Handling heterogeneous data in tabular datasets poses a significant challenge for deep learning models. While attention-based architectures and self-supervised learning have achieved notable success, their application to tabular data remains less effective over linear and tree based models. Although several breakthroughs have been achieved by models which transform tables into uni-modal transformations like image, language and graph, these models often underperform in the presence of feature heterogeneity. To address this gap, we introduce TabGLM (Tabular Graph Language Model), a novel multi-modal architecture designed to model both structural and semantic information from a table. TabGLM transforms each row of a table into a fully connected graph and serialized text, which are then encoded using a graph neural network (GNN) and a text encoder, respectively. By aligning these representations through a joint, multi-modal, self-supervised learning objective, TabGLM leverages complementary information from both modalities, thereby enhancing feature learning. TabGLM's flexible graph-text pipeline efficiently processes heterogeneous datasets with significantly fewer parameters over existing Deep Learning approaches. Evaluations across 25 benchmark datasets demonstrate substantial performance gains, with TabGLM achieving an average AUC-ROC improvement of up to 5.56% over State-of-the-Art (SoTA) tabular learning methods.

Data wrangling is a critical yet often labor-intensive process, essential for transforming raw data into formats suitable for downstream tasks such as machine learning or data analysis. Traditional data wrangling methods can be time-consuming, resource-intensive, and prone to errors, limiting the efficiency and effectiveness of subsequent downstream tasks. In this paper, we introduce AutoDW: an end-to-end solution for automatic data wrangling that leverages the power of Large Language Models (LLMs) to enhance automation and intelligence in data preparation. AutoDW distinguishes itself through several innovative features, including comprehensive automation that minimizes human intervention, the integration of LLMs to enable advanced data processing capabilities, and the generation of source code for the entire wrangling process, ensuring transparency and reproducibility. These advancements position AuoDW as a superior alternative to existing data wrangling tools, offering significant improvements in efficiency, accuracy, and flexibility. Through detailed performance evaluations, we demonstrate the effectiveness of AutoDW for data wrangling. We also discuss our experience and lessons learned from the industrial deployment of AutoDW, showcasing its potential to transform the landscape of automated data preparation.

The number of web APIs is growing rapidly. API adoption is increasing across all industries with executives prioritizing investments in the API economy. Each API provider offers API documentation which includes complex descriptions. In order to collect and understand the applications and operations of diverse APIs, software engineers read lengthy and complicated API documentations. Understanding the variety of API documentations is a labor intensive and error-prone process. In this paper, we introduce a data-centric web crawler service to collect, analyze, and construct a large corpus of API documentations. The generated API Corpus can be used in machine programming (i.e., code generation, code search). The proposed API web-crawler intelligently harvests more than 2.8M API documentation pages where it uses a machine-learning-based approach with an accuracy of 91.32% to select only web API pages (REST). We also conducted an extensive and end-to-end real-world evaluation, where the proposed API web-crawler not only collects a sheer number of API pages, but also successfully validates 1,222 APIs out of 1,521 target APIs with a success rate of 80.34%.

Visualization is very important for machine learning (ML) pipelines because it can show explorations of the data to inspire data scientists and show explanations of the pipeline to improve understandability. In this paper, we present a novel approach that automatically generates visualizations for ML pipelines by learning visualizations from highly-upvoted Kaggle pipelines. The solution extracts both code and dataset features from these high-quality human-written pipelines and corresponding training datasets, learns the mapping rules from code and dataset features to visualizations using association rule mining (ARM), and finally uses the learned rules to predict visualizations for unseen ML pipelines. The evaluation results show that the proposed solution is feasible and effective to generate visualizations for ML pipelines.

