System and method for generating recommendations

Patent metadata

Field	Value
Publication number	US-11392840-B2
Application number	US-201615094027-A
Country	US
Kind code	B2
Filing date	Apr 8, 2016
Priority date	Apr 10, 2015
Publication date	Jul 19, 2022
Grant date	Jul 19, 2022

Abstract

Disclosed is method and system for generating recommendations to a user. System receives real time data associated with users for scenarios and batch data associated with multiple users from different data sources, received from different data channels. The user is online user. System pre-processes batch data and real time data to generate pre-processed data and stores preprocessed data in distributed, scalable big data store. System filters pre-processed data based on rules to obtain filtered data. System applies combination of machine learning techniques on filtered data, based on the scenarios associated with the user, leveraging inter-play between machine learning techniques, to generate personalized recommendations for individual user and storing the personalized recommendations in distributed database. Machine learning techniques are customized to work in distributed processing mode to reduce overall processing time. System recommends user with the personalized recommendations comprising products or services.

First claim

What is claimed is:

1. A processor-implemented method for generating personalized recommendations in real time using predictive user data obtained from one or more users, the method comprising:
receiving, by one or more hardware processors,
real time data associated with a user including search terms by the user, click stream log details, location, time and date of browsing, wherein copy of the real time data is saved in a recommendation file through a sink operator for transmitting the copy of the real time data to a data repository, wherein the sink operator only in case of the real time data modifies a format of the real time data and transmits the formatted real time data to the recommendation file of the data repository, wherein the real-time data corresponds to a plurality of scenarios associated with the user identified as type of the user as an unknown user visiting a website for a first time or unregistered in the website, or a known user visiting the website after registering in the website, and wherein the plurality of scenarios are developed for the unknown user and the known user and comprise online browsing actions performed at a website including a home page, a search page, a product description page and an order confirmation page, wherein the online browsing actions are self-learnt by a system and most recent activities are learned for generating personalized recommendation and the learning ability of a recommendation engine in the system is improved by considering additional data sources and frequency of personalized recommendations refreshed for an online user; and
batch data associated with one or more users, from one or more data sources the batch data being, received from a plurality of data channels by the one or more data sources, wherein the real time data and the batch data are moved to the data repository through a data staging layer to temporarily store the real time data and the batch data;

pre-processing, by the one or more hardware processors, the batch data and the real time data, to generate pre-processed data, wherein the pre-processing comprises:
merging and cleansing the real-time data and the batch data, wherein merging includes transforming the real-time data and the batch data to sync with a system requirement, wherein transformation of the real time data extract only the search terms used by the user from stream of web logs; and
indexing the merged and cleansed real-time data and the batch data to obtain the pre-processed data;

storing, by the one or more hardware processors, the pre-processed data in a distributed database;
filtering, by the one or more hardware processors, the preprocessed data based on one or more rules to obtain filtered data that is moved to a modeling pane in a functional architecture, wherein the filtering comprises:
receiving inputs to create or modify the one or more rules associated with the plurality of scenarios, wherein the one or more rules are around a time frame for the batch data is considered for making product recommendations, or product categories is considered or excluded as part of the plurality of scenarios, wherein the one or more rules further comprises modifiable attributes associated with items or services, a time period, threshold values for a plurality of machine learning techniques, and assign weights to the modifiable attributes;
enabling a controlled simulation of the plurality of scenarios associated with the user, setting-up a recommendation strategy by the items or services, and a web content management process in sync with a recommendation approach;
mapping the user to a particular age bucket based on items bought in the past by the users, items added to wish list, items viewed, items reviewed, items liked and pinned on social media;
building a machine learning model for a combination of a plurality of machine learning techniques selected from a list of machine learning techniques based on the plurality of scenarios associated with the user, wherein each of the machine learning techniques is built on a MapReduce model to handle the filtered data from the distributed database and works in a distributed processing mode, wherein the plurality of machine learning techniques includes an user-based collaborative filtering, an item-based collaborative filtering, clustering, classification, pattern mining, Shingling, min hashing, locality sensitive hashing, a sentiment analysis or sentiment mining, and semantic analysis, a Pearson similarity, a cosine similarity and a Jaccard similarity technique,

wherein the machine learning techniques including Jaccard similarity and Shingling are custom coded on the MapReduce model to enhance capability to process large volumes of the real time data and the batch data and complete processing of the real time data and the batch data in short time frames by leveraging distributed capability of the MapReduce model, wherein an input file to the Jaccard similarity is provided in a Cartesian input format with each record in the input file containing a key and a value, wherein the key refers to a single user data for whom similarity value needs to be computed and the value refers to all other users against whom a similarity value is to be computed;
applying, by the one or more hardware processors, the machine learning model built with the combination of a plurality of machine learning techniques, to leverage interplay between the machine learning techniques, on the filtered data based on the particular age bucket mapped with the user;
generating, by the one or more hardware processors, a set of personalized recommendations associated with the known user differently from the unknown user based on the applied plurality of machine learning techniques;
implementing the machine leaning model built on one or more instances of a production server to obtain a score of impact of the set of personalized recommendations generated by the machine learning model;
refining the machine learning model by modifying combination of the plurality of machine learning techniques if the score of impact is less than a threshold value;
applying the modified combination of the plurality of machine learning techniques; and
generating a set of personalized recommendations based on the modified combination of plurality of machine learning techniques.

2. The method of claim 1 , wherein the one or more data sources, comprises a retailer database, a user database, a product database, and a service database, wherein the one or more data sources are received from a plurality of data channels comprising a store, mobile, web, and social media.

3. The method of claim 1 , wherein the batch data comprises user history data, user activity data, user transaction details, order history, product attributes, service attributes, pricing data, Inventory data, Clickstream data, click stream logs, device data, Internet Protocol (IP) address, time stamp data, user location details and Social data.

4. The method of claim 1 , further comprises applying the combination of the plurality of machine learning techniques of the machine learning model, if the score is more than or equal to the threshold value to generate the set of personalized recommendations.

5. The method of claim 1 , further comprises recommending the online user with the set of personalized recommendations in decreasing order of user's propensity to buy the recommendations, wherein the set of personalized recommendations comprises at least one of items or services.

6. The method of claim 1 , wherein the selection of the plurality of machine learning techniques based on the plurality of scenarios associated with the user, and application of combination of the plurality of machine learning techniques so customized to work on a scalable big data store leverages distri

Assignees

• Tata Consultancy Services Ltd
• Tata Consultancy Ltd Services

Inventors

• Santhanam Janardhan
• Motani Gaurav
• Joshua Allan

Classifications

• G06N20/20 Primary

  Ensemble learning · CPC title

• G06Q30/0631

  Recommending goods or services · CPC title

• G06N20/00

  Machine learning · CPC title

• G06N5/04 Primary

  Inference or reasoning models · CPC title