Item Anomaly Detection Based on Dynamic Partition for Time Series in Recommender Systems

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Mining

Background:

• Recommender systems combat information overload but face challenges like shilling attacks.
• Shilling attacks inject fake profiles to manipulate recommendations, characterized by extreme ratings and rapid injection.
• Existing detection methods require improvement in detection rate, false alarm rate, and universality.

Purpose of the Study:

• To propose a novel item anomaly detection method for recommender systems.
• To address the limitations of current shilling attack detection techniques.
• To enhance the robustness and accuracy of recommendation algorithms.

Main Methods:

• Dynamic partitioning of item-rating time series based on salient temporal points.
• Utilizing chi-square (χ2) distribution for detecting anomalous rating intervals.
• Evaluating the method on MovieLens 100K and 1M datasets.

Main Results:

• The proposed method achieves a high detection rate for shilling attacks.
• A low false alarm rate was observed, minimizing the misclassification of genuine users.
• The approach demonstrated stability across various attack models and filler sizes.

Conclusions:

• Dynamic partitioning combined with chi-square testing offers an effective solution for shilling attack detection.
• This method improves the reliability and accuracy of recommender systems.
• The findings contribute to more trustworthy and robust recommendation technologies.