overview kernels for sequences and graphs
play

Overview: Kernels for Sequences and Graphs String Kernels 8 - PowerPoint PPT Presentation

Oct 01, 2023 •310 likes •1.8k views

Memorial Sloan-Kettering Cancer Center Overview: Kernels for Sequences and Graphs String Kernels 8 Example Sequence Classification Position-(In)dependent Kernels Advanced Kernels Easysvm Kernels on Graphs 9 Basics Random Walks Subtrees

  1. Memorial Sloan-Kettering Cancer Center Regulatory Modules Kernel [Schultheiss et al., 2008] Search for overrepresented motifs m 1 , . . . , m M (colored bars) Find best match of motif m i in example x j ; extract windows s i , j at position p i , j around matches (boxed) Use a string kernel, e.g. k WDS , on all extracted sequence windows, and define a combined kernel for the sequences: k seq ( x j , x k ) = � M i =1 k WDS ( s i , j , s i , k ) Use a second kernel k pos , e.g. based on RBF kernel, on vector of pairwise distances between the motif matches: f j = ( p 1 , j − p 2 , j , p 1 , j − p 3 , j , . . . , p M − 1 , j − p M , j ) Regulatory Modules kernel: k RM ( x , x ′ ) := k seq ( x , x ′ ) + k pos ( x , x ′ ) @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 106

  2. Memorial Sloan-Kettering Cancer Center Regulatory Modules Kernel [Schultheiss et al., 2008] Search for overrepresented motifs m 1 , . . . , m M (colored bars) Find best match of motif m i in example x j ; extract windows s i , j at position p i , j around matches (boxed) Use a string kernel, e.g. k WDS , on all extracted sequence windows, and define a combined kernel for the sequences: k seq ( x j , x k ) = � M i =1 k WDS ( s i , j , s i , k ) Use a second kernel k pos , e.g. based on RBF kernel, on vector of pairwise distances between the motif matches: f j = ( p 1 , j − p 2 , j , p 1 , j − p 3 , j , . . . , p M − 1 , j − p M , j ) Regulatory Modules kernel: k RM ( x , x ′ ) := k seq ( x , x ′ ) + k pos ( x , x ′ ) @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 106

  3. Memorial Sloan-Kettering Cancer Center Regulatory Modules Kernel [Schultheiss et al., 2008] Search for overrepresented motifs m 1 , . . . , m M (colored bars) Find best match of motif m i in example x j ; extract windows s i , j at position p i , j around matches (boxed) Use a string kernel, e.g. k WDS , on all extracted sequence windows, and define a combined kernel for the sequences: k seq ( x j , x k ) = � M i =1 k WDS ( s i , j , s i , k ) Use a second kernel k pos , e.g. based on RBF kernel, on vector of pairwise distances between the motif matches: f j = ( p 1 , j − p 2 , j , p 1 , j − p 3 , j , . . . , p M − 1 , j − p M , j ) Regulatory Modules kernel: k RM ( x , x ′ ) := k seq ( x , x ′ ) + k pos ( x , x ′ ) @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 106

  4. Memorial Sloan-Kettering Cancer Center Regulatory Modules Kernel [Schultheiss et al., 2008] Search for overrepresented motifs m 1 , . . . , m M (colored bars) Find best match of motif m i in example x j ; extract windows s i , j at position p i , j around matches (boxed) Use a string kernel, e.g. k WDS , on all extracted sequence windows, and define a combined kernel for the sequences: k seq ( x j , x k ) = � M i =1 k WDS ( s i , j , s i , k ) Use a second kernel k pos , e.g. based on RBF kernel, on vector of pairwise distances between the motif matches: f j = ( p 1 , j − p 2 , j , p 1 , j − p 3 , j , . . . , p M − 1 , j − p M , j ) Regulatory Modules kernel: k RM ( x , x ′ ) := k seq ( x , x ′ ) + k pos ( x , x ′ ) @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 106

  5. Memorial Sloan-Kettering Cancer Center Local Alignment Kernel In order to compute the score of an alignment, one needs: gap penalty substitution matrix S ∈ R Σ × Σ g : N → R An alignment π is then scored as follows: CGGSLIAMM----WFGV |...|||||....|||| C---LIVMMNRLMWFGV s S , g ( π ) = S ( C , C ) + S ( L , L ) + S ( I , I ) + S ( A , V ) + 2 S ( M , M ) + S ( W , W ) + S ( F , F ) + S ( G , G ) + S ( V , V ) − g (3) − g (4) Smith-Waterman score (not positive definite) SW S , g ( x , y ) := max π ∈ Π( x , y ) s S , g ( π ) Local Alignment kernel [Vert et al., 2004] K β ( x , y ) = � π ∈ Π( x , y ) exp( β s S , g ( π )) @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 107

  6. Memorial Sloan-Kettering Cancer Center Local Alignment Kernel In order to compute the score of an alignment, one needs: gap penalty substitution matrix S ∈ R Σ × Σ g : N → R An alignment π is then scored as follows: CGGSLIAMM----WFGV |...|||||....|||| C---LIVMMNRLMWFGV s S , g ( π ) = S ( C , C ) + S ( L , L ) + S ( I , I ) + S ( A , V ) + 2 S ( M , M ) + S ( W , W ) + S ( F , F ) + S ( G , G ) + S ( V , V ) − g (3) − g (4) Smith-Waterman score (not positive definite) SW S , g ( x , y ) := max π ∈ Π( x , y ) s S , g ( π ) Local Alignment kernel [Vert et al., 2004] K β ( x , y ) = � π ∈ Π( x , y ) exp( β s S , g ( π )) @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 107

  7. Memorial Sloan-Kettering Cancer Center Local Alignment Kernel In order to compute the score of an alignment, one needs: gap penalty substitution matrix S ∈ R Σ × Σ g : N → R An alignment π is then scored as follows: CGGSLIAMM----WFGV |...|||||....|||| C---LIVMMNRLMWFGV s S , g ( π ) = S ( C , C ) + S ( L , L ) + S ( I , I ) + S ( A , V ) + 2 S ( M , M ) + S ( W , W ) + S ( F , F ) + S ( G , G ) + S ( V , V ) − g (3) − g (4) Smith-Waterman score (not positive definite) SW S , g ( x , y ) := max π ∈ Π( x , y ) s S , g ( π ) Local Alignment kernel [Vert et al., 2004] K β ( x , y ) = � π ∈ Π( x , y ) exp( β s S , g ( π )) @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 107

  8. Memorial Sloan-Kettering Cancer Center Locality-Improved Kernel Polynomial Kernel of degree d : Σ �� l � d k POLY ( x , x ′ ) = p =1 I p ( x , x ′ ) ⇒ Computes all d -th order monomials: global information A C d . (. .) G A T T d . Locality-Improved Kernel [Zien et al., 2000] (. .) C G T A d . � N (. .) T T k LI ( x , y ) = p =1 win p ( x , y ) G T d . C (. .) A A A �� + l � d d . (. .) A G win p ( x , y ) = j = − l p j I p + j ( x , y ) Sequence Sequence x x’ � x i = x ′ local/global information 1 , I i ( x , x ′ ) = i 0 , otherwise @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 108

  9. Memorial Sloan-Kettering Cancer Center Fisher & TOP Kernel General idea [Jaakkola et al., 2000; Tsuda et al., 2002a] Combine probabilistic models and SVMs Sequence representation Sequences s of arbitrary length Probabilistic model p ( s | θ ) (e.g. HMM, PSSMs) Maximum likelihood estimate θ ∗ ∈ R d Transformation into Fisher score features Φ( s ) ∈ R d Φ( s ) = ∂ log( p ( s | θ )) ∂ θ Describes contribution of every parameter to p ( s | θ ) k ( s , s ′ ) = � Φ( s ) , Φ( s ′ ) � @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 109

