35.txt

This complete matlab for neural network

ng personalization to a Web site.
The basic approach to handling personalized and other dynamic pages is to se
rve the base HTML page from the server, while caching embedded image files. 
This doesn't require new technology and is how proxy serving typically works
 on the Web today. For example, IBM WebSphere Performance Pack, installed at
 Deutsche Telekom as a proxy server, caches the embedded images of popular p
ages. Since image files tend to outnumber HTML pages, reasonable proxy hit r
ates are still possible. The drawback is that, even if personalized HTML pag
es represent significantly less than 50% of the requests and bytes requested
 from a Web site, the CPU overhead for generating the personalized pages can
 still be significant and can affect the throughput of the Web site. Where S
SL is used for secured pages, avoid encrypting and decrypting image GIF file
s to improve performance by increasing reuse of cached images.
Other strategies to maximize performance
IBM is developing technologies and techniques for reducing the overhead of s
erving dynamic data, such as personalized data (see Resources 3, 4, and 5). 
Figure 3 shows a multi-tiered Web site and the caching and personalization t
echniques suitable for each Web site component. The caching levels show that
 performance is maximized when cache hits occur close to the browsers. Simil
arly, more complex and sophisticated personalization techniques are introduc
ed as you move through the different tiers and get closer to the database la
yer. For example, at the ISP and router levels, rule-based and simple filter
ing may offer sufficient personalization capabilities for a relatively small
 investment of effort. When more is needed or wanted, more complex technique
s can be implemented. Note that all the techniques could be employed at the 
application server, while just the most complex techniques are in use at the
 database server. When data mining is needed to develop business intelligenc
e and offer highly sophisticated personalization, the processing occurs at t
he database layer.
Figure 3. Overview of Web site with personalization and intelligent content 
distribution
When database changes arrive rapidly, as they do during a sporting event or 
a trading day, a trigger monitor can be implemented to watch changes (see Re
sources 7, 8, and 9). Changes can then be propagated forward from the databa
se server to the browser. When a certain number of changes, or certain chang
es, occur, the trigger monitor rebuilds the affected Web pages and distribut
es updated pages to the caches. This technique ensures data is current and p
erformance is maximized, making it appropriate for use with dynamic personal
ized pages and ideally reducing a page's only dynamic content to, for exampl
e, personalized account information. The trigger monitor is the key technolo
gy at the heart of a robust implementation of intelligent content distributi
on.
IBM's sports Web sites efficiently create and serve dynamic Web data, includ
ing personalized data. These Web sites use new techniques for caching dynami
c data as well as for creating complex dynamic Web pages from simpler fragme
nts. Figure 4 depicts the evolution of IBM's sporting event Web sites. Curre
nt sites use an integrated cache to serve dynamic pages. An externalized API
 enables the server to load and invalidate pages as needed. A trigger monito
r keeps caches current while content is changing rapidly.
Figure 4. Evolution of IBM sporting event sites
Sites can benefit from content caches, as well as the trigger monitor. A con
tent cache can build certain types of personalized pages from fragments stor
ed in cache. For example, for the 2000 Olympics Web site, advertisements nee
d to be based on the country of origin of the client. This could be done bas
ed on the IP address of the client and advertising fragments for each countr
y. More generally, this could be done by partitioning clients into groups, w
ith each group being served pages for a specific URL, where some of the page
 fragments are personalized based on the client group. The client group coul
d be identified in various ways, for example, by source IP address, URL exte
nsions identifying the client or group, or cookies. Then, based on the clien
t group, the content cache combines specific page fragments, sometimes calle
d tagged content, to compose the personalized page. A tagged content design 
facilitates managing and reusing content fragments (see Resource 2). The lev
el of personalization possible with these intelligent content distribution t
echniques covers a significant subset of personalization requirements. Howev
er, the techniques at the cache are still more limited than personalization 
achievable at the server because available information about the client is l
imited, and performance requirements limit the degree of personalization.
