We study the problem of inferring substitutable and complementary items, which underpins applications such as alternative and follow-up purchase suggestions. Existing approaches typically learn from behavior-derived item-item associations using GNNs or leverage item content alone. However, these methods often overlook two key challenges: (i) user behaviors (e.g., co-view/co-purchase) only provide noisy weak supervision, and (ii) behavior signals are long-tailed, leaving many items with sparse associations. We propose MMSC, a self-supervised multi-modal relational representation learning framework that combines a multi-modal foundation model adapted to encode item metadata and a self-supervised denoising module that learns relationship-aware representations from noisy user behaviors, unified by a hierarchical aggregation mechanism. We further use LLM-assisted supervision to mitigate noise in behavior-derived supervision during training. Experiments on five real-world datasets show that MMSC consistently outperforms existing baselines by 26.1% for substitutable and 39.2% for complementary item inference, while remaining effective for cold-start items.

This paper proposes a novel pre-trained framework for zero-shot cross-domain sequential recommendation without auxiliary information. While using auxiliary information (e.g., item descriptions) seems promising for cross-domain transfer, a cross-domain adaptation of sequential recommenders can be challenging when the target domain differs from the source domain—item descriptions are in different languages; metadata modalities (e.g., audio, image, and text) differ across source and target domains. If we can learn universal item representations independent of the domain type (e.g., groceries, movies), we can achieve zero-shot cross-domain transfer without auxiliary information. Our critical insight is that user interaction sequences highlight shifting user preferences via the popularity dynamics of interacted items. We present a pre-trained sequential recommendation framework: PrepRec, which utilizes a novel popularity dynamics-aware transformer architecture. Through extensive experiments on five real-world datasets, we show that PrepRec, without any auxiliary information, can zero-shot adapt to new application domains and achieve competitive performance compared to state-of-the-art sequential recommender models. In addition, we show that PrepRec complements existing sequential recommenders. With a simple post-hoc interpolation, PrepRec improves the performance of existing sequential recommenders on average by 11.8% in Recall@10 and 22% in NDCG@10. We provide an anonymized implementation of PrepRec at https://github.com/CrowdDynamicsLab/preprec.

In recent times, deep learning methods have supplanted conventional collaborative filtering approaches as the backbone of modern recommender systems. However, their gains are skewed towards popular items with a drastic performance drop for the vast collection of long-tail items with sparse interactions. Moreover, we empirically show that prior neural recommenders lack the resolution power to accurately rank relevant items within the long-tail.

The quality variance in user-generated content is a major bottleneck to serving communities on online platforms. Current content ranking methods primarily evaluate text and non-textual content features of each user post in isolation. In this paper, we demonstrate the utility of considering the implicit and explicit relational aspects across user content to assess their quality. First, we develop a modular platform-agnostic framework to represent the contrastive (or competing) and similarity-based relational aspects of user-generated content via independently induced content graphs. Second, we develop two complementary graph convolutional operators that enable feature contrast for competing content and feature smoothing/sharing for similar content. Depending on the edge semantics of each content graph, we embed its nodes via one of the above two mechanisms. We also show that our contrastive operator creates discriminative magnification across the embeddings of competing posts. Third, we show a surprising result-applying classical boosting techniques to combine final-layer embeddings across the content graphs significantly outperforms the typical stacking, fusion, or neighborhood embedding aggregation methods in graph convolutional architectures. We exhaustively validate our method via accepted answer prediction over fifty diverse Stack-Exchange (https://stackexchange.com/) websites with consistent relative gains of over 5% accuracy over state-of-the-art neural, multi-relational and textual baselines.

We present InfoMotif, a new semi-supervised, motif-regularized, learning framework over graphs. We overcome two key limitations of message passing in popular graph neural networks (GNNs): localization (a k-layer GNN cannot utilize features outside the k-hop neighborhood of the labeled training nodes) and over-smoothed (structurally indistinguishable) representations. We propose the concept of attributed structural roles of nodes based on their occurrence in different network motifs, independent of network proximity. Two nodes share attributed structural roles if they participate in topologically similar motif instances over co-varying sets of attributes. Further, InfoMotif achieves architecture independence by regularizing the node representations of arbitrary GNNs via mutual information maximization. Our training curriculum dynamically prioritizes multiple motifs in the learning process without relying on distributional assumptions in the underlying graph or the learning task. We integrate three state-of-the-art GNNs in our framework, to show significant gains (3-10% accuracy) across six diverse, real-world datasets. We see stronger gains for nodes with sparse training labels and diverse attributes in local neighborhood structures.