PaperlayerView exact revision on arXiv

Source-linked AI summary

Pre-train, Prompt and Recommendation: A Comprehensive Survey of Language Modelling Paradigm Adaptations in Recommender Systems

Peng Liu, Lemei Zhang, Jon Atle Gulla

arXiv:2302.03735v3cs.IR

TL;DR

Sparse user-item interactions challenge deep recommendation models, and prior reviews have not systematically linked recommendation data types with language-modelling paradigms. This paper surveys and categorizes PLM-based recommendation strategies and their associated input types and objectives. Its evaluation discussion reports that cache management can accelerate training and yield the most improvements over baselines.

  • Problem

    Deep recommendation models can overfit and generalize poorly when user-item interactions are insufficient, while prior reviews did not systematically analyze how data types relate to training paradigms.

    Source: §1 · pp. 1–2

  • Method

    The paper systematically surveys PLM-based recommendation by training strategy and learning objective, and analyzes their connections to recommender-system input data types.

    Source: §1 · p. 2

  • Results

    The evaluation discussion reports that cache management can accelerate model training and yield the most improvements over baselines.

    Source: §7.2 · pp. 10–11

  • Takeaways & Limitations

    The survey provides an integrated account of recommendation training paradigms and input data types, alongside future research directions.

    Source: §1 · p. 2

    Factual consistency in generated recommendation sentences remains an urgent research problem that has not received sufficient attention.

    Source: §8 · p. 11

Abstract

from arXiv · show

The emergence of Pre-trained Language Models (PLMs) has achieved tremendous success in the field of Natural Language Processing (NLP) by learning universal representations on large corpora in a self-supervised manner. The pre-trained models and the learned representations can be beneficial to a series of downstream NLP tasks. This training paradigm has recently been adapted to the recommendation domain and is considered a promising approach by both academia and industry. In this paper, we systematically investigate how to extract and transfer knowledge from pre-trained models learned by different PLM-related training paradigms to improve recommendation performance from various perspectives, such as generality, sparsity, efficiency and effectiveness. Specifically, we propose a comprehensive taxonomy to divide existing PLM-based recommender systems w.r.t. their training strategies and objectives. Then, we analyze and summarize the connection between PLM-based training paradigms and different input data types for recommender systems. Finally, we elaborate on open issues and future research directions in this vibrant field.

1 Introduction

Deep recommendation models face a data-sparsity bottleneck, while prior reviews leave important aspects of language-modelling paradigms for recommendation insufficiently covered. This paper surveys these paradigms and their relationship to recommender-system data types.

1 Introduction

  • Deep recommendation models are data-hungry and can overfit or generalize poorly when user-item interactions are insufficient, making sparsity a major performance bottleneck.The text notes that this occurs, for example, when many new users have few interactions.Source: §1 · p. 1
  • Earlier recommendation reviews covered limited BERT-like work or purely self-supervised settings, and did not systematically analyze links between data types and training paradigms.The paper describes its survey as the first up-to-date, comprehensive review of Language Modelling Paradigm Adaptations for Recommender Systems (LMRS).Source: §1 · p. 2
  • The survey organizes PLM-based recommendation by training strategy and learning objective, covering pre-training and fine-tuning as well as prompting.It explains subcategories, concepts, formulations, methods, and recommendation training and inference processes.Source: §1 · p. 2
  • The paper also discusses limitations and future research directions and shares integrated resources for beginners and practitioners.The contribution is framed as helping readers learn more effectively about the field.Source: §1 · p. 2

2 Generic Architecture of LMRS

LMRS processes recommendation inputs through preprocessing, a pre-train/fine-tune or pre-train/prompt flow, and inference for recommendation tasks.

2 Generic Architecture of LMRS

  • LMRS takes preprocessed inputs into either a pre-train/fine-tune or a pre-train/prompt flow before using the trained model for recommendation inference.The described pipeline covers data input, pre-training, fine-tuning or prompting, and inference.Source: §2 · p. 2
  • Input objects can be relevant to both training and inference, and preprocessing can convert them into graphs, ordered sequences, or aligned text-image pairs.Source: §2 · p. 2
  • When inference relies solely on the pretrained model, the approach is end-to-end and leverages LM-based learning objectives.Source: §2 · p. 2

3 Data Types

LMRS uses diverse recommendation inputs, including text, sequences, graphs, and multimodal content, with encoding adapted to the input type.

