Scalable Web Usage Mining and Soft Computing Approaches for High Performance Intelligent Web Recomme

1
Scalable Web Usage Mining and Soft Computing
Approaches for High Performance Intelligent Web
Recommender Systems

• Olfa Nasraoui
• Research Assistants Cesar Cardona, Carlos Rojas,
  Fabio Gonzalez, Elizabeth Leon, Chris Petenes,
  Mrudula Pavuluri
• Dept. of Electrical Computer Engineering
• The University of Memphis
• E-mail onasraou_at_memphis.edu

Research sponsored by National Science Foundation
CAREER Award NSF-IIS 0133948
2
Outline

• Web Usage Mining
• Background Evolutionary Computation
• Mining Web Profiles with Hierarchical
  Unsupervised Niche Clustering (H-UNC)
• H-UNC Web Mining Experimental Results
• Need for scalability Background Artificial
  Immune Systems
• Scalable Web Usage Mining based on the Immune
  System Metaphor
• Scalable Web Usage Mining Preliminary Results
• Tying it All Together An Automated Real Time
  Recommendation System
• Conclusions and Future Prospects

3
Web Personalization

• WWW Personalization Tailor users interaction
  with Web info space based on information about
  the user
• eg recommend items/links based on prior
  ratings/visits
• Manually entered profiles are subjective, static,
  not always available, and raised privacy concerns
• Alternative Extract profiles based on all users
  
• access patterns Mass profiling
• ?anonymous profiles

4
Web Personalization System
Web Usage Mining
P R O F I L E S
Recomm-endation Engine
5
Web Mining Data Types Challenges

• Mining the Web Data
• Content Web pages
• Usage Web access log files
• Structure Link structure of Web pages
• Challenges
• Huge, semi/unstructured, highly dynamic data
• data corrupted with noise (not all info. is
  relevant)

6
Web Usage Mining Framework
7
Knowledge Discovery Process For Web Usage Mining

• Source of Data Web Clickstreams ? Web Log Files
• Goal Extract interesting user profiles by
  categorizing user sessions into groups or
  clusters
• Complete KDD process
• Preprocessing selecting and cleaning data
• Data Mining (Learning Phase)
• A model for session data (bag of URLs visited)
• Similarity assessment
• Clustering algorithm to categorize sessions
• Derivation and Interpretation of results
• Computing profiles
• Evaluating results

8
Different Ways to Mine Web User Profiles
Possible Solutions Problems

• Mining Web Usage Data using Clustering
• K Means (Shahabi et al. 1997) (problems w/
  Euclidean distance, feature vector
  representation, sparsity, noise, not knowing the
  of clusters, assumes user profiles do not
  overlap, sensitive to initialization local
  optima, etc)
• Relational Clustering (same as above, except that
  they use a distance/relation matrix instead of
  vector representation, huge memory and
  computation requirements, etc)
• Fuzzy Relational Clustering (Nasraoui et al.
  1999) (same as above, but allows overlapping
  clusters, etc)
• Robust relational clustering (Nasraoui et al.
  1999) (same as above, but can handle noise in
  data!)

9
Different Ways to Mine Web User Profiles
Possible Solutions Problems

• Evolutionary techniques can avoid the feature
  vector representation dilemma (but appropriate
  coding is required), flexible allow any
  similarity measure, any subjective fitness
  criterion measure, better global search
  (population based)
• Hierarchical Unsupervised Niche Clustering or
  HUNC (Nasraoui Krishnapuram, 2001) based on
  Darwinian evolution metaphor and niche
  speciation, handles noise, unknown of clusters,
  reliable w.r.t initialization
• Immune Based Clustering (Nasraoui et al., 2002)
  based on immune system metaphor a microcosm of
  evolution
• Need a Scalable Immune Based Clustering linear
  complexity even for huge clickstream data that
  cannot fit in main memory dynamic online
  learning of evolving user profiles (Nasraoui et
  al., 2003)

10
Different Ways to Mine Web User Profiles
Possible Solutions Problems

• Other Approaches (Web Session Transaction)
• Frequent Itemset Association Rule Mining
• Very sensitive to required parameters support
  and confidence thresholds
• Either too few profiles or too many including
  spurious profiles
• Exorbitant computational complexity for low
  support thresholds
• Association Rule Mining, followed by Hypergraph
  Partition based Clustering
• Above all drawbacks of hypergraph partition
  clustering (huge complexity)

