You are on page 1of 5

Optimization in Social Networks: Exercises

9th Winter School on Network Optimization, January 20-24, 2020

Exercise 1 Influence Maximization Problem (ICM)


Consider the influence maximization problem based on the independent cascade model and the
scenario modeling approach (slide 14/15). We stated a covering formulation for this problem (slide
21). Try to find alternative (mixed) integer linear programming formulations based on
(a) time-indexed variables (see slide 43),

(b) cycle-elimination (see slide 44).

Solution:
(a) time-indexed model:
X X
max pω xωi,|V |
ω∈Ω0 i∈V
X
zi ≤ K
i∈V
xωi,0 ≤ zi ∀i, ω
X
xωi,t ≤ xωj,t−1 + xωi,t−1 ∀i, t = 1, . . . , |V |, ω
(j,i)∈Aω

xωi,t ∈ {0, 1} ∀i, t, ω


zi ∈ {0, 1} ∀i

(b) cycle-elimination model:


X X
max pω xωi
ω∈Ω0 i∈V
X
zi ≤ K
i∈V
X
xωi ≤ ω
yji + zi ∀i, ω
(j,i)∈Aω

yijω ≤ xωi ∀(i, j) ∈ Aω , ω


X
yijω ≤ |C| − 1 ∀ cycles C ∈ Aω , ω
(i,j)∈C

xωi ∈ {0, 1} ∀i, ω


yijω ∈ {0, 1} ∀(i, j) ∈ Aω , ω
zi ∈ {0, 1} ∀i

Exercise 2 Competitive Influence Maximization (ICM)


Consider the competitive influence maximization problem (CIMP, slide 32/33). The covering formu-
lation for the non-competitive IMP can be easily adapted to the CIMP by re-defining the reachability
sets (slide 34). Try to adapt the formulations from Exercise 1 to the CIMP.
Solution:
Let Tiω be the earliest time at which the leader can activate node i in scenario ω. The follower needs
to reach it earlier if it wants to activate it.
(a) time-indexed model:
X X
max pω xωi,|V |
ω∈Ω0 i∈V
X
zi ≤ K
i∈V \L

xωi,0 ≤ zi ∀i ∈ V \ L, ω
X
xωi,t ≤ xωj,t−1 + xωi,t−1 ∀i, 1 ≤ t < Tiω , ω
(j,i)∈Aω

xωi,t ∈ {0, 1} ∀i, t, ω


zi ∈ {0, 1} ∀i ∈ V \ L

(b) cycle-elimination model: without notion of time (e.g., in time-indexed variables, reachability
sets) it is not obvious how to model the CIMP with arc variables and cycle-elimination in each
live-arc graph.

Exercise 3 Rumor Blocking (ICM)


Assume a similar setting as for the competitive influence maximization problem (slide 32), but here
the leader (let’s call him “Tromp”) is spreading “fake news” from his seed set L starting in time step
1. You realize this only in time step 5 and now want to block this spreading by selecting at most K
nodes which are convinced not to forward Tromp’s fake news. Your goal is to minimize the number
of nodes which are reached by Tromp’s fake news. Formulate an MILP (based on any modeling
concept) for this problem.

Solution:
For simplicity we use a time-indexed model but probably also some cover model might work. The
fact that we realize the fake news only in time step 5 is not a big issue, we just let grow the leader’s
seed set for 5 time steps and define the set of finally active nodes as new seed set L0 . We could
optionally also remove all nodes from the graph which are not reached by the leader since they are
not relevant here.
X X
min pω xωi,|V |
ω∈Ω0 i∈V
X
zi ≤ K
i∈V \L0

xωi,0 = 1 ∀i ∈ L0 , ω
xωj,t + zj ≥ xωi,t−1 ∀(i, j) ∈ Aω , t ≥ 1, ω
xωi,t ≥ xωi,t−1 ∀i, t ≥ 1, ω
xωi,t ∈ {0, 1} ∀i, t, ω
zi ∈ {0, 1} ∀i ∈ V \ L

Exercise 4 Budgeted Influence Maximization Problem (LTM)


We define a budgeted variant of the influence maximization problem (slide 14), but based on the
linear threshold model (slide 38): Each node i ∈ V has a cost ci ≥ 0 for including it in the seed set,
and a profit pi ≥ 0 which is obtained when the node gets active. We want to find a seed set with
costs not exceeding the budget B which maximizes the total profit obtained from activated nodes.
Formulate this problem as an MILP
(a) based on time-indexed variables (slide 43),
(b) based on cycle-elimination (slide 44),
(c) based on minimal influencing sets (slides 52/53).

