![Introducing ZESRec [1]
Zero Shot Recommender System
[1] āZero Shot Recommender Systemsā, Hao Ding, Anoop Deoras, Yuyang Wang, Hao Wang. ICLR Workshop 2022
ā¢ ZESRec learns the universal item embeddings based on domain-agnostic generic features ā text;
ā¢ ZESRec adopts sequential recommenders which generates the universal user embeddings
14](https://image.slidesharecdn.com/netflixprs2023anoopdeoras-230718045132-ad79b704/85/Foundation-Models-in-Recommender-Systems-14-320.jpg)
![Introducing LMRecSys[3]
Converting userās interaction history into a text inquiry ā Prompts
science fiction film directed by Peter Weir. The screenplay by Andrew Nicole was
adapted from Nicoleās 1997 novel of the same name. The film tells the story of
Truman Burbank, a man who is unwittingly placed in a televised reality show that
broadcasts every aspect of his life without his knowledge.
A user watched Jaws, Saving Private Ryan, The Good, the Bad, and the Ugly, Run Lola
Run, Goldfinger. Now the user may want to watch something funny and light-hearted
comfort him after having seen some horrors.
Knowledge
Reasoning
J1-Jumbo
Large Pre-trained Language
Model
(178B Parameters)
Bolded texts are generated by the
model.
A user watched Jaws, Saving Private Ryan, The Good, the Bad, and the Ugly, Run Lola Run, Goldfinger.
Now the user may want to watch __ __ __
p(d(xt)| f([d(x1), . . . , d(xtā1)]))
Item 372 Item 168 Item 413 Item 77 Item 952
p(xt |x1, . . . , xtā1)
Item 1
Item 2
Item N
ā¦
Recommended Item
Token 1
Token 2
Token V
ā¦
Token 1
Token 2
Token V
ā¦
Token 1
Token 2
Token V
ā¦
Item 1
Item 2
Item N
ā¦
Recommended Item
Predicted Token Distributions from Language Models
Enable zero-shot recommendation
Improve data eļ¬ciency
Goal
GRU4Rec
Traditional Recommender System
LMRecSys
PLMs as Recommender System
[3] āLanguage Models as Recommender Systems: Evaluations and Limitationsā,
Yuhui Zhang, Hao Ding, Zeren Shui, Yifei Ma, James Zou, Anoop Deoras, Hao Wang. NeurIPS Workshop 2021
23](https://image.slidesharecdn.com/netflixprs2023anoopdeoras-230718045132-ad79b704/85/Foundation-Models-in-Recommender-Systems-23-320.jpg)
![Recent Advances in Merging LLMs with RecSys
FineTuning an LLM
M6-Rec[5]:
P5[4]: designed a text to text
fi
ne-tuning
paradigm based on the pre-trained T5.
[4] āRecommendation as language processing (rlp): A uni
fi
ed pretrain, personalized prompt & predict paradigm (p5)ā,
Geng Shijie et.al.. RecSys 2022
[5] āM6-Rec: Generative Pretrained Language Models are Open-Ended Recommender Systemsā,
Zeyu Cui et.al.. ArXiv 2022
28](https://image.slidesharecdn.com/netflixprs2023anoopdeoras-230718045132-ad79b704/85/Foundation-Models-in-Recommender-Systems-28-320.jpg)
![Recent Advances in Merging LLMs with RecSys
Inference with LLM
[6] "Zero-Shot Next-Item Recommendation using Large Pretrained Language Models." Wang, Lei, and Ee-Peng Lim. ArXiv 2023.
[7] āChat-REC: Towards Interactive and Explainable LLMs-Augmented Recommender Systemā, Yunfan Gao et.al. ArXiv 2023
Zeyu Cui et.al.. ArXiv 2022
ā¢ NIR [6], Chat-REC[7] and [8] propose to directly recommend using LLMs ā Inference only.
