Posts with #Dataselection
Public posts on X that use #Dataselection.
Xiao Tian @snoidetx✨ 𝐈𝐂𝐌𝐋 2026 𝐒𝐩𝐨𝐭𝐥𝐢𝐠𝐡𝐭✨ In #DataSelection, a common practice is to use #DataValuation methods (e.g., #DataShapley, #Semivalues) to compute the worth of each data point, and select the top-m data points. But more than often, the selection is not even better than random. #DataValuation researchers can only demonstrate on limited datasets and settings. Is #DataShapley really not better than random in data selection? Our #ICML2026 Spotlight paper, NASH, answers NO 🙅♀️. What’s the problem? Imagine a company hiring from applicants who each receive an overall score. If the company simply hires the highest-scoring candidates, it may end up with a team full of excellent coders, but no product managers, salespeople, or designers. Top-m applicants ≠ complete team 🙅♂️, which needs to fulfil a variety of roles. NASH addresses this by (1) breaking down the complex objective function to multiple simpler components (≈ roles) so that the data values on these components accurately reflect each data point’s unique strengths and weaknesses 📊; (2) forming a new objective which non-linearly aggregates these components so that a data point is preferred when its strengths complement the other selected data points 🧑🧑🧒🧒. NASH consistently and substantially boosts the performance of Data Shapley with minimal additional runtime cost. Future research on #DataShapley would not need to limit their datasets and settings any more 😁. Drop by our poster to say hi or discuss more 🤝! 📍 Poster: #3202 | Hall A 🕔 Time: 8 July 2026 (Wed), 5:00 - 6:45 pm 📖 OpenReview: 📝 Authors: @snoidetx @YvonneFan12 @RachaelSim2 Zixuan Wang, Nancy F. Chen, @bryanklow #DataSelection #DataValuation #MachineLearning #LLM
Bryan Kian Hsiang Low @bryanklow
Xiao Tian @snoidetx✨ 𝐈𝐂𝐌𝐋 2026 𝐒𝐩𝐨𝐭𝐥𝐢𝐠𝐡𝐭✨ In #DataSelection, a common practice is to use #DataValuation methods (e.g., #DataShapley, #Semivalues) to compute the worth of each data point, and select the top-m data points. But more than often, the selection is not even better than random. #DataValuation researchers can only demonstrate on limited datasets and settings. Is #DataShapley really not better than random in data selection? Our #ICML2026 Spotlight paper, NASH, answers NO 🙅♀️. What’s the problem? Imagine a company hiring from applicants who each receive an overall score. If the company simply hires the highest-scoring candidates, it may end up with a team full of excellent coders, but no product managers, salespeople, or designers. Top-m applicants ≠ complete team 🙅♂️, which needs to fulfil a variety of roles. NASH addresses this by (1) breaking down the complex objective function to multiple simpler components (≈ roles) so that the data values on these components accurately reflect each data point’s unique strengths and weaknesses 📊; (2) forming a new objective which non-linearly aggregates these components so that a data point is preferred when its strengths complement the other selected data points 🧑🧑🧒🧒. NASH consistently and substantially boosts the performance of Data Shapley with minimal additional runtime cost. Future research on #DataShapley would not need to limit their datasets and settings any more 😁. Drop by our poster to say hi or discuss more 🤝! 📍 Poster: #3202 | Hall A 🕔 Time: 8 July 2026 (Wed), 5:00 - 6:45 pm 📖 OpenReview: 📝 Authors: @snoidetx @YvonneFan12 @RachaelSim2 Zixuan Wang, Nancy F. Chen, @bryanklow #DataSelection #DataValuation #MachineLearning #LLM
Open Data Watch @OpenDataWatch
