Informatics Publishing Limited

Magda B. Fayek ¹, Hatem M. El-Boghdadi ¹, Mohamed A. Gawad ²

Affiliations
1 Computer Engineering Department, Cairo University, Giza-12613, EG
2 IT Section, Ministry of Finance, Cairo, EG

Abstract

Information retrieval (IR) deals with the representation, storage, organization and access to information items. Often users' queries to search engines are not well formulated and hence donot express what the user is searching for exactly. Such poorly defined criteria result in the retrieval of documents that donot exactly meet user expectations. Many attempts have been made for refining document retrieval based on interaction with user. Mostly, those attempts provide the user with functionalities for editing queries and marking documents. To many users these functionalities are too complicated and hence users hardly use them.
In this paper we present an intelligent search engine that targets those poorly defined queries and interactively helps users fine tune their search. The user merely specifies those documents among initially retrieved documents that are most relevant to his request. Then the system makes use of users' relevance feedback in response to initial search results and automatically updates the search criteria initially submitted by the user. The search results are then updated to improve the selection of documents retrieved. The system adopts RBIR (Ranked Boolean IR), which is a modified Boolean model that estimates document relevance using keyword weights to rank search results. Its accuracy is comparable with Vector Space, while keeping processing overhead low.
Results show that a remarkable improvement in precision is achieved already at the first iteration after relevance feedback, especially at very poor criteria and low recall. As recall rate increases the improvement in precision drops, however improvement remains even at a recall rate of 100%. Generally, the average performance of RBIR with relevance feedback is always better than vector space and RBIR. The average improvement ranges between 12% and 60% relative to vector space and 32% and 25% relative to RBIR at low recall rates. As queries become less definitive the enhancement is more profound.

Keywords

Boolean IR Model, IR Evaluation, Relevance Feedback, Recall-Precision Measure, Vector Space Model.

Full Text

     

Magda B. Fayek ¹

Affiliations
1 Computer Engineering Department, Cairo University, EG

Abstract

The inability of GA to find multiple global maxima of a multimodal function has been a basic limitation. Many niching and diversity preserving techniques have been suggested to overcome it. Clearing is one of the common approaches proposed. Due to its promising performance it has attracted researchers to further enhance it. CBCI (Context Based Clearing) is an updated version of standard clearing procedure. This approach made use of context information to prevent clearing of some candidates that may lead to undetected optima. In this paper three advanced dynamic versions of CBC are proposed (CBCII, CBCIII, and CBCIV) to further enhance the clearing process. They are modified versions of CBCI, each by a different approach. The proposed techniques were tested using the M7 function; a massively multimodal deceptive optimization function typically used for testing the efficiency of finding global optima in a search space. The results are compared with those of CBCI. Results show that the new versions discover more optima in earlier generations than CBCI while having a comparable complexity with CBCI. Compared to other reported clearing techniques such as Spatially-Structured EAs (SSEAs) using ring topology, LC (Local clearing) and direct crowding (CD) CBCIV finds optima much faster. the suggested techniques outperform them as regards complexity and manages to find global M7 optima much faster.

Keywords

Clearing, Context, Genetic Algorithms, Multimodal Optimization, Niching Methods.

Full Text