Consistency Residuals as Drift Detectors in Preference Taxonomies

A user's interests are nested. Someone who spends a month on "cameras" moves between lenses, tripods, and bodies week to week, while the fact that they care about photography at all may hold for years. Recommenders and preference models usually track this with one memory length: a single forgetting rate, a single window, or one attention mechanism over the whole history. We propose to track a user's interest as a distribution over an item taxonomy (category, subcategory, item). Each level gets its own forgetting rate, and the estimates are forced to add up across levels, so the share of attention on "cameras" equals the sum of the shares on its subcategories. The tool for this is the least-squares reconciliation machinery that census agencies use to make noisy county counts sum to state counts.

The central observation is that the reconciliation step yields a useful signal before it corrects anything. Estimate a parent category with long memory and its children with short memory, and in a stable period the children's sum should match the parent up to predictable noise. When the gap exceeds that noise, the long memory at the parent level is stale: drift is happening between categories, not only inside them. This mismatch, which we call the consistency residual, gives a data-driven way to set per-level forgetting rates and to locate where in the taxonomy a user's taste is moving. After that, reconciliation borrows strength from well-estimated coarse levels to stabilize sparse, fast-moving leaf estimates.

Core Claim

Hierarchical consistency constraints have been treated only as a property to restore after noise is injected. We claim that, under temporal preference drift, violations of these constraints measured before correction are an informative statistic. They pick the memory length at each taxonomy level without knowing how much drift there is, and they tell apart within-category churn from between-category shifts. Combined with exact tree reconciliation, this should track user interest shares better than a single discount factor, better than independent per-level estimates, and better than an unstructured sequence model alone. The gain should be largest for users whose histories are sparse and whose drift sits at coarse levels.

What Each Direction Contributes, and What Is Missing

The hierarchical consistency literature supplies the method. The 2020 Census TopDown Algorithm (The 2020 Census Disclosure Avoidance System TopDown Algorithm) measures noisy counts at every level of a geographic tree. It then solves a weighted least-squares problem, level by level, so that children sum to their parent. Differential Privacy of Hierarchical Census Data: An Optimization Approach shows that such problems admit an exact dynamic program over the tree that runs up to two orders of magnitude faster than generic solvers, with zero constraint violations. Differentially Private Hierarchical Group Size Estimation shows that variance-weighted reconciliation cut national-level error by over 25x relative to bottom-up aggregation. That result is the strength-borrowing effect we want to reuse. Differentially Private Range Queries with Correlated Input Perturbation instead achieves consistency by construction, through correlated noise. That route is unavailable here, because our "noise" is sampling variability in user behavior, which we do not design. All of this work treats the tree as holding a static quantity. The one temporal member, OptStream, reconciles an electricity stream with hierarchical sums. It uses one fixed temporal treatment and never asks whether different levels change at different speeds.

The preference drift literature supplies the problem and a known weakness. Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift (NS-DPO) weights each training pair by $\gamma^{T-t_i-1}$ and shows this rescues reward accuracy from about 50–55% to above 60–70% when preferences flip. It also states that the best $\gamma$ depends on an unknown total variation budget, and that one scalar weight treats all data at a given time identically. Sequential Recommendation with User Evolving Preference Decomposition (SPLIT) splits a user's history into parallel interest threads, using time gaps to trigger splits. Its threads are flat, and nothing guarantees they aggregate back to the user's overall state. A Shared Low-Rank Adaptation Approach to Personalized RLHF supports our structural premise, since preference models split cleanly into a stable shared component and a variable individual one. Its setting is static, though. Hierarchical Conversational Preference Elicitation with Bandit Feedback is the closest bridge: it uses an item taxonomy to speed up preference learning. It assumes stationary rewards and does not enforce consistency between levels.

That last paper also constrains us. Its user study finds that people value a category by the average of its top items, not the mean over all of them. So summing utilities up a tree would be wrong. We therefore aggregate attention shares: the fraction of a user's recent interactions falling under each node. These are additive by definition, so additive reconciliation is correct for them. Utility aggregation, which is non-additive, stays out of scope and is flagged as follow-up.

Research Program

For a user, let $n_v(s)$ be the number of interactions at time step $s$ falling under taxonomy node $v$, and $N(s)$ the total. At level $\ell$ of the tree we form a discounted share

$$\hat p_v(t) = \frac{\sum_{s \le t} \gamma_\ell^{\,t-s}\, n_v(s)}{\sum_{s \le t} \gamma_\ell^{\,t-s}\, N(s)},$$

where $\gamma_\ell \in (0,1]$ is the forgetting rate for level $\ell$. The variance of $\hat p_v$, written $\sigma_v^2$, is approximately $\hat p_v(1-\hat p_v)$ divided by the effective sample size implied by the weights. This makes the trade-off explicit: short memory tracks change but is noisy, especially at sparse leaves.

Two things are computed from these estimates.

The first is the consistency residual at each parent $p$ with children $c$: the gap $\hat p_p - \sum_c \hat p_c$, divided by its standard deviation under no drift. Parent and children share events but use different weights, so this standard deviation is computed exactly from the two weight sequences rather than assumed independent. Under stationarity the residual has mean zero. When the parent's memory is longer than its children's and attention moves between parents, the parent lags and the residual becomes biased.

We choose each $\gamma_\ell$ online from a small grid. The rule takes the longest memory whose recent standardized residuals stay within a threshold calibrated on stationary segments, where the threshold is set by permuting events within short time blocks. This targets the problem NS-DPO leaves open: memory is set per level from observable evidence instead of from a guessed variation budget.