ā¢ Most e
ff
ort spent around āPrompt Engineeringā
ā¢ Optimal encoding of user context in the prompts
ā¢ āOut of Vocabularyā problems solved using techniques such as candidate pools, text-matching
ā¢ Mixed success. Still a long way to go.
[8] āIs ChatGPT a Good Recommender? A Preliminary Study ā, Junling Liu et.al. ArXiv 2023](https://image.slidesharecdn.com/netflixprs2023anoopdeoras-230718045132-ad79b704/85/Foundation-Models-in-Recommender-Systems-29-320.jpg)
Foundation Models in Recommender Systems
- 1. Zero Shot Recommenders, LLMs and Prompt Engineering PRS Workshop, Net fl ix, 2023 June 9th, 2023 Hao Ding (haodin haoding2019 ) and Anoop Deoras (adeoras ) AWS AI, Amazon 1 Towards Building Foundation Models in Recommender Systems
- 2. Our Mission at AWS Put Machine Learning in the Hands of Every Developer 2
- 3. The AWS ML Stack Broadest and Most Complete Set of ML Capabilities GenAI NEW Bedrock CodeWhisperer 3
- 4. Amazon Personalize Who are we in a nutshell ? ā¢ Customers can elevate the user experience with ML-powered personalization ā¢ We cater to many thousands of customers from many diverse domains ā¢ Such as: Retail, News and Media, Video on Demand, Travel and Hospitality, .. ā¢ We provide recommendations that respond in real-time to changing user behavior ā¢ In short, we provide the concierge service for all things personalization 4
- 5. Amazon Personalize Who are we in a nutshell ? 5
- 6. Customer Obsessed Science Applied Research at AWS AI ā¢ Constantly innovating on behalf of the customers ā¢ Amazon fundamentally believes that scienti fi c innovation is essential to being the most customer- centric company in the world ā¢ Science at Amazon enables new customer experiences, addresses existing customer pain points, complements engineering and product disciplines. 6
- 7. 3 Anchors for the Discussion Today ColdStart, Foundation Models in RecSys and LLMs ā¢ Cold Start Problems in Recommender Systems ā¢ Foundation Models in Recommender Systems ā¢ Role Large Language Models (LLMs) can play in Recommender Systems 7
- 8. 3 Cold Start Problems in Recommender System ā¢ Cold Users: Users during inference are unseen during training and model needs to generalize ā¢ Cold Items: New items get introduced to catalogue ā¢ Cold Domains: Target data available only for inference. No Models can be built. ā¢ Less extreme case: Domains with very little training data / less frequent training cadence ā¢ Performance of RecSys relies heavily on the amount of training data available 8
- 9. Foundation Models in Recommender Systems Why should we talk about them ? ā¢ De fi nition of a Foundation Model: A model trained on broad data that can be adapted to a wide range of downstream tasks. ā¢ Why Foundation Models in RecSys? Two main selling points: ā¢ They encode āworld knowledgeā, thus complementary to models on domainās behavioral data ā¢ LLM Foundation Modelsā interactive nature can potentially help with explaining away the recommendations 9
- 10. Two Approaches for Building Foundation Models RecSys from Other Domains, Large Language Models ā¢ We will talk about 2 research e ff ort ā¢ ZeroShot Learning: Can we leverage the knowledge in one domain to kick start a recommendation in a completely di ff erent domain ā¢ ZeroShot Inference: We will further assume that we have no source domain to rely on. How can we kick start a recommendation with large language models 10
- 11. ZeroShot Learning Kind of Like Domain Adaptation but with zero User/Item overlap 11