11
Outline

• Web Usage Mining
• Background Evolutionary Computation
• Mining Web Profiles with Hierarchical
  Unsupervised Niche Clustering (H-UNC)
• H-UNC Web Mining Experimental Results
• Need for scalability Background Artificial
  Immune Systems
• Scalable Web Usage Mining based on the Immune
  System Metaphor
• Scalable Web Usage Mining Preliminary Results
• Tying it All Together An Automated Real Time
  Recommendation System
• Conclusions and Future Prospects

12
Background Genetic Algorithms (Inspired by
Nature Darwinian Evolution)

• Genetic Algorithms (GAs) Evolve a population of
  individuals/solutions using selection, crossover,
  mutation as in nature

Operators
Operators
13
Outline

• Web Usage Mining
• Background Evolutionary Computation
• Mining Web Profiles with Hierarchical
  Unsupervised Niche Clustering (H-UNC)
• H-UNC Web Mining Experimental Results
• Need for scalability Background Artificial
  Immune Systems
• Scalable Web Usage Mining based on the Immune
  System Metaphor
• Scalable Web Usage Mining Preliminary Results
• Tying it All Together An Automated Real Time
  Recommendation System
• Conclusions and Future Prospects

14
Mining Web Profiles with Hierarchical
Unsupervised Niche Clustering Step 1 Data
Preprocessing

• Access log Record of all URLs accessed by users
  on a Web site
• Log entry access time, IP address, URL viewed,
  etc.
• ___________________________________________
• 171148 141.225.195.29 GET /graphics/griffin.jpg
  200
• 171148 141.225.195.29 GET /people/faculty/nasrao
  ui.html 200
• ___________________________________________
• Map NU URLs on site to indices
• User session vector s(i) temporally compact
  sequence of Web accesses by a user

15
Step 2 Clustering Sessions

• Clustering Dividing unlabeled data into groups
• Unsupervised clustering when number of
  categories unknown
• Robust clustering when data contains noise
  outliers
• Genetic Clustering Can deal with
  non-differentiable objective functions/similarity
  measures.

16
Unsupervised Niche Clustering (UNC)

• Representation binary chromosome strings (one
  substring per feature)
• Deterministic Crowding Selection Children
  replace closest parent if they have better
  fitness.
• Density fitness measure
• Robust weight

17
Evolution Example
18
Adaptation to Web Mining Hierarchical UNC (HUNC)

• Encode binary session vectors
• Perform UNC in hierarchical mode (HUNC)
• ? Fast Multi-resolution profiling Vary no. of
  levels (L)
• Start by applying UNC to entire data set w/ small
  pop. size (L 1)
• Focus on each cluster recursively Reapply UNC on
  data subset assigned to each cluster to extract
  more clusters at higher resolution (L gt 1)
• Repeat until cluster size or scale become too
  small

19
Outline

• Web Usage Mining
• Background Evolutionary Computation
• Mining Web Profiles with Hierarchical
  Unsupervised Niche Clustering (H-UNC)
• H-UNC Web Mining Experimental Results
• Need for scalability Background Artificial
  Immune Systems
• Scalable Web Usage Mining based on the Immune
  System Metaphor
• Scalable Web Usage Mining Preliminary Results
• Tying it All Together An Automated Real Time
  Recommendation System
• Conclusions and Future Prospects

20
Web Mining Experimental Results with HUNC
21
(No Transcript)
22
Level 2 examples

• General outside visitor Profiles 1 and 3
• Prospective students Profiles 2 and 4
• Insiders (students) Profiles 6, 7, etc

23
Main Site of Univ. Of Missouri
Profile 16 Example of discovering associations
w/out any prior knowledge of content
24
Outline

• Web Usage Mining
• Background Evolutionary Computation
• Mining Web Profiles with Hierarchical
  Unsupervised Niche Clustering (H-UNC)
• H-UNC Web Mining Experimental Results
• Need for scalability Background Artificial
  Immune Systems
• Scalable Web Usage Mining based on the Immune
  System Metaphor
• Scalable Web Usage Mining Preliminary Results
• Tying it All Together An Automated Real Time
  Recommendation System
• Conclusions and Future Prospects

