Combining semantic information with behavioral data is a crucial research area in recommender systems. A promising approach involves leveraging external knowledge to enrich behavioral-based recommender systems with abundant semantic information. However, this approach faces two primary challenges: (1) denoising raw external knowledge and (2) adapting semantic representations. To address these challenges, we propose exTernal knowledge-enhanced RecommendAtion With LLM assistance (TRAWL). This method utilizes large language models to extract relevant recommendation knowledge from raw external data and employs a contrastive learning strategy for adapter training. Experiments on public datasets and real-world online recommender systems validate the effectiveness of our approach.

Recommendation platforms commonly offer diverse content scenarios for users to interact with. Pre-training models are the most commonly used approach in recommendation systems to capture users' full-domain interests. Traditional ID-based pre-training models mainly capture user interests by leveraging collaborative signals. However, a prevalent drawback of those systems is the incapacity to handle cold-start scenarios. With the recent advent of large language models, there has been a significant increase in research efforts exploiting LLMs to extract semantic information for items. However, text-based recommendations highly rely on elaborate feature engineering and often fail to capture collaborative similarities.

In recent years, with the significant evolution of multi-modal large models, many recommender researchers realized the potential of multi-modal information for user interest modeling. In industry recommendation system, a wide-used modeling architecture is to first pre-train a multi-modal model to provide omnipotent representations and then encode to discrete semantic IDs for online model. Although such a paradigm achieves remarkable improvements, there still exist two problems that limit model performance: (1) Modalities Mapping Independence: Each modal representation is independently mapped to semantic spaces and then get the specific code, which ignores the consistency and complementarity of different modalities of the same item. (2) User-irrelevant Clustering Assignment: For the specific item, most of existing quantization methods assume that all users share the same cluster assignments, failing to account for the varying interpretations and emotional responses users may have toward an item.

In real-world recommendation scenarios, users typically engage with platforms through multiple types of behavioral interactions. Multi-behavior recommendation algorithms aim to leverage various auxiliary user behaviors to enhance prediction for target behaviors of primary interest (e.g., buy), thereby overcoming performance limitations caused by data sparsity in target behavior records. Current state-of-the-art approaches typically employ hierarchical design following either cascading (e.g., view$\rightarrow$cart$\rightarrow$buy) or parallel (unified$\rightarrow$behavior$\rightarrow$specific components) paradigms, to capture behavioral relationships. However, these methods still face two critical challenges: (1) severe distribution disparities across behaviors, and (2) negative transfer effects caused by noise in auxiliary behaviors. In this paper, we propose a novel model-agnostic Hierarchical Graph Information Bottleneck (HGIB) framework for multi-behavior recommendation to effectively address these challenges. Following information bottleneck principles, our framework optimizes the learning of compact yet sufficient representations that preserve essential information for target behavior prediction while eliminating task-irrelevant redundancies. To further mitigate interaction noise, we introduce a Graph Refinement Encoder (GRE) that dynamically prunes redundant edges through learnable edge dropout mechanisms. We conduct comprehensive experiments on three real-world public datasets, which demonstrate the superior effectiveness of our framework. Beyond these widely used datasets in the academic community, we further expand our evaluation on several real industrial scenarios and conduct an online A/B testing, showing again a significant improvement in multi-behavior recommendations. The source code of our proposed HGIB is available at https://github.com/zhy99426/HGIB.

In the digital era, users typically interact with diverse items across multiple domains (e.g., e-commerce, streaming platforms, and social networks), generating intricate heterogeneous interaction graphs. Leveraging multi-domain data can improve recommendation systems by enriching user insights and mitigating data sparsity in individual domains. However, integrating such multi-domain knowledge for cross-domain recommendation remains challenging due to inherent disparities in user behavior and item characteristics and the risk of negative transfer, where irrelevant or conflicting information from the source domains adversely impacts the target domain's performance. To tackle these challenges, we propose HAGO, a novel framework with Heterogeneous Adaptive Graph coOrdinators, which dynamically integrates multi-domain graphs into a cohesive structure. HAGO adaptively adjusts the connections between coordinators and multi-domain graph nodes to enhance beneficial inter-domain interactions while alleviating negative transfer. Furthermore, we introduce a universal multi-domain graph pre-training strategy alongside HAGO to collaboratively learn high-quality node representations across domains. Being compatible with various graph-based models and pre-training techniques, HAGO demonstrates broad applicability and effectiveness. Extensive experiments show that our framework outperforms state-of-the-art methods in cross-domain recommendation scenarios, underscoring its potential for real-world applications. The source code is available at https://github.com/zhy99426/HAGO.

