Skip to content
Search

Recommender system

lastmod 2026-03-24created 2024-04-2439 edits

A recommender system (推薦system) is a type of Korean: 정보 필터링 (IF) technology that recommends information a particular user is likely to be interested in (movies, music, books, news, images, web pages, etc.). Recommender systems mainly use collaborative filtering techniques. Examples include recommending links to people on social bookmarking sites and recommending movies from the MovieLens dataset.

Categories

Projects

  • starreco stands for State-of-The-Art Review Recommendation System
  • Gorse - open-source recommender system

Deep Learning based

  • DeepFM - parallel FM (second-order interactions) + DNN (high-order interactions) structure that effectively learns both sparse data and complex interactions; excellent for CTR prediction
  • Two-Tower - extracts embeddings from separate, independent neural networks for users and banners; strong for large-scale data and real-time serving; similarity-based recommendation
  • Flash-MoE - runs a 397-billion-parameter model on a laptop

Hierarchy

Recommender_system_-_hierarchy.png

Other Recommemder System

  • Context-based Recommendation
    • Context-aware Recommendaton System
    • Location-based Recommendaton System
    • Real-time or Time-Sensitive Recommendation System
  • Community-based Recommendation
    • Recommends based on the preferences of the user's friends or the communities the user belongs to
    • Uses SNS news feeds, SNS network data, etc.
  • Knowledge-based Recommendation
    • Recommendation using item Features based on specific domain knowledge
    • Case-based Recommendation : picks and recommends the best fit between the user's needs (current problems, etc.) and solutions
    • Constraint-based Recommendation : recommends to the user based on predefined rules

Content-based filtering vs Collaborative filtering

Content-based filtering

Pros

  • No dilemma about the first rater (Cold-start dilemma X)
  • Gets appropriate recommendations without help from other users (User liberty)
  • No need to rate a new item in order to recommend it
  • Not affected by Scarcity
  • Practical for handling many users' interests and changes in interest by monitoring changes and updates to user profiles.
  • Works with less data

Cons

  • Limited content analysis
  • Overspecialization (the system can only recommend items that score highly against the user profile; the user is recommended items similar to those already rated)
  • Cannot describe items that are too different or too similar.
  • Cannot interpret complex relationships
  • Computationally expensive
  • Runs into trouble because of synonyms or homonyms
  • Some types of item content are not easy to analyze
  • Users can only receive items similar to their past experience (cannot surprise the user)

Collaborative filtering

Pros

  • Does not require content information
  • Users can receive items they have never come into contact with before but are interested in (can surprise the user)
  • Because other users' scores reflect changes in the system, they are not used as deterministic results (Adaptive Quality that improves over time)

Cons

  • Requires Rating data
  • Cannot recommend with Sparse data
  • Scalability problems
  • Cold-start dilemma (new user (community) and new item problems)
  • Performance degrades with an enormous item set and a small number of users
  • Quality depends on collecting large amounts of data
  • Hard to recommend to users with unusual tastes
  • Hard to cluster and classify users whose preferences change or are included

Deep learning based

Pros of deep learning models in recommender systems

  • Non-Linear Transformation
    • Can capture complex User-Item interaction patterns
    • Existing Matrix Factorization models are linear, so they are limited in learning expressive power
  • Representation Learning
    • Effectively learns the features of Input data
    • Can reduce Hand-Craft Feature Design → lower Feature Engineering cost
    • Can effectively handle unstructured data such as images and text
  • Sequence Modeling
    • Works effectively even on sequential data (User behavior, Item changes) → can recommend the next Item likely to be purchased
  • Flexibility
    • Has a variety of convenient Frameworks such as Keras, Tensorflow, and Pytorch

Cons of deep learning models in recommender systems

  • Interpretability
    • Hard to interpret the predictions of deep learning models (Black-Box model)
    • Currently, Atttention models have opened up interpretability
  • Data Requirement
    • Training deep learning models requires a large amount of data
    • Collecting Labeled data in recommender systems is easier than collecting image and text data
  • Extensive Hyper-parameter Tuning
    • A very large number of Hyper-Parameters
    • A general limitation of machine learning
    • Research is underway on being able to use a single Hyper-Parameter

Models

Deep learning related:

  • MF - Matrix Factorization
  • GMF - Generalized Matrix Factorization
  • MLP - Multilayer Perceptrons
  • NeuMF - Neural Matrix Factorization
  • FM - Factorization Machine
  • NeuFM - Neural Factorization Machine
  • WDL - Wide & Deep Learning
  • DeepFM - Deep Factorization Machine
  • xDeepFM - Extreme Deep Factorization Machine
  • FGCNN - Feature Generation by using Convolutional Neural Network
  • ONCF - Outer-based Product Neural Collaborative Filtering
  • CNNDCF - Convolutional Neural Network based Deep Colloborative Filtering
  • ConvMF - Convolutional Matrix Factorization
  • AutoRec - AutoRec
  • DeepRec - DeepRec
  • CFN - Collaborative Filtering Network
  • CDAE - Collaborative Denoising AutoEncoder
  • CCAE - Collaborative Convolutional AutoEncoder
  • SDAECF - Stacked Denoising AutoEncoder for Collaborative Filtering
  • mDACF - marginalized Denoising AutoEncoder Collaborative Filtering
  • GMF++ - Generalized Matrix Factorization ++
  • MLP++ - Multilayer Perceptrons ++
  • NeuMF++ - Neural Matrix Factorization ++
  • Graph-based Recommendation
    • NGCF
  • Session-based Recommendation
    • NARM
    • BERT4Rec
  • Graph
    • Gprah Laplacian
    • DGCNN (Graph classification)
    • Link Prediction Based on Graph Neural Networks

