Multi-TAP: Multi-criteria Target Adaptive Persona Modeling for Cross-Domain Recommendation

📝 Paper Summary

Cross-Domain Recommendation (CDR) User Modeling

Multi-TAP improves cross-domain recommendation by decomposing user preferences into multi-criteria personas to capture intra-domain heterogeneity and using a target-adaptive doppelganger mechanism for selective knowledge transfer.

Core Problem

Existing CDR methods assume user preferences are homogeneous within domains, aggregating diverse behaviors into single representations that compress context-dependent signals.

Why it matters:

Compressed representations mask fine-grained preference shifts (e.g., price sensitivity changes across categories), leading to information loss before transfer even begins
Indiscriminate transfer of these coarse representations exacerbates the loss, failing to preserve domain-specific preference structures in the target domain
Behavioral science (Theory of Planned Behavior) suggests observed behaviors arise from context-dependent intentions, which standard monolithic embeddings fail to capture

Concrete Example: In the Electronics domain, a user might buy a high-priced 'Computer' but a budget 'Home Audio' cable. Standard methods average these into a 'medium price' preference, losing the nuance. Multi-TAP identifies distinct price sensitivities per category.

Key Novelty

Multi-criteria Target-Adaptive Persona (Multi-TAP)

explicitly decomposes user behavior into five distinct criteria (e.g., Price Sensitivity, Popularity Bias) using an LLM to generate granular textual personas rather than a single holistic profile
Uses a 'doppelganger' mechanism where a target-domain persona creates a proxy in the source domain; this proxy selectively absorbs only the source information relevant to the target context via attention
Aligns the target persona with its source-derived doppelganger using contrastive learning, avoiding the rigidity of direct source-target embedding alignment

Architecture

The overall architecture of Multi-TAP, divided into (a) Multi-criteria Persona Modeling and (b) Target-adaptive Doppelganger Transfer.

Evaluation Highlights

Outperforms five state-of-the-art CDR methods by up to 36.3% in HR@5 (Hit Rate) on real-world datasets
Empirical analysis reveals 'preservation ratio' of preferences across categories drops to ~31% in some pairs (e.g., Car vs. GPS), quantitatively proving the existence of intra-domain heterogeneity

Breakthrough Assessment

7/10

Identifies a fundamental flaw in CDR (homogeneity assumption) and proposes a coherent LLM-based solution. Strong quantitative motivation, though full experimental tables are not in the provided text.

⚙️ Technical Details

Problem Definition

Setting: Cross-domain recommendation with a source domain and target domain, sharing a subset of overlapping users

Inputs: User-item interaction logs, item metadata (category, price, ratings) from both source and target domains

Outputs: Recommended items for users in the target domain

Pipeline Flow

Data Preprocessing (Discretize metadata into bins)
Persona Generation (LLM creates text descriptions)
Persona Encoding (LLM embeds text to vectors)
User Aggregation (Self-attention combines personas)
Doppelganger Transfer (Cross-attention source -> target proxy)
Recommendation (LightGCN + Persona Fusion)

System Modules

Persona Database Builder

Organize interactions into structured bins for 5 criteria: Price Sensitivity, Quality Preference, Popularity Bias, Category Diversity, Category Familiarity

Model or implementation: Statistical Discretization (Tertiles)

Persona Generator (Persona Modeling)

Generate natural language descriptions for each preference criterion based on structured evidence

Model or implementation: gpt-4o

Persona Encoder (Persona Modeling)

Encode textual personas into dense vectors

Model or implementation: text-embedding-3-large

Doppelganger Transfer

Selectively transfer source knowledge by attending to source personas using a target-initialized query

Model or implementation: Cross-Attention Mechanism

Novel Architectural Elements

Doppelganger Transfer Mechanism: Instead of direct alignment, creates a 'doppelganger' proxy that selectively aggregates source info before contrasting with the target
Multi-criteria Persona Decomposition: Explicitly modeling 5 distinct dimensions of preference (Price, Quality, Popularity, Diversity, Familiarity) rather than one user vector

Modeling

Base Model: LightGCN (for behavioral backbone) + gpt-4o (for persona generation)

Training Method: Joint optimization of Recommendation Loss and Transfer Loss

Objective Functions:

Purpose: Optimize ranking accuracy in the target domain.

