Cold-Start Users as Points on a Community Simplex

A recommender system meeting a new user knows almost nothing about them: perhaps an age, an occupation, a sign-up profile, and no clicks or ratings. Most systems handle this by guessing from the profile and then, at some threshold of activity, switching to a model trained on the user's own behavior. This proposal treats the guess and the later personalized model as the same object. Each user is described by how much they belong to each of $K$ overlapping taste communities, written as a mixture whose weights sum to one. For example, a user might be 60% "arthouse," 30% "family animation," and 10% "action." A network trained on profile features predicts this mixture for a brand-new user. Each new interaction then updates the same mixture, plus a small personal correction. The cold-to-warm transition therefore needs no switch, gate, or threshold. It is a single posterior update whose prior comes from features and whose evidence comes from behavior.

The claim to be tested is specific. Communities discovered from the interaction graph, represented in a geometry built for mixtures, make a better prior for cold users than three alternatives: unconstrained feature embeddings, a single global initialization, or soft prototypes learned only from ratings. The same representation should also carry users from zero interactions to many without a loss in warm accuracy. If the graph-derived communities or the mixture geometry add nothing over a well-tuned feature encoder, the ablations below will show it.

Where this sits

Four lines of prior work bear on the idea. The first supplies the geometry. Aitchison Embeddings for Learning Compositional Graph Representations represents each graph node as a mixture over $K$ archetypal roles. It optimizes in isometric log-ratio coordinates, an invertible map from the mixture simplex to ordinary $(K-1)$-dimensional Euclidean space that preserves ratios between components. It reports a link-prediction AUC of 0.976 on AstroPh. That model is a transductive lookup table: a node with no edges has no embedding, and it uses no content features. We reuse its geometry and its code (AICoG implementation) but replace the lookup with a feature-conditioned prior, which makes it usable for cold nodes.

The second line motivates the problem. Hybrid cold-start methods move from group-level to individual models through discrete switches (An Improved Switching Hybrid Recommender System Using Naive Bayes Classifier and Collaborative Filtering) or through single-criterion demographic groups. Generalized Group Profiling for Content Customization reports that group granularity has an awkward "Goldilocks zone" and names multi-criteria grouping as open. Cold & Warm Net learns a gate between a profile expert and a history expert, using static user-group embeddings. All three decide how much to trust the group by a separate mechanism. In our model the amount of evidence decides it, and groups overlap by construction.

The third line is the closest competitor and must be faced directly. MAMO replaces the single global initialization of meta-learning with memory modules. Profile features attend over memory slots to produce a personalized parameter bias. That attention is already a soft mixture over $K$ prototypes. Our proposal differs in two testable ways. First, our prototypes are anchored to overlapping communities in the user co-interaction graph, not learned purely through the rating loss. Second, the mixture is the quantity updated as evidence arrives; it is not a one-time bias added before fine-tuning. Bayesian Matrix Factorization with Side Information and Dirichlet Process Mixtures combines feature regression with latent clusters, but it assigns each entity to one cluster per sample and relies on expensive Gibbs sampling. Fuzzy-Clustered Mixture-of-Experts uses soft feature-space memberships to gate experts on tabular data. It also constrains our design: its authors report that the soft prior improves calibration but acts as shrinkage that can hurt discrimination. Our design must therefore show it does not lose ranking quality.

The fourth line supplies the training trick and the main skeptic's baseline. DropoutNet randomly zeroes collaborative vectors during training so that a network learns to map content into the collaborative space. It reports a 6% Recall@100 gain over DeepMusic on CiteULike. We borrow its simulation of cold users during training, but apply it to truncated histories rather than dropped vectors. Inductive matrix completion (Recommending Tumblr Blogs to Follow with Inductive Matrix Completion) shows the cost of pure feature constraints: latent factors live in the span of the features. This is why our model keeps a free personal correction that grows with evidence.

