Fajie Yuan

Mingyang Hu, Fajie Yuan, Kevin K. Yang et al. · microsoft-research

Large-scale Protein Language Models (PLMs) have improved performance in protein prediction tasks, ranging from 3D structure prediction to various function predictions. In particular, AlphaFold, a ground-breaking AI system, could potentially reshape structural biology. However, the utility of the PLM module in AlphaFold, Evoformer, has not been explored beyond structure prediction. In this paper, we investigate the representation ability of three popular PLMs: ESM-1b (single sequence), MSA-Transformer (multiple sequence alignment) and Evoformer (structural), with a special focus on Evoformer. Specifically, we aim to answer the following key questions: (i) Does the Evoformer trained as part of AlphaFold produce representations amenable to predicting protein function? (ii) If yes, can Evoformer replace ESM-1b and MSA-Transformer? (ii) How much do these PLMs rely on evolution-related protein data? In this regard, are they complementary to each other? We compare these models by empirical study along with new insights and conclusions. All code and datasets for reproducibility are available at https://github.com/elttaes/Revisiting-PLMs.

Jin Su, Runnan Fang, Yeqiu Li et al.

Large language model (LLM)-based agents have been successfully deployed in many tool-augmented settings, but their scalability is fundamentally constrained by context length. Existing context-folding methods mitigate this issue by summarizing past interactions, yet they are typically designed for single-query or single-intent scenarios. In more realistic user-centric dialogues, we identify two major failure modes: (i) they irreversibly discard fine-grained constraints and intermediate facts that are crucial for later decisions, and (ii) their summaries fail to track evolving user intent, leading to omissions and erroneous actions. To address these limitations, we propose U-Fold, a dynamic context-folding framework tailored to user-centric tasks. U-Fold retains the full user--agent dialogue and tool-call history but, at each turn, uses two core components to produce an intent-aware, evolving dialogue summary and a compact, task-relevant tool log. Extensive experiments on $τ$-bench, $τ^2$-bench, VitaBench, and harder context-inflated settings show that U-Fold consistently outperforms ReAct (achieving a 71.4% win rate in long-context settings) and prior folding baselines (with improvements of up to 27.0%), particularly on long, noisy, multi-turn tasks. Our study demonstrates that U-Fold is a promising step toward transferring context-management techniques from single-query benchmarks to realistic user-centric applications.

Yu Cheng, Fajie Yuan

Recent advances in Latent Video Diffusion Models (LVDMs) have revolutionized video generation by leveraging Video Variational Autoencoders (Video VAEs) to compress intricate video data into a compact latent space. However, as LVDM training scales, the computational overhead of Video VAEs becomes a critical bottleneck, particularly for encoding high-resolution videos. To address this, we propose LeanVAE, a novel and ultra-efficient Video VAE framework that introduces two key innovations: (1) a lightweight architecture based on a Neighborhood-Aware Feedforward (NAF) module and non-overlapping patch operations, drastically reducing computational cost, and (2) the integration of wavelet transforms and compressed sensing techniques to enhance reconstruction quality. Extensive experiments validate LeanVAE's superiority in video reconstruction and generation, particularly in enhancing efficiency over existing Video VAEs. Our model offers up to 50x fewer FLOPs and 44x faster inference speed while maintaining competitive reconstruction quality, providing insights for scalable, efficient video generation. Our models and code are available at https://github.com/westlake-repl/LeanVAE

Ziqi Zhao, Zhaochun Ren, Liu Yang et al.

Offline reinforcement learning (RL) aims to learn policies without online explorations. To enlarge the training data, model-based offline RL learns a dynamics model which is utilized as a virtual environment to generate simulation data and enhance policy learning. However, existing data augmentation methods for offline RL suffer from (i) trivial improvement from short-horizon simulation; and (ii) the lack of evaluation and correction for generated data, leading to low-qualified augmentation. In this paper, we propose offline trajectory optimization for offline reinforcement learning (OTTO). The key motivation is to conduct long-horizon simulation and then utilize model uncertainty to evaluate and correct the augmented data. Specifically, we propose an ensemble of Transformers, a.k.a. World Transformers, to predict environment state dynamics and the reward function. Three strategies are proposed to use World Transformers to generate long-horizon trajectory simulation by perturbing the actions in the offline data. Then, an uncertainty-based World Evaluator is introduced to firstly evaluate the confidence of the generated trajectories and then perform the correction for low-confidence data. Finally, we jointly use the original data with the corrected augmentation data to train an offline RL algorithm. OTTO serves as a plug-in module and can be integrated with existing model-free offline RL methods. Experiments on various benchmarks show that OTTO can effectively improve the performance of representative offline RL algorithms, including in complex environments with sparse rewards like AntMaze. Codes are available at https://github.com/ZiqiZhao1/OTTO.