Solution:
(a) time-indexed model:
X
max pi xi,|V |
i∈V
X
ci xi,0 ≤ B
i∈V
X
hi xi,t−1 + dji xj,t−1 ≥ hi xi,t ∀i ∈ V, t ≥ 1
(j,i)∈A

xi,t ∈ {0, 1} ∀i, t

(b) cycle-elimination model:


X
max p i xi
i∈V
X
ci y i ≤ B
i∈V
X
hi yi + dji zji ≥ hi xi ∀i ∈ V
(j,i)∈A

zij ≤ xi ∀(i, j) ∈ A
X
zij ≤ |C| − 1 ∀ cycles C ⊆ A
(i,j)∈C

xi , yi , zij ∈ {0, 1}

(c) set cover model: incentives are set to Pi = {0, hi } for all i.
X
max pi x i
i∈V
X
ci λ∅i ≤ B
i∈V
X
λUi = xi ∀i ∈ V
U ∈Λi
X
λUj = zij ∀(i, j) ∈ A
U ∈Λj :i∈U

zij ≤ xi ∀(i, j) ∈ A
X
zij ≤ |C| − 1 ∀ cycles C ⊆ A
(i,j)∈C
Exercise 5 Non-Progressive Linear Threshold Model
Consider the following non-progressive variant of the linear threshold model (slide 38): Each node
i ∈ V has a hurdle hit ≥ 0 potentially different for each time step t = 1, . . . , T , within some time
horizon T . Node i is active in time step t if
X
dji ≥ hit ,
j∈Ui,t−1

where Ui,t−1 is the set of neighbors which are active in time step t − 1. Therefore, a node’s state
might go from active to inactive again if the hurdle increases or the set of active neighbors decreases
in a later time step.
We now state the influence maximization problem (slide 14) based on such a non-progressive LTM.
The given objective function of maximizing the number of active nodes does not make sense anymore
in the non-progressive context. Instead we want to maximize the sum of active time steps over all
nodes within time horizon T . Formulate this problem as an MILP with time-indexed variables (slide
43).

Solution:
X X
max xi,t
i∈V t=1,...,T
X
xi,0 ≤ K
i∈V
X
hi,t xi,0 + dji xj,t−1 ≥ hi,t xi,t ∀i ∈ V, 1 ≤ t ≤ T
(j,i)∈A

xi,t ∈ {0, 1} ∀i, t

Exercise 6 Competitive Influence Maximization (LTM)


Try to think of a reasonable propagation model with two players based on the linear threshold model
(slide 38) in the flavor of the competitive independent cascade model (slide 32/33). Based on this
propagation model formulate the competitive influence maximization problem as an MILP (based
on any modeling concept).

Solution:
A potential competitive LTM might be: It should be progressive, all influence values and hurdles
hold for both players, and in case both players have enough influence to activate some node in some
time step, then the leader is prioritized.
Time-indexed model:
X
max yi,|V |
i∈V
xi,0 = 1 ∀i ∈ L
X
yi,0 ≤ K
i∈V \L

xi,t ≤ xi,t+1 ∀i, t


yi,t ≤ yi,t+1 ∀i, t
X
hi xi,t−1 + dji xj,t−1 ≥ hi xi,t ∀i, t ≥ 1
(j,i)∈A
X
hi yi,t−1 + dji yj,t−1 ≥ hi yi,t ∀i, t ≥ 1
(j,i)∈A
X
dji xj,t−1 − hi + 1 ≤ M (xi,t + yi,t−1 ) ∀i, t ≥ 1
(j,i)∈A

xi,t , yi,t ∈ {0, 1} ∀i, t

Note that we have to force the leader propagation, otherwise nothing would happen since we are
maximizing only the follower. For this we use the last set of Big-M inequalities.

Exercise 7 Heuristics for Influence Maximization (LTM)


For the influence maximization problem based on the independent cascade model, there is an efficient
greedy heuristic by Kempe et al. (2003) on slide 17. The greedy criterium is based on a node’s
marginal gain. For the linear threshold model (slide 38) there is no such marginal gain concept for a
node since usually only multiple nodes together can activate another node. Think of efficient greedy
heuristics for the influence maximization problem based on the LTM.

Solution:
One possibility for a reasonable greedy criterium for some node i might be to count the abso-
lute/relative amount of other nodes’ hurdle gaps
S node i is able to close. We define the hurdle gap
for node j ∈ V \ U based on active nodes U = k Uk as
X
rj := hj − dkj .
k∈Uj

Then, an absolute greedy criterium for node i could be based only on the direct out-neighbors of
some node, i.e., X
max{rj , dij },
j:(i,j)∈A,j∈V \U

or on the out-neighbors of all nodes W ⊆ V \ U which get newly activated by node i, i.e.,
X X
max{rj , dkj }.
j∈V \(U ∪{i}) k:(k,j)∈A,k∈W

It probably makes sense to normalize each term by dividing by rj .

You might also like