ranking
改进desity计算
SpecGreedy: Uni?ed Dense Subgraph Detection
缺点：these algorithms still incur a prohibitive computational cost for the massive
graphs that arise in modern data science applications, without considering the
properties of real-world data.
G = (V, E) with |V | = n. Let S ? V and E(S) be the edges of subgraph G(S) induced by the subset S, i.e. E(S) = {
e
i
j
e_{ij}
eij? :
v
i
,
v
j
v_i , v_j
vi?,vj? ∈S ∧
e
i
j
e_{ij}
eij? ∈ E}. Let A = (
a
i
j
a_{ij}
aij?) ∈ R, n×n be the adjacency matrix of G and
a
i
j
a_{ij}
aij? ≥ 0.
G’: 同一组节点构造的差异较大的图。EG动态图在两个不同时刻的snapshot
- Generalized Densest Subgraph Problem
- remark: 可疑最密集群体检测问题的处理
新用户的cold start
Generating Behavior Features for Cold-Start Spam Review Detection
-
Real Behavior Features (RBFs)
- activity window (AW)
- maximum number of reviews
- percentage of positive reviews (PR)
- review count (RC)
- reviewer deviation (RD)
- maximum content similarity (MCS)
-
Easily Accessible Features (EAFs)
- text features (TF)
- rating features (RF)
- attribute features (AF).
Generator： The 1st three layers are used to do normalization and get EAFs. We then use three non-linear hidden layers to transform EAFs into SBFs.
Discriminator：judge the (EAF+, RBF) pairs from the realistic training data as real and the (EAF+, SBF) pairs from the generator as fake.
embedding
马尔科夫
ColluEagle: Collusive review spammer detection using Markov random fields