Introduction to Recommender System

Summary

The lecture introduces recommender systems as a form of online decision support. A recommender system suggests information, products, or services that may interest a user—for example, books, movies, news, research articles, or potential friends.

Why recommender systems are useful

• Reduce information overload: Instead of requiring users to sift through a large catalogue, the system surfaces relevant or related options.
• Offer suggestions beyond search results: A search engine returns items matching a query, while a recommender can also suggest related or complementary items chosen by users with similar interests.
• Personalize the experience: Suggestions can reflect a person’s past or current interests.
• Support online customer relationships: In e-commerce, recommendations can guide virtual customers when no salesperson is present.
• Potentially increase sales and loyalty: Useful or unexpected suggestions may encourage purchases and repeat visits.

Core data used

The lecture describes three broad kinds of data:

1. User preferences for items

   • Preferences can be organized as a matrix: rows represent users, columns represent items, and each cell records a user’s preference for an item.
   • Preferences may be collected explicitly, such as by asking users to rate an item on a discrete scale (for example, 1–5) or a continuous scale.
   • They may also be inferred implicitly from behavior, such as whether a user viewed or purchased an item, browsing patterns, or time spent on a page.
   • Depending on the data, preferences may be binary or numeric.

2. User characteristics

   • Some systems use demographic or other user information, such as age or profession.
   • This information may be supplied by users or inferred from available data.

3. Item features

   • Each item can be represented by a set of attributes.
   • For a book, these might include author, publisher, subject, page count, category, or keywords. For a movie, they might include genre, director, or actors.
   • Features can be explicitly listed or derived from the item’s content.

Together, these data can help a system estimate whether an item suits a particular user, identify users likely to want a new item, or recommend items to a new user. The lecture also notes that large services must handle enormous numbers of users and items while producing recommendations quickly.

Common recommendation tasks

• Predict a preference: Estimate how an active user might rate an item they have not encountered.
• Make a top-N recommendation: Suggest the N items an active user is most likely to prefer.
• Identify top-M users: Find users most likely to value a newly introduced item, potentially for targeted marketing.

Types of recommender systems introduced

• Popularity-based

  • Recommends broadly popular items using measures such as purchase frequency or average rating.
  • Can produce top-N lists.
  • Is generally not personalized; different users may receive the same recommendations.

• Content-based

  • Uses item features and a user’s preferences to find items with relevant content.
  • Can provide predictions, top-N item lists, or identify users likely to value an item.
  • Is personalized.

• Collaborative filtering

  • Uses patterns in user preferences to find similar users or similar items.
  • May be based on user-to-user or item-to-item similarity.
  • Can produce predictions and top-N recommendations and is personalized.

• Association-based

  • Recommends items that users have often viewed or purchased together.
  • Relies on item co-occurrence patterns.
  • Can support predictions and top-N recommendations.

• Demographic-based

  • Uses similarities in users’ demographic characteristics to inform recommendations.
  • Can produce predictions, top-N item lists, and lists of likely users for an item.
  • Is personalized.

• Reputation-based

  • Uses the opinions of users whom a community considers trustworthy or influential.
  • A reputation measure may draw on factors such as purchase history, site activity, loyalty, or how other users regard a reviewer.
  • Can produce predictions or top-N recommendations and is personalized.

Early examples mentioned

The lecture cites early research and commercial initiatives, including work at MIT Media Lab on recommending music, GroupLens for news, CiteSeer for research articles, Amazon’s recommendations, DoubleClick’s personalized advertising, and CDnow’s music recommendations. These are presented as early examples of a practice that later became common across websites.

Speakers and sources featured

• Speaker: One lecturer; no name is given in the subtitles.
• Organizations, services, and examples mentioned: MIT Media Lab, GroupLens, CiteSeer, Amazon, DoubleClick, CDnow, Flipkart, YouTube, and commercial websites generally.