Binzong Geng, Fajie Yuan, Qiancheng Xu et al.

This ability to learn consecutive tasks without forgetting how to perform previously trained problems is essential for developing an online dialogue system. This paper proposes an effective continual learning for the task-oriented dialogue system with iterative network pruning, expanding and masking (TPEM), which preserves performance on previously encountered tasks while accelerating learning progress on subsequent tasks. Specifically, TPEM (i) leverages network pruning to keep the knowledge for old tasks, (ii) adopts network expanding to create free weights for new tasks, and (iii) introduces task-specific network masking to alleviate the negative impact of fixed weights of old tasks on new tasks. We conduct extensive experiments on seven different tasks from three benchmark datasets and show empirically that TPEM leads to significantly improved results over the strong competitors. For reproducibility, we submit the code and data at: https://github.com/siat-nlp/TPEM

Binzong Geng, Min Yang, Fajie Yuan et al.

Lifelong learning capabilities are crucial for sentiment classifiers to process continuous streams of opinioned information on the Web. However, performing lifelong learning is non-trivial for deep neural networks as continually training of incrementally available information inevitably results in catastrophic forgetting or interference. In this paper, we propose a novel iterative network pruning with uncertainty regularization method for lifelong sentiment classification (IPRLS), which leverages the principles of network pruning and weight regularization. By performing network pruning with uncertainty regularization in an iterative manner, IPRLS can adapta single BERT model to work with continuously arriving data from multiple domains while avoiding catastrophic forgetting and interference. Specifically, we leverage an iterative pruning method to remove redundant parameters in large deep networks so that the freed-up space can then be employed to learn new tasks, tackling the catastrophic forgetting problem. Instead of keeping the old-tasks fixed when learning new tasks, we also use an uncertainty regularization based on the Bayesian online learning framework to constrain the update of old tasks weights in BERT, which enables positive backward transfer, i.e. learning new tasks improves performance on past tasks while protecting old knowledge from being lost. In addition, we propose a task-specific low-dimensional residual function in parallel to each layer of BERT, which makes IPRLS less prone to losing the knowledge saved in the base BERT network when learning a new task. Extensive experiments on 16 popular review corpora demonstrate that the proposed IPRLS method sig-nificantly outperforms the strong baselines for lifelong sentiment classification. For reproducibility, we submit the code and data at:https://github.com/siat-nlp/IPRLS.

Jiachun Wang, Fajie Yuan, Jian Chen et al.

Deep learning has brought great progress for the sequential recommendation (SR) tasks. With advanced network architectures, sequential recommender models can be stacked with many hidden layers, e.g., up to 100 layers on real-world recommendation datasets. Training such a deep network is difficult because it can be computationally very expensive and takes much longer time, especially in situations where there are tens of billions of user-item interactions. To deal with such a challenge, we present StackRec, a simple, yet very effective and efficient training framework for deep SR models by iterative layer stacking. Specifically, we first offer an important insight that hidden layers/blocks in a well-trained deep SR model have very similar distributions. Enlightened by this, we propose the stacking operation on the pre-trained layers/blocks to transfer knowledge from a shallower model to a deep model, then we perform iterative stacking so as to yield a much deeper but easier-to-train SR model. We validate the performance of StackRec by instantiating it with four state-of-the-art SR models in three practical scenarios with real-world datasets. Extensive experiments show that StackRec achieves not only comparable performance, but also substantial acceleration in training time, compared to SR models that are trained from scratch. Codes are available at https://github.com/wangjiachun0426/StackRec.

Fajie Yuan, Xiangnan He, Alexandros Karatzoglou et al.

