In recent years, there has been a burgeoning interest in multimodal recommender systems within the recommendation systems domain. These systems aim to understand user preferences by leveraging both user interaction data and multimodal information associated with items. This approach frequently results in superior recommendation accuracy compared to traditional models that rely solely on user-item interactions. Despite the advancements of these methods, there is a relatively low utilization of image features in propagating item-item characteristics, an overreliance on text feature similarity, and a frequent neglect of the deep relationships between items, users, and modalities. In response to these challenges, we introduce a novel model termed LLMs-Enhanced Hyper-Knowledge Graph Recommender for Multimodal Recommendation (DOGE). DOGE utilizes large language models (LLMs) to understand image information under the guidance of text information, generating cross-modal features that effectively enhance the relationship between text and image modalities. Subsequently, DOGE constructs a Hyper-Knowledge Graph (HKG) using user-item interaction information and modality features enhanced by large language models. This graph encompasses a wide range of item-item and user-user binary relations and hyper-relations, effectively expanding the feature propagation mechanisms and mitigating the overreliance on text modality. By learning on heterogeneous user-item graphs and homogeneous item-item, user-user graphs, DOGE enhances potential effective propagation between item features and user features, acquiring more effective feature representations of users and items. Comprehensive experimentation across three public real-world datasets illustrates that DOGE attains state-of-the-art (SOTA) performance, exhibiting a 7.2% improvement over the strongest baseline.

Multi-criteria decision making (MCDM) and preference learning (PL) are crucial subfields of intelligent decision-making, both aiming to aid decision-makers (DMs) in selecting, classifying, or ranking alternatives. While MCDM and PL can complement each other to some extent, existing approaches combining MCDM and PL often struggle with large data volumes and complex relational information. To address this, we propose a novel approach called ID-GMLM that integrates graph models and large language models (LLMs) for intelligent decision-making. It reformulates decision-making as a high-parallelism ranking function in the graph domain, using graph neural networks (GNNs) to learn and understand complex relationships between alternatives or criteria, and LLMs to parse and quantify the preferences of DMs. ID-GMLM features a multi-task learning framework that optimizes the primary task of predicting alternative rankings while modeling criterion interactions through the auxiliary task. Additionally, ID-GMLM incorporates a parameter tuning network based on criterion weights and an attention network, allowing the model to adaptively adjust to the context of the current task and the evolving preferences of DMs. Experiments on benchmark datasets demonstrate that ID-GMLM achieves significant performance improvements, inheriting the interpretability and intuitive appeal of MCDM while leveraging the computational efficiency and high accuracy of PL.

As sixth-generation (6G) networks extend to remote regions, efficient data collection from Internet of Things (IoT) devices faces challenges such as limited infrastructure, diverse latency requirements, and energy constraints. To address these, we propose a dual-mode data transmission strategy: (1) a UAV-satellite network for real-time delay-sensitive data, where wireless power transfer (WPT) is introduced to alleviate energy depletion caused by frequent data transmission, and a deadline τmaxis set in each timeslot to maintain data freshness. (2) a carry-store mode for delay-tolerant data. Both aim to minimize energy consumption while maximizing data collection. Specifically, we develop a time scheduling policy to allocate data and energy transmission time for delay-sensitive data within τmax, and propose a novel centralized control, distributed execution framework using decentralized deep reinforcement learning (DRL), termed DRL-Schedule, designed to optimize UAV navigation and ensure energy-efficient, timely data collection. Simulation results demonstrate that DRL-Schedule achieves an average improvement of approximately 11.2% and 12.8% over the best baseline PPO in energy efficiency, when varying different number of UAVs and IoTs, respectively, and notably surpassing GA, Greedy and Random.

