Recommender Systems

Are you struggling to choose the best marketing strategy or measure the effectiveness and adequacy of your marketing campaign? You are not alone I’m too. I’m no expert in marketing strategies so to set this straight before you go ahead and read the entire article, but I’m an expert in…

Paper Link Observation Existing methods (E.g. NRMS, NPA, LSTUR) usually first encode user’s clicked news into news embeddings independently and then aggregate the …

In recent years, the accessibility to the internet has been increased and many businesses like Netflix, Amazon, YouTube have grown as a result. But what is the secret to their success? One of the most important reasons can be improved customer experience. …

上一篇(1/3) 在這： [論文探討][Deep Neural Networks for YouTube Recommendations] 從YouTube看企業級推薦系統演算法設計 — (1/3) Medium 首發就獻給Deep Neural Networks for YouTube Recommendations 這篇論文啦！在讀這篇之前，小弟對於企業級大規模推薦系統的想像，都還以為只要靠一個Matrix…medium.com 在前一篇有提到Google的YouTube團隊將推薦系統視為一個two-stage information retrieval的任務，分成candidate generation以及ranking兩階段，這一篇就先來探討第一階段candidate generation是如何運作的吧！ 模型input與output 前一篇有提到，在candidate generation這個階段，處理的是一個多分類的任務，負責預測使用者下一個會看的影片，模型的input放的是與使用者相關的YouTube活動紀錄，可以放使用者最近看過的影片、搜尋過的關鍵字、使用者基本資料……等，而模型的output則可以分為training 階段以及serving 階段來介紹，在training 階段時，output會通過一層softmax來把output轉化為0~1間且和為1的機率值，並搭配cross-entropy 來optimize 模型，而在serving 階段則不通過softmax，以最後一層fully connected layer的output 當作使用者的embedding 來查詢對應到的候選集。

Medium 首發就獻給Deep Neural Networks for YouTube Recommendations 這篇論文啦！在讀這篇之前，小弟對於企業級大規模推薦系統的想像，都還以為只要靠一個Matrix Factorization 或者 Neural Network的算法就可以解決，但讀了這篇之後，實在讓我對Google的YouTube團隊感到相當佩服，顛覆我對於企業級推薦系統的認知，得到超多的啟發。雖然說這篇是在2016年發表的，以Machine Learning每年飛速更迭的角度來看或許有一點點久，但其實這樣的架構在現今業界作法中還是很常見，非常推薦大家研讀一下，況且整篇論文只有八頁(不加references剩七頁XD)，花一點時間獲取經典推薦算法不是很讚嗎XD 現在就讓我們來看看這個每日影片瀏覽量超過26億的全球最大影音分享平台，是如何規劃他們的推薦系統！ 不囉唆直接上圖，YouTube 團隊將推薦系統視為一個two-stage information retrieval的任務，第一階段candidate generation負責粗篩出候選集，第二階段ranking則從候選集中，預測使用者對各候選影片的愛好程度，細排出要推薦給使用者的topN 影片。這兩階段則分別都有一個neural network來負責所對應到的任務，candidate generation比較特別，YouTube團隊將它視為一個多分類的任務，負責預測使用者下一個會看的影片，ranking 則是負責預測點擊率的任務，大家在接觸推薦系統時所學的各種演算法，多半可以用在此階段任務上，像是常聽到的Matrix Factorization, Google的Wide&Deep, Huawei的DeepFM……等。

Recommandation systemler içerisinde kendimce en komplike gördüğüm konu olan kullanıcı tabanlı öneri yapmak üzerine yapmış olduğum bir örneği anlatmış ve aktarmış olacağım. Öneri sistemleri özünde şöyle çalışır diyebilirim: Bir sihir yapmadan var olan patternlere göre -ki bu ya user’ın bırakmış olduğu patternlerdir ya da önerilecek olan item’ın userlar üzerinde bıraktığı…

Series overview In the previous part, we reviewed the mechanics of a memory-based recommender system and built a custom collaborative filtering recommender. Today we will apply “out-of-box” recommenders from popular Python modules, evaluate their efficiency, and try some techniques to seamlessly improve the SVD prediction algorithm known as the Netflix prize winner…

Recommendations are all around us, from Youtube recommending what we should watch to Amazon proposing what book we should read. We’re constantly being suggested things, but have you ever thought how these recommendations work behind the curtains? Before we start, if recommender systems don’t mean anything to you, I highly…