![During forward pass or
inference, the model takes
user represented by his
vector of ratings from the
training set x ∈ ℝn, where
n is number of items
x = [s1, s2, s3,..., sn]
Recommender Model: Model
10](https://image.slidesharecdn.com/trainingdeepautoencodersforcollaborativefiltering4-181031032359/85/Training-deep-auto-encoders-for-collaborative-filtering-10-320.jpg)
Training deep auto encoders for collaborative filtering
- 1. Training Deep AutoEncoders for Collaborative Filtering Marlesson Santana Prof. Dr. Anderson Soares
- 2. • Introduction • Objective • Recommender Model • Results and discussion • Conclusion Outline 2
- 3. RecSys are a subclass of information filtering system that seek to predict the ‘rating’ or ‘preference’ that user would give to an item. Sites like amazon, Netflix and Spotify use RecSys to suggest items to users. Introduction: What is a RecSys ? 3
- 4. RecSys can be divided into different types: ● Content-Based ● Collaborative filter ● Cold-start ● Hybrid Introduction: Types of RecSys 4
- 5. The classic CF problem is what tries to infer the missing entries in an mxn matrix, whose (i,j) entry describes the ratings given by the ith user to the jth item. Introduction: Collaborative Filter 5
- 6. There are many non deep learning types of approaches to CF, Matrix factorization techniques is the most popular. Several recent approaches use autoencoders for RecSys, including the current state-of-the-art. Introduction: RecSys with DeepLearning 6
- 7. This paper proposes a Deep Autoencoder Model for Collaborative Filtering for rating prediction task in recommender system. Objective 7
- 8. The dataset used was the original Netflix Prize, separated into subsets of training and testing with differing periods Recommender Model: Dataset 8
- 9. Encoder and decoder parts of the autoencoder consist of feed-forward neural networks with classical fully connected layers computing: l = f(W * x + b) Recommender Model: Model 9
- 10. During forward pass or inference, the model takes user represented by his vector of ratings from the training set x ∈ ℝn, where n is number of items x = [s1, s2, s3,..., sn] Recommender Model: Model 10
- 11. The optimize Masked Mean Squared Error loss (MMSE) where, ri is actual rating, yi is reconstructed and mi is a mask function such that mi = 1 if ri != 0 else mi = 0. Recommender Model: Loss function 11
- 12. Results and discussion: Activation functions 12 There are two properties which seems to separate activations: ● non-zero negative part ● unbounded positive part
- 13. Results and discussion: Size of hidden units 13 Single layer autoencoder with 128, 256, 512 and 1024 hidden units in the coding layer. A: training RMSE per epoch; B: evaluation RMSE per epoch.
- 14. Results and discussion: Dropout 14 This model quickly over-fits if trained with no regularization. Very high values of drop probability (0.8) turned out to be the best
- 15. Results and discussion: Comparison 15
- 16. Results and discussion: Inference of RecSys 16 1. Given sparse x, x ∈ ℝn 2. Compute dense f(x) = decode(encode(x)) 3. Filters only f(x)i where xi = 0 4. Recommends the highest value of f(x)i
- 17. • This paper demonstrated how very deep autoencoders can be successfully trained for RecSys • On the task of future rating prediction, the model outperforms other approaches even without using additional temporal signals • It has been demonstrated the importance of the activation function and the regularization by dropout Conclusion 17
- 18. Obrigado(a)! 18