25
Towards Scalability Dynamic Web Mining

• Typically, data mining has to be completely
  re-applied periodically and offline on newly
  generated Web server logs in order to keep the
  discovered knowledge up to date.
• An intelligent Web mining system should be able
  to continuously learn evolving usage trends
  without ungraceful stoppages, reconfigurations,
  or restarting from scratch
• Need to make scalable to handle huge data sets
  given limited amounts of main memory
• We may view the problem as clustering data
  streams huge flux of data with very limited
  memory to store it need single pass learning.
• Applicable both to usage and content (text) data

26
Another Evolutionary System in Nature The
Immune System

• immune system parallel and distributed adaptive
  system w/ tremendous potential in many
  intelligent computing applications.
• Protects our bodies from foreign pathogens
  (viruses/bacteria)
• Innate Immune System (initial, limited, ex skin,
  tears, etc)
• Acquired Immune System (Learns how to respond to
  NEW threats adaptively through an evolutionary
  process)
• Primary immune response
• First response to invading pathogens
• Secondary immune response
• Encountering similar pathogen a second time
• Remember past encounters
• Faster and stronger response than primary response

27
Points of Strength of The Immune System

• Recognition (Anomaly detection, Noise tolerance)
• Robustness (Noise tolerance)
• Feature extraction
• Diversity (can face an entire repertoire of
  foreign invaders)
• Reinforcement learning
• Memory (remembers past encounters basis for
  vaccine)
• Distributed Detection (no single central system)
• Multi-layered (defense mechanisms at multiple
  levels)
• Adaptive (Self-regulated)

28
Learning in the Immune System

• Main purpose of the immune system recognize all
  cells (or molecules) within the body and
  categorize those cells as self or non-self.
• Non-self cells are further categorized in order
  to stimulate an appropriate type of defensive
  mechanism.
• The immune system learns through micro-scale
  evolution to distinguish between foreign antigens
  (e.g., bacteria, viruses, etc.) and the body's
  own cells or molecules.

29
B-Cells

• Through a process of recognition and stimulation,
  B-Cells will clone and mutate to produce a
  diverse set of antibodies adapted to different
  antigens
• B-Cells secrete antibodies that can bind to
  specific antigens and destroy their host invading
  agent through a KILL, SUICIDE, or INGEST signal.
• B-Cells antibody also can bind to antibodies on
  other B-Cells, hence sending a STIMULATE or
  SUPPRESS signal ? Network!

30
Immune Recognition

• Immune recognition based on complementarity
  between binding region of the receptor and a
  portion
• of the antigen called epitope.
• B-cell Antibodies present a single type of
  receptor, antigens might present several
  epitopes.
• This means that different antibodies can
  recognize a single antigen.
• Binding between B-cells and antigens is NOT exact
  (soft error-tolerant binding)

31
Artificial Immune Systems (AIS) Recent History

• Based on the Immune Network theory (Jerne, 1973)
• The system consists of a network of B-Cells
• Antigens represent data
• B-Cells represent clusters
• B-cells form a Network by interacting with each
  other through stimulation and suppression (to
  form memory of past antigens)
• Exponential explosion in B-Cell population!!!
• Huge immune network bottleneck against
  scalability!
• Unscalable (time and memory!)

32
Outline

• Web Usage Mining
• Background Evolutionary Computation
• Mining Web Profiles with Hierarchical
  Unsupervised Niche Clustering (H-UNC)
• H-UNC Web Mining Experimental Results
• Need for scalability Background Artificial
  Immune Systems
• Scalable Web Usage Mining based on the Immune
  System Metaphor
• Scalable Web Usage Mining Preliminary Results
• Tying it All Together An Automated Real Time
  Recommendation System
• Conclusions and Future Prospects

33
General Architecture of Proposed Approach

• Information in Immune network
• Stimulation (competition memory)
• Age (old vs. new)
• Co-stimulation /
• suppression
• ? network interactions
• - Outliers (based on activation)

1-Pass Adaptive Immune Learning
Evolving data ?
?
Evolving Immune Network (compressed into
subnetworks)
34
Model for Artificial Immune Cell

