This is another paper that applies deep neural network for collaborative filtering problem. Although there are a few outstanding deep learning models for CF problems such as CF-NADE and AutoRec, the author claims that those models are solving for explicit feedback and positioned this work to solve for ‘implicit feedback CF’ problem.

The model is straightforward and similar to Neural Factorization Machine. The idea is to find embedding vectors for users and items and model their interaction as a non-linear function via a multi-layers neural network. A non-linear interaction has been commonly used in many recent works such as Neural Factorization Machine [2], Deep Relevant Matching Model [3].

The authors proposed 3 models incrementally: (1) Generalized Matrix Factorization – which is basically MF with additional non-linear transformation. In this case, they use sigmoid function; (2) Multi-layer Perceptron – which concatenate user and item embedded vectors and transform them by a learned non-linear function; (3) Fusion model can be either shared the same embedding vectors and add them up at the last layers or learned separate embedding vectors and concatenate them at the output layer.

Since they want to predict either the given item is preferable or not, it is a binary classification problem. Then the loss function can be a binary cross-entropy. To me, implicit feedback seems to be an easier problem than rating prediction problem because we only need to make a binary prediction.

The baseline seems okay but I wish the authors include more recent deep learning models such as [4]. The AutoRec model is also applicable for an implicit feedback by forcing the output to be a binary output. Regardless of their baseline, the extensive experiments tried to convince readers that deep neural network can model a complex interaction and will improve the performance.

Speak of the performance, the author uses hit-ratio and NDCG. Basically, there is one test sample for each user. The model tries to give a high score for that sample. This metric is more practical than MSE and RMSE since the dataset is implicit feedback dataset.

Non-linear interaction is a simple extension to a traditional MF model. This author shows that this simple extension does work for MovieLens and Pinterest dataset.

**Reference:**

[1] He, Xiangnan, et al. “Neural collaborative filtering.” Proceedings of the 26th International Conference on World Wide Web. International World Wide Web Conferences Steering Committee, 2017.

http://www.comp.nus.edu.sg/~xiangnan/papers/ncf.pdf

[2] Xiangnan He, Tat-Seng Chua, “Neural Factorization Machines for Sparse Predictive Analytics”, ACM SIGIR 2017.

[3] Guo, Jiafeng, et al. “A deep relevance matching model for ad-hoc retrieval.” Proceedings of the 25th ACM International on Conference on Information and Knowledge Management. ACM, 2016.

[4] Zheng, Yin, et al. “Neural Autoregressive Collaborative Filtering for Implicit Feedback.” Proceedings of the 1st Workshop on Deep Learning for Recommender Systems. ACM, 2016. APA