Machine learning recommender systems in digital media companies

1. Introduction: towards a more personalized experience

Much of the success of digital media companies is based on their recommender systems. In recent times they have become an essential tool for personalizing the user experience on platforms, e-commerce and social networks.

These systems analyze behavior patterns and preferences to suggest products, movies, articles and more. As users, we already recognize the messages from streaming platforms. “If you liked this, we recommend…”.

However, we are just at the beginning of the journey. The increasing complexity of user preferences and expanding content catalogs, demand more sophisticated approaches.

This is where advanced machine learning (ML) technologies and large language models (LLMs) come into play. This guide explores how it is possible to use these technologies to develop recommender systems with the highest levels of accuracy and personalization.

2. Basics of recommender systems

In general, we identify two types of recommender systems.

 Collaborative filtering. This is the classical approach. It is based on the premise that if two users have had similar interests in the past, they are likely to repeat such matches in the future. These methods use the matrix of user-element interactions to record and learn from past interactions. However, on their own, they can be limited, particularly when encountering new users or elements (this is known as the “cold start problem”).

Content-based filtering. Unlike the previous one, it uses additional information about users and items. For example, in a movie recommendation system, this could include directors or actors, among other variables. These methods can provide more personalized recommendations. This is because they consider the specific characteristics of features that have appealed to individual users in the past.

3. New technologies: LLM and embeddings

 In a world where technologies are evolving rapidly, innovations are emerging that are set to change the rules of the game. Recommender systems are reaching new levels.

 – LLM: natural language understanding. These models, including GPT, Bert and Titan, are revolutionary. In particular because of their ability to understand and generate natural language.

 Based on the transformer architecture, they can process sequences of words, capturing complex contexts and relationships. In recommender systems, they play an essential role in better understanding descriptions, reviews and metadata. Thus, they provide a deeper understanding of both content and user preferences.

 – Embeddings: capturing semantic meaning. Embeddings are vector representations of text that capture semantic meanings and contextual associations. In a recommender system, converting item descriptions and user preferences into embeddings allows similarities and differences to be calculated efficiently. This not only improves the accuracy of recommendations but also helps to overcome the “cold start problem”. This is because it allows comparisons with new items or users. Let’s look at how this applies specifically to the digital media industry. These companies have valuable information about their content in both synopsis and metadata. For example, the actors that participate in each episode of each series or in each movie. Embeddings allow to capture all this information to produce a more assertive system.

4. Hands-on implementation

What are the next steps?

– Integration of LLM into recommender systems. To do so, it is first necessary to fit a pre-trained model with the organization’s own data. In addition to being pre-trained with a vast amount of content, many of these models allow us to perform fine tuning. This involves adapting that training to fit the organizational data. Therefore, it is possible to train these models with the historical data of the users so that they can learn about their tastes and, based on that, predict what will be the next content to be chosen by each of them. This is how the recommendation system is built.

– Construction and use of embeddings. To build a content-based approach, we have the help of embeddings. These are the ones that allow us to transform the texts related to the content (descriptions, reviews, among others) into numerical vectors. To do so, they use a specific model, such as Ada or Titan. Then, we are able to map all our available items in the same space, but always maintaining the semantic difference between them. Subsequently, it is possible to use these embeddings to feed machine learning algorithms that predict user preferences. For example, calculating the cosine similarity between the vector representing the user to each vector representing each different content, so as to recommend those closest in the semantic space and as a result those that have a higher similarity with the user’s profile.

5. How can a recommender help in an app?

Digital media companies that incorporate a recommender in their app gain access to the following benefits:

– Prioritize the user experience. Indeed, the user is at the center of the strategy, since the suggested content is accurate according to their tastes, needs and behaviors. This increases loyalty and satisfaction levels.

– Better strategic consumption of content. Combined with the specific needs of the business, the recommender can drive the consumption of certain strategic content with a high level of precision.

– More performing systems. At Nubiral we are working on developing a recommendation system that focuses on these new technologies. What we have been able to prove is that in this way, more performing systems are obtained. We take advantage of the ability of this new technology to capture the semantic meaning and include information about the content. This often comes in natural language, such as user reviews. This allowed us to build a recommender that considers all available information, both from users and from the available content to be recommended. After testing in scenarios with real users, we were able to improve by up to 20% the rate of clicks made by users on the recommendations. In other words, we obtained a 20% more assertive recommendation system.

6. Conclusions: looking into the future

Recommendation systems are an integral part of the user experience, and their importance will continue to grow as digital platforms do.

New technologies, as we have already seen, offer significant promises for improving the accuracy and personalization of these recommendations.

However, it is vital to address this concept with a balanced approach, recognizing both its potential and its limitations.

The field of ML and recommender systems is constantly evolving. Companies in the digital media industry need to keep up to date in this regard if they want to continue to succeed in their business.

Our experts can help you get the most value from these technological advances. We look forward to hearing from you: Schedule your meeting!