• Antigens data
• B-Cells clusters or patterns to be
  learned/extracted
• Dynamic environment antigens are presented to
  the immune network one at a time, with the
  stimulation and scale measures re-updated with
  each presentation.
• antigen index, j monotonically increasing with
  time antigens are presented in the following
  chronological order x1, x2, , xN,
• Dynamic Weighted B-cell (D-W-B-Cell)
• Represents neighborhood modeled by activation
  function robust weight/membership function

35
Model for Artificial Immune Cell

• Each D-W-B-Cell is allowed to have is own zone of
  influence with size/scale si
• D-W-B-Cells dynamically adapt their influence
  zones/hence stimulation level in a strife for
  survival.
• Activation Weight function dynamically adapts to
  evolving data (time decay)
• Outliers are easily detected through weak
  activations
• Flexible for different attributes types
  (numerical, categorical, etc)
• D-W-B-cells are cloned in proportion to their
  stimulation levels.

36
Incremental Update Eqs. Network Interactions

• Stimulation (fitness)
• Influence Zone scale
• Stimulation and Suppression from neighboring
  B-cells
• Positive suppression (competition), but no
  stimulation good population control and no
  redundancy, but no memory immune network will
  forget past encounters.
• Positive stimulation, but no suppression good
  memory but no competition proliferation of
  D-W-B-cell population maximum redundancy.
• Natural tradeoff between redundancy/memory and
  competition/reduced costs.

37
Divide and Conquer Compress Immune Network into
K Subnetworks

• Assuming that the network is divided into roughly
  K equal sized subnetworks (ex w/ 2-3 iter. Of K
  Means),
• Then the number of internal interactions in an
  immune network of NB D-W-B-cells, can drop from
  (NB)2 in the uncompressed network, to (NB /K)2
  intra-subnetwork interactions and (K-1)
  inter-subnetwork interactions in the compressed
  immune network.
• Can approach linear complexity as K ? (NB)1/2
• Significant savings in computation

38
Internal and External Immune Interactions
Before After Compression
39

• Memory Constraints

Start/ Reset
Activates subNet ?
Yes
No
Outlier?

• Domain Knowledge Constraints

Yes
B-cells gt MaxLimit?
Secondary storage
No
ImmuNet Stats Visualization
40
Immune Based Learning of Web profiles

• The Web server plays the role of the human body,
  and the incoming requests play the role of
  antigens that need to be detected
• The input data is similar to web log data (a
  record of all files/URLs accessed by users on a
  Web site)
• The data is pre-processed to produce session
  lists
• A session list Si for user i is a list of URLs
  visited by same user
• In discovery mode, a session is fed to the
  learning system as soon as it is available
• B-celli ith candidate profile
• List of URLs
• Each profile has its own influence zone defined
  by ?i

41
(No Transcript)
42
(No Transcript)
43
(No Transcript)
44
Single Pass Results (locationscale) on a Noisy
dataset presented one at a time in the same order
as clusters
350 samples 1125 samples 1925 samples
all 3200 samples
45
Ability to distinguish between core and outlier
points (wij lt0.001) for noisy data set presented
in the different orders
cluster 1 to cluster 5, cluster 5 to cluster
1, random order

• Blank areas first pts used to start the network
• Most recent noise pts (depending on order) shown
  in black their fate is still uncertain until
  future data confirms it

46
Outline

• Web Usage Mining
• Background Evolutionary Computation
• Mining Web Profiles with Hierarchical
  Unsupervised Niche Clustering (H-UNC)
• H-UNC Web Mining Experimental Results
• Need for scalability Background Artificial
  Immune Systems
• Scalable Web Usage Mining based on the Immune
  System Metaphor
• Scalable Web Usage Mining Preliminary Results
• Tying it All Together An Automated Real Time
  Recommendation System
• Conclusions and Future Prospects

47
Simulation Scenarios for Tracking Evolving Web
Usage Trends

• Scenario 1 We used 20 profiles previously
  discovered using Hierarchical Unsupervised Niche
  Clustering (HUNC) to partition the Web sessions
  into 20 distinct sets of sessions, each one
  assigned to the closest profile. Then we
  presented these sessions to the immune clustering
  algorithm one usage trend at a time from trend 0
  to 19. That is, we first present the sessions
  assigned to trend 0, then the sessions assigned
  to profile 1, , etc.
• Scenario 2 we used the same pre-partitioned
  session data set as the previous scenario, but
  presented the profiles in reverse order from
  trend 19 to 0. That is, we first present the
  sessions assigned to trend 19, then the sessions
  assigned to profile 18, , etc, ending with
  sessions from trend 0.
• Scenario 3 Natural chronological order exactly
  as they were received in real time by the web
  server.

