Balázs Hidasi, profile picture
Uploaded byBalázs Hidasi
738 views

ShiftTree: model based time series classifier (ECML/PKDD 2011 presentation)

The document presents 'shifttree', an interpretable model-based approach for time series classification that employs dynamic attribute operators and aims to improve classification accuracy through faster labeling and interpretability. It discusses the construction and evaluation of models using a sequence of simple operators as a decision tree, and highlights both advantages and drawbacks of the method. Additionally, the concept of 'shiftforest' is introduced as a way to combine models for enhanced accuracy while acknowledging potential losses in interpretability.

Embed presentation

SHIFTTREE: AN INTERPRETABLE MODEL-BASED APPROACH FOR TIME SERIES CLASSIFICATION Balázs Hidasi - hidasi@tmit.bme.hu Csaba Gáspár-Papanek - gaspar@tmit.bme.hu Budapest University of Technology and Economics, Hungary ECML/PKDD, 6TH SEPTEMBER 2011, ATHENS
OUTLINE  General concept  Task & adaptation  Labeling & learning  Evaluation  Forest methods  Conclusion 2/17
GENERAL CONCEPT  Not limited to time series  Time series  Semi-structured data  Graphs  Two questions:  „Where to look at?”  „What to observe?”  Operator families  EyeShifter Operator(s) (ESO) F(x)  ConditionBuilder Operators (CBO)  Dynamic attributes  Model: sequence of simple operatos / rules 3/17
TIME SERIES CLASSIFICATION  Time series Learning Model  With one observed variable  Evenly sampled  The algorithm has no such limitations  Standard classification task  Focusing on accuracy Model Labeling  Goals  Satisfying accuracy  Model-based instead of memory-based 4/17  Interpretable models
EXPLAINING THE GOALS  Why model-based?  Labeling is much faster   Can learn more general properties of the classes   Needs more training samples   What is interpretability good for?  Getting user trust   Helping in understanding the data  5/17
CONCEPT APPLIED TO TIME SERIES  Dynamic attributes of time series  EyeShifter Operator (ESO): „Where to look at?”  Moves a cursor along the time axis  To a specified point of the series  E.g.: „To the next local minimum”, „100 steps ahead”, etc.  Note: result of operator sequence depens on the order  ConditionBuilder Operator (CBO): „What to observe?” F(x)  Computes the dynamic attribute  E.g.: „Value at the given time”, „Length of the jump”, etc.  Decision tree as model Weighted average  Works well with the dynamic attribute concept 6/17  Operator sequence from root to leaf
LABELING Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
LABELING Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
LABELING Dynamic attribute value: 0,836848 Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
LABELING Dynamic attribute value: 0,836848 Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
LABELING Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
LABELING Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
LABELING Dynamic attribute value: 0,767139 Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
LABELING Dynamic attribute value: 0,767139 Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
LABELING Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
LEARNING (1/3)  Operator pool definition: ESOs, CBOs  Root node  Cursor starts at the beginning  May be overridden 8/17
F(x) LEARNING (2/3)  Available dynamic attributes (CBO pool)  At/around reachable positions  Selecting a…  Attribute  Threshold value Minimize weighted child node entropy 9/17
LEARNING (3/3)  Best split selected  Cursor is moved by ESO  Reachable positions will change  Train set split  Child nodes continue the learning process  Stop: homogenous nodes 10/17
INTERPRETABILITY  Depends on the operators  Helps undepstanding the data  E.g.: CBF dataset  Z-normalized, 3 class: Cylinder, Bell, Funnel  Distinguish cylinder from bell and funnel by global maxima  The data is z-normalized (standardized)  Distinguish bell from funnel by stepping back 25 steps + noise 11/17 filtering through weighted average  On which side is the peak
PERFORMANCE OF THE ALGORITHM  Databases  UCR: mostly smaller data sets (20)  Ford: larger data sets (2)  No optimizations in the experiments  Performance  Better on larger data sets (model-based) 100.00% 94.97% 96.52% 92.94% 90.00% 81.29% 83.19% 80.00% 74.67% 73.33% 69.11% 70.00% 64.93% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 12/17 0.00% Smaller UCR sets (15) Larger UCR sets (5) Ford (largest) sets (2) 1NN Eucledian 1NN DTW ShiftTree
