Learning to rank for information retrieval using layered multi-population genetic programming

2022

Preprint

Exploration and Exploitation methods in Evolutionary Algorithms (EA) play a fundamental role for obtaining an optimal solution. Exploration technique target is for search large portion of the search space, while Exploitation procedure is used for converge the current solution to the nearest local or global optimal solution. Example of Exploration procedure in EA is mutation process and Exploitation procedure can be accomplished using crossover process or better initialization at the beginning. This paper uses four probability distributions random generator for mutation process in (1+1)-Evolutionary Algorithms ((1+1)-EAs) for examining their mutation behaviors. In addition, this paper uses ranking models from Dependent Click Model and Linear Regression algorithms as initial solution in (1+1)-EAs for better Exploitation. From the experimental results, the exploration process using Gaussian Random Number (GRN) generator with exploitation procedure using Linear Regression (LR) model and Dependent Click (DC) model initialization in (1+1)-EAs outperformed other methods for evolving steps on training dataset. On the other hand, Levy Random Number generator with LR outperformed other methods used for results on unseen test datatset, while GRN is the strongest competitor compared to Levy for these Predictive rank results. Furthermore, state-of-the-art method in the previous studies Combining Simulated Annealing with (1+1)-Evolutionary Strategy outperformed the predictive ranking and evolving results on testing and training data for both with or without initialization by other Linear ranking compared to novel (1+1)-Evolutionary Gradient Strategy method and other (1+1)-Evolutionary Strategy. This paper demonstrated the results and findings obtained using MQ2008 dataset and providing their package codes.

Section: Related Workmentioning

confidence: 99%

Exploration and Exploitation Methods Using (1+1)-Evolutionary Algorithms for Better Ranking Models

2022

Preprint

“…The listwise approach takes the entire retrieved list of objects (the list of query-document pairs for each query) as the learning instance. Examples of the listwise approaches are ListNET (Listwise Neural Net) [Cao et al, 2007], RankGP [Lin et al, 2012, Mick, Accessed 2016, Coordinate Ascent [Metzler and Bruce Croft, 2007], AdaRank [Xu and Li, 2007] and RankGPES [Islam, 2013]. The proposed ES-Rank method described later in this thesis is a listwise approach because this type has been shown to perform better than pointwise and pairwise approaches [Cao et al, 2007].…”

Section: Evolving Similarity Matching Function Using Genetic Programmingmentioning

confidence: 99%

A Survey on Computational Intelligence Applications in Information Retrieval

2022

Preprint

Computational Intelligence (CI) techniques have been widely employed in information retrieval (IR). CI approaches include everything from evolutionary computation and metaheuristics to machine learning techniques. These methods may be classified based on the problem domain for which they are employed in conjunction with the IR dataset representation formats. When utilizing these methods, the Vector Term Model (VTM) and Vector Feature Model (VFM) are the data representation methods that were used. The VTM represents the document and the query as vectors of term-weights. CI approaches are commonly used in VTM in conjunction with the classical Vector Space Model (VSM), while Term-weights are substituted by feature-weights in the VFM to yield feature vectors of query-document pairs. These features represent the query-document similarity matching values (such as cosine similarity and BM25), the query-document term-weights (such as TF-IDF, Okapi, and Language Models), and the document's online reputation. This paper reviews some of the research works on applying CI approaches to the VTM and VFM.

“…The list-wise approach takes the entire list of objects retrieved (the list of query-document pairs for each query) as the learning instance. Examples of the list-wise approach are ListNET (Cao et al, 2007) which is based in neural networks, RankGP (Lin et al, 2012;Mick, 2016), Coordinate Ascent (Metzler and Bruce Croft, 2007), AdaRank (Xu and Li, 2007) and RankGPES (Islam, 2013).…”

Section: Related Workmentioning

confidence: 99%

An evolutionary strategy with machine learning for learning to rank in information retrieval

Landa-Silva

2018

Soft Comput

Learning to Rank (LTR) is one of the problems in Information Retrieval (IR) that nowadays attracts attention from researchers. The LTR problem refers to ranking the retrieved documents for users in search engines, question answering and product recommendation systems. There is a number of LTR approaches based on machine learning and computational intelligence techniques. Most existing LTR methods have limitations, like being too slow or not being very effective or requiring large computer memory to operate. This paper proposes a LTR method that combines a (1+1)-Evolutionary Strategy with machine learning. Three variants of the method are investigated: ES-Rank, IESR-Rank and IESVM-Rank. They differ on the mechanism to initialize the chromosome for the evolutionary process. ES-Rank simply sets all genes in the