Unmanned Aerial Vehicles (UAVs) possess advantages such as high-quality channel capabilities and high mobility, making them widely used in assisting wireless sensor networks (WSNs) for environmental monitoring, smart medical care, etc. However, wireless sensor networks face energy constraints. In order to address this challenge, a novel one-hop task scheduling in wireless sensor networks is proposed to maximize the data value based on one-hop task scheduling. Then we formulated it as a Constrained Markov Decision Process, aiming to maximize the data value of all the cluster heads under multi-dimensional constraints; Furthermore, a robust algorithm V-DQN (Value-deep Q-learning) based on deep reinforcement learning (DRL) is proposed to help the Q-value table out with the curse of dimensionality, and enhance data collection efficiency. Simulation results demonstrate that our work surpasses MCTS-based by 4.2%, and notably outperforms DTS-UAV, Greedy, Random about 30.12%, 62.15%, 82.28%, respectively. Furthermore, it surpasses BPSO by approximately 38. 64% compared to its scenario, demonstrating a data collection efficiency of approximately 93.8%.

Multimodal recommender systems enhance recommendation performance by integrating information from different modalities (e.g., text and images). A common approach is to link items with high modality similarity in modality graphs, helping users explore their interests more broadly. However, existing methods often introduce noise when enhancing modality graphs, making it challenging to effectively balance performance and accuracy. To address this issue, we propose an Interest Tree Augmented Modality Graph RecommendER for Multimodal Recommendation (TAMER). In this framework, we first redistribute item modality features using various component analysis methods to ensure more reliable item similarity within modality graphs. Next, we construct interest graphs based on reliable semantic relationships and prune the interest graphs into multiple interest trees. These interest trees are then applied to the multimodal item-item homogeneous graph to extend potential links within the modality homogeneous graph. The interest tree-based enhancement method effectively captures high-order relationships in the modality graph while avoiding noisy links. The effectiveness of the proposed method is demonstrated through comprehensive experiments on three real-world datasets. Compared with the strongest baseline methods, our method achieves an average improvement of 9.98% across four evaluation metrics. The source code is available at https://github.com/Z-last-ONE/TAMER.

In the era of digital personalization, multi-criteria recommender systems (MCRSs) play a vital role in capturing the multi-dimensional nature of user preferences by considering multiple evaluation criteria rather than relying on a single overall rating. However, existing approaches to MCRSs face challenges in managing graph sparsity, criterion independence, and leveraging semantic information for recommendation tasks. To address these limitations, we propose a novel framework named Large L anguage Models-augmented Light G raph Model for M ulti- c riteria Rec ommendation ( LGMcRec ). LGMcRec integrates the strengths of graph neural networks (GNNs) and large language models (LLMs) to improve the representation and recommendation capabilities of MCRSs. In our model, we construct a tripartite graph structure that captures user-item interactions, item-criterion associations, and criterion interdependencies, effectively addressing issues of sparsity and unmodeled correlations in multi-criteria data. We extend the LightGCN architecture to learn embeddings over this graph, which are further enriched through semantic alignment with embeddings generated by LLMs from textual user and item profiles. To bridge the gap between graph-based and LLM-based embeddings, we employ a contrastive learning approach that maximizes the mutual information between the two embedding spaces, ensuring cohesive and comprehensive user and item representations. Experimental results on three MCRS datasets demonstrate that LGMcRec achieves significant performance improvements over state-of-the-art methods.

In the context of natural language understanding for short texts, pipeline and joint learning models based on deep learning are two common approaches to address the challenges of intent recognition and slot filling. Due to the potential error propagation behavior of the pipeline model, its final effectiveness is often compromised by deviations in the initial stage. Although there have been some endeavors and contributions in the domain of intent recognition within joint learning, most methods do not explicitly focus on establishing the relationship between intent and slot. We propose a novel joint recognition model for intent and slot which incorporates an enhanced self-attention mechanism and a weight gating unit channel. This channel effectively extracts the correlation between intent and slot during the decoding process. Our model is evaluated on public datasets like ATIS and SNIPS, demonstrating an approximate 1.0% improvement in accuracy compared to the current mainstream models that utilize attention-based methods.

