eddf54908061ccf7d62c0b50d47f97ec.ppt

- Количество слайдов: 66

Зарегистрируйтесь, чтобы просмотреть полный документ!

РЕГИСТРАЦИЯ
Visualization and Data Mining techniques By- Group number- 14 Chidroop Madhavarapu(105644921) Deepanshu Sandhuria(105595184) Data Mining CSE 634 Prof. Anita Wasilewska 1

References n http: //coblitz. codeen. org: 3125/citeseer. ist. psu. edu/cache/papers/cs/10335/ftp: z. Szftp. cs. umn. edu z. Szdeptz. Szusersz. Szkumarz. Szdatavis. pdf/ganesh 96 visual. pdf n http: //www. ailab. si/blaz/predavanja/ozp/gradivo/2002 -Keim-Visualization%20 in%20 DMIEEE%20 Trans%20 Vis. pdf n http: //www. geocities. com/anand_palm/ n http: //citeseer. ist. psu. edu/cache/papers/cs/27216/http: z. Szwwwusers. cs. umn. eduz. Szz. Cz 7 Ectluz. Sz. Paper. Talk. Filez. Szits 02. pdf/shekhar 02 cubeview. pdf n http: //www. cs. umn. edu/Research/shashi-group/Book/sdb-chap 1. pdf n http: //www. cs. umn. edu/research/shashi-group/alan_planb. pdf n http: //coblitz. codeen. org: 3125/citeseer. ist. psu. edu/cache/papers/cs/27637/http: z. Szwwwusers. cs. umn. eduz. Szz. Cz 7 Epushengz. Szpubz. Szkdd 2001 z. Szkdd. pdf/shekhar 01 detecting. pdf 2

Motivation Visualization for Data Mining • Huge amounts of information • Limited display capacity of output devices Visual Data Mining (VDM) is a new approach for exploring very large data sets, combining traditional mining methods and information visualization techniques. 3

Why Visual Data Mining 4

Why Visual Data Mining 5

VDM Approach VDM takes advantage of both, n n The power of automatic calculations, and The capabilities of human processing. n Human perception offers phenomenal abilities to extract structures from pictures. 6

Levels of VDM n No or very limited integration Corresponds to the application of either traditional information visualization or automated data mining methods. n n Loose integration n Visualization and automated mining methods are applied sequentially. The result of one step can be used as input for another step. Full integration n Automated mining and visualization methods applied in parallel. n Combination of the results. 7

Methods of Data Visualization Different methods are available for visualization of data based on type of data Data can be n Univariate n Bivariate n Multivariate 8

Univariate data n Measurement of single quantitative variable n Characterize distribution n Represented using following methods n Histogram n Pie Chart 9

Histogram 10

Pie Chart 11

Bivariate Data n Constitutes of paired samples of two quantitative variables n Variables are related n Represented using following methods n Scatter plots n Line graphs 12

Scatter plots 13

Line graphs 14

Multivariate Data n Multi dimensional representation of multivariate data n Represented using following methods n Icon based methods n Pixel based methods n Dynamic parallel coordinate system 15

Icon based Methods 16

Pixel Based Methods n Approach: n n n Each attribute value is represented by one colored pixel (the value ranges of the attributes are mapped to a fixed color map). The values of each attribute are presented in separate sub windows. Examples: n Dense Pixel Displays 17

Dense Pixel Display Approach: n n Each attribute value is represented by one colored pixel (the value ranges of the attributes are mapped to a fixed color map). Different attributes are presented in separate sub windows. 18

Visual Data Mining: Framework and Algorithm Development Ganesh, M. , Han, E. H. , Kumar, V. , Shekar, S. , & Srivastava, J. (1996). Working Paper. Twin Cities, MN: University of Minnesota, Twin Cities Campus. 19

References n http: //coblitz. codeen. org: 3125/citeseer. ist. psu. edu/cache/papers/cs/10335/ftp : z. Szftp. cs. umn. eduz. Szdeptz. Szusersz. Szkumarz. Szdatavis. pdf/ganesh 96 visual. pdf n http: //www. ailab. si/blaz/predavanja/ozp/gradivo/2002 -Keim. Visualization%20 in%20 DM-IEEE%20 Trans%20 Vis. pdf n http: //www. geocities. com/anand_palm/ 20

Abstract n n n VDM refers to the use of visualization techniques in Data Mining process to n Evaluate n Monitor n Guide This paper provides a framework for VDM via the loose coupling of databases and visualization systems. The paper applies VDM towards designing new algorithms that can learn decision trees by manually refining some of the decisions made by well known algorithms such as C 4. 5. 21

