On-device session-based recommendation systems have been achieving increasing attention on account of the low energy/resource consumption and privacy protection while providing promising recommendation performance. To fit the powerful neural session-based recommendation models in resource-constrained mobile devices, tensor-train decomposition and its variants have been widely applied to reduce memory footprint by decomposing the embedding table into smaller tensors, showing great potential in compressing recommendation models. However, these model compression techniques significantly increase the local inference time due to the complex process of generating index lists and a series of tensor multiplications to form item embeddings, and the resultant on-device recommender fails to provide real-time response and recommendation. To improve the online recommendation efficiency, we propose to learn compositional encoding-based compact item representations. Specifically, each item is represented by a compositional code that consists of several codewords, and we learn embedding vectors to represent each codeword instead of each item. Then the composition of the codeword embedding vectors from different embedding matrices (i.e., codebooks) forms the item embedding. Since the size of codebooks can be extremely small, the recommender model is thus able to fit in resource-constrained devices and meanwhile can save the codebooks for fast local inference.Besides, to prevent the loss of model capacity caused by compression, we propose a bidirectional self-supervised knowledge distillation framework. Extensive experimental results on two benchmark datasets demonstrate that compared with existing methods, the proposed on-device recommender not only achieves an 8x inference speedup with a large compression ratio but also shows superior recommendation performance.
Computer systems hold a large amount of personal data over decades. On the one hand, such data abundance allows breakthroughs in artificial intelligence (AI), especially machine learning (ML) models. On the other hand, it can threaten the privacy of users and weaken the trust between humans and AI. Recent regulations require that private information about a user can be removed from computer systems in general and from ML models in particular upon request (e.g. the "right to be forgotten"). While removing data from back-end databases should be straightforward, it is not sufficient in the AI context as ML models often "remember" the old data. Existing adversarial attacks proved that we can learn private membership or attributes of the training data from the trained models. This phenomenon calls for a new paradigm, namely machine unlearning, to make ML models forget about particular data. It turns out that recent works on machine unlearning have not been able to solve the problem completely due to the lack of common frameworks and resources. In this survey paper, we seek to provide a thorough investigation of machine unlearning in its definitions, scenarios, mechanisms, and applications. Specifically, as a categorical collection of state-of-the-art research, we hope to provide a broad reference for those seeking a primer on machine unlearning and its various formulations, design requirements, removal requests, algorithms, and uses in a variety of ML applications. Furthermore, we hope to outline key findings and trends in the paradigm as well as highlight new areas of research that have yet to see the application of machine unlearning, but could nonetheless benefit immensely. We hope this survey provides a valuable reference for ML researchers as well as those seeking to innovate privacy technologies. Our resources are at https://github.com/tamlhp/awesome-machine-unlearning.
Online Knowledge Distillation (OKD) improves the involved models by reciprocally exploiting the difference between teacher and student. Several crucial bottlenecks over the gap between them -- e.g., Why and when does a large gap harm the performance, especially for student? How to quantify the gap between teacher and student? -- have received limited formal study. In this paper, we propose Switchable Online Knowledge Distillation (SwitOKD), to answer these questions. Instead of focusing on the accuracy gap at test phase by the existing arts, the core idea of SwitOKD is to adaptively calibrate the gap at training phase, namely distillation gap, via a switching strategy between two modes -- expert mode (pause the teacher while keep the student learning) and learning mode (restart the teacher). To possess an appropriate distillation gap, we further devise an adaptive switching threshold, which provides a formal criterion as to when to switch to learning mode or expert mode, and thus improves the student's performance. Meanwhile, the teacher benefits from our adaptive switching threshold and keeps basically on a par with other online arts. We further extend SwitOKD to multiple networks with two basis topologies. Finally, extensive experiments and analysis validate the merits of SwitOKD for classification over the state-of-the-arts. Our code is available at https://github.com/hfutqian/SwitOKD.