Inductive transfer learning has had a big impact on computer vision and NLP domains but has not been used in the area of recommender systems. Even though there has been a large body of research on generating recommendations based on modeling user-item interaction sequences, few of them attempt to represent and transfer these models for serving downstream tasks where only limited data exists. In this paper, we delve on the task of effectively learning a single user representation that can be applied to a diversity of tasks, from cross-domain recommendations to user profile predictions. Fine-tuning a large pre-trained network and adapting it to downstream tasks is an effective way to solve such tasks. However, fine-tuning is parameter inefficient considering that an entire model needs to be re-trained for every new task. To overcome this issue, we develop a parameter efficient transfer learning architecture, termed as PeterRec, which can be configured on-the-fly to various downstream tasks. Specifically, PeterRec allows the pre-trained parameters to remain unaltered during fine-tuning by injecting a series of re-learned neural networks, which are small but as expressive as learning the entire network. We perform extensive experimental ablation to show the effectiveness of the learned user representation in five downstream tasks. Moreover, we show that PeterRec performs efficient transfer learning in multiple domains, where it achieves comparable or sometimes better performance relative to fine-tuning the entire model parameters. Codes and datasets are available at https://github.com/fajieyuan/sigir2020_peterrec.

Jinhui Tang, Xiaoyu Du, Xiangnan He et al.

With the prevalence of multimedia content on the Web, developing recommender solutions that can effectively leverage the rich signal in multimedia data is in urgent need. Owing to the success of deep neural networks in representation learning, recent advance on multimedia recommendation has largely focused on exploring deep learning methods to improve the recommendation accuracy. To date, however, there has been little effort to investigate the robustness of multimedia representation and its impact on the performance of multimedia recommendation. In this paper, we shed light on the robustness of multimedia recommender system. Using the state-of-the-art recommendation framework and deep image features, we demonstrate that the overall system is not robust, such that a small (but purposeful) perturbation on the input image will severely decrease the recommendation accuracy. This implies the possible weakness of multimedia recommender system in predicting user preference, and more importantly, the potential of improvement by enhancing its robustness. To this end, we propose a novel solution named Adversarial Multimedia Recommendation (AMR), which can lead to a more robust multimedia recommender model by using adversarial learning. The idea is to train the model to defend an adversary, which adds perturbations to the target image with the purpose of decreasing the model's accuracy. We conduct experiments on two representative multimedia recommendation tasks, namely, image recommendation and visually-aware product recommendation. Extensive results verify the positive effect of adversarial learning and demonstrate the effectiveness of our AMR method. Source codes are available in https://github.com/duxy-me/AMR.

Binzong Geng, Zhaoxin Huan, Xiaolu Zhang et al.

With the rise of large language models (LLMs), recent works have leveraged LLMs to improve the performance of click-through rate (CTR) prediction. However, we argue that a critical obstacle remains in deploying LLMs for practical use: the efficiency of LLMs when processing long textual user behaviors. As user sequences grow longer, the current efficiency of LLMs is inadequate for training on billions of users and items. To break through the efficiency barrier of LLMs, we propose Behavior Aggregated Hierarchical Encoding (BAHE) to enhance the efficiency of LLM-based CTR modeling. Specifically, BAHE proposes a novel hierarchical architecture that decouples the encoding of user behaviors from inter-behavior interactions. Firstly, to prevent computational redundancy from repeated encoding of identical user behaviors, BAHE employs the LLM's pre-trained shallow layers to extract embeddings of the most granular, atomic user behaviors from extensive user sequences and stores them in the offline database. Subsequently, the deeper, trainable layers of the LLM facilitate intricate inter-behavior interactions, thereby generating comprehensive user embeddings. This separation allows the learning of high-level user representations to be independent of low-level behavior encoding, significantly reducing computational complexity. Finally, these refined user embeddings, in conjunction with correspondingly processed item embeddings, are incorporated into the CTR model to compute the CTR scores. Extensive experimental results show that BAHE reduces training time and memory by five times for CTR models using LLMs, especially with longer user sequences. BAHE has been deployed in a real-world system, allowing for daily updates of 50 million CTR data on 8 A100 GPUs, making LLMs practical for industrial CTR prediction.

Fengyuan Dai, Zifeng Zhuang, Yufei Huang et al.

Diffusion models have emerged as powerful generative tools across various domains, yet tailoring pre-trained models to exhibit specific desirable properties remains challenging. While reinforcement learning (RL) offers a promising solution,current methods struggle to simultaneously achieve stable, efficient fine-tuning and support non-differentiable rewards. Furthermore, their reliance on sparse rewards provides inadequate supervision during intermediate steps, often resulting in suboptimal generation quality. To address these limitations, dense and differentiable signals are required throughout the diffusion process. Hence, we propose VAlue-based Reinforced Diffusion (VARD): a novel approach that first learns a value function predicting expection of rewards from intermediate states, and subsequently uses this value function with KL regularization to provide dense supervision throughout the generation process. Our method maintains proximity to the pretrained model while enabling effective and stable training via backpropagation. Experimental results demonstrate that our approach facilitates better trajectory guidance, improves training efficiency and extends the applicability of RL to diffusion models optimized for complex, non-differentiable reward functions.