Image enhancement methods leveraging learning-based approaches have demonstrated impressive results when trained on synthetic degraded-clear image pairs. However, when deployed in real-world scenarios, such models often suffer significant performance degradation due to the inherent domain gap between synthetic and real degradations. To bridge this gap, we propose a novel Two-stage Contrastive Domain Adaptation image Enhancement (TCDAE) framework consisting of two key strategies: (1) Synthetic-to-Real Domain Transfer Learning (S2R-DTL) that effectively translates images from the synthetic degraded domain to the real degraded domain, aligning the domains at the pixel level, and (2) Degraded-to-Clear Domain Transfer Learning (D2C-DTL) that further adapts the enhancement model from the synthetic to the real domain by translating images from the real degraded domain to the real clean domain in both supervised and unsupervised branches. A unique aspect of our approach is the integration of a Domain Noise Contrastive Estimation (DoNCE) loss in both learning strategies. This specialized loss formulation enables TCDAE to robustly translate images across domains, even in scenarios lacking strong positive examples. Consequently, our framework can generate enhanced images with natural, realistic appearances akin to real clear images. Comprehensive experiments on real-world degraded scenes across diverse tasks, including dehazing, deraining, and deblurring, demonstrate the superiority of TCDAE over state-of-the-art methods, achieving improved visual quality, quantitative metrics, and downstream task performance.

Preference learning refers to learning the preferences for a collection of alternatives based on observed or revealed preference information, which are usually represented in the form of an order relation. Some of the existing preference learning methods are parametric in nature, and such methods are faster to train but less expressive. To address this issue, we map the preference learning onto the graph structure, introduce the concept of graph neural networks (GNNs), and propose a two-stage preference learning method based on GNNs, which consists of preference relation prediction stage and object preference ranking stage. The first stage is used to judge the preferences between pairs of objects, the second stage is used to correct the inconsistent preference information of the first stage and rank all objects. In Stage 1, we turn the prediction problem into an edge classification problem on the graph, design a multilayer perceptron (MLP) model to extract edge features, and mine preference information with the help of GNNs. In Stage 2, we construct a comparator neural network structure that takes pairwise preference information as input and generates a score for each object as output. The ranking of the object scores determines the objects’ preference order. Experiments conducted on preference learning datasets have shown that our method achieves significant performance improvements over existing preference learning methods when evaluated in the context of preference service.

Multi-person behavior recognition is an important task in intelligent video surveillance.In this paper, we propose a multi-person behavior recognition method based on multi descriptors aggregations (MBR-MDA) for real-time surveillance scenarios.Our method employs multi-object tracking to obtain consecutive frames of each person, and uses a 2D convolutional network with temporal shift module (TSM) for behavior recognition.To address the limitation of 2D convolutional network in capturing global temporal features, we introduce a plugand-play module called MDA that can be integrated into the 2D convolutional network.By applying data augmentation and embedding the MDA3D module, our method achieves a 4.8% improvement over TSM baseline on the HMDB51 dataset, with only a minimal speed loss of 0.3ms.We evaluate our method on several public datasets and demonstrate that embedding MDA into other methods can also enhance their performance.

Long-tailed data distribution is a common phenomenon observed in nature. However, this distribution poses challenges for neural network training, as it performs well on samples with head categories, but not so well on tail categories - precisely what requires attention in certain scenes. Although existing methods involve re-sampling data or designing loss functions, they often fail to adapt well to changing data distributions. In this article, we propose a long-tailed detection approach based on multi-expert aggregation (MEA). This method employs multi-expert networks to learn about head, medium, and tail categories from a fixed long-tailed distribution. Different data are modelled using different loss designs, and the results of multi-expert networks are aggregated to achieve good accuracy for both head and tail categories. The effectiveness of this method has been verified on LVIS.

