People change what they want at different speeds at different levels of generality. Someone may stay a runner for a decade while cycling through a new brand of shoe every season. Most recommenders and preference-alignment methods that handle drift use a single forgetting rate: they down-weight old interactions by $\gamma^{\Delta t}$ no matter what those interactions say. As a result, "likes running gear" fades exactly as fast as "clicked these particular shoes once." We propose to place items on a learned ultrametric tree and measure drift separately at each level of that tree. An ultrametric tree is a hierarchy in which the distance between two items is set only by how high up their closest shared ancestor sits. Each level then gets its own forgetting rate, chosen from how much that level actually drifts in the data instead of from a guessed hyperparameter.
The key enabling fact is that on such a tree, the optimal-transport distance between two distributions over items splits exactly into a sum over tree edges. It is the edge weight times the change in mass below that edge. So how much a user's taste moved between two time windows breaks down cleanly into how much moved near the root (broad interests) and how much moved near the leaves (specific items). That breakdown gives a per-level drift estimate. A standard bias–variance argument then turns each estimate into a per-level memory length. The result is a recommender whose forgetting schedule is measured rather than tuned. It also gives a diagnostic for whether multi-scale drift is real in a given dataset at all.
Core Claim
Uniform temporal discounting is misspecified whenever drift rates differ across levels of generality. We claim that the drift rate at each level of a learned ultrametric item tree can be estimated from interaction logs alone, and that discounting each level by its own measured rate beats uniform discounting on time-split next-item prediction. The gain should concentrate in users whose measured fine-level drift exceeds their coarse-level drift. If the per-level rates turn out to be roughly equal on real data, uniform decay is vindicated, and we will report that.
Concretely, let $\mathcal{T}$ be an ultrametric tree over item clusters. For each node $v$ at depth $h$, a user's profile mass at time $t$ is
$$p_t(v) = \sum_{i \in \text{sub}(v),\, t_i < t} \gamma_h^{\,t - t_i},$$
normalized within each depth. Here $\text{sub}(v)$ is the set of past interactions $i$ with items under $v$, $t_i$ is the time of interaction $i$, and $\gamma_h \in (0,1]$ is the forgetting rate for depth $h$. Write $V_h$ for the average change in level-$h$ mass per unit time, measured on held-out windows of training data, and $\sigma_h^2$ for the per-step noise in that mass. Averaging over a memory of length $m$ incurs bias of roughly $V_h m$ and variance of roughly $\sigma_h^2/m$. Minimizing the squared error gives the memory length
$$m_h^\star \propto (\sigma_h / V_h)^{2/3}, \qquad \gamma_h = 1 - 1/m_h^\star.$$
This rule replaces the a priori variation budget that Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift needs to set its single discount. A candidate item is scored by summing its ancestors' profile masses, weighted by the tree's level heights. The scoring can be used on its own or fed into a sequential model as features.
Relation to Prior Work
Two lines of work motivate the gap. NS-DPO shows that exponential discounting of preference pairs rescues alignment under drift: reward accuracy stays at 60–70% where standard DPO falls to 50–55% after a preference flip. Its own analysis, however, depends on knowing the variation budget, and it discounts every preference dimension at the same rate. Dual Contrastive Transformer for Hierarchical Preference Modeling in Sequential Recommendation (HPM) supplies evidence that drift is multi-scale. It runs separate transformers over item IDs and fixed categories and gains 1.53–17.61% in Hit Rate (HR) and Normalized Discounted Cumulative Gain (NDCG) across six Amazon domains. HPM's hierarchy has exactly two predefined levels, and its time kernels are learned implicitly inside the network rather than measured per level. We differ from both by allowing arbitrary depth, by learning distances rather than taking them from a fixed taxonomy, and by estimating forgetting rates explicitly. Preference or Intent? Double Disentangled Collaborative Filtering similarly separates coarse intent from item preference, but uses flat latent channels and has no time component.
The method for building the tree comes from Learning Ultrametric Trees for Optimal Transport Regression. That paper fits ultrametric distance matrices by projected gradient descent with a minimum-spanning-tree projection and reports orders-of-magnitude lower tree-Wasserstein error than quadtrees. Its two known limitations constrain our design. The projection costs $O(k\cdot n^2)$, and it struggles on very sparse signals. We therefore fit the tree over a few thousand item clusters built from dense item co-occurrence embeddings, not over raw items, and we keep the tree fixed within an evaluation period. All the temporal dynamics live in the user's mass on the tree. This choice deliberately sidesteps the unsolved problem of streaming tree projection discussed around Dynamic Maintenance of Low-Stretch Probabilistic Tree Embeddings with Applications and Bridging Arbitrary and Tree Metrics via Differentiable Gromov Hyperbolicity. Efficient Online Clustering with Moving Costs shows that decomposing a problem level by level on a tree makes online tracking tractable. We borrow that intuition, not its regret machinery. Finally, Tree of Preferences for Diversified Recommendation builds an LLM-prompted taxonomy that is static, unweighted, and has no notion of time. It is a reference point for the claim that hierarchy helps, not a competitor on drift.