Since people with close relationships are easily influenced by each other, social friends usually have more preferences of higher similarities than others. For this reason, social recommendation methods are proposed to adopt social links to improve the degree of preference matching between users and recommended items. Although current social recommendation methods have captured the general preference similarities among social friends, it is still difficult to model the evolution of dynamic social influence among friends, especially in session-based scenarios. In reality, when users’ dynamic preferences are changing, social feedback from their friends is also changing over time. So that the dynamic social feedback is an important social influence, which has not been considered in current studies. To this end, we propose a social feedback-enhanced session-based recommendation (SFRec) method, which not only utilizes the similarity of general preferences among friends but also captures the friends’ influence which reflects people’s dynamic preferences. Specifically, we first coordinate similarity relations via information propagation on social graph, item transition graph and user–item interaction graph. To capture social feedback based on users’ dynamic preferences, we then construct a social feedback generation module that consists of preference extraction, feedback generation and feedback aggregation. Finally, we construct a preference fusion module to obtain the final preference representation and make personalized recommendation. We conduct comprehensive experiments on three datasets. Results demonstrate that SFRec surpasses the state-of-the-art models on recommendation performance.

Recent years have witnessed tremendous interest in deep learning on graph-structured data. Due to the high cost of collecting labeled graph-structured data, domain adaptation is important to supervised graph learning tasks with limited samples. However, current graph domain adaptation methods are generally adopted from traditional domain adaptation tasks, and the properties of graph-structured data are not well utilized. For example, the observed social networks on different platforms are controlled not only by the different crowds or communities but also by domain-specific policies and background noise. Based on these properties in graph-structured data, we first assume that the graph-structured data generation process is controlled by three independent types of latent variables, i.e., the semantic latent variables, the domain latent variables, and the random latent variables. Based on this assumption, we propose a disentanglement-based unsupervised domain adaptation method for the graph-structured data, which applies variational graph auto-encoders to recover these latent variables and disentangles them via three supervised learning modules. Extensive experimental results on two real-world datasets in the graph classification task reveal that our method not only significantly outperforms the traditional domain adaptation methods and the disentangled-based domain adaptation methods but also outperforms the state-of-the-art graph domain adaptation algorithms. The code is available at https://github.com/rynewu224/GraphDA .

Micro-video platforms such as TikTok are extremely popular nowadays. One important feature is that users no longer select interested videos from a set; instead, they either watch the recommended video or skip to the next one. As a result, the time length of users’ watching behavior becomes the most important signal for identifying preferences. However, our empirical data analysis has shown a video-length effect that long videos can more easily receive a higher value of average view time, and thus adopting such view-time labels for measuring user preferences can easily induce a biased model that favors the longer videos. In this article, we propose a V ideo L ength D ebiasing Rec ommendation (VLDRec) method to alleviate such an effect for micro-video recommendation. VLDRec designs the data labeling approach and the sample generation module that better capture user preferences in a view-time-oriented manner. It further leverages the multi-task learning technique to jointly optimize the above samples with the original biased ones. Extensive experiments show that VLDRec can improve users’ view time by 1.81% and 11.32% on two real-world datasets, given a recommendation list of a fixed overall video length, compared with the best baseline method. Moreover, VLDRec is also more effective in matching users’ interests in terms of the video content.

