Dynamic pricing of products in e-commerce using ensemble of contextual bandits

A system for dynamic pricing of one or more products in an e-commerce retail comprising:

• a memory storing instructions;
• one or more Input/Output (I/O) interfaces;
• and
• one or more hardware processors coupled to the memory via the one or more I/O interfaces, wherein the one or more hardware processors are configured by the instructions to:
• receive an input data from two primary sources pertaining to a plurality of Stock Keeping Units (SKUs), wherein the input data from a first primary source is based on one or more historical transactions, an inventory, a price master, an online click stream, and one or more attributes of one or more customers, and wherein the second primary source includes dynamics of the e-commerce landscape;
• pre-process the received input data to obtain pre-processed input data, wherein the pre-processing includes removing one or more predefined outliers based on clipping technique, imputing missing values based on a moving average technique, and a dimensionality reduction based on principal component analysis and a combination of normalization and standardization techniques based on minimum and maximum value and one or more distribution characteristics of the input data;
• aggregate the obtained pre-processed input data to a predefined level of granularity to get an aggregated data using a set of clustering and segmentation techniques;
• enable at least one of one or more triggers, wherein the one or more triggers include a competitor price distance beyond a predefined threshold, an inventory level below threshold, a predefined special event, holidays, one or more competitor products go out of stock and a change in an input cost beyond a predefined threshold;
• obtain a contextual data from the aggregated data based on the enabled one or more triggers for a price change, wherein the contextual data includes a plurality of contextual features comprising competitor price leakage ratio coefficients, cross price elasticity coefficients, customer clusters and seasonality indices;
• train one or more Contextual Bandit (CB) models based on the plurality of contextual features with the enabled one or more triggers to obtain a price recommendation corresponding to each of the one or more CB models—wherein each of the one or more CB models are tweaked by employing an approximate inference including a variational inference or a Markov Chain Monte Carlo (MCMC) with samples to speed up inference, a caching of previously computed values, a model approximation, and a sampling reduction to speed up computation of the CB model, wherein caching of the previously computed values includes posterior distributions for price points, thereby avoiding recomputing the computed values;
• provide an Ensemble of Contextual bandits for dynamic pricing in the E-commerce retail (ENCODE) model with a portfolio of varied bandit models, wherein the CB model at a level 0 includes information pertaining to a context that is local to each CB model at a time instant, and the ENCODE model in a level 1 include information pertaining to the context at the time instant and the context for future time intervals in accordance with projections, wherein the ENCODE model makes an informed decision about the price recommendations that is optimal both with respect to a current context and a projected context and the ENCODE model adaptively navigates in varied competitive scenarios including shifts in competitor pricing strategies over time rather than adhering to a fixed strategy;
• compute an optimal price recommendation of each of the one or more CB models based on a predefined ensemble model, wherein the predefined ensemble model calculates a cumulative reward, and wherein the cumulative reward is a combination of a current reward and a future reward corresponding to the time period until the next price change based on the price recommendation of each of the plurality of CB models;
• and
• recommend at least one optimal price computed by on the predefined ensemble model, wherein the predefined ensemble model following one or more ensemble strategies comprising a bootstrap aggregation technique, boosting technique and a simple average technique, wherein the one or more ensemble strategies comprising:
• assigning a weight to each of the one or more CB models based on predefined historical performance and using the assigned weights to aggregate prices;
• predict a final price against proposed prices of each of the one or more CB models using an advanced algorithmic learning technique and a voting mechanism;
• continuously monitoring performance of the ensemble model to adjust ensemble approach to optimize performance;
• and
• posting the final price on an e-commerce website upon selecting the final price by the ensemble model and obtaining live feedback in a user interface,
• wherein dynamic pricing of products in the e-commerce website establishes closed-loop feedback to continuously monitor performance of the ensemble model and adjust the weight and the voting mechanisms to optimize performance of the ensemble model,
• transition varied price recommendations from multiple CB algorithms in the level 0 into level 1, and in the level 1, the ensemble model selects a price recommendation and then flows to the level 2, and in the level 2, the price is posted on the e-commerce website. and made available in the user interface,
• wherein the ENCODE decides between price recommendations offered by multiple CB algorithms and progressively reduces number of times that the price recommendation is adopted.