Research Program
The work has three parts. First, we build item trees in two ways: by ultrametric fitting over co-occurrence-embedding clusters, following Chen et al., and, where available, from the dataset's own category taxonomy. Second, we build the drift diagnostic. For every user, we split the training history into consecutive time windows and compute the tree-Wasserstein change between windows. We break that change down by depth, then pool it across users to get $V_h$ and $\sigma_h$. This produces a per-dataset drift curve over depth, which is a finding in its own right. A flat curve means uniform decay is adequate. A curve that rises toward the leaves means it is not. Third, we build the level-wise decayed profile scorer. We evaluate it both as a standalone recommender and as an extra input to SASRec, to test whether it adds information that self-attention does not already capture.
Baselines
The naive reference is an undiscounted profile on the same tree ($\gamma_h = 1$ at every level), alongside global popularity. Any method that cannot beat these is not modelling drift at all.
The most direct competitor is uniform scalar discounting: the same tree profile with one shared $\gamma$, which is the NS-DPO discounting rule moved to recommendation. It isolates our single change, which is level-dependence. It receives the same tuning budget as our method, a grid of the same size, selected on the same validation window. Beating it means statistically significant gains in NDCG@10 on time-split test data in at least four of the six Amazon domains. A second competitor is per-user learned decay in the style of the Time Decay Sampler from Learning Dynamic Preference Structure Embedding From Temporal Networks. It personalizes the forgetting rate by user rather than by level, and we will also test a version that varies by both user and level.
The strongest neural baselines are Self-Attentive Sequential Recommendation and HPM. SASRec reports average gains of 6.9% in Hit Rate and 9.6% in NDCG over Markov-chain, CNN, and RNN baselines on Amazon Beauty, Amazon Games, Steam, and MovieLens-1M. HPM reports 1.53–17.61% HR/NDCG gains over 13 baselines, including SASRec, on six Amazon domains. No reported absolute HR@10 or NDCG@10 value was found on Lacuna for either model, so we will re-run both under our time-based split with their published settings and an equal hyperparameter-search budget. Beating HPM outright is not required. The decisive neural comparison is whether SASRec plus our level-wise features beats SASRec alone and SASRec plus uniformly decayed features, both trained with identical compute.
Three ablations isolate the key ingredients. The first uses the fixed dataset taxonomy instead of the learned ultrametric tree. The second keeps the shape of the learned tree but randomly reassigns leaves. The third chooses per-level $\gamma_h$ by direct cross-validation instead of from the drift estimator, which tests whether the estimator is a usable substitute for tuning.
Why not just…?
Why not just use one tuned exponential decay? NS-DPO shows that one discount goes a long way. HPM's gains from separating category and item streams suggest it leaves signal unused, but that is indirect evidence. The direct test is the uniform-γ baseline on the identical tree. If the measured drift curve is flat on a dataset, we predict no gain there, and we will report that.
Why not just let self-attention learn what to forget? SASRec can in principle weight old versus recent items adaptively. However, it sees one flat sequence and has no notion of which level of generality an interaction belongs to. The SASRec-plus-features experiment settles this question. If level-wise features add nothing on top of attention, our method is redundant as a model and survives only as a diagnostic.
Why not just use the existing category tree with per-level decay? This is the most serious threat, and it may be enough. Amazon's taxonomy is free and interpretable. The fixed-taxonomy ablation is a full experimental arm, not a strawman. If it matches the learned tree, we will say the contribution is level-wise drift estimation and that the learned metric is unnecessary. We expect the learned tree to win only where the taxonomy is shallow or poorly matched to how people actually co-purchase, as the quadtree comparison in Chen et al. hints.
Why not just use a two-level model like HPM? If drift varies only between "category" and "item," a two-level model captures everything. Our per-depth drift curve tests this directly. Deeper trees are justified only if the intermediate levels show drift rates distinct from both ends.
Proposed Experiments
We use HPM's six Amazon domains (Grocery, Sports, Beauty, Clothing, Cellphones, Toys), plus MovieLens-1M and Steam from SASRec. Splits are strictly chronological: train, validation, and test periods are separated by global timestamps, because leave-last-out splits mix time periods across users and hide drift. Metrics are HR@10 and NDCG@10, also reported separately for users grouped by the ratio of their fine-level to coarse-level drift. Our hypothesis predicts gains that grow with that ratio, and that stratified trend matters more than the average.
The decisive check is a controlled injection. In MovieLens-1M, after a cutoff time we either permute leaves within subtrees, which creates pure fine-level drift, or swap top-level branches, which creates pure coarse-level drift. The estimator should place each injected drift at the correct depth. Level-wise discounting should beat the best single γ in both cases, whereas any single γ should fit one case and lose on the other. If the estimator cannot localize injected drift, or if a single tuned γ matches level-wise discounting on real data, the core claim fails.
References
- Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
- Dual Contrastive Transformer for Hierarchical Preference Modeling in Sequential Recommendation
- Preference or Intent? Double Disentangled Collaborative Filtering
- Learning Ultrametric Trees for Optimal Transport Regression
- Dynamic Maintenance of Low-Stretch Probabilistic Tree Embeddings with Applications
- Bridging Arbitrary and Tree Metrics via Differentiable Gromov Hyperbolicity
- Efficient Online Clustering with Moving Costs
- Tree of Preferences for Diversified Recommendation
- Learning Dynamic Preference Structure Embedding From Temporal Networks
- Self-Attentive Sequential Recommendation