Target-Anchored Signed Weighting for Sparse Multi-Behavior Graphs

Online shops log many kinds of user actions. Views and clicks number in the millions, while purchases, the action the business wants to predict, are orders of magnitude rarer. A common strategy splits the interaction log into one graph per behavior and lets the plentiful click graph support the scarce purchase graph. This helps, but only in aggregate. Some clicks predict purchases and many do not. Current models trust a whole behavior with one scalar weight, so they cannot tell the two kinds of click apart. The dense click graph also causes over-smoothing: repeated neighbor averaging blurs every user's representation toward the same vector, so users become indistinguishable.

We propose a small change to how each behavior graph passes messages. Every auxiliary edge, such as a click from user $u$ on item $i$, gets a weight computed from how much $u$ and $i$ already agree in the purchase embedding space. The weight is signed. Agreeing edges are averaged in, which is ordinary smoothing. Disagreeing edges are subtracted, which pushes the two endpoints apart and acts as a high-pass filter. The weight uses no projection matrices, only a cosine similarity and one temperature per behavior. This keeps the parameter-light design that works for recommenders whose nodes have no features beyond an ID. The central claim is testable. Harm from auxiliary behaviors lives at the level of individual edges, not whole behaviors. A projection-free, signed, target-anchored weight can use that fact to reduce both negative transfer and over-smoothing in dense views. It should help most for users with very few purchases.

Core Claim and Mechanism

The design has to satisfy two lessons from prior work at once.

The first lesson is about parameters. LightGCN showed that on ID-only graphs, learned transformation matrices and nonlinearities hurt collaborative filtering. It uses fixed degree-normalized averaging and averages over layers, and it reports about a 16% relative gain over NGCF. This constrains our design.

The second lesson is about what fixed weights cannot do. Graph Attention Networks and its descendants learn which neighbors matter, but through a projection $\mathbf{W}$ and an attention vector. Those are exactly the components LightGCN warns against. GTCF4MB is the one multi-behavior model in our file that adds such attention. It inserts full query/key/value transformer projections and relies only on supervised signal. Its authors note that it underperforms on sparse Taobao and that it needs self-supervision.

Our mechanism keeps LightGCN's linear propagation inside each behavior graph $b$ and changes only the edge coefficient. Let $\mathbf{z}_u$ and $\mathbf{z}_i$ be the current target-behavior (purchase) embeddings, used with gradients stopped. Let $d_u$ and $d_i$ be the node degrees in graph $b$. Let $\tau_b$ be a learned temperature for that behavior. The layer update is

$$ \mathbf{e}_u^{(k+1,b)} = \sum_{i \in \mathcal{N}_b(u)} \frac{\tanh\!\big(\cos(\mathbf{z}_u,\mathbf{z}_i)/\tau_b\big)}{\sqrt{d_u d_i}}\;\mathbf{e}_i^{(k,b)}, $$

where $\mathcal{N}_b(u)$ is $u$'s neighbor set in behavior $b$. The purchase graph itself uses plain LightGCN weights. The final representation averages over layers and fuses the per-behavior outputs into the target embedding.

The same coefficient plays three roles. It weights each neighbor individually, which is the dynamic-weighting direction. Its negative range acts like the high-pass channel that FB-GNN adds as a separate branch with fixed Laplacian filters. Here the high-pass action is applied edge by edge, where the purchase signal says it belongs. Its anchor is the purchase view, so a click only helps if it points toward the user's purchase preferences.

We stop gradients through $\mathbf{z}$ for two reasons. It prevents the weights from collapsing onto hub nodes or toward uniformity, a known failure of attention under sparse features. It also keeps the per-behavior parameter count at one scalar.

To handle target sparsity we add SimGCL's embedding-noise contrastive loss on the target embeddings. SimGCL regularizes embeddings toward uniformity without deleting edges. Deleting edges would be damaging in views that are already more than 99.9% empty. We treat this loss as shared infrastructure, not as a contribution. Every baseline that can use it will get it.

Relation to Prior Work

Two lines of multi-behavior work motivate the gap.

S-MBRec splits interactions into per-behavior graphs. It runs isotropic GCNs on each and fuses them with behavior-level coefficients plus a star-shaped contrastive loss centered on the target behavior. CML uses a meta-network to learn personalized weights per behavior. Both locate negative transfer at the behavior level, or at the user-by-behavior level. We locate it at the edge. If the edge-level hypothesis is false, per-user behavior weights should match us, and that is what the ablations test.

HGCL and MCCLK decompose data into social, knowledge, and interaction views and align them contrastively. Their within-view encoders are static LightGCN-style averaging. We reuse their decomposition setting but change what happens inside each view.

hpGAT enriches attention with structural proximity from matrix factorization. This supports the idea that a non-feature signal can drive attention. However, hpGAT needs offline factorization and has been tested only on homogeneous node classification. Our driving signal is cross-view agreement, which is computed online at no extra parameter cost.

Relation-aware heterogeneous GNNs motivate decomposing by relation in the first place. They also flag the lost cross-relation interaction that our target anchoring partly restores: each auxiliary view is read through the target view.

Research Program

The work has three parts.

We will build the model on a shared LightGCN codebase so that every variant differs only in its edge coefficient and fusion.

We will analyze over-smoothing per behavior graph as a function of that graph's density. We follow Are we measuring oversmoothing in graph neural networks correctly? and do not rely on a single cosine-similarity statistic. We will report embedding similarity measures alongside ranking accuracy as depth grows.

