Secure Personalized Recommendation System for Mobile User - Information Security and Cryptology-ICISC 2010 - page 275

In the experiment, the similarity measure between users is computed by means of
Pearson Correlation Coefficient. Similarly, the similarity value is weighted and
selected the neighboring users that have the highest similarity rating with the active
user by using Significant Weighting and then computing the Trust Weighting. Finally,
a prediction from the rating of neighbor is computed.

Comparison of Prediction Values for Push Attack

6

4

2

0

1 3 5 7 9 11131517192123252729
Neighborhood S ize

-2

Prediction without Attack Prediction with Attack

Fig. 6. Comparison of Prediction Values for Push Attack

Comparison of Prediction Values for Nuke Attack

6

4

2

0

1 3 5 7 9 11131517192123252729
Neighborhood Size

P re dic tion without Atta c k P re dic tion with Atta c k

Fig. 7. Comparison of Prediction Values for Nuke Attack

Fig. 6 and 7 show the comparison of prediction values of the active user with and
without shilling attack. Before applying attack detection algorithm, the prediction
values are very high for the target item which is roughly 5 on account of the push
attack. After detecting push attack, the prediction values are actual rating values for
the active user. As shown in the comparison of prediction values for nuke attack in
Fig. 7, the prediction value is low for the target item and after detecting the nuke
attack gives the actual rating value. The neighborhood of the attackers is very sparse
with genuine users. The highly neighborhood is removed from the neighbor and then
the prediction result is computed. The modified RPCF algorithm can make the actual
prediction to the active user by detecting the push or nuke attack.
Finally, the accuracy of a prediction is evaluated by using Mean Absolute Error
(MAE). The lower MAE presents the more accurate prediction value. As shown in
Fig.8 and 9, while the fluctuation of the MAE value with the attack is high, the MAE
value without attack is quite stable. After detecting the push attacks and nuke attacks
show the value of MAE is much closed to 0 (zero). This means that modified RPCF