Lane detection is a key task for autonomous driving vehicles. Currently, lane detection relies on a huge amount of annotated images, which is a heavy burden. Active learning has been proposed to reduce annotation in many computer vision tasks, but no effort has been made for lane detection. Through experiments, we find that existing active learning methods perform poorly for lane detection, and the reasons are twofold. On one hand, most methods evaluate data uncertainties based on entropy, which is undesirable in lane detection because it encourages to select images with very few lanes or even no lane at all. On the other hand, existing methods are not aware of the noise of lane annotations, which is caused by heavy occlusion and unclear lane marks. In this paper, we build a novel knowledge distillation framework and evaluate the uncertainty of images based on the knowledge learnt by the student model. We show that the proposed uncertainty metric overcomes the above two problems. To reduce data redundancy, we explore the influence sets of image samples, and propose a new diversity metric for data selection. Finally we incorporate the uncertainty and diversity metrics, and develop a greedy algorithm for data selection. The experiments show that our method achieves new state-of-the-art on the lane detection benchmarks. In addition, we extend this method to common 2D object detection and the results show that it is also effective.

Matching buyers with most suitable sellers providing relevant items (e.g., products) is essential for e-commerce platforms to guarantee customer experience. This matching process is usually achieved through modeling inter-group (buyer-seller) proximity by e-commerce ranking systems. However, current ranking systems often match buyers with sellers of various qualities, and the mismatch is detrimental to not only buyers' level of satisfaction but also the platforms' return on investment (ROI). In this paper, we address this problem by incorporating intra-group structural information (e.g., buyer-buyer proximity implied by buyer attributes) into the ranking systems. Specifically, we propose De ep Gr aph E mbe dding (DEGREE), a deep learning based method, to exploit both inter-group and intra-group proximities jointly for structural learning. With a sparse filtering technique, DEGREE can significantly improve the matching performance with computation resources less than that of alternative deep learning based methods. Experimental results demonstrate that DEGREE outperforms state-of-the-art graph embedding methods on real-world e-commence datasets. In particular, our solution boosts the average unit price in purchases during an online A/B test by up to 11.93%, leading to better operational efficiency and shopping experience.

Strategy design and analysis is important in team sports, such as basketball and soccer. In this paper, we take basketball as an example and study how to cluster movement trajectories in the games to identify the strategies. This problem is challenging in that the trajectories are diverse and that it is unknown how many or what strategies are employed in the games. As a result, traditional parametric clustering methods are not directly applicable to the raw trajectory data. Therefore, we propose to align trajectories around the basket and simplify them based on movement directions and game events, including dribbling, passing, and shooting. Furthermore, we propose a non-parametric density peak (NPDP) method to cluster these simplified event trajectories. Our experiments on an NBA game dataset of 50,000 offenses show that, without parameter tuning, NPDP clusters all trajectories into groups of high similarity and identifies distinguishing movement strategies.

Taxi fraud has become a serious problem in many large cities, where passengers are overcharged by taxi drivers in various ways. Researchers have developed a number of methods to detect taxi frauds with the assumption that fraudulent trips, among normal trips, are recorded by taximeters. In this paper, different from the previous work, we identify a new type of taxi fraud called unmetered taxi rides, where taxi drivers carry passengers without activating the taximeters. Since these fraudulent rides are not recorded by taximeters, previous detection approaches cannot directly apply to them. Hence, we propose a novel fraud detection system specifically designed for unmetered taxi rides. Our system uses a learning model to detect unmetered trajectory segments that are similar to metered rides, and introduces a heuristic algorithm to construct maximum fraudulent trajectories from the trajectory dataset. We have conducted detailed experiments on real-world datasets, and the results show that the proposed system can detect unmetered taxi rides effectively and efficiently.

Active learning has been widely used to select the most informative data for labeling in classification tasks, except for time series classification. The main challenge of active learning in time series classification is to evaluate the informativeness of a time series instance. Specifically, many informativeness metrics have been proposed for traditional active learning, however, none of them is particularly effective on time series data. In this paper, we design an informativeness metric that considers the characteristics of time series data in defining our instance uncertainty and utility. We prove that our informativeness metric is a submodular set function, and further develop an effective and efficient algorithm to select the most informative time series instances for training. In the experiment, we validate our method on a variety of datasets in the UCR Time Series Data Archive. The results show that our method achieves a higher classification accuracy than existing methods, using only 50% of the training instances.

Hash Ripple join is an online aggregation algorithm that can rapidly give good approximate join results increases with the progress of the join operation and converges to the real result when the join finishes. Luo et al. proposed a parallel hash ripple join (PHRJ) that runs in a distributed setting. However, the PHRJ has two drawbacks when handling large-scale data: 1) PHRJ updates approximate results in fine grain which induces extra communication cost in a distributed environment 2) When data is out of memory, PHRJ cannot provide unbiased approximate result. In this thesis, a scalable hash ripple join is proposed that 1) runs on a distributed framework that can process distributed data in coarse-grain to speed up the join performance; 2) continuously gives unbiased and consistent approximate join results even in the presence of memory overflow; and 3) has good scalability handling growing amount of data. We have implemented a prototype of the scalable hash ripple join (SHRJ) algorithm on Spark. Experiment results show that SHRJ can give good approximate join result while taking less than 10% of the time of Spark's own join operator.

The maintenance of batteries used in wireless mobile communication is an important practical problem. The experts can easily recognize the battery events, such as turning on, by watching the monitoring data. However it is infeasible to have experts watch the data all the time. There are devices that can report battery events. These devices sometimes report incorrect event. In order to solve this problem, we propose a time series classification framework to use the expert knowledge to build an accurate classifier, and then use the classifier to monitor the batteries in real time. We first propose an active learning method to efficiently collect the experts' labels for each event. Then we apply various feature extraction methods to convert each time series segment into a feature vector. Finally, we apply random forest classifier to perform the classification. Moreover, in practice, the labeled data is unbalanced, i.e. >99% of the data instances belong to a single label. We use bootstrap to solve this problem. We test our method on a dataset for 500 batteries in 3 months. The results show that our method achieves a very high classification accuracy, using only less than 1% of the dataset as training set.