Recommendation systems learn from clicks, but many clicks are mistakes: a misclick, a curious glance, a song skipped after two seconds. When a session holds only three or four clicks, one accidental click can be a large share of what the model sees. The one method in this file that directly removes these "false positive" clicks, FPAdaMetric, needs post-click telemetry such as skip times. Most click logs have no such signal. We propose to build the missing noise labels from the rest of the population. If a click is badly out of line with how millions of other sessions move between items, it is probably accidental. We call this "distant supervision", the relation-extraction trick of generating training labels automatically from an outside source instead of by hand (Mintz et al.).
Those automatic labels will be wrong in a predictable way, and handling that is the heart of the proposal. In relation extraction, "these two entities appear together" does not mean "this sentence states the relation." Here, "this transition is rare in the population" does not mean "this click was a mistake." Very often it only means the item is in the long tail and the population has little to say about it. Our claim is that population statistics become useful noise labels only when each label is softened according to how much evidence the population holds for that item pair. A student detector can then be trained on these soft labels with the noise-tolerant methods relation extraction developed. With this treatment, a recommender trained on clicks alone should recover much of the gain that skip telemetry provides, without deleting genuine niche interests.
Core Claim
We make three testable claims. First, on datasets with skip telemetry, noise labels built only from global transition statistics predict held-out skips better than labels from local, within-session detectors. The reason is that a short session does not contain enough context to judge its own outliers, the problem SSDRec calls over- and under-denoising. Second, labels that account for how much evidence backs them avoid the specific failure of flagging long-tail items as noise, which raw co-occurrence thresholds commit. Third, a recommender whose training loss down-weights the detected clicks closes a meaningful part of the gap between an ordinary click-trained model and the telemetry-supervised FPAdaMetric.
Positioning Against Prior Work
Several global-to-local recommenders already use a population-level item graph, but they use it to add information, not to judge clicks. Cross-session GNNs and session-adaptive KG propagation pull neighbor representations into the session. APGL4SR learns a perturbed global graph with a contrastive objective. All three treat the observed session as ground truth. The KG propagation paper builds its co-view and add-to-cart edges from fixed rules and never estimates how confident each edge is. These papers motivate our work, because they show that global structure helps short sessions. We use it for a different job: labeling which clicks to trust.
The two closest papers each cover half the problem. SSDRec uses a global multi-relation graph to insert items into sequences and then prunes noisy clicks with a two-stage heuristic filter. Its thresholds are fixed, and it is never checked against real ground truth on which clicks were noise. FPAdaMetric turns skips into negative signal inside a triplet-style metric loss, but it depends on telemetry. We reuse FPAdaMetric's evaluation data in an unusual way: skips are hidden during training and used only as an answer key for the noise labels. That gives this research area something it has lacked, a direct measurement of denoising accuracy rather than only its downstream effect.
We borrow the noise-handling machinery from relation extraction. Global Distant Supervision showed that consistency across many instances, including the rule that similar instances should get similar labels, lifted KBP F1 from 0.255 (Mintz++) to 0.322. It did this with Markov Logic Networks, which do not scale to click logs. We keep the neighbor-consistency idea and replace symbolic inference with a learned detector. Teacher-student training and soft distillation with adaptive focal loss supply the training recipe for learning from soft, partly wrong targets. HiRAM is a warning from the relation-extraction side: distant labels are weakest exactly where data is sparse. That is why labels that account for evidence are the central ingredient here, not an afterthought.
