www.umbc.edu personalized recommender systems in e-commerce and m- commerce: a comparative study...

21
www.umbc.edu Personalized Recommender Personalized Recommender Systems in e-Commerce and m- Systems in e-Commerce and m- Commerce: A Comparative Study Commerce: A Comparative Study Azene Zenebe, Ant Ozok and Anthony F. Norcio Department of Information Systems University of Maryland Baltimore County (UMBC) Baltimore, MD 21250 USA

Upload: britney-cox

Post on 01-Jan-2016

214 views

Category:

Documents


0 download

TRANSCRIPT

www.umbc.edu

Personalized Recommender Personalized Recommender Systems in e-Commerce and m-Systems in e-Commerce and m-

Commerce: A Comparative StudyCommerce: A Comparative Study

Azene Zenebe, Ant Ozok and Anthony F. NorcioDepartment of Information Systems

University of Maryland Baltimore County (UMBC)Baltimore, MD 21250 USA

OutlineOutline

• Introduction– m-commerce verse e-commerce– Personalized recommendations

services (PRS)• System Framework• recommender systems of Amazon and MovieLens

• Comparison – Factors for comparison– Requirement analysis for PRS for

mobile users and devices

• Conclusion & Future research

IntroductionIntroduction• E-commerce verse m-commerce• Challenges in m-commerce (Ghinea &

Angelides, 2004; Turban, King, Lee, & Viehland, 2004; Nielsen, Molich, Snyder, & Farrell, 2001 )

– limited data or query input capability– limited display capability (2-2.5’), resolution– limited processing speed and memory – customer confidence is still low to cell

phone transactions– limited data transmission capability speeds – low battery power of devices – customer confidence is still low

A summary of comparison between e-A summary of comparison between e-commerce and m-commercecommerce and m-commerce

Factor E-Commerce M-Commerce

Technology Device PC Smartphones, Pagers, PDAs, Cell phones

Operating System Windows, Unix, Linux Symbian (EPOC), PalmOS, Pocket PC, proprietary platforms.

Common Communication protocols in m-commerce are

Web’s Hyper Text Transfer Protocol (HTTP)

Wireless Application Protocol (WAP) and DoCoMo”s (Japan) proprietary protocol

Programming and presentation Standards

HTML, XML, JavaScript, Java, etc.

HTML, WML, HDML, i-Mode, Java support

Browser Microsoft Explorer, Netscape

Phone.com UP Browser, Nokia browser, MS Mobile Explorer and other micro-browsers

Bearer Networks TCP/IP & Fixed Wired-line Internet

GSM, GSM/GPRS, TDMA, CDMA, CDPD, paging, Wireless Fidelity (Wi-Fi) networks

Services Personalized Recommendation Well Developed Not Well Developed as e-commerce except a few location-based systems ???; Begins via wired Internet

Accessibility At desktop, workstation, etc.

Ubiquitous: Any time and anywhere

Customer Usage Motivation if they have good reasons or not

Only if they have good reasons

Usability relatively good number of studies

very few studies

Personalized Recommender Systems - Personalized Recommender Systems -

FrameworkFramework What is a Personalized RS?

•matches a customer’s interest, preference, etc. & the products’ attributes •Recommends products or services to customers tailored to their preferences

Personalized Recommender Personalized Recommender Systems - ExamplesSystems - Examples

• e-commerce:– Amazon’s personalized

recommendations that recommends books, DVDs, etc., and

– MovieLens (Sarwar, Karypis, Konstan, & Riedl, 2000) which is a movie recommender system.• Interested reader can refer (Herlocker,

Konstan, Terveen, & Riedl, 2004; Schafer, J, & Riedl, 2001)

Personalized Recommender Personalized Recommender Systems - ExamplesSystems - Examples

• m-commerce:– Amazon Anywhere for Palm PDAs

and WAP devices– Research systems:

• PocketLens (Miller, Knostant, & Riedl, 2004)

• MovieLens Unplugged (Miller, Albert, Lam, Knostant, & Riedl, 2003)

Personalized Recommender Personalized Recommender Systems – Current StatusSystems – Current Status

• Highly successful in e-commerce

• M-commerce?– No personalized recommendation

service for cell phones users in Amazon for digital access

– MovieLens are also not yet fully adapted to mobile access

• Challenges in m-commerce (why not matured?)

