Fig. 4.

Visualization of NDCG rank scores for different models reveals that, among the embedding approaches, FCGR serves as the baseline model and outperforms other methods, except the neural embeddings. The neural embeddings using ladder loss (convnet-ladder) exhibit the highest overall score, significantly surpassing the baseline and all other models. n represents the number of runs. For some approaches, n is smaller, e.g., the hashes, because there is no variance in the results compared to the deep learning architectures.