Knowledge base question answering (KBQA) aims to answer natural language questions using large-scale knowledge bases (KBs). Among various KBQA approaches, semantic parsing-based (SP-based) methods have demonstrated strong effectiveness by generating concise logical forms (LFs) that capture complex subgraph structures and semantic information. Recent research suggests that integrating large language models (LLMs) with SP can achieve significant improvements in the performance and efficiency of KBQA by facilitating the direct generation of LFs with minimal retrieval. However, generating complete LFs with LLMs continues to pose a challenge due to the complexity of the required graph structures and constraints, leading to the significant issue of non-executability. To address these challenges, we propose GCA-KBQA, a step-wise fine-tuned LLM-based framework that employs hop-wise generation, knowledge-assisted calibration, and path-level assembly to construct complete LFs for KBQA. Specifically, we decompose the complex SP process into manageable steps: first, we iteratively generate LFs for each topic entity one hop at a time using a fine-tuned LLM, leveraging KB knowledge to calibrate intermediate outputs and mitigate error propagation. Subsequently, we guide the LLM in assembling path-level LFs from different topic entities, resulting in optimized final LF. We evaluate the proposed method on four KBQA benchmarks spanning two distinct KBs, demonstrating its superior performance compared to state-of-the-art baselines. The code is available at https://github.com/pvfeldt/GCA-KBQA.

Practical applications often require forecasting the future states of short time series (STS) using multiple related long time series (LTS) as auxiliary data, a process known as few-shot prediction. The primary challenge, given the limited data on STS, is effectively capturing the pattern similarities between STS and LTS. Current methods, despite notable advancements, primarily focus on trans- ferring pattern characteristics from LTS to STS without explicitly addressing their similarities at various levels. To overcome this lim- itation, we propose a novel few-shot time series forecasting model called DANet. Drawing on the Retrieval-Augmented Generation (RAG) framework in large language models, DANet retrieves long and short sequences from LTS that closely resemble STS, thereby enhancing prediction accuracy while simultaneously reducing un- certainty through this retrieval process. First, we define two metrics to quantify pattern similarities between STS and LTS, addressing the issue of different representations of the same pattern due to variations in sequence length. Second, we propose a dual-attention mechanism which embeds the two similarities metrics to extract and integrate long and short sequences from LTS across variable and temporal levels for generating predictions. Our experiments across six scenarios show that DANet significantly outperforms six state-of-the-art (SOTA) methods.

The recent years have witnessed a surge in the development of traffic flow prediction methods, often deployed on cloud platforms to offer predictive services for entire transportation networks. However, the processes of training and executing a model for the entire traffic network are both time-consuming and computationally expensive. As a result, the utilization of edge servers for local sub-network prediction services has gained prominence. Nevertheless, training prediction models for numerous sub-networks within the extensive traffic network remains a time-intensive and computing resource-consuming task. To tackle this challenge, this article introduces the Pre-trained model REcommendation Framework for Edge computing enabled tRaffic flow prediction (PREFER). PREFER trains a set of traffic flow prediction models on selected sub-networks, then recommends optimal pre-trained models for edge servers. The recommendation is specifically based on performance prediction, integrating neural collaborative filtering and traffic flow characteristics. Experiments conducted on real datasets reveal that the pre-trained models recommended by PREFER perform close to the actual optimal ones and significantly outperform existing recommendation algorithms.

Event sequence anomaly detection has garnered considerable attention in research, encompassing applications such as identifying anomalies in system logs, anomalous transaction users, and so on. Yet, prevailing anomaly detection methods often rely solely on local data for training, potentially leading to imperfect detection performance. In this article, we introduce a personalized Federated anomaly detection framework for discrete event Sequences, named FeadSeq. Specifically, we propose a separate architecture for sequence reconstruction networks (SEPRE) which partitions the network into two parts: a shared part and a standalone part, better suited for federated learning schemes. In tandem, we propose a novel partial shared federated learning scheme that employs a mask strategy to alleviate communication overhead and produce personalized local models to address the statistical heterogeneity of data among clients. This scheme dictates that a subset of weights is communicated between clients and servers for collaborative training, while the remaining weights are trained exclusively locally. To evaluate the effectiveness of FeadSeq, we conduct extensive experiments on both system logs and business process event logs. The results affirm the superiority of FeadSeq over existing personalized federated learning algorithms, showcasing not only improved performance but also reduced communication overhead.