Improving recommender systems and search in the era of LLMs

Recommender systems and search have historically developed by taking inspiration from language models

  • Word2vec → learning item embeddings (embedding-based retrieval)
  • GRU, Transformer, BERT → predicting the next item to recommend (ranking)

The current large language model (LLM) paradigm is also evolving in the same direction

LLM/multimodal-augmented model architectures

  • Semantic IDs (YouTube)
  • M3CSR (Kuaishou)
  • FLIP (Huawei)
  • beeFormer
  • CALRec (Google)
  • EmbSum (Meta)

LLM-based data generation and analysis

  • Recommendation Quality Improvement (Bing)
  • Expected Bad Match (Indeed)
  • Query Understanding (Yelp)
  • Query Recommendations (Spotify)
  • Playlist Search (Amazon)

Scaling Laws, transfer learning, knowledge distillation, LoRA

  • Scaling Laws
  • PrepRec
  • E-CDCTR (Meituan)
  • Bridging the Gap (YouTube)
  • Self-Auxiliary Distillation (Google)
  • DLLM2Rec
  • MLoRA (Alibaba)
  • Taming One-Epoch (Pinterest)
  • Sliding Window Training (Netflix)

Unified search and recommendation architectures

  • Bridging Search & Recommendations (Spotify)
  • 360Brew (LinkedIn)
  • UniCoRn (Netflix)
  • Unified Embeddings (Etsy)
  • Embedding Long Tail (Best Buy)
  • User Behavioral Service (YouTube)
  • Modern Ranking Platform (Zalando)

See also

Favorite site

Articles

  • 20 Recommender System Tech Posts | GeekNews
    • How does Melon recommend music? - Jeon Jae-ho (Kakao recommendation team)
    • LINE Timeline's New Challenge Part 1 – Discover for exploring recommended content and Follow, a new subscription model - Lee Kyung-bok (LINE)
    • With the AI technology ‘RIYO’ applied to Naver's trending search terms, the search term chart is now personalized for each individual based on user choices - Naver Diary blog
    • Six Descriptions (記述) of the Power of Brunch Recommendations - Choi Kyu-min (Kakao recommendation team)
    • Deep learning recommender system in production - Jeon Mu-ik (Danggeun Market)
    • Deep learning personalized recommendation - Jeon Mu-ik (Danggeun Market)
    • The ‘movies worth watching together’ recommendation service powered by AiTEMS is now open. - Naver Search blog
    • Personalized ‘worth viewing together’ products too: introducing style topic-based shopping recommendation (AiTEMS). - Naver Search blog
    • Developing and adopting the Bunjang recommender system, a retrospective Part3 - Lee Myung-hwi (Bunjang)
    • Developing and adopting the Bunjang recommender system, a retrospective Part2 - Lee Myung-hwi (Bunjang)
    • Developing and adopting the Bunjang recommender system, a retrospective Part1 - Lee Myung-hwi (Bunjang)
    • [Technology Naver researches] AI-based place recommendation system AiRSPACE - Naver Search blog
    • Naver's automatic news recommendation technology presented at CIKM2017 - Jung Hoo-joong (Naver Search), Park Keun-chan (Naver AiRS)
    • The evolution of the Kakao I recommendation engine - Sung In-jae (Kakao recommendation team)
    • The technology of AiTEMS, which builds a shopping platform that satisfies both users who rarely shop and new sellers - Lee Jung-tae (Naver Personal Shopper), Park Byung-sun (Naver Communication Group)
    • ‘AiTEMS, a deep learning-based personalized product recommendation system that finds must-have items that hit your taste - Lee Jung-tae (Naver Personal Shopper), Park Byung-sun (Naver Communication Group)
    • The AI assistant in my hand, recommendation algorithms - Jeon Sang-hyuk, Kim Kwang-seop (Kakao recommendation team)
    • The secret of Kakao's news recommendation AI algorithm 'RUBICS' - Kakao Policy & Industry Research Brunch
    • Introducing AiRS, an AI-based recommender system - Naver Diary blog
    • RUBICS – kakao's real-time recommender system - Seo Sang-won (Kakao recommendation team)
  • How is the first screen you see decided? The story of MUSINSA's personalized home banner recommendations | GeekNews

Y.LAB's Recommender System/recommender system Tutorials

author Zer0 sitemap