When Friends From Different Circles Agree: Structural Diversity as the Trust Signal in Hybrid Recommenders

Recommenders that mix two kinds of evidence, what your contacts already did and how well an item's content matches your history, must decide how much to trust each one. Today that decision usually rests on how much data a user has. Research on how information spreads through social networks points to a better cue. People are much more likely to adopt something when the friends who exposed them to it come from several separate social circles than when the same number of exposures all come from one tight-knit group. Inside one group, repeated exposures go stale quickly. Exposures from independent circles act as independent votes that the item appeals across groups. This proposal turns that observation into the trust weight of a hybrid recommender. When the evidence from contacts spans many independent circles, the model should trust it even if the content looks off-profile. When it all comes from one dense circle, the model should fall back on content match.

The practical payoff is serendipity. Prior work estimates that 27–30% of resharing on microblogs is "unexpected relevance": items far from a user's usual topics that the user nonetheless finds relevant. A purely content-based ranker will bury these items, and a ranker that just counts adopters will overrate echo-chamber items. We propose a leakage-safe test, run on Sina Weibo diffusion data, of whether the shape of social exposure, beyond its volume, tells a hybrid model when to override content. Following the baseline-first discipline of recent audits, the test is built to report a null result honestly if simpler models already capture the effect.

Core Claim

For a user $u$ and a candidate item $i$ (a message they could forward), divide the contacts who already adopted $i$ into the distinct social contexts of $u$'s ego network, meaning the network of $u$'s contacts and the links among them. Our claim has two parts. First, the number of distinct contexts predicts adoption beyond the raw count of adopters, beyond collaborative embeddings, and beyond a calibrated per-user base rate. Second, and this is the new part, structural diversity should moderate how much the model trusts social evidence relative to content evidence. Concretely, the model we test is

$$ \Pr(\text{adopt}_{u,i}) = \sigma\Big(b_u + g(D_{u,i})\,\sum_{c \in C_{u,i}} f(n_c) + \big(1-g(D_{u,i})\big)\, s_{\text{content}}(u,i)\Big), $$

The symbols are as follows. $\sigma$ is the logistic function. $b_u$ is a calibrated per-user base rate. $C_{u,i}$ is the set of $u$'s contexts that contain at least one adopter of $i$, and $D_{u,i}=|C_{u,i}|$ is how many there are. $n_c$ is the number of adopters inside context $c$. $f$ is a learned saturating function; we expect it to flatten after about two exposures, following the within-community peak reported on Weibo. $s_{\text{content}}$ is a content-match score. $g$ is a learned monotone gate. The prediction is that $g$ increases with $D$, and that the cases where the gate overrides a low content score are mainly the serendipitous adoptions.

Relation to Prior Work

Three papers motivate the claim. Cumulative Effect in Information Diffusion shows on Weibo that forwarding probability peaks at two exposures and then declines. This is the reason for saturating within a context rather than counting. The Survey of social influence analysis and modeling summarizes evidence that the number of unconnected circles predicts adoption better than the number of influencers. Unexpected Relevance defines serendipity with likelihood-ratio tests over unigram language models and links it to weak, non-reciprocal ties. That study is purely observational and builds no ranker. We reuse its labelling procedure as an evaluation target rather than as a model.

Two papers come closest to our method. An Empirical Evaluation Of Social Influence Metrics already puts Louvain counts of active communities into Random Forest adoption classifiers. It frames the problem as balanced binary classification, with no content arm and no notion of trusting one evidence source over another. Our proposal differs in three ways: the interaction between diversity and content, per-user calibrated ranking, and a temporal split that removes leakage. Toward Order-of-Magnitude Cascade Prediction uses diversity features to predict cascade-level virality, not per-user preferences.

On the recommender side, Llama4Rec shifts trust between collaborative and language-model signals using a long-tail coefficient based on data volume. We replace volume with exposure topology. CDiff4Rec mixes real-user and text "pseudo-user" neighbors inside a denoising diffusion generator. It shares only the word "diffusion" with this proposal: its neighbors are static cosine-similarity sets with no social topology.

Two papers act as constraints. Do LLMs Memorize Recommendation Datasets? finds GPT-4o reproduces 80.76% of MovieLens-1M titles. This is why our content arm defaults to a model trained only on in-split text. More Context Is Not Validation shows a listener-grouped calibrated prior (Brier 0.216) beating a grounded LLM hybrid (Brier 0.235). This sets the bar we must clear and the audit style we adopt. LLM-Augmented Agent-Based Simulation of Cyber Social Agents studies how agent diversity shapes diffusion in simulation. Our observational estimates of $f$ and $g$ could serve as empirical targets that such simulators must reproduce, but we do not build a simulator here.

Research Program

We use the WISE 2012 Weibo data (16.6M messages, 58.6M users, 265.5M follow links) from Cumulative Effect in Information Diffusion. We split it by time. The follow graph, ego-network contexts, user base rates, collaborative embeddings, and content model are all fitted on data before a cutoff. Prediction targets are forwarding decisions after the cutoff.

An exposure event is a pair $(u,i)$ in which at least one followee of $u$ posted or forwarded $i$ within a time window. Weibo has no impression logs, a limitation Cumulative Effect shares. We therefore treat following as exposure, report sensitivity to the window length, and condition on recency, since users echo the most recent source 85.5% of the time.

