以類神經網路解決情境式推薦問題

Abstract

為了向用戶提供合適的商品推薦，推薦系統已在市場中廣泛應用。儘管市場上衝刺著各種可以使用的數據分析，但是冷啟動（Cold Start）問題對於新進用戶來說仍然是一個大問題。許多最新的推薦算法在假設用戶和商品保持線性關係的前提下設計了演算法，而實際上大多數情況下兩者間存在非線性關係。這項研究開發了一種使用神經網絡（NN）和情境式是推薦的演算法來處理非線性特徵和探索利用的權衡。推薦系統可以有效地預測新進用戶的喜好，還可以快速探索快速變化的喜好。通過將貝葉斯網絡（Bayesian networks）和自動編碼器（AE）集成到NN中，我們的系統, NN Contextual Bandit（NNCB）可以利用不同程度的探索和開發。因此，我們的系統能快速適應情境的變化。我們採用真實世界中的影片評分數據集來證明所提出系統的有效性，與傳統的情境是推薦演算法相比，該系統大約4％的優於就演算法。

Recommendations have been wildly applied in marketplaces to provide right items to users. While various heterogeneous data available in marketplaces, the cold start problem is still a big issue for newcomers. Many state-of-the-art recommendation algorithms were designed on the assumption that users and items remain a linear relationship, while most cases exist nonlinear relationship in reality. This study develops an algorithm using neural network (NN) and contextual bandit to deal with nonlinear context and explore-exploit tradeoff. The recommendation system could effectively predict newcomers’ preferences and also provide quick exploration for fast- changing preferences. By integrating Bayesian networks and AutoEncoder (AE) in the NN, our system, NN Contextual Bandit (NNCB), could leverage different levels of exploration and exploitation. Thus, the proposed recommendation can quickly adapt to the real-time context. We adopt real-world video rating dataset to demonstrate the effectiveness of the proposed system which improve 4% regret as the conventional bandit algorithms.