TopKGAT: Top-K Objective Recommender

• TopKGAT is a top-K objective-driven recommender that integrates a differentiable relaxation of Precision@K directly into its GNN layers.
• The architecture aligns its layerwise aggregation with evaluation metrics using gradient-ascent steps and a band-pass attention function near the top-K threshold.
• Empirical evaluations on benchmark datasets demonstrate statistically significant gains in NDCG@20 and Recall@20 over traditional models.

TopKGAT is a top-K objective-driven recommendation architecture that integrates a differentiable relaxation of the Precision@K metric directly into its graph neural network (GNN) layers. This approach enforces an inductive bias precisely aligned with the actual evaluation metrics used in recommender systems—specifically, Precision@K and Recall@K—rather than relying on surrogate or pairwise ranking losses. The design leverages graph attention mechanisms and efficiently adapts message passing for large-scale bipartite user–item interactions by focusing model capacity on scores near the top-K cutoff boundary (Chen et al., 26 Jan 2026).

1. Differentiable Relaxation of Top-K Metrics

TopKGAT is built upon a differentiable approximation to Precision@K and Recall@K. The standard discrete metrics for a user $u$ are defined as:

• Precision@K: $$\mathrm{Precision}@K(u) = \frac{|R_u^K \cap T_u|}{K}$$
• Recall@K: $$\mathrm{Recall}@K(u) = \frac{|R_u^K \cap T_u|}{|T_u|}$$

where $R_u^K$ is the set of K items with the highest predicted scores $s_{ui}$, and $T_u$ is the ground-truth test set for user $u$.

To make these metrics differentiable, TopKGAT introduces the $K$-quantile threshold $\beta_u^K = \inf\{s_{ui} : i \in R_u^K\}$, such that an item $i$ is in $R_u^K$ iff $s_{ui} \ge \beta_u^K$. The intersection is rewritten:

$$|R_u^K \cap T_u| = \sum_{i \in T_u} \mathbb{I}(s_{ui} - \beta_u^K \ge 0),$$

where the indicator $\mathbb{I}$ is replaced with a smooth sigmoid:

$$\mathbb{I}(x \ge 0) \approx \sigma(x),\quad \sigma(x) = \frac{1}{1 + e^{-x}}.$$

This yields a differentiable Precision@K:

$$\mathrm{Precision}@K(u) \approx \frac{1}{K} \sum_{i \in T_u} \sigma(s_{ui} - \beta_u^K).$$

Aggregating over all users and interactions, with degree-normalization and $L_2$ regularization, the global objective is:

$$\mathcal{J}_{\mathrm{Pre@K}} = \sum_{(u,i)\in D} \frac{\sigma(s_{ui} - \beta_u)}{\sqrt{d_u d_i}} - \lambda \|\mathbf{Z}\|_2^2,$$

where $D$ is the set of interactions, $d_u$ and $d_i$ are user/item degrees, and $\beta_u$ are learnable quantile approximations (Chen et al., 26 Jan 2026).

2. GNN Layers as Gradient-Ascent on Precision@K

Each layer in TopKGAT corresponds to a gradient-ascent step on $\mathcal{J}_{\mathrm{Pre@K}}$. For embeddings $\mathbf{Z}^{(l)}$,

$$\mathbf{Z}^{(l+1)} = \mathbf{Z}^{(l)} + \tau \frac{\partial \mathcal{J}_{\mathrm{Pre@K}}}{\partial\mathbf{Z}^{(l)}}.$$

Specializing for user embeddings and absorbing constants gives:

$$z_u^{(l+1)} = \sum_{i \in N_u} \frac{\omega((z_u^{(l)})^T z_i^{(l)} - \beta_u^{(l)})}{\sqrt{d_u d_i}} z_i^{(l)},$$

$$z_i^{(l+1)} = \sum_{u \in N_i} \frac{\omega((z_u^{(l)})^T z_i^{(l)} - \beta_u^{(l)})}{\sqrt{d_u d_i}} z_u^{(l)},$$

where $$\omega(x) = 4 \sigma'(x) = \frac{4}{(1 + e^{-x})(1 + e^{x})}$$, a band-pass function sharply peaking at the top-K threshold boundary. Thus, each GNN layer directly implements one step of (smoothed) top-K-aware optimization (Chen et al., 26 Jan 2026).

