A recommender has to decide how many distinct "interests" a user has before it can fetch candidates for each one. Most multi-interest systems fix that number in advance, or tie it to history length. A different approach models each user as a smooth preference surface over item embeddings using a Gaussian process (GP). A GP is a Bayesian regression model whose predictions come with uncertainty, and when fit to a user's history its surface grows one peak per cluster of interactions. This handles any number of interests without a fixed count. The cost is that exact GP fitting is cubic in the number of interactions, so the published version caps histories at about 160 items. That cap excludes the heavy, varied users who most need several interests.
This note proposes one object that addresses both problems. We fit each user's GP through a small set of "inducing points", representative items that summarise the history, and let the size of that set be chosen per user by the model's own evidence. The inducing points are then used directly as the user's interest vectors for candidate retrieval. The number of interests, the cost of fitting the user, and the number of retrieval queries become the same quantity, set by how spread out the user's behaviour actually is rather than by how long it is.
Core Claim
GPR4DUR fits an exact GP per user over pre-trained item embeddings. It reports the highest interest-wise coverage against 11 baselines, including MIND and ComiRec, on Amazon, MovieLens-1M and Taobao. It is the only model with a significant tail-exposure gain on Amazon. It stays cubic in history length, keeps users isolated from one another, and does not say how to retrieve from a catalog without scoring every item.
Our claim is that a sparse GP whose inducing set is chosen per user, with a fixed price per inducing point, does three things. It matches exact GPR4DUR on users short enough to fit exactly. It beats truncated GPR4DUR on long-history users. It yields a learned interest count that tracks semantic spread (category entropy) rather than history length. MIND's rule, by contrast, sets the interest count to roughly $\log_2$ of history length (MIND).
The main risk is the one raised by Approximation-Aware Bayesian Optimization: a finite inducing budget smooths away sharp local peaks, and niche interests are exactly such peaks. We argue this failure is specific to fixed budgets. The selection rule below places points where the approximation leaves the most unexplained variance. An isolated niche cluster, far from the user's other items, has large leftover variance and therefore gets its own point early. This is a testable claim, not an assumption.
Research Program
For user $u$ with interacted item embeddings $H_u$, we select the inducing set $Z_u \subseteq H_u$ by maximising
$$\mathcal{L}_u(Z_u) = \mathrm{ELBO}_u(Z_u) - \lambda\,|Z_u|.$$
Here $\mathrm{ELBO}_u$ is the collapsed variational lower bound on the log marginal likelihood of the sparse GP given inducing set $Z_u$, and $\lambda$ is a single global price, in nats, charged for each interest vector served.
We optimise this greedily. We add the history item with the largest remaining posterior variance, in the style of a pivoted Cholesky factorisation, and stop when the ELBO gain falls below $\lambda$. Per-user cost is $O(N_u m_u^2)$, where $N_u$ is history length and $m_u = |Z_u|$ is the selected count. This is the per-user analogue of the Poisson-point-process prior over inducing subsets in Probabilistic selection of inducing points in sparse Gaussian processes. That paper supplies the principle that the inducing count should shrink for simple functions and grow for complex ones. We replace its score-function (REINFORCE) optimisation with a deterministic greedy rule, because gradient variance across millions of users is the open concern the file flags. We will also run the stochastic version on a subset to check that greedy selection does not lose accuracy.
Kernel hyperparameters (lengthscale, noise, and the prior-mean offset) are shared across all users. They are estimated once on a small sample of users and then frozen. This follows the decoupling that lets GPnn scale to 1.6 million points: hyperparameters are learned cheaply on a subset, and prediction is local. It also gives users a minimal form of collaboration that GPR4DUR lacks, since everyone shares one learned geometry.
Retrieval works as follows. The posterior mean has the form $\mu_u(x) = \sum_{j} \beta_{uj}\, k(x, z_{uj})$, where $z_{uj}$ is the $j$-th inducing point, $\beta_{uj}$ its learned weight, and $k$ the kernel. Each inducing point with positive weight issues one approximate-nearest-neighbour query. Its candidate budget is proportional to its weight. The union of candidates is then reranked by the full posterior using an upper-confidence-bound score, the same exploration rule GPR4DUR uses. The number of retrieval queries therefore equals the learned interest count. We will analyse how $m_u$ relates to history length and to category entropy, and whether $\lambda$ can be set once, globally, as a compute–accuracy dial.