Multivariate time series forecasting is applied in many domains, such as finance, transportation, and industry. The main challenge of precise forecasting lies in accurately capturing latent dependencies. Recent studies develop various frameworks to reduce computational complexity or to enhance the learning of intricate relationships, while lacking interpretability and generality. In this article, we aim to elucidate the capture of dependencies as the recognition of patterns. We believe that patterns can be formally described from two aspects: the shapes of segments that frequently repeat and the corresponding forms of repetitions. Drawing upon this idea, we design a multivariate time series forecasting model named PRformer , 1 which incorporates a pattern-oriented attention mechanism and a pattern-based projector. The attention mechanism can perceive different forms of repetitions by embedded with various similarity evaluation metrics between segments, and filter out noise from segments to extract potential patterns with a statistical-driven weighting scheme. The pattern-based projector is employed to form the forecasting results by deriving the representative patterns from the set of potential ones. By incorporating explicit definitions of patterns, PRformer is interpretable and general to various time series scenarios. Experimental results on seven datasets demonstrate that PRformer outperforms six state-of-the-art models by about 10.7% in forecasting accuracy.

Data imputation is crucial in the analysis of incomplete time series, such as forecasting and classification, which involves learning dependencies among the observed values to infer missing ones. As there are no ground truths for missing values, the challenge of time series imputation lies in preventing the model from overfitting to spurious correlations. In this article, we believe that ensuring dependency consistency between observed and imputed values in a sequence is paramount for data imputation. Based on this idea, we propose a model called IR 2 -Net , 1 which combines an incomplete representation mechanism (IRM) with an iterative reconstruction framework (IRF) to establish a closed-loop learning-validation imputation paradigm. Firstly, IRM facilitates the representation of dependencies in incomplete sequences while preserving their distributions and semantics, effectively preventing the model from capturing spurious correlations. Secondly, IRF enables the model to reconstruct identical complete sequences separately based on imputed and observed values, ensuring that the dependencies of imputed values remain consistent with those of the observed ones. We conduct experiments on four datasets and compare IR 2 -Net with seven state-of-the-art imputation models. The experiment results show that IR 2 -Net outperforms all the baselines by 4.1%–23.4% in terms of accuracy. Moreover, IRF and IRM are two general modules that can be easily integrated into two existing models, significantly enhancing their performance by 18.3%–42.0%.

Effective management of business processes is crucial for organizational success. However, despite meticulous design and implementation, anomalies are inevitable and can result in inefficiencies, delays, or even significant financial losses. Numerous methods for detecting anomalies in business processes have been proposed recently. However, there is no comprehensive benchmark to evaluate these methods. Consequently, the relative merits of each method remain unclear due to differences in their experimental setup, choice of datasets and evaluation measures. In this paper, we present a systematic literature review and taxonomy of business process anomaly detection methods. Additionally, we select at least one method from each category, resulting in 16 methods that are cross-benchmarked against 32 synthetic logs and 19 real-life logs from different industry domains. Our analysis provides insights into the strengths and weaknesses of different anomaly detection methods. Ultimately, our findings can help researchers and practitioners in the field of process mining make informed decisions when selecting and applying anomaly detection methods to real-life business scenarios. Finally, some future directions are discussed in order to promote the evolution of business process anomaly detection.

