A recommender that has never seen a new user rate anything still has to place that user somewhere. A classic workaround uses the people the newcomer says they trust. In social matrix factorization, which describes each user and item as a short vector of learned numbers and predicts a rating as their dot product, the new user's vector is pulled toward the average vector of their trusted neighbors. The pull works like a spring, and its strength is fixed once and never revisited. The neighbor might have stopped rating years ago, or their taste might have drifted away from the people who trust them, and the spring would still pull just as hard. A user with many ratings can override a bad spring with their own data. A cold-start user cannot, so the error lands on exactly the users the spring was meant to help.
We propose to give each trusted user a running, time-aged reputation as an anchor. The reputation answers one question: when people who trust this user have recently rated items, did this user's vector predict those ratings? It is updated continuously as new ratings arrive, it decays during silence, and it is allowed to respond more sharply to failures than to successes. Cold-start users are then anchored mostly to neighbors whose reputation is currently high. The quantity is learned on warm users, the people who can check an edge against their own ratings, and transferred to cold users, who cannot. Our claim is that this outcome-based, time-varying reputation reduces cold-start rating error beyond what static trust weights or simple recency decay achieve, and that the gain is concentrated where theory says it should be: among users with zero or very few ratings.
Core Claim and How It Differs From Existing Work
Three bodies of work frame the gap. The first supplies the base method. SocialMF (A Transitivity Aware Matrix Factorization Model for Recommendation in Social Networks) regularizes each user's vector toward the trust-weighted average of their neighbors' vectors. That gives zero-rating users an informed position, and the paper reports an 11.5% RMSE improvement over STE for Epinions users with fewer than 5 ratings and 8.5% on Flixster. Its edges, however, are permanent and unweighted by any evidence of usefulness. Later static refinements change how the edge weight is computed but not when. Fang et al. (Leveraging Decomposed Trust in Probabilistic Matrix Factorization for Effective Recommendation) split trust into benevolence, integrity, competence and predictability and fuse them with a kernel regressor. TR-PMF (Research on Collaborative Filtering Recommendation Based on Trust Relationship and Rating Trust) adds a "rating trust" derived from prediction accuracy, and its authors name the missing temporal evolution as the main limitation. Our reputation is closest in spirit to TR-PMF's rating trust and Fang's competence facet. The difference is that ours is a per-trustee quantity tracked through time with aging, not a snapshot.
The second body is temporal and motivates the work. Dynamic matrix factorization with social influence (Dynamic matrix factorization with social influence) lets user vectors evolve over 11 time bins on Epinions under a linear transition, regularized at each step by the social graph. It shows that social and temporal signals are complementary: it reports RMSE 3.2783 against 3.3352 for the best static matrix factorization. That model treats the social graph in each bin as observed and fixed, uses one global balancing weight, and never evaluates cold-start users. We do the opposite. We keep user vectors simple and let the edge weights themselves move, and cold start is the target regime.
The third body supplies the update rule. Dagdanov et al. (Improving Trust Estimation in Human-Robot Collaboration Using Beta Reputation at Fine-grained Timescales) track trust as a Beta distribution updated with continuous rewards at every step and an exponential aging factor. They report 0.18% absolute error against human self-reports, compared with 18.17% for end-of-task binary updates. Their parameter estimates also show that failures weigh more than successes. That work involves one human and one robot and has never been connected to collaborative filtering. We borrow the update and treat the asymmetry as a hypothesis to test on rating data, not as an assumption.
The Mechanism
Let $u_i, v_j \in \mathbb{R}^D$ be the vectors for user $i$ and item $j$. Let $\mathcal{N}(k)$ be the users who trust user $k$. When a truster $i \in \mathcal{N}(k)$ submits a rating $r_{ij}$ at time $t$, we score how well $k$'s vector would have predicted it, $s = \exp(-(r_{ij} - u_k^\top v_j)^2/\tau)$, where $\tau$ sets the tolerance. Trustee $k$ keeps two pseudo-counts, $a_k$ for success evidence and $b_k$ for failure evidence. They are updated as
$$a_k \leftarrow \gamma^{\Delta t} a_k + s, \qquad b_k \leftarrow \gamma^{\Delta t} b_k + \eta\,(1-s), \qquad \rho_k(t) = \frac{a_k}{a_k + b_k},$$
Here $\Delta t$ is the time since $k$'s last update, $\gamma \in (0,1]$ is the aging rate, and $\eta \ge 1$ is the failure asymmetry. Under this rule, both counts shrink toward the prior when no truster rates anything, so the reputation $\rho_k$ becomes uncertain rather than staying high. A trustee who remains active but whose taste has diverged from their trusters loses reputation, which a pure recency rule cannot detect.
