µ¥ÀÌÅͺ£À̽º ¿¬±¸È¸Áö(SIGDB)
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
ºÐ»ê ȯ°æ¿¡¼ Ŭ¸¯·ü ¿¹ÃøÀ» À§ÇÑ °³³ä º¯È ÀûÀÀ |
¿µ¹®Á¦¸ñ(English Title) |
Concept Drift Adaptation for Click-Through Rate Prediction in a Distributed Environment |
ÀúÀÚ(Author) |
ÃÖ¼ºÁØ
¹®Ã¶ÇÑ
¹ÎÁرâ
Seongjun Choe
Cheolhan Moon
Junki Min
ÀÌÁ¾ÇÐ
±è»óÈ£
À̱âÈÆ
JongHak Lee
Sang-ho Kim
Ki-Hoon Lee
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 39 NO. 02 PP. 0017 ~ 0030 (2023. 08) |
Çѱ۳»¿ë (Korean Abstract) |
¿Â¶óÀÎ ±¤°í ½Ã½ºÅÛ¿¡¼ ¼öÀÍÀ» ±Ø´ëÈÇϱâ À§Çؼ´Â Ŭ¸¯·ü(click-through rate)À» Àß ¿¹ÃøÇÏ´Â °ÍÀÌ Áß¿ä
ÇÏ´Ù. Ŭ¸¯·üÀº »ç¿ëÀÚ¿¡°Ô ³ëÃâµÈ ±¤°í¸¦ »ç¿ëÀÚ°¡ Ŭ¸¯ÇÒ È®·üÀÌ´Ù. Ãֱ٠Ư¡ °£ÀÇ ÀúÂ÷ ¹× °íÂ÷ »óÈ£ÀÛ¿ë
À» ¸ðµÎ ÇнÀÇϱâ À§ÇÑ Å¬¸¯·ü ¿¹Ãø ¸ðµ¨µé(DeepFM µî)ÀÌ Á¦¾ÈµÇ¾úÁö¸¸, ½Ç½Ã°£À¸·Î »ý¼ºµÇ´Â Ŭ¸¯ µ¥ÀÌÅÍ
ÀÇ °³³ä º¯È(concept drift)¸¦ °í·ÁÇÏÁö ¾Ê¾Æ ¸ðµ¨ÀÇ ¿¹Ãø ¼º´ÉÀÌ ÀúÇ쵃 ¼ö ÀÖ´Ù. °³³ä º¯È´Â ½Ã°£¿¡ µû¶ó
µ¥ÀÌÅÍÀÇ Åë°èÀûÀΠƯ¼ºÀÌ º¯ÇÏ´Â Çö»óÀÌ´Ù. º» ³í¹®¿¡¼´Â ºÐ»ê ȯ°æ¿¡¼ °³³ä º¯È¿¡ ÀûÀÀÇÒ ¼ö Àִ Ŭ¸¯·ü
¿¹Ãø ¹æ¹ýÀ» Á¦¾ÈÇÑ´Ù. Á¦¾ÈÇÑ ¹æ¹ýÀº DeepFM ¸ðµ¨¿¡ ¿Â¶óÀÎ ¸Ó½Å·¯´×(online machine learning)À» Àû¿ëÇÑ
´Ù. ±×¸®°í Àü¹®°¡ È¥ÇÕ¸ðµ¨(mixture of experts)¸¦ »ç¿ëÇØ ¿Â¶óÀÎ ¸Ó½Å·¯´×À» Àû¿ëÇÑ DeepFM ¸ðµ¨µéÀ» ¾Ó
»óºíÇÑ´Ù. Àü¹®°¡ È¥ÇÕ¸ðµ¨ÀÇ Ã³¸®À²(throughput)À» Çâ»ó½ÃÅ°±â À§ÇØ ºÐ»ê 󸮸¦ Àû¿ëÇÑ´Ù. Criteo µ¥ÀÌÅÍ
¼ÂÀ» ÀÌ¿ëÇØ ½ÇÇèÇÑ °á°ú, Á¦¾ÈÇÑ ¹æ¹ýÀÌ DeepFM ¸ðµ¨º¸´Ù Æò±Õ AUC°¡ 4.6% ´õ ³ô¾Ò´Ù. ±×¸®°í 4´ëÀÇ ¿öÄ¿
³ëµå(worker node)·Î ºÐ»ê ó¸®ÇÔÀ¸·Î½á ó¸®À²À» 381% Çâ»ó½ÃÄ×´Ù. |
¿µ¹®³»¿ë (English Abstract) |
In an online advertising system, click-through rate (CTR) prediction is crucial for maximizing
revenue. The CTR represents the probability of a user clicking on a displayed advertisement.
Recently, CTR prediction models, such as DeepFM, have been proposed to learn both low- and
high-order feature interactions. However, these models may suffer from performance degradation
as they do not consider the concept drift of the click data generated in real-time. Concept drift is
a phenomenon where the statistical properties of data change over time. In this paper, we propose
a CTR prediction method that can adapt to the concept drift in a distributed environment. The
proposed method applies online machine learning to the DeepFM model and employs a mixture
of experts for the ensemble of the DeepFM models applied online machine learning. We performed
distributed processing to improve throughput of the mixture of experts. The experimental results
using the Criteo dataset show that the proposed method accomplishes 4.6% higher average AUC
compared with the DeepFM model. Furthermore, by distributing the processing across four worker
nodes, we improved throughput by 381% compared with processing on a single worker node. |
Å°¿öµå(Keyword) |
º´·Ä ó¸®
±×·¡ÇÈ Ã³¸® À¯´Ö
»èÁ¦ ÄÚµå
Äڽà ¸®µå-¼Ö·Î¸ó ÄÚµå
Parallel processing
Graphic processing unit
Eras
Ŭ¸¯·ü ¿¹Ãø
°³³ä º¯È
¿Â¶óÀÎ ¸Ó½Å·¯´×
Àü¹®°¡ È¥ÇÕ¸ðµ¨
ºÐ»ê ó¸®
click-through rate prediction
concept drift
online machine learning
mixture of experts
distributed environment
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|