Analyzing and Reducing Catastrophic Forgetting in Parameter Efficient Tuning

Existing research has shown that large language models (LLMs) exhibit remarkable performance in language understanding and generation. However, when LLMs are continuously fine-tuned on complex and diverse domain-specific downstream tasks, the inference performance on historical tasks decreases dramatically, which is known as a catastrophic forgetting problem. A trade-off needs to be kept between learning plasticity and memory stability. Plenty of existing works have explored strategies like memory replay, regularization and parameter isolation, but little is known about the geometric connection of various adjacent minima in the continual LLMs fine-tuning scenarios. In this work, we investigate the geometric connections of different minima through the lens of mode connectivity, which means different minima can be connected by a low-loss valley. Through extensive experiments, we uncover the mode connectivity phenomenon in the LLMs continual learning scenario and find that it can strike a balance between plasticity and stability. Building upon these findings, we propose a simple yet effective method called Interpolation-based LoRA (I-LoRA), which constructs a dual-memory experience replay framework based on LoRA parameter interpolations. Extensive experiments and analysis on eight domain-specific CL benchmarks demonstrate that I-LoRA consistently show significant improvement over the previous state-of-the-art approaches with up to $11\%$ performance gains, providing a strong baseline and insights for future research on the large language model continual learning problem. Our code is available at \url{https://github.com/which47/LLMCL}.

Disambiguated Node Classification with Graph Neural Networks

Graph Neural Networks (GNNs) have demonstrated significant success in learning from graph-structured data across various domains. Despite their great successful, one critical challenge is often overlooked by existing works, i.e., the learning of message propagation that can generalize effectively to underrepresented graph regions. These minority regions often exhibit irregular homophily/heterophily patterns and diverse neighborhood class distributions, resulting in ambiguity. In this work, we investigate the ambiguity problem within GNNs, its impact on representation learning, and the development of richer supervision signals to fight against this problem. We conduct a fine-grained evaluation of GNN, analyzing the existence of ambiguity in different graph regions and its relation with node positions. To disambiguate node embeddings, we propose a novel method, {\method}, which exploits additional optimization guidance to enhance representation learning, particularly for nodes in ambiguous regions. {\method} identifies ambiguous nodes based on temporal inconsistency of predictions and introduces a disambiguation regularization by employing contrastive learning in a topology-aware manner. {\method} promotes discriminativity of node representations and can alleviating semantic mixing caused by message propagation, effectively addressing the ambiguity problem. Empirical results validate the efficiency of {\method} and highlight its potential to improve GNN performance in underrepresented graph regions.

Distribution Consistency based Self-Training for Graph Neural Networks with Sparse Labels

Few-shot node classification poses a significant challenge for Graph Neural Networks (GNNs) due to insufficient supervision and potential distribution shifts between labeled and unlabeled nodes. Self-training has emerged as a widely popular framework to leverage the abundance of unlabeled data, which expands the training set by assigning pseudo-labels to selected unlabeled nodes. Efforts have been made to develop various selection strategies based on confidence, information gain, etc. However, none of these methods takes into account the distribution shift between the training and testing node sets. The pseudo-labeling step may amplify this shift and even introduce new ones, hindering the effectiveness of self-training. Therefore, in this work, we explore the potential of explicitly bridging the distribution shift between the expanded training set and test set during self-training. To this end, we propose a novel Distribution-Consistent Graph Self-Training (DC-GST) framework to identify pseudo-labeled nodes that are both informative and capable of redeeming the distribution discrepancy and formulate it as a differentiable optimization task. A distribution-shift-aware edge predictor is further adopted to augment the graph and increase the model's generalizability in assigning pseudo labels. We evaluate our proposed method on four publicly available benchmark datasets and extensive experiments demonstrate that our framework consistently outperforms state-of-the-art baselines.