Research Program
The first stage builds the distant labels. For each click $c$ on item $j$ in a session with preceding context items $i_1,\dots,i_k$, we compute a compatibility score from the global transition graph. The score is the average pointwise mutual information between $j$ and the context items, which measures how much more often they appear together than chance would predict. We then turn this score into a soft noise label that shrinks toward an uninformative prior when evidence is thin:
$$q_c = \frac{n_c}{n_c+\kappa}\,\sigma\!\big(-\beta\,(\mathrm{PMI}_c-\tau)\big) + \frac{\kappa}{n_c+\kappa}\,\pi .$$
In this formula, $n_c$ is the total global count of sessions containing $j$ together with any context item, $\kappa$ is a shrinkage strength, and $\sigma$ is the logistic function. $\beta$ sets how sharply the label switches, $\tau$ is a compatibility threshold, and $\pi$ is the base rate of noise. This is the usual shrinkage form for estimates built on small counts, so a tail item with little co-occurrence data receives a label near $\pi$ instead of a confident "noise." The hyperparameters $\kappa,\beta,\tau$ are tuned on a small validation split, never on the skip labels used for evaluation.
The second stage trains a student detector on the soft labels. It is a small network that sees the click's local position and dwell-free session features, the item's embedding, and aggregated global neighbor features. It is trained with a focal-weighted cross-entropy against $q_c$, following Tan et al., with a moving-average teacher as in Nagesh et al.. We add a neighbor-consistency term, taken from Global Distant Supervision, that asks clicks in similar session contexts to receive similar noise scores. The student can generalize past the rule that produced its labels. For example, it can learn that a click immediately abandoned for an unrelated item looks different from a deliberate step into a new category.
The third stage passes the detector's scores into an unchanged recommender backbone, SASRec or GRU4Rec. We compare two uses: down-weighting a click's contribution to the next-item loss, or treating high-scoring clicks as soft negatives in the style of FPAdaMetric. Nothing changes at inference time, so serving cost matches the backbone.
Baselines
The naive reference point is the backbone, SASRec and GRU4Rec, trained on all clicks with no denoising. Every other method is reported as a change from it under the same data splits, embedding size, epochs, and hyperparameter-search budget. We allow 30 trials per method, with the same random-search space for shared hyperparameters.
The strongest existing competitor without telemetry is SSDRec, which also uses global graph priors to prune clicks from short sequences. It reports +39% to +51% gains over SASRec on Amazon Beauty and Sports and +110% NDCG on ML-100K. These are relative gains, and no absolute values were found on Lacuna for SSDRec. We will rerun it on our splits with its own thresholds retuned under the shared budget. Beating it means higher NDCG@10 and Recall@20 on Amazon Beauty and Sports and on the telemetry datasets, together with higher skip-prediction AUC, since SSDRec's filter produces a keep-or-drop decision that can be scored against skips.
The strongest method overall is FPAdaMetric trained with real skip labels. On Spotify it reports raising HR@10 from about 29.9% to about 82.2%. It serves as an upper reference, not a target to beat. Success means recovering a substantial fraction of the gap between the naive backbone and FPAdaMetric without seeing any skip. Alongside it we run an oracle that simply deletes all skipped clicks. This separates the value of the labels themselves from the value of FPAdaMetric's loss.
For session benchmarks without telemetry, the published reference points are SR-GNN at P@20 = 50.73% on Diginetica and MRR@20 = 30.94 on Yoochoose 1/64 (SR-GNN), and FGNN-BCS at Recall@20 = 51.67% and MRR@20 = 18.69% on Diginetica (cross-session GNN). We will add our denoising to SR-GNN and report against its reproduced number.
Three ablations isolate the proposal's key ingredient. The first uses raw PMI hard thresholding: clicks below $\tau$ are deleted outright, with no shrinkage and no student. The second uses soft labels without the evidence term, that is, $n_c \to \infty$. The third trains the student without the neighbor-consistency term. If the full method does not beat the first two ablations in the tail-popularity buckets, the evidence-aware labeling claim fails.
Why not just…?
Why not just threshold the co-occurrence statistics and delete low-scoring clicks? This is the most serious threat. It is exactly our first ablation, and on randomly injected noise it may well match the full method, because random clicks are trivially incompatible with their context. We expect it to fail on real skips in the tail. Rare items have low co-occurrence counts for reasons that have nothing to do with noise. This is the recommendation version of the distant-supervision false positive, and the same sparsity failure HiRAM documents for rare relations. The popularity-bucketed skip-AUC comparison settles the question. If thresholding matches the full method there too, we will report that the whole benefit comes from global statistics and drop the learned stages.
