Social Knowledge for Cross-Domain User Preference Modeling

📝 Paper Summary

User modeling Cross-domain recommendation

Users can be modeled as a list of popular social media accounts they follow, allowing effective cross-domain recommendation by projecting them into a pre-trained social embedding space.

Core Problem

Traditional recommender systems struggle with 'cold-start' scenarios and cross-domain generalization because user feedback (ratings) is typically sparse and confined to single domains (e.g., only movies).

Why it matters:

Users have correlated preferences across domains (e.g., liking action movies and sports cars), but current systems fail to bridge these gaps without massive overlapping data.
New users often lack sufficient history for personalization, creating a need for lightweight, effective user modeling that works immediately.
Socio-demographic traits encoded in social behavior are powerful predictors of preference but are underutilized in cross-domain transfer.

Concrete Example: A user who likes action movies might also prefer sports cars. Traditional systems cannot predict this link without explicit feedback in both domains. This approach uses the user's Twitter follow list (e.g., following 'Fast & Furious' and 'Ferrari') to project them into a shared space where these entities are naturally close.

Key Novelty

Inductive Social User Modeling via Entity Co-following

Represents users as a 'bag of entities' (popular accounts they follow), projecting them into a pre-trained social embedding space derived from global Twitter co-following patterns.
Enables zero-shot cross-domain recommendation by measuring cosine similarity between the user's vector and candidate items in completely different domains (e.g., inferring music taste from news consumption).
Demonstrates that LLMs (Large Language Models) can perform personalized ranking simply by being prompted with a short list of entities the user likes.

Evaluation Highlights

+22% improvement in Mean Average Precision (MAP) over strong popularity baselines for cross-domain link prediction.
Effective personalization achieved with as few as 10 entities per user, showing rapid convergence of user preference signals.
+13% MAP improvement when prompting GPT-4o with just 12 example entities compared to a generic one-fit-all ranking.

Breakthrough Assessment

7/10

Offers a novel, lightweight inductive approach to user modeling that bridges domains effectively. While the method is straightforward, the validation on cross-domain tasks and LLM integration is significant.

⚙️ Technical Details

Problem Definition

Setting: Cross-domain link prediction: Rank candidate entities in a target domain based on a user's known entities in source domains.

Inputs: A user u represented by a set of entities {e_i} they follow (from source domains).

Outputs: Ranked list of candidate entities in a target domain (e.g., Music, Sports).

Pipeline Flow

Entity Aggregation
Projection
Similarity Scoring
Ranking

System Modules

Entity Aggregator (User Representation)

Collects pre-trained embeddings for the set of entities a user follows

Model or implementation: SocialVec (Lookup Table)

User Projector (User Representation)

Computes a single vector representation for the user

Model or implementation: Mean Pooling

Scorer

Calculates relevance of candidate entities in target domain

Model or implementation: Cosine Similarity

Novel Architectural Elements

Projection of users into pre-trained social entity space via averaging (inductive representation)
Zero-shot cross-domain ranking using entity-user cosine similarity without domain alignment training

Modeling

Base Model: SocialVec (pre-trained Word2Vec-style embeddings of Twitter accounts)

Comparison to Prior Work

vs. TwHIN: Inductive averaging of public pre-trained embeddings vs. transductive learning on proprietary production graph
vs. CF: Cross-domain zero-shot capability vs. single-domain interaction dependence
vs. Standard User Embeddings: Learned from global entity co-occurrence patterns rather than local graph structure alone

Limitations

Relies on users having a Twitter (X) account with public follow information
Biases in social media usage (demographics, political) may be encoded in embeddings
Static embeddings may not capture rapidly shifting trends or new entities without retraining
Entity mapping requires external knowledge bases (Wikidata/DBpedia) which may be incomplete

📊 Experiments & Results

Evaluation Setup

Link prediction task ranking held-out entities in a target domain based on known entities in source domains.

Benchmarks:

Social User Dataset (Cross-domain Entity Ranking) [New]

Metrics:

Mean Average Precision (MAP)
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
Social User Dataset	MAP	Not reported in the paper	Not reported in the paper	Not reported in the paper
GPT-4o Ranking	MAP	Not reported in the paper	Not reported in the paper	Not reported in the paper

Main Takeaways

Social user modeling significantly outperforms popularity baselines for cross-domain recommendations (+22% MAP).
Performance converges quickly; only ~10 entities are needed to build a useful user profile.
LLMs like GPT-4o can leverage this 'bag of entities' representation directly in prompts to personalize rankings effectively.
Socio-demographic traits (age, gender, politics) inferred from embeddings correlate strongly with domain preferences.

📚 Prerequisite Knowledge

Prerequisites

Word2Vec (Skip-gram model)
Collaborative Filtering
Cosine Similarity
Latent Space Embedding

Key Terms

_comment: REQUIRED: Define ALL technical terms, acronyms, and method names used ANYWHERE in the entire summary. After drafting the summary, perform a MANDATORY POST-DRAFT SCAN: check every section individually (Core.one_sentence_thesis, evaluation_highlights, core_problem, Technical_details, Experiments.key_results notes, Figures descriptions and key_insights). HIGH-VISIBILITY RULE: Terms appearing in one_sentence_thesis, evaluation_highlights, or figure key_insights MUST be defined—these are the first things readers see. COMMONLY MISSED: PPO, DPO, MARL, dense retrieval, silver labels, cosine schedule, clipped surrogate objective, Top-k, greedy decoding, beam search, logit, ViT, CLIP, Pareto improvement, BLEU, ROUGE, perplexity, attention heads, parameter sharing, warm start, convex combination, sawtooth profile, length-normalized attention ratio, NTP. If in doubt, define it.

SocialVec: Pre-trained embeddings of Twitter accounts (entities) where proximity indicates shared user followership

MAP: Mean Average Precision—a rank-aware metric that evaluates the quality of the entire ordered list of recommendations

Inductive: The ability to generate representations for new, unseen users without retraining the entire model

Transductive: Learning representations only for users present during the training phase; cannot handle new users directly

Cold-start: The problem of recommending items to new users who have practically no interaction history

Skip-gram: A neural network architecture that predicts context words (or entities) given a focus word (or entity), used here to learn embeddings

Homophily: The tendency of individuals to associate and bond with similar others

LLM: Large Language Model—AI models trained on vast text data capable of generating human-like text