Vardaan Pahuja, Yu Gu, Wenhu Chen, Mehdi Bahrami, Lei Liu, Wei-Peng Chen, Yu Su. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

In recent years, IF-This-Then-That (IFTTT) services are becoming more and more popular. Many platforms such as Zapier, IFTTT.com, and Workato provide such services, which allow users to create workflows with "triggers" and "actions" by using Web Application Programming Interfaces (APIs). However, the number of IFTTT recipes in the above platforms increases much slower than the growth of Web APIs. This is because human efforts are still largely required to build and deploy IFTTT recipes in the above platforms. To address this problem, in this paper, we present an automation tool to automatically generate the IFTTT mashup infrastructure. The proposed tool provides 5 REST APIs, which can automatically generate triggers, rules, and actions in AWS, and create a workflow XML to describe an IFTTT mashup by connecting the triggers, rules, and actions. This workflow XML is automatically sent to Fujitsu RunMyProcess (RMP) to set up and execute IFTTT mashup. The proposed tool, together with its associated method and procedure, enables an end-to-end solution for automatically creating, deploying, and executing IFTTT mashups in a few seconds, which can greatly reduce the development cycle and cost for new IFTTT mashups.

The number and variety of Web APIs is growing exponentially. Software engineers need to expend a significant amount of time and effort reading and understanding the accompanying documentation. In addition, system automation may use API to interact with each other. However, this is not always a simple task since the API documentation of a provider can be anything from a single HTML page description through to a complex structure with information spanning several pages. Understanding this wide variety of API documentation structures and styles is therefore a labor intensive and error-prone task for engineers. By providing a machine-learning platform that can extract and standardize API usage information, however, we believe we can accelerate the creation of API-enabled systems by using automation to simplify the task of understanding. In this paper we introduce a novel approach to automating and standardizing usage information about APIs, combining several machine-learning algorithms in order to extract key attributes from API documentation and generate a machine readable Open API Specification (OAS). We develop i) a content-based learning model that identifies the context of a block of extracted API features; ii) a signature-based machine learning model that recognizes a sequence of successful/unsuccessful extracted API endpoints; and iii) a deep mapping model that pinpoints fine-grained mapping of extracted API attributes to OAS objects. Results of our experiments show that the proposed approach successfully works with an accuracy of 99%, 94% and 97% for content-based learning, signature-based learning, and Deep Mapping of API attributes respectively. We then use the models to produce OAS compliant API Specifications for more than 2,585 public APIs, validate them via API calls and finally deploy the validated APIs to the RunMyProcess software automation platform.

The number of web APIs grows rapidly. Each API provider offers API documentations which comes with diversity and complexity of structures. In order to understand a large number of diverse web APIs, we employ deep learning that extracts key objects of a large number of web APIs and produce a unified API specification (Open API Specification). However, the unified API specification is not a simple task due to heterogeneous API documentations and it is required additional human evaluation to adjust incorrect information which produced by machine. In this paper, we introduce WATAPI which represents an automation tool for representing machine learning outcomes and adding a user as a humanin-the-loop to transparently interact with complex machinelearning components to process diverse API documentations. WATAPI allows the user to annotate API documentations and/or adjust the automated annotation process which produced by machine-learning models' predictions. WATAPI is also capable to perform as a semi-automated annotation tool where it records user's interactions. WATAPI is able to automatically apply user's interaction of one API annotation to another API documentations with a similar structure. The user's annotation adjustment provides feedback to machinelearning components for improving the accuracy of extracting Open API Specifications.

Stack Overflow (SO) has been a great source of natural language questions and their code solutions (i.e., question-code pairs), which are critical for many tasks including code retrieval and annotation. In most existing research, question-code pairs were collected heuristically and tend to have low quality. In this paper, we investigate a new problem of systematically mining question-code pairs from Stack Overflow (in contrast to heuristically collecting them). It is formulated as predicting whether or not a code snippet is a standalone solution to a question. We propose a novel Bi-View Hierarchical Neural Network which can capture both the programming content and the textual context of a code snippet (i.e., two views) to make a prediction. On two manually annotated datasets in Python and SQL domain, our framework substantially outperforms heuristic methods with at least 15% higher F1 and accuracy. Furthermore, we present StaQC (Stack Overflow Question-Code pairs), the largest dataset to date of ~148K Python and ~120K SQL question-code pairs, automatically mined from SO using our framework. Under various case studies, we demonstrate that StaQC can greatly help develop data-hungry models for associating natural language with programming language

