The article "Deep learning based recommender system: A survey and new perspectives. Zhang, S., Yao, L., Sun, A., & Tay, Y. (2019)." is a survey that reviews literature regarding the advances of deep learning on the recommender systems field. The purpose of the article is not to, merely present the information, but to foster innovation on this topic by presenting rough guidelines to help choose what deep neural networks to use on a given recommendation task.
Up to section 3.4 the reading has been very thorough on presentation and classification of deep learning techniques applied on recommender systems. The article first presents different paradigms on recommendation, and deep learning techniques in an independent manner before merging those fields together. Benefits and potential limitations are also presented but the authors present themselves on an optimistic position towards the usage of deep learning in recommender systems. Their position is evident in the presentation of potential limitations with their question: "Are there really any drawbacks and limitations with using deep learning for recommendation?" and by the fact that every potential limitation pointed out is accompanied by a counterargument.
Afterwards, deep learning techniques are ordered by categories and discussed later on. It is great how the authors not only show applications of DL on the field but they correctly integrate traditional RecSys models and techniques with the new DL techniques. The authors do not discard the previous technical knowledge on the field from the argument of the article. It is all the way round, the previous technical knowledge is presented on the same level of importance than the new DL techniques. For instance some models use the principles of matrix factorization to represent user and items with latent vectors that serve as input for deep learning models. Some other models are ensembles of factorization machines with MLP, others such as AutoSVD++, use the hidden layer of autoencoders to build the low rank matrices of Matrix Factorization models. Examples like this can be found throughout the reading.
The way the authors present the models indeed fulfills their purpose to encourage readers to be creative and innovate on new proposals that incorporate DL. For me, the model that got my attention was the MLP based model for makeup recommendation due to the fact that it was able to mix expert rules with labeled examples. I always thought that expert rules can be extremely useful to enchance the recommendations in certain domains such as content-based music generation by incorporating music theory or item recomendations in video games by incorporating professional players knowledge and rules (item statistics, synergies, etc..).
Lastly, I want to point out that the article despite being a survey that requires to cite thoroughly other articles, I believe that many in-text citations made the reading hard to understand. The narrative depends a bit too much on other readings, so in order to understand it completely, the reader is almost forced to check the references to keep up the reading or to know all the techniques in advance.