One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause:

• receiving, via an input/output interface, an input data from two primary sources pertaining to a plurality of Stock Keeping Units (SKUs), wherein the input data from a first primary source is based on one or more historical transactions, an inventory, a price master, an online click stream, and one or more attributes of one or more customers and SKUs, and wherein the second primary source includes dynamics of the e-commerce landscape;
• pre-processing the received input data to obtain pre-processed input data, wherein the pre-processing includes removing one or more predefined outliers based on clipping technique, imputing missing values based on a moving average technique, and a dimensionality reduction based on principal component analysis and a combination of normalization and standardization techniques based on minimum and maximum value and one or more distribution characteristics of the input data;
• aggregating the obtained pre-processed input data to a predefined level of granularity to get an aggregated data using a set of clustering and segmentation techniques;
• enabling at least one of one or more triggers, wherein the one or more triggers include a competitor price distance beyond a predefined threshold, an inventory level below threshold, a predefined special event, holidays, one or more competitor products going out of stock and a change in an input cost beyond a predefined threshold;
• obtaining a contextual data from the aggregated data based on the enabled one or more triggers for a price change, wherein the contextual data includes a plurality of contextual features comprising competitor price leakage ratio coefficients, cross price elasticity coefficients, customer clusters and seasonality indices;
• training one or more Contextual Bandit (CB) models based on the plurality of contextual features with the enabled one or more triggers to obtain a price recommendation corresponding to each of the one or more CB models wherein each of the one or more CB models are tweaked by employing an approximate inference including a variational inference or a Markov Chain Monte Carlo (MCMC) with samples to speed up inference, a caching of previously computed values, a model approximation and a sampling reduction to speed up computation of the CB model, wherein caching of the previously computed values includes posterior distributions for price points, thereby avoiding recomputing the computed values;
• providing an Ensemble of Contextual bandits for dynamic pricing in the E-commerce retail (ENCODE) model with a portfolio of varied bandit models, wherein the CB model at a level 0 includes information pertaining to a context that is local to each CB model at a time instant, and the ENCODE model in a level 1 include information pertaining to the context at the time instant and the context for future time intervals in accordance with projections, wherein the ENCODE model makes an informed decision about the price recommendations that is optimal both with respect to a current context and a projected context and the ENCODE model adaptively navigates in varied competitive scenarios including shifts in competitor pricing strategies over time rather than adhering to a fixed strategy;
• computing an optimal price recommendation of each of the one or more CB models based on a predefined ensemble model, wherein the predefined ensemble model calculates a cumulative reward, and wherein the cumulative reward is a combination of a current reward and a future reward corresponding to the time period until the next price change based on the price recommendation of each of the plurality of CB models;
• and
• recommending at least one optimal price computed by on the predefined ensemble model, wherein the predefined ensemble model following one or more ensemble strategies comprising a bootstrap aggregation technique, boosting technique, a simple average technique, predefined advanced algorithmic learning technique, and voting techniques wherein the one or more ensemble strategies comprising:
• assigning a weight to each of the one or more CB models based on predefined historical performance and using the assigned weights to aggregate prices;
• predicting a final price against proposed prices of each of the one or more CB models;
• continuously monitoring performance of the ensemble model to adjust weight and voting mechanisms to optimize performance;
• and
• posting the final price on an e-commerce website upon selecting the final price by the ensemble model and obtaining live feedback in a user interface,
• wherein dynamic pricing of products in the e-commerce website establishes closed-loop feedback to continuously monitor performance of the ensemble model and adjust the weight and the voting mechanisms to optimize performance of the ensemble model,
• transitioning varied price recommendations from multiple CB algorithms in the level 0 into level 1, and in the level 1, the ensemble model selects a price recommendation and then flows to the level 2, and in the level 2, the price is posted on the e-commerce website, and made available in the user interface,
• wherein the ENCODE decides between price recommendations offered by multiple CB algorithms and progressively reduces number of times that the price recommendation is adopted.