Heating, Ventilation and Air Conditioning (HVAC) system is largest energy consumer in buildings. Worldwide, buildings consume 20% of the total energy production. Therefore, increasing efficiency of the HVAC system will result in significant financial savings. As one solution, Thermal Energy Storage (TES) tanks are being utilized with buildings to store excess energy to be reused later. An optimal control strategy is crucial for optimal usage. Therefore, this paper presents a novel control framework based on Artificial Neural Networks (ANN) for optimally controlling a TES for achieving increased savings. The presented ANN controller utilizes 3 main inputs: 1) current TES energy availability, 2) predicted building power requirement, and 3) predicted utility load/price. In addition to the design details of the control framework, this paper presents implementation details of the ANN controller. Further, experiments on several test cases were carried out and the paper presents the experimental setup and obtained results for each test case. Performance of the presented ANN control framework was compared against a classical proportional derivative (PD) controller. It was observed that the presented framework resulted in better cost savings than the classical controller consistently for all the experimental test cases.

In this paper, we present a method for resource allocation for small cells that integrate licensed and unlicensed RF operations motivated by the widespread WiFi hotspots and the common inclusion of WiFi interface in most cellular terminals. Small cells have proven popular for cell coverage enhancement and traffic offloading from macrocells. We formulate an optimization problem that jointly allocates resources over both licensed and unlicensed bands with the goal of maximizing sum small cell user equipment (SUE) rate while achieving fairness among these user equipments and controlling inter-cell interference to neighboring macrocell users. The proposed solution further considers the quality of service (QoS) requirement of SUE traffics to be distributed over both licensed and unlicensed bands. We show the formulation of the proposed optimization problem as an efficient and low complexity linear programming. We further show that our problem formulation can be modified to maximize the revenue of mobile network operators. Our proposed solution achieves better performance than several existing solutions.

We consider a femto cellular network consisting of multiple neighboring femtocells, e.g., in an enterprise deployment such as shopping malls, stadiums, or corporate premises. We present a practical but suboptimal channel assignment and interference management algorithm for fractional frequency reuse (FFR) wireless networks. More specifically, we propose an adaptive graph coloring approach for resource allocation with the goal of interference management among femtocells as well as achieving fairness among users. While the global-optimum solution has exponential complexity, our proposed scheme has a linear complexity in the number of femtocells. Although suboptimal, we have evaluated our algorithm in small scenarios, where direct evaluation is possible, and found that the achieved minimum user rate using the proposed algorithm is 85% of the optimal minimum rate. Additionally, we have analyzed several practical design considerations of our proposal such as channel feedback, latency, and computational complexity. We demonstrate the performance of our proposed solution against various alternatives and show that it provides better performance under various environment parameters. For example, in a dense femtocell deployment, the performance was improved by 47% over a full frequency reuse scheme.

WiFi interfaces have been recently incorporated in most cellular user equipments (UEs). In current practice, the UE selects either the licensed band for cellular technologies or the unlicensed WiFi band depending on the signal quality of both bands. At the same time, small (pico or femto) cells have also become popular means to offload traffic from traditional macro-cell networks and to improve cell coverage. This work presents a method for dynamic switching and aggregation of licensed and unlicensed bands in small cells for traffic offloading and per-user throughput enhancement. Our proposed method allows small cells to jointly control transmission in both licensed and unlicensed bands in order to maximize the sum of small cell user throughputs over both bands while constraining the interference effect to maintain the Quality of Service (QoS) requirements for macrocell user equipments. Performance evaluation shows that our proposed scheme outperforms other existing solutions.