ADVANTAGES AND DRAWBACKS  Advantages  Advantages of being model based (fast labeling, etc)  Interpretable (depens on operators)  Can be used independently of application domain  Expert’s knowledge can be built in through operators  Disadvantages  Optimization of the operators is not trivial  Learning may take a long time when too many operators used  Disadvantages of the model based methods (larger training set required, etc) 13/17
SHIFTFOREST: COMBINING MULTIPLE MODELS  Combining models enhances accuracy  Boosting  Some issues on smaller datasets  The „XV” method  Splitting training set (random splits)  Building model on the first part, evaluating on the second  Model weight: the measured accuracy 14/17
SHIFTFOREST RESULTS  Accuracy increases  Interpretability is often lost  May the intersection of the trees can be interpreted 100.00% 96.52%97.77% 92.94%94.97% 88.87% 90.00% 85.93% 81.29% 83.19% 80.00% 74.67% 73.33% 69.11% 70.00% 64.93% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% Smaller UCR sets (15) Larger UCR sets (5) Ford (largest) sets (2) 15/17 1NN Eucledian 1NN DTW ShiftTree ShiftForest
CONCLUSION  Novel model-based algorithm  Uses cursors and operators  Cursor movement  Attribute computation  Sequence of simple operators as the model  Decision tree  Base of a whole algorithm family 16/17
THANK YOU FOR YOUR ATTENTION! Balázs Hidasi Csaba Gáspár-Papanek (hidasi@tmit.bme.hu) (gaspar@tmit.bme.hu) 17/17 For more ShiftTree related research visit my side: http://www.hidasi.eu
TIME OF MODEL BUILDING, SCALABILITY  O(N*log(N)) per node (ordering)  Time of model building  Depends on tree structure 200  Acceptable 180 160 140  Scales linearly 120 Time (s)  Experiments 100 80 60 40 20 0 Size of training set Min of 20 Max of 20 Mean of 20
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 27
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 28
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 29
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 30
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 31
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 32
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 33
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 34
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 35
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 36
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 Inf 0 Best score in the node (so far): 0,983634 0,805777 Inf 0 Best ordering: 37
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 Inf 0 Best score in the node (so far): 0,983634 0,805777 Inf 0 Best ordering: 38
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 Inf 0 Best score in the node (so far): 0,983634 0,805777 Inf 0 Best ordering: 39
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 40
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 41
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 42
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 43
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 44
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 45
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 46
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,600131 Inf 0 Best score in the node (so far): Inf 0 Best ordering: 47
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,600131 Inf 0 Best score in the node (so far): Inf 0 Best ordering: 48
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,600131 Inf 0 Best score in the node (so far): Inf 0 Best ordering: 49
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 - Score of the best split by current dyn. attribute : Inf 0 Best score in the node (so far): Inf 0 Best ordering: 50
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 - Score of the best split by current dyn. attribute : Inf 0 Best score in the node (so far): Inf 0 Best ordering: 51
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 - Score of the best split by current dyn. attribute : Inf 0 Best score in the node (so far): Inf 0 Best ordering: 52
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 Score of the best split by current dyn. attribute : 0 Best score in the node (so far): 0 Best ordering: 53
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 Score of the best split by current dyn. attribute : 0 Best score in the node (so far): 0 Best ordering: 54
LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: - CBO: - Dynamic attribute selection M Threshold: - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 Score of the best split by current dyn. attribute : 0 Best score in the node (so far): 0 Best ordering: 55