Yang Sun, Fajie Yuan, Min Yang et al.

Deep neural networks (DNN) have achieved great success in the recommender systems (RS) domain. However, to achieve remarkable performance, DNN-based recommender models often require numerous parameters, which inevitably bring redundant neurons and weights, a phenomenon referred to as over-parameterization. In this paper, we plan to exploit such redundancy phenomena to improve the performance of RS. Specifically, we propose PCRec, a top-N item \underline{rec}ommendation framework that leverages collaborative training of two DNN-based recommender models with the same network structure, termed \underline{p}eer \underline{c}ollaboration. PCRec can reactivate and strengthen the unimportant (redundant) weights during training, which achieves higher prediction accuracy but maintains its original inference efficiency. To realize this, we first introduce two criteria to identify the importance of weights of a given recommender model. Then, we rejuvenate the unimportant weights by transplanting outside information (i.e., weights) from its peer network. After such an operation and retraining, the original recommender model is endowed with more representation capacity by possessing more functional model parameters. To show its generality, we instantiate PCRec by using three well-known recommender models. We conduct extensive experiments on three real-world datasets, and show that PCRec yields significantly better recommendations than its counterpart with the same model (parameter) size.

Lei Chen, Fajie Yuan, Jiaxi Yang et al.

Sequential recommender systems (SRS) have become a research hotspot due to its power in modeling user dynamic interests and sequential behavioral patterns. To maximize model expressive ability, a default choice is to apply a larger and deeper network architecture, which, however, often brings high network latency when generating online recommendations. Naturally, we argue that compressing the heavy recommendation models into middle- or light- weight neural networks is of great importance for practical production systems. To realize such a goal, we propose AdaRec, a knowledge distillation (KD) framework which compresses knowledge of a teacher model into a student model adaptively according to its recommendation scene by using differentiable Neural Architecture Search (NAS). Specifically, we introduce a target-oriented distillation loss to guide the structure search process for finding the student network architecture, and a cost-sensitive loss as constraints for model size, which achieves a superior trade-off between recommendation effectiveness and efficiency. In addition, we leverage Earth Mover's Distance (EMD) to realize many-to-many layer mapping during knowledge distillation, which enables each intermediate student layer to learn from other intermediate teacher layers adaptively. Extensive experiments on real-world recommendation datasets demonstrate that our model achieves competitive or better accuracy with notable inference speedup comparing to strong counterparts, while discovering diverse neural architectures for sequential recommender models under different recommendation scenes.

Lei Chen, Fajie Yuan, Jiaxi Yang et al.

Making accurate recommendations for cold-start users has been a longstanding and critical challenge for recommender systems (RS). Cross-domain recommendations (CDR) offer a solution to tackle such a cold-start problem when there is no sufficient data for the users who have rarely used the system. An effective approach in CDR is to leverage the knowledge (e.g., user representations) learned from a related but different domain and transfer it to the target domain. Fine-tuning works as an effective transfer learning technique for this objective, which adapts the parameters of a pre-trained model from the source domain to the target domain. However, current methods are mainly based on the global fine-tuning strategy: the decision of which layers of the pre-trained model to freeze or fine-tune is taken for all users in the target domain. In this paper, we argue that users in RS are personalized and should have their own fine-tuning policies for better preference transfer learning. As such, we propose a novel User-specific Adaptive Fine-tuning method (UAF), selecting which layers of the pre-trained network to fine-tune, on a per-user basis. Specifically, we devise a policy network with three alternative strategies to automatically decide which layers to be fine-tuned and which layers to have their parameters frozen for each user. Extensive experiments show that the proposed UAF exhibits significantly better and more robust performance for user cold-start recommendation.

Fajie Yuan, Guoxiao Zhang, Alexandros Karatzoglou et al.

