A Study Of Search Satisfaction Evaluation Approach Based On Search Engine Log Mining

Search engines are one of the important ways for users to get the required resource information from the huge amount of data on the Internet.With the rapid expansion of the scale of data on the Internet,users require to obtain resource information more efficiently and conveniently.To provide better service to users,search engines need to optimize search systems continuously.Therefore,how to evaluate the quality of search engines effectively has become the focus of researchers and industries.Search satisfaction is a metric to evaluate the quality of search engines,which emphasizes the search experience of each user.A series of behaviors generated from the interaction processes between users and search engines are recorded in search engine log,e.g.,submitting a query,moving the mouse,and clicking a result,and there is a strong correlation between user behaviors and search satisfaction.Therefore,researchers propose to use user behaviors to evaluate search satisfaction.To deal with the challenges of how to exploit the time interval information effectively in search action sequences,how to exploit the trajectory information of mouse cursor movement effectively,and how to train an effective search satisfaction evaluation model with a small amount of labeled data,this thesis conducts in-depth studies on search-engine-log-mining based search satisfaction approaches,the main contents are:1)Propose a long short-term memory based search satisfaction evaluation approach.The proposed approach employs long short-term memory that performs end-to-end fine-tuning during the training to model search action sequences,and introduces dummy idle actions to represent the dwell time in search action sequences.In addition,a dwell time perturbation based data augmentation strategy is introduced to increase the pattern variations on search action sequences,so as to improve the generalization ability of the models for evaluating search satisfaction.The experimental results show that the proposed approach achieves significant performance improvement compared with several excellent search satisfaction evaluation approaches;2)Propose a region-action long short-term memory for search satisfaction evaluation.We first leverage regions together with actions to extract mouse interaction sequences from search engine log.Since the region information and the action information are modeled by the region gate and the action gate,respectively,the region-action long short-term memory could capture the interactive relations between regions and actions without subjecting the network to higher training complexity.A multi-factor perturbation based data augmentation strategy is introduced to increase the pattern variations on user interaction sequences.The experimental results show that the proposed approach achieves better performance than the state-of-the-art approach for search satisfaction evaluation in different search environments;3)Propose a multi-view semi-supervised search satisfaction evaluation approach.This approach exploits a semi-supervised learning approach to improve the performance of search satisfaction evaluation by combining a small amount of labeled data and a large amount of unlabeled data.On this basis,the idea of multi-view learning is introduced to overcome the problem that the single view semi-supervised learning approach easily falls into local maxima.Use the strategy of different parameter configurations to make enough disagreement between base classifiers,so as to alleviate the requirement of sufficient and redundant views.The experimental results show that the proposed approach achieves better performance than the state-of-the-art semi-supervised approach for evaluating search satisfaction;4)Propose a multi-view active semi-supervised search satisfaction evaluation approach.On the one hand,by exploiting unlabeled data,semi-supervised learning enhances the accuracy of the classifiers used in the active learning part;on the other hand,the active learning part uses human efforts in several stages and uses a regional density measurement to measure the representativeness of each candidate sample,and combines the idea of multi-view to measure the informativeness of each candidate sample,so as to select the most valuable data for querying labels.In this fashion,active learning provides a higher quality labeled training dataset to the semi-supervised part.The experimental results show that the combination of semi-supervised learning and active learning processes could achieve better performance than each individual approach for evaluating search satisfaction.