More Related Content

PPTX
What can we learn from learning analytics?
byMoira Sarsfield
 
PDF
Rr q3 2016_presentation_final_final
byRamirent Group
 
PPTX
Cho thue pg tai hcm, thuê pg, cung cap pg chuyen nghiep gia rẻ tai tp.hcm
byCông ty Tổ Chức Sự Kiện VietSky
 
PPTX
Actividad 2
byMaría de Jesús Rodríguez López
 
PPTX
Công ty tổ chức sự kiện chuyên nghiệp tại tp.hcm
byCông ty Tổ Chức Sự Kiện VietSky
 
PPTX
Cho thue pg tai hcm, thuê pg, cung cap pg chuyen nghiep tai hcm
byCông ty Tổ Chức Sự Kiện VietSky
 
PPT
ENJ-300 Modulo V Curso Etapa Inicial
byENJ
 
DOCX
ประวัติ Kim taehyung
bydit9jkp_14
 
What can we learn from learning analytics?
byMoira Sarsfield
 
Rr q3 2016_presentation_final_final
byRamirent Group
 
Cho thue pg tai hcm, thuê pg, cung cap pg chuyen nghiep gia rẻ tai tp.hcm
byCông ty Tổ Chức Sự Kiện VietSky
 
Actividad 2
byMaría de Jesús Rodríguez López
 
Công ty tổ chức sự kiện chuyên nghiệp tại tp.hcm
byCông ty Tổ Chức Sự Kiện VietSky
 
Cho thue pg tai hcm, thuê pg, cung cap pg chuyen nghiep tai hcm
byCông ty Tổ Chức Sự Kiện VietSky
 
ENJ-300 Modulo V Curso Etapa Inicial
byENJ
 
ประวัติ Kim taehyung
bydit9jkp_14
 

Viewers also liked

PDF
REX du MOOC Cuisine Afpa
byLoïc TOURNEDOUET
 
PPTX
De-identification in Learning Analytics
byMohammad Khalil
 
PPT
Personalidad
byAsesoriacreativa
 
PPTX
Drilling machine class
byPOLAYYA CHINTADA
 
PPTX
трускавец
byschool32
 
PPTX
Drilling machine 1
byashjm
 
REX du MOOC Cuisine Afpa
byLoïc TOURNEDOUET
 
De-identification in Learning Analytics
byMohammad Khalil
 
Personalidad
byAsesoriacreativa
 
Drilling machine class
byPOLAYYA CHINTADA
 
трускавец
byschool32
 
Drilling machine 1
byashjm
 

More from Balázs Hidasi

PDF
Egyedi termék kreatívok tömeges gyártása generatív AI segítségével
byBalázs Hidasi
 
PDF
The Effect of Third Party Implementations on Reproducibility
byBalázs Hidasi
 
PPTX
GRU4Rec v2 - Recurrent Neural Networks with Top-k Gains for Session-based Rec...
byBalázs Hidasi
 
PPTX
Deep Learning in Recommender Systems - RecSys Summer School 2017
byBalázs Hidasi
 
PPTX
Parallel Recurrent Neural Network Architectures for Feature-rich Session-base...
byBalázs Hidasi
 
PPTX
Context aware factorization methods for implicit feedback based recommendatio...
byBalázs Hidasi
 
PPTX
Deep learning to the rescue - solving long standing problems of recommender ...
byBalázs Hidasi
 
PPTX
Deep learning: the future of recommendations
byBalázs Hidasi
 
PDF
Context-aware preference modeling with factorization
byBalázs Hidasi
 
PPTX
Approximate modeling of continuous context in factorization algorithms (CaRR1...
byBalázs Hidasi
 
PPTX
Utilizing additional information in factorization methods (research overview,...
byBalázs Hidasi
 
PPT
Az implicit ajánlási probléma és néhány megoldása (BME TMIT szeminárium előad...
byBalázs Hidasi
 
PPTX
Context-aware similarities within the factorization framework (CaRR 2013 pres...
byBalázs Hidasi
 
