Uber探索利用圖學習模型解決欺詐檢測問題

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"由於 Uber 在用戶中的知名度和規模不斷擴大,它也受到了網絡空間中金融犯罪分子的關注。欺詐行爲的一種類型是勾結,即用戶之間的合作欺詐行爲。舉例來說,用戶串通起來,用盜取的信用卡進行虛假旅程,導致信用卡退單(由銀行發起的信用卡消費退款)。本文展示了一種應用前沿的、名爲關係圖卷積網絡(relational graph convolutional network,RGCN)[1] 的深度圖學習模型,用於檢測這種勾結的案例研究。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在欺詐檢測 [2] [3] 中,圖學習方法得到了廣泛應用。例如,在 Uber Eats,已經開發出一種圖學習技術,可以浮現最有可能吸引單個用戶的美食 [4]。圖學習是一種提高我們在 Uber 平臺上食物和餐館推薦的質量和相關性的方式。在檢測勾結行爲中也可採用類似技術。從圖 1 可以看出,欺詐性用戶往往是聯繫在一起並聚集在一起的,這有助於檢測。本文概述了一種關係圖學習模型的案例研究,該模型利用這些信息來檢測勾結用戶,並使用不同的連接類型來改進學習。其目的在於分享我們在這一案例研究中的發現,並將其擴展到解決其他相關欺詐檢測。需要注意的是,在 Uber 的產品平臺上並沒有使用本文開發的模型。"}]},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/8e\/bc\/8e53dbeb8d30193a51e56a03d4ba3dbc.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"圖 1:一個連接司機和乘客的圖示。紅色節點代表欺詐用戶,藍色節點代表合法用戶。用戶通過共享信息連接。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"關係圖學習"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"將 RGCN 模型應用於小樣本數據,可以預測使用者是否存在欺詐行爲。在用戶圖中,有兩種類型的節點:通過共享信息,司機和乘客可以相互連接。每一個用戶都可以被視爲圖中的一個節點,它由一個嵌入的向量表徵。這種表徵對用戶及其鄰近社區的屬性進行編碼,可方便地應用於機器學習任務,例如節點分類和邊緣預測。例如,爲了檢測用戶是否存在欺詐行爲,我們不僅會使用該用戶的特徵,而且也會使用幾個跳數內的鄰近用戶特徵。該模型是基於對圖進行操作的神經網絡,專門爲多關係圖數據建模而開發。這種類型的圖學習已經被證明能顯著改善節點分類 [5]。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"用戶在平臺上通過共享信息互相連接。結果表明,區分不同的連接類型能夠放大欺詐檢測的信號。所以,我們將連接類型用作圖學習的一種特徵。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"要更好地理解如何爲圖用戶數據進行建模並檢測勾結行爲,瞭解 RGCN 的基礎知識是很有幫助的。圖卷積網絡(Graph convolutional network,GCN)在編碼來自結構化鄰域的特徵時,證明了效率很高 [6],在鏈接到源節點的邊緣時,它們分配的權重相同。而 RGCN 則根據邊緣的類型和方向有特定關係的轉換。這樣,對每個節點計算的消息都會增加邊緣類型信息。圖 2 展示了 RGCN 模型的示意圖。該模型的輸入包括節點特徵和邊緣類型。節點特徵被傳遞到 RGCN 層,通過聚合從所鏈接的鄰居中學到的表徵轉換成對學習表徵的向量。從相連鄰居獲得的信息由邊緣類型加權。具體地說,該模型通過加權和歸一化累積鄰近節點的信息,把這些信息傳遞給目標節點,以便學習 RGCN 層的隱層表徵,然後把這些信息傳遞給激活函數(例如 ReLU)。RGCN 層通過消息傳遞和圖卷積的方式來提取高級節點表徵。將 Softmax 層作爲輸出層,將交叉熵作爲損失函數,RGCN 模型能夠學會節點的分數。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/0f\/4b\/0f50dd568cb4134d49808925e0c1ff4b.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"圖 2:RGCN 的建模流程:將用戶的節點特徵和邊緣類型的輸入傳遞給多個 RGCN 層,從而生成節點分數。邊緣的顏色代表不同的邊緣類型。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"鄰近節點的變換特徵向量取決於特定的邊緣變換,這些變換記錄了一條邊緣的類型和方向。對於第 l+1 層的節點,還可以通過第 l 層中相應的表徵告知它們,這是將單個自連接作爲一個特殊的邊緣類型添加到每個節點上的結果。可以用第 l+1 層計算的信息來表徵爲:"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/4f\/a7\/4f35552885e1841e3c80e5ba530060a7.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/cd\/6b\/cd26ecf2402fdc386b6a36de5f45ab6b.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"用於欺詐的 RGCN"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Uber 擁有多種風險模型和多個檢查點,用於發現欺詐用戶。要更好地爲這些風險模型服務,一種想法是將欺詐分數作爲下游風險模型的特徵。RGCN 模型爲每個用戶輸出欺詐分數,表示用戶的風險。欺詐分數的學習流程如圖 3 所示。圖中每個節點的隱層表徵被學習,通過最小化二值交叉熵(binary cross entropy)損失來預測一個用戶是否是欺詐性的。一個用戶可以是司機、乘客或兩者都是,因此我們將輸出兩個分數:一個是司機,一個是乘客。將這兩個分數作爲兩個特徵被注入到下游的風險模型中。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/35\/22\/35dee4b800515a55c086ab846859a022.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"圖 2:欺詐分數建模流程:分數是通過二值交叉熵損失來學習的。