Recently, cross-domain recommendation (CDR) has been widely studied in both research and industry since it can alleviate a long-standing challenge of traditional recommendation methods, i.e., data sparsity issue, by transferring the information from a relatively richer domain (termed source domain) to a sparser domain (termed target domain). To our best knowledge, most (if not all) existing CDR methods focus on transferring either the similar content information or the user preferences embedding from the source domain to the target domain. However, they fail to improve the recommendation performance in real-world recommendation scenarios where the items in the source domain are totally different from those in the target domain in terms of attributes. To solve the above issues, we analyzed the historical interactions of users from different domains in the WeChat platform, and found that if two users have similar interests (interactions) in one domain, they are very likely to have similar interests in another domain even though the items of these two domains are totally different in terms of attributes. Based on this observation, in this paper, we propose a novel model named Dual Interests-Aligned Graph Auto-Encoders (DIAGAE) by utilizing the inter-domain interest alignment of users. Besides, our proposed model DIAGAE also leverages graph decoding objectives to align intra-domain user interests, which makes the representation of two users who have similar interests in a single domain closer. Comprehensive experimental results demonstrate that our model DIAGAE outperforms state-of-the-art methods on both public benchmark datasets and online A/B tests in WeChat live-stream recommendation scenario. Our model DIAGAE now serves the major online traffic in WeChat live-streaming recommendation scenario.

Automated Machine Learning (AutoML) techniques have recently been introduced to design Collaborative Filtering (CF) models in a data-specific manner. However, existing works either search architectures or hyperparameters while ignoring the fact they are intrinsically related and should be considered together. This motivates us to consider a joint hyperparameter and architecture search method to design CF models. However, this is not easy because of the large search space and high evaluation cost. To solve these challenges, we reduce the space by screening out usefulness hyperparameter choices through a comprehensive understanding of individual hyperparameters. Next, we propose a two-stage search algorithm to find proper configurations from the reduced space. In the first stage, we leverage knowledge from subsampled datasets to reduce evaluation costs; in the second stage, we efficiently fine-tune top candidate models on the whole dataset. Extensive experiments on real-world datasets show better performance can be achieved compared with both hand-designed and previous searched models. Besides, ablation and case studies demonstrate the effectiveness of our search framework.

Graph Neural Network (GNN) based social recommendation models improve the prediction accuracy of user preference by leveraging GNN in exploiting preference similarity contained in social relations. However, in terms of both effectiveness and efficiency of recommendation, a large portion of social relations can be redundant or even noisy, e.g., it is quite normal that friends share no preference in a certain domain. Existing models do not fully solve this problem of relation redundancy and noise, as they directly characterize social influence over the full social network. In this paper, we instead propose to improve graph based social recommendation by only retaining the informative social relations to ensure an efficient and effective influence diffusion, i.e., graph denoising. Our designed denoising method is preference-guided to model social relation confidence and benefits user preference learning in return by providing a denoised but more informative social graph for recommendation models. Moreover, to avoid interference of noisy social relations, it designs a self-correcting curriculum learning module and an adaptive denoising strategy, both favoring highly-confident samples. Experimental results on three public datasets demonstrate its consistent capability of improving three state-of-the-art social recommendation models by robustly removing 10-40% of original relations. We release the source code at https://github.com/tsinghua-fib-lab/Graph-Denoising-SocialRec.

In recommender systems, some features directly affect whether an interaction would happen, making the happened interactions not necessarily indicate user preference. For instance, short videos are objectively easier to finish even though the user may not like the video. We term such feature as confounding feature , and video length is a confounding feature in video recommendation. If we fit a model on such interaction data, just as done by most data-driven recommender systems, the model will be biased to recommend short videos more, and deviate from user actual requirement. This work formulates and addresses the problem from the causal perspective. Assuming there are some factors affecting both the confounding feature and other item features, e.g., the video creator, we find the confounding feature opens a backdoor path behind user-item matching and introduces spurious correlation. To remove the effect of backdoor path, we propose a framework named Deconfounding Causal Recommendation (DCR) , which performs intervened inference with do-calculus . Nevertheless, evaluating do-calculus requires to sum over the prediction on all possible values of confounding feature, significantly increasing the time cost. To address the efficiency challenge, we further propose a mixture-of-experts (MoE) model architecture, modeling each value of confounding feature with a separate expert module. Through this way, we retain the model expressiveness with few additional costs. We demonstrate DCR on the backbone model of neural factorization machine (NFM) , showing that DCR leads to more accurate prediction of user preference with small inference time cost. We release our code at: https://github.com/zyang1580/DCR .