Components of VQLBCI n The three major components of VQLBCI are Visual Representations, Computations and Events. 22

Visual Development of Algorithms n Most interesting use of visual data mining is the development of new insights and algorithms. n The figure below shows the ER diagram for learning classification decision trees. n n This model allows the user to monitor the quality and impact of decisions made by the learning procedure. Learning procedure can be refined interactively via a visual interface. 23

ER diagram for the search space of decision tree learning algorithm 24

General Framework n n n Learning a classification decision tree from a training data set can be regarded as a process of searching for the best decision tree that meets user-provided goal constraints. The problem space of this search process consists of Model Candidates, Model Candidate Generator and Model Constraints. Many existing classification-learning algorithms like C 4. 5 and CDP fit nicely within this search framework. New learning algorithms that fit user’s requirements can be developed by defining the components of the problem space. 25

General Framework n n Model Candidate corresponds to the partial classification decision tree. Each node of the decision tree is a Model Atom Search process is the process of finding a final model candidate such that it meets user goal specifications. Model Candidate Generator transforms the current model candidate into a new model candidate by selecting one model atom to expand from the expandable leaf model atoms. Model Constraints (used by Model Candidate Generator) provide controls and boundaries to the search space. 26

Search Process 27

Acceptability Constraint n n Model Constraints consist of Acceptability constraints, Expandability constraints and a Data-Entropy calculation function. Acceptability constraint predicate specifies when a model candidate is acceptable and thus allows search process to stop. EX: n n n A 1) Total no of expandable leaf model atoms = 0. A 2) Overall error rate of the model candidate <= acceptable error rate. A 3) Total number of model atoms in the model candidate>= maximal allowable tree size. A 1 is used in C 4. 5 and CDP 28

Expandability Constraint n An Expandability constraint predicate specifies whether a leaf model atom is expandable or not. EX: n n C 4. 5 uses E 1 and E 2 CDP uses E 2 and E 3 29

Traversal Strategy n Traversal strategy ranks expandable leaf model atoms based on the model atom attributes. EX: n n n Increasing order of depth Decreasing order of depth Orders based on other model atom attributes. 30

Steps in Visual Algorithm Development n n No single algorithm is the best all the time, performance is highly data dependent. By changing different predicates of model constraints, users can construct new classificationlearning algorithm. This enables users to find an algorithm that works the best on a given data set. Two algorithms are developed : BF based on Best First search idea and CDP+ which is a modification of CDP 31

BF n n n This algorithm is based on the Best-First search idea. For Acceptability criteria, it includes A 1 and A 2 with a user specified acceptable error rate. The Traversal strategy chosen is T 3 In Best-First, expandable leaf model atoms are ranked according to the decreasing order of the number of misclassified training cases. (local error rate * size of subset training data set) The traversal strategy will expand a model atom that has the most misclassified training cases, thus reducing the overall error rate the most. 32

CDP + n CDP+ is a modification of CDP n CDP has dynamic pruning using expandability constraint E 3. n Here, the depth is modified according to the size of the training data set of the model atom. n n We set B is the branching factor of the decision tree, t is the size of training data set belonging to model atom, T is the whole training data set. 33

Comparison of different classification learning algorithms 34

Experiment n n The new BF and CDP+ algorithms are compared with the C 4. 5 and CDP algorithms. Various metrics are selected to compare the efficiency, accuracy and size of final decision trees of the classification algorithm. The generation efficiency of the nodes is measured in terms of the total number of nodes generated. To compare accuracy of the various algorithms, the mean classification error on the test data sets have been computed. 35

Classification error for 10 data sets 36

Nodes generated for 10 data sets 37

Final decision tree size 38

Results/Conclusion n n CDP has accuracy comparable to C 4. 5 while generating considerably fewer nodes. CDP+ outperformed CDP in error rate and number of nodes generated. Considering all performance metrics together, CDP+ is the best overall algorithm. Considering classification accuracy alone, C 4. 5 P is the winner. 39

Conclusion n Different datasets require different algorithms for best results. Diverse user requirements put different constraints on the final decision tree. The experiment shows that Interactive Visual Data Mining Framework can help find the most suitable algorithm for a given data set and group of user requirements. 40

Data Mining for Selective Visualization of Large Spatial Datasets Proceedings of 14 th IEEE International Conference on Tools with Artificial Intelligence (ICTAI'02), 2002. Washington (November 2002), DC, USA, Shashi Shekhar, Chang-Tien Lu, Pusheng Zhang, Rulin Liu Computer Science & Engineering Department University of Minnesota 41

