DYNAMIC PREDICTION OF ONLINE SHOPPER'S INTENT USING A COMBINATION OF PREDICTION MODELS

• US 20140222503A1
• Filed: 02/06/2014
• Published: 08/07/2014
• Est. Priority Date: 02/07/2013
• Status: Active Grant

• Current Assignee: [24]7.ai, Inc.
• Original Assignee: 24/7 Customer Incorporated
• Inventors: VIJAYARAGHAVAN, Ravi, ADUSUMILLI, Kranthi Mitra, KULKARNI, Subhash Ramchandra, PRAKASH, Ravi
• Granted Patent: US 10,373,177 B2
• Time in Patent Office: Days
• US Class Current: 705/7.29
• CPC Class Codes: G06Q 30/0201 Market modelling; Market analysis; Collecting market data

Abstract

Online browsing behavior is used to predict the intent of online users dynamically. The category of online user is predicted and the prediction can be used to provide assistance to the user, if required. Such prediction is based on a combination of a Naïve'"'"'s Bayes classifier and a Markov model.

Claims

1. A computer implemented method for user intent analysis, comprising:

providing a processor configured for receiving a Web server request from a user through at least one of a plurality of channels;

said processor collecting data traffic of the user at the Web server;

said processor analyzing the collected data traffic of the user to determine a Web path taken by the user at the Web site during a visiting session as said Web site;

said processor beginning to infer user intent from the Web path as the visiting session commences;

said processor classifying the user based on said inferred user intent before the user leaves the Web site;

said processor training an assistance model using the user intent that was inferred; and

said processor providing virtual assistance by proactively including search suggestions to the user based on the assistance model.

2. A computer implemented method for user intent analysis, comprising:

providing a processor configured for receiving a Web server request from a user through at least one of a plurality of channels;

said processor collecting data traffic of the user at the Web server;

said processor analyzing Web path taken by the user in real time while the user is visiting a Web site; and

based upon said path analysis, said processor using dynamic models to infer user intent from the Web path as the user visit to the Web site commences.

3. The method of claim 2, further comprising:

based on said path analysis, said processor categorizing said user as any of;

a browser comprising users who browse without purpose;

a knowledge seeker comprising users who browse to gather information;

a purchaser with assistance comprising users who want to buy but are not very clear and, therefore, need assistance; and

a purchaser with self-help comprising users who know exactly what is wanted and where to find it.

4. The method of claim 2, further comprising:

said processor developing at least two predictive models based on said Web path analysis; and

said processor testing said at least two predictive models with Web path data of another of a plurality of users.

5. The method of claim 3, further comprising:

using said predictive models to predict the intent of the user; and

updating said predictive models based upon both a user category and by how far along a search path the user has progressed.

6. The method of claim 5, further comprising:

using said predictive models to enable a virtual support system which is provided to the user.

7. The method of claim 2, wherein said user'"'"'s Web navigation reveals information about a user'"'"'s intent;

and

wherein said intent either changes or gets stronger as the user'"'"'s Web journey matures.

8. The method of claim 2, further comprising:

applying a Naï

ve Bayes'"'"' function to predict raw intent with static information only; and

combining results of said Naï

ve Bayes'"'"' function with first-order Markov information to capture a dynamic nature of user intent with each step of a user'"'"'s navigation during a Web journey.

9. The method of claim 2, further comprising:

analyzing data traffic at a Web server by collecting click-stream data related to a Web path that said user takes when navigating through a Web site of a provider.

10. The method of claim 9, wherein said click-stream comprises a record of said user'"'"'s activity on the Internet including factors that comprise any of every Web site and every page of every Web site that the user visits, how long the user was on a page or site, in what order the pages were visited, any newsgroups in which the user participates, and the addresses of any email that the user sends and receives;

and

wherein the path Web taken by the user reflects the user'"'"'s intents, needs, and goals.

11. The method of claim 2, wherein said user'"'"'s online search behavior comprise any of goal-oriented search and exploratory search.

12. The method of claim 2, further comprising:

categorizing said user'"'"'s online search behavior to develop a plurality of intents comprising any of directed-purchase visits, search and deliberation visits, hedonic-browsing visits, and knowledge-building visits.

13. The method of claim 2, wherein a user may belong to one of the following categories:

knowledge seeker, browser, prospect, or customer.

14. An apparatus for user intent analysis, comprising:

a processor executing instructions to implement a user classifier module, said user classifier module classifying users into one of a plurality of categories by taking into account browsing and cognitive behavior of said user, said user classifier module generating a response variable that captures the user'"'"'s intent;

said user classifier module transforming navigational data of said user into a new data set;

said processor implementing a user tracker module, said user tracker module tracking and monitoring user browsing behavior in real time from any of a plurality of sources and reporting user activities to said user classifier module;

said processor implementing a page categorizer module, said page categorizer module reducing Web site URL'"'"'s of pages visited by said user during an online session into a smaller group of categories;

said processor modeling captured user-related data concerning the user'"'"'s visits to a Web site and transforming said data;

said processor dynamically predicting the user'"'"'s intent once the data is modeled; and

said processor using resulting data to create training and testing data for a plurality of users.

