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

The paper proposes Multi-TAP, a multi-criteria target-adaptive persona framework that addresses data sparsity and intra-domain heterogeneity in cross-domain recommendation by explicitly modeling semantic personas and selectively transferring relevant source-domain signals, thereby outperforming state-of-the-art methods on real-world datasets.

The Gist

Imagine you are a personal shopper for a massive department store that sells everything from high-end electronics to cozy home goods and sports gear.

The Old Problem: The "One-Size-Fits-All" Mistake
Traditionally, recommendation systems (like the algorithms on Amazon or Netflix) treat a customer as a single, static person. They assume: "If you love buying expensive laptops, you must also love buying expensive speakers."

But in real life, people are messy and complex. You might be a "budget shopper" for home audio (buying cheap, reliable speakers) but a "luxury shopper" for computers (splurging on the best tech). If the system only sees your "average" spending habits, it gets confused. It tries to transfer your love for expensive computers to your home audio purchases, and suddenly, it's recommending $5,000 speakers to someone who just bought a $50 pair of headphones. This is called data sparsity (not enough data in one area) and intra-domain heterogeneity (people act differently in different sections of the same store).

The New Solution: Multi-TAP
The paper introduces Multi-TAP, a smart new system that stops treating you as one flat profile. Instead, it realizes you have multiple "personas" (different versions of yourself) depending on what you are shopping for.

Here is how Multi-TAP works, broken down into three simple steps:

1. The "Persona" Interview (Multi-criteria Modeling)

Instead of just looking at your purchase history, Multi-TAP uses a super-smart AI (a Large Language Model) to interview you based on your history. It doesn't just ask, "What did you buy?" It asks, "Why did you buy it?"

It creates five different "personas" for you:

• The Price Saver: "I only buy budget-friendly items."
• The Quality Seeker: "I only buy top-rated, premium items."
• The Trend Follower: "I buy what everyone else is buying."
• The Explorer: "I try new categories."
• The Expert: "I stick to categories I know well."

Analogy: Think of this like a chef who doesn't just make "soup." They realize you want a spicy soup for dinner but a creamy soup for lunch. Multi-TAP separates these flavors so it doesn't serve you a spicy soup when you wanted creamy.

2. The "Doppelgänger" Bridge (Target-Adaptive Transfer)

Now, imagine you are shopping in the Electronics section (Source) and the system wants to help you in the Home section (Target).

In the old days, the system would just copy your entire "Electronics profile" and paste it onto your "Home profile." This is messy because your "Electronics Expert" persona doesn't make sense in the "Home" section.

Multi-TAP uses a clever trick called Doppelgänger Transfer:

• It creates a "Doppelgänger" (a look-alike) of your Home persona.
• It then asks: "Hey, which parts of your Electronics self are actually useful for your Home self?"
• It selectively copies only the relevant traits. Maybe your "Quality Seeker" trait from electronics is useful for home goods, but your "Tech Expert" trait is not.

Analogy: Imagine you are moving to a new city. You don't pack your entire life into a box and dump it in the new house. You look at your new apartment, figure out what you actually need (a winter coat, but not a surfboard), and only pack those specific items. Multi-TAP does this with your preferences.

3. The Final Recommendation

Once the system has built this refined, multi-faceted profile of you for the specific section you are in, it makes recommendations. Because it understands that you are a "Budget Shopper" for speakers but a "Luxury Shopper" for cameras, it gives you perfect suggestions for both.

Why is this a big deal?

The researchers tested this on real Amazon data. They found that:

1. People are inconsistent: Your spending habits change wildly depending on the category (e.g., you might spend a lot on computers but very little on clothes).
2. Old systems fail here: They try to force a single average profile, which leads to bad recommendations.
3. Multi-TAP wins: By acknowledging that you have different "hats" for different tasks, it predicts what you want up to 36% better than the current best systems.

In a nutshell: Multi-TAP stops treating you like a robot with one set of instructions. It realizes you are a complex human with different moods and needs for different parts of your life, and it uses a "Doppelgänger" to borrow the right advice from your past to help you shop better in the future.

Technical Summary

Here is a detailed technical summary of the paper "Multi-TAP: Multi-criteria Target Adaptive Persona Modeling for Cross-Domain Recommendation."

1. Problem Statement

Cross-Domain Recommendation (CDR) aims to alleviate data sparsity in a target domain by leveraging user interaction data from a data-rich source domain. However, existing CDR methods suffer from two critical limitations:

1. Coarse-Grained Representations: Most methods assume user preferences are homogeneous within a domain, aggregating all behaviors into a single domain-level embedding. This compresses context-dependent signals.
2. Intra-Domain Heterogeneity (IDH): The paper argues that user preferences are not consistent even within a single domain. Users exhibit divergent behaviors across different categories (e.g., a user might buy high-priced computers but budget-friendly home audio). Existing methods fail to capture this nuance, leading to "indiscriminate" knowledge transfer that distorts target-domain representations.