Smart grid system can obtain users' daily load data, and by clustering, we can get users' load profiles to divide them into industrial, commercial and residential types. Load data has the characteristic of changing periodically. Within a period, the load profiles are relatively stable. However, load profiles often changes significantly according to reasons like holidays and season changing. When conducting a continuous clustering task for consecutive days, traditional clustering algorithms cannot consider the time-dimension features into analysis, which may make clustering results be very different even if user behaviors are almost the same. Evolutionary clustering (EC) can be taken into consideration. EC doesn't ignore historical clustering results and makes results more stable in a period. However, when dramatic changes happen in user behaviors, the quality of EC's clustering results will decrease significantly. This paper proposed an optimized evolutionary clustering algorithm: Timestamp-Based Self-Adaptive Evolutionary Clustering (TBSAEC). TBSAEC is based on evolutionary clustering, and takes a heuristic approach to pick the evolutionary parameter to maximize the total quality. TBSAEC maintains the stability of continuous-time clustering results while better adapting to changes in user behaviors. Besides, TBSAEC optimize the running efficiency of the algorithm by picking samples in equal portions from historical data instead of the whole data. We applied TBSAEC to the load data of a certain region in east China in 2015, and the results showed that TBSAEC is 3% to 9% higher than the ordinary evolutionary clustering algorithm in total quality, and 87% faster in running time.

Mobile edge computing (MEC) in 5G networks has recently emerged as a promising paradigm to enhance the data processing capabilities of mobile devices. Due to portability and cost considerations, mobile devices usually have limited battery and computing resources. Through MEC technology, computing tasks can be offloaded to remote servers, which helps to reduce computing latency and energy consumption. However, an im-proper computing offloading may produce additional overheads such as waiting time and wireless transmission latency, because of the limited resources of remote servers and extra wireless transmissions. In this paper, we propose a multi -agent deep reinforcement learning (MADRL) based decentralized cooperative offloading decision algorithm, which determines whether the computing tasks are executed locally or placed on an appropriate edge node to minimize system costs. Our goal is to learn an optimal online policy from experiences to solve a combinatorial optimization problem at a lower computational complexity, and the factors such as computation delay, energy consumption, communication latency, and waiting time in remote servers are all considered. Experiment results show that our approach is able to reduce by an average of 5.4% execution latency than traditional methods, and outperforms single-agent D RL algorithms.

Given the problem that the relations among alternatives or criteria cannot be handled well in multicriteria decision making, this paper applies the concept of Pythagorean fuzzy sets to the graph and develops a decision-making approach based on Pythagorean fuzzy graphs (PFGs). First, the weights obtained from the Laplacian energy of PFGs are taken as the subjective weights of criteria. Then, a new Pythagorean fuzzy entropy measure is defined to compute the objective weights of criteria. Meanwhile, a combined weighting method is presented, which makes the criteria weights consist of subjective weights and objective weights. Furthermore, combined weights are applied to the decision-making process and a graph-based Pythagorean fuzzy decision-making method is proposed. Compared with other existing techniques, the proposed approach considers the relations between alternatives and the relations between criteria simultaneously. Finally, an illustrative example is used to verify the approach and demonstrate its effectiveness. The results show that the alternative ranking obtained by the proposed approach is reliable and credible.

This article presents the solution that we use in the Global Road Damage Detection Challenge 2020, which is designed to recognize the road damages present in an image captured from three countries: India, Japan, and Czech. In this challenge, Cascade R-CNN is selected as a baseline model to detect objects in images. It is commonly known that making a precise annotation in a large dataset is crucial to the performance of object detection and placing bounding boxes for every object in each image is time-consuming and costs a lot. To make full use of available unlabeled data, the consistency filtering mechanism (CFM) with self-supervised methods is proposed to utilize high-confident samples with pseudo-labels for training. And we also apply a series of data augmentation techniques (road segmentation, flip, mixup, CLAHE) to labeled data in training phase. Moreover, we ensemble models with different tricks by weighted boxes fusion to produce the final prediction. Finally, our proposed method can achieve a great mean f1-score of 0.6290 on the test1 dataset and 0.6219 on the test2 dataset respectively, which wins the Bronze Prize (ranks 3rd place). Code and trained models are available at the following link: https://pan.baidu.com/s/1VjLuNBVJGS34mMMpDkDRGQ, password: xzc6.