This work studies dual-access-technology femtocells equipped with both cellular and non-cellular air interfaces with the goal of increasing per-user throughput as well as achieving user fairness. We target enterprise environments characterized by high data rates and/or dense user terminals such as corporate premises, shopping malls, or stadiums. To meet this high data demand, multiple femtocells are deployed for which the major challenge is interference management. We propose an architecture for interference management and resource allocation over both cellular and non-cellular bands. Numerical evaluation shows that our proposed solution outperforms other existing solutions.

We solve the power allocation problem for distributed antenna system (DAS) with a simple algorithm running in polynomial time. Simulations in two scenarios with single and three sectors per cell show that the proposed power allocation scheme in DAS achieves higher system capacity by 50.1%, and 45.9%, respectively, in comparison with macro base station. Our proposal provides a practical and cost effective solution to address the issues of deployment cost and system capacity.

As the demand for mobile broadband services are growing, cellular networks are heading towards their fourth generation (4G). Long Term Evolution (LTE) is emerging as a major candidate for 4G cellular networks. 4G networks are to support a myriad of applications which have different error tolerance and delay requirements. MIMO (Multiple Input Multiple Output) technology combined with OFDMA and more efficient modulation/coding schemes (MCS) are key physical layer technologies in LTE networks. However, in order to fully utilize the benefits of the advances in physical layer technologies MIMO configuration and MCS need to be dynamically adjusted to derive the promised gains of 4G at the application level. This paper proposes an application-profile based link adaptation architecture that adapts the physical layer transmission parameters based not only on channel conditions but also on application-specific requirements to transfer physical layer gains to the application layer.

P2P applications are among the most popular Internet applications and constitute a majority of Internet traffic today. From the perspective of network management, a thorough understanding of the stochastic properties of P2P traffic is thus crucial to designing better traffic/resource control algorithms. However, previous work on measuring and analyzing P2P traffic focuses either on a specific type of P2P applications or report only qualitative or quantitative findings in the measured traffic, without systematically analyzing and modeling P2P traffic. In this paper, we aim to bridge the gap and perform a detailed analysis and modeling of P2P traffic of different types. We first gather and identify P2P traffic on a trans-pacific link between Japan and US. Then based on the flow level (transport layer) information, we analyze and model the marginal distributions of the traffic volume, connection duration, and connection interarrival times for P2P connections of Napster, BitTorrent (BT), eDonkey, Gnutella, and Fasttrack, respectively. We also study the burstiness property of P2P traffic generated by these five different applications and the implication/impact of connections with zero traffic volumes. A major finding of our study is that due to the different parameters/strategies used in the file sharing mechanism, P2P traffic of different applications exhibits different traffic characteristics. As a result, the traffic metrics of interest (i.e., the per-connection traffic volume, the connection duration, and the connection interarrival times) cannot be characterized uniformly with a single model among the different types of applications, but instead have to be modeled with different probability distributions.

Push-to-talk over cellular (PoC) is considered as one of important applications in next generation networks (NGN). The main objective of this study is to investigate the performance of our PoC system evaluated over three operational 3G networks in the US. Measurements are conducted in two stages. First, several generic measurements such as round trip time (RTT), UDP, TCP, and dormant mode are made to understand the fundamental characteristics of 3G packet switched (PS) networks for the voice and data transmissions of the PoC service. Next, various parameters and features in our PoC software are tested to evaluate the performance of startup latency, voice latency, and voice quality. In conclusion, through empirical study we have shown that PoC applications can achieve high quality in current 3G PS networks. Furthermore from these results we study the network requirements to be solved in NGN for the voice and data communications.

Modeling and analysis of wireless traffic is fundamental to traffic engineering and resource management. The majority of existing work in this arena has focused on modeling/analyzing aggregate traffic, and little has been done in modeling/analyzing wireless traffic at the per-user level. In this paper, we bridge the gap and perform a detailed analysis and modeling on the traffic generated by individual wireless users, leveraging the data traces collected in a period of 4 months (November 2003 - February 2004) on the Dartmouth campus-wide 802.11 WLANs.Our study indicates that several parameters that characterize the wireless traffic generated by individual users are predictable, such as the traffic volume originating from a user and destined for a specific IP address, the set of destination IP addresses (for which connections initiated by a user are destined), and the patterns by which a wireless user connects to applications. We also model the per-connection traffic volume, the duration, and the interarrival time, of connections issued by a user using Weibull or Pareto distributions, and show that the burstiness of aggregate wireless traffic (which has been reported in the literature) is a direct consequence of the burstiness of the burstiness of traffic generated by individual users. These findings can be used to optimize traffic and resource management and provide better QoS (in terms of availability, resiliency, and performance).