Formally: L_rec = BPR_Loss(observed, unobserved)
Purpose: Transfer knowledge by aligning target persona with its source-derived doppelganger.

Formally: L_cl = -log(exp(sim(h_t, h_doppel)) / sum(exp(sim(h_t, h_neg))))

Training Data:

Source domain interaction graph
Target domain interaction graph
Overlapping users set

Compute: Not reported in the paper

Comparison to Prior Work

vs. EMCDR/UniCDR: These methods transfer compressed domain-level representations; Multi-TAP transfers fine-grained criterion-specific personas.
vs. General CDR: Most assume preference homogeneity (consistency); Multi-TAP explicitly models heterogeneity (divergence across categories).

Limitations

Relies on closed-source commercial LLMs (gpt-4o, text-embedding-3-large) for persona generation and encoding
Computationally expensive due to API calls for every user to generate text personas
Requires rich metadata (price, ratings) to build the initial persona database; may not work on opaque ID-only datasets

Reproducibility

Code: https://github.com/archivehee/Multi-TAP

Code is publicly available at https://github.com/archivehee/Multi-TAP. The paper specifies the exact LLMs used (gpt-4o, text-embedding-3-large) but relies on OpenAI APIs, which are closed-source dependencies.

📊 Experiments & Results

Evaluation Setup

Cross-domain recommendation on real-world e-commerce datasets (likely Amazon, based on context)

Benchmarks:

Amazon Datasets (implied) (Top-k Recommendation)

Metrics:

HR@5 (Hit Rate at 5)
NDCG@5 (Normalized Discounted Cumulative Gain at 5)
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
Empirical analysis results quantify the 'Intra-Domain Heterogeneity' problem, showing that user preferences are not consistent across categories.
Electronics Domain (Category Pair: Car -> GPS)	Preservation Ratio (Price)	100	31	-69

Experiment Figures

Heatmaps/Matrices showing the 'Conditional Preservation Ratio' of user preferences between different categories in Electronics and Sports domains.

Main Takeaways

Multi-TAP consistently outperforms state-of-the-art CDR methods, achieving up to 36.3% improvement in HR@5.
Intra-domain heterogeneity is a significant phenomenon; users frequently shift preferences (e.g., price sensitivity) across categories, invalidating the homogeneity assumption of prior CDR works.
Explicitly modeling multiple preference criteria (Price, Quality, Popularity, etc.) is more effective than learning a single latent user vector for cross-domain transfer.

📚 Prerequisite Knowledge

Prerequisites

Cross-Domain Recommendation (CDR) concepts
Basics of Large Language Models (LLMs) for text embedding
Contrastive Learning (InfoNCE loss)
Collaborative Filtering (LightGCN)

Key Terms

CDR: Cross-Domain Recommendation—leveraging data from a rich source domain to improve recommendations in a sparse target domain

Intra-domain heterogeneity (IDH): The phenomenon where a user exhibits different preference tendencies (e.g., price sensitivity) across different categories within the same domain

Doppelganger embedding: A constructed intermediate representation in the source domain initialized from the target domain, used to selectively absorb transferable signals

LightGCN: A graph convolutional network that learns user/item embeddings by propagating signals over the interaction graph, simplified by removing non-linearities

BPR: Bayesian Personalized Ranking—a loss function that optimizes the ranking order by maximizing the difference between positive (observed) and negative (unobserved) items

InfoNCE: A contrastive loss function used to maximize mutual information between related samples (positive pairs) while pushing away unrelated samples (negatives)