TSR: 針對SVD推薦系統之時效性推薦法設計與實現

TSR: A Time-Sensitive Recommendation Method for an SVD-based Recommender Systems

陳炯勛(Jiong-Xun Chen)

指導教授 ： 黃育綸

國立交通大學/電機學院/電機工程系所/碩士(2020年)

摘要

在資訊爆炸的時代，使用者需要推薦工具來幫助他們整理、篩選出他們比較會感興趣的資訊或產品，以便他們在有限的時間內去吸收新知或取得所需的商品。傳統的推薦系統會藉由分析個人資料，把相似的使用者組成一群。當被歸為同一群的使用者中有人瀏覽、購買一個新東西時，推薦系統就可以趁機把這個東西推薦給群內的其他人。然而，遇到使用者所提供的個人資訊過少的情況時，推薦系統將難以完成使用者分群，進而推薦資訊或商品。為了解決這類問題，我們設計了一個新的推薦法叫做TSR。TSR設計了一種新的與時間有關的特徵指標，分別是個人的時間特徵以及季節性的特徵。個人的時間特徵可用來描述使用者隨時間變化的喜好，例如在不同時間點所感興趣的物品類型等。季節性的特徵則用來描述季節事件對於使用者的影響程度，例如聖誕節前使用者可能會想採購的聖誕禮物等。TSR會透過分析每個使用者的行為紀錄自動給予每個使用者自己的個人時間特徵，透過每個人的個人時間特徵來串連興趣、特徵相似的使用者。TSR並改良原SVD模型，透過分析使用者特徵、參考其他擁有比較多歷史行為記錄的使用者等方式，藉以改良推薦的精準度。在這個研究中，我們設計了六個實驗，用以驗證 TSR 的可行性與實用性。實驗結果顯示，相較於原 SVD 模型，TSR 至少可以提升 1.5\% 的推薦精準度。

關鍵字

推薦系統 ； 時效性 ； 推薦 ； 推薦法

並列摘要

In the age of information exploration, people dive in the ocean of information. A recommender system becomes necessary for helping people filter out the items they are looking for. Usually, a recommender system tries to group up users who have similar features. Based on such grouping mechanisms, the recommender system can promote an item to users of the same group if one of the users in the group accesses a new item. However, in such cases, imprecise recommendations may come with the lack of user information and historical behaviors. To deal with the problem of lacking user features, we propose a recommendation method adopting time-sensitive features, called Time-Sensitive Recommendation (abbreviate to TSR), to improve the precision of recommendations. TSR defines a novel type of user feature (called time-sensitive feature), which includes subjective features and seasonal features. The subjective feature describes the user preferences changing over time. The seasonal feature represents the features affected by seasonal events, such as holiday celebrations, etc. TSR runs with a revision of SVD model. Bias from users and items is considered in the revision (we call the revision the bias SVD model) for generating predicting scores when making recommendations. When launching a recommender system running TSR, TSR can automatically create subjective features for each user based on the historical behaviors obtained from databases and calculate the similarity between users. With the user similarity measured by TSR, the recommender system can recommend an item to a user. When encountering the lack of user information, TSR uses the historical behaviors from other users in the system to build up a well-trained SVD model and improve the precision of recommendations. We conduct six experiments to prove the feasibility and practicalness of TSR. The experiment results show that TSR can improve recommendation precision by 1.5\%, compared to the recommender systems running with the original SVD model.

並列關鍵字

SVD ； Recommender ； Recommendation

參考文獻

[1] J. Lu, D. Wu, M. Mao, W. Wang, and G. Zhang, “Recommender system application developments: a survey,” Decision Support Systems, vol. 74, pp. 12–32, 2015.

[2] Y. Koren, “Factorization meets the neighborhood: a multifaceted collaborative filtering

model,” in Proceedings of the 14th ACM SIGKDD international conference on Knowledge

discovery and data mining, 2008, pp. 426–434.

[3] J. B. Schafer, D. Frankowski, J. Herlocker, and S. Sen, “Collaborative filtering recommender systems,” in The adaptive web. Springer, 2007, pp. 291–324.