Recommender systems are used extensively by online businesses. Systems using neural networks may not be as effective as advertised.
Over the last 4-5 years deep learning has become a popular technique in recommender systems. Recent research conducted by Paolo Cremonesi and Maurizio Ferrari Dacrema has found that many published findings are difficult to replicate and tend not to prove significant improvement over simpler methods like K-nearest neighbours.
The difficulty in reproducing research is a broad problem faced across science. Another problem for recommender systems is that there is not a single, large data set like Imagenet that is used by all researchers. Lots of published results use deep learning, but it is not the best for performance - at least some of the reason for this is that deep-learning is ‘trendy’ and more likely to get accepted by journals and conferences.
🛎️ Why this matters: Unfortunately, we should take all published research with a pinch of salt. Beware fashions!📖 Read more (1129 words) 📖