Search and recommender systems have produced highly relevant search results. A natural next step in the development of such systems in e-commerce is to rerank these results to increase the platform's revenue from paid promotion products. However, maximizing revenue alone may degrade the user experience by reducing relevance or increasing fraud risk. To avoid this, we state the reranking problem as an integer linear program (ILP) that maximizes revenue subject to per-query constraints on other metrics, e.g., relevance. Since solving ILP exactly for every query is slow for deployment to the online service, we propose a lightweight permutation-based reranking approximation algorithm PermR. At each step, the algorithm selects a pair of neighboring items and swaps them to either improve the objective or repair a violated constraint. We evaluate PermR across multiple categories of a large classified platform in offline and online settings. PermR achieves about 63% of the ILP revenue improvement, within production latency limits, preserving all constraints. In a 14-day online A/B test over 56 million search queries, PermR increased revenue by 2%.

Modern Deep Neural Networks (DNNs) require significant memory to store weight, activations, and other intermediate tensors during training. Hence, many models don’t fit one GPU device or can be trained using only a small per-GPU batch size. This survey provides a systematic overview of the approaches that enable more efficient DNNs training. We analyze techniques that save memory and make good use of computation and communication resources on architectures with a single or several GPUs. We summarize the main categories of strategies and compare strategies within and across categories. Along with approaches proposed in the literature, we discuss available implementations.

We propose a simple interpolation-based method for the efficient approximation of gradients in neural ODE models. We compare it with reverse dynamic method (known in literature as “adjoint method”) to train neural ODEs on classification, density estimation and inference approximation tasks. We also propose a theoretical justification of our approach using logarithmic norm formalism. As a result, our method allows faster model training than the reverse dynamic method what was confirmed and validated by extensive numerical experiments for several standard benchmarks.

This paper presents a new fast clustering algorithm RhoNet, based on the metric concenration location procedure. To locate the metric concentration, the algorithm uses a reduced matrix of pairwise ranks distances. The key feature of the proposed algorithm is that it doesn't need the exhaustive matr ix of pairwise distances. This feature reduces computational complexity. It is designed to solve the protein secondary structure recognition problem. The computational experiment collects tests and to hold performance analysis and analysis of dependency for the algorithm quality and structure parameters. The algorithm is compared with k-modes and tested on different metrics and data sets.