The criminal behavior of telecom fraud is increasing rapidly with the development of the communication industry, causing huge losses every year. The commonly used traditional fraud detection methods are less flexible. Currently, a more accurate and timely method is needed to deal with the evolving fraud. However, with the development of deep learning, more and more complex structures and a large number of parameters lead to the more time-consuming training process and poor interpretability which are unacceptable in the field of anti-fraud. Therefore, this paper proposes a shallow model called Encoding Broad Learning by incorporating the denoising autoencoder into the Broad Learning System (BLS). We use the first 15 seconds of the call content to process the text data of the fraudulent call by constructing TF-IDF, adding Gaussian noise to it, and combining with the denoising autoencoder to learn more general and robust features in the data. Then it is transformed into a neural network based on BLS, and fraudulent calls are identified on this model. This method is extremely suitable for fraud identification systems with few data features and high real-time training requirements. At the same time, in order to further improve the training efficiency and solve the potential memory explosion problem, we propose an integrated learning algorithm for parallel training of EBLS. Experiments and detailed analysis of the above methods are carried out. The results show that compared with the existing classic network algorithms, this method has a faster training speed, can ensure the accuracy and timeliness of online fraud identification, and help quickly identify fraudulent calls. The paper visualized the training results of these algorithms after processing the relevant parameters of EBLS training. These algorithms have strong interpretability and better meet the high security requirements in the field of anti-fraud.

Mining quantitative association rules is one of the most important tasks in data mining and exists in many real-world problems. Many researches have proved that particle swarm optimization(PSO) algorithm is suitable for quantitative association rule mining (ARM) and there are many successful cases in different fields. However, the method becomes inefficient even unavailable on huge datasets. This paper proposes a parallel PSO for quantitative association rule mining(PPQAR). The parallel algorithm designs two methods, particle-oriented and data-oriented parallelization, to fit different application scenarios. Experiments were conducted to evaluate these two methods. Results show that particle-oriented parallelization has a higher speedup, and data-oriented method is more general on large datasets.

Electricity consumption of diverse facilities can be recorded hourly or minutely due to the development of smart grid and smart home technologies. As a result, the traditional relationship analysis between electricity consumption and other external factors should be improved and conducted based on fine-grained rather than coarse-grained time series data. In that case, canonical correlation analysis (CCA) is an appropriate method to process two or more datasets containing multiple variables. However, the result of CCA is not unique, which leads to the challenge for batch-oriented data analysis. To solve this problem, we propose an optimal result selection mechanism for CCA and kernel CCA algorithms based on accuracy validation of canonical weights and components. An additional clustering is also provided to optimize the approach in terms of time complexity and accuracy performance. The approach is implemented on three multivariable datasets, referring to 960 non-residential electricity consumers in 60 towns or cities of the same district, to find the canonical correlations among electricity consumption, gas consumption and climate change for every consumer. The experimental results indicate that the proposed approach outperforms other related methods. We also find out three typical patterns of canonical correlation curves, which are relative stability, cyclic change and seasonal change.

Speech classification methods mainly focus on the content of the voice segment. To help better underestand the information in a segmented voice, the contents of other segments in the same paragraph should also be paid attention to. In our custom service speech classification problem, we are facing a problem of classification a series of voice segments in a conversation separately into category "custom" or "custom service". Sometimes the voice of both parties in the same conversation can be both sound like a "custom service" or both sound like "custom". In order to make the right prediction, the model needs to know not only the content of the voice segment that it's classifying, but both parties' voice in a conversation, the extra information can help the model to determine who is "more likely" to be a custom service in a conversation. We propose a method called I-CNN, which combines the info-feed layer with CNN. The Info-feed layer allows the CNN to use information from other samples in the same batch, which is helpful in improving the model's performance in our custom service speech classification problem.

