Recommender systems that use graphs connect each user to the items they touched and pass information along those connections. When a user's tastes change, those connections go stale. Today's systems handle this in one of two ways. Some fade every old interaction by its age, using a single forgetting rate. Others clean the graph once by deleting edges that look like noise, and then treat the cleaned graph as fixed. Neither approach can express the common case in which a three-year-old purchase still describes who the user is, while yesterday's accidental click does not. Neither can express an interest that goes dormant and later returns.
We propose to learn, for every interaction edge, whether it is part of the user's current preference graph. That on/off decision is allowed to change over time, but only rarely, and only a small fraction of edges are kept at any moment. Two ideas from sparse graph learning make this tractable. The first is an L1 penalty, a cost proportional to how many edges are kept, which drives most edge weights to exactly zero. The second is a penalty on how often each edge's on/off status flips. Together they produce a graph whose set of active edges changes in a few discrete jumps rather than fading smoothly. The testable consequence is that this model should beat age-based forgetting specifically on users whose next action revisits an old interest. It should do so without losing on users whose next action reflects a brand-new interest.
Core Claim
Edge relevance under preference drift is better modeled as a sparse set of active edges that switches on and off at a few change points than as a weight that decays with age. Concretely, let $g_e(t) \in [0,1]$ be a gate on edge $e$ at time $t$, where $e$ connects a user to an item or one item to another. The gate scales that edge's contribution to message passing in a graph neural network, a model that builds each node's representation by aggregating its neighbors. We train
$$\min_{\Theta,\phi}\; \mathcal{L}_{\text{rec}}(\Theta; G_{g_\phi}) + \lambda \sum_{e,t} g_e(t) + \mu \sum_{e,t} |g_e(t) - g_e(t-1)|,$$
where $\mathcal{L}_{\text{rec}}$ is the next-item ranking loss of the graph network with weights $\Theta$, run on the gated graph $G_{g_\phi}$. The gates come from a small network with parameters $\phi$. The term weighted by $\lambda$ is the L1 sparsity cost. The term weighted by $\mu$ is a total-variation cost, a penalty on the size of each change between consecutive time steps, which charges for every flip in an edge's status.
Exponential forgetting is a special case of this setup: the gates are forced to decrease monotonically in edge age. The claim is that relaxing this restriction, while keeping both penalties, recovers dormant-but-valid edges and drops recent noise. Neither can be done by an age-only rule.
Research Program
The gate network $\phi$ does not store a separate parameter for each edge at each time step, which would not scale. Instead it reads edge-level features. These include the edge's age, how many times the interaction was repeated, how recently the user's neighbors have been active on the same item, and the similarity between the item and the user's last few interactions. From these features it outputs the gate for that edge at that time.
Three techniques come from prior work, each in a specific role. First, from GraphTOP we take three pieces. That paper adapts pre-trained graph networks by rewiring edges through a Gumbel-Softmax relaxation, a way of making a discrete keep/drop choice differentiable. It adds an entropy loss that pushes edge probabilities toward 0 or 1, and it restricts rewiring to a local neighborhood around the target node. We reuse all three. GraphTOP is evaluated only on static graphs, so its losses contain no time term; we add the flip penalty. Second, from SRSG we take the idea of penalizing differences in neighbor sets. SRSG penalizes the difference between the neighbor sets of points that are close in space and shows that those neighbor sets stabilize during optimization. Our flip penalty applies the same idea across consecutive time steps of a single node rather than across nearby nodes. Third, from ConNIe we take a two-step procedure. ConNIe infers diffusion networks with an L1 penalty and then re-fits unpenalized weights on the selected edges to remove shrinkage bias. After the gates binarize, we re-fit the recommendation model on the selected active edges without the sparsity term.
Several papers motivate the design. GDMSR shows that deleting 10–40% of social edges improves recommendation, but it prunes once and assumes static interests. BISTRO documents drift that is both sharp and discrete: job seekers revise résumés every 7.28 days, and over 75% of those revisions are career switches. That pattern fits switching gates better than smooth decay. NS-DPO shows that recency weighting helps under drift, but it uses a single global discount that, by construction, cannot separate an old high-quality signal from a recent low-quality one.
Two papers constrain the design. DGSR reports over-smoothing when propagation layers are stacked over noisy neighbors, so we keep propagation shallow and let sparsity do the cleaning. GraphSASA improves long-tail users by adding virtual edges, which warns that aggressive pruning will hurt users with few interactions. We therefore scale $\lambda$ down with node degree, following GDMSR's degree-adaptive pruning, and report results separately for tail users.
The analysis deliverables are the learned gate trajectories themselves. For each user, these show which past interactions the model treats as current and when that changed. We also measure support recovery on data where the true change points are known.
Baselines
Naive reference points. We use two. The first is LightGCN-style propagation, a simplified graph network that averages neighbor embeddings, run on the full ungated history. The second is the same model trained only on each user's most recent $N$ interactions, with $N$ tuned. These indicate whether any form of edge selection matters on a given dataset.
Strongest sequential competitor. This is SASRec, a self-attention model that already learns to weight recent versus distant actions within a single user's sequence. It is the competitor most likely to absorb our gains through attention alone.
