480
Manouchehri MA, Helfroush MS, Danyali H (2021) A theoretically guaranteed approach to efficiently block the influence of misinformation in social networks. IEEE Trans Comput Soc Syst 8(3):716–727. https://doi.org/10.1109/TCSS.2021.3059430
Zhu J, Ni P, Wang G, Li Y (2021) Misinformation influence minimization problem based on group disbanded in social networks. Inf Sci 572:1–15
Yan R, Li D, Wu W, Du DZ (2020) Minimizing influence of rumors by blockers on social networks: algorithms and analysis. IEEE Trans Netw Sci Eng 7(3):1067–1078
Wang Z, Yang Y, Pei J, Chu L, Chen E (2017) Activity maximization by effective information diffusion in social networks. IEEE Trans Knowl Data Eng 29(11):2374–2387. https://doi.org/10.1109/TKDE.2017.2740284
Hosseini-Pozveh M, Zamanifar K, Naghsh-Nilchi AR, Dolog P (2016) Maximizing the spread of positive influence in signed social networks. Intell Data Anal 20(1):199–218
Ghoshal AK, Das N, Das S (2021) Influence of community structure on misinformation containment in online social networks. Knowl Based Syst 213:106693
Tripathy RM, Bagchi A, Mehta S (2010) A study of Rumor Control Strategies on Social Networks. New York, USA: Association for Computing Machinery
Berger J, Schwartz EM (2011) What drives immediate and ongoing word of mouth? J Mark Res 48(5):869–880. https://doi.org/10.1509/jmkr.48.5.869
Berger J (2013) Contagious: why things catch on. Simon and schuster
Zanette DH (2002) Dynamics of rumor propagation on small-world networks. Phys Rev E 65:041908. https://doi.org/10.1103/PhysRevE.65.041908
Vishnupad SP, Shin YC (1999) Adaptive tuning of fuzzy membership functions for non-linear optimization using gradient descent method. J Intell Fuzzy Syst 7:13–25
Dagum P, Karp R, Luby M, Ross S (2000) An optimal algorithm for monte carlo estimation. SIAM J Comput 29(5):1484–1496
Conforti M, Cornuéjols G (1984) Submodular set functions, matroids and the greedy algorithm: tight worst-case bounds and some generalizations of the rado-edmonds theorem. Discret Appl Math 7(3):251–274
Cornuéjols G, Fisher M, Nemhauser G (1977) Location of bank accounts to optimize float: an analytic study of exact and approximate algorithms. Manag Sci 04(23):789–810
Leskovec J, Kleinberg J, Faloutsos C (2005) Graphs over time: densification laws, shrinking diameters and possible explanations. In: Proceedings of the Eleventh ACM SIGKDD International Conference on Knowledge Discovery in Data Mining. KDD ’05. New York, NY, USA: Association for Computing Machinery. pp 177–187
Shu K, Wang S, Tang J, Zafarani R, Liu H (2017) User identity linkage across online social networks: a review. SIGKDD Explor Newsl 18(2):5–17. https://doi.org/10.1145/3068777.3068781
Deng K, Xing L, Zheng L, Wu H, Xie P, Gao F (2019) A user identification algorithm based on user behavior analysis in social networks. IEEE Access 7:47114–47123. https://doi.org/10.1109/ACCESS.2019.2909089
Nurgaliev I, Qu Q, Bamakan SMH, Muzammal M (2020) Matching user identities across social networks with limited profile data. Front Comput Sci 14(6):1–14
Rossi RA, Ahmed NK (2015) The network data repository with interactive graph analytics and visualization. In: Proceedings of the twenty-ninth AAAI conference on artificial intelligence. AAAI’15. AAAI Press. pp 4292–4293
Tang L, Wang X, Liu H (2009) Uncovering groups via heterogeneous interaction analysis. In: ICDM 2009 - The 9th IEEE International Conference on Data Mining. Proceedings - IEEE International Conference on Data Mining, ICDM; 2009. p. 503–512. 9th IEEE International Conference on Data Mining, ICDM 2009 ; Conference date: 06-12-2009 Through 09-12-2009
De Domenico M, Lima A, Mougel P, Musolesi M (2013) The anatomy of a scientific rumor. Scientific reports. vol 3(1)
Sabidussi G (1966) The centrality index of a graph. Psychometrika 31(4):581–603
Brandes U (2008) On variants of shortest-path betweenness centrality and their generic computation. Soc Netw 30(2):136–145
Page L, Brin S, Motwani R, Winograd T (1999) The pagerank citation ranking: bringing order to the web. Stanford infolab. pp 1999–66
Crucitti P, Latora V, Marchiori M, Rapisarda A (2004) Error and attack tolerance of complex networks. Phys A: Stat Mech Appl 340(1):388–394
Karp RM (2010) In: Reducibility among combinatorial problems. Springer, Berlin, pp 219–241
Zhu J, Ni P, Wang G (2020) Activity minimization of misinformation influence in online social networks. IEEE Trans Comput Soc Syst 06:1–10
Il’ev VP (2001) An approximation guarantee of the greedy descent algorithm for minimizing a supermodular set function. Discret Appl Math 114(1):131–146
Acknowledgements
This work was supported in part by the National Natural Science Foundation of China under Grant No. 72074203.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of Interests
The authors have no competing interests to declare that they are relevant to the content of this article.