3. Attention Mechanism and Personalized Thresholds

TopKGAT operates on a bipartite user–item graph $G=(U \cup I, D)$, where $D$ is the user–item interaction set. For each edge $(u, i)$, the similarity score at layer $l$ is $s_{ui}^{(l)} = (z_u^{(l)})^T z_i^{(l)}$. The edge attention weight is based on $\omega(s_{ui}^{(l)} - \beta_u^{(l)})$, which emphasizes interactions around the current top-K threshold. Each user and layer maintains a personalized learnable threshold $\beta_u^{(l)}$, allowing dynamic adaptation across both network depth and users. Contributions are degree-normalized by $\sqrt{d_u d_i}$ for stability. The aggregation rules constitute a graph attention mechanism specialized for the top-K objective, with the attention function and bias derived analytically rather than heuristically.

4. Training Objective and Optimization

While TopKGAT layers are derived to follow gradient steps on $\mathcal{J}_{\mathrm{Pre@K}}$, end-to-end model training uses the standard Bayesian Personalized Ranking (BPR) pairwise loss plus $L_2$ regularization, with the Adam optimizer. The differentiable top-K objective is structurally embedded in the architecture, but the outer loss remains BPR as in standard practice. This maintains compatibility with prevailing evaluation and negative sampling procedures, while the inductive bias of the layers continues to enforce attention to the top-K region. Embedding normalization is applied prior to score computation for numerical stability.

5. Computational Efficiency and Implementation Details

Each layer in TopKGAT requires $O(|D| \cdot d)$ computational complexity per layer, identical to LightGCN, since the operations are sparse-dense matrix multiplications over the bipartite adjacency. The threshold parameters $\beta_u^{(l)}$ add only $L \cdot |U|$ scalars to the model. All similarity scores and attention weights are calculated using batch-vectorized dot products and the pointwise band-pass function $\omega$, obviating the need for top-K sorting during forward or backward passes. Sparse adjacency structures and GPU-accelerated operations enable practical application to million-edge graphs.

6. Empirical Evaluation and Results

Experiments were conducted on four benchmark datasets (5-core, split 7/1/2 for train/validation/test): Ali-Display (17,730 users, 10,036 items, 173,111 interactions), Epinions (17,893 users, 17,659 items, 301,378 interactions), Food (14,382 users, 31,288 items, 456,925 interactions), and Gowalla (55,833 users, 118,744 items, 1,753,362 interactions). Evaluation metrics were Recall@20 and NDCG@20.

Baselines included non-attention methods (MF, LightGCN, LightGCN++, ReducedGCN) and attention/transformer-type models (GAT, NGAT4Rec, MGFormer, Rankformer). TopKGAT achieved consistent and significant ($p < 0.05$) improvements in both NDCG@20 and Recall@20:

Dataset	NDCG@20	Recall@20
Ali-Display	+5.33%	+4.10%
Epinions	+4.51%	+4.32%
Food	+3.09%	+1.80%
Gowalla	+1.19%	+1.13%

All improvements are statistically significant, demonstrating the advantage of aligning the layerwise inductive bias directly with the top-K recommendation objective.

7. Architectural Alignment and Core Insights

The central principle of TopKGAT is the direct alignment of GNN aggregation dynamics with the smoothed Precision@K objective. The band-pass attention activation $\omega(\cdot)$ promotes learning focus on items whose scores are near the current top-K threshold for each user, while the per-user, per-layer thresholds $\beta_u^{(l)}$ permit dynamic adaptation. The resulting architecture tightly couples model capacity with the evaluation metric, which yields non-trivial and consistent improvements over GCN- and GAT-based recommenders that do not incorporate the top-K cutoff explicitly. This mechanism sets TopKGAT apart from prior approaches, providing a targeted solution to the longstanding objective/architecture mismatch in recommendation models (Chen et al., 26 Jan 2026).