Strongest graph-based drift competitors. These are DGSR and GDERec. DGSR combines dynamic subgraph sampling with temporal-order attention. GDERec evolves embeddings continuously over a fixed graph. Between them they represent "drift handled in the embeddings, topology left alone." The file gives only relative numbers for DGSR: a 10.56% NDCG@10 gain over its strongest baseline on Amazon Beauty, and 3.92% on Games. GDERec reports average gains of 6.5% in Recall@5 and 8.8% in Recall@10 across five benchmarks. The only absolute figures are for other models: FCSRec reports Recall@10 of 0.09752 on ML-100K, and GraphSASA reports Recall@20 of 0.0263 on Taobao. Because the file contains no absolute DGSR numbers, we re-run every baseline ourselves.
Fairness protocol. All baselines use the same chronological split, filtering, and candidate set. All use the same embedding dimension and the same hyperparameter search budget, measured in trials. Our method's $\lambda$ and $\mu$ count against that same budget.
Key ablations. These isolate the main ingredient. The age-decay gate replaces our gate network with $g_e(t)=\sigma(a_u - b_u\cdot \text{age}_e)$, using a learned rate per user as in DPS. The static-prune variant sets $\mu \to \infty$, so each edge's status is fixed once, which reproduces GDMSR's regime on interaction edges. The no-flip-penalty variant sets $\mu=0$, which tests whether discrete switching, rather than gating alone, carries the gain.
Success criterion. Beating the baselines means a statistically significant Recall@10 and NDCG@10 improvement over both the age-decay gate and DGSR on at least three of the four main datasets, with no significant loss on the tail-user slice.
Why not just…?
Why not just decay old edges exponentially, with a learned rate per user? This is the main threat. DPS reports that rates per node beat a global rate by up to 8.4% in accuracy, so a well-tuned decay rule may capture most of the available gain. Our claim survives only if the age-decay ablation loses specifically on "returning-interest" test cases. These are cases where the held-out item's category was last seen long before the most recent interactions, which is exactly where monotone decay must assign low weight. If the age-decay gate matches us on that slice, we will report that switching gates add nothing beyond decay with per-user rates.
Why not just let SASRec's attention decide which past items matter? Attention does select over history. However, its selection is soft, it is computed separately for each user, and it never removes noisy neighbors from the cross-user graph. DGSR's gains over sequential models on Beauty suggest that cross-user signal matters. We test this directly by also running a gated SASRec that applies our gates to its history tokens. That comparison separates the value of sparse switching from the value of the graph.
Why not just denoise the graph once, as GDMSR does? One-shot pruning cannot reactivate an edge, and BISTRO's frequency of career switches suggests that edge relevance does change within the observation window. The static-prune ablation is the experiment that decides this. If it matches the full model, drift in the topology is not worth modeling on these datasets.
Why not just truncate to recent history? Truncation is the extreme form of decay and discards long-term identity outright. The recent-$N$ baseline measures how much that costs. We expect it to be competitive on dense, fast-moving data such as Taobao, and that would be informative rather than embarrassing.
Proposed Experiments
The main benchmarks are Amazon Beauty, Games, and CDs, used as in DGSR; ML-1M, used as in SASRec and GDERec; and Taobao with daily snapshots, used as in GraphSASA. We hold out the last interaction per user in chronological order and report Recall@10/20 and NDCG@10. We also report the fraction of edges kept and the number of flips per edge.
The first decisive check is the slice analysis described above. Test cases are split into returning-interest, new-interest, and continuing-interest groups using category metadata. The hypothesis predicts gains over the age-decay gate concentrated in the returning-interest group, and parity elsewhere.
The second check uses semi-synthetic support recovery. We take ML-1M users and splice in a second user's history at a known time, which creates known change points and known stale edges. We then measure F1 between the learned active edge set and the true one, comparing against age decay and static pruning. Following ConNIe, we also check whether re-fitting on the selected edges reduces ranking bias relative to keeping the shrunken gate values.
The third check is the cost of sparsity on tail users. For users with fewer than ten interactions, we compare against GraphSASA-style densification. If degree-scaled $\lambda$ does not close the gap, we will report the conditions under which pruning and densification should be combined.
If the BOSS Zhipin data used by BISTRO can be obtained alongside its public code, we will use résumé revision dates as external ground truth for change points. We will test whether the learned gate flips cluster around those dates. This would be the most direct evidence that the model's switches track real preference change rather than optimization noise.
References
- GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks
- Locally Regularized Sparse Graph by Fast Proximal Gradient Descent
- On the Convexity of Latent Social Network Inference
- Robust Preference-Guided Denoising for Graph based Social Recommendation
- Adapting Job Recommendations to User Preference Drift with Behavioral-Semantic Fusion Learning
- Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
- Dynamic Graph Neural Networks for Sequential Recommendation
- Dynamic Graph Recommendation via Sparse Augmentation and Singular Adaptation
- Self-Attentive Sequential Recommendation
- Learning Graph ODE for Continuous-Time Sequential Recommendation
- Flow-based Time-aware Causal Structure Learning for Sequential Recommendation
- Learning Dynamic Preference Structure Embedding From Temporal Networks