模型的輸出是兩個分數,一個是司機的分數,一個是乘客的分數。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們採用兩種輸入來源:節點特徵(面向用戶的)和邊緣類型。司機 - 乘客的內存圖是用 DGL 庫 [7] 構建的。欺詐的標籤是指用戶在某個時間範圍內,是否存在信用卡退單行爲。我們通過特徵工程的方法來幫助模型學習。例如,一個司機 - 乘客圖有兩種類型的節點:司機和乘客。每一種節點類型,司機或乘客可能具有不同的特徵。針對這一問題,我們採用零填充來保證輸入的特徵向量大小相等;其次,我們專門定義了邊緣類型,並在模型訓練期間爲每種類型學習不同的權重。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"爲了評估 RGCN 模型的性能和欺詐分數的效用,我們對歷史數據進行了 4 個月的訓練,直到一個特定的分割日期。然後,我們根據分割日期之後 6 周的數據,對模型進行了性能測試。具體來說,我們爲用戶輸出了欺詐分數,並計算了準確率、召回率 和 AUC。在實驗過程中,我們觀察到,通過在現有的生產模型中增加兩個欺詐分數特徵,準確率提高了 15%,而誤報率的增加卻很小。在下游模型的 200 個特徵中,這兩種欺詐的分數分別位於第 4 位和第 39 位。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/fb\/90\/fbb703dbd0aeddcab1f6ea083c999590.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"圖 4:下游風險模型所使用的特徵重要性:從 RGCN 學到的用戶分數 1 和用戶分數 0 分別排在第 4 和第 39 位。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"數據管道"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"數據獲取"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"此前一篇關於 Uber Eats 的美食發現的博文 [4] 中,解釋了我們如何利用離線圖生態系統來生成一個城市級別的用戶 - 餐廳圖。對於這一用例,我們的需求是建立一個巨大的圖,而非一些更小的城市級圖。我們通過重用許多組件,如 Spark 上的 Cypher,可以生成多關係的用戶圖。獲取框架將源 Hive 錶轉換爲節點表和關係表。節點表捕獲用戶特徵,而關係表捕獲用戶之間不同類型的邊。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"圖劃分"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"圖的大尺寸使得分佈式訓練和預測成爲必要。初始圖被劃分成一些更小的圖,這樣它就可以放入工作機器的內存中。我們只對最近使用過 Uber 平臺的用戶的 x-hop 子圖感興趣。這些最近的“種子用戶”隨後被隨機分配到一個分區號(0 到 n)。每個種子用戶的 x-hop 子圖也被拉入同一個分區中。一個用戶可能是多個分區的一部分,或者不在任何分區中的休眠用戶。每一個分區對應着一個訓練 \/ 預測工作機器。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"通過擴展 Cypher 語言,我們爲圖的創建添加了一個分區子句。以下查詢示例將自動生成多個由分區列劃分的圖。每個分區將包含種子用戶和他們的單跳鄰居。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/ae\/d7\/ae0e932f1c65ea1e486271yya7f6ffd7.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"超級節點"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"圖生成過程的一大挑戰是如何處理超級節點,即具有極高連接量的節點。這個問題可以分爲兩個階段來處理。第一,在創建關係表的同時,對具有高連接度的實體進行過濾。舉例來說,兩個用戶通過 1000 個共享實體進行連接,會產生 10002 個用戶 - 用戶關係。但是,我們將技術作爲一個節點特徵加進來。第二,在圖的劃分階段,有些用戶在他們的子圖上表現出了具有非常高程度的不同關係。這樣會增大分區大小的差異,有些分區甚至很大。基於閾值,我們將這類用戶限制爲他們的前幾跳。這類離羣值的情況可以用規則加以追蹤。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"訓練與批量預測"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"數據管線和訓練管道如圖 5 所示。從 Apache Hive 表中獲得數據,並將包含節點和邊緣信息的 Parquet 文件作爲 HDFS 輸入,這是管道的第一步。每一個節點和邊都由一個時間戳進行版本化。具有節點和邊的最新屬性的圖被保留下來,使用 Cypher 格式存儲在 HDFS 中,並給定一個特定的日期。在使用 Apache Spark 執行引擎中的 Cypher 查詢語言將圖劃分到模型中。圖的分區被直接送入 DGL 訓練和批量預測應用程序。生成的分數存儲在 Hive 中,用於操作和離線分析。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/f6\/aa\/f6d5ae0169b6d7e8a4a1e39a55c9f8aa.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"圖 5:用於學習欺詐分數的數據管道(上行)和用於改進欺詐檢測的訓練管道(下行)"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"未來方向"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"圖學習在學術界和工業界都受到廣泛關注。它提供了一種令人信服的欺詐檢測方法。儘管圖學習已經極大地提高了檢測質量和相關性,但是還需要進一步的工作來提高系統的可擴展性和實時性。特別是,我們正在探索一種更有效的方式來存儲大規模的圖,並進行分佈式訓練和實時服務。此外,由於司機 - 乘客圖是密集連接的,爲了使信息傳遞更加有效,我們將探索基於注意力的圖模型,它利用掩蔽的自注意力層,賦予鄰域不同節點不同的重要性。例如,圖注意力網絡 [8] [9] 與我們的應用相關。