PPTX
iTALS: implicit tensor factorization for context-aware recommendations (ECML/...
byBalázs Hidasi
 
PPTX
Initialization of matrix factorization (CaRR 2012 presentation)
byBalázs Hidasi
 
PPT
ShiftTree: model alapú idősor-osztályozó (VK 2009 előadás)
byBalázs Hidasi
 
PPTX
ShiftTree: model alapú idősor-osztályozó (ML@BP előadás, 2012)
byBalázs Hidasi
 
Egyedi termék kreatívok tömeges gyártása generatív AI segítségével
byBalázs Hidasi
 
The Effect of Third Party Implementations on Reproducibility
byBalázs Hidasi
 
GRU4Rec v2 - Recurrent Neural Networks with Top-k Gains for Session-based Rec...
byBalázs Hidasi
 
Deep Learning in Recommender Systems - RecSys Summer School 2017
byBalázs Hidasi
 
Parallel Recurrent Neural Network Architectures for Feature-rich Session-base...
byBalázs Hidasi
 
Context aware factorization methods for implicit feedback based recommendatio...
byBalázs Hidasi
 
Deep learning to the rescue - solving long standing problems of recommender ...
byBalázs Hidasi
 
Deep learning: the future of recommendations
byBalázs Hidasi
 
Context-aware preference modeling with factorization
byBalázs Hidasi
 
Approximate modeling of continuous context in factorization algorithms (CaRR1...
byBalázs Hidasi
 
Utilizing additional information in factorization methods (research overview,...
byBalázs Hidasi
 
Az implicit ajánlási probléma és néhány megoldása (BME TMIT szeminárium előad...
byBalázs Hidasi
 
Context-aware similarities within the factorization framework (CaRR 2013 pres...
byBalázs Hidasi
 
iTALS: implicit tensor factorization for context-aware recommendations (ECML/...
byBalázs Hidasi
 
Initialization of matrix factorization (CaRR 2012 presentation)
byBalázs Hidasi
 
ShiftTree: model alapú idősor-osztályozó (VK 2009 előadás)
byBalázs Hidasi
 
ShiftTree: model alapú idősor-osztályozó (ML@BP előadás, 2012)
byBalázs Hidasi
 

Recently uploaded

PPTX
TravelTech Paris 2025 | Beyond the pipes: Why Channel Management isn’t really...
byapidays
 
PDF
UiPath Automation Developer Associate Training Series 2026 - Session 2
byDianaGray10
 
PDF
Analyze and Preserve Logs - RHCSA (RH134).pdf
byLinuxCert Guru
 
PDF
Writing GPU-Ready AI Models in Pure Java with Babylon
byAna-Maria Mihalceanu
 
PPTX
How Does an ICO Launchpad Work Step-by-Step Breakdown.pptx
byTom Hardy S
 
PPTX
Introduction-------- to-------ARM Cortex
byPujaSengupta6
 
PDF
Designing a Blog Using Wordpress
bymarkzubi50
 
PDF
Advanced SELinux Management - RHCSA (RH134).pdf
byLinuxCert Guru
 
PDF
Mesh WiFi Router: The Smart Solution for Fast, Seamless, Whole-Home Internet
byAeroMesh Systems
 
PDF
UiPath Automation Developer Associate Training Series 2026 - Session 1
byDianaGray10
 
PDF
Post-Hackathon-Learnings-Maximizing-Impact-Beyond-the-Event.pdf
byishantyadav1111
 
PDF
NTG -Company Profile_Software_Vendor_Company_in_MENA
byMustafa Kuğu
 
PDF
The Emergence of Empathic XR: AWE Asia 2026 keynote
byMark Billinghurst
 
PDF
From DeFi POC to Production MVP - A 12-16 Week Blueprint.pdf
byniahiggins21
 
PDF
Explaining specific examples of purpose-specific BOM
byakipii ogaoga
 
PDF
GenerationAI Paris 2025 | City of Light (COL)
byapidays
 
PDF
AI for Risk Management & Fraud Detection ppt.pdf
byalexmartincaneda
 
PDF
Title Installing Windows, Linux, MacOS.pdf
byGeraldAbadajos
 
PDF
Rustici Software: eLearning standards in the age of AI
byRustici Software
 
PDF
Traditional-Security-Models-No-Longer-Work.pptx (1).pdf
byOhhproJunction
 
TravelTech Paris 2025 | Beyond the pipes: Why Channel Management isn’t really...
byapidays
 
