Profile-Aware KG Aggregation

• Profile-aware KG aggregation is a method that integrates user and entity profiles to tailor the aggregation of knowledge graph features, addressing data heterogeneity.
• It employs techniques such as decision trees, attention mechanisms, and reinforcement learning to dynamically select and weight KG features based on semantic relevance.
• Empirical results from models like KGUF, SPiKE, and DPAO demonstrate significant gains in recommendation accuracy and improved filtering of irrelevant information.

Profile-aware KG aggregation refers to a family of methodologies that adapt the aggregation of knowledge graph (KG) features according to user-specific or entity-specific profiles. These methods address the heterogeneity and granularity of user/item preferences in recommender systems and other graph-based tasks, leveraging both high-order graph connectivity and explicit profile information to optimize representations. Techniques vary in their construction of profiles (from semantic features and attributes to LLM-compressed rationales) and in their integration into aggregation, but all share the goal of tailoring KG information flow to subject-level interests or behavioral patterns.

1. Foundational Concepts and Objectives

Profile-aware KG aggregation is designed to overcome the limitations of coarse-grained knowledge integration in GNN-based recommender systems. Conventional graph collaborative filtering (GCF) models propagate signals uniformly across graph neighborhoods, failing to differentiate between informative and irrelevant KG features from the perspective of individual users or items. Recent approaches utilize explicit or latent profiles—ranging from pruned semantic feature sets, attention-weighted attributes, reinforcement-learned receptive fields, to LLM-generated semantic vectors—to selectively aggregate only those elements of the KG pertinent to the target entity's behavioral or semantic signature.

This paradigm is instantiated in models such as KGUF (Bufi et al., 2024), DPAO (Jung et al., 2023), AKGAN (Huai et al., 2021), and SPiKE (Ahn et al., 13 Jan 2026). Each method formalizes profile-awareness in aggregation via distinct mechanisms, balancing interpretability, scalability, and recommendation accuracy.

2. Mechanisms for Profile-aware Feature Selection

A central step in profile-aware aggregation is the identification and selection of KG features that align with user or entity profiles.

KGUF (Bufi et al., 2024) employs user-level decision trees (DTs) that analyze the semantic feature sets of positively and negatively interacted items. For each item $i$, its KG-derived feature set $$\mathcal{F}_i = \{\langle \rho, \omega \rangle \mid (i \xrightarrow{\rho} \omega) \in \mathcal{KG}\}$$ is pruned such that only features with positive information gain (i.e., those used as splits in any user's tree) are retained: $$\mathcal{F}_i^* = \mathcal{F}_i \cap (\bigcup_{u \in \mathcal{U}} \mathcal{F}_u^{T_u})$$. This ensures that only features predictive of user-level positive feedback are retained for downstream aggregation.

AKGAN (Huai et al., 2021) decomposes KG attributes into relation-specific coordinate blocks and utilizes an interest-aware attention mechanism. User embeddings are decomposed, and attention scores $f_{\mathrm{att}}(u, r_m)$ modulate attribute pools, reflecting how strongly a given user aligns with each attribute, unconstrained by a simplex normalization. This structure maintains the semantic independence of relation blocks.

SPiKE (Ahn et al., 13 Jan 2026) generates a semantic profile $p_e$ for every KG entity via LLM prompting, encodes these profiles, and injects them additively into the entity's KG embedding space via a learned projection $M(p_e)$ with scaling parameter $\lambda_p$.

DPAO (Jung et al., 2023), by contrast, does not explicitly select KG features but adaptively determines the aggregation depth (number of hops/layers) per node via dual policy Deep-Q Networks, reflecting a learned notion of how wide each profile's receptive field should be.

3. Profile-aware Aggregation and Embedding Propagation

After profile-based selection or weighting, aggregation proceeds via various graph neural network mechanisms, with modifications introduced by profile-awareness:

• KGUF: For item $i$, the propagation is controlled by a content-collaborative mixture (Eq. 6):

$$e_i^{(l)} = \alpha \cdot \frac{1}{|\mathcal{F}_i^*|} \sum_{f \in \mathcal{F}_i^*} e_f + (1-\alpha) \sum_{u \in N(i)} \frac{1}{\sqrt{|N(u)||N(i)|}} e_u^{(l-1)}$$

where $\alpha$ tunes the influence of filtered KG content vs. collaborative user signals.