  10. Memorial Sloan-Kettering Cancer Center Example: Fisher Kernel on PSSMs Sequences s ∈ Σ N of fixed length PSSMs: log p ( s | θ ) = log � N i =1 θ i , s i = � N i =1 log θ i , s i =: � N i =1 θ log i , s i Fisher score features: (Φ( s )) i ,σ = d p ( s | θ log ) = Id ( s i = σ ) d θ log i ,σ N � Kernel: k ( s , s ′ ) = � Φ( s ) , Φ( s ′ ) � = Id ( s i = s ′ i ) i =1 Identical to WD kernel with order 1 Note: Marginalized-count kernels [Tsuda et al., 2002b] can be understood as a generalization of Fisher kernels. See e.g. [Sonnenburg, 2002] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 110

  11. Memorial Sloan-Kettering Cancer Center Pairwise Comparison Kernels General idea [Liao and Noble, 2002] Employ empirical kernel map on Smith-Waterman/BLAST scores Advantage Utilizes decades of practical experience with BLAST High computational cost ( O ( N 3 )) Disadvantage Alleviation Employ Blast instead of Smith-Waterman Use a smaller subset for empirical map @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 111

  12. Memorial Sloan-Kettering Cancer Center Summary of String Kernels Kernel l x � = l x ′ Pr ( x | θ ) Posi- Scope Com- tional? plexity linear no no yes local O ( l x ) polynomial no no yes global O ( l x ) locality-improved no no yes local/global O ( l · l x ) sub-sequence yes no yes global O ( nl x l x ′ ) O ( l x ) n-gram/Spectrum yes no no global O ( l x ) WD no no yes local local/global O ( s · l x ) WD with shifts no no yes local/global O ( l x l x ′ ) Oligo yes no yes TOP yes/no yes yes/no local/global depends Fisher yes/no yes yes/no local/global depends @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 112

  13. Memorial Sloan-Kettering Cancer Center Live Demonstration Please check out instructions at http://raetschlab.org/lectures/ MLSSKernelTutorial2012/demo @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 113

  14. Memorial Sloan-Kettering Cancer Center Illustration Using Galaxy Web Service Task 1: Learn to classify acceptor splice sites with GC features 1 Train classifier and predict using 5-fold cross-validation ( SVM Toolbox → Train and Test SVM ) 2 Evaluate classifier ( SVM Toolbox → Evaluate Predictions ) Steps: 1 Use “ Upload file ” with URL http://svmcompbio.tuebingen. mpg.de/data/C_elegans_acc_gc.arff ; set file format to ARFF and upload; file appears in history on right 2 Use “ Train and Test SVM ” on uploaded data set (choose ARFF data format) tool; set the kernel to linear , execute and look at the result 3 Use “ Evaluate Predictions ” on predictions and the labeled data (choose ARFF format), select ROC Curve and execute; check out the evaluation summary and the ROC curves @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 114

  15. Memorial Sloan-Kettering Cancer Center Illustration Using Galaxy Web Service Task 1: Learn to classify acceptor splice sites with GC features 1 Train classifier and predict using 5-fold cross-validation ( SVM Toolbox → Train and Test SVM ) 2 Evaluate classifier ( SVM Toolbox → Evaluate Predictions ) Steps: 1 Use “ Upload file ” with URL http://svmcompbio.tuebingen. mpg.de/data/C_elegans_acc_gc.arff ; set file format to ARFF and upload; file appears in history on right 2 Use “ Train and Test SVM ” on uploaded data set (choose ARFF data format) tool; set the kernel to linear , execute and look at the result 3 Use “ Evaluate Predictions ” on predictions and the labeled data (choose ARFF format), select ROC Curve and execute; check out the evaluation summary and the ROC curves @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 114

  16. Memorial Sloan-Kettering Cancer Center Illustration Using Galaxy Web Service Task 1: Learn to classify acceptor splice sites with GC features 1 Train classifier and predict using 5-fold cross-validation ( SVM Toolbox → Train and Test SVM ) 2 Evaluate classifier ( SVM Toolbox → Evaluate Predictions ) Steps: 1 Use “ Upload file ” with URL http://svmcompbio.tuebingen. mpg.de/data/C_elegans_acc_gc.arff ; set file format to ARFF and upload; file appears in history on right 2 Use “ Train and Test SVM ” on uploaded data set (choose ARFF data format) tool; set the kernel to linear , execute and look at the result 3 Use “ Evaluate Predictions ” on predictions and the labeled data (choose ARFF format), select ROC Curve and execute; check out the evaluation summary and the ROC curves @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 114

  17. Memorial Sloan-Kettering Cancer Center Illustration Using Galaxy Web Service Task 1: Learn to classify acceptor splice sites with GC features 1 Train classifier and predict using 5-fold cross-validation ( SVM Toolbox → Train and Test SVM ) 2 Evaluate classifier ( SVM Toolbox → Evaluate Predictions ) Steps: 1 Use “ Upload file ” with URL http://svmcompbio.tuebingen. mpg.de/data/C_elegans_acc_gc.arff ; set file format to ARFF and upload; file appears in history on right 2 Use “ Train and Test SVM ” on uploaded data set (choose ARFF data format) tool; set the kernel to linear , execute and look at the result 3 Use “ Evaluate Predictions ” on predictions and the labeled data (choose ARFF format), select ROC Curve and execute; check out the evaluation summary and the ROC curves @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 114

  18. Memorial Sloan-Kettering Cancer Center Demonstration Using Galaxy Webservice Task 1: Learn to classify acceptor splice sites with sequences 1 Train classifier and predict, using 5-fold cross-validation ( SVM Toolbox → Train and Test SVM ) 2 Evaluate classifier ( SVM Toolbox → Evaluate Predictions ) Steps: 1 Use “ Upload file ” with URL http://svmcompbio.tuebingen. mpg.de/data/C_elegans_acc_seq.arff . Set file format to ARFF and upload. 2 Use “ Train and Test SVM ” on uploaded dataset (choose ARFF data format) tool. Set the kernel to a) Spectrum with degree=6 and b) Weight Degree with degree=6 and shift=0. Execute and look at the result. 3 Use “ Evaluate Predictions ” on predictions and the labeled data (choose ARFF format). Select ROC Curve and execute. Check out the evaluation summary and the ROC curves. @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 115

  19. Memorial Sloan-Kettering Cancer Center Demonstration Using Galaxy Webservice Task 1: Learn to classify acceptor splice sites with sequences 1 Train classifier and predict, using 5-fold cross-validation ( SVM Toolbox → Train and Test SVM ) 2 Evaluate classifier ( SVM Toolbox → Evaluate Predictions ) Steps: 1 Use “ Upload file ” with URL http://svmcompbio.tuebingen. mpg.de/data/C_elegans_acc_seq.arff . Set file format to ARFF and upload. 2 Use “ Train and Test SVM ” on uploaded dataset (choose ARFF data format) tool. Set the kernel to a) Spectrum with degree=6 and b) Weight Degree with degree=6 and shift=0. Execute and look at the result. 3 Use “ Evaluate Predictions ” on predictions and the labeled data (choose ARFF format). Select ROC Curve and execute. Check out the evaluation summary and the ROC curves. @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 115

  20. Memorial Sloan-Kettering Cancer Center Demonstration Using Galaxy Webservice Task 1: Learn to classify acceptor splice sites with sequences 1 Train classifier and predict, using 5-fold cross-validation ( SVM Toolbox → Train and Test SVM ) 2 Evaluate classifier ( SVM Toolbox → Evaluate Predictions ) Steps: 1 Use “ Upload file ” with URL http://svmcompbio.tuebingen. mpg.de/data/C_elegans_acc_seq.arff . Set file format to ARFF and upload. 2 Use “ Train and Test SVM ” on uploaded dataset (choose ARFF data format) tool. Set the kernel to a) Spectrum with degree=6 and b) Weight Degree with degree=6 and shift=0. Execute and look at the result. 3 Use “ Evaluate Predictions ” on predictions and the labeled data (choose ARFF format). Select ROC Curve and execute. Check out the evaluation summary and the ROC curves. @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 115