2. Methodology: Multi-TAP Framework

The authors propose Multi-TAP, a framework that explicitly models intra-domain heterogeneity through multi-criteria persona modeling and performs target-adaptive transfer. The framework consists of two main phases:

Phase 1: Multi-criteria Persona Modeling (P1)

This phase decomposes user preferences within each domain into fine-grained, criterion-specific personas rather than a single unified vector.

• Preference Criteria: The system identifies five key criteria:
  1. Price Sensitivity (PS)
  2. Quality Preference (QP)
  3. Popularity Bias (PB)
  4. Category Diversity (CD)
  5. Category Familiarity (CF)
• LLM-Assisted Construction:
  • User interaction histories and item metadata are processed into structured ordinal labels (Low/Medium/High) for each criterion.
  • A Large Language Model (LLM) (specifically GPT-4o) is used as an auxiliary module to generate natural language persona descriptions based on these structured signals. The LLM is not used for end-to-end inference but to verbalize explicit evidence, reducing hallucination risks.
• Encoding & Aggregation:
  • The generated persona descriptions are encoded into dense vectors using a text encoder (text-embedding-3-large).
  • A self-attention mechanism aggregates these heterogeneous persona embeddings into a single domain-specific user embedding ($h_u$), allowing the model to weigh different criteria dynamically based on the user and domain.

Phase 2: Target-adaptive Doppelganger Transfer (P2)

This phase handles the transfer of knowledge from the source to the target domain without forcing a direct alignment that might distort target preferences.

• Doppelganger Generation: For each overlapping user, a "doppelganger" embedding ($\tilde{h}_{ut}$) is created in the target domain. It is initialized as a copy of the target persona ($h_{ut}$) to preserve target characteristics.
• Selective Absorption: The doppelganger embedding undergoes a cross-attention operation where the target embedding acts as the query and the source persona ($h_{us}$) acts as the key/value. This allows the doppelganger to selectively absorb relevant signals from the source domain.
• Contrastive Learning: A contrastive loss (InfoNCE) is applied to align the original target persona ($h_{ut}$) with its doppelganger ($\tilde{h}_{ut}$). This ensures the target representation is enriched with transferable source knowledge while maintaining its unique identity, avoiding the "collapse" of distinct preferences.

Training Objective

The final model combines:

1. Recommendation Loss: Bayesian Personalized Ranking (BPR) on the target domain, using a hybrid representation of the persona embedding and ID-based behavioral embeddings (from LightGCN).
2. Transfer Loss: The doppelganger-based contrastive loss.
3. Joint Optimization: The total loss is a weighted sum of the recommendation and transfer objectives.

3. Key Contributions

1. Empirical Observation of IDH: The authors provide quantitative evidence (via preference preservation ratios) that user preferences shift significantly across categories within the same domain, invalidating the assumption of homogeneous domain-level representations.
2. Novel Framework (Multi-TAP): Introduction of a framework that explicitly decomposes preferences into multi-criteria personas using LLMs and employs a "doppelganger" mechanism for selective, target-adaptive transfer.
3. State-of-the-Art Performance: Extensive experiments show Multi-TAP outperforms existing CDR methods (including UniCDR, PPA, and DGCDR) significantly.

4. Experimental Results

• Datasets: Evaluated on five Amazon domains (Electronics, Home, Sports, Clothing, Toys) across six cross-domain transfer tasks.
• Performance:
  • Multi-TAP achieved the best performance in 5 out of 6 transfer tasks.
  • In the Home $\to$ Electronics task, it outperformed the best baseline (PPA) by 36.3% in HR@5 and 42.6% in NDCG@5.
  • It consistently outperformed single-domain methods and other CDR baselines.
• Ablation Studies:
  • Persona Decomposition: Multi-persona modeling with processed metadata significantly outperformed single-persona models using raw metadata.
  • Aggregation: The self-attention aggregation mechanism outperformed simple mean or concatenation strategies.
  • Transfer Mechanism: The doppelganger-based indirect transfer was more stable and effective than direct alignment strategies (Direct-ID, Direct-Persona), which often suffered from performance instability.
• Efficiency: The model is computationally efficient, training within 9–30 minutes per domain pair. The LLM persona generation is an offline process costing approximately $0.022 per user.

5. Significance

This paper fundamentally shifts the paradigm of Cross-Domain Recommendation by challenging the assumption that user preferences are static within a domain. By acknowledging and modeling intra-domain heterogeneity, Multi-TAP prevents the loss of fine-grained preference signals during the transfer process. The use of LLMs for structured persona generation and the doppelganger mechanism for selective transfer offers a robust solution to the "negative transfer" problem, where irrelevant source knowledge degrades target performance. This approach provides a more interpretable and controllable way to leverage cross-domain data, making it highly relevant for large-scale e-commerce platforms with diverse product categories.