胡斌斌-基于大模型的推薦算法及應用.pdf

《胡斌斌-基于大模型的推薦算法及應用.pdf》由會員分享，可在線閱讀，更多相關《胡斌斌-基于大模型的推薦算法及應用.pdf（30頁珍藏版）》請在三個皮匠報告上搜索。

1、DataFunSummit#2024Large language model based Recommender System and Application胡斌斌螞蟻集團算法專家01Background0302LLM as Knowledge Extractor04LLM as a Reasoning Pool目錄 CONTENTLLM as Teacher RecommenderWorkflow of RSs1.Train recommender on collected interaction data to capture user preferences.2.Recommender

2、generates recommendations based on estimated preferences.3.User engage with the recommended items,forming new data,affected by open world.4.train recommender with new data again,either refining user interests or capturing new ones.Current recommender systems are oftentrainedonaclosed-loopuser-itemin

3、teraction dataset,inevitably sufferingfrom severe exposure bias and popularitybias.Jizhi Zhang et al.Large Language Models for Recommendation:Progresses and Future Directions.WWW 2024.Jiawei Chen et al.Bias and debias in recommender system:A survey and future directions.TOIS 2023.Development of LMsL

4、arge Language Model：billions of parameters,emergent capabilities Rich knowledge&Language Capabilities Instruction following In-context learning Chain-of-thought Planning Jizhi Zhang et al.Large Language Models for Recommendation:Progresses and Future Directions.WWW 2024.From of CF to LLM based RSFro

5、m Shallow Models,to Deep Models,to Large ModelsShallow ModelsDeep ModelsLarge ModelsKoren et al.Matrix factorization techniques for recommender systems.Computer 2009.Heng-Tze Cheng et al.Wide&deep learning for recommender systems.”DLRS 2016.Jizhi Zhang et al.Large Language Models for Recommendation:

6、Progresses and Future Directions.WWW 2024.Key Challengesp Tend to rely on semantics,and another important aspect of recommendation tasks is collaborative information.p Balance the trade-off between the cost and effectiveness.p Adapt the reasoning ability of LLMs to Recommendation.LLM for Recommendat

7、ionEffortlessly Integrating LLM Insights into RSKnowledge ExtractorReasoning PoolTeacher Recommenderdistillation01Background0302LLM as Knowledge Extractor04LLM as a Reasoning Pool目錄 CONTENTLLM as Teacher RecommenderEntity/RelationExtractionRepresentationLearningCurrent corpusTarget-oriented KGTradit

8、ional MethodsLLM as Knowledge ExtractorXiangnan He et al.Neural Collaborative Filtering.WWW 2018.Xiang Wang et al.Neural Graph Collaborative Filtering.SIGR 2019.Guorui Zhou et al.Deep Interest Network for Click-Through Rate Prediction.KDD 2018.LLM as Knowledge Extractor關系生成實體擴散Prior KnowledgeFormula

9、tion知識關系過濾目標實體生成LLM通常很難感知營銷場景中給定實體的真正含義，例如，運動鞋相關的實體可能更喜歡給定實體“Air”。為此我們準備了圖譜級別的結構知識和描述式的本知識純粹的關系成過程是理想的，但通常是不可控的。我們借助LLM的強能，通過設計的prompt從預定義的關系集 R 中過濾所需的關系 r。由于先驗知識的多樣性，再加上提示在輕微修改下的敏感性，單的提示可能并不總是產預期的結果。因此我們通過漸進式的增提示來多次驅動LLM，并設計可靠的聚合策略來進實體擴散。Chunjing Gan et al.Making Large Language Models Better Knowle

10、dge Miners for Online Marketing with Progressive Prompting Augmentation.Arxiv2024.Making LLM as better Knowledge Miner關系檢索考慮到整個關系集R（數100+）不能直接輸到單個Prompt中，先通過查詢原始圖譜來定位實體的類型。給定實體類型，我們可以粗略的過濾個候選關系集（數在1020之間）。我們將LLM視為個關系過濾器，其中的prompt最有結構知識(“Struc_KG”)和描述性(“Desc_KG”)知識增強，產出 i）符合需求的關系集和 ii）個潛在的標實體，可以作為實體擴

11、散的提示。關系過濾漸進式的prompting增廣聚合策略知識多樣性：-結構式知識：kS-描述式知識：kD-繼承式知識：kI組合式prompting LLMs語義相關性KG-BERT融合致性結論：-效果：我們的模型在所有核心指標上都能取得最優（或者接近最優）的效果。-對比圖譜補全的方法：雖然這類方法可以在acc上取得和我們的模型相當的效果，但無法為圖譜補充新的實體，且豐富度較低。-對比圖譜生成的方法：這類方法可以為圖譜帶來的知識，但在ACC指標上普遍較低-消融實驗：驗證了聚合策略、漸進式的prompt增廣和關系過濾的有效性結論：-我們發現，攜帶知識產出的目標實體與對應的源實體更加相關。-知識可以