Learning user representations is a vital technique toward effective user modeling and personalized recommender systems. Existing approaches often derive an individual set of model parameters for each task by training on separate data. However, the representation of the same user potentially has some commonalities, such as preference and personality, even in different tasks. As such, these separately trained representations could be suboptimal in performance as well as inefficient in terms of parameter sharing. In this paper, we delve on research to continually learn user representations task by task, whereby new tasks are learned while using partial parameters from old ones. A new problem arises since when new tasks are trained, previously learned parameters are very likely to be modified, and as a result, an artificial neural network (ANN)-based model may lose its capacity to serve for well-trained previous tasks forever, this issue is termed catastrophic forgetting. To address this issue, we present \emph{Conure} the first \underline{con}tinual, or lifelong, \underline{u}ser \underline{re}presentation learner -- i.e., learning new tasks over time without forgetting old ones. Specifically, we propose iteratively removing less important weights of old tasks in a deep user representation model, motivated by the fact that neural network models are usually over-parameterized. In this way, we could learn many tasks with a single model by reusing the important weights, and modifying the less important weights to adapt to new tasks. We conduct extensive experiments on two real-world datasets with nine tasks and show that \emph{Conure} largely exceeds the standard model that does not purposely preserve such old "knowledge", and performs competitively or sometimes better than models which are trained either individually for each task or simultaneously by merging all task data.

Shilin Qu, Fajie Yuan, Guibing Guo et al.

Recently, Memory-based Neural Recommenders (MNR) have demonstrated superior predictive accuracy in the task of sequential recommendations, particularly for modeling long-term item dependencies. However, typical MNR requires complex memory access operations, i.e., both writing and reading via a controller (e.g., RNN) at every time step. Those frequent operations will dramatically increase the network training time, resulting in the difficulty in being deployed on industrial-scale recommender systems. In this paper, we present a novel general Chunk framework to accelerate MNR significantly. Specifically, our framework divides proximal information units into chunks, and performs memory access at certain time steps, whereby the number of memory operations can be greatly reduced. We investigate two ways to implement effective chunking, i.e., PEriodic Chunk (PEC) and Time-Sensitive Chunk (TSC), to preserve and recover important recurrent signals in the sequence. Since chunk-accelerated MNR models take into account more proximal information units than that from a single timestep, it can remove the influence of noise in the item sequence to a large extent, and thus improve the stability of MNR. In this way, the proposed chunk mechanism can lead to not only faster training and prediction, but even slightly better results. The experimental results on three real-world datasets (weishi, ml-10M and ml-latest) show that our chunk framework notably reduces the running time (e.g., with up to 7x for training & 10x for inference on ml-latest) of MNR, and meantime achieves competitive performance.

Yang Sun, Fajie Yuan, Min Yang et al.

Sequential recommender systems (SRS) have become the key technology in capturing user's dynamic interests and generating high-quality recommendations. Current state-of-the-art sequential recommender models are typically based on a sandwich-structured deep neural network, where one or more middle (hidden) layers are placed between the input embedding layer and output softmax layer. In general, these models require a large number of parameters (such as using a large embedding dimension or a deep network architecture) to obtain their optimal performance. Despite the effectiveness, at some point, further increasing model size may be harder for model deployment in resource-constraint devices, resulting in longer responding time and larger memory footprint. To resolve the issues, we propose a compressed sequential recommendation framework, termed as CpRec, where two generic model shrinking techniques are employed. Specifically, we first propose a block-wise adaptive decomposition to approximate the input and softmax matrices by exploiting the fact that items in SRS obey a long-tailed distribution. To reduce the parameters of the middle layers, we introduce three layer-wise parameter sharing schemes. We instantiate CpRec using deep convolutional neural network with dilated kernels given consideration to both recommendation accuracy and efficiency. By the extensive ablation studies, we demonstrate that the proposed CpRec can achieve up to 4$\sim$8 times compression rates in real-world SRS datasets. Meanwhile, CpRec is faster during training\inference, and in most cases outperforms its uncompressed counterpart.

Xiaoyu Du, Xiangnan He, Fajie Yuan et al.

As the core of recommender system, collaborative filtering (CF) models the affinity between a user and an item from historical user-item interactions, such as clicks, purchases, and so on. Benefited from the strong representation power, neural networks have recently revolutionized the recommendation research, setting up a new standard for CF. However, existing neural recommender models do not explicitly consider the correlations among embedding dimensions, making them less effective in modeling the interaction function between users and items. In this work, we emphasize on modeling the correlations among embedding dimensions in neural networks to pursue higher effectiveness for CF. We propose a novel and general neural collaborative filtering framework, namely ConvNCF, which is featured with two designs: 1) applying outer product on user embedding and item embedding to explicitly model the pairwise correlations between embedding dimensions, and 2) employing convolutional neural network above the outer product to learn the high-order correlations among embedding dimensions. To justify our proposal, we present three instantiations of ConvNCF by using different inputs to represent a user and conduct experiments on two real-world datasets. Extensive results verify the utility of modeling embedding dimension correlations with ConvNCF, which outperforms several competitive CF methods.