- 12. The Status-Quo Collaborative Filtering, Item IDs and their Embeddings ā¢ Current RecSys models learn item ID embeddings through interactions ā¢ Item ID Embeddings are parameters of your neural network and we learn them via BackProp ā¢ These embeddings are indexed by categorical domain speci fi c item ID ā¢ These are transductional and not generalizable to unseen items 12
- 13. Concept of Universal Item Embeddings Collaborative Filtering, Item IDs and their Embeddings ā¢ The idea behind universal item embeddings is to tap into itemās content information. ā¢ e.g. Natural Language product description / movie synopsis etc ā¢ Strong NLP models are used to obtain continuous universal item representations ā¢ Universal user representations can then be built on top of these universal item representations. 13
- 14. Introducing ZESRec [1] Zero Shot Recommender System [1] āZero Shot Recommender Systemsā, Hao Ding, Anoop Deoras, Yuyang Wang, Hao Wang. ICLR Workshop 2022 ā¢ ZESRec learns the universal item embeddings based on domain-agnostic generic features ā text; ā¢ ZESRec adopts sequential recommenders which generates the universal user embeddings 14
- 15. We want to ask 2 questions about ZESRec Relevance, Lead Time ā¢ How relevant are ZESRec recommendations compared to a fully trained systems ? ā¢ How much in domain data is needed to outperform ZESRec ā¢ How much is the lead time ? 15
- 16. High Level Approach ZESRec Training SEQ SEQ SEQ ā¦ User Universal Embedding 1-Layer NN Pretrained BERT Model X 1-Layer NN Pretrained BERT Model ā¦ 0.36 0.29 ā¦ 0.09 0.02 Prediction Score Item Universal Embedding Pretrained BERT Model 1-Layer NN Item Universal Embedding Pretrained BERT Model 1-Layer NN Item Universal Embedding Item Universal Embedding Item Universal Embedding ā¦ ā¦ Latent Item Offset Vector + Latent Item Offset Vector + Latent Item Offset Vector + Latent Item Offset Vector ā¦ Latent Item Offset Vector + + Latent User Offset Vector 16
- 17. High Level Approach ZESRec Inference SEQ SEQ SEQ ā¦ User Universal Embedding 1-Layer NN Pretrained BERT Model X 1-Layer NN Pretrained BERT Model ā¦ 0.36 0.29 ā¦ 0.09 0.02 Prediction Score Item Universal Embedding Pretrained BERT Model 1-Layer NN Item Universal Embedding Pretrained BERT Model 1-Layer NN Item Universal Embedding Item Universal Embedding Item Universal Embedding ā¦ ā¦ 17
- 19. Results How long before In-Domain Model Takes over ? 19 10K 10K 5K 5K 2.5K 2.5K 0 0 Number of Interactions Number of Interactions 0.04 0.02 0 0.04 0.02 0 0.06 0.08 Recall@20 Recall@20 MIND dataset Amazon dataset
- 20. ZeroShot Inference No reference recommender system at hand 20
- 21. From ZeroShot Learning to ZeroShot Inference Task and Limitations ā¢ Now lets imagine we donāt have the luxury of even having any source domain RecSys ā¢ How realistic this assumption is ? Answer: Quite Realistic (startups, new business lines ..) ā¢ What can we do ? ā¢ There is no learning part left for ZeroShot Learning ā¢ We need to resort to ZeroShot Inference 21
- 22. LLM Foundation Models to the rescue Can we kick start recommendations using Large Language Models ? ā¢ Pre-trained language models such as BERT and GPT learn general text representations ā¢ They encode āworld knowledgeā ā¢ Question we want to ask: Can we leverage these powerful LLMs as recommender systems ā¢ Use prompts to reformulate session based recommendation task 22
