Neural Collaborative Filtering

Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, Tat-Seng Chua
National University of Singapore

WWW'17

Motivation (1/2)

• The exploration of deep Learning techniques has received relatively less scrutiny on recommender systems comparing to other research domains.​
• Previous works only apply deep learning techniques to model auxiliary information.
  • Textual descriptions of items
  • Acoustic feature of musics
  • Visual content of images
• Few works focus on modeling user-item interaction using deep learning techniques.

Motivation (2/2)

• Matrix factorization models user-item interaction via a simple inner product.
• However, inner product may not be sufficient for modeling user-item interaction.
  • Inner product is no more than a linear function.
  • The importance of every latent dimension is the same.
  • Neural network has been proven to be capable of approximating any continuous function.
• Try to learn a user-item interaction function via deep neural network on implicit feedback.

sim(2, 3) > sim(1, 2) > sim(1, 3)

sim(4, 1) > sim(4, 3) > sim(4, 2)

Methodology (1/13)

• Neural Collaborative Filtering (NCF) Framework
  • Generalized Matrix Factorization (GMF)
    • Linearity
  • Multi-Layer Perceptron (MLP)
    • Non-linearity

Methodology (2/13)

• General Framework
  • Input layer
    • One-hot encoded vector
  • Embedding layer
    • Fully-connected
  • Neural collaborative filtering layer
    • The dimension of the last hidden layer X (termed as predictive factors) determines the model’s capability.

Methodology (3/13)

• General Framework
  • Output layer
    • Pointwise Loss (this work)
    • Pairwise Loss
  • User-item interaction function                                                        

Methodology (4/13)

• General Framework
  • Objective function (Pointwise)
    • Previous work: Regression with weighted squared loss
      • Assumption
        • Ratings are generated from Gaussian distribution.
      • It may not tally well with implicit feedback.

Methodology (5/13)

• General Framework
  • Objective function (Pointwise)
    • ​Binary cross-entropy loss (Log loss)
      • Address recommendation with implicit feedback as a binary classification problem.

Methodology (6/13)

• Generalized Matrix Factorization (GMF)
  • MF can be interpreted as a special case of our NCF framework.
  • Setting
    • Activation function: identity function (i.e.                    )
    • Weight vector: uniform vector of 1 (i.e. <1,1,...,1>)

Methodology (7/13)

• Generalized Matrix Factorization (GMF)
  • A generalized and extended MF
    • Activation function
      • Select ​sigmoid function to enforce non-linearity.
    • ​Weight vector
      • ​Learning from data with log loss.
      • Allow varying importance of latent dimensions.

Methodology (8/13)

• Multi-Layer Perceptron (MLP)
  • Simple concatenation of user embedding and item embedding does not account for any user-item interaction.
  • Adding hidden layers may be a better choice.

Methodology (9/13)

• Multi-Layer Perceptron (MLP)
  • Activation function
    • ReLU
      • Non-saturated
      • Sparse activation
        • Well suited for sparse data and prevent overfitting.
  • Network structure
    • Tower structure
      • Halve the layer size for each successive higher layer.
      • More abstractive features of data can be learned by using a small number of hidden units for higher layers.

Methodology (10/13)

• Fusion of GMF and MLP
  • Goal
    • Design a structure that allows GMF and MLP to mutually reinforce each other.
  • Staightforward solution
    • Let GMF and MLP to share the same embedding layer.
      • Cons
        • They must use the same size of embeddings.

Methodology (11/13)

• Fusion of GMF and MLP
  • Proposed solution
    • Neurual Matrix Factorization (NeuMF)
      • Let GMF and MLP to learn separate embeddings and combine them by concatenating their last hidden layer.
      • Use back-propagation to learn model parameters.

Methodology (12/13)

• Fusion of GMF and MLP
  • Pre-training
    • Train GMF and MLP with random initializations until convergence.
    • Use their model parameters as the initialization for the corresponding parts of NeuMF’s parameters.
    • Trade-off between the two pre-trained models.

Methodology (13/13)

• Fusion of GMF and MLP
  • Learning rate adaption
    • Pre-training
      • Adaptive Moment Estimation (Adam)
        • Adapt the learning rate by performing smaller updates for frequent and larger updates for infrequent parameters.
    • Training
      • Vanilla SGD

Experiments (1/7)

• Datasets
  
  
  
   
  • Binarize the rating data to implicit data.
  • Remove users with #(feedbacks) < 20.
• Evaluation
  • Leave-one-out evaluation
  • Randomly sample unobserved 100 items for each user and rank them with the test item.

Experiments (2/7)

• Evaluation metrics
  • HR @ 10
  • NDCG @ 10
• Competitors
  • ItemPop
  • ItemKNN
  • Bayesian Personalized Ranking (BPR)
  • element-wise Alternating Least Squares (eALS)
    • Optimize weighted squared loss.
    • Treat all unobserved feedback as negative feedback and non-uniformly give weights by item's popularity.

Experiments (3/7)

• Implementation
  • Github By Keras
• Parameter Settings
  • Negative sampling: 4 negative instances per positive instance.
  • Embedding size: 16
  • Batch size: {128, 256, 512, 1024}
  • Learning rate: {0.0001, 0.0005, 0.001, 0.005}
  • Predictive factors: {8, 16, 32, 64}
  • #(hidden layers): 3
  • Trade-off parameter for pre-training: 0.5

Experiments (4/7)

• NCF outperforms the state-of-the-art implicit collaborative filtering methods.
  • For BPR and eALS, the number of predictive factors is equal to the number of latent factors.
  • GMF outperforms BPR.
    • It shows the effectiveness of the classification aware log loss for the recommendation task.

Experiments (5/7)

• Pre-training is important for NCF.

Experiments (6/7)

• Log loss with negative sampling does work for recommendation tasks.
  • The sampling method for pointwise objective may be more flexible than pairwise objective.
    • Pointwise: arbitrary number per positive instance
    • Pairwise: only one per positive instance
  • Setting the sampling ratio too aggressively may adversely hurt the performance.

Experiments (7/7)

• Deep learning is helpful for recommendation tasks.
  • Stacking more layers are beneficial to performance.
  • But the authors only show results of #(hidden layers) <= 4...
• Transforming latent factors with hidden layers are essential.
  • ​Simple concatenation (MLP-0 in the table) yields poor results.

Conclusion and Future Works

• This work opens up a new avenue of research possibilities for recommendation based on deep learning.
  • A general framework NCF and three instantiations are proposed to solve the key collaborative filtering task.
    • GMF, MLP and NeuMF
• Future work
  • Model auxiliary information
  • Efficient online recommendation