Quantum has always gained considerable attention in scientific studies as it defies common perception. Quantum Computing- recently evolving- has also started to gain considerable attention. Quantum Computers can solve, some of the distinct challenges in Nondeterministic polynomials (NP)-Hard problems faster than traditional computers. This work is based on the implementation of Task1-feature selection in a shared task, QCLEF2024 with MQ2007 dataset containing 46 features. The task is performed using simulated annealing and Quantum Annealing. The performance of both annealing methods is analyzed based on ndcg@10 (Normalized Discounted Cumulative Gain) and Annealing Time. We obtained a result of 0.3621 and 27222 milliseconds for ndcg@10 and Annealing Time respectively while using Quantum Annealing. We obtained a result of 0.4024 and 284106 milliseconds for ndcg@10 and Annealing Time respectively while using simulated Annealing.
Feature selection is an essential process in ML to select the subset of relevant features to improve the
model performance from the available features. It reduces the dimensionality, complexity, overfitting,
and computational cost, which improves model generalization. Similarly QML feature selection
decreases the curse of dimensionality, computational complexity, and efect of overfitting and solved
through heuristics [
Wang [
Poggiali et al. [
Hellstern et al. [
Bhagawati and Subramanian [
This experimental research is carried out in the workspace provided by the QCLEF2024 using the
MQ2007 Dataset for Task1A.[
The MQ2007 dataset contains 47 features among those features one feature is the target feature with the size 41955. The label named target of the MQ2007 dataset is labeled by human experts as “relevant”, “partially relevant”, and “not relevant”. MQ2007 dataset is in text format, and all the features are of datatype float64 except target, which is of int64 datatype containing values 0,1 or 2. The experimental research of shared Task 1A, feature selection of QCLEF2024 was carried out by Simulated Annealing(SA) and Quantum Annealing(QA), and the performance comparison was made using diferent parameters and selected the most relevant 5 features from the available features.
The Binary Quadratic Model(BQM) is a collection of binary-valued variables that can hold any two values associated with linear and quadratic biases. Quadratic Unconstrained Binary Optimization (QUBO) models are used by samplers such as the D-Wave system. D-wave uses the process of Quantum Annealing to discover results to solve some distinct types of NP-Hard problems. The problem can be minimized using the minimization formula 1https://www.microsoft.com/en-us/research/project/letor-learning-rank-information-retrieval/ ∑︀ + ∑︀< , and , denotes linear and quadratic coeficients.
QUBO is the minimization energy. To extract useful information from the dataset, mutual information and conditional information are used, The dataset is quantified using the Shannon Entropy formula H(X)=− ∑︀ p(x) () p(x) denotes the probability of occurrence of an event.
For Conditional Shannon Entropy (CSE) is calculated if Y is known and X is unknown. H(X|Y) = H(X,Y) - H(Y) H(X|Y)=− ∑︀ p(x,y) (, ) - ( ) P(x,y) denotes the probability of x and y.
H(X, Y) denotes information X and Y together. . The features interact with each other using qubits. So the features are quantified using Shannon Entropy(SE) and normalized using Conditional Shannon Entropy(CSE). The correlation or dependency of the target features with the other features is calculated using the mutual information formula I(X;Y)=∑︀ ∑︀ p(x,y) ((),()) p(x) and p(y) denotes marginal probability p(x,y) denotes joint probability I(X;Y|Z)=(|) - (|, ) H(X|Z) denotes CSE of X on Z H(X|Y,Z) denotes CSE of X on Y and Z
The variable of interest between the target feature and other features is calculated and stored as
another feature. This calculation is performed by the formula
The MQ2007, a LETOR (LEarning TO Rank) document retrieval dataset consists of user queries and the
corresponding retrieved documents fetched by the query[
This Task1A is accomplished in the workspace provided by the QCLEF2024 using the MQ2007 Dataset. The dataset has 47 features, among them one is used as a target. We selected only 5 important features from the 46 features, which are more relevant for future problem solving. The performance metrics of Quantum Annealing are shown in Table 1 and Simulated Annealing is shown in Table 2. The result shows the performance of Simulated Annealing is better when compared with Quantum Annealing in terms of ndcg@10 and Annealing Time. The mutual information between the feature and target is visualized according to rank in Figure 1.
The shared task, Task 1A by QCLEF2024 intended to perform feature selection from MQ2007_train dataset. The most prominent 5 features are selected from the available 46 features. This implementation is carried out by simulated Annealing and Quantum Annealing and we observed that the annealing time of the Quantum model is less when compared to Simulated Annealing.