GPT4Rec Framework Overview

• GPT4Rec is a framework that leverages generative modeling and prompt-based adaptation to reframe recommendation as query generation and retrieval.
• It employs multi-query beam search to capture diverse user interests, thereby improving recall, diversity, and coverage across textual, graph, and streaming data.
• The framework integrates graph prompt tuning and reinforcement learning to enhance adaptability and interpretability in continual recommendation scenarios.

GPT4Rec denotes a family of frameworks that leverage generative modeling and prompt-based adaptation—using both autoregressive LLMs and graph neural network (GNN) backbones—within recommender systems. The frameworks aim to improve the personalization, adaptability, and interpretability of recommendations, operating in various modalities including textual, graph-structured, and streaming data. Notable variants include the original generative NLP-based GPT4Rec (Li et al., 2023), the reinforcement learning–aligned GPTRec (Petrov et al., 7 Mar 2024), and the graph prompt tuning–based GPT4Rec for streaming recommendation (Zhang et al., 12 Jun 2024).

1. Generative LLM–Based Recommendation

GPT4Rec’s core innovation is to reframe personalized recommendation as a "query generation plus retrieval" task in the language space. The key steps are:

• Query Generation: Given a user's history $H = [i_1, ..., i_T]$, each item $i_t$ is represented by its title, concatenated into a natural-language prompt $W^u$, such as “Previously the customer bought: Title₁. Title₂.… In the future, the customer wants to buy”. An autoregressive LLM (e.g., GPT-2, fine-tuned on historical purchase sequences) generates $M$ diverse search-style queries via beam search (see Section 3), modeling $P(q|H)$ as

$$P(q_1,\dots,q_L \mid H) = \prod_{t=1}^L P(q_t \mid q_{<t}, H)$$

The language modeling objective is to minimize negative log-likelihood over target item titles.

• Item Retrieval: Each generated query $q^m$ is used to retrieve top-$(K/M)$ items using a BM25 search engine over item titles, formalized as:

$$\mathrm{score}(q, d) = \sum_{w \in q} \mathrm{IDF}(w) \cdot \frac{f_{w,d}(k_1{+}1)}{f_{w,d} + k_1(1{-}b+b\,|d|/\mathrm{avgdl})}$$

The final recommendation list is constructed by merging per-query retrieved lists in a round-robin, diversity-enhancing manner (Li et al., 2023).

This approach allows both enhanced utilization of content information and direct interpretability: generated queries serve as human-readable approximations of user intent.

2. Multi-Query Beam Search and Interest Coverage

Instead of generating a single query, GPT4Rec employs multi-query beam search to address the multi-faceted nature of user interests. The beam search algorithm produces $M$ distinct high-probability queries by maintaining $B$ beams and promoting hypotheses that capture distinct semantic aspects of user history, e.g., focusing separately on subcategories or brands within the user's profile (Li et al., 2023).

Ablation studies demonstrate a monotonic increase in Recall@K, Diversity@K, and Coverage@K with larger $M$, reflecting improved relevance and coverage of diverse user interests.

Example:

• For a user with a makeup and skincare history, queries might include “hydrating face cream for dry skin” and “nude eyeshadow palette set”, reflecting disjoint interests (Li et al., 2023).

3. Shared Embedding Space and Interpretability

By fine-tuning all GPT-2 parameters on item titles and prompts, both user and item representations are mapped into a shared semantic space:

• Item $i$: $\mathrm{Encode}(Title_i) \in \mathbb{R}^d$
• User $u$: Aggregated hidden states over user prompt $W^u$

This coupling enables the generator to compose linguistically meaningful queries and facilitates semantic retrieval on new/cold-start items based solely on titles—improving adaptiveness without requiring model retraining (Li et al., 2023).

4. Graph Prompt Tuning for Streaming and Continual Recommendation

The graph-based GPT4Rec variant adapts to streaming user-item interaction graphs where edges and nodes arrive incrementally, and prior data replay is infeasible. The framework is structured as follows (Zhang et al., 12 Jun 2024):

• Graph Disentanglement: Incoming graph increment $\Delta G_t = ( \Delta A_t, \Delta X_t )$ is projected into $V$ disentangled “views” via linear projections. Each view isolates specific types of interaction patterns.
• Prompt-Based Adaptation:
  • Node-level prompts $(P^{(v)} \in \mathbb{R}^{L \times d})$: Modulate node features to accommodate attribute drift or new users/items.
  • Structure-level prompts $(Q^{(v)} \in \mathbb{R}^{K \times d})$: Guide adaptation to changes in connectivity via attention-weighted message passing.
  • View-level (cross-view) prompts $(J = [j_1,...,j_V])$: Aggregate view-specific embeddings into a final node representation using learnable, context-dependent aggregation weights.