Contrastive learning (CL) has recently been demonstrated critical in improving recommendation performance. The fundamental idea of CL-based recommendation models is to maximize the consistency between representations learned from different graph augmentations of the user-item bipartite graph. In such a self-supervised manner, CL-based recommendation models are expected to extract general features from the raw data to tackle the data sparsity issue. Despite the effectiveness of this paradigm, we still have no clue what underlies the performance gains. In this paper, we first reveal that CL enhances recommendation through endowing the model with the ability to learn more evenly distributed user/item representations, which can implicitly alleviate the pervasive popularity bias and promote long-tail items. Meanwhile, we find that the graph augmentations, which were considered a necessity in prior studies, are relatively unreliable and less significant in CL-based recommendation. On top of these findings, we put forward an eXtremely Simple Graph Contrastive Learning method (XSimGCL) for recommendation, which discards the ineffective graph augmentations and instead employs a simple yet effective noise-based embedding augmentation to create views for CL. A comprehensive experimental study on three large and highly sparse benchmark datasets demonstrates that, though the proposed method is extremely simple, it can smoothly adjust the uniformity of learned representations and outperforms its graph augmentation-based counterparts by a large margin in both recommendation accuracy and training efficiency. The code is released at https://github.com/Coder-Yu/SELFRec.
Overfitting has long been considered a common issue to large neural network models in sequential recommendation. In our study, an interesting phenomenon is observed that overfitting is temporary. When the model scale is increased, the trend of the performance firstly ascends, then descends (i.e., overfitting) and finally ascends again, which is named as double ascent in this paper. We therefore raise an assumption that a considerably larger model will generalise better with a higher performance. In an extreme case to infinite-width, performance is expected to reach the limit of this specific structure. Unfortunately, it is impractical to directly build a huge model due to the limit of resources. In this paper, we propose the Overparameterised Recommender (OverRec), which utilises a recurrent neural tangent kernel (RNTK) as a similarity measurement for user sequences to successfully bypass the restriction of hardware for huge models. We further prove that the RNTK for the tied input-output embeddings in recommendation is the same as the RNTK for general untied input-output embeddings, which makes RNTK theoretically suitable for recommendation. Since the RNTK is analytically derived, OverRec does not require any training, avoiding physically building the huge model. Extensive experiments are conducted on four datasets, which verifies the state-of-the-art performance of OverRec.
Multi-modal knowledge graphs (MKGs) include not only the relation triplets, but also related multi-modal auxiliary data (i.e., texts and images), which enhance the diversity of knowledge. However, the natural incompleteness has significantly hindered the applications of MKGs. To tackle the problem, existing studies employ the embedding-based reasoning models to infer the missing knowledge after fusing the multi-modal features. However, the reasoning performance of these methods is limited due to the following problems: (1) ineffective fusion of multi-modal auxiliary features; (2) lack of complex reasoning ability as well as inability to conduct the multi-hop reasoning which is able to infer more missing knowledge. To overcome these problems, we propose a novel model entitled MMKGR (Multi-hop Multi-modal Knowledge Graph Reasoning). Specifically, the model contains the following two components: (1) a unified gate-attention network which is designed to generate effective multi-modal complementary features through sufficient attention interaction and noise reduction; (2) a complementary feature-aware reinforcement learning method which is proposed to predict missing elements by performing the multi-hop reasoning process, based on the features obtained in component (1). The experimental results demonstrate that MMKGR outperforms the state-of-the-art approaches in the MKG reasoning task.