48
Distribution of input sessions

• Trends 5, 9, 13, 14, 15, and 19 appear to be
  weaker and noisier.
• Also trends 6 and 7 emerge late in the 12-day
  access log, while trend 0 weakens in the last days

49
Noise sessions
50
Evaluation Method

• In each scenario, we track the actual composition
  of the B cells in the immune network, i.e., the
  URLs present in the B cell (profile).
• We also track the number of B cells that succeed
  in learning each of the 20 ground truth profiles
  after each session is presented
• by computing an evolving number of hits per
  expected usage trend number of B-cells within
  0.4 radius of the ground truth profile.
• distance is computed as (1 - cosine similarity)2.

51
Hits per usage trend vs. time
52
Hits per usage trend vs. time scenario 1 (from
trend 0 to 19)
53
Hits per usage trend vs. time scenario 2
(reverse order)
54
Hits per usage trend vs. time scenario 3
(chronological)
55
If centroids of compressed sub-networks are
allowed to clone
56
Hits Based on High Precision and Coverage
B-cells form a faithful synopsis of input usage
data in 1 pass.
Compare to Input Data (below)
57
Suitability for Real Time Web Mining

• Single pass over all 1704 Web user sessions
  (non-optimized Java code) lt 7 seconds on 2 GHz
  Pentium IV PC on Linux.
• Ability to learn an unknown number of evolving
  profiles in real time average of 4 milliseconds
  per user session ? suitable for real time
  personalization.
• Old profiles can be handled in a variety of ways
  They may either be discarded, moved to secondary
  storage, or cached for possible re-emergence.
• Even if discarded, older profiles that re-emerge
  later, would be re-learned from scratch.
• Logistics of maintaining old profiles are not
  crucial.
• Used same technique successfully to track learn
  evolving topic categories/clusters in text data

58
Outline

• Web Usage Mining
• Background Evolutionary Computation
• Mining Web Profiles with Hierarchical
  Unsupervised Niche Clustering (H-UNC)
• H-UNC Web Mining Experimental Results
• Need for scalability Background Artificial
  Immune Systems
• Scalable Web Usage Mining based on the Immune
  System Metaphor
• Scalable Web Usage Mining Preliminary Results
• Tying it All Together An Automated Real Time
  Recommendation System
• Conclusions and Future Prospects

59
Recommender Systems Motivations

• The move from traditional commerce ? e-commerce
• Space limited physical inventory ? Huge virtual
  inventory (Information Overload)
• Recommender systems
• Help in navigation (adaptive websites)
• Enhance e-commerce sales (cross-sell, customer
  loyalty, turning browsers into buyers,etc)
• Web request prediction prefetching, caching,
  load balancing (for single websites ISPs)
• E-commerce sites simply cannot survive without
  recommender systems!
• Same for huge web portals (Yahoo!)
• Same for huge digital libraries and Web
  information systems
• On Search engines context awareness modify
  query or order/rank of search results

60
Web Personalization System
Web Usage Mining
P R O F I L E S
Recomm-endation Engine
61
Recommender Systems

• K- Nearest Neighbor based Collaborative
  Filtering Recommend items preferred by K
  nearest neighbors (usually top N items)
• Association Rule based
• Discover association rules
• items_L ? items_R (support, confidence)
• At recommendation time find all rules supported
  by customer (rated items included in items_L)
• Sort rules by confidence
• Recommend first N ranked products
• Challenges Scalability, sparsity, low coverage
  or precision

62
Fuzzy Inference (IF x is A THEN y is B) via fuzzy
relation matrix

• R X ? Y ? 0,1 - encodes strength of
  relationship between x and y
• (x,y) ? mR (x,y)
• R P ? U ? 0,1 - encodes strength of
  relationship between Profile i and URL j
• (Pi,urlj) ? mR (Pi,urlj) Pij

R (profiles)
A mA(x)
B mB(y)
s (session)
ms(i)
R
r (recommendations) mr(j)
similarity
Input membership computation
Fuzzy Inference
Pi (profiles)
63
Fuzzy Recommendation Engine

