Recommender systems (RSs) have been the most important technology for increasing the business in Taobao, the largest online consumer-to-consumer (C2C) platform in China. There are three major challenges facing RS in Taobao: scalability, sparsity and cold start. In this paper, we present our technical solutions to address these three challenges. The methods are based on a wellknown graph embedding framework. We first construct an item graph from users' behavior history, and learn the embeddings of all items in the graph. The item embeddings are employed to compute pairwise similarities between all items, which are then used in the recommendation process. To alleviate the sparsity and cold start problems, side information is incorporated into the graph embedding framework. We propose two aggregation methods to integrate the embeddings of items and the corresponding side information. Experimental results from offline experiments show that methods incorporating side information are superior to those that do not. Further, we describe the platform upon which the embedding methods are deployed and the workflow to process the billion-scale data in Taobao. Using A/B test, we show that the online Click-Through-Rates (CTRs) are improved comparing to the previous collaborative filtering based methods widely used in Taobao, further demonstrating the effectiveness and feasibility of our proposed methods in Taobao's live production environment.

SM2miR is freely available at http://bioinfo.hrbmu.edu.cn/SM2miR/.

Respiratory motion (RM) and cardiac motion (CM) degrade the quality and resolution in cardiac PET scans. We have developed non-rigid motion estimation methods to estimate both RM and CM based on 4D cardiac gated PET data alone, and compensate the dual respiratory and cardiac (R&C) motions after (MCAR), during (MCDR), and before (MCBR) image reconstruction. In all three R&C motion correction methods, attenuation-activity mismatch effect was modeled by using transformed attenuation maps using the estimated RM. The difference of using activity preserving and non-activity preserving models in R&C correction was also studied. Realistic Monte Carlo simulated 4D cardiac PET data using the 4D XCAT phantom and accurate models of the scanner design parameters and performance characteristics at different noise levels were employed as the known truth and for method development and evaluation. Results from the simulation study suggested that all three dual R&C motion correction methods provide substantial improvement in the quality of 4D cardiac gated PET images as compared with no motion correction. Specifically, the MCDR method yields the best performance for all different noise levels compared with the MCAR and MCBR methods. While MCBR reduces computational time dramatically but the resultant 4D cardiac gated PET images has overall inferior image quality when compared to that from the MCAR and MCDR approaches in the 'almost' noise free case. Also, the MCBR method has better noise handling properties when compared with MCAR and provides better quantitative results in high noise cases. When the goal is to reduce scan time or patient radiation dose, MCDR and MCBR provide a good compromise between image quality and computational times.

The goal is to develop an adaptive center-of-mass (COM)-based approach for device-less respiratory gating of list-mode positron emission tomography (PET) data. Our method contains two steps. The first is to automatically extract an optimized respiratory motion signal from the list-mode data during acquisition. The respiratory motion signal was calculated by tracking the location of COM within a volume of interest (VOI). The signal prominence (SP) was calculated based on Fourier analysis of the signal. The VOI was adaptively optimized to maximize SP. The second step is to automatically correct signal-flipping effects. The sign of the signal was determined based on the assumption that the average patient spends more time during expiration than inspiration. To validate our methods, thirty-one F-FDG patient scans were included in this paper. An external device-based signal was used as the gold standard, and the correlation coefficient of the data-driven signal with the device-based signal was measured. Our method successfully extracted respiratory signal from 30 out of 31 datasets. The failure case was due to lack of uptake in the field of view. Moreover, our sign determination method obtained correct results for all scans excluding the failure case. Quantitatively, the proposed signal extraction approach achieved a median correlation of 0.85 with the device-based signal. Gated images using optimized data-driven signal showed improved lesion contrast over static image and were comparable to those using device-based signal. We presented a new data-driven method to automatically extract respiratory motion signal from list-mode PET data by optimizing VOI for COM calculation, as well as determine motion direction from signal asymmetry. Successful application of the proposed method on most clinical datasets and comparison with device-based signal suggests its potential of serving as an alternative to external respiratory monitors.

No abstract