Why not just denoise locally with the recommender's own self-attention? SSDRec reports that local-only denoisers such as HSD and STEAM over- or under-denoise in short sequences, because a single outlier cannot be told apart from a real interest shift. We test this directly by scoring an attention-based local filter against held-out skips and breaking the results down by session length. The prediction is that it matches our detector on long sessions and falls behind on sessions of three items or fewer.
Why not just use SSDRec as is? SSDRec is the closest method, and it may win on downstream accuracy because it also inserts items. Insertion is a separate effect from removing noise. To keep the comparison fair, we also run SSDRec with insertion disabled and report skip-AUC, which SSDRec has never been evaluated on. Our contribution holds if our labels identify accidental clicks more accurately, even in cases where SSDRec's insertion adds its own downstream gains.
Why not just collect telemetry or ask a large language model which clicks look accidental? Telemetry is simply absent from logs like Yoochoose and Diginetica. Prompting a language model per click is expensive, and KDSR argues that large-model signals are best distilled offline for this reason. A model that reads item metadata also cannot see the population behavior that distinguishes a misclick from an idiosyncratic but real choice. We treat telemetry as the upper reference and leave language-model labeling as a possible extra feature, not a replacement.
Proposed Experiments
The decisive experiment uses the post-click benchmarks from FPAdaMetric, Spotify and LastFM, with Amazon Movie as a third domain. We remove all skip information from training and measure three things: noise-detection AUC and precision-at-recall against held-out skip labels, broken down by item-popularity quintile and session length; the tail items' share of flagged clicks compared with their share of true skips; and downstream HR@10 and NDCG@10 against the naive backbone, the telemetry oracle, and FPAdaMetric. A positive result looks like this: the global labels beat local filters on skip-AUC, most clearly in short sessions; evidence-aware labels keep the false-flag rate for tail items close to the rate for head items while the raw threshold does not; and the downstream model recovers a clear share of FPAdaMetric's gain.
The second experiment checks transfer to logs without telemetry. We use Diginetica and Yoochoose 1/64 with SR-GNN and SASRec backbones, and Amazon Beauty and Sports with SASRec against SSDRec. Here we can measure only downstream accuracy, so we add a controlled test. We inject non-random "plausible" noise by sampling clicks from the same category as the context, which makes it harder to catch than uniform noise, and measure detection recall. We report uniform-noise results separately and do not treat them as evidence, since any co-occurrence rule catches them.
The final check is falsification. If raw thresholding matches the full method on tail-bucket skip-AUC, or if detection gains on Spotify and LastFM do not produce downstream gains, then the distant-supervision treatment is not doing the work we claim. We will report whichever of these outcomes occurs.
References
- FPAdaMetric: False-Positive-Aware Adaptive Metric Learning for Session-Based Recommendation
- Distant supervision for relation extraction without labeled data
- SSDRec: Self-Augmented Sequence Denoising for Sequential Recommendation
- Exploiting Cross-session Information for Session-based Recommendation with Graph Neural Networks
- Knowledge Graph-based Session Recommendation with Session-Adaptive Propagation
- APGL4SR: A Generic Framework with Adaptive and Personalized Global Collaborative Information in Sequential Recommendation
- Global Distant Supervision for Relation Extraction
- Semi-Supervised Teacher-Student Architecture for Relation Extraction
- Document-Level Relation Extraction with Adaptive Focal Loss and Knowledge Distillation
- Hierarchical Relation-Guided Type-Sentence Alignment for Long-Tail Relation Extraction with Distant Supervision
- Session-Based Recommendation with Graph Neural Networks.
- Lightweight Modality Adaptation to Sequential Recommendation via Correlation Supervision