Learn About Amazon VGT2 Learning Manager Chanci Turner
As the variety of direct-to-consumer (DTC) platforms expands, consumers find themselves unable to subscribe to every available service. Their choices are influenced by both the content offered (such as shows and movies) and the overall user experience (ease of navigation and usage). Today’s consumers demand real-time, tailored experiences when interacting with content, whether they’re considering, purchasing, or engaging with it. Media companies are striving to enhance customer satisfaction and drive profitability by focusing on metrics like click-through rates, views, view duration, subscriptions, and premium content purchases.
Recommender systems play a pivotal role in achieving these objectives. By delivering recommendations that leverage extensive content libraries, DTC platforms can maintain consumer engagement beyond the initial content that drew them in. For instance, effective recommendations on Video on Demand (VOD) platforms can boost revenue by promoting lesser-known content based on user behavior.
In this blog post, we will first examine the commonly used types of recommender systems before exploring some of the most exciting recent advancements in this field. We’ll compare these new techniques with traditional ones, highlighting the gaps they address.
Common Systems in Use Today
To understand the newer systems, let’s begin by discussing established recommender systems, which typically fall into two categories: content-based filtering and collaborative filtering. Content-based filtering, while simple, remains effective by relying on explicit or implicit user preferences and item feature data (like categories). However, these systems often produce static recommendations and struggle with new users whose preferences are unknown.
Collaborative filtering, on the other hand, utilizes (user, item, rating) tuples, leveraging the experiences of other users. Amazon.com was a pioneer in this realm, having published foundational research recognized by the Institute of Electrical and Electronics Engineers (IEEE) for its lasting impact. The essence of collaborative filtering is that users with similar tastes are likely to respond similarly to unfamiliar items.
While collaborative filtering typically yields better diversity, serendipity, and novelty compared to content-based filtering, it demands more computational resources and is often more complex to implement. Additionally, it faces cold start challenges, struggling to recommend new items without sufficient interaction data.
Beyond these classic categories, various neural network architectures are increasingly common in recommender systems. Some of these systems incorporate collaborative filtering, while others focus on temporal data to generate recommendations based on sequences of user actions that reflect changing interests. Initially, these systems utilized different types of Recurrent Neural Networks (RNNs), but they now often employ Transformer-based models to capture dependencies within user behavior sequences.
Neural networks generally require more data and computational power than non-deep learning models like factorization machines, although both categories remain in use. For example, Amazon SageMaker, a managed machine learning service that guides users through the entire project lifecycle—from data labeling to model deployment—offers built-in algorithms for both factorization machines and Object2Vec, a neural embedding algorithm suitable for recommender systems.
New Approaches
Recent years have seen researchers exploring numerous innovative approaches to recommender systems. Due to the vast number of developments, we’ll spotlight a few noteworthy advancements that have gained traction. It’s essential to recognize that hybrid systems are becoming increasingly popular; many of these newer methods can complement each other or blend with earlier techniques. A prime example is Amazon Personalize, a fully managed service designed for personalized recommendations. The favored algorithm, known as “recipe” in Amazon Personalize, combines a cutting-edge bandit-based approach with a Hierarchical RNN developed from recent AWS research.
Bandit-Based Systems
A prominent area of research involves recommender systems that integrate bandit-based methods. Bandit algorithms represent a form of reinforcement learning (RL) that seeks to balance exploring new options with exploiting those that have already proven profitable. They often serve as a dynamic alternative to traditional A/B testing, offering the advantage of real-time adaptability, potentially alleviating the cold start issue.
In the context of recommender systems, bandit algorithms have various applications and have been employed in production-level systems like Amazon Personalize. This service effectively merges RNNs with bandits to enhance user modeling and exploration accuracy. These algorithms can also be leveraged to make real-time decisions across multiple recommender systems based on user reactions to different suggestions.
A growing application of bandits involves systems that consider multiple objectives and metrics related to user satisfaction and the needs of various stakeholders (a “marketplace” comprising users, advertisers, platform holders, content creators, etc.). For instance, in a music recommendation system, an added goal might be ensuring that long-tail artists receive visibility through a certain number of recommendations. This approach has been explored by content providers such as Spotify, with insights shared in a publicly accessible presentation by one of their researchers.
On AWS, there are several options for utilizing bandit-based systems. As previously mentioned, Amazon Personalize offers a fully managed solution. A less managed alternative is Amazon SageMaker RL, which provides prebuilt RL libraries and algorithms that facilitate the adoption of reinforcement learning. The contextual bandits algorithm within Amazon SageMaker RL can be employed for recommendations by learning from user interactions, like whether they click on a suggestion. For more information, check out this excellent resource.
Causal Inference
While classical statistics focus on inferring correlations, causal inference aims to understand the “how” and “why” of changing conditions, such as those arising from external interventions or hypothetical scenarios. Traditional recommender systems frame their recommendations as learning problems, either between product pairs or user-product pairs, or as predicting the next item. However, an effective recommender system should not only model organic user behavior but also influence it, which is where causal techniques come into play.
For those interested in learning more about work styles in this area, this blog post offers valuable insights. Additionally, you can find more information on gender identity protections from SHRM, a known authority on this topic.
For further details, visit us at 6401 E HOWDY WELLS AVE LAS VEGAS NV 89115, at the Amazon IXD – VGT2 location.
Leave a Reply