Machine Learning

Data Mining

(A special program to Text Mining)

jdwang@asia.edu.tw, Room:I517, ext:1847

Time: Monday, 1:10pm~4:00pm, Room: I528

OutlinePreCourse

(1) PubMed Articles Download and Analysis (50%)

PubMed Articles Statistics.
- Topic: your own keywords?
- The statistics of PubMed articles according to Year, Month and day.
- Compute the TF, IDF, TF*IDF of each word.
- Select the representative words according to TF*IDF.
- Obtain the trend of your research from the web site significant pattern history
  - (Having your own comments and conclusions)
(2014.11.17) Presentation (PPT)
- Please summarize your experimental results and present using PPT (15min~20 min).
(2014.11.24 Report (Please just email your report to me, without have class at 11/24)
- Report with MS-Word format (Title, keywords, Introduction, approaches, experimental result, conclusion and discussion.
- Report format (Title, Background, methods, Experimental result, conclusion and future works)

(2) PubMed Articles Classification (50%)

Source Program : program13_SVMClassifier_jdwang2014_12_30_Pattern10000Test.7z
- Example : ReadMe_Pattern_Top_Log(TF)IDF_ICF_LimitCnt10000_5foldCV.txt
- Result: program13_SVMClassifier_jdwang2014_12_30_Pattern10000.7z
Experimental Resource
- DataSet1:
  - XML format
    - (From jdwang)pubmed_result_WilsonDisease_jdwang2014_10_9.xml
    - (From David) pubmed_result_DNAmethylation_microRNA_cancer.xml
    - (From Heri) pubmed_result_H5N1.xml
    - (From Ezra) epi_mirna.xml

DataSet2:
- Download at least three types (keywords) of PubMed Articles. (Created by yourself)

Parse the XML format articles.
Extract the features according to tf*idf weighting method.
Select features according to their tf*idf weighting.
Transform all of articles into vectors with class-labels via selected features.
Try SVM classifier with k-fold Cross-Validation(CV).
Find the accuracy, Recall, Precision and F1, Confusion Matrix, and show your observation.
Comparison:

Different number of selected features. (1000,3000,5000,10000,20000)
Different value k of k-fold CV. (k=5,10)
Computation Time = Training Time + Testing Time
Accuracy + Computation Time =>LIBSVM vs. LIBLINEAR -- A Library for Large Linear Classification
Confusion Matrix
Do you have another feature selection or weighting method?

(2015.1.19) Presentation (PPT)
(2015.1.21) Report （Email to jdwang@asia.edu.tw）
- Title Page (Final Project, Your Name, ID, Date)
- Method
  - LIBSVM
  - LIBLINEAR