Presentation is loading. Please wait.

Presentation is loading. Please wait.

From Feature Construction, to Simple but Effective Modeling, to Domain Transfer Wei Fan IBM T.J.Watson

Similar presentations

Presentation on theme: "From Feature Construction, to Simple but Effective Modeling, to Domain Transfer Wei Fan IBM T.J.Watson"— Presentation transcript:

1 From Feature Construction, to Simple but Effective Modeling, to Domain Transfer Wei Fan IBM T.J.Watson

2 Feature Vector Most data mining and machine learning model assume the following structured data: (x 1, x 2,..., x k ) -> y where xis are independent variable y is dependent variable. y drawn from discrete set: classification y drawn from continuous variable: regression

3 Frequent Pattern-Based Feature Construction Data not in the pre-defined feature vectors Transactions Biological sequence Graph database Frequent pattern is a good candidate for discriminative features So, how to mine them?

4 FP: Sub-graph A discovered pattern NSC 4960 NSC NSC NSC NSC (example borrowed from George Karypis presentation)

5 Computational Issues Measured by its frequency or support. E.g. frequent subgraphs with sup 10% Cannot enumerate sup = 10% without first enumerating all patterns > 10%. Random sampling not work since it is not exhaustive. NP hard problem

6 1. Mine frequent patterns (>sup) Frequent Patterns DataSet mine Mined Discriminative Patterns select 2. Select most discriminative patterns; 3. Represent data in the feature space using such patterns; 4. Build classification models. F1 F2 F4 Data Data Data Data ……… represent | Petal.Width< 1.75 setosa versicolor virginica Petal.Length< 2.45 Any classifiers you can name NN DT SVM LR Conventional Procedure Feature Construction followed by Selection Two-Step Batch Method

7 Two Problems Mine step combinatorial explosion Frequent Patterns DataSe t mine 1. exponential explosion 2. patterns not considered if minsupport isnt small enough

8 Two Problems Select step Issue of discriminative power Frequent Patterns Mined Discriminative Patterns select 3. InfoGain against the complete dataset, NOT on subset of examples 4. Correlation not directly evaluated on their joint predictability

9 Direct Mining & Selection via Model- based Search Tree Basic Flow Mined Discriminative Patterns Compact set of highly discriminative patterns Divide-and-Conquer Based Frequent Pattern Mining 2 Mine & Select P: 20% Y 3 Y 6 Y + Y Y 4 N Few Data N N + N 5 N Mine & Select P:20% 7 N … … Y dataset 1 Mine & Select P: 20% Most discriminative F based on IG Feature Miner Classifier Global Support: 10*20%/10000 =0.02%

10 Analyses (I) 1. Scalability of pattern enumeration Upper bound (Theorem 1): Scale down ratio: 2. Bound on number of returned features

11 Analyses (II) 3. Subspace pattern selection Original set: Subset: 4. Non-overfitting 5. Optimality under exhaustive search

12 Experimental Studies: Itemset Mining (I) Scalability Comparison Datasets#Pat using MbT sup Ratio (MbT #Pat / #Pat using MbT sup) Adult % Chess + ~0% Hypo % Sick % Sonar % 2 Mine & Select P: 20% Y 3 Y + Y Y Few Data N + N dataset 1 Mine & Select P: 20% Most discriminative F based on IG Global Support: 10*20%/10000 =0.02% 6 Y 5 N Mine & Select P:20% 7 N 4 N 2 Y 3 Y + Y Y Few Data N + N dataset 1 Mine & Select P: 20% Most discriminative F based on IG Global Support: 10*20%/10000 =0.02% 6 Y 5 N Mine & Select P:20% 7 N 4 N

13 Experimental Studies: Itemset Mining (II) Accuracy of Mined Itemsets 4 Wins 1 loss But, much smaller number of patterns

14 Experimental Studies: Itemset Mining (III) Convergence

15 Experimental Studies: Graph Mining (I) 9 NCI anti-cancer screen datasets The PubChem Project. URL: Active (Positive) class : around 1% - 8.3% 2 AIDS anti-viral screen datasets URL: H1: CM+CA – 3.5% H2: CA – 1%