We will run an edge-level diagnostic. Using training data only, we label each click edge by whether the same user also purchased that item or a co-purchased neighbor. We then check whether the learned signed weights separate these groups, and whether that separation predicts accuracy gains on users with few purchases.

We will also run a transfer check on social recommendation. There, the user-user view plays the auxiliary role, which tests whether the mechanism generalizes beyond behavior types.

Baselines

The naive reference is LightGCN trained on the purchase graph alone. We also run a second naive reference: LightGCN on the union of all behaviors treated as one graph. The union run shows whether decomposition matters at all.

The strongest simple competitor is the union-graph LightGCN plus SimGCL's noise contrastive loss. SimGCL reports improving Recall@20 by over 25% relative to LightGCN on Amazon-Book, per SimGCL. It is cheap and robust, so a proposal that cannot beat it is not worth pursuing.

The strongest published multi-behavior methods are S-MBRec and GTCF4MB.

  • S-MBRec: it reports improvements "up to 20%" over baselines such as LightGCN and MB-GMN on Taobao.
  • GTCF4MB: it reports HR@10 improvements "up to 32.96%" over MBGCN, MB-GMN, and GNMR on Taobao.
  • Absolute numbers: our file contains only these relative figures. No reported absolute HR@10 or NDCG@10 value was found on Lacuna for either method on Taobao or Beibei. We will therefore reproduce both methods rather than copy numbers.

CML is included as the personalized behavior-weighting method; no reported result was found on Lacuna for CML.

Fairness rules apply to all runs. All methods use the same splits and the same embedding dimension. Each gets the same hyperparameter search budget, measured in trials, over its own key knobs. All train under the same wall-clock limit. The SimGCL loss is added to S-MBRec and GTCF4MB as well, so that gains cannot come from self-supervision alone.

The ablations isolate the key ingredient. The first replaces edge weights with one learned scalar per behavior, or per user and behavior, which tests edge-level against behavior-level weighting. The second clips weights to be nonnegative, removing the high-pass role. The third anchors the cosine on the auxiliary view's own embeddings instead of the target view. The fourth replaces the cosine with a GAT-style projected attention, which tests the projection-free claim. The fifth uses a fixed FB-GNN-style low-plus-high-pass filter per view.

Beating the field means two things. The method must significantly outperform both the tuned SimGCL-union baseline and the SimGCL-augmented S-MBRec on HR@10 and NDCG@10 on Taobao and Beibei, across five seeds. The gain must also be concentrated in the users with the fewest purchases.

Why not just…?

Why not just learn one weight per behavior, or per user and behavior, as CML and S-MBRec do? This is the most serious alternative. If harmful clicks are mostly specific to users, a per-user behavior weight captures most of the benefit at lower cost. The first ablation settles this directly. If per-user behavior weights come within noise of edge-level weights, and the edge diagnostic shows no separation among clicks by the same user, we will drop the edge-level claim. We would then report only the signed high-pass component if that component still helps.

Why not just run SimGCL on the union graph and skip decomposition? SimGCL's reported gains come from single-behavior graphs. Its fixed averaging cannot tell a purchase from a click. The multi-behavior gains that GTCF4MB and S-MBRec report over LightGCN suggest that behavior identity carries signal. Still, those comparisons did not give LightGCN a uniformity loss, so the evidence is inconclusive. Our SimGCL-union baseline is tuned with the same budget as our method, which makes the comparison clean.

Why not just prefilter clicks, keeping only those consistent with training purchases? This heuristic may capture much of the edge-level benefit for users with many purchases. For users with one or two purchases it deletes nearly all their clicks, and those are the users sparsity hurts most. We will run it as an additional baseline. Our method needs to win in the lowest-purchase group, even if it only ties in the upper groups.

Why not just use full attention, as GTCF4MB does? LightGCN's ablations show that projections hurt on ID-only graphs, and GTCF4MB's own authors report weakness on sparse Taobao. The fourth ablation keeps everything else fixed and swaps in projected attention. It tests whether the projection-free form actually matters.

Proposed Experiments

The main benchmarks are the multi-behavior datasets Taobao and Beibei, used by both S-MBRec and GTCF4MB, plus multi-behavior Yelp. We evaluate with HR@10, NDCG@10, and Recall@20, using full ranking rather than sampled negatives. Results are stratified by each user's number of training purchases. The transfer check uses Ciao and Epinions with the trust graph as the auxiliary view, against HGCL. HGCL reports a 7.3% NDCG@10 gain over its strongest baseline on Ciao; our file has no absolute value for it.

The first decisive check is a depth sweep from 1 to 8 layers in the click view. If the signed weights counter over-smoothing, accuracy should keep holding or improving beyond the 2 to 4 layers where LightGCN-style models plateau. The nonnegative ablation should degrade across the same depth range.

The second decisive check correlates learned edge weights with the training-only purchase-consistency labels, then tests whether that separation predicts the accuracy gain for sparse users.

The third check injects noise by adding 20% random click edges, following SGL's protocol for random interaction noise. Our method should assign negative or near-zero weight to the injected edges, and its accuracy should degrade less than that of scalar-weighted baselines.

A negative outcome would falsify the core claim. That outcome is equal accuracy together with no weight separation among edges within a behavior.