The social term of the objective then becomes a reputation-weighted SocialMF penalty,
$$\mathcal{L}_{\text{soc}}(t) = \frac{\beta}{2}\sum_i \Big\| u_i - \sum_{k \in T(i)} \tilde{w}_{ik}(t)\, u_k \Big\|^2, \qquad \tilde{w}_{ik}(t) = \frac{\rho_k(t)\, c_{ik}}{\sum_{k'} \rho_{k'}(t)\, c_{ik'}},$$
where $T(i)$ is the set of users $i$ trusts and $c_{ik}$ is the original trust value. Training alternates between gradient steps on the vectors and a pass over newly arrived ratings in time order that updates the counts. Because the reputation is computed from warm trusters' outcomes, a newcomer with no ratings inherits well-calibrated spring strengths for free. The parameters $\gamma$, $\eta$ and $\tau$ are fit on a held-out validation time bin. We report the learned values of $\gamma$ and $\eta$ as findings: a fitted $\eta \approx 1$ would mean the robotics asymmetry does not carry over to rating data.
Research Program
The program has three parts. The first is the model and a temporal evaluation protocol. Every comparison uses forward-in-time splits: train on everything before bin $t$ and predict ratings in bin $t$. Cold users are defined at prediction time, as users with zero training ratings and at least one trust edge to a training-time user, plus a separate stratum with 1–4 training ratings. The second is an analysis of the reputation signal itself, checking whether trustees whose reputation falls really did become worse anchors. The third is a secondary question raised by the literature. TrustGNN (TrustGNN: Graph Neural Network-Based Trust Evaluation via Learnable Propagative and Composable Nature) finds that trust paths longer than two hops add noise, while SocialMF's recursion propagates across the whole network. We test whether reputation weighting reduces the harm of deep propagation, by comparing truncated and full propagation with and without reputation.
Epinions is the primary dataset because its ratings carry dates (the dynamic matrix factorization study bins them from 1999 to 2011) and its trust network is large and sparse. Ciao is the second dataset if rating timestamps can be confirmed in the release we obtain. Flixster is attractive because of its many socially connected users with few or no ratings, and we will use it if timestamps are available. FilmTrust lacks the temporal signal, so it serves only as a static sanity check.
Baselines
The naive reference point is a global-mean-plus-item-bias predictor, which is what a cold user effectively receives without social information. We add item-based collaborative filtering, which GlobalTrustWalker's paper uses as its cold-start reference (MAE 1.1895, coverage 22.59% on Epinions). Every proposed method must beat these clearly on the cold strata, or the social channel is not doing anything.
The strongest existing competitor for this specific question is SocialMF (A Transitivity Aware Matrix Factorization Model for Recommendation in Social Networks). It is the original cold-start social anchor, our method is a strict modification of it, and it reports the 11.5% Epinions cold-user improvement over STE cited above. That number comes from a random split, so it is not directly comparable. We re-run SocialMF under our temporal splits with the same latent dimension, the same optimizer, and the same tuning budget of an equal number of hyperparameter trials on the validation bin. Beating it means a statistically significant RMSE and MAE reduction on the zero-rating and 1–4-rating strata, tested with a Wilcoxon signed-rank test over users and across test bins (the style of test PLF uses (A Posterior-Neighborhood-Regularized Latent Factor Model for Highly Accurate Web Service QoS Prediction)), with no significant loss on warm users.
We also run GlobalTrustWalker (Using Trust of Social Ties for Recommendation), which reports cold-start MAE 0.8925 and 77.75% coverage on its pruned Epinions subset. We re-implement it on our splits and compare both MAE and coverage, since a factorization model gives full coverage and the fair comparison is MAE restricted to items both methods can score. Dynamic MF with social influence (Dynamic matrix factorization with social influence) is the strongest temporal-social competitor. Its reported RMSE of 3.2783 is on a scale that indicates different preprocessing, so we treat it only as evidence that the model works and re-run it ourselves on the same bins. For static trust refinement we include TR-PMF and Fang et al.'s decomposed trust, both of which report gains on Ciao and FilmTrust; TR-PMF reports a 1.63% MAE gain on FilmTrust. For warm users we include UserReg (UserReg: A Simple but Strong Model for Rating Prediction), which reports RMSE 0.901 on MovieLens-100K. It has no social component and is undefined for zero-rating users, so on the cold stratum it falls back to bias terms.