  21. Memorial Sloan-Kettering Cancer Center Demonstration Using Galaxy Webservice Task 1: Learn to classify acceptor splice sites with sequences 1 Train classifier and predict, using 5-fold cross-validation ( SVM Toolbox → Train and Test SVM ) 2 Evaluate classifier ( SVM Toolbox → Evaluate Predictions ) Steps: 1 Use “ Upload file ” with URL http://svmcompbio.tuebingen. mpg.de/data/C_elegans_acc_seq.arff . Set file format to ARFF and upload. 2 Use “ Train and Test SVM ” on uploaded dataset (choose ARFF data format) tool. Set the kernel to a) Spectrum with degree=6 and b) Weight Degree with degree=6 and shift=0. Execute and look at the result. 3 Use “ Evaluate Predictions ” on predictions and the labeled data (choose ARFF format). Select ROC Curve and execute. Check out the evaluation summary and the ROC curves. @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 115

  22. Memorial Sloan-Kettering Cancer Center Illustration Using Galaxy Web Service Task 2: Determine the best combination of polynomial degree d = 1 , . . . , 5 and SVMs C = { 0 . 1 , 1 , 10 } using 5-fold cross-validation ( SVM Toolbox → SVM Model Selection ) Steps: 1 Reuse the uploaded file from Task 1. 2 Use “ SVM Model Selection ” with uploaded data (choose ARFF format), set the number of cross-validation rounds to 5, set C ’s as 0 . 1 , 1 , 10, select the polynomial kernel and choose the degrees as 1 , 2 , 3 , 4 , 5. Execute and check the results. @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 116

  23. Memorial Sloan-Kettering Cancer Center Illustration Using Galaxy Web Service Task 2: Determine the best combination of polynomial degree d = 1 , . . . , 5 and SVMs C = { 0 . 1 , 1 , 10 } using 5-fold cross-validation ( SVM Toolbox → SVM Model Selection ) Steps: 1 Reuse the uploaded file from Task 1. 2 Use “ SVM Model Selection ” with uploaded data (choose ARFF format), set the number of cross-validation rounds to 5, set C ’s as 0 . 1 , 1 , 10, select the polynomial kernel and choose the degrees as 1 , 2 , 3 , 4 , 5. Execute and check the results. @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 116

  24. Memorial Sloan-Kettering Cancer Center Illustration Using Galaxy Web Service Task 2: Determine the best combination of polynomial degree d = 1 , . . . , 5 and SVMs C = { 0 . 1 , 1 , 10 } using 5-fold cross-validation ( SVM Toolbox → SVM Model Selection ) Steps: 1 Reuse the uploaded file from Task 1. 2 Use “ SVM Model Selection ” with uploaded data (choose ARFF format), set the number of cross-validation rounds to 5, set C ’s as 0 . 1 , 1 , 10, select the polynomial kernel and choose the degrees as 1 , 2 , 3 , 4 , 5. Execute and check the results. @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 116

  25. Memorial Sloan-Kettering Cancer Center Do-it-yourself with Easysvm (Prep) Install Shogun toolbox: wget http://shogun-toolbox.org/archives/shogun/releases/0.7/sources/shogun-0.7.3.tar.bz2 tar xjf shogun-0.7.3.tar.bz2 cd shogun-0.7.3/src ./configure --interfaces=python_modular,libshogun,libshogunui --prefix=~/mylibs make && make install && cd ../.. export PYTHONPATH=~/mylibs/lib/python2.?/site-packages; export LD_LIBRARY_PATH=~/mylibs/lib; export DYLD_LIBRARY_PATH=$LD_LIBRARY_PATH Install Easysvm and get data: wget http://www.fml.tuebingen.mpg.de/raetsch/projects/easysvm/easysvm-0.3.1.tar.gz tar xzf easysvm-0.3.1.tar.gz cd easysvm-0.3.1 && python setup.py install --prefix=~/mylibs && cd .. wget http://svmcompbio.tuebingen.mpg.de/data/C_elegans_acc_gc.arff wget http://svmcompbio.tuebingen.mpg.de/data/C_elegans_acc_seq.arff @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 117

  26. Memorial Sloan-Kettering Cancer Center Do-it-yourself with Easysvm Task 1: Learn to classify acceptor splice sites with GC features 1 Train classifier and predict using 5-fold cross-validation ( cv ) 2 Evaluate classifier ( eval ) SVM C predictions kernel data format and file ���� � �� � � �� � � �� � ~/mylibs/bin/easysvm.py cv 5 1 linear arff C_elegans_acc_gc.arff lin_gc.out 2 features, 2200 examples Using 5-fold crossvalidation head -4 lin_gc.out #example output split 0 -0.8740213 0 1 -0.9755172 2 2 -0.9060478 1 predictions output file data format and file � �� � � �� � � �� � ~/mylibs/bin/easysvm.py eval lin_gc.out arff C_elegans_acc_gc.arff lin_gc.perf tail -6 lin_gc.perf Averages Number of positive examples = 40 Number of negative examples = 400 Area under ROC curve = 91.3 % Area under PRC curve = 55.8 % Accuracy (at threshold 0) = 90.9 % @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 118

  27. Memorial Sloan-Kettering Cancer Center Do-it-yourself with Easysvm Task 2: Determine the best combination of polynomial degree d = 1 , . . . , 5 and SVMs C = { 0 . 1 , 1 , 10 } using 5-fold cross-validation (modelsel) kernel & parameters SVM C ’s � �� � � �� � ~/mylibs/bin/easysvm.py modelsel 5 0 . 1 , 1 , 10 poly 1,2,3,4,5 true false \ output file data format and file � �� � � �� � arff C_elegans_acc_gc.arff poly_gc.modelsel 2 features, 2200 examples Using 5-fold crossvalidation ... head -8 poly_gc.modelsel Best model(s) according to ROC measure: C=10.0 degree=1 Best model(s) according to PRC measure: C=1.0 degree=1 Best model(s) according to accuracy measure: C=10.0 degree=1 ... @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 119

  28. Memorial Sloan-Kettering Cancer Center Demonstration with Easysvm Task 1: Learn to classify acceptor splice sites with sequences 1 Train classifier and predict using 5-fold cross-validation ( cv ) 2 Evaluate classifier ( eval ) SVM C predictions kernel data format and file ���� � �� � � �� � � �� � ~/mylibs/bin/easysvm.py cv 5 1 spec 6 arff C_elegans_acc_seq.arff spec_seq.out predictions output file data format and file � �� � � �� � � �� � ~/mylibs/bin/easysvm.py eval spec_seq.out arff C_elegans_acc_seq.arff spec_seq.perf tail -3 spec_seq.perf Area under ROC curve = 80.4 % Area under PRC curve = 33.7 % accuracy (at threshold 0) = 90.8 % SVM C predictions kernel data format and file ���� � �� � � �� � � �� � ~/mylibs/bin/easysvm.py cv 5 1 WD 6 0 arff C_elegans_acc_seq.arff wd_seq.out predictions output file data format and file � �� � � �� � � �� � ~/mylibs/bin/easysvm.py eval wd_seq.out arff C_elegans_acc_gc.arff wd_seq.perf tail -6 wd_seq.perf Area under ROC curve = 98.8 % Area under PRC curve = 87.5 % Accuracy (at threshold 0) = 97.0 % @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 120

