大模型推薦技術(shù)及展望

1958術(shù)大1958119581.推薦及LLM簡介2.LLM賦能推薦系統(tǒng)3.大模型推薦展望3)2000“大語言模型…”規(guī)劃能力about“大語言模型…”規(guī)劃能力abouttheLLM…”4泛化能力泛化能力“新華社電…”得世界杯冠軍的歷程”學(xué)會了！學(xué)會了！《戰(zhàn)狼》講述的勾股定理是直角《戰(zhàn)狼》講述的勾股定理是直角三角形兩個(gè)直角邊的平方和等于斜邊的平方愛新覺羅.玄燁?Item：在物理世界中存在千絲萬縷的關(guān)聯(lián)，很多關(guān)聯(lián)都很低頻未知任務(wù)：個(gè)性化推薦，冷啟動(dòng)推薦…傳統(tǒng)范式大模型范式未知任務(wù)：個(gè)性化推薦，冷啟動(dòng)推薦…未知任務(wù)：個(gè)性化推薦，冷啟動(dòng)推薦…傳統(tǒng)范式大模型范式未知任務(wù)：個(gè)性化推薦，冷啟動(dòng)推薦… 序列推薦：籃球à?球à梅西à內(nèi)馬爾 對話推薦：今天沒有籃球比賽可以看，對話推薦：今天沒有籃球比賽可以看，推薦魯棒性、用戶歷史?為推薦序列推薦：籃球序列推薦：籃球à?球à梅西à內(nèi)馬爾 對話推薦：今天沒有籃球比賽可以看，對話推薦：今天沒有籃球比賽可以看，推薦魯棒性、用戶歷史?為推薦[7]Gao,etal.“Chat-REC:LLMs-Augmented[8]Liu,etal.“IsChatGPTagoodrecommender?ApreliminaryStudy”2023.[9]BaoKeqinetal.“LLM4Rec:LargeLangua[11]LiJinmingetal.“GPT4Rec:AGenerativeFrameworkforPersonalizedRecommendationandUserInterestsInterpretation”arXiv2023.[12]HouYupengetal.88919581.推薦及LLM簡介2.LLM賦能推薦系統(tǒng)3.大模型推薦展望UsenaturelanguageRecommendation“KDD2023.Recommendation“KDD2023.Recommendation“KDD2023.Recommendation“KDD2023.Recommendation“KDD2023.LiuPengfei,etal."Pre-train,Prompt,andPredict:ASystematicSurveyofPromptingMethodsinNaturalLanguageProcessing"arXiv2021.ZhangYuhui,etal."LanguageModelsasRecommenderSystems:EvaluationsandLimitations“NeurIPSWorkshop2021.16Systems“arXiv2022.Systems“arXiv2022.Systems“arXiv2022.Systems“arXiv2022. Gao,etal.Chat-REC:LLMs-AugmentedRecommenderSystem.2023.22In-contextlearning+recomModels“arXiv2023.KeqinBaoetal.“LLM4Rec:LargeLanguageModelsforRecommendationviaALightweightTuningFramework”24Recsys2023KeqinBaoetal.“LLM4Rec:LargeLanguageModelsforRecommendationviaALightweightTuningFramework”25Recsys2023○○ ◆ Dragon(Wuhuzanglong)..widerangeofentitiesthatfulfillthSystems“arXiv2023.○○ ◆ Dragon(Wuhuzanglong)..Systems“arXiv2023.Systems“arXiv2023.-BaselinesexhibitsignificantlyworseperformancethanB-ImprovementofBIGReci?possiblyduetothevaryingproperSystems“arXiv2023.Systems“arXiv2023.?IncorporatingpopularityinformaBIGRecachmetrics,particularlyforlargervaluesof“differentconventionalmodel.Systems“arXiv2023.19581.推薦及LLM簡介2.LLM賦能推薦系統(tǒng)3.大模型推薦展望候選物品排序角色扮演思維鏈Action?調(diào)用Action?調(diào)用LLMAPI?更新獎(jiǎng)勵(lì)函數(shù)reward?大模型黑盒特性，參數(shù)不可訪問yTrainingSamples[Prompt]Lightweightlaptopwithnewamazingfeatures,batterylifeisawesome.Accuracy=78%feedbackStep1yTrainingSamples[Prompt]Lightweightlaptopwithnewamazingfeatures,batterylifeisawesome.Accuracy=78%feedbackStep1ProvidefeedbackonvariousproductsStepnProvideasentimentyAccuracy=85%yTestsamplesw/thesamedistribution(sourcegroup)Testsamplesw/adifferentdistribution(targetgroup)PromptOptimizationDeploymentDeploymentoptimizedpromptoptimizedpromptMoxinLietal.,SubmittedtoEMNLP,AnonymousTitle.2023新推薦范式（GenerativeRecommendation）應(yīng)用場景：o2.AI作畫o3.AI改畫o4.AI換裝WenjieWangetal.arxiv,GenerativeRecommendation:TowardsNext-generationRecommenderParadigm.2023薦會繼承現(xiàn)有的社會偏見！JizhiZhangetal.Recsys,CanChatGPTMakeFairRecommendation?AFairnessEvaluationBenchmarkforRecommendationwithLargeLanguageModel.2023?Item：在物理世界中存在千絲萬縷的關(guān)聯(lián)，很多關(guān)聯(lián)