Age-Calibrated Co-Location Graphs for Sparse Recommendation

When two people keep turning up at the same place at the same time, they might be friends, or they might both just use the busiest station in town. Systems that rebuild missing social networks from location logs (GPS traces, card swipes, check-ins) need a way to tell these cases apart. Recommenders working with very sparse user-item data need that rebuilt network to be clean, because a noisy social graph hurts them. We propose using age homophily, the well-documented tendency of real social ties to link people of similar age, as an aggregate calibration signal for each place and time slot, not as a filter on individual pairs. The test is simple. If the pairs of people who co-occur at a venue are no more similar in age than the venue's visitors are overall, co-occurrence there is mostly chance. If they are much more similar, co-occurrence there is mostly social.

The opportunity comes from a gap between two literatures. Location-based tie inference has fought "chance co-location" with null models and diversity heuristics, but it has never had an independent signal of which co-occurrences are social. Social recommendation has learned to prune noisy ties, but it does so with the same sparse interaction data the recommender is trying to compensate for. Age labels on a small fraction of users can supply that independent signal. Because the signal is estimated per context and then applied to every user, it works for users whose age is unknown. It also does not punish ties between people of different ages, such as family members.

Core Claim

For each co-location context $c$ (a venue crossed with a time-of-day band and a dwell type), the age gaps among co-located pairs follow a mixture:

$$p_c(g) = \pi_c\, f_{\text{tie}}(g) + (1-\pi_c)\, f^{\text{null}}_c(g).$$

Here $g$ is the absolute age gap of a co-located pair. $f_{\text{tie}}$ is the gap distribution of genuine ties, shared across contexts. $f^{\text{null}}_c$ is the gap distribution obtained by randomly reshuffling ages among the users present in $c$; this preserves the venue's own age mix. $\pi_c$ is the fraction of co-occurrences in $c$ that are social. We estimate $\pi_c$ by maximum likelihood on age-labeled users, with empirical-Bayes shrinkage toward a global prior for small contexts. We then weight each candidate edge as

$$w_{uv} = \sum_c \pi_c\, n_{uv,c},$$

where $n_{uv,c}$ counts the co-occurrences of users $u$ and $v$ in context $c$. Age enters only through $\pi_c$. No pair is scored by its own age gap.

The claim is that this graph recovers true ties better, and helps sparse recommenders more, than three alternatives: a co-location graph thresholded against a movement null, one downweighted by venue popularity, or one pruned by preference similarity. The advantage should be largest for users with few interactions.

The mechanism has an honest failure mode. When a context's visitors are all nearly the same age, as in a first-year dining hall, $f^{\text{null}}_c$ and $f_{\text{tie}}$ coincide, so $\pi_c$ cannot be identified and falls back to the prior. We will report how often this happens rather than hide it.

Relation to Prior Work

Three works motivate the gap by inferring ties from co-occurrence without any demographic signal. HUBS builds an unsupervised student network from campus co-occurrences that exceed a statistical null and feeds it to attention-based sequence models, but treats every edge as equally influential. A Significance Test for Inferring Affiliation Networks from Spatio-Temporal Data tests sheep co-occurrences against Markovian movement nulls to remove "water hole effects." GEOSO separates co-visit diversity ("compatibility") from co-visit frequency ("commitment"), but is purely theoretical. All three ask whether a co-occurrence is more frequent than independent movement predicts. We ask instead whether co-occurrences in a context look like ties on an attribute that real ties are known to share. This second question can separate two places with identical traffic. Our method supplies a different evidence source and will be combined with their nulls, not substituted for them.

Cold start link prediction supplies a method we reuse. It downweights large shared groups in Adamic/Adar style and propagates link probabilities with Probabilistic Katz on Flickr. Popularity downweighting is the closest cheap rival to $\pi_c$, and we treat it as such below.

On the homophily side, MAIF infers Twitter ages from text plus community structure and reports that 93.6% of reciprocal friend pairs differ by under five years. That figure tells us $f_{\text{tie}}$ is sharply concentrated, which is what makes $\pi_c$ identifiable. Its text-based age inference also gives us a route to noisy age labels on review platforms. SHaRe constrains the design. It shows that explicit social graphs carry very little preference homophily (a ratio of 0.0154 on Yelp), and that pruning bad ties matters more than adding new ones. It does its pruning with warm-up item embeddings. In a sparse regime this is circular, because the graph is meant to supply what the interactions lack. Our calibration uses a signal that is independent of interactions.

On the recommender side, LightGCN, SimGCL, and RAU show that linear propagation and embedding-space regularization beat heavier graph networks under sparsity. We therefore hold the recommender fixed at these simple backbones and vary only the social graph.

Research Program

The first step builds the context-level estimator. We discretize visits into contexts and compute $f^{\text{null}}_c$ by within-context age permutation. We estimate $f_{\text{tie}}$ in one of two ways: from a small set of known ties where available, or jointly with all the $\pi_c$ by EM (expectation-maximization) under the constraint that $f_{\text{tie}}$ is shared across contexts. We report a bootstrap interval for each $\pi_c$ and flag contexts where the two distributions are too close to identify the mixture.