15. The apparatus of claim 14, further comprising:

based upon said classifying, said processor offering a virtual support system to said user as any of chat any interactive voice response (IVR) while said user navigates a provider'"'"'s Web site.

16. The apparatus of claim 14, further comprising:

a database capturing said user-related data, said user-related data comprising any of user'"'"'s ID, a browser used, referral, search words, start time, a navigational journey, and related information comprising any of sequence of pages (URLs) visited during a session, and time spent on the page.

17. The apparatus of claim 14, further comprising:

said processor employing a model comprising a combination of a Naï

ve Bayes classifier and a Markov model.

18. A computer implemented method for user intent analysis, comprising:

providing a processor executing instructions for preprocessing raw data;

said processor obtaining transformed data;

said processor measuring a set of predictors for an online user who is visiting a Web site, wherein the set of predictors is available at the outset of the user visit;

said processor considering a variable length set of variables comprising a sequence of uniform resource locators (URLs) that are viewed by said user during said visit up to a particular click;

said processor developing a model to predict the intent of the user by computing probabilities at said particular click and basing a decision on a comparison with a pre-determined threshold;

as the user'"'"'s journey of navigating the Web site progresses, said processor updating and evolving an understanding of the user'"'"'s intent; and

said processor engaging the user by offering to the user a suitable form of interaction in response to the user'"'"'s queries.

19. The method of claim 18, further comprising

discriminating between a fixed and variable part of user information, wherein said fixed set of variables is available during the beginning of the Web site visit;

approximating the probability of a sequence of page visits that is conditional on the user'"'"'s intent by applying a first order Markov chain, assuming that a next page visit of the user depends on a previous visit;

computing transition probabilities of the user visiting a particular page by using observed counts in each of a plurality of user intent counts;

training a model after computing required probabilities and transition probabilities matrices corresponding to training data available at a particular click;

wherein a final trained model comprises a collection of models corresponding to each page of said Web site visited by said user, along with decision criterion that is used to predict th user'"'"'s intent;

evaluating a combined Naï

ve Baye'"'"'s classifier model;

deploying said model; and

making a decision to determine a category of intent.

20. The method of claim 18, further comprising:

evaluating a next training and test data required by;

selecting training and test data at a selected click to provide a balanced training set comprising selecting a predetermined percentage of users from available users at said selected click for training the model, and selecting a remaining percentage of said available users for testing the model;

training said model using said training data;

estimating said model at each click and classifying user intent into one of a plurality of types;

wherein a resulting type is a maximum value of a modified Naï

ve Bayes classifier computed at each click; and

wherein the intent type may be changed when the user proceeds to visit a next Web page.

21. The method of claim 20, further comprising:

determining the intent of a user by taking the following rules take into account;

a decision is not taken until the user completes at least two page visits;

if the model output based on the two successive clicks indicates the intent as browser, the model continues to look for further information, no decision is taken;

if the model output based on the two successive clicks indicates the intent as self-help purchaser, a decision is not taken; and

in all other cases, the user needs assistance and the system makes a chat offer, and thereafter, the combined model is not evaluated.

22. An apparatus for user intent analysis, comprising:

a processor receiving a Web server request from a user through at least one of a plurality of channels;

said processor collecting data traffic of the user at the Web server;

said processor analyzing the collected data traffic of the user to determine a Web path taken by the user at the Web site during a visiting session as said Web site;

said processor beginning to infer user intent from the Web path as the visiting session commences;

said processor classifying the user based on inferred the user intent before the user leaves the Web site;

said processor training an assistance model using the user intent that was inferred; and

said processor providing virtual assistance proactively including search suggestions to the user based on the assistance model.

23. An apparatus for user intent analysis, comprising:

a processor receiving a Web server request from a user through at least one of a plurality of channels;

said processor collecting data traffic of the user at the Web server;

said processor analyzing Web path taken by the user in real time while the user is visiting a Web site; and

based upon said path analysis, said processor using dynamic models to infer user intent from the Web path as the user visit to the Web site commences.

24. An apparatus for user intent analysis, comprising:

a processor categorizing intent of one or more users in real time while they are searching one or more online commerce sites, wherein each user connects to a Web site with an intention of performing one more specific tasks;

said processer inferring said user'"'"'s intent early in a browsing session by tracking a Web path that the user follows on the Web site;

said processer using one or more dynamic models to predict user intent before the user abandons the Web site, based on a combination of models comprising a Naï

ve Bayes Classifier and a Markov model; and

said processer adapting said combined models depending on the user'"'"'s location along a search path;

wherein as the user progresses farther along the Web path, the model'"'"'s prediction of the user'"'"'s intent becomes more accurate.

25. A computer implemented method for user intent analysis, comprising:

providing a processor executing instructions for categorizing intent of one or more users in real time while they are searching one or more online commerce sites, wherein each user connects to a Web site with an intention of performing one more specific tasks;

said processer inferring said user'"'"'s intent early in a browsing session by tracking a Web path that the user follows on the Web site;

said processer using one or more dynamic models to predict user intent before the user abandons the Web site, based on a combination of models comprising a Naï

ve Bayes Classifier and a Markov model; and

said processer adapting said combined models depending on the user'"'"'s location along a search path;

wherein as the user progresses farther along the Web path, the model'"'"'s prediction of the user'"'"'s intent becomes more accurate.