The Science Behind the Recommendation Model

Adapted from 6estates CS Seminar Series: An introduction to Neural Recommender Models

Data from machine learning and artificial intelligence can be the stepping stone to provide a personalised experience for your consumers. Recommender systems have long been used by e-commerce and e-publications to suggest appropriate products and content for people. To facilitate the sharing of technical knowledge, 6Estates holds CS (Computer Science) Seminar to promote AI & Computer Science community in Singapore.

The Innovation of Neural Recommender Models

Recommender systems, or RecSys, might come across as an unfamiliar term, but they have been in fact deeply embedded in the numerous facets of our daily lives, especially in Information Age that we currently live in. Recommender systems are adopted in environments where items are often being “recommended” to users, and they seek to predict the “rating” that users will give to a specific item after use. The concept of “Recommendation” is actually commonplace – While shopping online, we might go on a shopping spree because buying a particular shirt online will bring about recommendations of similar products right onto our webpage.

Recommendation Systems in a Nutshell

For business owners, RecSys are of immense potential as a major monetization tool for customer-oriented online services like e-commerce and social networks. Huge margins of profits can be accrued through advertisements which are supported by recommendation solutions that help predict the click-through rate of users. For example, the click-through rate predictions have accounted to more than 80% of Netflix’s movie watch traffic, contributing more than $1 billion of profit per year.

So how do RecSys work? In the modern two-stage RecSys architecture, all items will be listed in the database, which can often number into the millions. In the first stage, potential candidates, or product users will be identified through the collaborative filtering method which makes predictions through matrixes based on data sourced from the user’s history and context. For example, a user’s movie preferences can be predicted by collecting information on the preferences of other users, hence the term collaborative. These “other” users will also placed within the matrix and matched on the similarity in preferences towards other movies. This method then reduces the number of relevant items that might be of interest to the user into the hundreds, before the second stage is launched, where more complicated feature-based ranking models are applied to produce a narrower and much more relevant list of results to the user.

The value of Recommendation Systems

        The innovation of RecSys overcomes a significant challenge of the concept of recommendation, which is to prevent any overlap between user features and item features. In matching users and items together, queries are sent in word form on platforms, which might result in overlaps occurring during retrieval. In RecSys, academia and industry professionals alike have proposed the use of the “learning to recommend” paradigm, where machines can learn matching and scoring functions from historical interaction data. This allows RecSys users to estimate the matching scores of the user and the item, moving them closer to answering the question of “How likely is the user going to purchase this item?”

Learn More about Artificial Intelligence

*If you are interested to find out more about Recommendation Systems and various deep-tech systems, visit 6estates’ website at 6estates is a AI-driven data analysis company focusing on marketing and finance intelligence. Through the use of processing algorithms, 6estates helps businesses users make better strategic decisions through a more comprehensive understanding of internal and external data sources.