We compare three definitions of "context" for computing $D_{u,i}$: connected components of the ego network, Louvain communities, and maximal connected $k$-trusses (subgraphs in which every link sits in at least $k-2$ triangles), as in Truss-Based Structural Diversity Search. Their TSD-index makes the truss computation feasible at this scale. Because contexts are computed once on the frozen pre-cutoff graph, the open problem of maintaining them on a changing graph does not arise.

The content score comes from a text model trained on each user's pre-cutoff posts. In a secondary arm we use a sentence-embedding model, after checking for memorization of the 2012 messages. Serendipity labels for held-out adoptions come from re-running the unexpectedness-plus-relevance likelihood-ratio tests of Unexpected Relevance on pre-cutoff text.

Baselines

The naive reference point is a calibrated per-user forwarding rate. It is fitted with five-fold user-grouped calibration, mirroring the prior that beat the LLM hybrid in More Context Is Not Validation (Brier 0.216 vs. 0.235 on audio listeners). We found no reported result on Lacuna for this prior on Weibo forwarding, so we compute it ourselves. Every other model must beat it in Brier score with a bootstrap interval that excludes zero, or it does not count as an improvement.

The strongest existing method for this question is the multi-measurement Random Forest of An Empirical Evaluation Of Social Influence Metrics. It reports Precision 0.95 and F1 0.948, against 0.679 and 0.622 for the LRC-Q baseline, on balanced Weibo samples. We reimplement it with its features: Louvain active-community count, influence-locality measures, timing, cascade depth, and hashtag/URL/mention metadata. We add our content score so that it sees exactly the same information as our model. It is trained and evaluated on our temporal split with the natural class ratio, and gets the same hyperparameter search budget. We expect its reported F1 to fall substantially under the natural class ratio and the temporal split. Our comparison is therefore against the reimplementation, not the published number. A gradient-boosted tree with identical features is run alongside it, because trees can learn the diversity-by-content interaction implicitly.

The collaborative baseline is LightGCN on the user–author forwarding graph, combined with the content score by learned late fusion in the style of Paulheim et al.. This tests whether latent embeddings already absorb community structure.

Two ablations isolate the key ingredient. The first replaces $D_{u,i}$ and the per-context sums with the raw adopter count, keeping the same functional form. The second fixes $g$ to a constant, which turns our model into a plain additive hybrid.

Beating the baselines means two things. Overall, Brier and NDCG@10 over each user's exposure set must improve over the boosted tree with matched features. On the serendipitous subset, recall@10 must improve over the constant-gate ablation. If the boosted tree matches us, the contribution reduces to the features and the measured interaction, and we will report it that way.

Why not just…?

Why not just count exposing friends and let a nonlinear model learn saturation? A count cannot tell three friends from one clique apart from three friends from three circles. That distinction is the whole hypothesis. An Empirical Evaluation Of Social Influence Metrics found active-community counts beat raw neighbor counts. The decisive check here is the matched comparison described below, which holds the adopter count fixed and varies the number of contexts.

Why not just hand everything to a gradient-boosted tree? This is a real threat. A tree given the diversity and content features may learn the gate on its own. We include it as a primary baseline. If it matches the gated model, we will say the explicit gate adds interpretability but not accuracy. The scientific result, that the trust given to social evidence rises with $D$, can still be checked through partial-dependence analysis of the tree.

Why not just diversify the list with a determinantal point process? A determinantal point process is a probabilistic model that rewards item sets whose members are dissimilar. The low-rank version in Low-Rank Factorization of Determinantal Point Processes makes it fast, but it spreads items apart by their own traits. It says nothing about whether independent social circles vouch for a given item. It answers a different question, list diversity rather than adoption likelihood, and could be stacked on top of either model.

Why not just use an LLM to judge relevance and generate serendipity? Do LLMs Memorize Recommendation Datasets? shows memorization skewed toward popular items (89% of head items vs. 64% of tail items). More Context Is Not Validation shows added LLM context failing a calibrated prior. We include an LLM content scorer only as a secondary arm, after a memorization probe, to test whether it substitutes for the structural signal.

Proposed Experiments and Decisive Checks

The first decisive check is a matched-exposure analysis. Among exposure events with exactly two, three, or four adopting contacts, we compare forwarding rates across $D=1$ to $D=4$, after stratifying on recency, tie reciprocity, and author popularity. If forwarding rates do not rise with $D$ at a fixed count, the claim fails before any modelling.

The second check tests the interaction. In held-out data we bin events by content score and by $D$. If structural diversity matters mostly when content match is low, as the gate predicts, the gap in forwarding rate between $D\geq 3$ and $D=1$ should be largest in the lowest content-score tercile.

The third check compares the three context definitions. If $k$-trusses do not outperform simple connected components here, we will recommend the cheaper definition.

The fourth check is the serendipity test. Among held-out adoptions labelled serendipitous, we expect a share near the 27–30% reported by Unexpected Relevance. The gated model should raise their rank relative to the constant-gate ablation without lowering NDCG on the rest.

Every comparison reports a 95% bootstrap interval resampled by user. Gains that hold only on balanced samples, the setting in which the published Random Forest figures were obtained, will not count.

Workshop authors to connect with