- 23. Introducing LMRecSys[3] Converting userās interaction history into a text inquiry ā Prompts science fiction film directed by Peter Weir. The screenplay by Andrew Nicole was adapted from Nicoleās 1997 novel of the same name. The film tells the story of Truman Burbank, a man who is unwittingly placed in a televised reality show that broadcasts every aspect of his life without his knowledge. A user watched Jaws, Saving Private Ryan, The Good, the Bad, and the Ugly, Run Lola Run, Goldfinger. Now the user may want to watch something funny and light-hearted comfort him after having seen some horrors. Knowledge Reasoning J1-Jumbo Large Pre-trained Language Model (178B Parameters) Bolded texts are generated by the model. A user watched Jaws, Saving Private Ryan, The Good, the Bad, and the Ugly, Run Lola Run, Goldfinger. Now the user may want to watch __ __ __ p(d(xt)| f([d(x1), . . . , d(xtā1)])) Item 372 Item 168 Item 413 Item 77 Item 952 p(xt |x1, . . . , xtā1) Item 1 Item 2 Item N ā¦ Recommended Item Token 1 Token 2 Token V ā¦ Token 1 Token 2 Token V ā¦ Token 1 Token 2 Token V ā¦ Item 1 Item 2 Item N ā¦ Recommended Item Predicted Token Distributions from Language Models Enable zero-shot recommendation Improve data eļ¬ciency Goal GRU4Rec Traditional Recommender System LMRecSys PLMs as Recommender System [3] āLanguage Models as Recommender Systems: Evaluations and Limitationsā, Yuhui Zhang, Hao Ding, Zeren Shui, Yifei Ma, James Zou, Anoop Deoras, Hao Wang. NeurIPS Workshop 2021 23
- 24. Generation OR Multi-Token Inference Answering the question of how to be faithful to oneās catalogue ā¢ Sequence of item ID can be mapped to a long prompt ā¢ How do we obtain ranked list of next item recommendation ? ā¢ Generation of free form text ā Need to be careful with Hallucination ā¢ Probability Assignment on available catalogue 24
- 25. A Few Open Questions Linguistic & Seq. Length Biases, Scales of LM and Creative Prompts ā¢ Multi-Token Inference: Length normalization is important. Recommendations highly sensitive to inference methods. ā¢ Linguistic Biases Disentanglement: Item names need not be fl uent English. ā¢ Scales of Language Models: Model size has signi fi cant impact on performance and latency ā¢ Prompt Engineering: Its important to design the right prompts 25
- 26. Some Results Experiments, Setup and Observations 26 ML 1M
- 27. The world after ChatGPT Unleashing the immense power of Large Language Models 27
- 28. Recent Advances in Merging LLMs with RecSys FineTuning an LLM M6-Rec[5]: P5[4]: designed a text to text fi ne-tuning paradigm based on the pre-trained T5. [4] āRecommendation as language processing (rlp): A uni fi ed pretrain, personalized prompt & predict paradigm (p5)ā, Geng Shijie et.al.. RecSys 2022 [5] āM6-Rec: Generative Pretrained Language Models are Open-Ended Recommender Systemsā, Zeyu Cui et.al.. ArXiv 2022 28
- 29. Recent Advances in Merging LLMs with RecSys Inference with LLM [6] "Zero-Shot Next-Item Recommendation using Large Pretrained Language Models." Wang, Lei, and Ee-Peng Lim. ArXiv 2023. [7] āChat-REC: Towards Interactive and Explainable LLMs-Augmented Recommender Systemā, Yunfan Gao et.al. ArXiv 2023 Zeyu Cui et.al.. ArXiv 2022 ā¢ NIR [6], Chat-REC[7] and [8] propose to directly recommend using LLMs ā Inference only. ā¢ Most e ff ort spent around āPrompt Engineeringā ā¢ Optimal encoding of user context in the prompts ā¢ āOut of Vocabularyā problems solved using techniques such as candidate pools, text-matching ā¢ Mixed success. Still a long way to go. [8] āIs ChatGPT a Good Recommender? A Preliminary Study ā, Junling Liu et.al. ArXiv 2023
- 30. Concluding Remarks ā¢ With the goal of building foundation models in RecSys, our e ff orts have been made in two directions: ā¢ Extract Knowledge from data in similar domains ā¢ Use Generic World Knowledge ā¢ We believe, the ultimate path is the hybrid of both: ZESRec + LMRecSys 30
- 31. Thank you Happy to take questions now 31