Placing Clicks on the Purchase Ladder: Learned Ranking Margins from Shared Behavior Factors

Online shops log far more cheap actions (views, clicks, add-to-cart) than purchases. A recommender trained only on purchases sees almost nothing per user. One trained on everything treats a curious click as if it were a purchase. Current multi-behavior recommenders handle this in one of two ways. Some fix a coarse ordering: purchased items rank above clicked items, which rank above unseen items. Others reweight whole behavior types per user. Neither asks the question that matters for each individual click: how much does it actually say about what this user will buy?

We propose to answer that question per user–item pair and to use the answer directly as a ranking margin. The model splits its representation into factors shared across all behaviors and factors specific to each behavior. A clicked item whose click is explained mainly by shared factors (the same factors that explain the user's purchases) is placed close to purchases on the training "ladder". A click explained mainly by click-specific factors, such as browsing habits or eye-catching thumbnails, is placed close to unobserved items. The ladder position becomes the margin in a pairwise loss. The result is a graded, data-driven ordering, which replaces both synthetic noise hierarchies and fixed behavior tiers.

Core Claim

Pairwise ranking losses make a model score a preferred item above a less-preferred one by at least some margin. The size of that margin matters a great deal. Graph Attention Collaborative Similarity Embedding for Recommender System (GACSE) adapts the margin to structural similarity in a single interaction graph, and removing that adaptive margin costs up to 15%. This finding motivates our work. Under sparse purchase data, the margin is where supervision is won or lost.

Our claim is that, for multi-behavior data, the right margin for an auxiliary interaction is set by how much of it lives in a subspace shared with the target behavior. This per-pair margin should beat two simpler choices: a fixed per-behavior tier, and per-user behavior weights. It should help most for users with very few purchases.

How It Differs From the Closest Work

Three lines of work sit nearest to this proposal.

Multi-view Multi-behavior Contrastive Learning in Recommendation (MMCLR) already uses a ranking-style contrastive term with the fixed order target > auxiliary > negative. Every clicked item gets the same tier, and there is no shared-versus-specific decomposition. Our proposal replaces the fixed tier with a learned position for each clicked item.

From Pairwise to Ranking: Climbing the Ladder to Ideal Collaborative Filtering with Pseudo-Ranking (PRP) builds multi-level orderings by adding Gaussian noise to positive item embeddings. It reports that gains fade beyond list length 4–5, which suggests synthetic levels lose meaning. This result constrains us. We use real behavior data rather than invented intermediate rungs.

Contrastive Meta Learning with Behavior Multiplicity for Recommendation (CML) learns one weight per user per behavior type. It cannot tell two clicks by the same user apart, which is exactly the distinction our margin targets.

Factor-disentangling methods do separate latent factors, but they score with plain dot products or uniform pairwise losses. These are Latent Factor Modeling With Expert Network for Multi-Behavior Recommendation (MBLFE, gated experts) and Disentangled Cascaded Graph Convolution Networks for Multi-Behavior Recommendation (Disen-CGCN, factor blocks with distance-correlation penalties). We borrow the shared/specific split from the multi-view literature. Shared Subspace Learning for Latent Representation of Multi-View Data supplies a shared coefficient matrix with view-specific bases. Multi-View Information-Bottleneck Representation Learning supplies shared plus view-specific variables. Both were built for dense clustering and classification, and neither was used to set ranking margins.

Method

A LightGCN-style encoder (a graph network that averages neighbor embeddings without nonlinearities) runs over the multi-behavior graph. Each user $u$ and item $i$ get a shared vector $s_u, s_i$ and, for each behavior $b$, a specific vector $p^b_u, p^b_i$. The model's score for behavior $b$ is $\hat y^b_{ui} = s_u^\top s_i + (p^b_u)^\top p^b_i$, and the target-behavior (purchase) score is $\hat y_{ui}$.

All behaviors are fit jointly. Because purchases must be explained mostly through the shared factors, those factors are pushed toward purchase-relevant structure. An independence penalty (distance correlation, as in Disen-CGCN) keeps shared and specific factors from duplicating each other.

For each auxiliary-only pair $(u,j)$, meaning clicked or carted but not purchased, we compute an intent estimate. It is computed by a slowly updated copy of the model (an exponential moving average, or EMA, "teacher"), and gradients are not passed through it:

$$\rho_{uj} = \frac{\exp(\bar s_u^\top \bar s_j)}{\exp(\bar s_u^\top \bar s_j) + \exp((\bar p^b_u)^\top \bar p^b_j)}$$

Here bars denote teacher embeddings. $\rho_{uj}\in(0,1)$ is the fraction of the auxiliary interaction attributed to shared factors.

With $i$ a purchased item, $j$ an auxiliary-only item, $k$ an unobserved item, and $\gamma$ a single global margin scale, the ranking loss is

$$\mathcal{L} = -\log\sigma\big(\hat y_{ui}-\hat y_{uj}-\gamma(1-\rho_{uj})\big) - \log\sigma\big(\hat y_{uj}-\hat y_{uk}-\gamma\rho_{uj}\big)$$

The total gap between a purchase and an unseen item stays at $\gamma$. The clicked item's position inside that gap is learned per pair.