  29. Memorial Sloan-Kettering Cancer Center Kernels on Graphs Graphs are everywhere . . . Graphs in Reality Graphs model objects and their relationships. Also referred to as networks . All common data structures can be modelled as graphs. Graphs in Bioinformatics Molecular biology studies relationships between molecular components. Graphs are ideal to model: Molecules Protein-protein interaction networks Metabolic networks @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 121

  30. Memorial Sloan-Kettering Cancer Center Kernels on Graphs Graphs are everywhere . . . Graphs in Reality Graphs model objects and their relationships. Also referred to as networks . All common data structures can be modelled as graphs. Graphs in Bioinformatics Molecular biology studies relationships between molecular components. Graphs are ideal to model: Molecules Protein-protein interaction networks Metabolic networks @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 121

  31. Memorial Sloan-Kettering Cancer Center Central Questions How similar are two graphs? Graph similarity is the central problem for all learning tasks such as clustering and classification on graphs. Applications Function prediction for molecules, in particular, proteins Comparison of protein-protein interaction networks Challenges Subgraph isomorphism is NP-complete. Comparing graphs via isomorphism checking is thus prohibitively expensive! Graph kernels offer a faster, yet one based on sound principles. @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 122

  32. Memorial Sloan-Kettering Cancer Center Central Questions How similar are two graphs? Graph similarity is the central problem for all learning tasks such as clustering and classification on graphs. Applications Function prediction for molecules, in particular, proteins Comparison of protein-protein interaction networks Challenges Subgraph isomorphism is NP-complete. Comparing graphs via isomorphism checking is thus prohibitively expensive! Graph kernels offer a faster, yet one based on sound principles. @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 122

  33. Memorial Sloan-Kettering Cancer Center Central Questions How similar are two graphs? Graph similarity is the central problem for all learning tasks such as clustering and classification on graphs. Applications Function prediction for molecules, in particular, proteins Comparison of protein-protein interaction networks Challenges Subgraph isomorphism is NP-complete. Comparing graphs via isomorphism checking is thus prohibitively expensive! Graph kernels offer a faster, yet one based on sound principles. @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 122

  34. Memorial Sloan-Kettering Cancer Center From the beginning . . . Definition of a Graph A graph G is a set of nodes (or vertices) V and edges E , where E ⊂ V 2 . An attributed graph is a graph with labels on nodes and/or edges; we refer to labels as attributes . The adjacency matrix A of G is defined as � 1 if ( v i , v j ) ∈ E , [ A ] ij = otherwise , 0 where v i and v j are nodes in G . A walk w of length k − 1 in a graph is a sequence of nodes w = ( v 1 , v 2 , · · · , v k ) where ( v i − 1 , v i ) ∈ E for 1 ≤ i ≤ k . w is a path if v i � = v j for i � = j . @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 123

  35. Memorial Sloan-Kettering Cancer Center From the beginning . . . Definition of a Graph A graph G is a set of nodes (or vertices) V and edges E , where E ⊂ V 2 . An attributed graph is a graph with labels on nodes and/or edges; we refer to labels as attributes . The adjacency matrix A of G is defined as � 1 if ( v i , v j ) ∈ E , [ A ] ij = otherwise , 0 where v i and v j are nodes in G . A walk w of length k − 1 in a graph is a sequence of nodes w = ( v 1 , v 2 , · · · , v k ) where ( v i − 1 , v i ) ∈ E for 1 ≤ i ≤ k . w is a path if v i � = v j for i � = j . @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 123

  36. Memorial Sloan-Kettering Cancer Center From the beginning . . . Definition of a Graph A graph G is a set of nodes (or vertices) V and edges E , where E ⊂ V 2 . An attributed graph is a graph with labels on nodes and/or edges; we refer to labels as attributes . The adjacency matrix A of G is defined as � 1 if ( v i , v j ) ∈ E , [ A ] ij = otherwise , 0 where v i and v j are nodes in G . A walk w of length k − 1 in a graph is a sequence of nodes w = ( v 1 , v 2 , · · · , v k ) where ( v i − 1 , v i ) ∈ E for 1 ≤ i ≤ k . w is a path if v i � = v j for i � = j . @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 123

  37. Memorial Sloan-Kettering Cancer Center Graph Isomorphism Graph isomorphism (cf. Skiena, 1998) Find a mapping f of the vertices of G to the vertices of H such that G and H are identical; i.e. ( x , y ) is an edge of G iff ( f ( x ) , f ( y )) is an edge of H. Then f is an isomorphism, and G and F are called isomorphic. No polynomial-time algorithm is known for graph isomorphism Neither is it known to be NP-complete Subgraph isomorphism Subgraph isomorpism asks if there is a subset of edges and vertices of G that is isomorphic to a smaller graph H. Subgraph isomorphism is NP-complete @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 124

  38. Memorial Sloan-Kettering Cancer Center Graph Isomorphism Graph isomorphism (cf. Skiena, 1998) Find a mapping f of the vertices of G to the vertices of H such that G and H are identical; i.e. ( x , y ) is an edge of G iff ( f ( x ) , f ( y )) is an edge of H. Then f is an isomorphism, and G and F are called isomorphic. No polynomial-time algorithm is known for graph isomorphism Neither is it known to be NP-complete Subgraph isomorphism Subgraph isomorpism asks if there is a subset of edges and vertices of G that is isomorphic to a smaller graph H. Subgraph isomorphism is NP-complete @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 124

  39. Memorial Sloan-Kettering Cancer Center Polynomial Alternatives Graph kernels Compare substructures of graphs that are computable in polynomial time Examples: walks, paths, cyclic patterns, trees Criteria for a good graph kernel Expressive Efficient to compute Positive definite Applicable to wide range of graphs @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 125

  40. Memorial Sloan-Kettering Cancer Center Polynomial Alternatives Graph kernels Compare substructures of graphs that are computable in polynomial time Examples: walks, paths, cyclic patterns, trees Criteria for a good graph kernel Expressive Efficient to compute Positive definite Applicable to wide range of graphs @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 125

  41. Memorial Sloan-Kettering Cancer Center Random Walks Principle Compare walks in two input graphs Walks are sequences of nodes that allow repetitions of nodes Important trick Walks of length k can be computed by taking the adjacency matrix A to the power of k A k ( i , j ) = c means that c walks of length k exist between vertex i and vertex j @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 126

  42. Memorial Sloan-Kettering Cancer Center Product Graph How to find common walks in two graphs? Use the product graph of G 1 and G 2 Definition G × = ( V × , E × ), defined via V × ( G 1 × G 2 ) = { ( v 1 , w 1 ) ∈ V 1 × V 2 : label ( v 1 ) = label ( w 1 ) } { (( v 1 , w 1 ) , ( v 2 , w 2 )) ∈ V 2 ( G 1 × G 2 ) E × ( G 1 × G 2 ) = ( v 1 , v 2 ) ∈ E 1 ∧ ( w 1 , w 2 ) ∈ E 2 ∧ ( label ( v 1 , v 2 ) = label ( w 1 , w 2 )) } Meaning Product graph consists of pairs of identically labeled nodes and edges from G 1 and G 2 @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 127

  43. Memorial Sloan-Kettering Cancer Center Random Walk Kernel The trick Common walks can now be computed from A k × Definition of random walk kernel | V × | ∞ � � λ n A n k × ( G 1 , G 2 ) = [ × ] ij , i , j =1 n =0 Meaning Random walk kernel counts all pairs of matching walks λ is decaying factor for the sum to converge @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 128

  44. Memorial Sloan-Kettering Cancer Center Runtime of Random Walk Kernels Notation given two graphs G 1 and G 2 n is the number of nodes in G 1 and G 2 Computing product graph requires comparison of all pairs of edges in G 1 and G 2 runtime O ( n 4 ) Powers of adjacency matrix matrix multiplication or inversion for n 2 * n 2 matrix runtime O ( n 6 ) Total runtime O ( n 6 ) @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 129