The ablations isolate the key ingredient. Fixing $\gamma = 1$ yields a static outcome-based reputation, removing aging. Fixing $\eta = 1$ removes the asymmetry. Replacing $\rho_k(t)$ with a pure recency weight, the exponential decay of time since trustee $k$'s last rating, removes the outcome signal while keeping time. All share the same code path and tuning budget.
Why not just…?
Why not just decay each trust edge by how long ago the neighbor was last active? This is the most serious threat, because staleness is plausibly the main way anchors go bad, and recency costs nothing to compute. Our argument is that recency cannot detect an active neighbor whose taste has drifted from their trusters, and the outcome-based reputation can. That argument is only worth something if such neighbors are common. The recency ablation settles it directly. We also run a controlled check: we take a set of active trustees and replace their later ratings with those of an unrelated user. Recency will keep these trustees at full weight while reputation should drop them, and the size of the resulting cold-start RMSE gap tells us whether the effect matters. If recency matches the full model on the natural data, we will report that the simpler rule suffices.
Why not just use a static reliability score like TR-PMF's rating trust or Fang's competence facet? These already reward neighbors whose ratings agree with others, but they average over the whole history, which TR-PMF's own authors identify as a limitation. The $\gamma = 1$ ablation is exactly this alternative inside our model. The bin-by-bin results show whether the static score degrades in later bins, where old agreement is least informative.
Why not just retrain dynamic matrix factorization with social influence, which already handles time? That model moves the user vectors but holds each bin's social graph fixed with one global weight, and its paper does not evaluate cold users. A zero-rating user has no trajectory for a linear transition to follow, so in that model all of their information comes through fixed edges. We run it on the same bins. Our expectation is that it is competitive on warm users and weaker on the zero-rating stratum, and the stratified results will show whether that holds.
Why not just add embedding noise or contrastive regularization as SimGCL does? SimGCL (Are Graph Augmentations Necessary?: Simple Graph Contrastive Learning for Recommendation) improves Recall@20 by over 25% on Amazon-Book, but it regularizes embeddings that already have interactions behind them. As the research file notes, it gives a zero-rating user no anchor. It addresses the sparse-warm regime, not the cold one, so it is orthogonal to our question and we do not treat it as a cold-start competitor.
Decisive Checks
The proposal succeeds if four things hold together. First, the reputation-weighted model beats SocialMF and the recency ablation on the zero-rating stratum across most test bins. Second, the gain shrinks as users accumulate ratings, which is the signature of a mechanism that matters exactly when a user cannot correct a bad spring. Third, trustees whose reputation drops are measurably worse predictors of their trusters' next-bin ratings than trustees whose reputation stays high, which validates the signal outside the training loss. Fourth, in the controlled drift test, reputation suppresses the drifted anchors while recency does not. The fitted value of $\eta$ answers, as a side result, whether the steep-loss, slow-gain pattern seen in human-robot trust also appears in rating data. If the full model fails to separate from recency decay, the result is still informative, and we will report that rating platforms' trust edges go bad mainly through abandonment rather than drift.
References
- A Transitivity Aware Matrix Factorization Model for Recommendation in Social Networks
- Leveraging Decomposed Trust in Probabilistic Matrix Factorization for Effective Recommendation
- Research on Collaborative Filtering Recommendation Based on Trust Relationship and Rating Trust
- Dynamic matrix factorization with social influence
- Improving Trust Estimation in Human-Robot Collaboration Using Beta Reputation at Fine-Grained Timescales
- TrustGNN: Graph Neural Network-Based Trust Evaluation via Learnable Propagative and Composable Nature
- A Posterior-Neighborhood-Regularized Latent Factor Model for Highly Accurate Web Service QoS Prediction
- Using Trust of Social Ties for Recommendation
- UserReg: A Simple but Strong Model for Rating Prediction
- Are Graph Augmentations Necessary?: Simple Graph Contrastive Learning for Recommendation