• AKGAN: Profile-aware user embeddings combine static blocks and attribute-pooled blocks modulated by attention:

$$\mathbf{e}_u^* = \mathbf{e}_u \;\Vert\; [f_{\mathrm{att}}(u, r_m) \cdot \frac{1}{|\mathcal{N}_u|}\sum_{i \in \mathcal{N}_u} \mathbf{e}_i^{r_m}]_{m=1}^M$$

• SPiKE: KG embeddings are profile-injected for $L$ layers, with the profile bias removed after aggregation:

$$z_e = \sum_{\ell=0}^L \overline{e}_e^{(\ell)} - \lambda_p M(p_e)$$

where message passing uses both structural and profile information (see section 2 of (Ahn et al., 13 Jan 2026)).

• DPAO: Embeddings are pooled over a policy-chosen number of GNN layers, with dual policy RL governing the aggregation kernel for each node.

4. Alignment and Optimization Objectives

SPiKE introduces a dedicated pairwise profile-preference matching loss that aligns the structure of profile-generated and KG-learned embeddings by minimizing the Frobenius norm of their pairwise cosine similarity matrices (Eq. 8 in (Ahn et al., 13 Jan 2026)). This regularization enforces semantic consistency between entities' LLM-derived and KG-derived profiles, supplementing standard recommendation losses such as Bayesian Personalized Ranking (BPR).

All methods ultimately produce a user-item scoring function via (possibly profile-aware) inner product, trained using BPR or similar loss structures. Regularization and optimization strategies include layer-wise combination, $L_2$ weight decay, early stopping, and in RL-based models, experience replay and target networks.

5. Training Workflow and Pseudocode

A typical training epoch in profile-aware KG aggregation incorporates profile extraction, feature filtering or injection, embedding initialization, graph message passing, and loss optimization.

For KGUF, the steps are:

1. For each user, build positive/negative item sets, train personal decision tree, collect relevant KG features.
2. For each item, prune KG feature set to retain only features relevant for some user split.
3. Initialize embeddings.
4. Propagate user and item embeddings with profile-aware mixture at each layer.
5. Combine layers.
6. Update parameters by BPR loss over sampled triplets.

SPiKE's workflow includes LLM profile encoding, profile-biased embedding initialization, multi-layer relation-aware aggregation, removal of injected profiles before scoring, and pairwise profile matching loss.

DPAO incorporates RL rollouts for both user and item MDPs, dynamic GNN depth, joint DQN and GNN updates, and plug-and-play adaptation to various base GNN-R models.

6. Empirical Performance and Ablation Insights

Profile-aware KG aggregation methods consistently outperform non-profiled baselines across multiple standard benchmarks.

• KGUF achieves superior nDCG@10 on MovieLens 1M (0.3277), Yahoo! Movies (0.2561), and competitive results on Facebook Books, outperforming SOTA GNN and KGCF methods where appropriate (Bufi et al., 2024).
• SPiKE demonstrates up to +2.9% absolute gain in Recall@10 over both KG-only and LLM-only recommenders, with ablation showing each profile component—including injection-removal and pairwise matching—is critical for full performance (Ahn et al., 13 Jan 2026).
• DPAO yields up to 63.7% relative improvement in nDCG@20 and 42.9% in Recall@20 on KG-based Amazon-Book, establishing adaptive aggregation as beneficial (Jung et al., 2023).
• In AKGAN, removal of independent attention or coarse attribute merging results in diluted or less interpretable recommendations (Huai et al., 2021).

Ablation studies in KGUF and SPiKE confirm that both collaborative propagation and semantic feature injection contribute to optimal accuracy, and that excessive aggregation depth can be detrimental due to over-smoothing.

7. Methodological Innovations and Future Directions

Methods such as SPiKE represent an emerging trend of combining LLM-driven profile synthesis with KG-based graph propagation, exploiting the expressive depth of transformer models for profile generation and the scalability/effectiveness of KGs for local and global reasoning.

The use of decision trees, attention mechanisms unconstrained by simplex normalization, and reinforcement learning for aggregation control reflects a landscape focused on interpretability, computational efficiency, and adaptability to the structural and behavioral heterogeneity present in practical recommender deployments.

Future research is likely to explore further the integration of multimodal profiles, dynamic profile updating, transferability across domains, and causal profile selection mechanisms, given observed sensitivities in ablation studies and continued performance gains over both deep and simple baselines. Empirical evidence favors profile-aware aggregation as a foundational component in next-generation KG-aware recommender systems (Bufi et al., 2024, Huai et al., 2021, Jung et al., 2023, Ahn et al., 13 Jan 2026).