# Neural Collaborative Filtering

This paper discusses the application of deep neural networks (DNNs) in creating a <mark style="color:yellow;">more sophisticated model for collaborative filtering</mark>, a key technique in recommender systems.&#x20;

{% embed url="<https://arxiv.org/abs/1708.05031>" %}
Neural Collaborative Filtering - cited over 6,000 times
{% endembed %}

#### <mark style="color:green;">**Context and Importance of Recommender Systems**</mark>

In today's world, where users are bombarded with vast amounts of information online, <mark style="color:yellow;">recommender systems play a crucial role in helping users navigate this information overload</mark>. These systems are widely used in various online platforms like e-commerce, online news, and social media to personalise user experiences by recommending items based on past interactions.

#### <mark style="color:green;">**Collaborative Filtering and Matrix Factorization**</mark>

The core of personalised recommender systems lies in <mark style="color:blue;">**collaborative filtering**</mark>, which models users' preferences on items based on their past interactions (like ratings and clicks).&#x20;

One popular technique in collaborative filtering is <mark style="color:blue;">**Matrix Factorization (MF)**</mark>, which *<mark style="color:yellow;">**represents both users and items as vectors in a shared latent space**</mark>*.&#x20;

The interaction between a user and an item is predicted by the inner product of their respective latent vectors.&#x20;

Despite its popularity, the effectiveness of MF is limited by the simplicity of the inner product as the interaction function.

#### <mark style="color:green;">**Enhancements to Matrix Factorization**</mark>

Various efforts have been made to enhance MF's performance.

Some approaches integrate MF with neighbour-based models, others combine it with topic models of item content, and some extend MF to factorization machines for a more general feature modelling.&#x20;

For example, incorporating user and item bias terms into the interaction function has been shown to improve MF's performance in rating prediction tasks.

#### <mark style="color:green;">**Deep Neural Networks for Interaction Function Learning**</mark>

The paper proposes using deep neural networks to learn the interaction function between users and items directly from the data, rather than relying on handcrafted functions.&#x20;

DNNs have demonstrated effectiveness in approximating any continuous function and have shown promising results in various domains like computer vision and speech recognition.

### <mark style="color:purple;">**Neural Collaborative Filtering (NCF)**</mark>

The authors introduce a general framework called <mark style="color:blue;">**Neural Collaborative Filtering (NCF)**</mark> that employs a neural network architecture to model the latent features of users and items for collaborative filtering.&#x20;

They illustrate that traditional MF can be seen as a specific case of NCF.&#x20;

By using a multi-layer perceptron, NCF can introduce a higher level of non-linearities into the model, potentially capturing more complex user-item interactions than MF.

<mark style="color:green;">**Focus on Implicit Feedback**</mark>

The paper focuses on <mark style="color:yellow;">**implicit feedback**</mark>, which reflects users' preferences through their behaviour (like clicks or purchases) rather than explicit feedback (like ratings or reviews). Implicit feedback is more readily available but challenging to use due to the absence of direct user satisfaction indicators and a scarcity of negative feedback.

<mark style="color:green;">**Contributions and Experiments**</mark>

The main contributions of this work include presenting the NCF framework that uses DNNs for collaborative filtering and demonstrating that MF is a specialisation within NCF.

Extensive experiments on real-world datasets validate the effectiveness of NCF and highlight the potential of deep learning in enhancing collaborative filtering techniques.

In summary, this paper advances the field of recommender systems by integrating deep learning with collaborative filtering, offering a more nuanced and powerful approach to modeling user-item interactions, particularly with implicit feedback data.