The ability to detect anomalies in business processes is crucial for achieving success in business operations. While unsupervised anomaly detection approaches have gained popularity in recent years due to their label-free nature, in some cases, a limited number of labelled anomalies can be provided and using them can improve the performance of anomaly detection. To address this issue, we propose a novel framework for anomaly detection that uses a pre-trained autoencoder to extract feature representations of traces. An anomaly score generator based on a multi-layer perceptron is utilized to evaluate the extracted features. The entire framework is trained using a joint loss that ensures the generated anomaly scores satisfy a specific distribution without compromising the autoencoder's ability to reconstruct normal traces. The feature encoder is fine-tuned to provide insights into the cause of anomalies. Additionally, we design a novel technique for calculating anomaly scores to mitigate the effects of varying numbers of potential attribute values. We conduct extensive experiments on both synthetic and real-life logs, and our results demonstrate that our proposed method, WAKE, outperforms state-of-the-art unsupervised deep business process anomaly detection methods by a significant margin. Additionally, it outperforms other weakly supervised anomaly detection methods as well

Accurately forecasting workloads in terms of throughput that is quantified as queries per second (QPS) is essential for microservices to elastically adjust their resource allocations. However, long-term QPS prediction is challenging in two aspects: 1) generality across various services with different temporal patterns, 2) characterization of intricate QPS sequences which are entangled by multiple components. In this paper, we propose a knowledge auto-embedding Informer network (KAE-Informer) for forecasting the long-term QPS sequences of microservices. By analyzing a large number of microservice traces, we discover that there are two main decomposable and predictable components in QPS sequences, namely global trend & dominant periodicity (TP) and low-frequency residual patterns with long-range dependencies. These two components are important for accurately forecasting long-term QPS. First, KAE-Informer embeds the knowledge of TP components through mathematical modeling. Second, KAE-Informer designs a convolution ProbSparse self-attention mechanism and a multi-layer event discrimination scheme to extract and embed the knowledge of local context awareness and event regression effect implied in residual components, respectively. We conduct experiments based on three real datasets including a QPS dataset collected from 40 microservices. The experiment results show that KAE-Informer achieves a reduction of MAPE, MAE and RMSE by about 16.6%, 17.6% and 23.1% respectively, compared to the state-of-the-art models.

A session-based recommendation system (SRS) tries to predict the next possible choice of anonymous users. In recent years, graph neural network (GNN) models have been successfully applied to SRSs and have achieved great success. Using GNN models in SRSs, each session graph is processed successively to obtain the embedding of the node (i.e, each action on an item), which is then imported into the prediction module to generate recommendation results. However, solely depending on the session graph to obtain the node embeddings is not sufficient because each session only involves a few items. Therefore, neighbor sessions have been used to extend the session graph to learn more informative node representations. In this paper, we introduce a S ession-based recommendation MO del based on N eighbor sessions with similar probabilistic int E ntions(SMONE). SMONE models the intentions behind sessions in a probabilistic way and retrieves the neighbor sessions with similar intentions. After the neighbor sessions are found, the target session and its neighbor sessions are modeled as a hypyergraph to learn the contextualized embeddings, which are combined with item embeddings through GNN to produce the final item recommendations. Experiments on real-world datasets prove the effectiveness and superiority of SMONE.

Multi-task learning (MTL) aims at solving multiple related tasks simultaneously and has experienced rapid growth in recent years. However, MTL models often suffer from performance degeneration with negative transfer due to learning several tasks simultaneously. Some related work attributed the source of the problem is the conflicting gradients. In this case, it is needed to select useful gradient updates for all tasks carefully. To this end, we propose a novel optimization approach for MTL, named GDOD, which manipulates gradients of each task using an orthogonal basis decomposed from the span of all task gradients. GDOD decomposes gradients into task-shared and task-conflict components explicitly and adopts a general update rule for avoiding interference across all task gradients. This allows guiding the update directions depending on the task-shared components. Moreover, we prove the convergence of GDOD theoretically under both convex and non-convex assumptions. Experiment results on several multi-task datasets not only demonstrate the significant improvement of GDOD performed to existing MTL models but also prove that our algorithm outperforms state-of-the-art optimization methods in terms of AUC and Logloss metrics.