  45. Memorial Sloan-Kettering Cancer Center Tottering Artificially high similarity scores Walk kernels allow walks to visit same edges and nodes multiple times → artificially high similarity scores by repeated visits to same two nodes Additional node labels Mah´ e et al. [2004] add additional node labels to reduce number of matching nodes → improved classification accuracy Forbidding cycles with 2 nodes Mah´ e et al. [2004] redefine walk kernel to forbid subcycles consisting of two nodes → no practical improvement @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 130

  46. Memorial Sloan-Kettering Cancer Center Limitations of Walks Different graphs mapped to identical points in walks feature space [Ramon and G¨ artner, 2003] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 131

  47. Memorial Sloan-Kettering Cancer Center Subtree Kernel (Idea only) Motivation Compare tree-like substructures of graphs May distinguish between substructures that the walk kernel deems identical Algorithmic principle For all pairs of nodes r from V 1 ( G 1 ) and s from V 2 ( G 2 ) and a predefined height h of subtrees: recursively compare neighbors (of neighbors) of r and s subtree kernel on graphs is sum of subtree kernels on nodes Subtree kernels suffer from tottering as well! @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 132

  48. Memorial Sloan-Kettering Cancer Center Subtree Kernel (Idea only) Motivation Compare tree-like substructures of graphs May distinguish between substructures that the walk kernel deems identical Algorithmic principle For all pairs of nodes r from V 1 ( G 1 ) and s from V 2 ( G 2 ) and a predefined height h of subtrees: recursively compare neighbors (of neighbors) of r and s subtree kernel on graphs is sum of subtree kernels on nodes Subtree kernels suffer from tottering as well! @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 132

  49. Memorial Sloan-Kettering Cancer Center All-paths Kernel? Idea Determine all paths from two graphs Compare paths pairwise to yield kernel Advantage No tottering Problem All-paths kernel is NP-hard to compute. Longest paths? Also NP-hard – same reason as for all paths Shortest Paths! computable in O ( n 3 ) by the classic Floyd-Warshall algorithm ’all-pairs shortest paths’ @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 133

  50. Memorial Sloan-Kettering Cancer Center All-paths Kernel? Idea Determine all paths from two graphs Compare paths pairwise to yield kernel Advantage No tottering Problem All-paths kernel is NP-hard to compute. Longest paths? Also NP-hard – same reason as for all paths Shortest Paths! computable in O ( n 3 ) by the classic Floyd-Warshall algorithm ’all-pairs shortest paths’ @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 133

  51. Memorial Sloan-Kettering Cancer Center Shortest-path Kernels Kernel computation Determine all shortest paths in two input graphs Compare all shortest distances in G 1 to all shortest distances in G 2 Sum over kernels on all pairs of shortest distances gives shortest-path kernel Runtime Given two graphs G 1 and G 2 n is the number of nodes in G 1 and G 2 Determine shortest paths in G 1 and G 2 separately: O ( n 3 ) Compare these pairwise: O ( n 4 ) Hence: Total runtime complexity O ( n 4 ) [Borgwardt and Kriegel, 2005] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 134

  52. Memorial Sloan-Kettering Cancer Center Shortest-path Kernels Kernel computation Determine all shortest paths in two input graphs Compare all shortest distances in G 1 to all shortest distances in G 2 Sum over kernels on all pairs of shortest distances gives shortest-path kernel Runtime Given two graphs G 1 and G 2 n is the number of nodes in G 1 and G 2 Determine shortest paths in G 1 and G 2 separately: O ( n 3 ) Compare these pairwise: O ( n 4 ) Hence: Total runtime complexity O ( n 4 ) [Borgwardt and Kriegel, 2005] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 134

  53. Memorial Sloan-Kettering Cancer Center Applications in Bioinformatics Current Comparing structures of proteins Comparing structures of RNA Measuring similarity between metabolic networks Measuring similarity between protein interaction networks Measuring similarity between gene regulatory networks Future Detecting conserved paths in interspecies networks Finding differences in individual or interspecies networks Finding common motifs in biological networks [Borgwardt et al., 2005; Ralaivola et al., 2005] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 135

  54. Memorial Sloan-Kettering Cancer Center Applications in Bioinformatics Current Comparing structures of proteins Comparing structures of RNA Measuring similarity between metabolic networks Measuring similarity between protein interaction networks Measuring similarity between gene regulatory networks Future Detecting conserved paths in interspecies networks Finding differences in individual or interspecies networks Finding common motifs in biological networks [Borgwardt et al., 2005; Ralaivola et al., 2005] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 135

  55. Memorial Sloan-Kettering Cancer Center Image Classification (Caltech 101 dataset, [Fei-Fei et al., 2004] ) Bag-of-visual-words representation is standard practice for object classification systems [Nowak et al., 2006] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 136

  56. Memorial Sloan-Kettering Cancer Center Image Basics [Nowak et al., 2006] Describing key points in images, e.g. using SIFT features [Lowe, 2004] : 8x8 field leads to four 8- dimensional vectors ⇒ 32-dimensional SIFT feature vector describing the point in the image 1 Generate a set of key-points and corresponding vectors 2 Generate a set of representative “code vectors” 3 Record which code vector is closest to key-point vectors 4 Quantize image into histograms h ⇒ ⇒ { f 1 , . . . , f m } ⇒ ⇒ SVM @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 137

  57. Memorial Sloan-Kettering Cancer Center Image Basics [Nowak et al., 2006] Describing key points in images, e.g. using SIFT features [Lowe, 2004] : 8x8 field leads to four 8- dimensional vectors ⇒ 32-dimensional SIFT feature vector describing the point in the image 1 Generate a set of key-points and corresponding vectors 2 Generate a set of representative “code vectors” 3 Record which code vector is closest to key-point vectors 4 Quantize image into histograms h ⇒ ⇒ { f 1 , . . . , f m } ⇒ ⇒ SVM @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 137

  58. Memorial Sloan-Kettering Cancer Center Image Basics [Nowak et al., 2006] Describing key points in images, e.g. using SIFT features [Lowe, 2004] : 8x8 field leads to four 8- dimensional vectors ⇒ 32-dimensional SIFT feature vector describing the point in the image 1 Generate a set of key-points and corresponding vectors 2 Generate a set of representative “code vectors” 3 Record which code vector is closest to key-point vectors 4 Quantize image into histograms h ⇒ ⇒ { f 1 , . . . , f m } ⇒ ⇒ SVM @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 137

  59. Memorial Sloan-Kettering Cancer Center Image Basics [Nowak et al., 2006] Describing key points in images, e.g. using SIFT features [Lowe, 2004] : 8x8 field leads to four 8- dimensional vectors ⇒ 32-dimensional SIFT feature vector describing the point in the image 1 Generate a set of key-points and corresponding vectors 2 Generate a set of representative “code vectors” 3 Record which code vector is closest to key-point vectors 4 Quantize image into histograms h ⇒ ⇒ { f 1 , . . . , f m } ⇒ ⇒ SVM @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 137

  60. Memorial Sloan-Kettering Cancer Center Image Basics [Nowak et al., 2006] Describing key points in images, e.g. using SIFT features [Lowe, 2004] : 8x8 field leads to four 8- dimensional vectors ⇒ 32-dimensional SIFT feature vector describing the point in the image 1 Generate a set of key-points and corresponding vectors 2 Generate a set of representative “code vectors” 3 Record which code vector is closest to key-point vectors 4 Quantize image into histograms h ⇒ ⇒ { f 1 , . . . , f m } ⇒ ⇒ SVM @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 137