UiPath Automation Developer Associate Training Series 2026 - Session 2
byDianaGray10
 
Analyze and Preserve Logs - RHCSA (RH134).pdf
byLinuxCert Guru
 
Writing GPU-Ready AI Models in Pure Java with Babylon
byAna-Maria Mihalceanu
 
How Does an ICO Launchpad Work Step-by-Step Breakdown.pptx
byTom Hardy S
 
Introduction-------- to-------ARM Cortex
byPujaSengupta6
 
Designing a Blog Using Wordpress
bymarkzubi50
 
Advanced SELinux Management - RHCSA (RH134).pdf
byLinuxCert Guru
 
Mesh WiFi Router: The Smart Solution for Fast, Seamless, Whole-Home Internet
byAeroMesh Systems
 
UiPath Automation Developer Associate Training Series 2026 - Session 1
byDianaGray10
 
Post-Hackathon-Learnings-Maximizing-Impact-Beyond-the-Event.pdf
byishantyadav1111
 
NTG -Company Profile_Software_Vendor_Company_in_MENA
byMustafa Kuğu
 
The Emergence of Empathic XR: AWE Asia 2026 keynote
byMark Billinghurst
 
From DeFi POC to Production MVP - A 12-16 Week Blueprint.pdf
byniahiggins21
 
Explaining specific examples of purpose-specific BOM
byakipii ogaoga
 
GenerationAI Paris 2025 | City of Light (COL)
byapidays
 
AI for Risk Management & Fraud Detection ppt.pdf
byalexmartincaneda
 
Title Installing Windows, Linux, MacOS.pdf
byGeraldAbadajos
 
Rustici Software: eLearning standards in the age of AI
byRustici Software
 
Traditional-Security-Models-No-Longer-Work.pptx (1).pdf
byOhhproJunction
 

ShiftTree: model based time series classifier (ECML/PKDD 2011 presentation)