References n http: //citeseer. ist. psu. edu/cache/papers/cs/27216/http: z. Szwwwusers. cs. umn. eduz. Szz. Cz 7 Ectluz. Sz. Paper. Talk. Filez. Szits 02. pdf/shekhar 02 cubeview. pdf n http: //www. cs. umn. edu/Research/shashi-group/Book/sdb-chap 1. pdf n http: //www. cs. umn. edu/research/shashi-group/alan_planb. pdf n http: //coblitz. codeen. org: 3125/citeseer. ist. psu. edu/cache/papers/cs/2 7637/http: z. Szwwwusers. cs. umn. eduz. Szz. Cz 7 Epushengz. Szpubz. Szkdd 2001 z. Szkdd. pdf/shek har 01 detecting. pdf 42

Basic Terminology n Spatial databases n n Spatial mining n n Mining of spatial databases Spatial datawarehouse n n Alphanumeric data + geographical cordinates Contains geographical data Spatial outliers n Observations that appear to be inconsistent with the remainder of that set of data 43

Spatial Cluster 44

Contribution n n Propose and implement the Cube. View visualization system General data cube operations Built on the concept of spatial data warehouse to support data mining and data visualization Efficient and scalable spatial outlier detection algorithms 45

Challenges in spatial data mining n Classical data mining - numbers and categories. Spatial data – n n more complex and extended objects such as points, lines and polygons. Second, classical data mining works with explicit inputs, whereas spatial predicates and attributes are often implicit. Third, classical data mining treats each input independently of other inputs. 46

Application Domain n n The Traffic Management Center - Minnesota Department of Transportation (MNDOT) has a database to archive sensor network. Sensor network includes n n n about nine hundred stations each of which contains one to four loop detector Measurement of Volume and occupancy. n n Volume is # vehicles passing through station in 5 minute interval Occupancy is percentage of time station is occupied with vehicles 47

Basic Concepts n n n Spatial Data Warehouse Spatial Data Mining Spatial Outliers Detection 48

Spatial Data Warehouse n n n Employs data cube structure Outputs - albums of maps. Traffic data warehouse n n Measures - volume and occupancy Dimensions - time and space. 49

Spatial Data Mining n n n Process of discovering interesting and useful but implicit spatial patterns. key goal is to partially ‘automate’ knowledge discovery Search for “nuggets” of information embedded in very large quantities of spatial data. 50

Spatial Outliers Detection n Suspiciously deviating observations Local instability Each Station n n Spatial attributes – time, space Non spatial attributes – volume, occupancy 51

Basic Structure – Cube. View 52

Cube. View Visualization System n Each node in cube – a visualization style n n n S - Traffic volume of station at all times. TTD – Time of the day TDW – Day of the week STTD – Daily traffic volume of each station TTD TDWS– Traffic volume at each station at different times on different days 53

Dimension Lattice 54

Cube. View Visualization System 55

Cube. View Visualization System 56

Cube. View Visualization System 57

Data Mining Algorithms for Visualization n Problem Definition n n n n Given a spatial graph G ={ S , E } S - s 1, s 2, s 3, s 4……. . E – edges (neighborhood of stations) f ( x ) - attribute value for a data record N ( x )- fixed cardinality set of neighbors of x ) - Average attribute value of x neighbors S( x ) - difference of the attribute value of each data object and the average attribute value of neighbors. 58

Data Mining Algorithms for Visualization n Problem Definition cont… n S( x ) - difference of the attribute value of each data object and the average attribute value of neighbors. Test for detecting an outlier n confidence level threshold θ n n n 59

Data Mining Algorithms for Visualization n Few points n n First, the neighborhood can be selected based on a fixed cardinality or a fixed graph distance or a fixed Euclidean distance. Second, the choice of neighborhood aggregate function can be mean, variance, or auto-correlation. Third, the choice for comparing a location with its neighbors can be either just a number or a vector of attribute values. Finally, the statistic for the base distribution can be selected as normal distribution. 60

Data Mining Algorithms for Visualization n Algorithms n Test Parameters Computation(TPC) Algorithm n Route Outlier Detection(ROD) Algorithm 61

Data Mining Algorithms for Visualization 62

Data Mining Algorithms for Visualization 63

Data Mining Algorithms for Visualization 64

Software n http: //www. cs. umn. edu/research/shashigroup/vis/traffic_volumemap 2. htm n http: //www. cs. umn. edu/research/shashigroup/vis/Data. Cube. htm 65

Visualization and Data Mining techniques Thank you!!!! 66

eddf54908061ccf7d62c0b50d47f97ec.ppt

- Количество слайдов: 66