Fajie Yuan, Xiangnan He, Haochuan Jiang et al.

Session-based recommender systems have attracted much attention recently. To capture the sequential dependencies, existing methods resort either to data augmentation techniques or left-to-right style autoregressive training.Since these methods are aimed to model the sequential nature of user behaviors, they ignore the future data of a target interaction when constructing the prediction model for it. However, we argue that the future interactions after a target interaction, which are also available during training, provide valuable signal on user preference and can be used to enhance the recommendation quality. Properly integrating future data into model training, however, is non-trivial to achieve, since it disobeys machine learning principles and can easily cause data leakage. To this end, we propose a new encoder-decoder framework named Gap-filling based Recommender (GRec), which trains the encoder and decoder by a gap-filling mechanism. Specifically, the encoder takes a partially-complete session sequence (where some items are masked by purpose) as input, and the decoder predicts these masked items conditioned on the encoded representation. We instantiate the general GRec framework using convolutional neural network with sparse kernels, giving consideration to both accuracy and efficiency. We conduct experiments on two real-world datasets covering short-, medium-, and long-range user sessions, showing that GRec significantly outperforms the state-of-the-art sequential recommendation methods. More empirical studies verify the high utility of modeling future contexts under our GRec framework.

Fajie Yuan, Alexandros Karatzoglou, Ioannis Arapakis et al.

Convolutional Neural Networks (CNNs) have been recently introduced in the domain of session-based next item recommendation. An ordered collection of past items the user has interacted with in a session (or sequence) are embedded into a 2-dimensional latent matrix, and treated as an image. The convolution and pooling operations are then applied to the mapped item embeddings. In this paper, we first examine the typical session-based CNN recommender and show that both the generative model and network architecture are suboptimal when modeling long-range dependencies in the item sequence. To address the issues, we introduce a simple, but very effective generative model that is capable of learning high-level representation from both short- and long-range item dependencies. The network architecture of the proposed model is formed of a stack of \emph{holed} convolutional layers, which can efficiently increase the receptive fields without relying on the pooling operation. Another contribution is the effective use of residual block structure in recommender systems, which can ease the optimization for much deeper networks. The proposed generative model attains state-of-the-art accuracy with less training time in the next item recommendation task. It accordingly can be used as a powerful recommendation baseline to beat in future, especially when there are long sequences of user feedback.

Guibing Guo, Songlin Zhai, Fajie Yuan et al.

Jointing visual-semantic embeddings (VSE) have become a research hotpot for the task of image annotation, which suffers from the issue of semantic gap, i.e., the gap between images' visual features (low-level) and labels' semantic features (high-level). This issue will be even more challenging if visual features cannot be retrieved from images, that is, when images are only denoted by numerical IDs as given in some real datasets. The typical way of existing VSE methods is to perform a uniform sampling method for negative examples that violate the ranking order against positive examples, which requires a time-consuming search in the whole label space. In this paper, we propose a fast adaptive negative sampler that can work well in the settings of no figure pixels available. Our sampling strategy is to choose the negative examples that are most likely to meet the requirements of violation according to the latent factors of images. In this way, our approach can linearly scale up to large datasets. The experiments demonstrate that our approach converges 5.02x faster than the state-of-the-art approaches on OpenImages, 2.5x on IAPR-TCI2 and 2.06x on NUS-WIDE datasets, as well as better ranking accuracy across datasets.

Long Chen, Fajie Yuan, Joemon M. Jose et al.

Although the word-popularity based negative sampler has shown superb performance in the skip-gram model, the theoretical motivation behind oversampling popular (non-observed) words as negative samples is still not well understood. In this paper, we start from an investigation of the gradient vanishing issue in the skipgram model without a proper negative sampler. By performing an insightful analysis from the stochastic gradient descent (SGD) learning perspective, we demonstrate that, both theoretically and intuitively, negative samples with larger inner product scores are more informative than those with lower scores for the SGD learner in terms of both convergence rate and accuracy. Understanding this, we propose an alternative sampling algorithm that dynamically selects informative negative samples during each SGD update. More importantly, the proposed sampler accounts for multi-dimensional self-embedded features during the sampling process, which essentially makes it more effective than the original popularity-based (one-dimensional) sampler. Empirical experiments further verify our observations, and show that our fine-grained samplers gain significant improvement over the existing ones without increasing computational complexity.