Quantization framework apparatus for efficient matrix decomposition in recommender system and learning method thereof
According to the present invention, a method for training a quantization framework for efficient matrix decomposition in a recommendation system comprises the steps of: mapping users and items to a k-dimensional latent feature space; reading, from a memory, corresponding user and item latent feature...
Gespeichert in:
Hauptverfasser: | , , , , |
---|---|
Format: | Patent |
Sprache: | eng ; kor |
Schlagworte: | |
Online-Zugang: | Volltext bestellen |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | According to the present invention, a method for training a quantization framework for efficient matrix decomposition in a recommendation system comprises the steps of: mapping users and items to a k-dimensional latent feature space; reading, from a memory, corresponding user and item latent feature vectors when random grade/user and item information belonging to training data is received; calculating a prediction grade by using the inner product of the two vectors, and calculating a prediction error based on the same; and updating user and item latent features to reduce the prediction error. Therefore, the present invention can provide a quantization framework for memory-, user-, and item-group-based matrix decomposition techniques, based on the unique characteristics of a matrix factorization model.
본 발명의 추천시스템에서의 효율적인 행렬 분해를 위한 양자화 프레임워크 학습 방법은 유저 및 아이템을 k 차원의 잠재 특징(latent feature) 공간에 매핑(mapping) 시키는 단계. 학습 데이터에 속하는 임의의 평점/유저 및 아이템 정보가 들어오면 대응되는 유저 및 아이템 잠재 특징(latent feature) 벡터를 각각 메모리에서 읽는 단계, 두 벡터의 내적으로 예측 평점을 구하고, 이를 기반으로 예측 오차를 구하는 단계 및 예측 오차를 줄이는 방향으로 유저 및 아이템 잠재 특징(latent feature)를 업데이트하는 단계를 수행하도록 구성된다. |
---|