Method and system for generating customer decision tree through machine learning

A system (100) for generating a Customer Decision Tree (CDT) in accordance with an attribute value (AV) based Demand Transfer (DT) estimation for a product category, the system (100) comprising:

• a data lake (106), as data source for aggregating data at a plurality of aggregation levels and provides data associated with the product category for all branches of an entity, wherein data structure and requirements are not defined until the data is considered as required data by the CDT system;
• a CDT system (102), directly connected with the data lake (106) comprising one or more hardware processors (202) executing computer-readable instructions stored in a memory (204), the stored computer-readable instructions further comprising a data matrix generator (212), a model generator (214), a prediction model (216), a scenario generator (220), a DT Estimator (218), a hierarchy generator (222) and a CDT generator (224), wherein:
• the data matrix generator (212) is configured to:
• aggregate data associated with a plurality of AVs of the product category managed by the entity from the data lake, wherein aggregation of the data is at a plurality of aggregation levels and the product category is identified with a plurality of attributes (ATs) with each AT having the plurality of AVs, wherein the plurality of aggregation levels include aggregation of data per second, per minute, per hour, per day, per week and per month for the product category under consideration, wherein a predefined aggregation level is approximated from the plurality of aggregation levels based on rate of sales movement of an inventory stock for the entity;
• generate a data matrix to represent data in a format for machine learning as a multivariate multi-dependent structure from the data aggregated at the predefined aggregation level for the product category for the plurality of AVs, wherein:
• a plurality of columns of the data matrix correspond to entity identity (ID) of the entity, the predefined aggregation level, sales drivers and recorded AV sales, and wherein the sales drivers comprise demographic data associated with the product category, inventory availability of the AV and inventory stock of the AV, wherein the plurality of columns of the data matrix are incremented as required to dynamically accommodate the sales drivers that affect the sales of the product category and enables processing large volume of data corresponding to all sales drivers and recorded attribute value (AV) sales for the product category;
• and
• a plurality of rows of the data matrix correspond to the data aggregated for a plurality of branches of the entity for the predefined aggregation level;
• the model generator (214) configured to:
• generate the prediction model (216), using the data aggregated as the multivariate multi-dependent structure and provided as an input data for machine learning process, to determine predicted AV sales for each AV of the product category at the predefined aggregation level, wherein the prediction model is a multivariate multi-structure machine learning model, wherein the recorded AV sales are represented in the matrix format that act as a multivariate dependent structure for the machine learning process to generate the prediction model, wherein the multivariate dependent structure captures interaction effects within dependent variables, and major sales drivers that includes inventory details of AVs providing the inventory availability of the AV and the inventory stock of the AV are represented in the matrix format that acts as multivariate independent structure for machine learning process, wherein the dependent structure and the independent structure captures interaction effect that exists in business scenario, wherein the recorded AV sales represent the dependent variables;
• and
• optimize the prediction model (216) to reduce error between the predicted AV sales and the recorded AV sales to minimum, wherein the optimization comprises:
• generate the data matrix for each aggregation level among the plurality of aggregation levels;
• determine the predicted AV sales at each aggregation level using the prediction model;
• determine error between the recorded AV sales and the predicted AV sales at each aggregation level;
• identify an optimum aggregation level from the plurality of aggregation levels based on the determined error, wherein the determined error for the identified aggregation level is minimum;
• refine the prediction model with the identified optimum aggregation level that is a data aggregation level for which the predicted AV sales for all AVs for the product category as predicted by the prediction model provides minimum variation from the recorded AV sales, wherein an optimized prediction model is further used for prediction of AV sales;
• and
• the CDT generator (224) configured to:
• generate the CDT utilizing the optimized prediction model, wherein the prediction model is refined with additional sales drivers and the CDT is generated for scenarios with sales patterns changing at a faster pace,
• wherein the optimum aggregation level reduces computation for generating the optimized prediction model and still maintain accuracy of AV sales prediction to generate more accurate CDT,
• wherein the machine learning process enhances scalability by handling high volume of aggregated data corresponding to sales drivers to generate the prediction model for predicting AV sales and increased number of sales drivers considered to improve accuracy of CDT as sales drivers have direct impact on AV sales,
• wherein the CDT is a graphical representation of customer's decision making process in form of a tree with a top node and a plurality of sub nodes, wherein the top node represents most important AT preferred by the customer and sub nodes represent a next important AT or AV for the customer, wherein the top node and the plurality of sub nodes are generated dynamically while generating the CDT and also creates new scenarios as per changing outcomes of at least one of the top node and the plurality of sub nodes resulting in reduction of total scenarios for generating CDT, wherein the CDT generation finds application in multiple domains where the decision is made based on multiple common criteria and each criteria includes multiple choice sets,
• wherein the domain includes a retail sector in which a group of mobile phones form the product category managed by a megastore and the plurality of attributes (ATs) associated with the mobile phone include brand, color, memory size, wherein the DT happens across varied combination of attributes that are influenced by customer lifestyle, demographics, seasonality and sales promotion, and sales drivers for all branches of the megastore across a city, a state, a country are considered,
• wherein the domain includes an insurance sector in which varied insurance policy options form the product category,
• wherein the domain include a banking sector and the CDT is generated for loan options provided by banks.

