AI 용어집
인공지능 완전 사전
Deep Learning Recommendation Systems
Recommendation systems using deep neural networks to model complex relationships between users and items. These systems outperform traditional methods by capturing non-linear interactions and latent patterns in the data.
Embedding Layers
Neural network layers that transform sparse categorical variables into low-dimensional dense vectors. Embeddings capture semantic similarities between items and users in a continuous vector space.
Neural Collaborative Filtering
Neural network architecture replacing traditional factorization models with deep neural networks to model user-item interactions. NCF learns complex interaction functions beyond simple matrix multiplication.
Wide & Deep Learning
Hybrid architecture combining a wide model (logistic regression) for memorization and a deep model (neural network) for generalization. This approach efficiently captures both explicit and implicit patterns in the data.
DeepFM (Deep Factorization Machine)
Model unifying Factorization Machines and neural networks for simultaneous low-level and high-level feature learning. DeepFM shares embeddings between FM and DNN components, optimizing efficiency and performance.
Autoencoders for Recommendations
Unsupervised neural networks learning compressed representations of user preferences for collaborative recommendation. Denoising autoencoders are particularly effective at handling sparse and noisy data.
Session-based Recommendations
Recommendation systems using recurrent neural networks (RNN) to model interaction sequences within a user session. These models capture temporal and contextual intent without requiring historical user profiles.
Graph Neural Networks for Recommendations
Approach representing recommendation systems as heterogeneous graphs with user, item, and attribute nodes. GNNs propagate information through graph structures to capture high-order relationships.
Attention Mechanism in Recommendations
Mechanism allowing recommendation models to weight historical items differently based on their relevance for the current prediction. Attention significantly improves performance in sequential and contextual recommendations.
Transformer Models for Recommendations
Architecture based on multi-head attention mechanisms to model long-distance dependencies in user behavior sequences. Transformers outperform RNNs in capturing complex and dynamic patterns.
Two-Tower Architecture
Dual model with separate towers to encode user and item features in a common embedding space. This architecture scales efficiently for millions of items thanks to pre-computed item embeddings.
Sequential Recommendation Models
Deep learning models capturing the dynamic evolution of user preferences through temporal sequences of interactions. These architectures use RNNs, Transformers, or GNNs to model sequential dependencies.
Deep Cross Network
Architecture specially designed to efficiently learn explicit cross interactions of arbitrary degree between features. DCN combines efficient cross layers with deep layers for generalization.
Variational Autoencoders for Recommendations
Probabilistic generative models learning latent distributions of user preferences for robust recommendations. VAEs naturally handle uncertainty and improve recommendation diversity.
Reinforcement Learning for Recommendations
Approach formulating recommendation as a Markov decision process optimizing long-term rewards. RL agents learn adaptive recommendation policies maximizing sustained user engagement.
Multi-task Learning for Recommendations
Learning paradigm of simultaneously training multiple objectives (CTR, CVR, session time) to improve generalization and efficiency. MTL shares representations while specializing in specific tasks.
Cold Start Problem with Deep Learning
Challenge addressed by deep learning architectures using metadata and neural networks to generate initial embeddings. Transfer learning models and GNNs on content graphs are particularly effective.
Neural Factorization Machines
Extension of Factorization Machines integrating neural networks to capture complex non-linear interactions between features. NFM combines the efficiency of FM with the expressive power of deep learning.