16 Experimental Studies: Graph Mining (II) Scalability 2 Mine & Select P: 20% Y 3 Y + Y Y Few Data N + N dataset 1 Mine & Select P: 20% Most discriminative F based on IG Global Support: 10*20%/10000 =0.02% 6 Y 5 N Mine & Select P:20% 7 N 4 N 2 Y 3 Y + Y Y Few Data N + N dataset 1 Mine & Select P: 20% Most discriminative F based on IG Global Support: 10*20%/10000 =0.02% 6 Y 5 N Mine & Select P:20% 7 N 4 N

17 Experimental Studies: Graph Mining (III) AUC and Accuracy AUC 11 Wins 10 Wins 1 Loss

18 AUC of MbT, DT MbT VS Benchmarks Experimental Studies: Graph Mining (IV) 7 Wins, 4 losses

19 Summary Model-based Search Tree Integrated feature mining and construction. Dynamic support Can mine extremely small support patterns Both a feature construction and a classifier Not limited to one type of frequent pattern: plug-play Experiment Results Itemset Mining Graph Mining New: Found a DNA sequence not previously reported but can be explained in biology. Code and dataset available for download

20 Even the true distribution is unknown, still assume that the data is generated by some known function. Estimate parameters inside the function via training data CV on the training data Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Some unknown distribution How to train models? There probably will always be mistakes unless: 1.The chosen model indeed generates the distribution 2.Data is sufficient to estimate those parameters But how about, you dont know which to choose or use the wrong one? List of methods: Logistic Regression Probit models Naïve Bayes Kernel Methods Linear Regression RBF Mixture models After structure is prefixed, learning becomes optimization to minimize errors: quadratic loss exponential loss slack variables

21 How to train models II Not quite sure the exact function, but use a family of free-form functions given some preference criteria. There probably will always be mistakes unless: the training data is sufficiently large. free form function/criteria is appropriate. List of methods: Decision Trees RIPPER rule learner CBA: association rule clustering-based methods … … Preference criteria Simplest hypothesis that fits the data is the best. Heuristics: info gain, gini index, Kearns-Mansour, etc pruning: MDL pruning, reduced error-pruning, cost-based pruning. Truth: none of purity check functions guarantee accuracy on unseen test data, it only tries to build a smaller model

22 Can Data Speak for Themselves? Make no assumption about the true model, neither parametric form nor free form. Encode the data in some rather neutral representations: Think of it like encoding numbers in computers binary representation. Always cannot represent some numbers, but overall accurate enough. Main challenge: Avoid rote learning: do not remember all the details Generalization Evenly representing numbers – Evenly encoding the data.

23 Potential Advantages If the accuracy is quite good, then Method is quite automatic and easy to use No Brainer – DM can be everybodys tool.

24 Encoding Data for Major Problems Classification: Given a set of labeled data items, such as, (amt, merchant category, outstanding balance, date/time, ……,) and the label is whether it is a fraud or non-fraud. Label: set of discrete values classifier: predict if a transaction is a fraud or non-fraud. Probability Estimation: Similar to the above setting: estimate the probability that a transaction is a fraud. Difference: no truth is given, i.e., no true probability Regression: Given a set of valued data items, such as (zipcode, capital gain, education, …), interested value is annual gross income. Target value: continuous values. Several other on-going problems

25 Encoding Data in Decision Trees Think of each tree as a way to encode the training data. Why tree? a decision tree records some common characteristic of the data, but not every piece of trivial detail Obviously, each tree encodes the data differently. Subjective criteria that prefers some encodings than others are always adhoc. Do not prefer anything then – just do it randomly Minimizes the difference by multiple encodings, and then average them.

26 Random Decision Tree to Encode Data - classification, regression, probability estimation At each node, an un-used feature is chosen randomly A discrete feature is un-used if it has never been chosen previously on a given decision path starting from the root to the current node. A continuous feature can be chosen multiple times on the same decision path, but each time a different threshold value is chosen

27 Continued We stop when one of the following happens: A node becomes too small (<= 3 examples). Or the total height of the tree exceeds some limits: Such as the total number of features.