  61. Memorial Sloan-Kettering Cancer Center Image Basics [Nowak et al., 2006] Describing key points in images, e.g. using SIFT features [Lowe, 2004] : 8x8 field leads to four 8- dimensional vectors ⇒ 32-dimensional SIFT feature vector describing the point in the image 1 Generate a set of key-points and corresponding vectors 2 Generate a set of representative “code vectors” 3 Record which code vector is closest to key-point vectors 4 Quantize image into histograms h ⇒ ⇒ { f 1 , . . . , f m } ⇒ ⇒ SVM @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 137

  62. Memorial Sloan-Kettering Cancer Center χ 2 -Kernel for Histograms ⇒ ⇒ { f 1 , . . . , f m } ⇒ ⇒ SVM Image described by histogram h C implied by code book C of size d Kernel for comparing two histograms : � � k γ, C ( h C , h ′ − γχ 2 ( h C , h ′ C ) = exp C ) , where γ is a hyper-parameter, d i ) 2 ( h i − h ′ � χ 2 ( h , h ′ ) := , h i + h ′ i i =1 and we use the convention x / 0 := 0 @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 138

  63. Memorial Sloan-Kettering Cancer Center Spatial Pyramid Kernels Decompose image into a pyramid of L levels � � � � 1 , = , k pyr 8 k � � � 1 + 1 + 4 k , 4 k , � � � 1 + 1 + 4 k , 4 k , � � 1 + 2 k , + . . . [Lazebnik et al., 2006] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 139

  64. Memorial Sloan-Kettering Cancer Center General Spatial Kernels Use general spatial kernel with subwindow B � � k γ, B ( h , h ′ ; { γ, B } ) = exp − γ 2 χ 2 B ( h , h ′ ) . where χ 2 B ( h , h ′ ) only considers the key-points within region B Example regions: 1000 subwindows @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 140

  65. Memorial Sloan-Kettering Cancer Center Application of Multiple Kernel Learn- ing Consider set of code books C 1 , . . . , C K or regions B 1 , . . . , B K Each code book C p or region B p leads to a kernel k p ( x , x ′ ). Which kernel is best suited for classification? Define kernel as linear combination K � k ( x , x ′ ) = β p k p ( x , x ′ ) p =1 Use multiple kernel learning to determine the optimal β ’s. [Gehler and Nowozin, 2009] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 141

  66. Memorial Sloan-Kettering Cancer Center Application of Multiple Kernel Learn- ing Consider set of code books C 1 , . . . , C K or regions B 1 , . . . , B K Each code book C p or region B p leads to a kernel k p ( x , x ′ ). Which kernel is best suited for classification? Define kernel as linear combination K � k ( x , x ′ ) = β p k p ( x , x ′ ) p =1 Use multiple kernel learning to determine the optimal β ’s. [Gehler and Nowozin, 2009] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 141

  67. Memorial Sloan-Kettering Cancer Center Application of Multiple Kernel Learn- ing Consider set of code books C 1 , . . . , C K or regions B 1 , . . . , B K Each code book C p or region B p leads to a kernel k p ( x , x ′ ). Which kernel is best suited for classification? Define kernel as linear combination K � k ( x , x ′ ) = β p k p ( x , x ′ ) p =1 Use multiple kernel learning to determine the optimal β ’s. [Gehler and Nowozin, 2009] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 141

  68. Memorial Sloan-Kettering Cancer Center Example: Scene 13 Datasets Classify images into the following categories: CALsuburb kitchen bedroom livingroom MITcoast MITinsidecity MITopencountry MITtallbuilding Each class has between 210-410 example images [Fei-Fei and Perona, 2005] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 142

  69. Memorial Sloan-Kettering Cancer Center Example: Optimal Spatial Kernel of Scene 13 1000 subwindows livingroom 27 subwindows MITcoast 19 subwindows bedroom 26 subwindows CALsuburb 15 subwindows MITtallbuilding 19 subwindows For each class differently shaped regions are optimal @ MLSS 2012, Santa Cruz [Gehler and Nowozin, 2009] � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 143

  70. Memorial Sloan-Kettering Cancer Center Example: Optimal Spatial Kernel of Scene 13 ⇒ 1000 subwindows livingroom 27 subwindows MITcoast 19 subwindows bedroom 26 subwindows CALsuburb 15 subwindows MITtallbuilding 19 subwindows For each class differently shaped regions are optimal @ MLSS 2012, Santa Cruz [Gehler and Nowozin, 2009] � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 143

  71. Memorial Sloan-Kettering Cancer Center Why Are SVMs Hard to Interpret? SVM decision function is α -weighting of training points N � s ( x ) = α i y i k( x i , x ) + b i =1 α 1 · α 2 · α 3 · . . . . . . α N · But we are interested in weights of features @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 144

  72. Memorial Sloan-Kettering Cancer Center Understanding Linear SVMs Support Vector Machine � N � � f ( x ) = sign y i α i k( x , x i ) + b , i =1 Use SVM w from feature space Recall SVM decision function in kernel feature space: N � f ( x ) = y i α i Φ( x ) · Φ( x i ) + b � �� � i =1 =k( x , x i ) Explicitly compute w = � N i =1 α i Φ( x i ) @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 145

  73. Memorial Sloan-Kettering Cancer Center Understanding Linear SVMs Explicitly compute N � w = α i Φ( x i ) i =1 Use w to rank importance | w dim | dim 17 +27.21 For linear SVMs Φ( x ) = x 30 +13.1 For polynomial SVMs, e.g. degree 2: 5 -10.5 · · · · · · � � � 1 1 1 Φ( x ) = ( x 1 x 1 , 2 x 1 x 2 , . . . 2 x 1 x d , 2 x 2 x 3 . . . x d x d ) @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 145

  74. Memorial Sloan-Kettering Cancer Center Understanding String Kernel based SVMs Understanding SVMs with sequence kernels is considerably more difficult For PWMs we have sequence logos: Goal: We would like to have similar means to understand Support Vector Machines @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 146

  75. Memorial Sloan-Kettering Cancer Center SVM Scoring Function - Examples L − k +1 N K � � � � � x [ i ] k , i w = α i y i Φ( x i ) s ( x ) := + b w i =1 k =1 i =1 · · · k-mer pos. 1 pos. 2 pos. 3 pos. 4 A +0.1 -0.3 -0.2 +0.2 · · · C 0.0 -0.1 +2.4 -0.2 · · · G +0.1 -0.7 0.0 -0.5 · · · T -0.2 -0.2 0.1 +0.5 · · · AA +0.1 -0.3 +0.1 0.0 · · · AC +0.2 0.0 -0.2 +0.2 · · · . . . . . ... . . . . . . . . . . TT 0.0 -0.1 +1.7 -0.2 · · · · · · AAA +0.1 0.0 0.0 +0.1 · · · AAC 0.0 -0.1 +1.2 -0.2 . . . . . ... . . . . . . . . . . TTT +0.2 -0.7 0.0 0.0 · · · @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 147

  76. Memorial Sloan-Kettering Cancer Center SVM Scoring Function - Examples L − k +1 K � � � � x [ i ] k , i s ( x ) := + b w k =1 i =1 Examples: WD kernel (R¨ atsch, Sonnenburg, 2005) WD kernel with shifts (R¨ atsch, Sonnenburg, 2005) Spectrum kernel (Leslie, Eskin, Noble, 2002) Oligo kernel (Meinicke et al., 2004) Not limited to SVM’s: Markov chains (higher order/heterogeneous/mixed order) @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 147

