1. Intelligent Database Systems Lab 國立雲林科技大學 National Yunlin University of Science and Technology 1 Feature selection for text categorization on imbalanced data Advisor : Dr. Hsu Presenter : Zih-Hui Lin Author :Zhaohui Zheng, Xiaoyun Wu, Rohini Srihari

2. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 2  Motivation  Objective  Introduction  Feature selection framework  Experimental setup  Primary results and analysis  Conclusions  Future work Outline

3. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 3 Motivation  Feature selection using one-sided metrics selects the features most indicative of membership only  Feature selection using two-sided metrics implicitly combines the features most indicative of membership (e.g. positive features) and non-membership (e.g. negative features) by ignoring the signs of features.

4. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 4 Objective  We investigate the usefulness of explicit control of that combination within a proposed feature selection framework.  Our experiments show both great potential and actual merits of explicitly combining positive and negative features in a nearly optimal fashion according to the imbalanced data.

5. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 5 Introduction  One choice in the feature selection policy is whether to rule out all negative features.  Others believe that negative features are numerous, given the imbalanced data set, and quite valuable in practical experience.

6. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 6 Introduction (cont.)  negative features are useful because their presence in a document highly indicates its non-relevance.  They help to confidently reject non-relevant documents.  When deprived of negative features, the performance of all feature selection metrics degrades, which indicates negative features are essential to high quality classification.

7. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 7 Introduction (cont.)  neither one-sided nor two-sided metrics themselves allow control of the combination.  The focus in this paper is to answer the following three questions with empirical evidence: ─ How sub-optimal are two sided metrics? ─ To what extent can the performance be improved by better combination of positive and negative features? ─ How can the optimal combination be learned in practice?

8. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 8 Feature selection metrics  Correlation Coefficient (CC), and Odds Ratio (OR) are one-sided metrics which select the features most indicative of membership for a category only,  IG and CHI are two-sided metrics, which consider the features most indicative of either membership (e.g. positive features) or non- membership (e.g. negative features).

9. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 9 Feature selection metrics (cont.) positive negative t 陳水扁 t 無「陳水扁」 Ci: 政治類 PositiveNegative Ci: 非政治類 NegativePositive

10. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 10 Feature selection metrics (cont.)  Information gain (IG) ─ Measures the number of bits of information obtained for category prediction by knowing the presence or absence of a term in a document.  Chi-square (CHI) ─ measures the lack of independence between a term t and a category c i and can be compared to the chi-square distribution with one degree of freedom to judge extremeness

11. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 11 Feature selection metrics (cont.)  Correlation coefficient (CC) ─ Correlation coefficient of a word t with a category ci  Odds ratio (OR) ─ the odds of the word occurring in the positive class normalized by that of the negative class.

12. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 12 Feature selection metrics (cont.)  IG and CHI are two-sided, whose values are non-negative.  CC and OR are one-sided metrics, whose positive and negative values correspond to the positive and negative features respectively  We can easily obtain that the sign for a one- sided metric, e.g. CC or OR, is sign (AD-BC). A positiveCnegative B negative D positive

13. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 13 Feature selection metrics (cont.)  A one-sided metric could be converted to its two- sided counterpart by ignoring the sign, while a two- sided metric could be converted to its one-sided counterpart by recovering the sign.  We propose the two-sided counterpart of OR, namely OR square, and the one-sided counterpart of IG, namely signed IG as follows.

14. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 14 The imbalanced data problem  The imbalanced data problem occurs when the training examples are unevenly distributed among different classes.  there is an overwhelming number of non-relevant training documents especially when there is a large collection of categories with each assigned to a small number of documents,  This problem presents a particular challenge to classification algorithms, which can achieve high accuracy by simply classifying every example as negative.

15. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 15 The imbalanced data problem ( cont.)  The impact of imbalanced data problem on the standard feature selection can be illustrated as follows, which primarily answers the first question of “how sub-optimal are two side metric”: ─ How to confidently reject the non-relevant documents is important in that case. ─ given a two-sided metric, the values of positive features are not necessarily comparable with those of negative features.

16. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 16 The imbalanced data problem ( cont.) ─ TP,FP, FN and TN denote true positives, false positives, false negatives and true negatives respectively. Positive features have more effect on TP and FN, negative features have more effect on TN and FP. feature selection using a two-sided metric combines the positive and negative features so as to optimize the accuracy, which is defined to be F1 has been widely used in information retrieval, which is: Finally, some performance measures themselves,

17. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 17 Feature selection framework – General formulation  For each category c i : ─ l is the size of feature set ─ 0 < l 1 <= 1 is the key parameter of the framework to be set. ─ The function =ﮒ ( ． ; c i ) should be defined so that the larger = ﮒ(t; c i ) is, the more likely the term t belongs to the category c i.

18. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 18 Feature selection framework – General formulation (cont.)  Obviously, one-sided metrics like SIG, CC, and OR can serve as such functions. ─ we can easily obtain: SIG(t; c i ) = - SIG(t; c i ); CC(t; c i ) = - CC(t; c i ); OR(t; c i ) = - OR(t; c i );

19. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 19 Feature selection framework – General formulation (cont.)  the second step can be rewritten as:  the framework combines ─ the l 1 terms with largest ﮒ =( ． ; ci) ─ the l 2 = l - l 1 terms with smallest = ﮒ ( ． ; ci).

20. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 20 Feature selection framework – two special cases  The standard feature selection methods generally fall into one of the following two groups: ─ select the positive features only using one-sided metrics, e.g. SIG, CC, and OR. For convenience, we will use CC as the representative of this group. ─ implicitly combine the positive and negative features using two-sided metrics, e.g. IG, CHI, and ORS. CHI will be chosen to represent this group.

21. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 21 Feature selection framework – two special cases (cont.)  The positive subset of F i is  ─

22. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 22 Feature selection framework – Optimization  The feature selection framework facilitates the control on explicit combination of the positive and negative features through the parameter l 1 =l (size ratio).  How to optimize the size ratio? ─ Ideal scenario use the training data to learn different models per category according to different size ratios ranging from 0 to 1 and select the ratio having best performance, ─ Practical scenario we tried in this paper is to empirically select the size ratio per category having best performance on the training set.

23. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 23 Feature selection framework – Optimization (cont.)  The efficient implementation of optimization within the framework is as follows: ─ select l positive features with greatest ﮒ (t; ci) in a decreasing order. ─ select l negative features with smallest ﮒ (t; ci) in an increasing order. ─ empirically choose the size ratio l 1 / l such that the feature set constructed by combining the first l 1 ; 0 < l 1 < l; positive features and the first l - l 1 negative features has the optimal performance.

24. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 24 Experimental setup – data collection  We conduct experiments on standard text categorization data using two popular classifiers: ─ Naiive Bayes (NB) and logistic regression (LR).  Data collection ─ Reuters-21578 (ModApte split) is used as our data collection, ─ contains 90 categories, with 7769 training documents and 3019 test documents. ─ After filtering out all numbers, stop-words and words occurring less than 3 times, we have 9484 indexing words in the vocabulary. ─ Words in document titles are treated same as in document ─ body.

25. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 25 Experimental setup – Classifiers  A NB score between a document d and the category c i can be calculated as: ─ f j is the feature appearing in the document P(c i ) and P(c i ) represent prior probabilities of relevant and non-relevant respectively and P(f j | c i ) and P(f j | c i ) are conditional probabilities estimated with Laplacian smoothing.

26. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 26  Logistic regression tries to model the conditional probability as: ─  The optimization problem for LR is to minimize: ─ d i is the ith training example, w is the weight vector, y i Є {-1, 1} is the label associated with d i, λ is an appropriately chosen regularization parameter Experimental setup – Classifiers (cont.)

27. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 27 Experimental setup – performance measure  To measure the performance, we use both precision (p) and recall (r) in their combined from F1 :. To remain compatible with other results, the F1 value at Break Even Point (BEP).  BEP is defined to be the point where precision equals recall. ─ It corresponds to the minimum of |FP – FN| in practice.  We report both micro and macro-averaged BEP F1. ─ The micro-averaged F1 is largely dependent on the most common categories while the rare categories influence macro-averaged F1.

28. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 28 Experimental setup – performance measure

29. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 29 Primary results and analysis – ideal scenario  Micro F1(BEP) values for NB with the feature selection methods at different sizes of features over the 58 categories: 3rd-60th. The micro-averaged F1(BEP) for NB without feature selection (using all 9484 features ) is.641

30. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 30 Primary results and analysis – ideal scenario (cont.)  As table1 but for macro-averaged F1(BEP). The macro-averaged F1(BEP) for NB without feature selection is.483

31. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 31 Primary results and analysis – ideal scenario (cont.)  Micro-averaged F1(BEP) values for LR with the feature selection methods at different sizes of features over the 58 categories. The micro-averaged F1(BEP) for LR without feature selection is.766

32. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 32 Primary results and analysis – ideal scenario (cont.)  As table 3, but for macro-averaged F1(BEP). The macro-averaged F1(BEP) for LR without feature selection is.676

33. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 33 Primary results and analysis – ideal scenario (cont.)  Figure 2: Size ratios implicitly decided by using two-sided metrics: IG, CHI and ORS respectively (58 categories:3rd- 60th, feature size = 50)  confirms that feature selection using a two-sided metric is similar to its one-sided counterpart (size ratio = 1) when the feature size is small.

34. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 34 Primary results and analysis – ideal scenario (cont.)  Figure 3: BEP F1 for test over the first two and last two categories (out of 58 categories) at different l 1 =l values (NB, ﮒ= CC, feature size = 50). The optimal size ratios for money-fx, grain, dmk and lumber are 0.95, 0.95, 0.1 and 0.2 respectively

35. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 35 Primary results and analysis – ideal scenario (cont.)  Figure 4: Optimal size ratios of iSIG, iCC and iOR (NB, 58 categories:3rd- 60th, feature size = 50)

36. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 36 Primary results and analysis – ideal scenario (cont.)  Figure 5: As gure 4, but for LR

37. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 37 Primary results and analysis – practical scenario (cont.)

38. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 38 Primary results and analysis – Additional results

39. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 39 Primary results and analysis – Additional results (cont.)

40. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 40 Primary results and analysis – Additional results (cont.)

41. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 41 Conclusions  a novel feature selection framework was presented, in which the positive and negative features are separately selected and explicitly combined.  We explored three special cases of the framework: ─ 1. consider the positive features only by using one-sided metrics; ─ 2. implicitly combine the positive and negative features by using two-sided metrics; ─ 3. combine the two kinds of features explicitly and choose the size ratio empirically such that optimal performance is obtained.

42. Intelligent Database Systems Lab N.Y.U.S.T. I. M. 42 Conclusions (cont.)  Implicitly combining positive and negative features using two-sided metrics is not necessarily optimal, especially on imbalanced data.  A judicious combination shows great potential and practical merits.  A good feature selection method should take into consideration the data set, performance measure, and classification methods.  Feature selection can significantly improve the performance of both naiive Bayes and regularized logistic regression on imbalanced data.