At present, the crime of telecom fraud, with advanced communications and Internet technologies, is growing rapidly and causing huge losses every year. The traditional fraud detection methods are less flexible. In this paper, we used the signaling data to train a clustering model, which can discover the hidden user characteristics of fraud phones. The paper puts forward the extraction method of behavior characteristics, reduce the dimension of features with principal component analysis and select the appropriate clustering parameters through grid search, then present the K-Means-based behavior identification system, which can help to distinguish the frauds and identify the fraud phone numbers. Finally, the feasibility of this model is verified by the actual sample dataset.

With the continuous development of communications industry, the majority of users gradually enjoy a variety of communications services. In the meantime, however, more and more fraud phones appears in the user's daily life. Although there are currently many interception schemes for fraud phones, they all belong to passive interception and rely on intelligent terminals. Therefore, this paper proposes a fraud phone calls analysis method based on label propagation community detection algorithm (LPA). Call content data are transformed into complex network. The LPA algorithm is used to generate fraud communities on this complex network. Detail analysis is also carried out for extracting the detail of communities. Results show that the methods can help to quickly identity the fraud phone calls.

The electricity load data recorded by smart meters contain plenty of knowledge that contributes to obtaining load patterns and consumer categories. Generally, the daily load curves are clustered first in order to obtain load patterns of each consumer. However, due to the volume and high dimensions of load curves, existing clustering algorithms are not appropriate in this situation. Thus, a fused load curve clustering algorithm based on wavelet transform (FCCWT) is proposed to solve this problem. The algorithm includes two main phases. First, FCCWT applies multilevel discrete wavelet transform (DWT) to convert the daily load curves for dimensionality reduction. Second, it detects clusters at two outputs of the first phase, and then fuses two groups of clusters with a sub-algorithm named cluster fusion to achieve the optimized clusters. FCCWT is implemented on datasets of both China and United States. Their clustering performances are evaluated by diverse validity indices comparing with four typical clustering methods. The experimental results show that FCCWT outperforms other comparison methods. Additionally, case analysis of two datasets are also provided to discuss the significance of load patterns.

PaaS is known as an application engine which third party developers can deploy their application onto. Security of PaaS becomes important as applications shares resources. How to secure and isolation the resources become an important topic. In this paper, a security PaaS container is proposed which is based on a customized JVM. This container is fully implemented and evaluated in real setting.

In recent years, RESTful Web services have been rapidly developed and deployed, because of the advantages of lightweight, flexibility and extensibility, etc. However, most RESTful services are described in heterogeneous and ordinary HTML pages, which makes them really difficult to be identified and crawled automatically from the Internet. In this paper we propose a hybrid classifier framework called co-NV for automatic identification of RESTful services on the Web. In our framework, web pages are analyzed and filtered according to the contents and structure characteristics of HTML documents, with Naïve Bayes classifier and Vector Space Model (VSM) respectively. Experiments with real RESTful services prove that our framework works effectively with high precision and recall rate, and is very practical.

As cloud computing goes, How to provide a security cloud becomes an important problem. Virtual networking plays an important role in cloud computing infrastructure. To identify which node is a critical node become an important research question. In this study, our team analyzed the fat tree network and small-world network, and proposed a network modeling method for the virtual networking. On this basis, we analyze time performance and detecting accuracy of the two critical nodes detecting algorithms. One is based on depth-first search, while the other concentricity analysis.

Recently it receives extensive concerns that mining tags from WSDL documents, since service tags are widely used for better utilization of web services. However, most WSDL documents lack contents, which lead to great limitation for service tag mining. In this paper, we propose a novel approach to retrieve and extract service tags by leveraging information from the Internet. Our approach retrieves service related web pages by searching online for the URL of the services, then executes a Vector Space Model (VSM) based procedure to fulfill structure analysis and content extraction of the pages, finally mines service related tags together with WSDL documents. Experiments on real web services demonstrate the effectiveness of our approach, which is able to extract more number of abundant and meaningful tags.