The second is reconciliation. We solve

$$\min_{q \ge 0} \; \sum_{v} \frac{(q_v - \hat p_v)^2}{\sigma_v^2} \quad \text{subject to} \quad q_p = \sum_{c \in \text{children}(p)} q_c \ \text{for every parent } p,$$

where $q_v$ are the reconciled shares. On a tree this problem is solved exactly in linear time by the dynamic-programming approach of Differential Privacy of Hierarchical Census Data: An Optimization Approach, adapted from integer counts to continuous shares. Down-weighting noisy nodes lets a sparse leaf lean on its well-estimated parent. The census literature warns that non-negativity biases sparse leaves upward (The 2020 Census Disclosure Avoidance System TopDown Algorithm), so we report leaf bias with and without the constraint.

To turn shares into recommendations, the reconciled masses along an item's taxonomy path are added as a log-prior to the scores of a fixed sequence model, SASRec (Self-Attentive Sequential Recommendation). This isolates the value of the taxonomy tracker from the choice of backbone.

A side benefit comes from the census setting. Reconciliation is post-processing, so the same pipeline applies unchanged when node counts are privatized before release, for example when a platform publishes category-trend statistics. We include one such run to show the estimator degrades gracefully, but privacy is not the core claim.

Baselines

The naive reference points are global popularity and an undiscounted per-user share. Popularity is not a straw man in session data: on Amazon-M2 it reached overall MRR@100 of 0.2426 against 0.2159 for the best deep model, STAMP, while losing on Recall@100 (0.4243 against 0.5591 for CORE) (Amazon-M2). Our method must beat both of popularity's numbers on Amazon-M2 to count as useful there.

The strongest existing sequence models are SASRec and SPLIT. SASRec learns its own weighting over recent and distant items and reported average improvements of 6.9% in hit rate and 9.6% in NDCG over earlier sequence models (Self-Attentive Sequential Recommendation). SPLIT is the closest competitor on the drift question, since it explicitly decomposes evolving interests. It reported average relative gains of +10.32% NDCG@5 and +10.08% Recall@5 over BPR, GRU4Rec and MCPRN across six datasets (Sequential Recommendation with User Evolving Preference Decomposition). The file reports only relative gains, so we found no absolute NDCG@5 value on Lacuna for SPLIT or SASRec on these datasets, and we will reproduce both. All models get the same splits, the same item vocabularies, and an equal hyperparameter-search budget measured in trials.

Beating SPLIT means a statistically significant NDCG@5 gain over the best of SASRec and SPLIT on at least four of the six datasets, when our prior is added to SASRec. We will also report the same prior added to SPLIT.

The ablations isolate each ingredient. The first uses a single $\gamma$ shared by all levels, tuned on validation data; this is NS-DPO-style discounting (Right Now, Wrong Then) transplanted to shares. The second uses per-level $\gamma$ tuned by grid search on validation data rather than chosen by residuals, which tests whether the residual rule matters or whether per-level memory alone does the work. The third uses independent per-level estimates without reconciliation. The fourth reconciles with a fixed $\gamma$ and no adaptation.

Why not just…?

Why not just let SASRec learn the timescales? Self-attention with positional embeddings can in principle weight recent items heavily within a category and old items for the category itself. On dense datasets such as MovieLens-1M and Steam it may do exactly this, and that is a real threat. Our bet is that sparse users do not give attention enough signal to separate the two timescales, while the tree prior pools across levels explicitly. We settle this by reporting results stratified by user history length, and with a synthetic benchmark in which drift is injected at a known level. If SASRec plus a category embedding matches us on sparse users and on the synthetic localization task, the explicit tracker is not justified.

Why not just tune one forgetting factor, or use a sliding window? NS-DPO itself shows that the right $\gamma$ depends on an unknown drift budget. It also shows that ignoring drift collapses accuracy to near random under abrupt flips (Right Now, Wrong Then). A window short enough for leaf churn also discards the long-lived interests that SPLIT's thread analysis shows users return to. The single-$\gamma$ ablation tests this directly. If a validation-tuned single $\gamma$ matches per-level memory, our structural premise is wrong.

Why not just shrink each child toward its parent and run a two-window change detector at each node? This is the most serious alternative. Hierarchical shrinkage plus per-node change detection is simple, and it resembles hierarchical forecast reconciliation from the time-series statistics literature (not indexed in this file). Our method differs in two ways. The residual compares levels rather than windows at a single node, which lets it assign drift to a level. And reconciliation produces shares that are exactly coherent, which shrinkage does not. We will implement the shrinkage-plus-detector baseline under the same budget. If it matches both our recommendation accuracy and our drift-localization accuracy, we will report that the contribution reduces to a coherent version of known tools.

Proposed Experiments

The first experiment uses synthetic users on a three-level taxonomy with ground-truth share trajectories. Drift is injected at the root, the middle level, or the leaves, abruptly or gradually. We measure tracking error of the reconciled shares, the accuracy of assigning drift to the correct level, and the false alarm rate on stationary segments. The decisive check is whether residual-chosen $\gamma_\ell$ come within a small margin of oracle per-level rates, which are known from the simulator, while the single-$\gamma$ baseline cannot.

The second experiment uses real sequential data with natural taxonomies: Amazon Video, Garden and Music, Foursquare venue categories, MovieLens-1M genres, Steam tags, and Amazon-M2 sessions. We measure NDCG@5 and Recall@5 for next-item prediction, and KL divergence between predicted and realized category shares in the next time window. All results are stratified by history length.

The third experiment adds Laplace noise to the node counts to mimic a private trend release, in the spirit of OptStream, and checks that residual calibration and reconciliation still hold.

The overall claim fails if per-level memory chosen by residuals gives no gain over a tuned single $\gamma$ on the synthetic data, or if the gains on real data vanish for sparse users once the shrinkage-plus-detector baseline is included.