Neural Network models I am using for the outlier detection are ANN, CNN, LSTM and GRU. CNN is correctly predicting the normal data but missing too much outlier data, due to which Recall comes too low.
Mukhtar Hussain I hope this helps you consider several helpful approaches to improving the recall of your neural network models for outlier detection on the ALOI dataset, particularly when CNNs are missing too many outliers.
One strategy is to address the issue of class imbalance. Outliers are typically rare, and this imbalance may cause the model to focus on the majority class, leading to low recall. Techniques such as class weighting, undersampling the majority class, or oversampling the minority class can help the model become more sensitive to outliers. Data augmentation is also valuable in this scenario. Since outliers are scarce, generating synthetic outliers using methods like SMOTE or generative models (like GANs) can assist in teaching the model how to recognize these anomalous (strange) patterns more effectively.
Using a different loss function might also help. Instead of relying on a standard loss like cross-entropy, applying loss functions tailored to anomaly detection, such as Focal Loss, will place more emphasis on harder-to-classify samples—like outliers. Similarly, adjusting the classification threshold can improve recall by making the model more sensitive to detecting outliers. For example, lowering the threshold from the typical 0.5 to something lower may increase sensitivity.
Another way to approach this issue is by incorporating Autoencoders, which are particularly suited for outlier detection tasks. By training an autoencoder on normal data and using the reconstruction error as an outlier score, it becomes easier to detect outliers based on deviations. Modifying the model architecture can also yield improvements. For instance, combining CNNs with LSTM or GRU layers could enhance the model's ability to capture patterns that indicate anomalies. Feature engineering can also play a key role. Extracting features that explicitly capture anomalies, such as through dimensionality reduction techniques like PCA or t-SNE, may help the model focus on more relevant patterns.
With these strategies, you should see an improvement in recall and overall performance in detecting outliers within the ALOI dataset.
I can send you a bulleted outline if you like me to! Cheers