Abstract: Systems, methods, and computer-readable storage media for reinforcing and modifying the search engine behavior using aspects of federated learning. A system can receive from a user at a first time, a query for a product, then generate a vector representation of the query, and execute a search of an embedding space for the vector representation, resulting in at least one similarity score for at least one product. The system can present to the user search results based on the at least one similarity score, and receive from the user a selected product from within the search results. The system can update, based on the selected product, the at least one similarity score for the selected product, resulting in at least one updated similarity score, where when the query is received again the search of the embedding space for the vector representation produces the at least one updated similarity score.

Abstract: Aspects of the subject technology relate to systems, methods, and computer-readable media for predicting energy states of devices operating in an environment. An operational constraint and a device setting for a device operating in an environment can be identified. An environmental state of the environment can be predicted based on the device operating at the device setting and under the operational constraint through application of an environmental state model that maps varying operational constraints and varying device settings to varying environmental states in the environment. An energy state for operating the device in the environment at the device setting and under the operational constraint can be predicted based on the predicted environmental state through application of an energy consumption model that maps the varying environmental states to varying energy states.

Abstract: A computer method for identifying product in a distributor's inventory system that fulfills a product request made via a natural language query. the natural language query is received as a product request including multiple words in sequential order. The words are vectorized into word-vectors that are concatenated and used to generate a query embedding. The query embedding is processed utilizing a trained product category classifier that predicts which product category the requested product belongs. Forward and backward sequence vectors are generated from the sequentially ordered words of the query that are concatenated and processed using a trained model specific to the predicted product category. The sequence vectors represent positional relationships between the words of the natural language query. Thereafter, the system identifies product attribute(s) embodied in the natural language query that each correspond to a predetermined key-characteristic of the category.

Abstract: A computer implemented method for identifying substitute products for a target product via a distributor's product management computer system, the method includes: receiving, at a distributor's product management computer system, data representing a target product of interest to a user of the system, wherein the target product is described by a plurality of variable-value attributes associated therewith; processing the received data through a machine-learning-produced algorithm and thereby generating data representing a first set of substitute product candidates for the target product and wherein each substitute product candidate has a corresponding match-score that represents a degree of determined similarity between that substitute product candidate and the target product; transforming the generated data that represents the first set of substitute product candidates into data representing a refined set of substitute product candidates by processing the generated data utilizing a constraint-based algorithm;

Abstract: A computer method for identifying product in a distributor's inventory system that fulfills a product request made via a natural language query. the natural language query is received as a product request including multiple words in sequential order. The words are vectorized into word-vectors that are concatenated and used to generate a query embedding. The query embedding is processed utilizing a trained product category classifier that predicts which product category the requested product belongs. Forward and backward sequence vectors are generated from the sequentially ordered words of the query that are concatenated and processed using a trained model specific to the predicted product category. The sequence vectors represent positional relationships between the words of the natural language query. Thereafter, the system identifies product attribute(s) embodied in the natural language query that each correspond to a predetermined key-characteristic of the category.