Estimate the Number of Relevant Images Using Two-Order Markov Chain Presented by: WANG Xiaoling Supervisor: Clement LEUNG
Outline Introduction Introduction Objective Objective Methodology Methodology Experiment Results Experiment Results Conclusion and Future Work Conclusion and Future Work
Introduction Large collections of images have been made available on web. Large collections of images have been made available on web. Retrieval effectiveness becomes one of the most important parameters to measure the performance of image retrieval systems. Retrieval effectiveness becomes one of the most important parameters to measure the performance of image retrieval systems.
Measures: Measures: Precision Precision Recall Recall Significant Challenge: the total number of relevant images is not directly observable Significant Challenge: the total number of relevant images is not directly observable
Basic Models Basic Models Regression Model Regression Model Markov Chain Markov Chain Two-Order Markov Chain Two-Order Markov Chain
Objective To investigate the probabilistic behavior of the distribution of relevant images among the returned results for the image search engines using two-order markov chain To investigate the probabilistic behavior of the distribution of relevant images among the returned results for the image search engines using two-order markov chain
Methodology Test Image Search Engine: Test Image Search Engine: Query Design Query Design 70% provided by authors 70% provided by authors One word query One word query Two word query Two word query Three word query Three word query 30% suggestive term 30% suggestive term Suggestive term with largest returned results Suggestive term with largest returned results Suggestive term with least returned results Suggestive term with least returned results
Methodology Database Setup: Database Setup: Stochastic process {X 1, X 2, …, X J } Stochastic process {X 1, X 2, …, X J } where X J denotes the aggregate relevance of all the images in page J where X J denotes the aggregate relevance of all the images in page J Equation: Equation: where Y Ji =1 if the i th image on page J is relevant, and Y Ji =0 if the i th image on page J is not relevant. where Y Ji =1 if the i th image on page J is relevant, and Y Ji =0 if the i th image on page J is not relevant.
Page J XJXJXJXJ
Forecast Using Two-Order Markov Chain Forecast Using Two-Order Markov Chain Markov Chain: Stochastic process {X J, J≥1} with state space S={0,1,2,…20} , Markov Chain: Stochastic process {X J, J≥1} with state space S={0,1,2,…20} , Two-Order Markov Chain: State space change to S 2 , Two-Order Markov Chain: State space change to S 2 , Forecast the state probability distribution of next page π(J) based on the original state probability distribution π(1) and transition probability matrix P. An Example Forecast the state probability distribution of next page π(J) based on the original state probability distribution π(1) and transition probability matrix P. An ExampleAn ExampleAn Example Model Test Model Test Mean Absolute Error Mean Absolute Error
Experiment Results Forecast Results Using Two-Order Markov Chain Forecast Results Using Two-Order Markov Chain PageGoogleYahooBing
Test Results--Google Test Results--Google
Test Results--Yahoo Test Results--Yahoo
Test Results--Bing Test Results--Bing
Measure for Forecast Accuracy Measure for Forecast Accuracy Mean Absolute Deviation (MAD): Mean Absolute Deviation (MAD):One-wordTwo-wordThree-wordGoogle Yahoo Bing
Comparative Results Comparative Results Best Model: Two- Order Markov Chain Best Model: Two- Order Markov Chain Worst Model: Regression Model Worst Model: Regression Model
Conclusion Two-Order Markov Chain could well represent the distribution of relevant images among the results pages for the major web image search engine. Two-Order Markov Chain could well represent the distribution of relevant images among the results pages for the major web image search engine. Two-Order Markov Chain is the best model among three models we have worked. Two-Order Markov Chain is the best model among three models we have worked.
Future Work Our future work will try to apply Hidden Markov Chain to this topic Our future work will try to apply Hidden Markov Chain to this topic
Thank you! Q & A
Two-Order Markov Chain An example (cont ’ ) Suppose the stochastic process {X t, t>=0} with a state space S={A, B, C} Suppose the stochastic process {X t, t>=0} with a state space S={A, B, C} As to two-order Markov chain, the state space: As to two-order Markov chain, the state space: S 2 ={AA, AB, AC, BA, BB, BC, CA, CB, CC} S 2 ={AA, AB, AC, BA, BB, BC, CA, CB, CC} The state probabilities distribution of period zero: The state probabilities distribution of period zero: (0)= ( AA, AB, AC, BA, BB, BC, CA, CB, CC ) (0)= ( AA, AB, AC, BA, BB, BC, CA, CB, CC )
An example (cont ’ ) The transition probability matrix: The transition probability matrix: P AA,BA =0
An example Therefore, the probability distribution of states for page J will be compute as: Therefore, the probability distribution of states for page J will be compute as: π(J)=π(J-1)*P π(J)=π(J-1)*P [Return] [Return]Return