T-SaS: Toward Shift-aware Dynamic Adaptation for Streaming Data

In many real-world scenarios, distribution shifts exist in the streaming data across time steps. Many complex sequential data can be effectively divided into distinct regimes that exhibit persistent dynamics. Discovering the shifted behaviors and the evolving patterns underlying the streaming data are important to understand the dynamic system. Existing methods typically train one robust model to work for the evolving data of distinct distributions or sequentially adapt the model utilizing explicitly given regime boundaries. However, there are two challenges: (1) shifts in data streams could happen drastically and abruptly without precursors. Boundaries of distribution shifts are usually unavailable, and (2) training a shared model for all domains could fail to capture varying patterns. This paper aims to solve the problem of sequential data modeling in the presence of sudden distribution shifts that occur without any precursors. Specifically, we design a Bayesian framework, dubbed as T-SaS, with a discrete distribution-modeling variable to capture abrupt shifts of data. Then, we design a model that enable adaptation with dynamic network selection conditioned on that discrete variable. The proposed method learns specific model parameters for each distribution by learning which neurons should be activated in the full network. A dynamic masking strategy is adopted here to support inter-distribution transfer through the overlapping of a set of sparse networks. Extensive experiments show that our proposed method is superior in both accurately detecting shift boundaries to get segments of varying distributions and effectively adapting to downstream forecast or classification tasks.

Skill Disentanglement for Imitation Learning from Suboptimal Demonstrations

Imitation learning has achieved great success in many sequential decision-making tasks, in which a neural agent is learned by imitating collected human demonstrations. However, existing algorithms typically require a large number of high-quality demonstrations that are difficult and expensive to collect. Usually, a trade-off needs to be made between demonstration quality and quantity in practice. Targeting this problem, in this work we consider the imitation of sub-optimal demonstrations, with both a small clean demonstration set and a large noisy set. Some pioneering works have been proposed, but they suffer from many limitations, e.g., assuming a demonstration to be of the same optimality throughout time steps and failing to provide any interpretation w.r.t knowledge learned from the noisy set. Addressing these problems, we propose {\method} by evaluating and imitating at the sub-demonstration level, encoding action primitives of varying quality into different skills. Concretely, {\method} consists of a high-level controller to discover skills and a skill-conditioned module to capture action-taking policies, and is trained following a two-phase pipeline by first discovering skills with all demonstrations and then adapting the controller to only the clean set. A mutual-information-based regularization and a dynamic sub-demonstration optimality estimator are designed to promote disentanglement in the skill space. Extensive experiments are conducted over two gym environments and a real-world healthcare dataset to demonstrate the superiority of {\method} in learning from sub-optimal demonstrations and its improved interpretability by examining learned skills.

Faithful and Consistent Graph Neural Network Explanations with Rationale Alignment

Uncovering rationales behind predictions of graph neural networks (GNNs) has received increasing attention over recent years. Instance-level GNN explanation aims to discover critical input elements, like nodes or edges, that the target GNN relies upon for making predictions. %These identified sub-structures can provide interpretations of GNN's behavior. Though various algorithms are proposed, most of them formalize this task by searching the minimal subgraph which can preserve original predictions. However, an inductive bias is deep-rooted in this framework: several subgraphs can result in the same or similar outputs as the original graphs. Consequently, they have the danger of providing spurious explanations and failing to provide consistent explanations. Applying them to explain weakly-performed GNNs would further amplify these issues. To address this problem, we theoretically examine the predictions of GNNs from the causality perspective. Two typical reasons for spurious explanations are identified: confounding effect of latent variables like distribution shift, and causal factors distinct from the original input. Observing that both confounding effects and diverse causal rationales are encoded in internal representations, \tianxiang{we propose a new explanation framework with an auxiliary alignment loss, which is theoretically proven to be optimizing a more faithful explanation objective intrinsically. Concretely for this alignment loss, a set of different perspectives are explored: anchor-based alignment, distributional alignment based on Gaussian mixture models, mutual-information-based alignment, etc. A comprehensive study is conducted both on the effectiveness of this new framework in terms of explanation faithfulness/consistency and on the advantages of these variants.

TopoImb: Toward Topology-level Imbalance in Learning from Graphs

Graph serves as a powerful tool for modeling data that has an underlying structure in non-Euclidean space, by encoding relations as edges and entities as nodes. Despite developments in learning from graph-structured data over the years, one obstacle persists: graph imbalance. Although several attempts have been made to target this problem, they are limited to considering only class-level imbalance. In this work, we argue that for graphs, the imbalance is likely to exist at the sub-class topology group level. Due to the flexibility of topology structures, graphs could be highly diverse, and learning a generalizable classification boundary would be difficult. Therefore, several majority topology groups may dominate the learning process, rendering others under-represented. To address this problem, we propose a new framework {\method} and design (1 a topology extractor, which automatically identifies the topology group for each instance with explicit memory cells, (2 a training modulator, which modulates the learning process of the target GNN model to prevent the case of topology-group-wise under-representation. {\method} can be used as a key component in GNN models to improve their performances under the data imbalance setting. Analyses on both topology-level imbalance and the proposed {\method} are provided theoretically, and we empirically verify its effectiveness with both node-level and graph-level classification as the target tasks.