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix A:: Proof
Appendix A:: Proof
1.1 A. 1 Proof of theorem 1
Proof
The problem of MIE-m tries to minimize the number of entities ultimately activated by the misinformation,
which is equivalent to maximizing the number of entities that are not influenced by \(\mathfrak {R}\) such that \(g({\varLambda })=\vert V \vert - \mathbb {E} \left [\sigma _{\mathfrak {R}}({\varLambda })\right ]\). We prove this by reducing the problem from the NP-complete set cover problem [50]. We set |V | = m. Let a ground set V = {v1,v2,⋯ ,vm} and a collection of sets Λ = {Λ1,Λ2,⋯ ,Λy}, where \(\cup _{j=1}^{y}{\varLambda }_{j} = V\). The set cover problem is to determine whether the union of K sets in Λ is equal to V. Next, we will show that the set cover problem can be regarded as a special case of the MIE-m problem. Given an arbitrary instance of the set cover problem, we construct a directed graph with m(n + 1) + y nodes. For each subset Λj we construct a related node aj, for each element vz (1 ≤ z ≤ m), construct n + 1 nodes \(u_{z},{u_{z}^{1}},\cdots ,{u_{z}^{n}}\), and create a directed edge \((u_{z},{u_{z}^{i}})\) for each node \({u_{z}^{i}} (1\leq i\leq n)\) with probability \(p_{u_{z}{u_{z}^{i}}}=1\). When element vz belongs to Λj, we create a directed edge (aj,uz) with probability \(p_{a_{j} u_{z}}=1\). Since the influence probability between nodes is 1, the dissemination of information is a fixed process. Therefore, the set covering problem is equivalent to deciding whether there are K nodes of Λ. □
1.2 A.2 Proof of theorem 2
Proof
When the number of online social networks is one, that is, n = 1, the account in an online social network can be equivalent to the entity. At this time, the problem of MIE-m is equivalent to the problem of misinformation influence minimization in single online social networks. We already know that the influence spread computation problem in online social network under the IC model is #P-hard [15]. Since the problem of misinformation influence minimization in online social networks is a special case of the problem of MIE-m, the influence spread computation problem in multi-social networks is also #P-hard, that is, computing \(\sigma _{\mathfrak {R}}({\varLambda })\) is #P-hard in multi-social networks. □
1.3 A.3 Proof of proposition 2
Proof
We use the formula [51] to calculate the probability \(\vartheta (w, \mathfrak {R})\) of w being activated by \(\mathfrak {R}\) to prove it. In multi-social networks G(G1,G2,⋯ ,Gn), given \(\mathfrak {R}\), we can get \(\vartheta (w, \mathfrak {R})= 1-{\prod }_{i=1}^{n} {\prod }_{v^{i}\in N^{in}(w^{i})} [1-P^{fwd}(v^{i}) P^{inf}(v^{i},w^{i})]\) for any w ∈ V, where Nin(wi) is the set of parent neighbours of entity w in online social network Gi. Then, we obtain that \(\vartheta (w, \mathfrak {R}) =1- {\prod }_{v\in N^{in}(w)} {\prod }_{i=1}^{n} [1-P^{fwd}(v^{i}) P^{inf}(v^{i},w^{i})] = 