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"參考文獻"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"《用圖卷積網絡對關係型數據進行建模》(Modeling Relational Data with Graph Convolutional Networks),Michael Schlichtkrull、Thomas N. Kipf、Peter Bloem、Rianne van den、Ivan Titov、Max Welling,ESWC 2018。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"《異構圖神經網絡在惡意賬戶檢測中的應用》(Heterogeneous Graph Neural Network),Ziqi Liu、Chaochao Chen 等人,CIKM 2018。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"《Web 級推薦系統中的圖卷積神經網絡》(Graph Convolutional Neural Networks for Web-Scale Recommender Systems),Rex Ying、Ruining He、Kaifeng Chen、Pong Eksombatchai、William L. Hamilton、Jure Leskovec,KDD 2018。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"《Uber Eats 的美食發現:利用圖學習爲推薦賦能》(Food Discovery with Uber Eats: Using Graph Learning to Power Recommendations),Ankit Jain、Isaac Liu、Ankur Sarda、Piero Molino。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"《欺詐檢測:基於圖的異常檢測方法的系統文獻綜述》(Fraud detection: A systematic literature review of graph-based anomaly detection approaches),Tahereh Pourhabibi、Kok-Leong Ong、Booi H.Kam、Yee Ling Boo,Decision Support Systems 2020。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"《基於圖卷積網絡的半監督分類》(Semi-Supervised Classification with Graph Convolutional Networks),Thomas N. Kipf、Max Welling,ICLR 2017。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"《Deep Graph Library:面向圖的高效可擴展深度學習》(Deep Graph Library: Towards Efficient and Scalable Deep Learning on Graphs),Minjie Wang、Lingfan Yu,ICLR 2019。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"《圖注意力網絡》(Graph attention networks),Petar Velickovi、Guillem Cucurull、Arantxa Casanova、Adriana Romero、Pietro Lio、Yoshua Bengio,ICLR 2018。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"《圖注意網絡的自適應結構指紋》(Adaptive structural fingerprints for graph attention networks),Kai Zhang、Yaokang Zhu、Jun Wang、Jie Zhang,ICLR 2020。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"作者介紹:"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Xinyu Hu,Uber 人工智能參與團隊研究科學家。致力於機器學習、因果推理和圖學習的交叉項目。擁有哥倫比亞大學生物統計學博士學位。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Chengliang Yang,Uber 風險工程團隊機器學習工程師。致力於建立基於機器學習的解決方案,以識別 Uber 平臺上的欺詐行爲。擁有佛羅里達大學計算機科學博士學位,專業爲機器學習。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Lawrence Murray,Uber 人工智能研究科學家經理,從事概率編程和貝葉斯推理的蒙特卡洛方法。擁有愛丁堡大學的信息學博士學位。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Ankur Sarda,Uber 風險工程團隊軟件工程師,致力於 Uber 的圖應用。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Ankit Jain,Uber 人工智能的前研究科學家。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Piero Molino,斯坦福大學 Hazy 研究小組成員,研究科學家。Uber 人工智能的前創始成員,創建了 Ludwig,從事應用項目(COTA、Uber Eats 的圖學習,Uber 的對話系統),並發表了關於自然語言處理、對話、可視化、圖學習、強化學習和計算機世界的研究論文。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"原文鏈接:"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"https:\/\/eng.uber.com\/fraud-detection\/"}]}]}
發表評論
所有評論
還沒有人評論,想成為第一個評論的人麼? 請在上方評論欄輸入並且點擊發布.
相關文章