# Deep recommenders - Scalable and can learn complex interactions - Swap some design choices (number of neighbours, representations, etc) with others (type of NN, loss functions, etc) - Reproducibility crisis: Only 1 (MultVAE) of 11 deep models work as claimed [2] - Typical to use Implicit Feedback (lots of data to learn from!) ## MultVAE - [[Variational Autoencoders]] for [[Collaborative filtering]] [3] - Input: Users are represented using a binary vector - Dimension = Item is 1 if an interaction exists / 0 otherwise - Assume output distribution is Multinomial - Regular VAE Loss - Reconstruction - try to reconstruct the input vector - Regularisation - KL Divergence between prior (Isotropic Gaussian) and posterior distribution ### Experiments - ML-20M, Netflix, Million Songs Dataset - Explicit -> mplicit - If Ratings >=3.5 set to 1 or 0 otherwise - Baselines are non-neural (WMF and SLIM) and neural CDAE $ \begin{aligned} &\begin{array}{lccc} & \text { Recall@20 } & \text { Recall@50 } & \text { NDCG@100 } \\ \hline \text { Mult-VAE }^{\text {Pi }} & 0.266 & 0.364 & 0.316 \\ \text { Mult-DAE } & 0.266 & 0.363 & 0.313 \\ \hline \text { WMF } & 0.211 & 0.312 & 0.257 \\ \text { SLIM } & - & - & - \\ \text { CDAE } & 0.188 & 0.283 & 0.237 \\ \hline \end{array} \end{aligned} $ ### Advantages - All the advantages that come with DL - Powerful, Flexible, Scalable, etc - Cross-pollination - adoption of advances from ML/DL ### Drawbacks - Cold start still a problem - No consensus if performance gains are significant - Reproducibility [2] - 'Deep' models might not be needed - EASE - Embarrassingly shallow autoencoders [8] outperforms (most) neural baselines - Recent models are promising see RecVAE [9] (current SOTA) --- ## References [2] Dacrema, Maurizio Ferrari, Paolo Cremonesi, and Dietmar Jannach. "Are we really making much progress? A worrying analysis of recent neural recommendation approaches." Proceedings of the 13th ACM Conference on Recommender Systems. 2019 [3] Liang, Dawen, et al. "Variational autoencoders for collaborative filtering." Proceedings of the 2018 world wide web conference. 2018 .