1- {\prod }_{v\in N^{in}(w)} 1-\hat {p}(v,w) = 1- {\prod }_{\hat {v}\in N^{in}(\hat {w})} 1-\hat {p}(\hat {v},\hat {w}) = \vartheta (\hat {w}, \mathfrak {\hat {R}})\) for any \(\hat {w}\in \hat {V}\), where Nin(w) is the set of parent neighbors of entity w. Since \(V=\hat {V}\), we have \(\sigma _{\mathfrak {R}}({\varLambda })={\sum }_{w\in V} \vartheta _{{\varLambda }}(w, \mathfrak {R})={\sum }_{\hat {w}\in \hat {V}} \vartheta _{{\varLambda }}(\hat {w}, \mathfrak {R})=\sigma _{\mathfrak {\hat {R}}}(\hat {{\varLambda }})\) for all \(\hat {{\varLambda }}\subseteq \hat {V}\), and the proposition follows. □
1.4 A.4 Proof of theorem 6
The proof framework is based on [52], but the supermodular function that makes the proof applicable to the problem of MIE-c requires some changes in the following. Given a coupled social network \(G_{cou}(\hat {V}, \hat {E})\), an initial influence entities \(\mathfrak {R}\) and a nonnegative nonincreasing supermodular function F(⋅) with supermodular curvature ϱF. We set \(U = \hat {V} \backslash \mathfrak {R}\), then F(U) = 0. Let \(W,M \subseteq U\), W = {w1,w2,⋯,wa} and M = {m1,m2,⋯ ,mb}, where Wi = {w1,w2,⋯ ,wi} (i = 1,⋯ ,a) and Ms = {m1,m2,⋯ ,ms} (s = 1,⋯ ,b) are sequences. For any \(W\subseteq U\), we set bw(W) = F(W∖w) − F(W).
Lemma 2
For any \(W\subseteq U\), \(F(W)={\sum }_{w_{j} \in U\backslash W} b_{w_{j}} (W_{a+j})\).
Proof
By the definition of b⋅(⋅), we can get \(F(W)= F(W_{a})= F(W_{a+1} )+ b_{w_{a+1}}(W_{a+1}) =b_{w_{a+1}}(W_{a+1})+ F(W_{a+2} ) + b_{w_{a+2}}(W_{a+2}) ={\cdots } = F(U) + {\sum }_{w_{j} \in U\backslash W} b_{w_{j}} (W_{a+j})\). Since F(U) = 0, the lemma is proved immediately. □
Lemma 3
For any \(W,M \subseteq U\), it holds
Proof
F(M ∩ W) = F(M ∩ Wa) = F((M ∩ Wa+ 1)∖wa+ 1). If wa+ 1 ∈ M∖W, we obtain F((M ∩ Wa+ 1)∖wa+ 1 \()=F(M\cap W_{a+1} )+ b_{w_{a+1}}(M \cap W_{a+1} )\). For any wa+ 1∉M, we have F({M ∩ Wa+ 1}∖wa+ 1) = F(M ∩ Wa+ 1),⋯. Finally, we can deduce that \(F(M\cap W)=F(\{ M \cap W_{a} \} \cup \{ M\backslash W \}) + {\sum }_{w_{j}\in M\backslash W} b_{w_{a+j}} (M \cap \) \(W_{a+j})= {\sum }_{w_{j} \in M\backslash W} b_{w_{a+j}} (M\cap W_{a+j}) + F(M)\). Then the first equation proof is complete. Using a similar method, we can get \(F(W\cap M)=F(W)+\sum \nolimits _{m_{z} \in W\backslash M} b_{m_{z}} (W\cap M_{b+z})\).
Given that Opt is an optimal solution to the problem of MIE-c, the TG algorithm consecutively acquires the sequences \(C_{0} = \varnothing \), C1 = {c1}, ⋯, Ci = {c1,c2,⋯ ,ci}, ⋯, CK = {c1,c2,⋯ ,cK}. Suppose r = |U|−|Ci|, \(\overline {W} =U \backslash W\), \(\overline {W}_{i} =U \backslash W_{i}\), and \(w_{j} \in \overline {W}_{i}\). Without causing ambiguity, we abbreviate \(b_{c_{i}}(C_{i})\) as bi. Then, we can derive Theorem 7. □
Theorem 7
For i = 1, 2,⋯ ,K, it holds
where \(s_{i} = \vert \overline {Opt} \cap \overline {C_{i}} \vert \) and \(\eta = \frac {{\varrho }^{F} }{1- {\varrho }^{F}}\).