Multi-task learning (MTL) has experienced rapid growth in recent years. A typical way of conducting MTL with deep neural networks (DNNs) is either establishing a sort of global feature sharing mechanism across all tasks or assigning each task an individual set of parameters with cross-connections. However, these existing approaches leverage DNNs only to share features of a certain order. Several modelsdemonstrated that explicitly modeling feature sharing with both low-order and high-order features can boost performance. To this end, we propose a model-agnostic sparse routing architecture called MASR, which emphasizes arbitrary order feature sharing for multi-task learning. It is able to choose specific orders of features to route for a given task through learnable latent variables. Moreover, MASR is model-agnostic and can be combined with existing MTL models to share features of both low-order and high-order. Extensive experimental results on several real-world datasets not only confirm the significant improvement of MASR performed to existing MTL models but also outperform existing hybrid architectures in terms of AUC metric.

With the development of cloud computing and the advent of the Web 2.0 era, composing a set of Web services as a service process is becoming a common practice to provide more functional services. However, a service process involves multiple service invocations over the network, which incurs a huge time cost and could become a bottleneck to performance. To accelerate its execution, we propose an engine-side cache-block-based service process caching strategy (CBPCS). It is based on, and derives its advantages from, three key ideas. First, the invocation of Web services embodies semantics, which enables the application of semantic-based caching. Second, cache blocks are identified from a service process, and each block is equipped with a separate cache so that the time overhead of service invocation and caching can be minimized. Third, a replacement strategy is introduced taking into account time and space factors to manage the space allocation for a process with multiple caches. The algorithms and methods used in CBPCS are introduced in detail. Moreover, how CBPCS can be applied to multiple service process models is also investigated. Finally, CBPCS is validated via comparison experiments, which shows the considerable improvements of CBPCS over other strategies.

To improve the performance of recommender systems in a practical manner, several hybrid approaches have been developed by considering item ratings and content information simultaneously. However, most of these hybrid approaches make recommendations based on aggregating different recommendation techniques using various strategies, rather than considering joint modeling of the item’s ratings and content, and thus fail to detect many latent factors that could potentially improve the performance of the recommender systems. For this reason, these approaches continue to suffer from data sparsity and do not work well for recommending items to individual users. A few studies try to describe a user’s preference by detecting items’ latent features from content-description texts as compensation for the sparse ratings. Unfortunately, most of these methods are still generally unable to accomplish recommendation tasks well for two reasons: (1) they learn latent factors from text descriptions or user--item ratings independently, rather than combining them together; and (2) influences of latent factors hidden in texts and ratings are not fully explored. In this study, we propose a probabilistic approach that we denote as latent random walk (LRW) based on the combination of an integrated latent topic model and random walk (RW) with the restart method, which can be used to rank items according to expected user preferences by detecting both their explicit and implicit correlative information, in order to recommend top-ranked items to potentially interested users. As presented in this article, the goal of this work is to comprehensively discover latent factors hidden in items’ ratings and content in order to alleviate the data sparsity problem and to improve the performance of recommender systems. The proposed topic model provides a generative probabilistic framework that discovers users’ implicit preferences and items’ latent features simultaneously by exploiting both ratings and item content information. On the basis of this probabilistic framework, RW can predict a user’s preference for unrated items by discovering global latent relations. In order to show the efficiency of the proposed approach, we test LRW and other state-of-the-art methods on three real-world datasets, namely, CAMRa2011, Yahoo!, and APP. The experiments indicate that our approach outperforms all comparative methods and, in addition, that it is less sensitive to the data sparsity problem, thus demonstrating the robustness of LRW for recommendation tasks.

Recommending routes for a group of competing taxi drivers is almost untouched in most route recommender systems. For this kind of problem, recommendation fairness and driving efficiency are two fundamental aspects. In the paper, we propose SCRAM, a sharing considered route assignment mechanism for fair taxi route recommendations. SCRAM aims to provide recommendation fairness for a group of competing taxi drivers, without sacrificing driving efficiency. By designing a concise route assignment mechanism, SCRAM achieves better recommendation fairness for competing taxis. By considering the sharing of road sections to avoid unnecessary competition, SCRAM is more efficient in terms of driving cost per customer (DCC). We test SCRAM based on a large number of historical taxi trajectories and validate the recommendation fairness and driving efficiency of SCRAM with extensive evaluations. Experimental results show that SCRAM achieves better recommendation fairness and higher driving efficiency than three compared approaches.