3 Data Types

  • Recommendation inputs include textual data, chronologically or otherwise ordered interaction sequences, graphs, and multimodal content such as images, videos, and audio.Item metadata and user profiles are treated as textual data in the paper.Source: §3 · pp. 2–3
  • Different input types carry different information: sequences represent ordered interactions, while graphs provide structural knowledge such as social or knowledge-graph relations.Source: §3 · p. 3
  • Multimodal data can improve recommendation, but its use in LMRS papers is scarce, possibly because accessible datasets are lacking.Some researchers have collected datasets for tri-modal music recommendation and shopping benchmarks.Source: §3 · p. 3

4 Training Strategies of LMRS

LMRS training strategies fall into pre-train/fine-tune and prompting paradigms, with subcategories reflecting which model components are trained and how recommendation inputs are used.

4 Training Strategies of LMRS

  • LMRS training strategies are grouped into pre-train/fine-tune and prompt-learning classes, each further divided by which parts of the recommendation model are trained.The survey also presents publication statistics by strategy and year, alongside representative methods.Source: §4 · p. 3, Figure 2 · p. 4

4.1 Pre-train, fine-tune paradigm for RS

  • Pre-training can improve initialization and generalization, provide transferable knowledge, and regularize training on low-resource datasets.The survey describes pre-train, fine-tune as a common strategy and cites BERT4Rec and Transformers4Rec as examples.Source: §4.1 · p. 3
  • Whole-model fine-tuning can outperform partial tuning, while partial tuning offers a balance between training overhead and recommendation performance.Hou et al. found that tuning a small proportion of parameters could quickly adapt a model to unseen domains with cold-start or new items.Source: §4.1 · p. 4
  • Some methods fine-tune task-specific layers or initialize a new recommendation model with parameters learned during pre-training.Examples include medication recommendation and item-embedding pre-training followed by fine-tuning with pairwise rank loss.Source: §4.1 · p. 4

4.2 Prompting paradigm for RSs

  • Prompting reformulates downstream recommendations using hard or soft prompts and can sometimes use the pretrained model directly without fine-tuning.Prompt tuning requires choosing templates and verbalizers, which can substantially affect recommendation performance.Source: §4.2 · p. 5
  • Tuning-free prompting uses input prompts without changing pretrained-model parameters, and has been reported effective for new users or items in same-domain and cross-domain settings.The described unified approach uses discrete prompts at inference to produce recommendations and related subtasks without a fine-tuning stage.Source: §4.2 · p. 5
  • Prompt+PTM tuning optimizes both prompt-related and model parameters, using prompts as additional bootstrapping for task-specific training.The approach can also tune prompts and model parameters in stages or learn soft prompts from reward-state pairs.Source: §4.2 · p. 6

5 Learning Objectives of LMRS

This section surveys language-model objectives used in recommendation and how they are adapted to different data types and recommendation tasks. It covers autoregressive, reconstruction, and auxiliary objectives, with adaptations spanning sequences, graphs, and multimodal data.

5.1 Language modelling objectives to recommendation

  • Self-supervised language-modelling objectives for recommendation include autoregressive, reconstruction, and auxiliary categories, reducing reliance on manually annotated labels.The survey frames these objectives as probabilistic density-estimation problems.Source: §5.1 · p. 6
  • Autoregressive objectives use left-to-right context, while partial autoregressive modelling permits span-based factorization; UniLMv2 initializes a news-embedding model.The survey describes GPT-2 and DialoGPT as common backbones for explainable and conversational recommendation, respectively.Source: §5.1 · p. 6
  • MLM predicts masked tokens from surrounding context, while NSP and SOP model segment relationships; these objectives are used for news, conversational, and code recommendation.BERT-based MLM is used to capture user-news matching signals, and NSP/SOP adaptations include conversational recommendation and StackOverflow tag recommendation.Source: §5.1 · pp. 6–7
  • RTD achieved the best performance among MLM and AM for session-based recommendation, possibly because it models the full interaction sequence bidirectionally.The passage presents the bidirectional-context explanation as a probability, not a demonstrated cause.Source: §5.1 · p. 8

5.2 Adaptive objectives to recommendation

  • Recommendation objectives adapt to sequential data: clicked histories support next-click prediction, masked historical records support item prediction, and graph paths support end-node generation.Zheng et al. (2022) and Xiao et al. (2022) model clicked news histories autoregressively; Yuan et al. (2020b) mask historical records, and Geng et al. (2022b) model graph paths.Source: §5.2 · p. 8
  • MLM adaptations include masked-node and masked-edge prediction on heterogeneous graphs and masked multimodal modelling of news text and image regions.NBP adapts NSP/SOP to CTR prediction by inferring whether a candidate is a user's next behavior.Source: §5.2 · p. 9

6 Formulating Training with Data Types

The section connects training strategies and objectives with recommender-system data types, using representative research and public datasets to organize the landscape. It observes that data can be reformulated and combined, while prompting remains less explored beyond text and sequential inputs.

