AltFS: Agency-light Feature Selection with Large Language Models in Deep Recommender Systems

Feature selection is crucial in recommender systems for improving model efficiency and predictive performance. Traditional methods rely on agency models, such as decision trees or neural networks, to estimate feature importance. However, this approach is inherently limited, as the agency models may fail to learn effectively in all scenarios due to suboptimal training conditions (e.g., feature collinearity, high-dimensional sparsity, and data insufficiency). In this paper, we propose AltFS, an Agency-light Feature Selection method for deep recommender systems. AltFS integrates semantic reasoning from Large Language Models (LLMs) with task-specific learning from agency models. Initially, LLMs will generate a semantic ranking of feature importance, which is then refined by an agency model, combining world knowledge with task-specific insights. Extensive experiments on three public datasets from real-world recommender platforms demonstrate the effectiveness of AltFS. Our code is publicly available for reproducibility.

Bridging Relevance and Reasoning: Rationale Distillation in Retrieval-Augmented Generation

The reranker and generator are two critical components in the Retrieval-Augmented Generation (i.e., RAG) pipeline, responsible for ranking relevant documents and generating responses. However, due to differences in pre-training data and objectives, there is an inevitable gap between the documents ranked as relevant by the reranker and those required by the generator to support answering the query. To address this gap, we propose RADIO, a novel and practical preference alignment framework with RAtionale DIstillatiOn. Specifically, We first propose a rationale extraction method that leverages the reasoning capabilities of Large Language Models (LLMs) to extract the rationales necessary for answering the query. Subsequently, a rationale-based alignment process is designed to rerank the documents based on the extracted rationales, and fine-tune the reranker to align the preferences. We conduct extensive experiments on two tasks across three datasets to demonstrate the effectiveness of our approach compared to baseline methods. Our code is released online to ease reproduction.

Pre-train, Align, and Disentangle: Empowering Sequential Recommendation with Large Language Models

Sequential recommendation (SR) aims to model the sequential dependencies in users' historical interactions to better capture their evolving interests. However, existing SR approaches primarily rely on collaborative data, which leads to limitations such as the cold-start problem and sub-optimal performance. Meanwhile, despite the success of large language models (LLMs), their application in industrial recommender systems is hindered by high inference latency, inability to capture all distribution statistics, and catastrophic forgetting. To this end, we propose a novel Pre-train, Align, and Disentangle (PAD) paradigm to empower recommendation models with LLMs. Specifically, we first pre-train both the SR and LLM models to get collaborative and textual embeddings. Next, a characteristic recommendation-anchored alignment loss is proposed using multi-kernel maximum mean discrepancy with Gaussian kernels. Finally, a triple-experts architecture, consisting aligned and modality-specific experts with disentangled embeddings, is fine-tuned in a frequency-aware manner. Experiments conducted on three public datasets demonstrate the effectiveness of PAD, showing significant improvements and compatibility with various SR backbone models, especially on cold items. The implementation code and datasets will be publicly available.

Tackling Data Heterogeneity in Federated Time Series Forecasting

Time series forecasting plays a critical role in various real-world applications, including energy consumption prediction, disease transmission monitoring, and weather forecasting. Although substantial progress has been made in time series forecasting, most existing methods rely on a centralized training paradigm, where large amounts of data are collected from distributed devices (e.g., sensors, wearables) to a central cloud server. However, this paradigm has overloaded communication networks and raised privacy concerns. Federated learning, a popular privacy-preserving technique, enables collaborative model training across distributed data sources. However, directly applying federated learning to time series forecasting often yields suboptimal results, as time series data generated by different devices are inherently heterogeneous. In this paper, we propose a novel framework, Fed-TREND, to address data heterogeneity by generating informative synthetic data as auxiliary knowledge carriers. Specifically, Fed-TREND generates two types of synthetic data. The first type of synthetic data captures the representative distribution information from clients' uploaded model updates and enhances clients' local training consensus. The second kind of synthetic data extracts long-term influence insights from global model update trajectories and is used to refine the global model after aggregation. Fed-TREND is compatible with most time series forecasting models and can be seamlessly integrated into existing federated learning frameworks to improve prediction performance. Extensive experiments on eight datasets, using several federated learning baselines and four popular time series forecasting models, demonstrate the effectiveness and generalizability of Fed-TREND.

ScalingNote: Scaling up Retrievers with Large Language Models for Real-World Dense Retrieval

Dense retrieval in most industries employs dual-tower architectures to retrieve query-relevant documents. Due to online deployment requirements, existing real-world dense retrieval systems mainly enhance performance by designing negative sampling strategies, overlooking the advantages of scaling up. Recently, Large Language Models (LLMs) have exhibited superior performance that can be leveraged for scaling up dense retrieval. However, scaling up retrieval models significantly increases online query latency. To address this challenge, we propose ScalingNote, a two-stage method to exploit the scaling potential of LLMs for retrieval while maintaining online query latency. The first stage is training dual towers, both initialized from the same LLM, to unlock the potential of LLMs for dense retrieval. Then, we distill only the query tower using mean squared error loss and cosine similarity to reduce online costs. Through theoretical analysis and comprehensive offline and online experiments, we show the effectiveness and efficiency of ScalingNote. Our two-stage scaling method outperforms end-to-end models and verifies the scaling law of dense retrieval with LLMs in industrial scenarios, enabling cost-effective scaling of dense retrieval systems. Our online method incorporating ScalingNote significantly enhances the relevance between retrieved documents and queries.