28 Illustration of RDT B1: {0,1} B2: {0,1} B3: continuous B2: {0,1} B3: continuous B2: {0,1} B3: continuous B3: continous Random threshold 0.3 Random threshold 0.6 B1 chosen randomly B2 chosen randomly B3 chosen randomly

29 Classification | Petal.Width< 1.75 setosa 50/0/0 versicolor 0/49/5 virginica 0/1/45 Petal.Length< 2.45 P( setosa |x,θ) = 0 P( versicolor |x,θ) = 49/54 P( virginica |x,θ) = 5/54

30 Regression | Petal.Width< 1.75 setosa Height=10in versicolor Height=15 in virginica Height=12in Petal.Length< in average value of all examples In this leaf node

31 Prediction Simply Averaging over multiple trees

32 Potential Advantage Training can be very efficient. Particularly true for very large datasets. No cross-validation based estimation of parameters for some parametric methods. Natural multi-class probability. Natural multi-label classification and probability estimation. Imposes very little about the structures of the model.

33 Reasons The true distribution P(y|X) is never known. Is it an elephant? Every random tree is not a random guess of this P(y|X). Their structure is, but not the node statistics Every random tree is consistent with the training data. Each tree is quite strong, not weak. In other words, if the distribution is the same, each random tree itself is a rather decent model.

34 Expected Error Reduction Proven that for quadratic loss, such as: for probability estimation: ( P(y|X) – P(y|X, θ) ) 2 regression problems ( y – f(x) ) 2 General theorem: the expected quadratic loss of RDT (and any other model averaging) is less than any combined model chosen at random.

35 Theorem Summary

36 Number of trees Sampling theory: The random decision tree can be thought as sampling from a large (infinite when continuous features exist) population of trees. Unless the data is highly skewed, 30 to 50 gives pretty good estimate with reasonably small variance. In most cases, 10 are usually enough.

37 Variance Reduction

38 Optimal Decision Boundary from Tony Lius thesis (supervised by Kai Ming Ting)

39 RDT looks like the optimal boundary

40 Regression Decision Boundary (GUIDE) Properties Broken and Discontinuous Some points are far from truth Some wrong ups and downs

41 RDT Computed Function Properties Smooth and Continuous Close to true function All ups and downs caught

42 Hidden Variable

43 Limitation of GUIDE Need to decide grouping variables and independent variables. A non-trivial task. If all variables are categorical, GUIDE becomes a single CART regression tree. Strong assumption and greedy-based search. Sometimes, can lead to very unexpected results.

44 It grows like …

45 ICDM08 Cup Crown Winner Nuclear ban monitoring RDT based approach is the highest award winner.

46 Ozone Level Prediction (ICDM06 Best Application Paper) Daily summary maps of two datasets from Texas Commission on Environmental Quality (TCEQ)

47 SVM: 1-hr criteria CV

48 AdaBoost: 1-hr criteria CV

49 SVM: 8-hr criteria CV

50 AdaBoost: 8-hr criteria CV

51 Other Applications Credit Card Fraud Detection Late and Default Payment Prediction Intrusion Detection Semi Conductor Process Control Trading anomaly detection

52 Conclusion Imposing a particular form of model may not be a good idea to train highly-accurate models for general purpose of DM. It may not even be efficient for some forms of models. RDT has been show to solve all three major problems in data mining, classification, probability estimation and regressions, simply, efficiently and accurately. When physical truth is unknown, RDT is highly recommended Code and dataset is available for download.

53 Standard Supervised Learning New York Times training (labeled) test (unlabeled) Classifier 85.5% New York Times

54 In Reality…… New York Times training (labeled) test (unlabeled) Classifier 64.1% New York Times Labeled data not available! Reuters

55 Domain Difference Performance Drop traintest NYT New York Times Classifier 85.5% Reuters NYT ReutersNew York Times Classifier 64.1% ideal setting realistic setting

56 A Synthetic Example Training (have conflicting concepts) Test Partially overlapping

57 Goal Source Domain Target Domain Source Domain Source Domain To unify knowledge that are consistent with the test domain from multiple source domains (models)

58 Summary Transfer from one or multiple source domains Target domain has no labeled examples Do not need to re-train Rely on base models trained from each domain The base models are not necessarily developed for transfer learning applications

59 Locally Weighted Ensemble M1M1 M2M2 MkMk …… Training set 1 Test example x Training set 2 Training set k …… x-feature value y-class label Training set

60 Modified Bayesian Model Averaging M1M1 M2M2 MkMk …… Test set Bayesian Model Averaging M1M1 M2M2 MkMk …… Test set Modified for Transfer Learning

