In search of better deep Recommender Systems
- 1. In search of a better Recommender System using Deep Learning by SK Reddy Chief Product Officer AI linkedin.com/in/sk-reddy/
- 2. where R - observed ratings, X - user embeddings, Y - item embeddings, λ is regularization parameter Explicit feedback including general bias and biases for user and item Implicit Feedback X, Y are user and item embeddings http://dsnotes.com/post/2017-05-28-matrix-factorization-for-recommender-systems/ Basics m users, n items, and an extremely sparse rating matrix R Additional side information matrix of user and item are denoted by The objective function of matrix factorization can be written as
- 3. Types of machine learning algorithms used in recommender systems https://arxiv.org/pdf/1511.05263.pdf
- 4. Google Play Recommendation System https://arxiv.org/pdf/1606.07792.pdf Apps recommendation pipeline overview Wide & Deep model Over 500 billion examples Model’s prediction using logistic regression where Y is the binary class label, σ(·) is the sigmoid function, φ(x) are the cross product transformations of the original features x, and b is the bias term. A. is the vector of all wide model weights, and are the weights applied on the final activations a (lf ) . Memorization: learning the frequent co-occurrence of items and exploiting the correlation. Generalization: transitivity of correlation and explores new feature combinations that have never or rarely occurred in the past.
- 5. http://www.kdd.org/kdd2017/papers/view/embedding-based-news-recommendation-for-millions-of-users Yahoo News Recommender Yahoo! JAPAN’s homepage on smartphones Encoder for triplets of articlesBrowsing history and session LSTM-based model • Start with distributed representations of articles based on a variant of the denoising autoencoder • Generate user representations by using an RNN with browsing histories as input sequences • Match and list articles for each user based on the inner product of article-user for relevance and article-article for de-duplication
- 7. Additional Stacked Denoising Autoencoder(aSDAE) Additional Denoising Autoencoder(aDAE) Ref: A Hybrid Collaborative Filtering Model with Deep Structure for Recommender Systems. The structure of the proposed hybrid model. The model contains three components: the upper component and the lower component are two aSDAEs which extract latent factor vectors for users and items respectively; the middle component decomposes the rating matrix R into two latent factor matrices Hybrid model based on Autoencoders Loss function of Autoencoder Loss function of Denoising Autoencoder (ADE) Loss function of additional Denoising Autoencoder (aDAE)