AI Recommendation Engine

An AI recommendation engine is a system that uses data and algorithms to suggest items to users. Building a recommendation engine typically involves collecting data on users and their interactions with items, cleaning and preprocessing the data, choosing a recommendation algorithm, training the algorithm, deploying the recommendation engine, monitoring and evaluating its performance, and making adjustments as necessary. Techniques such as A/B testing and advanced machine learning methods can also be used to improve the recommendation engine's performance. Its usage includes

• Content-based filtering: Recommends items based on their similarity to items the user has previously interacted with.

• Collaborative filtering: Recommends items based on the interactions of similar users.

• Hybrid: Combines the methods of content-based and collaborative filtering.

• Knowledge-based: Recommends items based on the user's profile and pre-defined characteristics.

We typically take the following steps to make the recommendation engine:

• Collect data on user interactions with items (such as views, purchases, ratings, etc.).

• Clean and preprocess the data to remove any errors or inconsistencies.

• Choose a recommendation algorithm that best fits the data and desires recommendations.

• Train the algorithm on the data and fine-tune it for optimal performance.

• Deploy the recommendation engine and integrate it into the application or system where recommendations will be provided.

• Continuously monitor and evaluate the recommendation engine's performance and make adjustments as necessary.

• Use A/B testing to improve the performance of recommendations.

• Use other techniques like reinforcement learning, deep learning, and matrix factorization to improve performance.

It's worth noting that the specific details of building a recommendation engine will vary depending on the type of data and the intended use case.

Understanding the data

The objective of a Recommendation Engine is to generate recommendations with a high probability of interaction with the user. While this is a requirement in many industries spanning from e-commerce to OTTs as an example we’ll stick to content recommendation. We’ll be using synthetic data to demonstrate a Recommendation Engine in a content-consuming platform. The dataset we’ll be working with has 900+ users and their corresponding ratings to a handful of films from a collection of 1500+ titles. The goal is to recommend films that are personalized to the user.

Setting up the training

Since the dataset that we’re working with is an academic dataset, there’re not many data irregularities that one might have to worry about. However, we do perform our standard data pre-processing to make sure that everything is consistent with the assumptions. For this demonstration, we’ll be using a certain variant of Graph Convolution Network(GCN). GCNs are one of the top-performing models for Recommendation Tasks at the moment. Nonetheless, when it comes to actual implementation we do consider other traditional techniques such as Matrix Factorization, and Cosine Similarity-based filtering methods, depending on the data available, to maximize the performance of the system.

Customization made easy

When it comes to model training, we train various models spanning something as simple as Logistic Regression to something as sophisticated as a Neural Network, all while varying the hyperparameters of the models, so you don't have to worry if you are using the right model for the job. We operate in a completely transparent manner making sure that there are no black boxes throughout the blueprint. We make use of open-source libraries and frameworks so that you can easily modify, replace and optimize for much more customization should you need it. Thanks to our transparent approach, you can get your hands dirty by playing around with the hyperparameters should you feel the need for experimentation. The performance of the models corresponding to different sets of hyperparameters is logged so you don’t need to make a note every time you intend to tweak the models. 

Evaluation Metrics

Mean Absolute Precision (MAP)

Measures the ability of the Recommendation Engine to rank the relevant items higher in the recommendation list. 

Normalized Discounted Cumulative Gain (nDCG)

Measures the ability of the Recommendation Engine to rank the relevant items higher than the non-relevant items. It gives higher scores to the items that are highly relevant and well-ranked.

MAP: 0.135305
NDCG: 0.453004

The Model’s Perspective

Model explainability is crucial for building trust and ensuring the responsible use of AI, as it allows stakeholders to understand and interpret the decisions made by models, and make informed choices about their deployment. Here’s a sneak peek into one such technique. Language Models, to understand the relationship between entities, use something called an Embedding space. These are the current embedding spaces of our model. The embeddings under-the-hood, are computed in 64 dimensions because it isn’t feasible to visualize 64 dimensions, we’ve made use of UMAP (a dimensionality reduction library) to help visualize these embeddings.

The Last Mile

We now have the model ready, and the next big step is to take it live. Model Monitoring and Responsibility are things that are often overlooked. Models in production come with their own set of challenges some noteworthy ones are model drift and data drift. It is also important to make sure that the models are responsible, that is, the prediction of the models shouldn’t biased to sensitive elements such as race, and gender. Our blueprint is further integrated with Censius- a platform for end-to-end AI Observability.