  • 1.
    SHIFTTREE: AN INTERPRETABLE MODEL-BASED APPROACH FOR TIME SERIES CLASSIFICATION Balázs Hidasi - hidasi@tmit.bme.hu Csaba Gáspár-Papanek - gaspar@tmit.bme.hu Budapest University of Technology and Economics, Hungary ECML/PKDD, 6TH SEPTEMBER 2011, ATHENS
  • 2.
    OUTLINE  General concept Task & adaptation  Labeling & learning  Evaluation  Forest methods  Conclusion 2/17
  • 3.
    GENERAL CONCEPT  Not limited to time series  Time series  Semi-structured data  Graphs  Two questions:  „Where to look at?”  „What to observe?”  Operator families  EyeShifter Operator(s) (ESO) F(x)  ConditionBuilder Operators (CBO)  Dynamic attributes  Model: sequence of simple operatos / rules 3/17
  • 4.
    TIME SERIES CLASSIFICATION  Time series Learning Model  With one observed variable  Evenly sampled  The algorithm has no such limitations  Standard classification task  Focusing on accuracy Model Labeling  Goals  Satisfying accuracy  Model-based instead of memory-based 4/17  Interpretable models
  • 5.
    EXPLAINING THE GOALS  Why model-based?  Labeling is much faster   Can learn more general properties of the classes   Needs more training samples   What is interpretability good for?  Getting user trust   Helping in understanding the data  5/17
  • 6.
    CONCEPT APPLIED TOTIME SERIES  Dynamic attributes of time series  EyeShifter Operator (ESO): „Where to look at?”  Moves a cursor along the time axis  To a specified point of the series  E.g.: „To the next local minimum”, „100 steps ahead”, etc.  Note: result of operator sequence depens on the order  ConditionBuilder Operator (CBO): „What to observe?” F(x)  Computes the dynamic attribute  E.g.: „Value at the given time”, „Length of the jump”, etc.  Decision tree as model Weighted average  Works well with the dynamic attribute concept 6/17  Operator sequence from root to leaf
  • 7.
    LABELING Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
  • 8.
    LABELING Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
  • 9.
    LABELING Dynamic attribute value: 0,836848 Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
  • 10.
    LABELING Dynamic attribute value: 0,836848 Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
  • 11.
    LABELING Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
  • 12.
    LABELING Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
  • 13.
    LABELING Dynamic attribute value: 0,767139 Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
  • 14.
    LABELING Dynamic attribute value: 0,767139 Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
  • 15.
    LABELING Level 0 ESO: To global max CBO: Value at time Threshold: 1,20775 Level 1 Level 1 ESO: Step forward 100 ESO: To next local max CBO: Value at time CBO: Value at time Threshold : 0,390093 Threshold : -0,700739 Level 2 Level 2 Level 2 Level 2 7/17 Leaf Leaf Leaf Leaf
  • 16.
    LEARNING (1/3)  Operatorpool definition: ESOs, CBOs  Root node  Cursor starts at the beginning  May be overridden 8/17
  • 17.
    F(x) LEARNING (2/3)  Available dynamic attributes (CBO pool)  At/around reachable positions  Selecting a…  Attribute  Threshold value Minimize weighted child node entropy 9/17
  • 18.
    LEARNING (3/3)  Best split selected  Cursor is moved by ESO  Reachable positions will change  Train set split  Child nodes continue the learning process  Stop: homogenous nodes 10/17
  • 19.
    INTERPRETABILITY  Depends on the operators  Helps undepstanding the data  E.g.: CBF dataset  Z-normalized, 3 class: Cylinder, Bell, Funnel  Distinguish cylinder from bell and funnel by global maxima  The data is z-normalized (standardized)  Distinguish bell from funnel by stepping back 25 steps + noise 11/17 filtering through weighted average  On which side is the peak
  • 20.
    PERFORMANCE OF THEALGORITHM  Databases  UCR: mostly smaller data sets (20)  Ford: larger data sets (2)  No optimizations in the experiments  Performance  Better on larger data sets (model-based) 100.00% 94.97% 96.52% 92.94% 90.00% 81.29% 83.19% 80.00% 74.67% 73.33% 69.11% 70.00% 64.93% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 12/17 0.00% Smaller UCR sets (15) Larger UCR sets (5) Ford (largest) sets (2) 1NN Eucledian 1NN DTW ShiftTree
  • 21.
    ADVANTAGES AND DRAWBACKS  Advantages  Advantages of being model based (fast labeling, etc)  Interpretable (depens on operators)  Can be used independently of application domain  Expert’s knowledge can be built in through operators  Disadvantages  Optimization of the operators is not trivial  Learning may take a long time when too many operators used  Disadvantages of the model based methods (larger training set required, etc) 13/17
  • 22.
    SHIFTFOREST: COMBINING MULTIPLEMODELS  Combining models enhances accuracy  Boosting  Some issues on smaller datasets  The „XV” method  Splitting training set (random splits)  Building model on the first part, evaluating on the second  Model weight: the measured accuracy 14/17
  • 23.
    SHIFTFOREST RESULTS  Accuracyincreases  Interpretability is often lost  May the intersection of the trees can be interpreted 100.00% 96.52%97.77% 92.94%94.97% 88.87% 90.00% 85.93% 81.29% 83.19% 80.00% 74.67% 73.33% 69.11% 70.00% 64.93% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% Smaller UCR sets (15) Larger UCR sets (5) Ford (largest) sets (2) 15/17 1NN Eucledian 1NN DTW ShiftTree ShiftForest