Core Claim

Let $z_u \in \mathbb{R}^{K-1}$ be user $u$'s community mixture in log-ratio coordinates, and let $\pi_u = \mathrm{ilr}^{-1}(z_u)$ be the corresponding mixture weights on the simplex. Let $a_k$ be the taste vector of community $k$, $q_i$ item $i$'s embedding, $b_i$ an item bias, and $\delta_u$ a personal residual. The score is

$$s(u,i) = \Big(\sum_{k=1}^{K} \pi_{uk}\, a_k + \delta_u\Big)^{\top} q_i + b_i .$$

The prior is $z_u \sim \mathcal{N}(g(x_u), \Sigma(x_u))$, where $g$ and $\Sigma$ are networks over profile features $x_u$, and $\delta_u \sim \mathcal{N}(0, \tau^2 I)$. Given a user's observed interactions $D_u$, with $|D_u| = n$, the user's state is the maximum-a-posteriori estimate

$$(\hat z_u, \hat \delta_u) = \arg\min_{z,\delta}\; \sum_{(i,r)\in D_u} \ell\big(r, s(u,i)\big) + \tfrac12 (z-g(x_u))^{\top}\Sigma(x_u)^{-1}(z-g(x_u)) + \tfrac{1}{2\tau^2}\lVert\delta\rVert^2 .$$

With $n=0$, the estimate is the feature-predicted mixture. As $n$ grows, the data term dominates, and the residual $\delta_u$ is freed only once there is evidence to pay for it. The cold-to-warm handoff therefore follows from the likelihood rather than a tuned threshold. A Gaussian prior is natural in log-ratio coordinates because it corresponds to a logistic-normal distribution on the simplex. The same Gaussian placed directly on softmax weights, or on an unnormalized Euclidean embedding, would mix overall scale (here, user activity level) with the trade-offs between communities. That is the distortion Nakis et al. identify for role compositions.

Research Program

Training has two coupled parts. On warm training users, community mixtures are fit jointly against two objectives. One is the rating loss. The other is the AICoG link likelihood on a user–user co-interaction graph, where two users are connected with weight proportional to their shared items. This ties the $K$ archetypes to real overlapping structure in behavior, not to arbitrary slots. In parallel, the feature networks $g$ and $\Sigma$ are trained episodically, in the spirit of DropoutNet. In each episode, a warm user's history is truncated to $n$ interactions, with $n$ drawn from a schedule covering 0 to 50. The MAP update above is approximated by a few unrolled gradient steps, and the loss is taken on that user's held-out interactions. Gradients flow into $g$, $\Sigma$, the archetypes $a_k$, and the item side. The variance network $\Sigma$ matters: it lets the model trust profile features for some demographics and distrust them for others.

The analyses go beyond accuracy. We will measure how far the cold-predicted mixture $g(x_u)$ lies from the mixture the same user eventually reaches when warm, using Aitchison distance as a function of $n$. This tests whether features actually predict community position. We will also inspect whether the learned archetypes are interpretable, by listing the top items and enriched demographics for each community. The number of communities $K$ is a sweep, from 8 to 128, so that the granularity question raised by group profiling is answered empirically rather than by hand.

Baselines

The naive reference is global popularity, which falls back to the most-interacted items, and a feature-free version of our model in which every cold user receives the population-mean mixture. The Tumblr study found popularity hard to beat for transductive models under extreme sparsity, so any method that fails to beat it at $n=0$ is disqualified.

The strongest existing method for user cold start on MovieLens 1M is MAMO. It reports a mean absolute error of 0.8894 in the cold-user, cold-item scenario, against 0.9299 for MeLU (MAMO). We will rerun MAMO's released configuration on MAMO's own splits, confirm we reproduce roughly that number, and then give it the same feature inputs and the same hyperparameter budget as our model: 50 search trials per method, the same item embedding dimension, and matched wall-clock compute. Beating it means a lower MAE in that scenario under that protocol, with a confidence interval over five seeds that excludes MAMO's reproduced mean.

Cold & Warm Net is the strongest gated hybrid. It reports a 23.94% HR@100 improvement and a 35.22% NDCG@10 improvement for cold users on the proprietary Little-World data, and the file gives no number for its MovieLens 1M evaluation. We will reimplement its two experts, gate, and distillation with our encoders. It is the right competitor for the transition curve: a gate that learns when to trust history is exactly what our posterior update claims to make unnecessary. Beating it means matching or exceeding its ranking accuracy at every history length while showing no visible discontinuity in the curve. DropoutNet is run with the same warm base model as ours, since its cold performance is capped by that base. For DropoutNet on our chosen splits, only the relative CiteULike gain quoted above is available on Lacuna; there is no absolute value.

