def compute_shapley_value(self,t,idxs,**kwargs):
V_S_t=kwargs['V_func']
util={}
powerset=list(powersettool(idxs))
# TMR round truncation below
S_0=()
util[S_0]=V_S_t(t=t,S=S_0)
S_all=powerset[-1]
util[S_all]=V_S_t(t=t,S=S_all)
if abs(util[S_all]-util[S_0])<=self.round_trunc_threshold:
sv_dict={id:0 for id in idxs}
return sv_dict
# TMR round truncation above
for S in powerset[1:-1]:
util[S]=V_S_t(t=t,S=S)
self.SV_t[t]=self.shapley_value(util,idxs)
self.Ut[t]=copy.deepcopy(util)
return self.SV_t[t]
def compute_shapley_value(self, idxs, **kwargs):
V_S_0 = kwargs['V_func']
N = len(idxs)
self.Contribution_records = []
powerset = list(powersettool(idxs))
util = {}
S_0 = ()
util[S_0] = V_S_0(S=S_0)
S_all = powerset[-1]
util[S_all] = V_S_0(S=S_all)
k = 0
while self.isnotconverge(k):
k += 1
v = [0 for i in range(N + 1)]
v[0] = util[S_0]
marginal_contribution_k = [0 for i in range(N)]
idxs_k = np.random.permutation(idxs)
for j in range(1, N + 1):
# key = C subset
C = idxs_k[:j]
C = tuple(np.sort(C, kind='mergesort'))
#truncation
if abs(util[S_all] - v[j - 1]) >= self.eps:
if util.get(C) != None:
v[j] = util[C]
else:
v[j] = V_S_0(S=C)
else:
v[j] = v[j - 1]
# record calculated V(C)
util[C] = v[j]
# update SV
marginal_contribution_k[idxs_k[j - 1] - 1] = v[j] - v[j - 1]
self.Contribution_records.append(marginal_contribution_k)
# shapley value calculation
shapley_value = (
np.cumsum(self.Contribution_records, 0) /
np.reshape(np.arange(1,
len(self.Contribution_records) + 1),
(-1, 1)))[-1:].tolist()[0]
# store round t results
self.SV = {key + 1: sv for key, sv in enumerate(shapley_value)}
self.Ut = copy.deepcopy(util)
return self.SV
def compute_shapley_value(self, idxs, **kwargs):
V_S_D = kwargs['V_func']
util = {}
powerset = list(powersettool(idxs))
for S in powerset:
util[S] = V_S_D(S=S)
self.SV = self.shapley_value(util, idxs)
self.Ut = copy.deepcopy(util)
return self.SV
def compute_shapley_value(self, t, idxs, **kwargs):
#timer only
self.st_t = time.time()
V_S_t = kwargs['V_func']
util = {}
powerset = list(powersettool(idxs))
for S in powerset:
util[S] = V_S_t(t=t, S=S)
#for print only
self.full_set = powerset[-1]
self.SV_t[t] = self.shapley_value(util, idxs)
self.Ut[t] = copy.deepcopy(util)
self.print_results(t)
return self.SV_t[t]
def compute_shapley_value(self,t,idxs,**kwargs):
V_S_t=kwargs['V_func']
N=len(idxs)
powerset=list(powersettool(idxs))
util={}
S_0=()
util[S_0]=V_S_t(t=t,S=S_0)
S_all=powerset[-1]
util[S_all]=V_S_t(t=t,S=S_all)
# group test relate
last_uds=[]
Z=0
for n in range(1,N):
Z+=1/n
Z*=2
UD=np.zeros([N,N],dtype=np.float32)
p=np.array([N/(i*(N-i)*Z) for i in range(1,N)])
k=0
while self.isnotconverge_Group(last_uds,UD) or k<self.CONVERGE_MIN_K:
k+=1
len_k=0
# 1. draw len_K ~ q(len_k)
len_k=np.random.choice(np.arange(1,N),p=p)
# 2. sample S with len_k
S=np.random.choice(idxs,size=len_k,replace=False)
# 3. M(S) + V(S)
S=tuple(np.sort(S,kind='mergesort'))
if util.get(S)!=None:
u_S=util[S]
else:
u_S=V_S_t(t=t,S=S)
# 4. Group Testing update UD
UD=(k-1)/k*UD
for i in range(0,N):
for j in range(0,N):
delta_beta=S.count(i+1)-S.count(j+1)
if delta_beta!=0:
value=delta_beta*u_S*Z/k
UD[i,j]+=value
last_uds.append(UD)
u_N=util[S_all]
#timer
st=time.time()
#timer
shapley_value=self.solveFeasible(N,u_N,UD)
#timer
dura=time.time()-st
print('Solve Feasible using %.3f seconds'%dura)
#timer
self.Ut[t]=copy.deepcopy(util)
self.SV_t[t]={key+1:sv for key,sv in enumerate(shapley_value)}
return self.SV_t[t]