  • 24.
    CONCLUSION  Novel model-basedalgorithm  Uses cursors and operators  Cursor movement  Attribute computation  Sequence of simple operators as the model  Decision tree  Base of a whole algorithm family 16/17
  • 25.
    THANK YOU FORYOUR ATTENTION! Balázs Hidasi Csaba Gáspár-Papanek (hidasi@tmit.bme.hu) (gaspar@tmit.bme.hu) 17/17 For more ShiftTree related research visit my side: http://www.hidasi.eu
  • 26.
    TIME OF MODELBUILDING, SCALABILITY  O(N*log(N)) per node (ordering)  Time of model building  Depends on tree structure 200  Acceptable 180 160 140  Scales linearly 120 Time (s)  Experiments 100 80 60 40 20 0 Size of training set Min of 20 Max of 20 Mean of 20
  • 27.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 27
  • 28.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 28
  • 29.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 29
  • 30.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 30
  • 31.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 31
  • 32.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 32
  • 33.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,332858 -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,985078 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 33
  • 34.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 34
  • 35.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 35
  • 36.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 -1,16155 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 1,32624 Inf 0 Best score in the node (so far): 0,985078 0,983634 0,805777 Inf 0 Best ordering: 36
  • 37.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,332858 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 Inf 0 Best score in the node (so far): 0,983634 0,805777 Inf 0 Best ordering: 37
  • 38.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 ESOMax - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 1,02775 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 Inf 0 Best score in the node (so far): 0,983634 0,805777 Inf 0 Best ordering: 38
  • 39.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 1,02775 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,983634 0,600131 Inf 0 Best score in the node (so far): 0,983634 0,805777 Inf 0 Best ordering: 39
  • 40.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 40
  • 41.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 41
  • 42.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 42
  • 43.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: CBOSimple - Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 43
  • 44.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 44
  • 45.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 45
  • 46.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,739969 -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,805777 0,600131 Inf 0 Best score in the node (so far): 0,805777 Inf 0 Best ordering: 46
  • 47.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,739969 -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,600131 Inf 0 Best score in the node (so far): Inf 0 Best ordering: 47
  • 48.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax ESONext100 - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 0,390093 - Best threshold by current dynamic attribute: -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,600131 Inf 0 Best score in the node (so far): Inf 0 Best ordering: 48
  • 49.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,390093 0,369887 0,77183 2,48622 - Score of the best split by current dyn. attribute : 0,600131 Inf 0 Best score in the node (so far): Inf 0 Best ordering: 49
  • 50.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 - Score of the best split by current dyn. attribute : Inf 0 Best score in the node (so far): Inf 0 Best ordering: 50
  • 51.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 - Score of the best split by current dyn. attribute : Inf 0 Best score in the node (so far): Inf 0 Best ordering: 51
  • 52.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 - Score of the best split by current dyn. attribute : Inf 0 Best score in the node (so far): Inf 0 Best ordering: 52
  • 53.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 Score of the best split by current dyn. attribute : 0 Best score in the node (so far): 0 Best ordering: 53
  • 54.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: ESONextMax - CBO: - CBOSimple Dynamic attribute selection M Threshold: -0,700739 - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 Score of the best split by current dyn. attribute : 0 Best score in the node (so far): 0 Best ordering: 54
  • 55.
    LEARNING THE MODEL Defined operators EyeShifter – Cursor position ESONextMax ESONextMax To the next local maxima ESO ESONext100 ESONext100 Step forward 100 ESOMax ESOMax To the global maxima CBO CBSimple CBOSimple Value at cursor position ShiftTree F(x) ConditionBuilder - Ordering Best split in node (so far) ESO: - CBO: - Dynamic attribute selection M Threshold: - Best threshold by current dynamic attribute: -0,700739 0,369887 2,48622 Score of the best split by current dyn. attribute : 0 Best score in the node (so far): 0 Best ordering: 55