Researchers from IBM Quantum and Paypal have developed a novel method for dealing with feature selection, one of the most challenging problems in the world. They used the variational quantum feature selection (VarQFS) selection algorithm to create a predictive model for a credit risk data set with 20 and 59 input features (qubits). They trained the model using quantum hardware and tensor-network-based numerical simulations, respectively.
The study, published in Quantum (the open journal for Quantum Science) last 25 January 2023, was initiated by researchers from IBM Quantum, IBM Research Europe, and Paypal, San Jose. Christa Zoufal, Ryan V. Mishmash, Nitin Sharma, Niraj Kumar, Aashish Sheshadri, Amol Deshmukh, Noelle Ibrahim, Julien Gacon, and Stefan Woerner were the experts behind this discovery.
In recent years, quantum computing has emerged as a possible platform for heuristic approaches to computationally demanding combinatorial optimization challenges. Most of these approaches are based on a Quadratic Unconstrained Binary Optimization (QUBO) problem. With this, the Feature selection system aids in the simplification of learning issues and the improvement of training properties.
Training a machine learning (ML) model on a training data set with hundreds of characteristics may become prohibitively expensive or perhaps impossible. Feature selection refers to picking a subset of features that will be used to train a specific model. The primary purpose of this strategy is to accelerate the model training process, reduce memory costs during training, and avoid overfitting.
The Variational Quantum Algorithm Approach
The difficulty of selecting a subset of relevant features to employ in constructing a predictive model, such as fraud detection, is referred to as feature selection. When defined as a general loss function, optimal feature selection provides a minimal framework for building conventional heuristics, mainly resulting in ‘greedy techniques.’
The work presented a variational quantum algorithm for solving unconstrained black box binary optimization problems. The paper also illustrated the theoretical explanation for the strategy adopted based on quantum imaginary time evolution convergence guarantees. The study’s primary goal is to design and test the variational quantum optimization technique for feature selection (VarQFS), which involves training a parameterized quantum circuit to yield good feature subsets through a promising, scalable optimizer, i.e., Quantum Natural SPSA (QN-SPSA).
The ibmq_montreal superconducting qubit processor with 27 qubits was chosen as the Quantum Processing Unit (QPU).
The results demonstrated that the ansatze, which eventually leads to increased entanglement in the system with greater depth, can improve specific characteristics of the trained model. Thus, the ibmq_montreal-based VarQFS model explicitly outperforms traditional approaches even when error mitigation is not used and can perform better when presented with other experimental situations.
Moreover, the study emphasized that when compared to traditional benchmarks, VarQFS can identify feature subsets that result in improved fitness results with regard to a predetermined metric.
The potential of the VarQFS method
In conclusion, researcher data reveals that the quantum method achieved competitive, and in some cases even better, performance than standard feature selection strategies in today’s industry.
According to the researchers, the findings provide the groundwork for future studies of VarQFS experiments on a broader scale using actual quantum technology. While this simulation has been infrequently used for variational quantum algorithms, the researchers believe it has much potential in this context, particularly for understanding the impact that rising entanglement may have in such algorithms.
Read the full article here.