ComparisonComparison• Goal

– Elicit additional requirements to adapt the technology developed & advanced in e-commerce RS to m-commerce RS

• Factors/Components– Customer/user, product and service

model– Recommender engine/algorithms– User interface (I/O and interaction)– Confidence and uncertainty model– Acceptance/Trust

Customer & Product ModelCustomer & Product Model

• Facts/assumptions about a customer:– personal facets; behavioral facets;

cognitive facets

– contextual facets-include physical location, past interaction, hardware and software available, tasks, and other users in the environment

• Representation of Products’ information• m-commerce:

– the contextual facets are more essential for effective and useful recommendation decisions

– Concise and easy way of representation of product

I/O and Interaction I/O and Interaction • Input

– individual user's implicit navigation– explicit ratings– purchase history and keywords – comments from community

• M-commerce– initially customers have to sign in wired web– location information needs to be gathered

using devices like GPS– less opportunity for gathering data during

interaction • MovieLens Unplugged (Miller et al., 2003)

attempts to provide a link on the mobile device, later found it to be rarely used.

I/O and Interaction I/O and Interaction • Output

– Customers need as much information as possible about a product or service • to get movie synopsis or reviews on movies• To present images, clips, etc. of products• explanations of how those

recommendations are generated

• M-commerce– Is it feasible to display in effective ways all

these outputs in mobile devices’ display? – optimal number of items to be displayed is

limited usually in range 1 to 5, • e.g. 4 items in MovieLens Unplugged

compared to 10 to 20 items in e-commerce

Methods and Algorithms Methods and Algorithms • Approaches and steps used for

– identifying and generating information and assumptions about customers,

– recommendations • Content-based or action-based

• Amazon Eyes and eBay Personal Shopper (Schafer et al., 2001)

• Collaborative Filtering (CF) • User – user CF; Item – item CF

– Amazon Your Recommendations – Amazon Customers who Bought

• Hybrid • CF - performed offline using a dedicated

server

Methods and Algorithms Methods and Algorithms

• Algorithms of e-commerce need to be adapted using the input, process and output requirements of mobile users and mobile devices– need to support localization for

location-specific recommendations– need to support for updating customer

model, and for generating recommender on fly during customer-system interaction

Confidence/Uncertainty and Confidence/Uncertainty and ExplanationExplanation

• Refers to degree of doubt associated in making recommendations for users – the incompleteness, imprecision, vagueness,

randomness or ambiguity

• Confidence/uncertainty information – level of confidence in user and product

model estimates, about the results of inference or reasoning, and in the recommendations

• Explanation on how are the recommendation obtained?– creating an accurate mental model of the

recommender system and its process

Confidence/UncertaintyConfidence/Uncertainty• Uncertainty originates from during:

– representing interest using crisp values; – representing the product attributes: genre– expressing true relationship among the

products as well as users’ preference to products

• Proposed a Methodology for PRS using Fuzzy and Possibility theory - fuzzy set membership function– to represent and handle uncertainty that

exists in product attributes (e.g. movie genre), user attributes (e.g. ratings) and their relationship in recommender systems.

Results of Evaluation Results of Evaluation • Simulated Movie Recommender System• Empirical evaluation:

– Datasets from MovieLens and IMDb– Compared to best reported results

• Results:– Faster

• nearly 1/10 seconds to infer a customer’s interest for a movie (model time)

• nearly 1/5 seconds to recommend a movie (recommendation time)

– Higher precision (increase by 141%),– 3 to 5 recommendations verse 10– require a few (5 to 10) initial ratings (model

size) from a customer verse 10 to 20

ConclusionConclusion

• Most important dimensions/components

• More similarities in the components

• Additional requirements for m-commerce

• Using fuzzy set and possibility theory for handling uncertainty in e-commerce showed a great potential for m-commerce

Future ResearchFuture Research

• Implement an actual recommender system to e-commerce and m-commerce customers

• Usability study– input and output interfaces of the

different mobile devices– Usefulness of explanation and

confidence information– Trust

www.umbc.edu

Appendix IAppendix I

• FTMax-best and FTMin-worst from Fuzzy Theoretic Approach• CMMax-best and CMMin-worst results from conventional approach

  P R F1

CMMin 0.220 0.131 0.120

CMMax 0.220 0.271 0.240

FTMin 0.509 0.199 0.239

FTMax 0.527 0.284 0.316