Aixin Sun 孙爱欣
@AixinSG
Followers
1K
Following
2K
Statuses
2K
Do a bit of research in information retrieval and a bit in recommender systems, mostly for fun.
Singapore
Joined July 2011
RT @RecSys_c: It is almost a year ago that #recsperts released a new episode about #recsys. @MarcelKurovski, when will there be a new one?
0
1
0
RT @ZouJie_IR: Anyone interested in doing a PhD/PostDoc in China? Come join my research team!! A full international PhD scholarship and a…
0
1
0
RT @DavenCheung: 🚨 Call for Contributions! 🚨 #CallForPapers Our #ACM_TORS special issue on Generative AI for Recommender Systems (#GenAI…
0
1
0
📢New #ACM_TORS #CfP for a special issue in generative models for #recsys (1 July, 2025). Plus, the deadline for the user interaction design for human-centred recommender systems is extended. #LLMs #GenerativeAI #RecommenderSystems #ACM
0
0
3
RecSys presents an intriguing challenge where the ground truth is not well-defined. Users make decisions based on the candidates suggested by the existing system (along with other influencing factors), which in turn shape their choices and, by extension, the so-called ground truth. I encountered a similar issue while working on the ticket routing problem, where the ground truth is equally ambiguous. When a user encounters an issue, they report it to the service provider as a ticket. This ticket is then forwarded to the support team responsible for resolving the problem. Often, a support team specializing in databases may identify the issue as a networking problem and route the ticket to the networking team, which might further route it to the GUI team for resolution. We are provided with these routing sequences, and our goal is to optimize the process by minimizing the number of steps, thereby reducing delays and unnecessary work. But is there a definitive ground truth? Certainly not—the existing process is shaped by human decisions. Should we aim to solve the issue by routing the ticket directly to the final team that resolves it? Not necessarily, as input from intermediate teams may be crucial for the final team to accurately identify and address the problem. Exploring these nuances is part of what makes RecSys—and related problems like ticket routing—so compelling. -- revised by ChatGPT
0
1
10
Thank you for the nice summary. I especially like the last sentence, and my group will certainly work toward enhancing diversity in retrieval mechanisms.
0
0
2
@jenova_ai_ @omarsar0 Fully agree that LC and RAG are two complementary tools that can be effectively used together.
0
0
0
We briefly discussed this point at the end of Section 6.1 in the paper: "realistic and synthetic long texts can only serve as proxies to reflect context relevance to some extent. The scope of the context is question-dependent and difficult to define clearly." I believe that not all contexts can be effectively captured through a pre-defined relevance measure.
@omarsar0 Ig the problem mainly would stem from the fact that the retrieval sysyem itself is not perfect, i.e not all relevant information was retrieved, otherwise a very well made retrieval system should perform as good if not better than throwing everything in the context
0
0
3
@jobergum @tanmay_patil It is relevant for fragmented context; but the key issue of RAG is that the so called "context" is defined by the similarity measures
0
0
0
The 'Unrelated to the current video' option in the feedback form aligns perfectly with my earlier illustration of RecSys dynamics: the next recommendation often depends on what you're currently watching.
0
0
2