Keras Recommenders: reliable, state-of-the-art recommendations for ranking and retrieval

Caris Recommenders for Building High-Quality Recommendation Systems

Key Concepts:

• Caris Recommenders (Caris RS): A Python library for building recommendation systems, built on top of Caris.
• Large Language Models (LLMs): Powerful models often prone to "hallucinations" (generating incorrect or nonsensical information), posing a challenge for reliable recommendations.
• Sequential Retrieval: A recommendation task focused on predicting the next item in a sequence (e.g., next movie a user will watch).
• Gated Recurrent Unit (GRU): A type of recurrent neural network (RNN) unit used for processing sequential data.
• Embedding Table: A lookup table that maps items (e.g., movies) to vector representations.
• Brute Force Retrieval: A layer in Caris Recommenders used to decode query embeddings and retrieve predictions.
• MovieLens Dataset: A widely used dataset for evaluating recommendation algorithms, containing movie ratings from various users.

Introduction & The Need for Recommendation Systems

The video highlights the pervasive nature of recommendation systems in modern online experiences – from e-commerce and social media to video streaming. While Large Language Models (LLMs) offer potential for building these systems, their tendency to "hallucinate" (generate inaccurate or misleading information) presents a significant challenge. Caris Recommenders (Caris RS) is presented as a solution to create reliable and high-performing recommendation systems, specifically addressing the hallucination issue. The speaker, Eufango, demonstrates this using a movie recommendation example.

What is Caris Recommenders?

Caris Recommenders is a Python library developed by the Caris team, designed to provide developers with building blocks for creating performant and accurate recommendation systems. It focuses on tasks like ranking and retrieval. A key benefit is its compatibility with TensorFlow, Jax, and PyTorch, leveraging the advantages of the underlying Caris framework. The library and its documentation can be found at caris.io/caris_rs. Installation is straightforward using pip install caris-rs. The documentation includes both web page examples and interactive Colab notebooks.

Sequential Retrieval & GRU Implementation

The example focuses on sequential retrieval, predicting the next item in a user’s sequence of interactions. This is applicable to scenarios like recommending products after a user browses a website, suggesting songs on a streaming service, or predicting the next social media content a user will engage with. The implementation utilizes a Gated Recurrent Unit (GRU), a type of recurrent neural network (RNN) unit, to process the sequential data and predict the next item. RNNs are designed to handle sequential data by maintaining a hidden state that captures information about past inputs. GRUs are a specific type of RNN known for their efficiency and ability to capture long-range dependencies in the data.

Data Preparation & Processing: The MovieLens Dataset

The demonstration uses the MovieLens dataset, a collection of movie ratings. The initial step involves transforming the data into sequences of movies for each user, ordered by timestamp. This is achieved by creating a dictionary mapping user IDs to lists of their movie reviews, sorted chronologically.

A crucial technique is then employed to expand the training data. For each user, multiple sequences are created by progressively increasing the length of the context used for prediction. This is likened to the game "snake," where each new review adds to the sequence length. For example, the first three reviews are used to predict the fourth, then the first four to predict the fifth, and so on. The final review for each user is reserved for a dedicated test dataset, ensuring it’s never seen during training.

The prepared data is then formatted into batches suitable for training, with each example consisting of a sequence of movie IDs (the context) and the next movie ID (the label).

Model Architecture & Training

The model architecture consists of two "towers":

1. Historical Movie Sequence Tower: Utilizes the GRU unit to process the sequence of past movies.
2. Candidate Movie Tower: An embedding table that maps each movie to a vector representation.

A brute force retrieval layer from Caris Recommenders is used to decode the query embeddings (generated by the GRU tower) and retrieve predictions from the candidate embeddings. This layer essentially calculates the similarity between the query embedding and the embeddings of all candidate movies, selecting the most similar ones as predictions.

The training process follows standard Caris procedures, involving instantiation, compilation, and execution of the training loop.

Prediction & Evaluation

After training, the model is used to generate predictions based on data from the test set. Users can also input their own sequences of favorite movies to evaluate the model’s recommendations. The video emphasizes that the Caris team provides numerous examples beyond this one, including deep and cross network models and two-tower embedding models for retrieval and ranking.

Notable Quote

“Since it’s built on top of Caris 3, it shares many of the benefits including compatibility with your choice of TensorFlow Jax or PyTorch.” – Eufango, highlighting the flexibility of Caris Recommenders.

Logical Connections

The video progresses logically from introducing the problem of hallucinations in LLM-based recommendation systems to presenting Caris Recommenders as a solution. It then details the specific implementation using sequential retrieval and a GRU model, emphasizing the importance of data preparation and model architecture. Finally, it demonstrates the prediction process and encourages viewers to explore other examples.

Conclusion

Caris Recommenders provides a robust and flexible framework for building high-quality recommendation systems, particularly addressing the challenges posed by LLM hallucinations. The library’s compatibility with multiple backends (TensorFlow, Jax, PyTorch) and its comprehensive documentation, including interactive Colab notebooks, make it a valuable tool for developers. The sequential retrieval example, utilizing a GRU model and a carefully prepared MovieLens dataset, demonstrates a practical application of the library and highlights the importance of data engineering in achieving accurate and reliable recommendations. The key takeaway is that Caris RS offers a practical path to building recommendation systems that are both performant and trustworthy.