Recommender systems

Neural Collaborative Filtering (NCF) is an advanced recommendation method that improves on traditional collaborative filtering by using deep learning models. Instead of relying purely on user-item interactions (like a traditional collaborative filtering matrix), NCF uses the power of neural networks to learn complex relationships between users and items.

Let’s use the following matrix of user ratings as an example:

In a traditional User-based Collaborative Filtering, we might recommend a movie to Li Wei based on Jake’s similar ratings(see example in collaborative filtering).For example, since Jake and Li Wei have both rated M1, M2, M3, and M5 similarly, the system would recommend M4 to Li Wei, which Jake has rated highly but Li Wei has not yet watched. However, NCF takes this a step further by using deep learning to better model interactions between users and items. It does so by embedding both users and items into dense vectors and then learning a neural network that captures their complex relationships.

In NCF, each user (Jake, Li Wei, Carlos, and Miguel) is mapped to a dense vector representation. For example, Jake might be represented by a vector like [0.5, 1.2, -0.3], while Li Wei might be represented as [0.4, 1.0, -0.2]. These vectors capture their interaction history in a compressed form. Each movie (M1, M2, M3, M4, M5) is also mapped to a dense vector representation. For instance, M1 could have a vector like [0.7, -0.1, 0.9], and M4 might be represented as [0.9, 0.2, 0.4].

The embeddings for a user (say Li Wei) and a movie (say M4) are fed into a neural network. The network combines the user and item vectors and learns to predict the likelihood of Li Wei enjoying M4 based on their previous interaction patterns with other items (or users). Once trained, the NCF model can predict the rating Li Wei would give to M4 by learning nonlinear patterns that are harder to capture in traditional methods. This goes beyond just looking at similar users; it can understand deeper, more nuanced relationships between users and items. In this case, NCF may predict that Li Wei would give M4 a high rating (say 4 or 5), confirming that M4 is a suitable recommendation based on more complex learned patterns.

NCF (Neural Collaborative Filtering) can predict user-item relationships more accurately by capturing nonlinear patterns and contextual information. It analyzes user interactions with similar items and adapts to changes in user behavior over time. NCF also considers implicit signals and social connections, as well as demographic and psychographic information to provide tailored recommendations.

An autoencoder is a neural network that learns how to compress and reconstruct input data. In the context of recommendation systems, its purpose is to take a user-item interaction matrix, like a matrix of user ratings for movies, and reduce it to a more straightforward, compressed form that captures the essential information.

The encoder takes the rating matrix, and reduces it to a smaller version, named a latent representation.This smaller version highlights the essential patterns in the data, such as user preferences or movie characteristics. The decoder takes this data and tries to reconstruct the orginal input, filling in missing values. The key idea behind an autoencoder is that by forcing the network to learn a compressed version of the data, it also knows the most important relationships between users and items, which can then be used to make recommendations. Essentially, the autoencoder is learning to represent complex data (like user ratings) in a more straightforward form that still captures the most critical features and then uses that representation to predict missing information. The information that needs to be added can be ratings or preferences. For example, if Jake has rated movies M1, M2, and M3 but hasn’t rated M4, the autoencoder will try to predict Jake’s rating for M4. This prediction is based on how Jake’s preferences (captured in the latent space) relate to the characteristics of M4, as well as how similar users have rated that movie. By reconstructing the entire user-item matrix, including the missing ratings, the autoencoder helps fill in gaps in the data, which can then be used to generate personalized recommendations for each user.

Recurrent Neural Networks (RNNs) are well-suited for sequential recommendation systems where the order of user interactions is important. In many scenarios, user preferences evolve over time, and understanding the sequence of interactions helps in predicting what users might be interested in next. RNNs are able to maintain a memory of previous interactions, which is useful for capturing temporal dynamics in user behavior.

Consider a user who watched M1, then M2, and then M3. The RNN model processes these interactions in a sequence, learning to predict the next possible movie the user might want to watch. This prediction is based on how similar sequences of movies have been watched by other users. In our example, if Li Wei watched M1, M2, and M3, the RNN might predict that M4 would be a suitable recommendation, as it has often followed similar sequences for other users. By maintaining a memory, RNNs can handle changing user interests more effectively.

RNNs are powerful in contexts like video streaming services or e-commerce, where the order of interactions significantly influences user behavior. By learning sequences, RNN-based recommenders can provide more personalized and timely suggestions, anticipating what users are likely to do next.

The Joker movie poster is an example of how Convolutional Neural Networks (CNNs) extract features. CNNs are are good at identifying essential details from different types of content, like visual and written data. The feature extraction process has a few stages, starting with the original movie poster as the input. Convolution Layers find features such as edges, colors, and textures at different levels of complexity. Then, Pooling Layers summarize and condense the information, followed by a Flattening Layer that turns the 2D feature maps into a 1D vector of 100,352 features. Finally, the Fully Connected Layer combines all the extracted features for final processing. The results from the CNN can be used to identify a movie’s visual style, themes, and genre characteristics. CNNs help recommendation systems understand rich multimedia content, resulting in more accurate and context-aware suggestions.

Transformers are now being used in recommendation systems to process sequences of user actions effectively. Initially designed for language processing, the Transformer architecture uses self-attention mechanisms to understand relationships between inputs. Transformers are very effective in recommendation tasks where understanding sequential patterns is crucial. Transformers can analyze how users like or skip content and adjust their context-based recommendations. A Transformer-based recommendation system can process user actions in sequence and assign varying levels of importance based on the context.

In our movie recommendation example, an attention-based model focuses more on recent interactions highly relevant to a user’s interests. For instance, if Miguel recently watched M4 and rated it highly, an attention mechanism would assign greater weight to this interaction when predicting what he might watch next. For transformers to work effectively, they need a variety of data inputs, such as user profile information (demographics), contextual data like time of day, type of device, and location, sequence data (a series of actions over time), and information about the content the user engaged with, such as categories, descriptions, and content details.