The document discusses a content-based approach to classify spam SMS messages using graph-based text representation and KNN algorithms. It details the process of feature selection, including mutual information and chi-statistic, to weigh tokens that indicate whether a message is spam or normal. The methodology also highlights the evolution of spam messages and the need to consider communication patterns alongside content for accurate classification.

1. Tran Phuc Ho, Ho-Seok Kang, Sung-Ryul Kim
Journal of Universal Computer Science, vol. 19, no. 16 (2013)
*

2. *
* Spam SMS : advertisements by commercial
companies, hacking messages for cheating and
stealing personal information.
* Content-based approach
Graph-based
Text representation
KNN
algorithm

3. spam
normal
Labeled
small
message
groups
5 messages (in real time, only 1 message)
Tokenize them by white spaces
and punctuations
*

4. *
* remove the noisy features and select the good
ones
Mutual information(MI),
X2-Statistic (CHI)

5. *
The dependence between a word(t) and a type of message(c)
t : token (word or phrase)
c : class (type of message – spam or ham)
The probability that t and c
co-occur
The conditional probability of t in c
Probability of t

6. *
The lack of independence between a word(t) and a type of message(c)
t : token (word or phrase)
c : class (type of message – spam or ham)
Probability of t
The probability that
t and c co-occur
t t
Probability that the text belong to c

7. *
* calculate the weight of each feature
*Use the high weighted words for constructing
the graphs
CHI(X2-statistic)
MI(Mutual Information)

8. *
Token selected
by feature selection
- unique word
G = (V, E, FWN)
V :set of nodes
E :set of weighted edges linking the nodes
FWN :feature weight matrix – weight of edges and nodes
The order &
Co-occurrence relationship
Between two feature words
(If feature words co-occur
within a step length, assign
an edge)

9. *
G = (V, E, FWN)
V :set of nodes
E :set of weighted edges linking the nodes
FWN :feature weight matrix – weight of edges and nodes
Weight of edges, Probability of tokens represented by nodes
W_ij : co-occurrence frequency of two feature words
f_i and f_j within a step length
Only calculate the
weight W_ij (i>j).
Ex) scientific paper
Zero
Ex) paper scientific
Frequency of single words

10. *
in K nearest neighbors of the text T to be classified, the class of T is the most
frequently appearing class in this collection
1. Build sample graphs (elements)
2. New message comes in
3. Build a testing graph
Similarity
Of two graphs
-> Feature Weight :
Weights of the edges
+ weight of the edge itself
(appear in the two graphs)

11. *
Testing graph (g) Sample graphs (sg_1, sg_2, … sg_n)
….
List (RL)
1 FW(tg,sg1)=2 Spam
2 FW(tg,sg2)=3 Spam
… FW(tg,sg3)=4 Normal
K FW(tg,sg4)=5 Spam
(Nfp : how many nodes in the sample
graph with their weights larger than 0
also appear in the test graph)
If Nfp > threshold, calculate FW(tg,sg1)
0.0001
3

12. *
Testing graph (g) Sample graphs (sg_1, sg_2, … sg_n)
….
List (RL)
1 FW(tg,sg5)=6 Spam
2 FW(tg,sg2)=3 Spam
… FW(tg,sg3)=4 Normal
K FW(tg,sg4)=5 Spam
If Nfp > threshold, calculate FW(tg,sg5)
6
Spam message

13. *
NUS SMS Corpus (5,574 messages)
– 4,827 normal(86.6%), 747 spam(13.4%)
[Uysal and Yildiz] SMS
collection
(875 messages)
- 450 normal, 425 spam

14. *

15. *
(%)
(seconds)

16. *
* Spam SMS messages are evolving.. Hard to
capture keywords.
* ex) 대★출, 이ㅈr, <<통>> / <<장>>, no space or
punctuation, no specific keyword, same content
with other phone numbers, no words only with
image …
* Graph patterns of communication between
sender and receiver should be added with
content-based approach.