Inductive link prediction (ILP) is to predict links for unseen entities in emerging knowledge graphs (KGs), considering the evolving nature of KGs. A more challenging scenario is that emerging KGs consist of only unseen entities, called as disconnected emerging KGs (DEKGs). Existing studies for DEKGs only focus on predicting enclosing links, i.e., predicting links inside the emerging KG. The bridging links, which carry the evolutionary information from the original KG to DEKG, have not been investigated by previous work so far. To fill in the gap, we propose a novel model entitled DEKG-ILP (Disconnected Emerging Knowledge Graph Oriented Inductive Link Prediction) that consists of the following two components. (1) The module CLRM (Contrastive Learning-based Relation-specific Feature Modeling) is developed to extract global relation-based semantic features that are shared between original KGs and DEKGs with a novel sampling strategy. (2) The module GSM (GNN-based Subgraph Modeling) is proposed to extract the local subgraph topological information around each link in KGs. The extensive experiments conducted on several benchmark datasets demonstrate that DEKG-ILP has obvious performance improvements compared with state-of-the-art methods for both enclosing and bridging link prediction. The source code is available online.
Owing to its nature of scalability and privacy by design, federated learning (FL) has received increasing interest in decentralized deep learning. FL has also facilitated recent research on upscaling and privatizing personalized recommendation services, using on-device data to learn recommender models locally. These models are then aggregated globally to obtain a more performant model, while maintaining data privacy. Typically, federated recommender systems (FRSs) do not consider the lack of resources and data availability at the end-devices. In addition, they assume that the interaction data between users and items is i.i.d. and stationary across end-devices, and that all local recommender models can be directly averaged without considering the user's behavioral diversity. However, in real scenarios, recommendations have to be made on end-devices with sparse interaction data and limited resources. Furthermore, users' preferences are heterogeneous and they frequently visit new items. This makes their personal preferences highly skewed, and the straightforwardly aggregated model is thus ill-posed for such non-i.i.d. data. In this paper, we propose Resource Efficient Federated Recommender System (ReFRS) to enable decentralized recommendation with dynamic and diversified user preferences. On the device side, ReFRS consists of a lightweight self-supervised local model built upon the variational autoencoder for learning a user's temporal preference from a sequence of interacted items. On the server side, ReFRS utilizes a semantic sampler to adaptively perform model aggregation within each identified user cluster. The clustering module operates in an asynchronous and dynamic manner to support efficient global model update and cope with shifting user interests. As a result, ReFRS achieves superior performance in terms of both accuracy and scalability, as demonstrated by comparative experiments.
The propagation of rumours on social media poses an important threat to societies, so that various techniques for rumour detection have been proposed recently. Yet, existing work focuses on \emph{what} entities constitute a rumour, but provides little support to understand \emph{why} the entities have been classified as such. This prevents an effective evaluation of the detected rumours as well as the design of countermeasures. In this work, we argue that explanations for detected rumours may be given in terms of examples of related rumours detected in the past. A diverse set of similar rumours helps users to generalize, i.e., to understand the properties that govern the detection of rumours. Since the spread of rumours in social media is commonly modelled using feature-annotated graphs, we propose a query-by-example approach that, given a rumour graph, extracts the $k$ most similar and diverse subgraphs from past rumours. The challenge is that all of the computations require fast assessment of similarities between graphs. To achieve an efficient and adaptive realization of the approach in a streaming setting, we present a novel graph representation learning technique and report on implementation considerations. Our evaluation experiments show that our approach outperforms baseline techniques in delivering meaningful explanations for various rumour propagation behaviours.
Dynamic graphs refer to graphs whose structure dynamically changes over time. Despite the benefits of learning vertex representations (i.e., embeddings) for dynamic graphs, existing works merely view a dynamic graph as a sequence of changes within the vertex connections, neglecting the crucial asynchronous nature of such dynamics where the evolution of each local structure starts at different times and lasts for various durations. To maintain asynchronous structural evolutions within the graph, we innovatively formulate dynamic graphs as temporal edge sequences associated with joining time of vertices (ToV) and timespan of edges (ToE). Then, a time-aware Transformer is proposed to embed vertices' dynamic connections and ToEs into the learned vertex representations. Meanwhile, we treat each edge sequence as a whole and embed its ToV of the first vertex to further encode the time-sensitive information. Extensive evaluations on several datasets show that our approach outperforms the state-of-the-art in a wide range of graph mining tasks. At the same time, it is very efficient and scalable for embedding large-scale dynamic graphs.