61 Global versus Local Weights …… xy …100001… M1M … M2M … wgwg 0.3 … wlwl … wgwg 0.7 … wlwl … Locally weighting scheme Weight of each model is computed per example Weights are determined according to models performance on the test set, not training set Training

62 Synthetic Example Revisited Training (have conflicting concepts) Test Partially overlapping M1M1 M2M2 M1M1 M2M2

63 Optimal Local Weights C1C1 C2C2 Test example x Higher Weight Optimal weights Solution to a regression problem w1w1 w2w2 = H wf

64 Approximate Optimal Weights How to approximate the optimal weights M should be assigned a higher weight at x if P(y|M,x) is closer to the true P(y|x) Have some labeled examples in the target domain Use these examples to compute weights None of the examples in the target domain are labeled Need to make some assumptions about the relationship between feature values and class labels Optimal weights Impossible to get since f is unknown!

65 Clustering-Manifold Assumption Test examples that are closer in feature space are more likely to share the same class label.

66 Graph-based Heuristics Graph-based weights approximation Map the structures of models onto test domain Clustering Structure M1M1 M2M2 weight on x

67 Graph-based Heuristics Local weights calculation Weight of a model is proportional to the similarity between its neighborhood graph and the clustering structure around x. Higher Weight

68 Local Structure Based Adjustment Why adjustment is needed? It is possible that no models structures are similar to the clustering structure at x Simply means that the training information are conflicting with the true target distribution at x Clustering Structure M1M1 M2M2 Error

69 Local Structure Based Adjustment How to adjust? Check if is below a threshold Ignore the training information and propagate the labels of neighbors in the test set to x Clustering Structure M1M1 M2M2

70 Verify the Assumption Need to check the validity of this assumption Still, P(y|x) is unknown How to choose the appropriate clustering algorithm Findings from real data sets This property is usually determined by the nature of the task Positive cases: Document categorization Negative cases: Sentiment classification Could validate this assumption on the training set

71 Algorithm Check Assumption Neighborhood Graph Construction Model Weight Computation Weight Adjustment

72 Data Sets Different applications Synthetic data sets Spam filtering: public collection personal inboxes (u01, u02, u03) (ECML/PKDD 2006) Text classification: same top-level classification problems with different sub-fields in the training and test sets (Newsgroup, Reuters) Intrusion detection data: different types of intrusions in training and test sets.

73 Baseline Methods One source domain: single models Winnow (WNN), Logistic Regression (LR), Support Vector Machine (SVM) Transductive SVM (TSVM) Multiple source domains: SVM on each of the domains TSVM on each of the domains Merge all source domains into one: ALL SVM, TSVM Simple averaging ensemble: SMA Locally weighted ensemble without local structure based adjustment: pLWE Locally weighted ensemble: LWE Implementation Package: Classification: SNoW, BBR, LibSVM, SVMlight Clustering: CLUTO package

74 Performance Measure Prediction Accuracy 0-1 loss: accuracy Squared loss: mean squared error Area Under ROC Curve (AUC) Tradeoff between true positive rate and false positive rate Should be 1 ideally

75 A Synthetic Example Training (have conflicting concepts) Test Partially overlapping

76 Experiments on Synthetic Data

77 Spam Filtering Problems Training set: public s Test set: personal s from three users: U00, U01, U02 pLWE LR SVM SMA TSVM WNN LWE pLWE LR SVM SMA TSVM WNN LWE Accuracy MSE

78 20 Newsgroup C vs S R vs T R vs S C vs T C vs R S vs T


80 Reuters pLWE LR SVM SMA TSVM WNN LWE pLWE LR SVM SMA TSVM WNN LWE Accuracy MSE Problems Orgs vs People (O vs Pe) Orgs vs Places (O vs Pl) People vs Places (Pe vs Pl)

81 Intrusion Detection Problems (Normal vs Intrusions) Normal vs R2L (1) Normal vs Probing (2) Normal vs DOS (3) Tasks > 3 (DOS) > 2 (Probing) > 1 (R2L)

82 Conclusions Locally weighted ensemble framework transfer useful knowledge from multiple source domains Graph-based heuristics to compute weights Make the framework practical and effective Code and Dataset available for download

83 More information or For code, dataset and papers

Download ppt "From Feature Construction, to Simple but Effective Modeling, to Domain Transfer Wei Fan IBM T.J.Watson"

Similar presentations

Ads by Google