What is the curse of directionality in aggregation-based recommenders?
This explores a failure mode of recommenders that build predictions by aggregating signal from neighbors (graph propagation, summed co-occurrence): the corpus doesn't name 'curse of directionality' literally, but it has several notes on the underlying problem — aggregation tends to collapse relationships into one symmetric direction of 'similarity,' losing asymmetry, anti-affinity, and self-prediction control.
This explores a failure mode of aggregation-based recommenders — systems that predict by pooling signal from a user's or item's neighbors, like graph propagation or summed co-occurrence. None of the notes here use the exact phrase 'curse of directionality,' so treat this as triangulation rather than a direct hit. But the corpus circles the same wound from several angles: when you predict by aggregating neighbors, you tend to flatten relationships into a single symmetric notion of 'things that go together,' and you lose the directional structure — who predicts whom, and what should actively repel what.
The sharpest piece of evidence is the linear autoencoder work in Can a linear model beat deep collaborative filtering?. Its central trick is a zero-diagonal constraint: an item is forbidden from predicting itself, forcing every prediction through item-to-item relationships rather than trivial self-aggregation. Just as important, ESLER finds that negative weights — encoding *anti-affinity*, the signal that one item should suppress another — are essential to performance. That's exactly the thing naive aggregation can't represent: summing or averaging neighbors only ever adds positive 'these belong together' mass, so it's structurally blind to repulsion and to the fact that influence runs one way more than the other.
You can see the same instinct in Why does multinomial likelihood work better for ranking recommendations?, where switching the VAE's likelihood to multinomial wins because it forces items to *compete* for probability rather than each being scored in isolation. Competition is a directional relationship — one item's gain is another's loss — and it aligns training with ranking in a way that independent per-item aggregation doesn't. Both notes point at the same lesson: structural bias about how signal flows between items matters more than raw model capacity.
The graph-propagation notes show where the curse originates and why it's seductive. Can graphs unify collaborative filtering and side information? propagates signal across a combined user-item-attribute graph to capture high-order connections, and Can autoencoders solve the cold-start problem in recommendations? uses graph autoencoders to fuse ratings with side information. These methods are powerful precisely because they aggregate over many hops — but the more hops you pool, the more everything blurs toward a generic 'central, popular' direction. That blur shows up elsewhere in the corpus as popularity overfitting in Does embedding dimensionality secretly drive popularity bias in recommenders? and as the crowding-out of minority interests that needs post-hoc fixing in Why do accuracy-optimized recommenders crowd out minority interests?.
The takeaway you might not have gone looking for: the cure in this corpus is never 'aggregate harder.' It's adding back *structure* the aggregation throws away — forbidding self-prediction, allowing negative/repulsive weights, making items compete instead of accumulate. The directionality and asymmetry of preference is signal, and pooling is the operation most likely to average it into mush.
Sources 6 notes
ESLER, a single-layer linear autoencoder constrained so items cannot predict themselves, outperforms most deep CF models. The constraint forces prediction through item relationships, and negative weights encoding anti-affinity prove essential—structural bias matters more than model capacity.
Liang et al. show that switching VAE likelihoods from Gaussian/logistic to multinomial achieves state-of-the-art results because enforced probability competition between items directly aligns training with top-N ranking objectives. Rebalancing KL regularization further improves performance.
KGAT merges user-item interaction graphs with item knowledge graphs into a Collaborative Knowledge Graph, using attention-based propagation to capture both user-similarity and attribute-similarity signals simultaneously—including high-order connections that standard supervised learning methods miss.
GHRS uses graph features and deep autoencoders to integrate rating history with side information, enabling predictions for new users and items by discovering non-linear relationships that linear hybrid methods miss.
Research shows that when user/item embedding dimensions are too small, recommender systems overfit toward popular items to maximize ranking quality. This compounds over time as niche items receive insufficient exposure, and cannot be fixed post-hoc without treating dimensionality as a fairness hyperparameter.
Accuracy-optimized models systematically miscalibrate by over-weighting dominant user interests. A post-processing reranking algorithm that enforces calibration constraints can restore proportional representation without retraining the underlying model.