Recently, graph neural network (GNN) approaches have received huge interests in recommendation tasks due to their ability of learning more effective user and item representations. However, existing GNN-based recommendation models cannot support real-time recommendation where the model keeps its freshness by continuously training the streaming data that users produced, leading to negative impact on recommendation performance. To fully support graph-enhanced large-scale recommendation in real-time scenarios, a deep graph learning system is required to dynamically store the streaming data as a graph structure and enable the development of any GNN model incorporated with the capabilities of real-time training and online inference. However, such requirements rule out existing deep graph learning solutions. In this paper, we propose a new deep graph learning system called PlatoGL, where (1) an effective block-based graph storage is designed with non-trivial insertion/deletion mechanism for updating the graph topology in-milliseconds, (2) a non-trivial multi-blocks neighbour sampling method is proposed for efficient graph query, and (3) a cache technique is exploited to improve the storage stability. We have deployed PlatoGL in Wechat, and leveraged its capability in various content recommendation scenarios including live-streaming, article and micro-video. Comprehensive experiments on both deployment performance and benchmark performance~(w.r.t. its key features) demonstrate its effectiveness and scalability. One real-time GNN-based model, developed with PlatoGL, now serves the major online traffic in WeChat live-streaming recommendation scenario.

Sequential recommender systems (SRSs) have become a research hotspot recently due to its powerful ability in capturing users' dynamic preferences. The key idea behind SRSs is to model the sequential dependencies over the user-item interactions. However, we argue that users' preferences are not only determined by their view or purchase items but also affected by the item-providers with which users have interacted. For instance, in a short-video scenario, a user may click on a video because he/she is attracted to either the video content or simply the video-providers as the vloggers are his/her idols. Motivated by the above observations, in this paper, we propose IPSRec, a novel Item-Provider co-learning framework for Sequential Recommendation. Specifically, we propose two representation learning methods (single-steam and cross-stream) to learn comprehensive item and user representations based on the user's historical item sequence and provider sequence. Then, contrastive learning is employed to further enhance the user embeddings in a self-supervised manner, which treats the representations of a specific user learned from the item side as well as the item-provider side as the positive pair and treats the representations of different users in the batch as the negative samples. Extensive experiments on three real-world SRS datasets demonstrate that IPSRec achieves substantially better results than the strong competitors. For reproducibility, our code and data are available at https://github.com/siat-nlp/IPSRec.

A dataset which is highly-dynamic and recency-sensitive means new data are generated in high volumes with a fast speed and of higher priority for the subsequent applications. Embedding technique is a popular research topic in recent years which aims to represent any data into low-dimensional vector space, which is widely used in different data types and have multiple applications. Generating embeddings on such data in a high-speed way is a challenging problem to consider the high dynamics and the recency sensitiveness together with both effectiveness and efficient. Popular embedding methods are usually time-consuming. As well as the common optimization methods are limited since it may not have enough time to converge or deal with recency-sensitive sample weights. This problem is still an open problem. In this paper, we propose a novel optimization method named Diffused Stochastic Gradient Descent for such highly-dynamic and recency-sensitive data. The notion of our idea is to assign recency-sensitive weights to different samples, and select samples according to their weights in calculating gradients. And after updating the embedding of the selected sample, the related samples are also updated in a diffusion strategy. We propose a Nested Segment Tree to improve the recency-sensitive weight method and the diffusion strategy into a complexity no slower than the iteration step in practice. We also theoretically prove the convergence rate of D-SGD for independent data samples, and empirically prove the efficacy of D-SGD in large-scale real datasets.