Browse Prior Art Database

Detection of focused vs. wanderer customers based on in-store movement trackings

IP.com Disclosure Number: IPCOM000234099D
Publication Date: 2014-Jan-12
Document File: 3 page(s) / 41K

Publishing Venue

The IP.com Prior Art Database

Abstract

Understanding and classifying shoppers in physical stores is a challenge, since it requires analyzing data about unique individuals. Most in-store analytics provide traffic pattern analysis in the aggregate, since identifying unique individual has been a challenge. With the broad use of smartphones, it is now easier to identify unique individuals in stores by tracking signals such as Bluetooth, WIFI, RFID, etc emitted from their mobile devices. In-store tracking solutions such as IBM Presence Zones and others, provide location tracking for unique individuals based on the mobile device carried by the user. This provides new sources of data for analyzing customer behavior, which can be used to improve store decision making and marketing strategies. Utilizing such data can be used to learn various customer segments based on machine learning techniques to discover various customer behavioral models and derive business rules based on that. An extremely important customer behavioral discovery task is the one of detection of focused customers (with high chance to buy) vs. wanderer customers (which usually just "waste" their spare time in-store). Such unique detection solely based on customer movement data has a lot of business value in the retail industry, and could help to reshape and optimize marketing decision making in-store. Existing solutions are lacking of such detection capabilities, and existing BI is focused on simple reporting based on MAC id analysis. Given records of customer movements in store in various store areas (zones), an unsupervised learning method is proposed for prediction of focused vs. "wanderer" customers in store. Detected segments can be further analyzed to derive business rules based on focus propensity. Finally, detected customer segments can be utilized to predict customer segment association for online decision making.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 50% of the total text.

Page 01 of 3

-store movement trackings

store movement trackings

We assume the input of customer in-store movements data. Such data can be obtained by means of wi-fi monitoring, RFIDs, etc and other technologies that provide information about unique individuals.

We assume that each tracking event is associated with a unique customer id, zone id (e.g., Mens cloths), and enter and exit times.

The following are the various analytics steps for learning focused vs. wanderer

customer segments using various machine learning techniques.

Step 1: Data processing:


Raw zone event data records were transformed into customer visit records

A visit is all events in which the exit of event N is at the same time (or at most 20 seconds before) as the enter in event N+1


Excluded "spotty" data - i.e. where mobile device is not captured consistently

throughout the store visit (such as with devices that don't ping the network often)


A visit is valid if it starts with a store event, includes at least one zone event, and its last zone event exit time agrees with the store event exit time


A co-visit is group of visits, each at most 5 seconds apart. i.e. People shopping together

A returning visit is a visit of same customer within 20 days of her last visit


Zones were further grouped based on their type, e.g.: Mens Zone: Mens-T-Shirts, Mens-Shoes, Mens-Pents

Step 2: Feature Extraction:


Global Features

Store visit time (sec)

Visit start time of day (hour)

GroupSize A visitor is in a group if entered and left the store exactly the same

time as at least an additional customer. A size of 0 means that the user is not in a group.

WeekDay 1..7, 1 for Sunday, 7 for Saturday.

IsWeekend 1 if so, 0 otherwise. Set to 1 for Saturday or Sunday.

ReVisitStrength20 0 for non returning customer, 1 if returned after 20 days, 2 if

returned after 19 days, … 19 if returned after 1 day, 20 if returned at the same day.

IsReturningCustomer 1 if ReVisitStrength20 is greater than 0.

CashiersTime time (seconds) spent in cashier zones in the last event of the

visit.

IsPurchase 1 if CashiersTime is at least 60 seconds, 0 otherwise.


Areal Features (Zone and Zone Group)

Num. area visits


Avg. area visit time


Stdev. area visit time


Area change rate = #area visits / total visit time in seconds


Num. unique area visits


Num area visits diversity - based on Entropy, computed as: − Pa · log(Pa)

wanderer customers based on in

wanderer customers based on in -

Detection of focused vs . .


Page 02 of 3

,when: Pa= num−visits−in−area−a / num−area−visits

Diversity of accumulated area visit time - Similar to previous diversity, just that

    here:Pa = total−time−spent−in−area−a / total−visit−time Step 3: Cluster Analysis (Customer segments detection)

A K-Means with k=2 provided the most coherent clusters (Silhouette ~ 0.8, 14 features)

Manual cluster exploration reveals clear behavioral differences among the two

customer segments, labeled hereinafter "focused" and "wanderer" customers.


Cluste...