12、有效防止幻覺實體和謬誤實體。PerformanceRAG Enhanced Knowledge MinerRAG with Similarity v.s.Utility Similarity UtilitySummaryEndowing RAG with Multilayered ThoughtIssues Capability of RAG not solely hinges on similarity Fail to abstract enough useful information for downstream applicationChunjing Gan et al.Similar

13、ity is Not All You Need:Endowing Retrieval Augmented Generation with Multi Layered Thoughts.Arxiv2024.Downstream ApplicationXiaoling Zang et al.Commonsense Knowledge Graph towards Super APP and Its Applications in Alipay.KDD 2023.Chunjing Gan et al.PEACE:Prototype lEarning Augmented transferable fra

14、mework for Cross-domain rEcommendation.WSDM 2023.Weifang Wang et al.GARCIA:Powering Representations of Long-tail Query with Multi-granularity Contrastive Learning.ICDE 2023.Weifang Wang et al.The Devil is in the Sources!Knowledge enhanced Cross-domain Recommendation in a Information Bottleneck Persp

15、ective.CIKM 2024(under review)Weifang Wang et al.Embracing Disentanglement:Graph-based Recommenders Can Be Easily Distracted by Redundant Knowledge.CIKM 2024(under review)01Background0302LLM as Knowledge Extractor04LLM as a Reasoning Pool目錄 CONTENTLLM as Teacher RecommenderStep-by-step knowledge dis

16、tillation(SLIM)Yuling Wang et al.Can Small Language Models be Good Reasoners for Sequential Recommendation?WWW 2024.CoT Prompting for Larger Teacher ModelGuiding the teacher LLMs in generatingcritical reasoning,all of which are essentialfor providing appropriate recommendations.Step1.Summarize user

17、preferences basedon the historical behavior sequences.Step2.Recommendcategories/brandsbased on the summarized preferences.Step3.Recommend products that align withthe recommended categories/brands.Extracting Rationales from LLMsSFT with Recommendation RationalesFor a given input instruction!,we train

18、 the smaller model with parameters to generate the correspondingrationale!from the larger teacher model.Knowledge distillation transfer the recommendation reasoning capabilities of larger teacher models to smallerstudent models,reducing the computational overhead.Empowering RS with Reasoning Knowled

19、ge Utilizing Recommendation Rationales.Encoding Recommendation RationalesID-Based RecommendationID-Agnostic RecommendationExperimentsPromising performanceCold-start issuePopularity biasq Knowledge distillation matters in LLM4RecqThe high inference latency of LLMs significantly restricts their practi

20、cal deployment in RS.q Challenges:q1）The teachers knowledge may not always be reliable.q2）The capacity gap makes it difficult for the student to assimilate the teachers knowledge.q3）The divergence in semantic space poses a challenge to distill the knowledge from embeddings.2222The inference latency

21、of BIGRec far exceeds that of DROS.BIGRec does not always outperform DROS.Distill into More Smaller Model DLLM2Rec23231)Importance-aware Ranking Distillation filter reliable and student-friendly knowledge by weighting instances2)Collaborative Embedding Distillation integrate knowledge from teacher e

22、mbeddings with studentsRanking PositionHigher ranked items by teachers are more reliableTeacher-Student Consensus The items recommended by both teacher and student are more likely to be positiveConfidence of LLMsThe distance between the generated descriptions with the target itemDLLM2RecYu Cui et al

23、.Distillation Matters:Empowering Sequential Recommenders to Match the Performance of Large Language Model.arXiv 20242424qDLLM2Rec boosting three typical sequential models with an average improvement of 47.97%,even enabling them to surpass LLM-based recommenders in some cases.DLLM2Rec01Background0302

24、LLM as Knowledge Extractor04LLM as a Reasoning Pool目錄 CONTENTLLM as Teacher RecommenderOnly a Reasoning Pool is EnoughInsightsNot all users need LLM reasoning Adaptive sampling Efficient retrievalChangxin Tian et al.Only a Controllable Reasoning Pool is Enough!Effortlessly Integrating Large Language

25、 Models Insights into Industrial Recommenders.RecSys 2024(under review)Extracting LLM Insights with Seed UsersPrompt Template Representation aware sampling Graph Pretrain-K-Means Importance aware sampling Prioritizes users with more interactionsAdaptive SamplingScaling LLM Insights via RetrievalRetr

26、ieve the rationales of seed users to serve asthe rationale for their corresponding users.Enhancing Recommender with LLMs InsightsPerformanceEmpowering current recommendation backbonesMore seed usersBetter performanceHigher quality of user representationsApplication Content recommendation Click-PV+x.xx%,CTR+x.xx%Marketing Click-UV+x.xx%Member match/rank CTR+x.xx%,CVR+x.xx%Better performance for users with lower activity感謝觀看