We use the teacher and the stop-gradient deliberately. Without them, the student could shrink its margins by shrinking its own shared scores. The behavior-fitting loss is added on top. Only $\gamma$ and the teacher rate need tuning, compared with the many loss weights MBLFE reports as a burden.

Research Program

The work has three parts.

The first is building the encoder, the factor split, and the ladder loss, and confirming that $\rho$ does not collapse to a constant. We will track the variance of $\rho$ over training.

The second is a validity analysis of $\rho$ itself. In standard splits of these e-commerce datasets, many test purchases are items the user had clicked during training. We can therefore ask whether $\rho$, computed only from training data, predicts which clicked items get bought at test time. We measure this with AUC of $\rho$ among clicked items. Doing so separates "the margin carries intent information" from "the loss happened to regularize well".

The third is the geometric question raised by Are Graph Augmentations Necessary?: Simple Graph Contrastive Learning for Recommendation (SimGCL). SimGCL shows that contrastive losses help sparse recommendation by spreading embeddings uniformly over the sphere, whereas a shared subspace wants structured, non-uniform directions. We will test applying SimGCL-style noise contrast to the specific factors only versus to everything. The aim is to see whether uniformity and the shared subspace conflict in practice.

Baselines

The naive reference points are LightGCN trained with standard Bayesian Personalized Ranking (BPR, a pairwise loss ranking observed above unobserved items) on purchases only, and the same model trained with all behaviors merged into one graph and treated as positives. The merged version is important. In datasets where buyers usually click first, it can be surprisingly strong, and it is the "do nothing clever" bar that any ladder must clear.

The strongest existing competitors are MBLFE and Disen-CGCN, the most recent factor-based multi-behavior models in the file. CML and MMCLR are the closest in mechanism.

We have no comparable absolute numbers to quote. MBLFE reports a 2.69% relative HR@10 gain and a 6.92% relative NDCG@20 gain over its best baseline on Taobao. Disen-CGCN reports an average 9.00% gain on Tmall. CML reports 25.4% HR@10 on Tmall. The file contains no reported absolute HR@10 or NDCG@10 values for any of these methods on Tmall, Taobao, or Beibei, and their splits differ. We will therefore rerun all baselines from released code on identical splits with the same embedding size, the same epoch budget, and the same number of hyperparameter trials per method (we propose 30, random search on a validation set). Beating them means a statistically significant gain in NDCG@10 over the best of them on at least three of four datasets, measured over five seeds.

Three ablations isolate the key ingredient, the shared-subspace margin. The first fixes $\rho$ to one tuned constant per behavior, an MMCLR-style tier. The second computes $\rho$ from the full click score with no shared/specific split. The third replaces per-pair $\rho$ with a per-user, per-behavior learned weight, as in CML. A fourth applies GACSE's similarity-based adaptive margin to the merged graph. This tests whether any adaptive margin, rather than our specific one, explains the gain.

Why not just…?

Why not just use a fixed three-tier ordering with a tuned middle margin? This is the most serious threat. If most clicks carry similar intent, a single tuned constant could match a learned per-pair $\rho$. The fixed-$\rho$ ablation settles it directly. The $\rho$-validity analysis explains the result either way: if $\rho$ has near-chance AUC for predicting which clicks turn into purchases, the per-pair idea fails and we will report that.

Why not just reweight behaviors per user, as CML does? Per-user weights treat all of a user's clicks alike. Our hypothesis is that intent varies more within a user's clicks than between users. The CML-style ablation measures this. We will also report how much of the variance in $\rho$ is within-user versus between-user. If between-user variance dominates, CML's simpler design suffices.

Why not just merge all behaviors and add SimGCL's embedding-noise contrast? SimGCL gives large single-behavior gains, over 25% in Recall@20 versus LightGCN on Amazon-Book, at modest cost. Its noise, however, has no notion of which interactions are high-intent. We include merged-graph SimGCL as a baseline. If it matches us on overall metrics, the remaining test is the sparse-user slice, described next.

Proposed Experiments

We will use Tmall, Taobao, Beibei, and Retail Rocket, with purchase as the target behavior. Metrics are HR@10 and NDCG@10 under full ranking over all items rather than sampled negatives.

The decisive checks are three. The first is the gain over the fixed-tier ablation, reported overall and in the bin of users with fewer than four purchases, the regime where CML reported its own advantage. The second is the AUC of training-time $\rho$ for predicting test-time conversion of clicked items. The third is a sparsity sweep that randomly removes 25–75% of training purchases. The proposal predicts that the margin of the learned ladder over fixed tiers widens as purchases thin out, because sparser targets make the quality of auxiliary supervision more important.

A null result on the first two checks, while overall metrics still improve, would mean the gain comes from regularization rather than intent modeling. We would report it as such.