Industrial systems such as recommender systems and online advertising, have been widely equipped with multi-stage architectures, which are divided into several cascaded modules, including matching, pre-ranking, ranking and re-ranking. As a critical bridge between matching and ranking, existing pre-ranking approaches mainly endure sample selection bias (SSB) problem owing to ignoring the entire-chain data dependence, resulting in sub-optimal performances. In this paper, we rethink pre-ranking system from the perspective of the entire sample space, and propose Entire-chain Cross-domain Models (ECM), which leverage samples from the whole cascaded stages to effectively alleviate SSB problem. Besides, we design a fine-grained neural structure named ECMM to further improve the pre-ranking accuracy. Specifically, we propose a cross-domain multi-tower neural network to comprehensively predict for each stage result, and introduce the sub-networking routing strategy with L0 regularization to reduce computational costs. Evaluations on real-world large-scale traffic logs demonstrate that our pre-ranking models outperform SOTA methods while time consumption is maintained within an acceptable level, which achieves better trade-off between efficiency and effectiveness.

Fall detection is an important public healthcare problem. Timely detection could enable instant delivery of medical service to the injured. A popular nonintrusive solution for fall detection is based on videos obtained through ambient camera, and the corresponding methods usually require a large dataset to train a classifier and are inclined to be influenced by the image quality. However, it is hard to collect fall data and instead simulated falls are recorded to construct the training dataset, which is restricted to limited quantity. To address these problems, a three-dimensional convolutional neural network (3-D CNN) based method for fall detection is developed, which only uses video kinematic data to train an automatic feature extractor and could circumvent the requirement for large fall dataset of deep learning solution. 2-D CNN could only encode spatial information, and the employed 3-D convolution could extract motion feature from temporal sequence, which is important for fall detection. To further locate the region of interest in each frame, a long short-term memory (LSTM) based spatial visual attention scheme is incorporated. Sports dataset Sports-1 M with no fall examples is employed to train the 3-D CNN, which is then combined with LSTM to train a classifier with fall dataset. Experiments have verified the proposed scheme on fall detection benchmark with high accuracy as 100%. Superior performance has also been obtained on other activity databases.