The second step combines $\pi_c$ with existing evidence. Candidate edges must first pass a HUBS-style null threshold. They are then weighted by $w_{uv}$, and the GEOSO compatibility count is kept as a separate edge feature.

The third step plugs the graph into a fixed recommender. We use LightGCN and SimGCL with social propagation, and SHaRe's homophilic contrastive term applied over our graph's edges. We train and tune each graph variant identically.

The fourth step is an analysis of which contexts receive high versus low $\pi_c$, and whether this ranking differs from venue popularity. If the two rankings correlate above roughly 0.9, the age signal adds little, and we will say so.

Baselines

The naive reference is the recommender with no social graph: LightGCN, and RAU-enhanced matrix factorization. RAU reports up to 41.76% relative Recall@20 gain over LightGCN on Gowalla, so it is a strong interaction-only bar that any social graph must clear.

The first graph baseline is the raw co-location graph built with fixed spatial and time windows and no null model. The second is the HUBS-style null-thresholded graph. The third is the Adamic/Adar popularity-downweighted graph followed by Probabilistic Katz from Cold start link prediction, which reports roughly 31% precision at 1% recall on Flickr; that is a tie-recovery reference, not a figure transferable to our data.

The strongest existing method is SHaRe-style rewiring applied to the null-thresholded co-location graph. SHaRe reports over 4% average relative gain in Recall@10 and NDCG@10 on Yelp across four base models. The file gives no absolute Recall figures for Yelp on Lacuna, so we will reproduce SHaRe ourselves on the same splits.

Every variant shares the same candidate edge set, recommender backbone, embedding size, epochs, and an equal hyperparameter-search budget. That budget covers SHaRe's pruning threshold and our shrinkage strength.

Two ablations isolate the key ingredient. The placebo replaces age with a randomly permuted attribute, so $\pi_c$ should collapse to the prior. The per-pair version filters edges by individual age gap instead of using context-level $\pi_c$.

Beating the field means two things. First, a statistically significant improvement over both popularity weighting and SHaRe in tie recovery (AUC and precision at fixed recall against held-out friend links), measured over five seeds. Second, a Recall@20 gain for the sparsest user quintile, with no loss in recall of cross-age true ties relative to the null-thresholded graph.

Why not just…?

Why not just filter co-location edges by each pair's age gap? This approach uses age only where it is observed. It mechanically removes intergenerational ties and makes recommendations age-segregated. The per-pair ablation measures this directly: we expect it to match our method on same-age ties and to lose on cross-age ties and on users without age labels. If it wins overall, the aggregate design is unnecessary, and the experiment would show that.

Why not just downweight popular venues, as Adamic/Adar does? This is the most serious threat, because busy hubs are often also places where ages mix. The two signals diverge in specific cases. A small pharmacy is unpopular but age-mixed, so its co-occurrences are chance. A large campus seminar room is popular but cohort-segregated, so its co-occurrences are social. The ranking-correlation analysis, together with head-to-head tie recovery, settles this. If popularity weighting matches $\pi_c$, we report that the cheaper method suffices.

Why not just prune with preference similarity, as SHaRe does? Under sparsity, warm-up embeddings for cold users are close to random, so preference pruning should degrade exactly where social information matters most. We test this by comparing the two methods across interaction-count quintiles. A crossover, where SHaRe wins for dense users and our method wins for sparse ones, would be an informative result in its own right. Combining the two methods is a natural final variant.

Why not just add age as a user feature in the recommender? Doing so does nothing to denoise the graph and is unavailable for unlabeled users. We include it as an extra input to every variant so that any gain from our graph is measured on top of it.

Proposed Experiments

No dataset in the file contains location traces, ages, ground-truth ties, and item interactions together, so the evaluation uses three complementary settings.

The first is semi-synthetic. We plant age-homophilous ties over real Gowalla check-in venues, inject chance co-presence at hubs, and vary how mixed each venue's age distribution is. This checks that $\pi_c$ is recovered where it is identifiable and falls back to the prior where it is not.

The second uses Yelp, which has check-ins, friend links, and reviews. Ages are inferred from review text for a labeled subset with a MAIF-style model. Because $\pi_c$ is an aggregate quantity, age noise should mostly shrink $\Delta$-type contrasts rather than reorder contexts. We will verify this by deliberately degrading the labels. Friend links serve as tie ground truth, and recommendation is evaluated with Recall@20 and NDCG@20 under a sparsified-interaction protocol.

The third is a campus smart-card setting like that of HUBS, with enrollment cohort as an age proxy and point-of-sale co-presence as in the White et al. lunch-buddy reconstruction. If access can be obtained, we evaluate on HUBS's downstream tasks, where it reports a 6.1% relative MSE improvement and a Macro-F1 of 0.9652.

The decisive checks, in order, are as follows. The placebo attribute must yield no gain. Our graph must beat popularity weighting on tie recovery on Yelp. The gain over SHaRe must grow as users' interaction counts shrink. Recall of cross-age ties must not fall. If the second check fails, the proposal reduces to a sophisticated reimplementation of popularity downweighting, and we would report it as a negative result.