Cloud computing provide computing resource on demand which helps people to make full use of legacy computing asset. Resource provision is transparent to user but not to provider. Computing provider need to aware of resource utilization. How to monitoring resource utilization become an important problem in cloud computing. In this study, we propose a monitoring architecture and a monitoring model. An Auto window filter algorithm is introduced to help reduce the network traffic.

Hadoop, the open-source implementation of MapReduce, has been widely used in different projects. However, when users want to use this parallel computing framework, they have to spend time on the Hadoop cluster configuration, learning the programming API, and the MapReduce job operations. This paper proposes the Parallel Computing Framework as a Cloud Service (PCFCS) to provide the users parallel computing cluster, and simplify the configuration, programming, uploading, and operating procedures. Especially, PCFCS defines a set of annotations, with which users can quickly build their own MapReduce job.

Web2.0 and Cloud Computing are two hot topics in current internet research. PaaS (Platform as a Service) hosts the users' service in an elasticity way and provides the load balance support, but there is not optimizing in load balance for web2.0 application, especially for SNS like website. Load balance optimizing can be divided in two aspects: web cache optimizing and load balance strategy designing. We propose a PUR-SNS (prediction on user requests for social networking services) algorithm for web cache optimizing, which organize web cache based on the social relations. Simulation shows that PUR-SNS can improve caching hit ratio in SNS like environment.

As it is known, service is the key element of the new generation telecommunication network. Though network technology develops so quickly, the technology that uses in the application remains in a low degree. In other hand, current service model is also a problem that prevent telecom application from developing.Meanwhile, people demand for more personal experiences, such as context aware demand. In this paper, the author will analyze current problems in service platform and then propose a new platform with social factors and ubiquitous feature. And an example application will set up to validate the framework.

BACKGROUND: Recently, microarray data analyses using functional pathway information, e.g., gene set enrichment analysis (GSEA) and significance analysis of function and expression (SAFE), have gained recognition as a way to identify biological pathways/processes associated with a phenotypic endpoint. In these analyses, a local statistic is used to assess the association between the expression level of a gene and the value of a phenotypic endpoint. Then these gene-specific local statistics are combined to evaluate association for pre-selected sets of genes. Commonly used local statistics include t-statistics for binary phenotypes and correlation coefficients that assume a linear or monotone relationship between a continuous phenotype and gene expression level. Methods applicable to continuous non-monotone relationships are needed. Furthermore, for multiple experimental categories, methods that combine multiple GSEA/SAFE analyses are needed. RESULTS: For continuous or ordinal phenotypic outcome, we propose to use as the local statistic the coefficient of multiple determination (i.e., the square of multiple correlation coefficient) R2 from fitting natural cubic spline models to the phenotype-expression relationship. Next, we incorporate this association measure into the GSEA/SAFE framework to identify significant gene sets. Unsigned local statistics, signed global statistics and one-sided p-values are used to reflect our inferential interest. Furthermore, we describe a procedure for inference across multiple GSEA/SAFE analyses. We illustrate our approach using gene expression and liver injury data from liver and blood samples from rats treated with eight hepatotoxicants under multiple time and dose combinations. We set out to identify biological pathways/processes associated with liver injury as manifested by increased blood levels of alanine transaminase in common for most of the eight compounds. Potential statistical dependency resulting from the experimental design is addressed in permutation based hypothesis testing. CONCLUSION: The proposed framework captures both linear and non-linear association between gene expression level and a phenotypic endpoint and thus can be viewed as extending the current GSEA/SAFE methodology. The framework for combining results from multiple GSEA/SAFE analyses is flexible to address practical inference interests. Our methods can be applied to microarray data with continuous phenotypes with multi-level design or the meta-analysis of multiple microarray data sets.