多模態

VIP內容

題目:End-to-End Entity Classification on Multimodal Knowledge Graphs

簡介:

知識圖的端到端多模式學習在很大程度上尚未解決。取而代之的是,大多數端到端模型(例如消息傳遞網絡)僅從圖形結構中編碼的關係信息中學習:原始值或文字要麼被完全省略,要麼從其值中剝離而被視為常規節點。無論哪種情況,我們都會丟失潛在的相關信息,而這些信息本來可以被我們的學習方法所利用。為避免這種情況,我們必須將文字和非文字視為單獨的情況。我們還必須分別並相應地處理每種形式:數字,文本,圖像,幾何形狀等等。我們提出了一種多模態消息傳遞網絡,該網絡不僅可以從圖的結構中端到端學習,而且可以從它們的多模態節點特征集合中學習。我們的模型使用專用的(神經)編碼器來自然學習節點特征的嵌入,這些節點特征屬於五種不同類型的模態,包括圖像和幾何圖形,這些圖像連同其關係信息被投影到聯合表示空間中。我們在節點分類任務上演示我們的模型,並評估每種模式對整體性能的影響。我們的結果支持我們的假設,即包含來自多種模式的信息可以幫助我們的模型獲得更好的整體性能。

成為VIP會員查看完整內容
0
19
3

最新論文

Due to limited transit network coverage and infrequent service, suburban commuters often face the transit first mile/last mile (FMLM) problem. To deal with this, they either drive to a park-and-ride location to take transit, use carpooling, or drive directly to their destination to avoid inconvenience. Ridesharing, an emerging mode of transportation, can solve the transit first mile/last mile problem. In this setup, a driver can drive a ride-seeker to a transit station, from where the rider can take transit to her respective destination. The problem requires solving a ridesharing matching problem with the routing of riders in a multimodal transportation network. We develop a transit-based ridesharing matching algorithm to solve this problem. The method leverages the schedule-based transit shortest path to generate feasible matches and then solves a matching optimization program to find an optimal match between riders and drivers. The proposed method not only assigns an optimal driver to the rider but also assigns an optimal transit stop and a transit vehicle trip departing from that stop for the rest of the rider's itinerary. We also introduce the application of space-time prism (STP) (the geographical area which can be reached by a traveler given the time constraints) in the context of ridesharing to reduce the computational time by reducing the network search. An algorithm to solve this problem dynamically using a rolling horizon approach is also presented. We use simulated data obtained from the activity-based travel demand model of Twin Cities, MN to show that the transit-based ridesharing can solve the FMLM problem and save a significant number of vehicle-hours spent in the system.

0
0
0
下載
預覽
Top