  77. Memorial Sloan-Kettering Cancer Center The SVM Weight Vector w T A A A T T A A A A A T A A A C A A A A T T T C A G T A A G A T A A T A A A A A C A A C T T A G A C C C C C C C T T C T A A T T T T T T T A T A T A T T C A C G G C G C C T G T A T C A C T T A C T T C C T C A G C C C C T T G G T C A G A C G G T A C C A G G C T C G G G G C A C T 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 5 ′ 3 ′ weblogo.berkeley.edu A C G T 5 10 15 20 25 30 35 40 45 50 Position AA AC AG AT CA CC CG CT GA GC GG GT TA TC TG TT 5 10 15 20 25 30 35 40 45 50 Position . . . Explicit representation of w allows (some) interpretation! String kernel SVMs capable of efficiently dealing with large k-mers k > 10 But: Weights for substrings not independent @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 148

  78. Memorial Sloan-Kettering Cancer Center Interdependence of k − mer Weights AACGTACGTACACAC T w T What is the score for TA w TA TAC? TAC w TAC Take w TAC ? GT w GT But substrings and . CGT . w CGT overlapping strings . contribute, too! AACGTACG w ... Problem The SVM-w does not reflect the score for a motif @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 149

  79. Memorial Sloan-Kettering Cancer Center Positional Oligomer Importance Matrices (POIMs) Idea: Given k − mer z at position j in the sequence, compute expected score E [ s ( x ) | x [ j ] = z ] ( for small k ) AAAAAAAAAATACAAAAAAAAAA AAAAAAAAAATACAAAAAAAAAC . AAAAAAAAAATACAAAAAAAAAG . . TTTTTTTTTTTACTTTTTTTTTT Normalize with expected score over all sequences POIMs Q ( z , j ) := E [ s ( x ) | x [ j ] = z ] − E [ s ( x ) ] ⇒ Needs efficient algorithm for computation [Sonnenburg et al., 2008] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 150

  80. Memorial Sloan-Kettering Cancer Center Positional Oligomer Importance Matrices (POIMs) Idea: Given k − mer z at position j in the sequence, compute expected score E [ s ( x ) | x [ j ] = z ] ( for small k ) AAAAAAAAAATACAAAAAAAAAA AAAAAAAAAATACAAAAAAAAAC . AAAAAAAAAATACAAAAAAAAAG . . TTTTTTTTTTTACTTTTTTTTTT Normalize with expected score over all sequences POIMs Q ( z , j ) := E [ s ( x ) | x [ j ] = z ] − E [ s ( x ) ] ⇒ Needs efficient algorithm for computation [Sonnenburg et al., 2008] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 150

  81. Memorial Sloan-Kettering Cancer Center Positional Oligomer Importance Matrices (POIMs) Idea: Given k − mer z at position j in the sequence, compute expected score E [ s ( x ) | x [ j ] = z ] ( for small k ) AAAAAAAAAATACAAAAAAAAAA AAAAAAAAAATACAAAAAAAAAC . AAAAAAAAAATACAAAAAAAAAG . . TTTTTTTTTTTACTTTTTTTTTT Normalize with expected score over all sequences POIMs Q ( z , j ) := E [ s ( x ) | x [ j ] = z ] − E [ s ( x ) ] ⇒ Needs efficient algorithm for computation [Sonnenburg et al., 2008] @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 150

  82. Memorial Sloan-Kettering Cancer Center Ranking Features and Condensing Information z i Q ( z , i ) GATTACA 10 +30 Obtain highest scoring z from AGTAGTG 30 +20 Q ( z , i ) (Enhancer or Silencer) AAAAAAA 10 -10 . . . . . . . . . Visualize POIM as heat map; POIM − GATTACA (Subst. 0) Order 1 x-axis: position A y-axis: k-mer C color: importance G T 5 10 15 20 25 30 35 40 45 50 Position Differential POIM Overview − GATTACA (Subst. 0) For large k : differential POIMs; 8 7 x-axis: position 6 Motif Length (k) 5 4 y-axis: k-mer length 3 2 color: importance 1 5 10 15 20 25 30 35 40 45 50 Position @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 151

  83. Memorial Sloan-Kettering Cancer Center GATTACA and AGTAGTG at Fixed Positions 10 and w Q @ MLSS 2012, Santa Cruz � Gunnar R¨ c atsch ( cBio@MSKCC ) Introduction to Kernels 152

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend

Kernels, Sequences and Approximations P . Zinterhof Department of Computer Sciences University

Kernels, Sequences and Approximations P . Zinterhof Department of Computer Sciences University

Kernels, Sequences and Approximations P . Zinterhof Department of Computer Sciences University of Salzburg Smolenice, 07.09 - 09.09.16 Overview Introduction 1 Hausdorff-Distances and Dispersion 2 Quality of Interpolation 3 Computational

314 views • 28 slides
Parametrizations of manifolds with heat kernels, multiscale analysis on graphs, and applications

Parametrizations of manifolds with heat kernels, multiscale analysis on graphs, and applications

Parametrizations of manifolds with heat kernels, multiscale analysis on graphs, and applications to analysis of data sets Mauro Maggioni Mathematics and Computer Science Duke University U.S.C./I.M.I., Columbia, 3/5/08 In collaboration with

350 views • 33 slides
From Monte Carlo to Mountain Passes Moments of Random Graphs With Fixed Degree Sequences Phil

From Monte Carlo to Mountain Passes Moments of Random Graphs With Fixed Degree Sequences Phil

From Monte Carlo to Mountain Passes Moments of Random Graphs With Fixed Degree Sequences Phil Chodrow, MIT ORC February 28th, 2020 1 Community Detection in Graphs Figure from Erika Legara, Community Detection with Networkx . Link 2

638 views • 49 slides
15-150 Fall 2020 Stephen Brookes Lecture 17 Sequences and cost graphs Halloween, a full moon

15-150 Fall 2020 Stephen Brookes Lecture 17 Sequences and cost graphs Halloween, a full moon

15-150 Fall 2020 Stephen Brookes Lecture 17 Sequences and cost graphs Halloween, a full moon and a time change all happening simultaneously announcements Next Tuesday (3 Nov) is ELECTION DAY Class will be held as usual (on zoom)

829 views • 68 slides
2 More Paper Goals The L4 Microkernel Is this actually useful? Is the Operations:

2 More Paper Goals The L4 Microkernel Is this actually useful? Is the Operations:

A short history of kernels Early kernel: a library of device drivers, support for threads (QNX) Operating System Kernels Monolithic kernels: Unix, VMS, OS 360 Unstructured but fast Over time, became very large

318 views • 10 slides
20-03-06 7. Learning Sequences/Behaviors How to use sequences/behaviors? Sequences and more

20-03-06 7. Learning Sequences/Behaviors How to use sequences/behaviors? Sequences and more

20-03-06 7. Learning Sequences/Behaviors How to use sequences/behaviors? Sequences and more generally behaviors are about Sequences are used integrating the concept of time into what is learned. In } to analyze time dependent data general,

298 views • 6 slides
The Gray Code Kernels The Gray Code Kernels The Gray Code Kernels Gil Ben-Artzi Hagit Hel-Or

The Gray Code Kernels The Gray Code Kernels The Gray Code Kernels Gil Ben-Artzi Hagit Hel-Or

The Gray Code Kernels The Gray Code Kernels The Gray Code Kernels Gil Ben-Artzi Hagit Hel-Or Yacov Hel-Or Bar-Ilan University Haifa University IDC 1 Motivation Motivation Image filtering with a successive set of kernels is very

1.06k views • 58 slides
Sequences Sequences are ordered lists of elements, e.g. 2, 3, 5, 7, 11, 13, 17, 19, . . . or a , b

Sequences Sequences are ordered lists of elements, e.g. 2, 3, 5, 7, 11, 13, 17, 19, . . . or a , b

Sequences, sums, cardinality 1 Myrto Arapinis School of Informatics University of Edinburgh October 1, 2014 1 Slides mainly borrowed from Richard Mayr 1 / 21 Sequences Sequences are ordered lists of elements, e.g. 2, 3, 5, 7, 11, 13, 17, 19, .

