Mitigating Negative Transfer in Multi‑Target Cross‑Domain Recommendation with Unbiased Information Extraction and Adaptation (UIEA)

Background

Modern recommender systems suffer from severe data imbalance: popular domains (e.g., blockbuster movies) contain abundant user‑item interactions, while niche domains (e.g., obscure electronics) are extremely sparse, making it difficult for models trained solely on a single domain to learn stable user preferences. Cross‑Domain Recommendation (CDR) aims to transfer knowledge from data‑rich domains to data‑scarce target domains to alleviate cold‑start and sparsity issues. In multi‑target CDR, two key limitations exist: (1) most methods simply average domain losses during representation learning, implicitly assuming all domains are equally important and share the same distribution, which lets data‑rich domains dominate optimization and cause negative transfer to sparse domains; (2) existing adaptation stages rely only on user‑side information and ignore item‑side representations that capture domain‑specific preferences.

UIEA: Unbiased Information Extraction and Adaptation

To overcome these problems, we propose UIEA, a novel multi‑target CDR framework that explicitly considers inter‑domain differences. UIEA consists of three stages: a pre‑training stage, an unbiased information extraction stage, and an unbiased information adaptation stage. The method extracts unbiased cross‑domain information and performs thorough target‑domain adaptation, reducing negative transfer while leveraging rich‑domain data.

Problem Definition

In multi‑target CDR, each domain contains a distinct item set and shares a user set (or, in multi‑country scenarios, each domain has its own user set but shares items). Let denote the shared user set and the item set of domain . The interaction matrix in domain records whether a user has interacted with an item. Given multiple interaction matrices , the goal is to simultaneously improve recommendation performance on all domains.

Pre‑training Stage

For each domain , we apply Bayesian Matrix Factorization to decompose its interaction matrix into a user representation matrix and an item representation matrix , where is the latent dimension. The inner product of user and item representations yields a preference score. The stage minimizes the following BPR pre‑training loss (image):

. Here denotes the set of items a user has interacted with in domain , and is a regularization term. Minimizing this loss yields domain‑specific user representations that capture each domain’s preferences.

Unbiased Information Extraction Stage

We employ an auto‑encoder as a representation generator. It takes all domain‑specific user representations as input and produces a global representation that captures domain‑invariant information. Inspired by Distributionally Robust Optimization (DRO), we focus on the worst‑case domain during joint training, updating both model parameters and domain weights.

The auto‑encoder maps concatenated domain representations to a latent space and reconstructs the input as follows (images):

. The extraction loss comprises a reconstruction loss and a fine‑tuning loss. The reconstruction loss measures the discrepancy between local user representations and reconstructed representations (image):

. The fine‑tuning loss evaluates the performance of the global user representation within each domain (image):

.

Traditional average‑loss objectives assume equal importance across domains, ignoring significant inter‑domain differences and potentially injecting biased information from data‑rich domains. To remedy this, we formulate a min‑max optimization that optimizes model parameters on the domain with the highest extraction loss (image):

. Domain weights are updated jointly with model parameters using stochastic gradients (images):

. The update rules involve a learning rate and Bregman divergences and to measure similarity between vectors (image):

. After rounds, we obtain final model parameters and domain weights. For each user, a decoder with parameters generates the final global representation.

Unbiased Information Adaptation Stage

Directly applying the global representation to a target domain is insufficient because it only captures domain‑invariant information. We therefore design a novel user‑item attention module that leverages domain‑specific item information from historical interactions.

For each user, we first average the item representations of all historical interactions in the target domain to obtain an item representation (image):

. We then combine the global user representation and the local user representation as inputs to the attention mechanism (image):

. The final user representation is computed via a scaled‑dot‑product attention (image):

. For any item in the target domain, we calculate a preference score . The attention module is trained by fixing local user and item representations and minimizing the BPR loss on the target domain.

Experimental Validation

We evaluate UIEA on three real‑world datasets—Amazon, Douban, and IE—against six baselines. UIEA consistently outperforms competitors, especially in sparse domains (e.g., books and electronics in Amazon). We also report average and worst‑case performance, showing that while methods such as EDDA/CAT‑ART improve over traditional single‑target CDR approaches, they still lag behind UIEA under the hardest conditions.

To assess the contribution of each module, we conduct an ablation study on the Amazon dataset with the following variants (results shown in the figure): Rec + MLP (uses only reconstruction loss and an MLP for adaptation), AVG + MLP (optimizes average loss across domains), UIE + MLP (uses UIE for extraction), and AVG + UIA (optimizes average loss and uses UIA for adaptation). The full UIEA model achieves the best performance.

Conclusion

We introduced UIEA, a novel multi‑target cross‑domain recommendation framework that mitigates negative transfer by extracting unbiased cross‑domain information and adapting it with a user‑item attention module. Extensive experiments on three large‑scale datasets demonstrate that UIEA achieves significant improvements over strong baselines.