Proof
Given any \(W,\ M \subseteq U\), from Lemma 3, we have \(F(M) = F(W)+ {\sum }_{m_{z} \in W \backslash M} b_{m_{z}} (W \cap M_{b+z})\) \(-{\sum }_{w_{j} \in M \backslash W} b_{w_{j}}(M \cap W_{a+j})\). By the definition of supermodular and ϱF, we have \(b_{w_{j}} (M \cap W_{a+j}) \leq b_{w_{j}}(w_{j}) \leq \frac {1}{1 - {\varrho }^{F}} b_{w_{j}}(U) \leq \frac {1}{1 - {\varrho }^{F}} b_{w_{j}}(W_{a+j})\). Since \(b_{m_{z}} (W \cap M_{b+z}) \geq b_{m_{z}} (W)\) for all mz ∈ W∖M. Then, we can obtain \(F(M) \geq {\sum }_{w_{j} \in \overline {W} } b_{w_{j}} (W_{a+j}) +{\sum }_{m_{z} \in \overline {M} \backslash \overline {W}} b_{m_{z}} (W)- \frac {1}{1-{\varrho }^{F}} {\sum }_{w_{j} \in \overline {W} \backslash \overline {M}} b_{w_{j}}(W_{a+j})\) \(={\sum }_{w_{j} \in \overline {W} \cap \overline {M} } b_{w_{j}} (W_{a+j}) +{\sum }_{m_{z} \in \overline {M} \backslash \overline {W}} b_{m_{z}} (W)- \frac {{\varrho }^{F} }{1- {\varrho }^{F}} {\sum }_{w_{j} \in \overline {W} \backslash \overline {M}} b_{w_{j}}(W_{a+j})\).
Let M = Opt, W = Ci and \(\eta = \frac {{\varrho }^{F} }{1- {\varrho }^{F}}\), we can get \(F(Opt) \geq {\sum }_{m_{z} \in \overline {Opt} \backslash \overline {C_{i}}} b_{m_{z}} (C_{i}) + {\sum }_{c_{j} \in \overline {C_{i}} \cap \overline {Opt} } b_{c_{j}} (C_{i+j}) \) \(-\eta {\sum }_{c_{j} \in \overline {C_{i}} \backslash \overline {Opt}} b_{c_{j}}(C_{i+j})\). By definition of Opt and Ci, for any \(m_{z} \in \overline {Opt} \backslash \overline {C_{i}}\), \(b_{m_{z}} (C_{i}) \geq b_{c_{i}} (C_{i})\). Therefore, \({\sum }_{m_{z} \in \overline {Opt} \backslash \overline {C_{i}}} b_{m_{z}} (C_{i}) \geq \vert \overline {Opt} \backslash \overline {C_{i}} \vert \cdot b_{c_{i}} (C_{i}) = \vert \overline {Opt} \backslash \{ \overline {Opt} \cap \overline {C_{i}} \} \vert \cdot b_{i} = (r-s_{i}) b_{i}\), where \(s_{i} = \vert \overline {Opt} \cap \overline {C_{i}} \vert \). Finally, we get \(F(Opt) \geq (r-s_{i}) b_{i} + {\sum }_{j: c_{j} \in \overline {C_{i}} \cap \overline {Opt} } b_{j} - \eta {\sum }_{j: c_{j} \in \overline {C_{i}} \backslash \overline {Opt}} b_{j}\). □
Given C = CK, \(\overline {C}=\{c_{1},\cdots ,c_{j},\cdots , c_{r}\}\) and \(\overline {C_{j}} =\{c_{1},c_{2},\cdots ,c_{j} \}\). Let \(\overline {Opt} \cap \overline {C} = \{ u_{1}, u_{2}, \cdots , u_{s}\}\), where s ≤ r, that is {u1,u2,⋯ ,us} be the elements not contained in Opt or C. Suppose \(b_{j}= F(C\cup \{\overline {C_{j}}\backslash c_{j}\})-F(C\cup \overline {C_{j}})\). By Lemma 2, we can obtain \(F(C)={\sum }_{j: c_{j} \in \overline {C} } b_{j}\). Then, the approximation ratio is defined as \(\frac {F(C)}{F(Opt)} = {\sum }_{j: c_{j} \in \overline {C} } \frac {b_{j}}{F(Opt)}\). Define \(y_{j} := \frac {b_{j}}{F(Opt)}, j \in [r]\). Since bj ≥ 0, then yi ≥ 0. We define \(L(\overline {Opt} \cap \overline {C}) = {\sum }_{j: c_{j} \in \overline {C} } \frac {b_{j}}{F(Opt)}\).
Considering 1 ≤ s ≤ r, for the variables yj, there are r constraints. Hence, the worst-case approximate ratio of \(\frac {F(C)}{F(Opt)}\) can be expressed as the following Linear Programming (LP).