• Given current session s, infer recommendation r.
  Hence, the following implication
• s ? r.
• Given the Web user profiles discovered by mining
  the Web logs, the relation R is defined as
  follows
• Rik pik
• The input fuzzy set is derived from the current
  session x s by computing the similarity value
  between s defined in (1) and each profile, pi, as
  follows msi ms(i) sim (s, pi)
• inferrence procedure concludes the recommendation
  r as the possibility for URL relevance via the
  following composition
• mrk mr(URLk) ms(i) ? R

t-norm (intersection/AND, e.g. min)
t-conorm (union/OR, e.g. max)
64
(No Transcript)
65
Simulation Experiments

• Given the prediscovered profiles, and a set of
  Web sessions extracted from the same Web log
  file, we treat every complete session as
  ground-truth session.
• For each such ground-truth session, all possible
  subsets of this session consisting of between 1
  and 9 URLs are considered as current test
  subsessions,
• recommendations are generated for each test
  subsession, and
• coverage and precision measures are averaged for
  each subsession size,
• roughly 380,000 separate recommendation tests,
  each tested using 16 different recommendation
  scenarios, corresponding to different
  combinations of input membership generation,
  composition, ...etc,

66
Tested Recommendation Scenarios

• 2 types of similarities cosine or Web session
  (denoted as Cosine and WS respectively in the
  plots).
• 2 different compositions were tested Max-Min and
  Bounded Sum-Min (denoted as MM and BS in the
  plots).
• 2 different types of profiles were tested
• raw profiles generate real similarities (denoted
  as Real Cosine and Real WS in plots), and
• crisp a cuts with a0.2 (denoted as Binary
  Cosine - .2 Thresholded Profile or Binary WS - .2
  Thresholded Profile in plots, and was only tested
  for max-min composition).
• Either raw recommendations or crisp a cuts with
  a0.001, 0.2, and 0.3 (denoted Type of Input
  Similarity - a Thresholded To Bin in plots).

67
Evaluation Measures

• actual completed session, sT treated as
  ground-truth,
• a subset of this session is treated as incomplete
  current sub-session, sj,
• rj fuzzy recommendations
• Then rj rj - sj recommendations obtained
  after omitting all URLs that are part of the
  current subsession.
• Also, sj sT - sj ground truth URLs, not
  including the ones in the current subsession
  being processed for recommendations.
• Precision is given by
• Coverage is given by

68
Coverage Comparison Fuzzy vs. Nearest-Profile
K-NN
69
Precision Fuzzy (better for longer sessions) vs.
Nearest Profile (middle performance) K-NN
70
F1 Measure Fuzzy (better for longer sessions)
vs. Nearest Profile K-NN
71
Discussion of Results Comparison with Nearest
Profile K-NN

• Fuzzy Recommendations have better Precision at
  larger session sizes (starting at 3 to 4 URLs)
• K-NN have highest Precision at small session
  sizes (lt4)
• However Nearest-Profile performs midway between
  k-NN and Fuzzy at small session sizes (lt4)
• Fuzzy Recommendations have superior Coverage
  regardless of session length
• Overall F1 measure Fuzzy Recommendations are
  better when the session size is larger and
  Nearest-Profile performs midway between k-NN and
  Fuzzy at small session sizes (lt4)
• Best to combine Nearest-Profile and Fuzzy and
  alternate between max-min and Bounded-sum
  depending on session length!!!

72
Time and Memory Complexity

• Offline training takes longer with the profile
  based approach.
• However this can be done on a back end computer
  and not on the server, and is therefore an
  offline process that does not affect the
  operation of the web server.
• On the other hand, both fuzzy and nearest profile
  based approaches are extremely fast at
  recommendation time and require a minimal amount
  of main memory to function
• (a mere summary of the previous usage history
  instead of the entire history as in collaborative
  filtering).
• In our simulations, fuzzy recommendations with
  non-optimized Perl script (non-compiled) code
  running on a 2 GHz Pentium 4 Linux PC operated at
  an average of 48 recommendations per second.
• The far more computationally complex K-Nearest
  Neighbor generated recommendations at a leisurely
  2 recommendations per second.