The backbone GNN parameters are frozen; only prompt sets are updated, minimizing catastrophic forgetting and avoiding model expansion.

• Optimization: For each time segment, the Bayesian Personalized Ranking (BPR) loss is minimized on $\Delta G_t$:

$$\mathcal{L}_{\mathrm{BPR}} = \sum_{(u,i,j) \in D_t} -\ln \sigma ( \hat{y}_{u,i} - \hat{y}_{u,j} )$$

where only prompt parameters are updated.

Experiments on e-commerce, video, and POI datasets show state-of-the-art results for streaming recommendation: 1–5% absolute gains in Recall@20 and NDCG@20 over parameter-isolation or experience-replay baselines, with consistent cross-domain stability (Zhang et al., 12 Jun 2024).

5. Next-K Generation and Reinforcement Learning Alignment

In contrast to score-and-rank (Top-K) recommenders, the GPTRec/Next-K approach generates recommendation slates sequentially, modeling:

$$\pi_\Theta ( h, g^{(i)}, x ) \approx P( g_i = x \mid h, g^{(i)} )$$

at each position $i$. This autoregressive construction supports optimization for listwise, beyond-accuracy objectives (Petrov et al., 7 Mar 2024).

Two-stage alignment procedure:

1. Imitation Pre-training: GPTRec is first fit to teacher Top-K slates (e.g., BERT4Rec) via next-token likelihood and possibly knowledge distillation.
2. Reinforcement Learning (RL) Fine-tuning: The policy is further refined using Proximal Policy Optimization (PPO) on arbitrary objective functions, including accuracy (NDCG), diversity (ILD@K), and popularity-bias reduction, using custom reward decompositions.

Notably, Next-K enables learning list-level dependencies impossible with independent Top-K scoring, yielding improved trade-offs between NDCG and ILD@K or nPCOUNT (normalized popularity), as empirically demonstrated across MovieLens and Steam datasets (Petrov et al., 7 Mar 2024).

6. Experimental Results and Comparative Analysis

A summary of quantitative findings across the frameworks:

Dataset/Task	Baseline (Recall/NDCG)	GPT4Rec Variant (Best)	Relative Gain
Amazon Beauty (Recall@40)	BERT4Rec: 0.1161	GPT4Rec: 0.2040	+75.7%
Amazon Electronics (Recall@40)	BERT4Rec: 0.0751	GPT4Rec: 0.0918	+22.2%
Taobao, Netflix, Foursquare (Recall@20/NDCG@20, Streaming)	Various baselines	GPT4Rec (graph prompt tuning): +1–5 pts	Consistent gain
MovieLens-1M (NDCG@10, Diversity ILD@10)	BERT4Rec: 0.1617/0.2746	GPTRec-RL-Diversity: 0.1499/0.3621	Substantial diversity with little NDCG loss

These results reflect strong improvements in recall, diversity, and coverage, while ablation studies confirm the importance of multi-query generation, prompt types, and multi-view disentanglement (Li et al., 2023, Zhang et al., 12 Jun 2024, Petrov et al., 7 Mar 2024).

7. Limitations, Interpretability, and Future Directions

Identified limitations include reliance on rich item titles for NLP-based variants, non-differentiability of retrieval modules (e.g., BM25), and separate hyperparameter tuning requirements. The frameworks' strengths are interpretability (natural-language queries as user intent), immediate cold-start robustness (item-title-based retrieval), and continual adaptation in non-stationary environments without replay (Li et al., 2023, Zhang et al., 12 Jun 2024).

Potential future enhancements entail:

• Substituting larger LLM backbones or neural semantic retrieval layers in place of BM25.
• Multi-modal prompt conditioning (e.g., integrating image or audio features).
• Joint end-to-end optimization of generation and retrieval using reinforcement learning to maximize task-specific metrics (Li et al., 2023, Petrov et al., 7 Mar 2024).
• Further advancements in prompt architectures for finer-grained adaptation in graph-based, streaming settings (Zhang et al., 12 Jun 2024).

GPT4Rec frameworks thus broadly recast recommendation as a generative and language-oriented (or prompt-adapted) problem, merging neural text and graph learning with classical retrieval algorithms to achieve interpretable, adaptive, and performant personalization in both static and streaming environments.