近年來推薦系統十分活躍並應用在很多領域，像是電影、音樂和 電商產品等等...... 我們選擇食物作為我們推薦系統的物品。食物推薦系統最大的挑戰為食物種類的稀疏度過大，稀疏度過大導致食物推薦系統預測太具有隨機性，因此我們提出群聚優化推薦系統來解決此問題。群聚優化推薦系統的輸入為使用者的飲食紀錄，輸出為明天使用者最有機率吃的食物。在這篇論文裡面，我們提出迭代式群聚演算法不僅可以解決食物稀疏度的問題，也可以讓群聚的效果更有意義性。迭代式群聚演算法包含著模型預測、群聚演算法。在模型預測方面，我們使用 LSTNET[4] 來訓練和預測使用者的飲食群聚; 在群聚演算法方面，我們重新群聚食物基於模型預測的群聚和正確答案的群聚。上述的二個流程會迭代的執行直到結果飽和。我們的方法評估於使用者 飲食紀錄，這份資料來自於 MyFitnessPal[1] 應用程式，並由新加坡管理大學所提供資料。我們的演算法在使用者飲食測試紀錄資料上達到 0.323 的 map[11] 分數，也贏過其他群聚演算法。

Recommendation system begin very popular in recent years and are adopted in many fields, such as movies, musics, and E-commerce items. We choose food items for our recommendation items. The biggest challenge of food recommendation is the sparsity of any food-item set under considerations. Therefore, we propose Clustering items for better food recommendation as ourthesis. Clustering items for better food recommendation takes users’ dietary records as the input and then predicts food which users most likely to eat next day. In this thesis, we propose an iterative cluster method (ITC) which cannot only solve the sparsity problem but also derive more meaningful clustering for food recommendation. The proposed ITC includes model prediction and reclustering method. For model prediction, we adopt LSTNET[4] to predict clusters based on users’ past dietary records. There clustering method reclusters food items based on the prediction of clusters and the groundtruth. The above two stages are repeated until the results are saturated. Our proposed method was evaluated over users’dietary records dataset, which is from MyFitnessPal[1]append are provided by Singapore Management University. It was shown that our method achieved the map[11] score of 0.323.

誌謝 i
摘要 ii
Abstract iii
1 Introduction 1
2 Related Works 2
3 ITC Model 3
3.0.1 Problem Formulation . . . 3
3.0.2 Model Overview . . . 4
3.0.3 Initial kmeans cluster . . . 5
3.0.4 Model Prediction . . . 5
3.0.5 Reclustering Algorithm . . . 7
4 Experiments 11
4.1 Datasets . . . 11
4.2 Evaluation Function . . . 12
4.3 Methods and Baseline . . . 13
4.4 Training Details . . . 14
4.5 Results . . . 14
5 Conclusions 18
6 Future Work 19
Bibliography 20

[1] L.AlbertandL.Mike. Myfitnesspal. 2005.
[2] N. Bhawna, A. Poorvi, S. Sonal, and V. Swati. Document classification using expectationmaximizationwithsemisupervisedlearning. arXiv:1112.2028,2011.
[3] D. Christopher, R. Prabhakar, and S. Hinrich. Hierarchical clustering. pages 377– 401,2009.
[4] L. Guokun, C. WeiCheng, Y. Yiming, and L. Hanxiao. Modeling long- and shorttermtemporalpatternswithdeepneuralnetworks. arXiv:1703.07015,2017.
[5] X.Hong-Jian,D.Xinyu,Z.Jianbing,H.Shujian,andC.Jiajun. Deepmatrixfactorizationmodelsforrecommendersystems. 2017.
[6] W. Hongwei, Z. Fuzheng, W. Jialin, Z. Miao, L. Wenjie, X. Xing, and G. Minyi. Ripplenet: Propagating user preferences on the knowledge graph for recommender systems. arXiv:1803.03467,2018.
[7] J.B.MacQueen. Somemethodsforclassificationandanalysisofmultivariateobservations. pages281–297,1967.
[8] T.JiaxiandW.Ke. Personalizedtop-nsequentialrecommendationviaconvolutional sequenceembedding. arXiv:1809.07426,2018.
[9] C. Junyoung, G. Caglar, C. KyungHyun, and B. Yoshua. Empirical evaluation of gatedrecurrentneuralnetworksonsequencemodeling. arXiv:1412.3555,2014.
[10] N. Kamal, M. Andrew, and M. M. Tom. Semi-supervised text classification using em. page33–56,2006.
[11] K.Ralf,F.Peter,andN.Wolfgang. Latentdirichletallocationfortagrecommendation. pages61–68,2009.
[12] O. Sergio, O. Vito Claudio, D. N. Tommaso, S. Xavier, and D. S. Eugenio. Sound andmusicrecommendationwithknowledgegraphs. pageArticleNo.21,2017.
[13] Z.Shuai,T.Yi,Y.Lina,andS.Aixin.Nextitemrecommendationwithself-attention. arXiv:1808.06414,2018.
[14] M.T.K. Theexpectation-maximizationalgorithm. pages47–60,1996.
[15] Z.XiaojinandB.Andrew.Introductiontosemi-supervisedlearning.page130,2009.