Baselines
Truncated GPR4DUR. The "do nothing" reference is GPR4DUR exactly as published, with histories truncated to the most recent 160 items. It is the strongest existing method on the coverage and tail metrics we care about. We run it on the same embeddings, the same splits, and the same shared-hyperparameter tuning budget. On users with at most 160 interactions, where no truncation happens, "beating" it means matching it within noise. On users with more than 160 interactions, it means a significant gain in Recall@50 and interest-wise coverage. The file does not transcribe the paper's values: no reported result found on Lacuna for GPR4DUR's numeric interest-wise coverage, Recall or NDCG. We will therefore reproduce it ourselves.
Fixed-budget sparse GP. This is the ablation that isolates the adaptive budget. It uses the same sparse GP and greedy selection, but every user gets the same inducing count $m$. We set $m$ so that the average compute equals the adaptive model's average. If adaptive allocation does not beat this at equal mean $m$, the proposal's key ingredient fails.
Point-based multi-interest models. These are MIND with its log-length rule (MIND), which reported gains of over 60% in HitRate@50 over YouTube DNN on Tmall and found 5 to 7 interests optimal there, and DMI-GNN (DMI-GNN), which reports HR@20 of 40.26 and 94.00% coverage on Tmall. These are trained with the authors' recommended settings and an equal hyperparameter search budget.
Local GP without a summary. This is a GPnn-style model: each candidate is scored with a GP conditioned only on the user's nearest history items, and candidates are fetched by item-to-item nearest-neighbour search from every history item.
Why not just…?
Why not just cluster the history with k-means (choosing k by an information criterion) and use the centroids as interests? This is the most serious threat, since it gives adaptive cardinality cheaply. Our view is that clustering criteria reward compressing dense regions. The ELBO-with-price rule instead rewards reducing predictive variance, which should protect small, isolated niches. We will run BIC-selected k-means centroids, both as interest vectors and as fixed inducing points, at matched mean count. If they match our method on the niche-recall test below, we will report that the GP machinery is unnecessary for choosing the interest count, and that its value is confined to the uncertainty used for exploration.
Why not just truncate or subsample history and keep exact GPR? Truncation is our first baseline. The prediction is that it loses long-dormant interests. We test this directly by holding out items from categories that appear only in the truncated-away part of a user's history.
Why not just use a local nearest-neighbour GP, which GPnn shows is fast and robust? Local GPs score candidates well, but they have no compact summary of the user. Retrieval then needs one query per history item, $N_u$ in total, rather than $m_u$. GPnn also degrades in high-dimensional spaces. We compare recall at an equal number of approximate-nearest-neighbour queries and report end-to-end latency.
Proposed Experiments
We use GPR4DUR's datasets (Amazon, MovieLens-1M, Taobao) with its interest-wise coverage, tail exposure, Recall@50 and NDCG@50 metrics. We add Amazon Office from SIGMA, chosen for its long sequences, where SIGMA reports a 9.25% relative Recall@20 gain over its best baseline. All results are stratified by history length (≤160, 160–1000, >1000) and by category-entropy quintile, mirroring the diversity tiers of Can One Embedding Fit All?.
Decisive check 1: interest count follows spread, not length. The Spearman correlation of the learned $m_u$ with category entropy, controlling for history length, should clearly exceed its correlation with length alone. MIND's rule has this reversed.
Decisive check 2: niche peaks survive compression. We hold out items from categories that make up less than 5% of a user's history. The adaptive model's recall on these items should be at least that of exact GPR on users short enough to fit exactly, and above the fixed-budget and k-means variants at equal mean $m$. If it falls short, the smoothing concern from Approximation-Aware Bayesian Optimization applies here, and the selection rule must be redesigned.
Decisive check 3: cost. We measure per-user fit time and retrieval latency as functions of $N_u$, and trace the curve of accuracy against compute as $\lambda$ is varied. The adaptive model should dominate the fixed-$m$ curve.
References
- Density-based User Representation using Gaussian Process Regression for Multi-interest Personalized Retrieval
- Multi-Interest Network with Dynamic Routing for Recommendation at Tmall
- Approximation-Aware Bayesian Optimization
- Probabilistic selection of inducing points in sparse Gaussian processes
- Leveraging Locality and Robustness to Achieve Massively Scalable Gaussian Process Regression
- Dynamic Multi-Interest Graph Neural Network for Session-Based Recommendation
- Semantic Gaussian Mixture Variational Autoencoder for Sequential Recommendation
- Can One Embedding Fit All? A Multi-Interest Learning Paradigm Towards Improving User Interest Diversity Fairness