Disentangled HAN: GNN Recommender for HIN

• The paper introduces a novel graph neural recommendation method that explicitly disentangles node representations into semantically independent aspect subspaces.
• It employs an EM-style propagation layer and attention mechanisms to aggregate intra- and inter-relation signals, isolating diverse collaborative factors like popularity and social influence.
• Empirical results on datasets such as Yelp, Amazon-Electronics, and MovieLens-HetRec demonstrate significant improvements in precision, recall, and interpretability over conventional methods.

Disentangled HAN (DisenHAN) is a graph neural recommendation architecture for heterogeneous information networks (HINs) that explicitly disentangles node representations into semantically independent aspect subspaces, learning aspect-wise propagation weights for each meta-relation. Compared to conventional heterogeneous graph neural networks, DisenHAN enables both higher accuracy and greater interpretability by identifying the main aspect to which each meta-relation contributes in the user/item embedding, thus isolating collaborative signals specific to popularity, category-preference, brand-loyalty, social influence, and related latent factors (Wang et al., 2021).

1. Problem Formulation and Notation

DisenHAN operates on an HIN represented as $\mathcal{G} = (\mathcal{V}, \mathcal{E})$, where $\mathcal{V}$ is a set of nodes and $\mathcal{E}$ is a set of edges. Each node $v \in \mathcal{V}$ is assigned a type via $\phi: \mathcal{V} \to \mathcal{A}$ ($\mathcal{A}$ is the set of node types, e.g., User, Item, Brand, Category), and each edge $e \in \mathcal{E}$ is assigned a meta-relation via $\psi: \mathcal{E} \to \mathcal{R}$ ($\mathcal{R}$ is the set of edge types, e.g., interacts, belongs_to, friend_of).

The main entities of interest are sets of users $$\mathcal{U} = \{u_1, \ldots, u_M\} \subset \mathcal{V}$$ and items $\mathcal{V}$0; the observed user-item implicit feedback is denoted $\mathcal{V}$1. The central goal is to learn a scoring function

$\mathcal{V}$2

that ranks items $\mathcal{V}$3 for each user $\mathcal{V}$4 using disentangled (aspect-wise) representations.

2. Disentangled Embedding Representation

Each node $\mathcal{V}$5 embeds into $\mathcal{V}$6 latent “channels” (aspects), separating sources of collaborative signal into independent subspaces. Given node features $\mathcal{V}$7, the projection for channel $\mathcal{V}$8 at layer $\mathcal{V}$9 is

$\mathcal{E}$0

where $\mathcal{E}$1 is type- and channel-specific, and $\mathcal{E}$2 is an activation (e.g., ReLU). This normalization confines each channel’s embedding to a unit hypersphere, supporting intra- and inter-node compatibility learning.

Disentangling these embeddings is critical for robustness and for interpretability, as it isolates how much each meta-relation contributes to particular aspects (e.g., distinguishing social and category influences).

3. Meta-Relation Decomposition

DisenHAN leverages meta-relations at the edge level, not extended meta-paths. Each edge $\mathcal{E}$3 is labeled as $\mathcal{E}$4. For target node $\mathcal{E}$5, the set $\mathcal{E}$6 contains all $\mathcal{E}$7 such that $\mathcal{E}$8 is of meta-relation $\mathcal{E}$9.

Rather than building an explicit adjacency tensor $v \in \mathcal{V}$0, DisenHAN handles edges grouped by relation type. This design supports multi-layer composition of meta-relational evidence, enabling automatic construction of higher-order semantic signals through propagation, avoiding hand-crafted meta-path enumeration.

4. Disentangled Embedding Propagation Layer

Each layer propagates aspect-specific signals in two phases: intra-relation (within one meta-relation) and inter-relation (across meta-relations), using learnable attention and an EM-style clustering approach for aspect assignment.

Intra-Relation Aggregation

Attention weights are computed for channel $v \in \mathcal{V}$1 from $v \in \mathcal{V}$2 to neighbor $v \in \mathcal{V}$3, per meta-relation $v \in \mathcal{V}$4: $v \in \mathcal{V}$5 where $v \in \mathcal{V}$6 is a relation-specific vector, and $v \in \mathcal{V}$7 denotes concatenation.

A prior aspect weight $v \in \mathcal{V}$8 modulates the aggregation: $v \in \mathcal{V}$9 Softmax normalization yields attention $\phi: \mathcal{V} \to \mathcal{A}$0, and neighbor feature aggregation per channel is

$\phi: \mathcal{V} \to \mathcal{A}$1

Inter-Relation Aggregation

Relation-channel scores are computed per channel $\phi: \mathcal{V} \to \mathcal{A}$2: $\phi: \mathcal{V} \to \mathcal{A}$3 Aspect-assignments are normalized: $\phi: \mathcal{V} \to \mathcal{A}$4 The updated channel embedding (with self-loop) is

$\phi: \mathcal{V} \to \mathcal{A}$5

This EM-style process aligns each meta-relation with a dominant aspect assignment, promoting stable, semantically meaningful clustering of relational signal into the disentangled channels.

5. Multi-Layer Architecture and Aspect-wise Collaborative Filtering

By stacking $\phi: \mathcal{V} \to \mathcal{A}$6 such disentangled propagation layers, each node’s channels $\phi: \mathcal{V} \to \mathcal{A}$7 can integrate $\phi: \mathcal{V} \to \mathcal{A}$8-hop information with increasing semantic abstraction (often decreasing channel count by layer: $\phi: \mathcal{V} \to \mathcal{A}$9).

The final user-item (aspect-wise) matching is

$\mathcal{A}$0

Each aspect channel thereby contributes additively to the total affinity score, enforcing explicit aspect attribution in the recommendation.

6. Training Objective

For top-$\mathcal{A}$1 recommendation, DisenHAN uses binary cross-entropy loss with negative sampling. Probabilities are computed as $\mathcal{A}$2. The training loss is

$\mathcal{A}$3

where $\mathcal{A}$4 is the set of observed positives, $\mathcal{A}$5 are sampled negatives, and $\mathcal{A}$6 regularizes all parameters $\mathcal{A}$7. Mini-batch Adam is used for optimization.

7. Empirical Results and Interpretability

DisenHAN was benchmarked on Yelp, Amazon-Electronics, and MovieLens-HetRec datasets, compared to classical CF (BPR-MF, NeuMF), HIN models with fixed meta-paths (FMG$\mathcal{A}$8, MCRec, NeuACF), and GNNs without disentanglement (GC-MC, NGCF) or with fixed meta-paths (HAN) (Wang et al., 2021). DisenHAN achieved the highest Prec@10, Recall@10, and NDCG@10 on all datasets, exceeding the best prior methods by $\mathcal{A}$9–$e \in \mathcal{E}$0 in relative improvement.

Its advantage is most pronounced in sparse regimes (cold-start conditions with $e \in \mathcal{E}$1–$e \in \mathcal{E}$2 observed interactions), validating that aspect-wise disentanglement of propagation delivers more robust collaborative signals. Ablation studies reveal that failing to disentangle ($e \in \mathcal{E}$3) significantly degrades performance, and an aspect count $e \in \mathcal{E}$4 is optimal; excessive channels lead to overfitting. Only a small number ($e \in \mathcal{E}$5) of EM iterations per layer are necessary, with two propagation layers ($e \in \mathcal{E}$6) typically optimal.

Aspect assignment weights $e \in \mathcal{E}$7 provide direct interpretability: for example, in Amazon-Electronics, user-item relations cluster under one aspect