Future Directions for Anomaly Detection

In this note, I summarize a list of promising future directions for anomaly detection based on the survey

Future Work

Rare and Imbalanced Data

Recent works have addressed this problem usually employing either data augmentation or a tailored learning method, where a higher weight is attributed to the least represented class.

Data Augmentation

Data augmentation is gerenated by either replication or simulation. Previous works attempted to utilize GAN for generate virtual data categorically. In addition, a novel method, based on Dynamic Time Warping (DTW) has been proposed, which can be achieved by the following steps:

1. Randomly sample multiple instances from a category
2. Rescale each instance
3. Randomly generate a weight vector
4. Compute the weighted DTW average using the weight vector and the intances

Weighted Learning

Researchers have previously attempted to utilize the calss weights in objective function while training to cope with the issue of data imbalance. An elaborated approach is proposed to map the input to clusters in an embedding space, balancing the learning by incorporating inter-cluster and inter-class margins. Then, the approach uses a novel objective function to learn the embeddings, qualified as triple-header cross-entropy. This approach has proven to be effective for SED with imbalanced data.

Recently, researchers also proposed a novel approach for hazardous events detection using type-2 fuzzy set. Since the hazardous events are a minority in the dataset, the weights of such classes have been manipulated in the type-2 fuzzy membership function, such that the weights of each class are inversely proportional to the class samples. Then, a membership degree is calculated using two components, an upper/optimistic membership component and a lower/pessimistic component. Finally, interval comparison is performed to select the event for which the membership is the highest.

Temporal Dimension

A previous research attempted to study the effect of temporal evolution of the input signal on the performance of audio surveillance RNN models. The results of the research show that sequential models oare not neccessarily the best fit for temporal data and that optimization of the temporal dimension is still an open problem.

Computational Efficiency

A recent work developed a novel parallel method for extracting features from spectrograms using MapReduce model in Hadoop. The obtained results show a better computational time efficiency and a higher recognition rate of critical acoustic events in different noisy conditions.