Recommender Systems Quiz

Recommender Systems

• Recommender systems are used to predict user preferences and make recommendations for items, like products or videos.
• Local shops often focus on popular items but have limited stock, while online shops have unlimited stock and a wider range of niche products.
• Online shops need recommender systems to suggest products to users.

Motivation

• Local shops focus on popular items due to limited stock, rarely having rarer books.
• Online shops stock a much wider variety of products, including many niche items.
• Recommendations are crucial for online shops to help users find products they might not otherwise discover.

Examples

• Amazon recommends products.
• Netflix recommends videos, with a $1 million prize for improving their algorithms.
• YouTube recommends similar videos.
• News applications suggest similar news stories.

Types of Recommender Systems

• Content-based filtering: Builds a user profile based on user behavior, such as search history, watch history, or time spent reading an article. The profile predicts preferences and makes recommendations.
• Collaborative filtering: Relies on the behavior of other users toward items. Users provide preferences (ratings or Boolean values) regarding products, allowing recommendations based on assumed similarities between users. Recommends items based on observed similarities of preferences amongst other users.
• Hybrid: Combines content-based and collaborative filtering approaches.

Content-Based Filtering

• Capturing user behavior to create user profiles.
• Utilizing a user's search history, viewing, or reading habits.
• Using user profiles to predict preferences and make recommendations.

Overview

• Data capture is crucial for building effective recommender systems.
• User profiles are built using the captured data.
• Comparisons of items and user profiles are made.
• Combining data sources is often needed for a more well-rounded approach.

Capturing Data

• Websites collect user data using cookies (stored locally on the user's computer) and account information (stored on the website server).
• User data includes order history, time spent reading articles or watching videos, browsing history, and location (time and spatial data.)

Build User Profile

• Use captured data to create a user profile.
• User profiles include video types watched, and description of items purchased.
• User profiles are used for product or service recommendations.
• More data leads to better recommendations and reduces uncertainty.

Item-Based Comparisons

• Users select an item.
• Vector-space models are used to compare the selected item to other items.
• The selected item is considered the query.
• Comparing items based on similar properties via cosine similarity.

User-Profile Comparisons

• User profiles are created from various data sources.
• Based on user preferences of items.
• Vector-space models are used to compare profiles.
• Similarities in user profiles are found.
• Dot-product matching is a common method.
• A user's profile may contain more information than the item itself.

Combining Data Sources

• Combining inputs from different sources (social media, sales records) is common.
• These sources are often hard to combine with simple linear methods.
• Deep learning/neural networks can assist in combining diverse data sources.

Deep Learning Approach: Benefits

• Modeling non-linear data.
• Combining varied data sources.
• Effort reduction in creating suitable features for data.
• Integrating various data formats (e.g., images, text).
• Modeling sequential behavior patterns.
• Flexibility to adapt to changes in requirements.
• Efficiently combining different recommendation approaches (e.g., content-based and collaborative).

Deep Learning Approach: Limitations

• Requires large datasets.
• Complex systems can be difficult to interpret.
• Many hyperparameters necessitate fine-tuning of the model.

Collaborative Filtering

• Records the behavior of numerous users.
• Data includes binary information (watched/not watched), item ratings, and viewing/reading time.
• A matrix records user interactions with items.
• Matrix analysis aids in extracting general patterns and results.

Users and Behavior: Boolean

• Example matrix showing user interaction with items using Boolean values (1 for interaction, 0 for no interaction).

Users and Behavior: Ratings

• Example matrix with user interaction with items using ratings from 1 to 5.

Users and Behavior: Similarities

• Example showing similarities between users based on common item choices.
• Sparsity is a common characteristic of user-item interaction matrices.

Users and Behavior: Sparse Matrix

• Matrices are often sparse, meaning users only interact with a small subset of items.
• This sparsity is offset by large numbers of users.

Collaborative Filtering Approaches

• Memory-based: Comparison between users and items based on their known interactions. Neighbour-based methods.
• Model-based: Prediction from a model (e.g., matrix factorization, SVD, neural networks).

Memory-Based Predictions

• User-based: Compare input users with other known users to identify similarities and recommend products.
• Item-based: Compare input items with known items based on similarities in preferences expressed by other users to recommend related items.

User-Based Predictions

• Implement using a vector-space model.
• User-profile data forms the query (Q).
• Find matching users.
• Simple dot-product matching.
• Others may have rated many more items.

Matrix Factorization

• Users and items are similar.
• Reduces the space needed to store the matrix.
• Predicts missing values.
• Numeric factorization methods used (e.g., alternating least squares, gradient descent).
• Introduces latent factors.

Matrix Factorization (Continued)

• Data is split into training and testing sets.
• Root Mean Square Error (RMSE) is used to evaluate the model's effectiveness.
• The derivative of the RMSE is minimized to improve the model.
• Adding a regularization term to the RMSE reduces overfitting (overly complex model).

Cold-Start Problem

• New products or users are added, but insufficient historical data exists.

Cold Start: New Products or Services

• New products or services are added without any user interaction with them.
• Collaborative filtering data might not exist initially.
• User interactions can be used to build a profile of these new items.
• Content-based filtering approaches may be better utilized.
• Increase ratings initially as an offset to get appropriate recommendations.
• Random selection of recommendations until data is collected for future accurate predictions.

Cold Start: New Users

• New users create a similar cold start problem.
• Content-based filtering might be an initial option.
• Extracting data from interactions with other websites.

Netflix Recommender System

• Netflix uses multiple algorithms working together for improved recommendations.
• The system now uses multiple screens for choice to accommodate the limited concentration time of users browsing.
• Optimized for user experience.

Motivation (Netflix)

• Users find it hard to choose amongst many options.
• Internet choice overwhelms users.
• Users' average concentration times viewing videos on Netflix is 60 to 90 seconds.

Netflix Homepage Layout

• Provides multiple recommendation types on the user interface.
• Matrix layout with many rows of options.
• Number of rows varies depending on the device used.

Algorithms (Netflix)

• Personalized Video Ranker (PVR): Combines personal preferences with general popularity.
• Top-N Video Ranker: Best personalized recommendations from catalogue.
• Trending Now: Uses short-term trends like Valentine's day or weather storms.
• Continue Watching: Estimates if a viewer will continue a video based on the time spent and stopping point.
• Video-Video Similarity: Similarity algorithm is not personalized, arranges in a personalized manner.

Page Generation (User Interface)

• Output from all algorithms is used.
• Accounts for recommendations and diversity.
• Fully personalized.
• Search history is considered in recommendations..

Test your knowledge on the mechanics of recommender systems, including collaborative filtering, hybrid systems, and data usage. This quiz covers essential concepts that help understand user preferences and behaviors in recommendation technology.