Distinct from wireless ad hoc networks, wireless sensor networks are data-centric, application-oriented, collaborative, and energy-constrained in nature. In this paper, formulate the problem of data transport in sensor networks as an optimization problem whose objective function is to maximize the amount of information (utility) collected at sinks (subscribers), subject to the flow, energy and channel bandwidth constraints. Also, based on a Markov model extended from [3], we derive the link delay and the node capacity in both the single and multi-hop environments, and figure them in the problem formulation. We study three special cases under the problem formulation. In particular, we consider the energy-aware flow control problem, derive an energy aware flow control solution, and investigate via ns-2 simulation its performance. The simulation results show that the proposed energy-aware flow control solution can achieve high utility and low delay without congesting the network.

We devise and evaluate a fully decentralized, light-weight, dynamic clustering algorithm for target tracking. Instead of assuming the same role for all the sensors, we envision a hierarchical sensor network that is composed of 1) a static backbone of sparsely placed high-capability sensors which assume the role of a cluster head (CH) upon triggered by certain signal events and 2) moderately to densely populated low-end sensors whose function is to provide sensor information to CHs upon request. A cluster is formed and a CH becomes active, when the acoustic signal strength detected by the CH exceeds a predetermined threshold. The active CH then broadcasts an information solicitation packet, asking sensors in its vicinity to join the cluster and provide their sensing information. We address and devise solution approaches (with the use of Voronoi diagram) to realize dynamic clustering: (I1) how CHs operate with one another to ensure that only one CH (preferably the CH that is closes to the target) is active with high probability, (I2) when the active CH solicits for sensor information, instead of having all the sensors in its vicinity reply, only a sufficient number of sensors respond with nonredundant, essential information to determine the target location, and (I3) both the packets that sensors send to their CHs and packets that CHs report to subscribers do not incur significant collision. Through both probabilistic analysis and ns-2 simulation, we use with the use of Voronoi diagram, the CH that is usually closes to the target is (implicitly) selected as the leader and that the proposed dynamic clustering algorithm effectively eliminates contention among sensors and renders more accurate estimates of target locations as a result of better quality data collected and less collision incurred.

In the paper, we devise and evaluate a fully decentralized, light-weight, dynamic clustering algorithm for target tracking. Instead of assuming the same role for all the sensors, we envision a hierarchical sensor network that is composed of (a) a static backbone of sparsely placed high-capability sensors which assume the role of a cluster head (CH) upon triggered by certain signal events; and (b) moderately to densely populated low-end sensors whose function is to provide sensor information to CHs upon request. A cluster is formed and a CH becomes active, when the acoustic signal strength detected by the CH exceeds a pre-determined threshold. The active CH then broadcasts an information solicitation packet, asking sensors in its vicinity to join the cluster and provide their sensing information. We address and devise solution approaches (with the use of Voronoi diagram) to realize dynamic clustering: (I1) how CHs cooperate with one another to ensure that for the most of time only one CH (preferably the CH that is closest to the target) is active; (I2) when the active CH solicits for sensor information, instead of having all the sensors in its vicinity reply, only a sufficient number of sensors respond with non-redundant, essential information to determine the target location; and (I3) both packets with which sensors respond to their CHs and packets that CHs report to subscribers do not incur significant collision. Through both probabilistic analysis and ns-2 simulation, we show with the use of Voronoi diagram, the CH that is usually closest to the target is (implicitly) selected as the leader and that the proposed dynamic clustering algorithm effectively eliminates contention among sensors and renders more accurate estimates of target locations as a result of better quality data collected and less collision incurred.