6 Formulating Training with Data Types

  • Table 1 links representative training strategies and objectives to input data types, while the survey’s resource list points to further research.The authors select typical strategies and objectives; the table includes only part of recent work because of page limits.Source: §6 · p. 9
  • Public datasets are catalogued by popularity and data diversity alongside their recommendation tasks, training strategies, and data types.The authors direct readers to shared online resources for broader coverage.Source: §6 · p. 9
  • Datasets can be transformed into different data types, and integrating data types can better serve different recommendation goals.The survey also notes that strategies can transfer across tasks using parameter tuning, added components, or prompts; prompting is mainly explored with text and sequential data, leaving multimodal and graph inputs as future directions.Source: §6 · p. 9

7 Evaluation

The evaluation discussion spans recommendation utility, generated-text quality, and comparisons across datasets. It reports differing model outcomes by benchmark and cautions that automatic metrics may not match human feedback.

7.1 Evaluation metrics

  • Recommendation utility is measured offline with metrics including RMSE, MAP, AUC, Recall, Precision, MRR, NDCG, and HitRate; group recommendation also uses Group AUC or User Group AUC, and online tests use Conversion rate or CTR.The passage also lists MAE, F1-score, and other familiar metrics.Source: §7.1 · pp. 9–10
  • Generated-text evaluation uses overlap, fluency, diversity, informativeness, factuality, and alignment measures, while conversational recommendation can be assessed with Item Ratio.Examples include BLEU and ROUGE, Perplexity, Distinct-n, USR, Entailment Ratio, MAUVE, Feature Diversity, and CLIPScore; end-to-end evaluation checks whether recommended items are inserted appropriately into dialogue.Source: §7.1 · p. 10
  • Human and automatic evaluations can yield opposite results, limiting automatic metrics for generated explanations and dialogues.Figure 3 reports usage frequencies of metrics across their respective tasks.Source: §7.1 · p. 10, Figure 3 · p. 10

7.2 Discussion on evaluation across datasets

  • On ReDial, fixed-prompt PTM tuning shows the largest improvements among the four conversational recommenders assessed.Table 3 presents the ReDial benchmark comparison.Source: §7.2 · p. 10, Table 3 · p. 11
  • On Amazon, ChatGPT zero- and few-shot recommendation underperform supervised baselines, while P5 shows the most improvement on Amazon and Yelp.Tables 4 and 5 present the Amazon Beauty and Yelp comparisons.Source: §7.2 · p. 10, Table 4 · p. 11, Table 5 · p. 11
  • On MIND, Xiao et al.’s cache-management framework accelerates training and yields the most improvement over baselines.Table 6 presents the MIND benchmark comparison.Source: §7.2 · p. 11, Table 6 · p. 11

8 Discussion and Future Directions

The discussion identifies open challenges in generated language, knowledge transfer, model scalability, and balancing multiple objectives. It also flags limited exploration of newer PLMs, instruction tuning, and privacy risks.

8 Discussion and Future Directions

  • Generated recommendation explanations and replies may favor generic or repeated safe language, leaving diversity, pertinence, and factual consistency as unresolved challenges.The survey calls for maintaining fluency while improving diversity and pertinence, and says factual consistency has not received sufficient attention.Source: §8 · p. 11
  • How much domain knowledge pretrained models contain and how to transfer it effectively remain open questions; a simple domain-adaptive pretraining approach yielded limited improvements.The survey also identifies cross-task transfer, measuring transferred knowledge, and possible upper bounds as issues for further study.Source: §8 · pp. 11–12
  • Scaling recommendation pretraining requires maintaining update efficiency and accuracy, but partial fine-tuning can perform poorly or yield unstable gains depending on the model and task.Proposed directions include student-model compression and faster fine-tuning with reduced GPU memory use.Source: §8 · p. 12

Balancing multiple objectives in pre-training

  • Multi-task recommendation objectives require trade-offs, and sequentially solving tasks can cause error propagation and declining performance.Suggested remedies for error propagation still need further verification.Source: Balancing multiple objectives in… · p. 12
  • Studies of ChatGPT-based recommendation remain initial explorations, with broader testing across tasks, PLMs, domains, and prompt designs still needed.The survey also identifies instruction tuning as an unexplored direction in LMRS studies.Source: Balancing multiple objectives in… · p. 12
  • Pretrained models can infer user profiles from learned representations, while web-crawled pretraining data may expose sensitive information, making privacy protection an open issue.The examples of inferred profile attributes include gender, age, and marital status.Source: Balancing multiple objectives in… · p. 12
Loading 2302.03735v3…