The trigger monitor keeps track of the scores and statistics that arrive rap
idly during sporting events. As the database updates arrive, the trigger mon
itor keeps track of changes, rebuilds the affected Web pages, and distribute
s updated pages to the caches, assuring they are kept current and the person
alized Web sites are updated as well.
You can reduce the overhead of personalization by reducing the degree of per
sonalization. For example, instead of creating pages specialized for each in
dividual client, you could create sets of pages specialized (tagged) to grou
ps of visitors. This could significantly reduce the total number of pages an
d allow reuse of some pages, thus increasing the utility of caching. This re
duced level of personalization can be provided at a content cache.
You can also vary the degree of personalization based on server load. When s
ervers are heavily loaded, the amount of personalization could be minimized.
 For example, for personalized advertisements, when the server is highly loa
ded, random advertisements can be included in the page, while at lower loads
 the advertisements can be highly targeted. You could combine this technique
 with content caches, where a lower degree of personalization pages could be
 served from the cache when the server is highly loaded, while deep personal
ization could be done at the server, when server load permits.
Personalized Web pages can be assembled at the client if it is enabled with 
Java. Some sites even provide Java to the client to optimize personalization
 and performance.
Among the multiple recommendation engines, one uses a new content-based coll
aborative filtering approach, where the object content is captured in making
 collaborative filtering. This technique achieves the advantages of both con
tent-based and collaborative filtering approaches. The content-based collabo
rative filtering technique is applicable to both product and document recomm
endations.
Because collecting visitor information can be an expensive effort and also a
ffect the performance, you should be able to measure its effectiveness. The 
issue is not only what to recommend, but also when and how. The personalized
 recommendation engines deal with the issue of what to recommend given a set
 of alternatives, but a more sophisticated application would decide when to 
invoke the recommendation engine and how to apply it, for example, whether t
o send the customer an e-mail or e-coupon, or add a Web link on the personal
ized Web page.
Personalizing your site based on site classification
IBM's IT experts have been working with customers to analyze many of the wor
ld's largest Internet and intranet sites, including IBM's own, to determine 
which attributes affect scalability and to help customers implement scalable
 Web sites. IBM has determined that:
Large sites are distinguished primarily by workload pattern
Based on workload patterns, Web sites can generally be classified into five 
types: publish/subscribe, online shopping, customer self-service, trading, a
nd business-to-business
Scaling techniques must be selected and applied based on workload pattern
If you're unfamiliar with the Web site classifications, refer to Appendix B:
 Summary of high-volume Web site classifications.
In the same way that a workload pattern suggests appropriate scaling techniq
ues, it can also suggest the most effective personalization techniques. Whil
e it's possible to implement any or all of the techniques at each site type,
 some techniques require significant effort and may degrade performance; you
 may or may not need that level of investment.
For each type of Web site, Figure 5 shows the personalization techniques tha
t would be most effective. Note, for example, that rule-based techniques app
ly to all site types except publish/subscribe, while all techniques apply to
 the self-service and business-to-business sites. After you determine which 
type of site you have, use this table to identify the personalization techni
ques you should consider. Note that at least one effective, relatively simpl
e technique is suggested for each type. From another perspective, consider A
mazon.com, one of the most successful and "smartest" online shopping sites (
see Resource 2). Given the volume and attributes of Amazon's objects, conten
t-based filtering would require excessive effort and so would not be conside
red effective.
Figure 5. Personalization techniques mapped to workload patterns
 Site type
Technique Publish/ subscribe Online shopping Self-service Trading Business-t
o- business
Rule based  X X X X
Simple filtering X X X X X
Content-based filtering X  X X X
Collaborative filtering  X X  X
Summary
Quite simply, personalization has become a required, expected feature of an 
e-business Web site. The presence and quality of site personalization determ
ines whether visitors find your site attractive and return to it with an int
ention to buy. The real question is not whether to personalize, but how and 
how much, and how to implement personalization while maximizing performance,
 which can be as important as the business effectiveness of the techniques y
ou choose. In this paper, you've learned about current personalization techn
iques and the significance of intelligent content distribution and other tec
hniques to maximize site performance. During site design, be sure to conside
r your workload pattern and to insist that your personalization and caching 
strategies be considered early and in relation to each other.
