Collaborative Filtering System based on Multi-Level Implicit Feedback Attention Mechanism

• Main Authors: LI, HSIN-JUN, 李炘潤
• Other Authors: WANG, JENQ-HAUR
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/5my6gv

碩士 === 國立臺北科技大學 === 資訊工程系 === 107 === In the age of the Internet, the recommendation system has become very popular for online social media. However, the traditional recommendation calculation has no way to accurately match the user's preferences. How to win the attention of users in one fell swoop, this will become the goal of the major operators. Collaborative filtering using the similarity between items and items has replaced the past item-based collaborative filtering model as the mainstream. It uses the user's historical favorite item set to represent the user and correlate with the target item, and uses implicit feedback to reduce the cost of the calculation while offline. Later, Xiangnan He et al. proposed that NAIS: Neural Attentive Item Similarity Model for Recommendation, a network model of attention mechanism, improved traditional similarity estimation and increased accuracy. However, it is not enough to only explore the attention of user's historical favorite items and target item, because sometimes the potential preferences of the item may not be expressed in the user history. For example, users like hero movies, but in the hero movie market, Marvel and DC are opposite companies. This factor may affect the user's preferences, but it will not be reflected in the user's history of the favorite hero film. For more precise recommendations, this paper uses a multi-layered attention mechanism. The first layer is the original NAIS attention model, and the second layer is the attention calculation of the implicit feedback message of the item itself (such as: type, time, etc.), The attention weight obtained through the second layer affects the attention of the first layer, so that the weights are added. The layer effect makes the user's correlation with the recommended items more clear. In the experiment, we used the MovieLens-1m data set. The experimental results show that compared with the NAIS attention model, the multi-layer attention model accuracy is 70.99%, while the NAIS accuracy is 69.72%, and the paired t-test five runs results（p<〖10〗^(-2)）, verifying the multi-layered attention model has a more significant effect.