73
Two-Step Recommendation Process based on a
Committee of Profile-Specific URL-Predictor
Neural Networks
74
Average precision, coverage, F1, cosine for the
Two-Step Profile-Specific URL-Predictor
Recommender model
75
Outline

• Web Usage Mining
• Background Evolutionary Computation Artificial
  Immune Systems
• Mining Web Profiles with Hierarchical
  Unsupervised Niche Clustering (H-UNC)
• H-UNC Web Mining Experimental Results
• Scalable Web Usage Mining based on the Immune
  System Metaphor
• Scalable Web Usage Mining Preliminary Results
• Tying it All Together An Automated Real Time
  Recommendation System
• Conclusions and Future Prospects

76
Conclusion And Future Prospects

• Ill-Posed Feature Representation problems,
  Subjective dissimilarity between Web sessions and
  subjective multi-modal fitness functions handled
  well using Evolutionary computation
• Why Evolutionary computation?
• Deals w/ ill-Posed Feature Representation
  problems,
• Handles subjective dissimilarity between Web
  sessions
• Handles subjective multi-modal fitness functions
• Insensitive to initialization
• HUNC first evolutionary based technique for Web
  usage mining
• Why NOT Evolutionary computation?
• Scalability problems
• Immune System Clustering scalable and flexible
  in the face of huge dynamic web usage access
  trends

77
Conclusion And Future Prospects

• Immune System Clustering Also used successfully
  to extract topics/cluster text documents in 1
  pass.
• Also used successfully to cluster and perform
  anomaly detection on kdd-cup99 network data in 1
  pass (unsupervised learning using only the normal
  data with no labels)
• Results in both attack detection false alarms
  superior to best results (kdd-cup winner
  supervised learning trained with data labeled in
  normal all different attacks)
• Completely Automated real-time web
  personalization system is feasible
• Soft computing techniques for intelligent
  recommender systems
• Web usage mining Fuzzy Inference Based better
  for longer sessions (increased uncertainty and
  noise)
• Web usage mining Two Stage Neural Network
  Unorecedented performance, but slow in training

78
Impact on WWW

• Scalable and adaptive Recommendation engine for
  personalization
• Improve search results by taking profile/context
  into account
• Improve design of dynamic Web sites
• Facilitate navigation

79
Some Related Publications

• O. Nasraoui and R. Krishnapuram, and A. Joshi.
  Mining Web Access Logs Using a Relational
  Clustering Algorithm Based on a Robust Estimator,
  8th International World Wide Web Conference,
  Toronto, pp. 40-41, 1999.
• O. Nasraoui, and R. Krishnapuram, A Novel
  Approach to Unsupervised Robust Clustering using
  Genetic Niching, Proc. of the 9th IEEE
  International Conf. on Fuzzy Systems, San
  Antonio, TX, May 2000, pp. 170-175.
• O. Nasraoui and R. Krishnapuram. A New
  Evolutionary Approach to Web Usage and Context
  Sensitive Associations Mining, International
  Journal on Computational Intelligence and
  Applications - Special Issue on Internet
  Intelligent Systems, Vol. 2, No. 3, pp. 339-348,
  Sep. 2002.
• Nasraoui O., Cardona C., Rojas C., and Gonzalez
  F, "TECNO-STREAMS Tracking Evolving Clusters in
  Noisy Data Streams with a Scalable Immune System
  Learning Model", in Proc. of Third IEEE
  International Conference on Data Mining
  (ICDM'03), Melbourne, FL, November 2003.
• Nasraoui O., Petenes C., "Combining Web Usage
  Mining and Fuzzy Inference for Website
  Personalization", in Proc. of WebKDD 2003 KDD
  Workshop on Web mining as a Premise to Effective
  and Intelligent Web Applications, Washington DC,
  August 2003, p. 37.
• Nasraoui O., Gonzalez F., Cardona C., Rojas C.,
  and Dasgupta D., "A Scalable Artificial Immune
  System Model for Dynamic Unsupervised Learning",
  Proc. of the Genetic and Evolutionary Computation
  Conference (GECCO), Chicago, IL, July 2003, p.
  219,
• Nasraoui O., Cardona C., Rojas C., and Gonzalez
  F., "Mining Evolving User Profiles in Noisy Web
  Clickstream Data with a Scalable Immune System
  Clustering Algorithm", in Proc. of WebKDD 2003
  KDD Workshop on Web mining as a Premise to
  Effective and Intelligent Web Applications,
  Washington DC, August 2003, p. 71,