Four ablations isolate the key ingredients, each with everything else held fixed. The first replaces log-ratio coordinates with an unconstrained Euclidean latent vector of the same dimension, which tests the geometry. The second drops the graph link-likelihood term so that communities come from ratings alone, which tests whether graph structure matters and is the direct contrast with MAMO-style prototypes. The third replaces soft mixtures with hard K-means assignments of the kind TDRO uses (Temporally and Distributionally Robust Optimization for Cold-Start Recommendation), which tests overlap. The fourth replaces the posterior update with a learned gate between the cold and warm estimates, which tests the transition mechanism.

Why not just…?

Why not just use DropoutNet with a good feature encoder? This is a serious threat, because a sufficiently flexible network mapping features to an unconstrained embedding could learn everything a community mixture encodes. The Euclidean-latent ablation answers this directly, since it is essentially DropoutNet trained with our history-truncation episodes. If it matches the full model at every $n$, the simplex structure buys only interpretability, and we will report it that way. Our expectation is that the gap appears at small but nonzero $n$, roughly 1 to 10 interactions. In that range a well-shaped prior matters most, and DropoutNet has no principled way to weigh three ratings against a profile.

Why not just use MAMO's memory, which is already a soft mixture over prototypes? This is the closest prior work, and the difference has to show up in the numbers rather than the description. MAMO's slots are shaped only by the rating loss and act once as a bias before fine-tuning. Our communities are anchored to co-interaction structure and remain the object that evidence updates. The ablation without the graph term converts our model into something close to MAMO's prototype idea while keeping our update rule. Comparing it with both the full model and MAMO itself separates the effect of the graph from the effect of the update. If graph anchoring does not help, the contribution narrows to the transition mechanism.

Why not just soft-cluster the profile features, for example with fuzzy C-means, and use cluster-level priors? This avoids the graph entirely and is what FC-MoE-GR does for tabular data. Its own authors report that such priors improve calibration while hurting discrimination. The reason is that clusters found in feature space need not align with taste. MovieLens demographics are coarse, so feature-space clusters will likely group users who share an age bracket but not a taste. We will include fuzzy C-means memberships over features as a prior-source variant. We will also report the Aitchison-distance analysis, which shows directly whether graph-derived or feature-derived mixtures better predict where users end up when warm.

Why not just gate between a cold model and a warm model? Gating works in production, as Cold & Warm Net's online gains show. However, a gate is trained on the average user at each activity level and introduces a second learned component that can be miscalibrated during the handoff. The gate ablation and the per-$n$ curves will show whether removing it costs accuracy. If the gated version wins, the posterior formulation is not justified and we will say so.

Proposed Experiments

The main benchmark is MovieLens 1M, which has age, gender, and occupation features, using MAMO's four warm and cold user–item scenarios with MAE and RMSE for rating prediction plus HR@10 and NDCG@10 for ranking. Book-Crossing, again with MAMO's protocol, tests transfer to a sparser domain; MAMO reports gains there but the file gives no number. The XING RecSys 2017 data, with 1.5 million users and rich professional profiles, tests scale and whether richer features sharpen the community prior. There we follow DropoutNet's cold-user split and report Recall@100. For item cold start, the same model is mirrored to give items community mixtures predicted from titles and abstracts on CiteULike, with Recall@100 following DropoutNet's protocol.

The decisive check is the transition curve. Test users are held out, revealed $n \in \{0,1,2,3,5,10,20,50\}$ interactions in time order, and scored on the remainder, for every method. The proposal succeeds if three things hold together. It must lie at or above every baseline across the whole curve and beat MAMO's reproduced cold-user, cold-item MAE. Its warm-end accuracy must be within noise of an unconstrained matrix factorization trained on the same data, showing that the residual term removes the inductive-constraint penalty. And the graph and geometry ablations must each lose accuracy somewhere on the curve. A secondary check uses the chronological 8:1:1 splits from TDRO's Amazon Clothing benchmark. It asks whether $\Sigma(x_u)$ widens for users whose profiles resemble those of recently arrived users, which would show that the prior's uncertainty tracks shift rather than being fixed.