A problem met when running ExpoMF

[Coder-Yu]

When I was trying to train ExpoMF with small datasets (less than 500,000 records, 3,000 users and 15,000 items), I met a problem that all the mu_i tend to approximate 1 with the iteration increasing, which recovers standard matrix factorization. I conducted the experiment with your code and the only change i made was just importing my dataset. Can you give me some suggestions to deal with the problem? Thanks.

[Coder-Yu]

After reading a related paper published in CIKM 2018, I consider the idea that one of the possible reasons is that the value of a_ij is very sensitive to the selected Gaussian precision parameter λ_y and may not reflect the true exposure of users. That paper also reveals that it is better to use Bernoulli distribution to model the implicit feedback instead of Gauss distribution.

[Coder-Yu]

很久之前看的，现在已经记不得细节了... 不好意思。paper本身没有问题，你可以再推导一下

[Coder-Yu]

你可以对照一下AAAI 18的一篇，参照着看看可能能帮助理解
Collaborative Filtering with Social Exposure: A Modular Approach to Social Recommendation

[eminemichael]

多谢多谢，费心了~

…

------------------ 原始邮件 ------------------ 发件人: "Yu"<[email protected]>; 发送时间: 2019年8月20日(星期二) 晚上11:15 收件人: "dawenl/expo-mf"<[email protected]>; 抄送: "mfyuufv"<[email protected]>; "Comment"<[email protected]>; 主题: Re: [dawenl/expo-mf] A problem met when running ExpoMF (#3) 你可以对照一下AAAI 18的一篇，参照着看看可能能帮助理解 Collaborative Filtering with Social Exposure: A Modular Approach to Social Recommendation — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.