570 views • 39 slides
SVM . . . if Pr(+1|v) > 0.5 then t (v) = +1 else t(v) = -1 G RAPH K ERNELS R ESEARCH IN A N

SVM . . . if Pr(+1|v) > 0.5 then t (v) = +1 else t(v) = -1 G RAPH K ERNELS R ESEARCH IN A N

G RAPHLET K ERNELS FOR V ERTEX C LASSIFICATION Presenter: Jos Lugo-Martnez Phd Candidate jlugomar@indiana.edu Jun 12, 2015 O UTLINE Overview of classification problems on graphs Graphlet kernels for vertex classification Case

561 views • 39 slides
Scalable Machine Learning 6. Kernels Alex Smola Yahoo! Research and ANU

Scalable Machine Learning 6. Kernels Alex Smola Yahoo! Research and ANU

Scalable Machine Learning 6. Kernels Alex Smola Yahoo! Research and ANU http://alex.smola.org/teaching/berkeley2012 Stat 260 SP 12 6. Kernels Outline Kernels Hilbert Spaces Regularization theory Kernels on strings,

2.07k views • 171 slides
Kernels on structures Andrea Passerini passerini@disi.unitn.it Machine Learning Kernels on

Kernels on structures Andrea Passerini passerini@disi.unitn.it Machine Learning Kernels on

Kernels on structures Andrea Passerini passerini@disi.unitn.it Machine Learning Kernels on structures Kernels on structures Similarity between structured data Kernels allow to generalize notion of dot product (i.e. similarity) to arbitrary

396 views • 37 slides
SVM Kernels COMPSCI 371D Machine Learning COMPSCI 371D Machine Learning SVM Kernels 1 /

SVM Kernels COMPSCI 371D Machine Learning COMPSCI 371D Machine Learning SVM Kernels 1 /

SVM Kernels COMPSCI 371D Machine Learning COMPSCI 371D Machine Learning SVM Kernels 1 / 27 Outline 1 Linear Separability and Feature Augmentation 2 Sample Complexity 3 Computational Complexity 4 Kernels and Nonlinear SVMs 5 Mercers

600 views • 27 slides
GRAPH MINING AND GRAPH KERNELS Part II: Graph Kernels Karsten Borgwardt^ and Xifeng Yan*

GRAPH MINING AND GRAPH KERNELS Part II: Graph Kernels Karsten Borgwardt^ and Xifeng Yan*

Graph Mining and Graph Kernels GRAPH MINING AND GRAPH KERNELS Part II: Graph Kernels Karsten Borgwardt^ and Xifeng Yan* ^University of Cambridge *IBM T. J. Watson Research Center August 24, 2008 | ACM SIG KDD, Las Vegas Graph Mining and

1.16k views • 48 slides
Beta kernels and transformed kernels applications to copulas and quantiles Arthur Charpentier

Beta kernels and transformed kernels applications to copulas and quantiles Arthur Charpentier

Arthur CHARPENTIER, transformed kernels and beta kernels Beta kernels and transformed kernels applications to copulas and quantiles Arthur Charpentier Universit Rennes 1 arthur.charpentier@univ-rennes1.fr http

1.18k views • 96 slides
Graphs () Graphs () Graphs Graphs Graphs are collections of nodes

Graphs () Graphs () Graphs Graphs Graphs are collections of nodes

Graphs () Graphs () Graphs Graphs Graphs are collections of nodes in which various pairs are connected by K08 line segments. The

627 views • 17 slides
Kernel methods and Graph kernels Social and Technological Networks Rik Sarkar University of

Kernel methods and Graph kernels Social and Technological Networks Rik Sarkar University of

Kernel methods and Graph kernels Social and Technological Networks Rik Sarkar University of Edinburgh, 2019. Kernels Kernels are a type of measures of similarity Important technique in Machine learning Used to increase power of many

568 views • 25 slides
Alexander Lee: C: elegans metabolic network Graph of C. elegans metabolic network. Note that

Alexander Lee: C: elegans metabolic network Graph of C. elegans metabolic network. Note that

Alexander Lee: C: elegans metabolic network Graph of C. elegans metabolic network. Note that some nodes very clearly have many more edges than others. Red nodes nodes with degree > 100 Blue nodes nodes with degree > 50

660 views • 21 slides
Liu et al.: Controllability of complex References networks Sandbox slides. Peter Sheridan Dodds

Liu et al.: Controllability of complex References networks Sandbox slides. Peter Sheridan Dodds

Liu et al.: Controllability of complex networks Liu et al.: Controllability of complex References networks Sandbox slides. Peter Sheridan Dodds Department of Mathematics & Statistics Center for Complex Systems Vermont Advanced

329 views • 12 slides
A TOPOLOGISTS VIEW OF SYMMETRIC AND QUADRATIC FORMS Andrew Ranicki (Edinburgh)

A TOPOLOGISTS VIEW OF SYMMETRIC AND QUADRATIC FORMS Andrew Ranicki (Edinburgh)

1 A TOPOLOGISTS VIEW OF SYMMETRIC AND QUADRATIC FORMS Andrew Ranicki (Edinburgh) http://www.maths.ed.ac.uk/ aar Patterson 60++, G ottingen, 27 July 2009 2 The mathematical ancestors of S.J.Patterson Augustus Edward Hough Love

723 views • 33 slides
Hebbian Learning, Hebbian Learning Principal Component Analysis, and Independent Component

Hebbian Learning, Hebbian Learning Principal Component Analysis, and Independent Component

Hebbian Learning, Hebbian Learning Principal Component Analysis, and Independent Component Analysis Donald Hebb wrote in 1949: When an axon in cell A is near enough to excite cell B and repeat- edly and persistently takes part in firing it,

250 views • 13 slides
Increasing Physical Activity through Health-Enabling Technologies: the Project Being Strong

Increasing Physical Activity through Health-Enabling Technologies: the Project Being Strong

Increasing Physical Activity through Health-Enabling Technologies: the Project Being Strong Without Violence Corinna Scharnweber, Wolfram Ludwig, Michael Marschollek, Wolfgang Pein, Peter Schack, Reiner Schubert, Reinhold Haux 29/08/2011

217 views • 9 slides
NICHD NICHD Research Mission & Research Mission & Research Mission & Research

NICHD NICHD Research Mission & Research Mission & Research Mission & Research

Presented at Mailman School of Public Health, Columbia University December 9, 2010 NICHD NICHD Research Mission & Research Mission & Research Mission & Research Mission & Eugene G. Hayunga, Ph.D. Eugene G. Hayunga, Ph.D.

542 views • 26 slides
Leveraging State Pilot Experience in Health Home Programs August 2, 2012; 3:00 4:00PM (ET)

Leveraging State Pilot Experience in Health Home Programs August 2, 2012; 3:00 4:00PM (ET)

Exploring Medicaid Health Homes: Leveraging State Pilot Experience in Health Home Programs August 2, 2012; 3:00 4:00PM (ET) For audio, dial: 1-800-273-7043; Passcode: 596413 A video archive will be posted on http://www.medicaid.gov.

367 views • 35 slides
Bioinformatics: Network Analysis Flux Balance Analysis and Metabolic Control Analysis COMP 572

Bioinformatics: Network Analysis Flux Balance Analysis and Metabolic Control Analysis COMP 572

Bioinformatics: Network Analysis Flux Balance Analysis and Metabolic Control Analysis COMP 572 (BIOS 572 / BIOE 564) - Fall 2013 Luay Nakhleh, Rice University 1 Flux Balance Analysis (FBA) Flux balance analysis (FBA), an optimality-base

955 views • 52 slides

More recommend