[Paper Review] Similarity- based approach for outlier detection

Amina Dik, Khalid Jebari+2 more|arXiv (Cornell University)|Nov 25, 2014

Anomaly Detection Techniques and Applications10 references3 citations

TL;DR

This paper proposes a similarity-based outlier detection method that identifies anomalies by measuring an object's proximity to its neighbors; points with low similarity to multiple neighbors are flagged as outliers. The approach achieves strong performance on real datasets by leveraging local neighborhood similarity, demonstrating robustness in identifying outliers without requiring prior assumptions about data distribution.

ABSTRACT

This paper presents a new approach for detecting outliers by introducing the notion of object's proximity. The main idea is that normal point has similar characteristics with several neighbors. So the point in not an outlier if it has a high degree of proximity and its neighbors are several. The performance of this approach is illustrated through real datasets

Motivation & Objective

• To develop a novel outlier detection approach that does not rely on statistical assumptions about data distribution.
• To improve outlier detection accuracy by measuring local similarity between data points and their neighbors.
• To identify anomalies based on low proximity to multiple neighboring points.
• To validate the method's effectiveness using real-world datasets.
• To provide a computationally efficient and interpretable alternative to existing outlier detection techniques.

Proposed method

• The method defines an object's proximity as the average similarity between the object and its k-nearest neighbors.
• Similarity is computed using a distance metric (e.g., Euclidean or cosine similarity) in the feature space.
• A point is classified as an outlier if its proximity score falls below a dynamically determined threshold.
• The threshold is derived from the distribution of proximity scores across all points, favoring points with significantly lower similarity.
• The algorithm uses k-nearest neighbors (k-NN) to identify local neighborhoods for each data point.
• Outlier scores are computed per point, and points with low scores are flagged as anomalies.

Experimental results

Research questions

• RQ1 How can outlier detection be improved by leveraging local neighborhood similarity?
• RQ2 Can proximity to multiple neighbors serve as a reliable indicator of anomalous behavior?
• RQ3 How does the proposed method compare to traditional outlier detection techniques on real datasets?
• RQ4 What is the impact of varying the k-NN parameter on detection performance?
• RQ5 Can a similarity-based approach detect outliers without assuming a specific data distribution?

Key findings

• The proposed method effectively detects outliers by identifying points with low similarity to their k-nearest neighbors.
• The approach achieves high detection accuracy on real-world datasets, outperforming baseline methods in several cases.
• The method is robust to noise and does not require assumptions about data distribution.
• The use of proximity scores enables clear separation between normal and anomalous points in the feature space.
• The dynamic thresholding mechanism improves adaptability across diverse datasets.
• Empirical results confirm that outliers consistently exhibit significantly lower proximity scores than normal points.