The Web’s growing role in electronic and business transactions has pushed the advancement of recommender systems. A key reason for this progress is how easily the Web lets users share their preferences. Take Netflix, for example. Users can quickly click to rate content they like or dislike. Typically, users rate items using a numerical system, like a five-star rating, to express their preferences.
Other types of feedback, while less direct, are even simpler to gather in the online world. For instance, when a user buys or browses an item, it can be seen as an endorsement for that item. This kind of feedback is common on sites like Amazon.com, and it’s effortless for customers to provide. The main idea behind recommender systems is to use these various data sources to understand what customers are interested in.
In these systems, the user who receives the recommendation is called the “user,” and the item being recommended is called an “item.” Therefore, recommendation analysis often looks at past interactions between users and items because previous interests are often good predictors of future choices.
There is an exception to the rule called knowledge-based recommender systems, which suggest recommendations based on user-specified requirements rather than past user behavior.
The fundamental principle behind recommendation algorithms is the significant relationship between user preferences and the items they interact with. For example, a user interested in historical documentaries is likely to be interested in other historical documentaries or educational programs, rather than action movies. These correlations between different categories of items or individual items can be learned from the ratings matrix and used to make predictions for users.
The more items a user has rated, the easier it is to predict their future behavior. Various learning models, such as collaborative filtering, can be used to achieve this. Collaborative filtering involves using ratings from multiple users to predict missing ratings. Neighborhood models are a simple type of collaborative filtering where cohorts of similar users are created based on their collective buying or rating behavior, and recommendations are made to individual members of these cohorts.
Recommender systems can also be more complex, incorporating different types of data. Content-based recommender systems, for example, use attributes of items and user ratings to make predictions. Knowledge-based systems, on the other hand, rely on user-specified interests combined with domain knowledge to provide recommendations. Advanced models may also consider contextual data such as temporal information, location, social factors, or network information.
This post will cover all types of recommender systems, including collaborative, content-based, and knowledge-based systems, discussing both basic and enhanced models in various domains. It will also explore the robustness of recommender systems, evaluation methods, and hybridization models. This chapter aims to provide an overview of the diverse work in recommender systems and how it relates to the rest of the post.
Before we delve into the goals of recommender systems, let’s first look at how the recommendation problem can be approached. There are two primary models:
- Prediction Version: This involves predicting the rating a user would give to an item. With training data on user-item preferences, represented as an incomplete m×n matrix (m users, n items), the missing values are predicted using this model. This is akin to solving a matrix completion problem.
- Ranking Version: Rather than predicting specific ratings, this approach focuses on recommending the top-k items for a user. This is also known as the top-k recommendation problem, where the goal is to rank items for a user, increasing the likelihood of sales.
In the ranking version, the exact ratings are less important than the relative rankings of items. While the first formulation is more general, as solutions for the ranking version can be derived from it, designing methods for the ranking problem directly is often more straightforward.
The primary goal of a recommender system is to increase product sales. Merchants use these systems to recommend items that are likely to be of interest to users, thus boosting sales and profits. However, achieving this goal involves more than just recommending relevant items; other factors such as novelty also play a crucial role:
- Relevance: Recommender systems aim to recommend items that are relevant to the user. This increases the likelihood of user engagement and consumption.
- Novelty: Recommending items that are new or unseen by the user can enhance their experience. For instance, recommending popular items of a preferred genre may not be novel to the user and could lead to a decrease in sales diversity.
While relevance is the primary operational goal, achieving a balance between relevance and novelty is essential for an effective recommender system.
- Serendipity: Serendipity adds an element of surprise to recommendations, suggesting items that are unexpected but enjoyable. It goes beyond novelty, providing users with recommendations that they may not have known about or considered before. This can lead to increased sales diversity and spark new interests in users, benefiting the merchant in the long term. While serendipitous recommendations may sometimes be irrelevant, the strategic benefits often outweigh these short-term drawbacks.
- Increasing Recommendation Diversity: Recommender systems often present a list of top-k items to users. When these items are too similar, there’s a risk that the user might not find any of them appealing. On the other hand, including diverse items in the list increases the chances of the user finding something they like. Diversity helps prevent user boredom from repeated recommendations of similar items.
In addition to these concrete goals, the recommendation process also fulfills soft goals that benefit both users and merchants. For users, recommendations can enhance overall satisfaction with a website, leading to increased loyalty and return visits. Merchants can gain insights into user needs and further customize the user experience. Providing explanations for recommendations can also be valuable, helping users understand why certain items are recommended to them.
Recommender systems vary widely in the types of products they recommend. Some systems, like Facebook, recommend social connections rather than products directly. This indirect recommendation can enhance site usability and advertising profits. Understanding these goals helps to appreciate the diversity of recommender systems, from research prototypes to commercial systems, in various problem settings.
GroupLens Recommender System:
GroupLens was a groundbreaking recommender system initially developed for recommending Usenet news articles. It collected ratings from Usenet readers to predict whether other readers would like an article before reading it. This system played a significant role in developing early collaborative filtering algorithms, which were later extended to recommend books and movies through systems known as BookLens and MovieLens, respectively. Notably, GroupLens contributed to the field by releasing datasets, including MovieLens datasets with increasing sizes of ratings.
Amazon.com Recommender System:
Amazon.com was an early adopter of recommender systems in the commercial sector. Initially focused on selling books, Amazon expanded its business to sell a wide range of products. Recommendations on Amazon are based on explicit ratings, buying behavior, and browsing behavior. Ratings are given on a 5-point scale, and recommendations are provided to users when they log into their accounts. Amazon also offers explanations for recommendations, such as showing how a recommended item relates to previously purchased items.
Netflix Movie Recommender System:
Netflix started as a DVD rental service and later transitioned to streaming movies and TV shows on a subscription basis. Users can rate movies and TV shows on a 5-point scale, and Netflix uses these ratings, along with user actions like watching history, to make recommendations. Netflix excels in providing explanations for recommendations, showing examples based on the user’s viewing history. This approach helps users make informed decisions and enhances their experience.
Netflix also made significant contributions to the research community through the Netflix Prize contest. This contest challenged participants to predict ratings of user-item combinations using a large dataset of ratings. The contest popularized many recommendation algorithms and highlighted the importance of proper evaluation design in recommendation systems.
Google News Personalization System:
Google News recommends news articles to users based on their click history. The system uses identification mechanisms linked to Gmail accounts to associate clicks with specific users. Clicking on a news article is considered a positive rating for that article. These ratings are unary and implicit, as there is no mechanism for users to express dislike explicitly. Collaborative filtering algorithms are then used to make personalized article recommendations for users.
Facebook Friend Recommendations:
Social networking sites, like Facebook, recommend potential friends to users to increase social connections on the platform. Unlike product recommendations that directly increase merchant profit, friend recommendations improve the user experience on the social network, leading to network growth and increased advertising revenue. These recommendations are based on structural relationships within the social network, rather than ratings data, and are crucial for the growth and connectivity of the network.
The Spectrum of Recommendation Applications:
Recommender systems have diverse applications, including traditional e-commerce for products like books, movies, travel, and other goods and services. However, they have expanded beyond product recommendations to include areas like computational advertising, where products are advertised alongside search results. Facebook also recommends friends, and online recruitment sites recommend employers and job-seekers to each other, illustrating the broad spectrum of recommendation applications.