Recommender Systems: Lecture 1

Lecture 1: Recommender Systems (RS)

• Presented by Georgios Peikos from the University of Milano-Bicocca, Milan, Italy, at the Department of Informatics, Systems, and Communication (DISCo).
• The lecture covers Information Retrieval and Recommender Systems.

Recommender System (RS)

• Information filtering systems that aim to push potentially useful information, goods, or services to users.
• They provide fresh recommendations based on a compilation of preferences
• Became more popular with the growth of websites, such as YouTube, Amazon, and Netflix
• Can generate revenue when used effectively and rely on machine learning methods

Course Schedule

• Lecture 1: Introduction to Recommender Systems (3 hours)
• Lecture 2: Non-Personalized RS, Bias, Fairness (2 hours)
• Lecture 3: Evaluation of RS (3 hours)
• Lecture 4: Content-Based Approach (3 hours)
• Also includes 4 Labs on RS
• Lecture 5: Collaborative Approach (3 hours)
• Lecture 6: Cold Start Problem and Recap (2 hours)

Information Seeking vs. RS

• Seeking information involves actively looking for specific data or items based on a query
• Recommending stimulates user actions, such as buying a book and helps with information overload
• Searching relies on a user-provided query matched with available documents or items
• Recommending does not necessarily need detailed item information
• Search engines focus on discriminating between relevant and irrelevant documents, ranking documents based on content
• Content-based recommenders and classical Information Retrieval (IR) methods are related; personalization blurs the boundary
• Searching provides a list of results matching the query, ranked by relevance
• Recommending offers suggestions that might interest the user, directing them to new or different categories

Purpose of RS

• Help match users with items, such as job portals
• Ease information overload, like in academic research
• Assist in sales through guidance and persuasion
• Personalize entertainment, such as with streaming services
• Aid in education and learning
• Boost social media engagement of users

Long tail

• They should recommend widely unknown or lesser-rated items that users might enjoy, promoting diversity and serendipity
• In the MovieLens 100K dataset, only 20% of items accumulate 74% of all positive ratings (rated above 3)
• A well-functioning RS could discover and recommend the other 80% of overlooked items that are still relevant

How RS Work

• Users interact with a user interface
• Recommendation system accesses a database
• Recommendations are processed via recommendation engine
• Users Receive Top Recommendations

History of RS

• 1990s: First systems (e.g., GroupLens) with basic algorithms
• 1995-2000: Rapid commercialization and challenges of scale
• 2000-2005: Research explosion and mainstream applications
• 2006: Netflix prize
• 2007: The first Recommender Systems conference
• 2010s: Applications become common with active research and many applications
• 2020s: Continued innovation, ethical considerations, and personalization at scale

RS Paradigms

• The goal of the algorithm is to reduce information overload

RS Algorithm Categories:

• Non-personalized. They offer general recommendations (most popular or manually curated)
• Collaborative. They use "Tell me what's popular among my peers"
• Content-based. They show users more of what they've liked
• Knowledge-based. They tell users "Tell me what fits based on my needs"
• Hybrid. They offer a blended collection of input/outputs

Knowledge Model Components for RS:

• User Knowledge: Preferences include destination type and Demographics, such as family and age. Also travel history.
• Item Knowledge: Destination attributes include climate, attractions, safety, cost, and package details.
• Domain Knowledge: Seasonal aspects best times to visit, Cultural, customs language, fees, and Travel Regulations like visa requirements

Benefits of Employing RS in Business Include:

• Boost sales and raise conversion rates.
• Assist in decision-making
• Analyze and provide personalized recommendations to customers
• Recommend as per user's interest
• Access data from prior sessions and enhance buyer experience

RS Applications by Sector:

• E-commerce and Retail: Leaders are Amazon, Alibaba, eBay. Provides similar products and relevant keywords.
• Entertainment and Media: Netflix and Spotify, Help users to create playlists, recommend movies.
• Social Platforms: LinkedIn, Instagram, Facebook, Offer personalized suggestions, direct to similar pages.
• Content Searching Platforms: Google, YouTube offer automated browsing recommendations.
• Banking and Finance: American Express, J.P. Morgan include the features give personalized advice, aids saving.
• Education: Coursera, Udemy, Khan Academy, Suggest studies, studies, papers.

RS Impact:

• Amazon's RS generates 35% of its purchases.
• YouTube's RS accounts for 70% of peoples' viewing time
• Netflix's RS is responsible for 75% of viewing, saving $1 Billion each year
• Alibaba boosted their conversions 20% using personalized pages

The Recommender Problem:

• The set of all Users is defined as "C"
• The set of the recommendations is defined as "S"
• The utility function is "u:CXS ->R" and defines usefulness of the recommendations

Data Inputs for Recommender Systems:

• Items data
• User data
• Interactions between Users & Items.
• Context such as user information

RS Taxonomy

• The algorithms used by the system can be Personalized or Non-Personalized
• Algorithms then used Memory and Model based
• Within these are further breakdowns like User Based, Item Based Matrix Factorization

Key Matrices in RS

• Includes the User Rating Matrix (URM) and the Item-Content Matrix (ICM).
• The matrices analyze user preferences and item traits

Recommender Matrix Definitions

• Foundations that use collaborative and content-based techniques to capture and store information about users, items, and the relationships between them.
• Used to organize and analyze large quantities of data and are important for many recommendation algorithms.
• Types of matrices includes: User-Feature Matrix, Context, Item-Similarity, and User-Similarity Matrix.

Item-Content Matrix (ICM)

• Captures the features or attributes of items.
• Rows are items while columns are specific attributes or features, such as genre or recipe ingredients.
• Values indicate the presence or level of presence of a particular item feature

User Rating Matrix (URM)

• It represents the interactions between users and items, such as ratings, clicks, and purchases.
• Rows for users and columns are for items can give rating value

Density vs. Sparsity

• Density (D) calculates the proportion of the user-item matrix with known values, where D = R/(U*I).
• Sparsity (S) calculates the proportion of the matrix that's unknown. Is the opposite of density
• S = 1 - D.

Feedback:

• Enables understanding of users. Leads to enhanced user experience
• Can be explicit or implicit

Explicit feedback

• Most commonly scales from 1-5 or 1-7 (like and dislike are options)
• Clear on users and delivers correct feedback, but can have ratings too small = poor data