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:

• aggregating data associated with a plurality of AVs of a product category managed by an entity from a data lake directly connected to a CDT system, wherein data structure and requirements are not defined until the data is considered as required data by the CDT system, wherein aggregation of the data is at a plurality of aggregation levels and the product category is identified with a plurality of attributes (ATs) with each AT having the plurality of AVs (302), wherein the plurality of aggregation levels include aggregation of data per second, per minute, per hour, per day, per week and per month for the product category under consideration, wherein a predefined aggregation level is approximated from the plurality of aggregation levels based on rate of sales movement of an inventory stock for the entity, wherein the data lake is a data source providing data associated with the product category for all branches of the entity;
• generating a data matrix to represent data in a format for machine learning as a multivariate multi-dependent structure from the data aggregated at the predefined aggregation level for the product category for the plurality of AVs, wherein:
• a plurality of columns of the data matrix correspond to entity identity (ID) of the entity, the predefined aggregation level, sales drivers, and recorded AV sales, and wherein the sales drivers comprises demographic data associated with the product category, inventory availability of the AV and inventory stock of the AV, wherein the plurality of columns of the data matrix are incremented as required to dynamically accommodate the sales drivers that affect the sales of the product category and enables processing large volume of data corresponding to all sales drivers and recorded attribute value (AV) sales for the product category;
• and
• a plurality of rows of the data matrix correspond to the data aggregated for a plurality of branches of the entity for the predefined aggregation level;
• generating a prediction model, using the data aggregated from the data lake as the multivariate multi-dependent structure and provided as an input data for machine learning process, to determine predicted AV sales for each of the AV of the product category at the predefined aggregation level, wherein the prediction model is a multivariate multi-structure machine learning model, wherein the recorded AV sales are represented in the matrix format that act as a multivariate dependent structure for the machine learning process to generate the prediction model, wherein the multivariate dependent structure captures interaction effects within dependent variables, and major sales drivers that includes inventory details of AVs providing the inventory availability of the AV and the inventory stock of the AV are represented in the matrix format that acts as multivariate independent structure for machine learning process, wherein the dependent structure and the independent structure captures interaction effect that exists in business scenario, wherein the recorded AV sales represent the dependent variables;
• optimizing the prediction model to reduce error between the predicted AV sales and the recorded AV sales to minimum, wherein optimizing the prediction model comprises:
• generating the data matrix for each aggregation level among the plurality of aggregation levels;
• determining the predicted AV sales at each aggregation level using the prediction model;
• determining error between the recorded AV sales and the predicted AV sales at each aggregation level;
• identifying an optimum aggregation level from the plurality of aggregation levels based on the determined error, wherein the determined error for the identified aggregation level is minimum;
• refining the prediction model with the identified optimum aggregation level that is a data aggregation level for which the predicted AV sales for all AVs for the product category as predicted by the prediction model provides minimum variation from the recorded AV sales, wherein an optimized prediction model is further used for prediction of AV sales;
• and
• generating the CDT utilizing the optimized prediction model, wherein the prediction model is refined with additional sales drivers and the CDT is generated for scenarios with sales patterns changing at a faster pace,
• wherein the optimum aggregation level reduces computation for generating the optimized prediction model and still maintain accuracy of AV sales prediction to generate more accurate CDT,
• wherein the machine learning process enhances scalability by handling high volume of aggregated data corresponding to sales drivers to generate the prediction model for predicting AV sales and increased number of sales drivers considered to improve accuracy of CDT as sales drivers have direct impact on AV sales,
• wherein the CDT is a graphical representation of customer's decision making process in form of a tree with a top node and a plurality of sub nodes, wherein the top node represents most important AT preferred by the customer and sub nodes represent a next important AT or AV for the customer, wherein the top node and the plurality of sub nodes are generated dynamically while generating the CDT and also creates new scenarios as per changing outcomes of at least one of the top node and the plurality of sub nodes resulting in reduction of total scenarios for generating CDT, wherein the CDT generation finds application in multiple domains where the decision is made based on multiple common criteria and each criteria includes multiple choice sets,
• wherein the domain includes a retail sector in which a group of mobile phones form the product category managed by a megastore and the plurality of attributes (ATs) associated with the mobile phone include brand, color, memory size, wherein the DT happens across varied combination of attributes that are influenced by customer lifestyle, demographics, seasonality and sales promotion, and sales drivers for all branches of the megastore across a city, a state, a country are considered,
• wherein the domain includes an insurance sector in which varied insurance policy options form the product category,
• wherein the domain include a banking sector and the CDT is generated for loan options provided by banks.