Lemma 4
L({u1,u2,⋯ ,us− 1}) ≥ L({u1,u2,⋯ ,us}) for any s = 1, 2,⋯ ,r.
Proof
We abbreviate L({u1,u2,⋯ ,us}) as Ls for simplicity. Let cj = us. When yj > 0, an optimal solution for LP Ls has the following form \(Y_{s}^{*} = \{y_{1},\cdots , y_{j-1}, y_{j}, y_{j}, y_{j}\frac {r-s+\eta }{r-s}, \cdots , y_{j} (\frac {r-s+\eta }{r-s})^{r-j-1} \}\). We construct a feasible solution Ys− 1 of Ls− 1. Suppose the first j elements of Ys− 1 are consistent with the first j elements of \(Y_{s}^{*}\), then Ys− 1 can be expressed as \( \{ y_{1},\cdots , y_{j-1}, y_{j}, y_{j}\frac {r-s+\eta +1}{r-s+1}, \cdots , y_{j}(\frac {r-s+\eta +1}{r-s+1})^{r-j} \}\). We can obtain \(L^{*}_{s}- L_{s-1}= y_{j}+ y_{j}\frac {r-s+\eta }{r-s}+ \cdots + y_{j} (\frac {r-s+\eta }{r-s})^{r-j-1} - y_{j}\frac {r-s+\eta +1}{r-s+1}- \cdots - y_{j}(\frac {r-s+\eta +1}{r-s+1})^{r-j} = y_{j}[1- \frac {r-s+\eta +1}{r-s+1}] +{\cdots } + y_{j}[(\frac {r-s+\eta }{r-s})^{r-j-1} -(\frac {r-s+\eta +1}{r-s+1})^{r-j}] \leq 0\). When yj = 0, the form of the optimal solution of LP Ls can be written as \(Y_{s}^{*} = \{y_{1},\cdots , y_{j-1}, 0, y_{j-1}\frac {r-s+\eta +1}{r-s},\) \( y_{j-1}\frac {r-s+\eta +1}{r-s} \frac {r-s+\eta }{r-s}, \cdots , y_{j-1}\frac {r-s+\eta +1}{r-s} (\frac {r-s+\eta }{r-s})^{r-j-1} \}\). Similarly, let the first j − 1 items of Ys− 1 be consistent with the first j − 1 items of \(Y_{s}^{*}\), then Ys− 1 can be expressed as \( \{y_{1},\cdots , y_{j-1}, y_{j-1}\frac {r-s+\eta +1}{r-s+1}, \cdots , y_{j-1} (\frac {r-s+\eta +1}{r-s+1})^{r-j+1} \}\). Then \(L^{*}_{s}- L_{s-1} \leq 0\). Hence, L({u1,u2,⋯ ,us− 1}) ≥ L({u1,u2,⋯ ,us}) for any s = 1, 2,⋯ ,r. □
Theorem 8
Given that C is the solution obtained by the greedy algorithm, Opt is the optimal solution of MIE-c, which satisfies
where ϱF is the supermodular curvature of set function F.
Proof
From LP (A1), it can be inferred that F(C) ≤ F(Opt) ⋅ L({u1,⋯ ,us}) for any s = 1,⋯ ,r. By lemma 4, we can obtain \(L(\{ u_{1}, \cdots , u_{s}\}) \leq L(\{ u_{1}, \cdots , u_{s-1}\}) \leq {\cdots } \leq L(\varnothing )\). Hence, \(F(C) \leq F(Opt)\cdot L(\varnothing )\). When \(\overline {Opt} \bigcap \overline {C}= \varnothing \), based on LP (A1), we can derive \(L(\varnothing )= \frac {1}{r} + \frac {r+\eta }{r \cdot r}+{\cdots } + \frac {(r+\eta )^{j}}{r \cdot r^{j}} +{\cdots } + \frac {(r+\eta )^{r-1}}{r \cdot r^{r-1}} = \frac {1}{\eta } [\frac {(r+\eta )^{r}}{r^{r}} -1] = \frac {1}{\eta } [(1+\frac {\eta }{r})^{r} -1] \leq \frac {1}{\eta } [e^{\eta } -1]\). Let \(\eta = \frac {{\varrho }^{F}}{1-{\varrho }^{F}}\), and the theorem is proved immediately. □
Rights and permissions
About this article
Cite this article
Ni, P., Zhu, J. & Wang, G. Misinformation influence minimization by entity protection on multi-social networks. Appl Intell 53, 6401–6420 (2023). https://doi.org/10.1007/s10489-022-03798-5
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10489-022-03798-5