IBM has products and services that can help you get started today and positi
on your site for enhancements as your business rules and requirements change
 and additional personalization techniques are developed.
Appendix A: More on filtering techniques
Content-based filtering
Content-based filtering works by analyzing the content of the objects to for
m a representation of the visitor's interests. Generally, the analysis needs
 to identify a set of key attributes for each object and then fill in the at
tribute values.
Recommending video purchases is an example of content-based filtering. The e
xample below uses seven attributes to analyze video content: action, drama, 
sex, violence, suspense, humor, and offbeat. The rating goes from 0 to 10 in
dicating the intensity. For example, a violence rating of 10 means extreme v
iolence and 0 means no violence.
Figure 6. An example of content-based filtering
Video / Attribute Action Drama Humor Sex Violence Suspense Offbeat
(A) Silence of the Lambs  7 3 1 9 10
(B) Seven 5 5 1 2 10 9 5
(C) Cape Fear 5 7 4 5 9 9 3
(D) Casablanca 2 10 5 0 1 8
(E) Waterboy 4 2 6 3 4 3 1
(F) L.A. Confidential 8 9 6 6 9 9 6
(G) West Side Story 3 5 4 0 1 3 1
Using a concept known as "Euclidean distance" or nearest neighbor, content-b
ased filtering analyzes the ratings to determine for any one video, which ot
her video has the closest ratings and could be recommended to a visitor who 
ordered the first video. For example, Silence of the Lambs could be found to
 come closest in content to Seven, in which case Seven could be a candidate 
to recommend to customers interested in Silence of the Lambs.
Collaborative filtering
Collaborative filtering collects visitors' opinions on a set of items, using
 either explicit or implicit ratings, to form like-minded peer groups and th
en learns from the peer groups to predict a particular visitor's interest in
 an item. Instead of finding objects similar to those a visitor liked in the
 past, as in content-based filtering, collaborative filtering develops recom
mendations by finding visitors with similar tastes.
Below is an example of collaborative filtering. Assume each person can rate 
a video from 1 to 7, where 7 means strongly like, 4 is neutral, and 1 means 
strongly dislike. Videos A through G represent the seven videos shown in the
 previous table.
Figure 7. An example of collaborative filtering
Video / Visitor A B C D E F G
Adam 7   6 2   2
Bill 7     1   2 5
Jennifer 4 2         2
John 6 2 7   7
Mary 2 7   7
Rose 1 7       6
Susan 2 6   7     6
For ease of illustration, we again use a nearest neighbor measure of closene
ss. When measuring the distance between two persons, only videos both have r
ated are considered. For example, when considering the distance between Adam
 and John, only the ratings on videos A and C are considered. Adam's close p
eers are Bill and John. For Adam, we can recommend video E based on John's l
iking. The point to note here is that the content of video E, Waterboy, can 
be quite different from the content of videos rated highly by Adam. Although
 with similar content to videos A and C, video B, Seven, will not be recomme
nded to Adam, because his peer, John, does not like it. However, content-bas
ed collaborative filtering will recommend video B to Adam, based solely on t
he fact that its content is similar to videos A and C, which are liked by Ad
am.
Comparing the techniques
Rule-based techniques and simple filtering offer significant personalization
 capabilities for an investment of effort relatively smaller than content-ba
sed and collaborative filtering.
Content-based filtering is most suitable when the objects are easily analyze
d by computer and the visitor's decision about object suitability is not sub
jective. For some objects, such as the videos, analyzing content cannot be a
utomated today, and the effort to identify attributes and evaluate each obje
ct can be considerable and require specific knowledge or skills. Recommendat
ions are limited to objects related to those the visitor has tried, with 

--
※ 来源:．南京大学小百合站 bbs.nju.edu.cn．[FROM: 61.144.60.26]