Explanation Guided Contrastive Learning for Sequential Recommendation

Recently, contrastive learning has been applied to the sequential recommendation task to address data sparsity caused by users with few item interactions and items with few user adoptions. Nevertheless, the existing contrastive learning-based methods fail to ensure that the positive (or negative) sequence obtained by some random augmentation (or sequence sampling) on a given anchor user sequence remains to be semantically similar (or different). When the positive and negative sequences turn out to be false positive and false negative respectively, it may lead to degraded recommendation performance. In this work, we address the above problem by proposing Explanation Guided Augmentations (EGA) and Explanation Guided Contrastive Learning for Sequential Recommendation (EC4SRec) model framework. The key idea behind EGA is to utilize explanation method(s) to determine items' importance in a user sequence and derive the positive and negative sequences accordingly. EC4SRec then combines both self-supervised and supervised contrastive learning over the positive and negative sequences generated by EGA operations to improve sequence representation learning for more accurate recommendation results. Extensive experiments on four real-world benchmark datasets demonstrate that EC4SRec outperforms the state-of-the-art sequential recommendation methods and two recent contrastive learning-based sequential recommendation methods, CL4SRec and DuoRec. Our experiments also show that EC4SRec can be easily adapted for different sequence encoder backbones (e.g., GRU4Rec and Caser), and improve their recommendation performance.

Synthetic Over-sampling for Imbalanced Node Classification with Graph Neural Networks

In recent years, graph neural networks (GNNs) have achieved state-of-the-art performance for node classification. However, most existing GNNs would suffer from the graph imbalance problem. In many real-world scenarios, node classes are imbalanced, with some majority classes making up most parts of the graph. The message propagation mechanism in GNNs would further amplify the dominance of those majority classes, resulting in sub-optimal classification performance. In this work, we seek to address this problem by generating pseudo instances of minority classes to balance the training data, extending previous over-sampling-based techniques. This task is non-trivial, as those techniques are designed with the assumption that instances are independent. Neglection of relation information would complicate this oversampling process. Furthermore, the node classification task typically takes the semi-supervised setting with only a few labeled nodes, providing insufficient supervision for the generation of minority instances. Generated new nodes of low quality would harm the trained classifier. In this work, we address these difficulties by synthesizing new nodes in a constructed embedding space, which encodes both node attributes and topology information. Furthermore, an edge generator is trained simultaneously to model the graph structure and provide relations for new samples. To further improve the data efficiency, we also explore synthesizing mixed ``in-between'' nodes to utilize nodes from the majority class in this over-sampling process. Experiments on real-world datasets validate the effectiveness of our proposed framework.

On Consistency in Graph Neural Network Interpretation

Uncovering rationales behind predictions of graph neural networks (GNNs) has received increasing attention over recent years. Instance-level GNN explanation aims to discover critical input elements, like nodes or edges, that the target GNN relies upon for making predictions. These identified sub-structures can provide interpretations of GNN's behavior. Though various algorithms are proposed, most of them formalize this task by searching the minimal subgraph which can preserve original predictions. An inductive bias is deep-rooted in this framework: the same output cannot guarantee that two inputs are processed under the same rationale. Consequently, they have the danger of providing spurious explanations and fail to provide consistent explanations. Applying them to explain weakly-performed GNNs would further amplify these issues. To address the issues, we propose to obtain more faithful and consistent explanations of GNNs. After a close examination on predictions of GNNs from the causality perspective, we attribute spurious explanations to two typical reasons: confounding effect of latent variables like distribution shift, and causal factors distinct from the original input. Motivated by the observation that both confounding effects and diverse causal rationales are encoded in internal representations, we propose a simple yet effective countermeasure by aligning embeddings. This new objective can be incorporated into existing GNN explanation algorithms with no effort. We implement both a simplified version based on absolute distance and a distribution-aware version based on anchors. Experiments on $5$ datasets validate its effectiveness, and theoretical analysis shows that it is in effect optimizing a more faithful explanation objective in design, which further justifies the proposed approach.