Sparser Training for On-Device Recommendation Systems

Recommender systems often rely on large embedding tables that map users and items to dense vectors of uniform size, leading to substantial memory consumption and inefficiencies. This is particularly problematic in memory-constrained environments like mobile and Web of Things (WoT) applications, where scalability and real-time performance are critical. Various research efforts have sought to address these issues. Although embedding pruning methods utilizing Dynamic Sparse Training (DST) stand out due to their low training and inference costs, consistent sparsity, and end-to-end differentiability, they face key challenges. Firstly, they typically initializes the mask matrix, which is used to prune redundant parameters, with random uniform sparse initialization. This strategy often results in suboptimal performance as it creates unstructured and inefficient connections. Secondly, they tend to favor the users/items sampled in the single batch immediately before weight exploration when they reactivate pruned parameters with large gradient magnitudes, which does not necessarily improve the overall performance. Thirdly, while they use sparse weights during forward passes, they still need to compute dense gradients during backward passes. In this paper, we propose SparseRec, an lightweight embedding method based on DST, to address these issues. Specifically, SparseRec initializes the mask matrix using Nonnegative Matrix Factorization. It accumulates gradients to identify the inactive parameters that can better improve the model performance after activation. Furthermore, it avoids dense gradients during backpropagation by sampling a subset of important vectors. Gradients are calculated only for parameters in this subset, thus maintaining sparsity during training in both forward and backward passes.

DiffSR: Learning Radar Reflectivity Synthesis via Diffusion Model from Satellite Observations

Weather radar data synthesis can fill in data for areas where ground observations are missing. Existing methods often employ reconstruction-based approaches with MSE loss to reconstruct radar data from satellite observation. However, such methods lead to over-smoothing, which hinders the generation of high-frequency details or high-value observation areas associated with convective weather. To address this issue, we propose a two-stage diffusion-based method called DiffSR. We first pre-train a reconstruction model on global-scale data to obtain radar estimation and then synthesize radar reflectivity by combining radar estimation results with satellite data as conditions for the diffusion model. Extensive experiments show that our method achieves state-of-the-art (SOTA) results, demonstrating the ability to generate high-frequency details and high-value areas.

WeatherGFM: Learning A Weather Generalist Foundation Model via In-context Learning

The Earth's weather system encompasses intricate weather data modalities and diverse weather understanding tasks, which hold significant value to human life. Existing data-driven models focus on single weather understanding tasks (e.g., weather forecasting). Although these models have achieved promising results, they fail to tackle various complex tasks within a single and unified model. Moreover, the paradigm that relies on limited real observations for a single scenario hinders the model's performance upper bound. In response to these limitations, we draw inspiration from the in-context learning paradigm employed in state-of-the-art visual foundation models and large language models. In this paper, we introduce the first generalist weather foundation model (WeatherGFM), designed to address a wide spectrum of weather understanding tasks in a unified manner. More specifically, we initially unify the representation and definition of the diverse weather understanding tasks. Subsequently, we devised weather prompt formats to manage different weather data modalities, namely single, multiple, and temporal modalities. Finally, we adopt a visual prompting question-answering paradigm for the training of unified weather understanding tasks. Extensive experiments indicate that our WeatherGFM can effectively handle up to ten weather understanding tasks, including weather forecasting, super-resolution, weather image translation, and post-processing. Our method also showcases generalization ability on unseen tasks.

UniTraj: Universal Human Trajectory Modeling from Billion-Scale Worldwide Traces

Human trajectory modeling is essential for deciphering movement patterns and supporting advanced applications across various domains. However, existing methods are often tailored to specific tasks and regions, resulting in limitations related to task specificity, regional dependency, and data quality sensitivity. Addressing these challenges requires a universal human trajectory foundation model capable of generalizing and scaling across diverse tasks and geographic contexts. To this end, we propose UniTraj, a Universal human Trajectory foundation model that is task-adaptive, region-independent, and highly generalizable. To further enhance performance, we construct WorldTrace, the first large-scale, high-quality, globally distributed dataset sourced from open web platforms, encompassing 2.45 million trajectories with billions of points across 70 countries. Through multiple resampling and masking strategies designed for pre-training, UniTraj effectively overcomes geographic and task constraints, adapting to heterogeneous data quality. Extensive experiments across multiple trajectory analysis tasks and real-world datasets demonstrate that UniTraj consistently outperforms existing approaches in terms of scalability and adaptability. These results underscore the potential of UniTraj as a versatile, robust solution for a wide range of trajectory analysis applications, with WorldTrace serving as an ideal but non-exclusive foundation for training.

LinRec: Linear Attention Mechanism for Long-term Sequential Recommender Systems

Transformer models have achieved remarkable success in sequential recommender systems (SRSs). However, computing the attention matrix in traditional dot-product attention mechanisms results in a quadratic complexity with sequence lengths, leading to high computational costs for long-term sequential recommendation. Motivated by the above observation, we propose a novel L2-Normalized Linear Attention for the Transformer-based Sequential Recommender Systems (LinRec), which theoretically improves efficiency while preserving the learning capabilities of the traditional dot-product attention. Specifically, by thoroughly examining the equivalence conditions of efficient attention mechanisms, we show that LinRec possesses linear complexity while preserving the property of attention mechanisms. In addition, we reveal its latent efficiency properties by interpreting the proposed LinRec mechanism through a statistical lens. Extensive experiments are conducted based on two public benchmark datasets, demonstrating that the combination of LinRec and Transformer models achieves comparable or even superior performance than state-of-the-art Transformer-based SRS models while significantly improving time and memory efficiency.