We propose an enhancement to DSR, called DSR with Connection-Aware Link-state Exchange aNd DiffERentiation (DSR-CALENDER). Specifically, to effectively collect and disseminate neighbor link states to nodes which may potentially use them, we devise a neighbor link-state information exchange mechanism: once a connection has been established, the neighbor link-state information is exchanged among nodes along the route from the source to the destination. As the information of the neighbor lists is piggybacked in data packets, the nodes on the source route are able to learn the partial topology around the neighborhood of the connection. In this manner, DSR-CALENDER balances the gain obtained from additional neighbor link states and the overhead incurred to distribute them. In the case that link failure occurs, an alternate route can be more effectively located in the route cache of the source or the intermediate nodes, and the number of times the route discovery phase is invoked can be reduced. To improve the quality of route caches, we investigate all possible sources from which a node may learn of its link states, and associate each of the sources with a different level of fidelity. In the case that a new/alternate route has to be computed, the level of fidelity is figured into the link cost so as to locate a route that likely exists. The simulation results show that with limited overhead incurred in neighbor list dissemination, DSR-CALENDER outperforms DSR with either path or link caches in terms of packet delivery ratio and route discovery overhead.

Abstract It is well known that the performance of TCP deteriorates in a mobile wireless environment. This is due to the fact that although the majority of packet losses are results of transmission errors over the wireless links, TCP senders still take packet loss as an indication of congestion, and adjust their congestion windows according to the additive increase and multiplicative decrease (AIMD) algorithm. As a result, the throughput attained by TCP connections in the wireless environment is much less than it should be. The key problem that leads to the performance degradation is that TCP senders are unable to distinguish whether packet loss is a result of congestion in the wireline network or transmission errors on the wireless links. In this paper, we propose a light‐weight approach, called syndrome, to improving TCP performance in mobile wireless environments. In syndrome, the BS simply counts, for each TCP connection, the number of packets that it relays to the destination host so far, and attaches this number in the TCP header. Based on the combination of the TCP sequence number and the BS‐attached number and a solid theoretical base, the destination host will be able to tell where (on the wireline or wireless networks) packet loss (if any) occurs, and notify TCP senders (via explicit loss notification, ELN) to take appropriate actions. If packet loss is a result of transmission errors on the wireless link, the sender does not have to reduce its congestion window. Syndrome is grounded on a rigorous, analytic foundation, does not require the base station to buffer packets or keep an enormous amount of states, and can be easily incorporated into the current protocol stack as a software patch. Through simulation studies in ns‐2 (UCB, LBNL, VINT network simulator, http://www‐mash.cs.berkeley.edu/ns/ ), we also show that syndrome significantly improves the TCP performance in wireless environments and the performance gain is comparable to the heavy‐weight SNOOP approach (either with local retransmission or with ELN) that requires the base station to buffer, in the worst case, a window worth of packets or states. Copyright © 2001 John Wiley & Sons, Ltd.

In this paper, we document our design of a MAC protocol, called OSU-MAC, subject to the physical layer characteristics and constraints of a narrow-band wireless modem testbed currently being built at the Ohio State University. The narrow-band wireless modem testbed is expected to support both real-time (bus location tracking) and non-real-time (regular) data applications. A number of techniques are proposed to support QoS imposed by the real-time applications, to deal with the asymmetry on the forward and reverse channels and the half-duplex transmission constraint imposed by the physical layer, and to enhance the error control capability of OSU-MAC. We also present simulation results to demonstrate the key functional characteristics of OSU-MAC.

This correspondence presents the first known results of complete recognition of continuous Mandarin speech for the Chinese language with very large vocabulary but very limited training data. Various acoustic and linguistic processing techniques were developed, and a prototype system of a continuous speech Mandarin dictation machine has been successfully implemented. The best recognition accuracy achieved is 92.2% for finally decoded Chinese characters.