TOWARDS EFFICIENT RECOMMENDER SYSTEMS USING COMPUTATIONAL INTELLIGENCE

Abstract

Due to the nature of digital marketplaces, where the platforms are not bound by physical space, companies are adding more and more content and products and thus increasing the options available to customers manifold. To illustrate, Amazon today has around 12 million products available, and Netflix has approximately 6000 movies and shows on its platform. It is practically impossible for the user to scroll through millions of such options. This problem makes E commerce shopping a daunting task. To solve the problem, recommender systems have emerged as an essential tool. Using recommender systems, the problem of item selection can be offloaded by implementing machine learning algorithms. These algorithms learn and predict what a user is likely to buy, hence reducing the set of products that the user has to go through to find an item that is relevant to his needs. Users often visit a platform just for window shopping, where they have no particular product or need in mind. In such a scenario, recommender systems become even more critical as they bear the entire burden of customer conversion. The system has to pick a comparatively much smaller subset of items to show the user from a much bigger list of products or services. If the system does not select relevant products, it loses potential sales despite having the right products. Thus, recommender systems are essential for working with a large pool of data. It considers the user’s liking, previous purchase/like history, the social network of the user, and much more. We faced the computational challenges of cold-start problem and data sparsity problem in implementing the recommender systems. We addressed these problems in our research work. Following are the objectives of this research work, the methodology we used to carry out these objectives which also construe the research contribution, and the results achieved after performing the research studies. Objectives: The following four objectives have been charted out for this research study: • To research and implement various methods and techniques of performing recommendations. • To improve and optimize the results of recommendations by implementing evolutionary algorithms. vi • To implement deep learning techniques for improving recommendations and design an algorithm which provides best results using these techniques. • To develop a novel recommendation algorithm which could improve the accuracy of the recommendations while remediating the cold start problem and data sparsity problem. Methodology: For achieving the mentioned objectives, this study utilizes machine learning and deep learning techniques like evolutionary algorithms, Neural Networks (NN), Natural Language Processing (NLP), and Topic Modeling approaches due to the tremendous applicability to solving the natural world problems. The following strategies are used to achieve the targeted objectives: • For achieving the first objective, in this objective, the implementation of several benchmark machine learning and deep learning techniques in recommender systems are studied. An extensive literature survey is performed to understand and analyse the holistic application domains of recommender systems. A novel application of recommender systems in the domain of culinary science is carried out to generate efficient results. • In the second objective, the results of recommender systems are gauged under the microscopic lens of evolutionary algorithm. To understand the proof of concept, ensemble learning is applied on benchmark machine learning techniques and the results are further optimized using Particle Swarm Optimization and after generating successful results, In another implementation of recommender systems, the results are optimized by the implementation of genetic algorithm. • For the third objective, after performing a thorough literature survey and several implementations of recommender systems, it was deduced that deep learning produces the most optimum results. Hence, in this objective, several benchmark deep learning techniques were implemented in an ensemble learning setup to produce efficient results. In another implementation of deep learning, Recurrent Neural Network, and its application was implemented to improve the results. • For the final objective, after performing the extensive survey and several implementations of recommender systems, it was discovered that recommender systems greatly suffer from the problem of cold-start and data sparsity. To remediate this problem, Recurrent Recommender Network, a spin-off of RNN, was implemented on real-world Point-of-Sale dataset to generate dynamic fruit recommendations. vii Results: The following research outcomes were attained after performing this research study: • An in-depth analysis of 120 research papers was performed that implemented deep learning techniques in recommender systems. • A novel framework to solve two-fold recommendation problem of food-wine parings where novel features were extracted using text mining and sentiment analysis. Two novel datasets were created and compiled for the process of feature extraction and the results showed the resulting food-wine recommendations aligned with wine sommelier’s food wine recommendations • An AutoML framework for Ensemble Learning Recommendations (EnPSO: Ensemble with Particle Swarm Optimization) was proposed and evolutionary algorithm Particle Swarm Optimization (PSO) for finding the best model was employed. The algorithm was analyzed on three publicly available benchmark MovieLens datasets and five benchmark machine learning techniques were implemented to create ensembles. • A recommender system was proposed to implement AutoML framework for Ensemble Learning (En-DLR: Ensemble based Deep Learning Recommender). Genetic Algorithm was employed to identify the most optimal model in the search space and four benchmark deep learning techniques were implemented as base recommenders. • A dynamic recommender system was implemented to dynamically incorporate the temporal changes in fruit seasonality variations and user preferences using deep learning based LSTM network. • Alleviated the problem of data sparsity with the implementation of Recurrent Recommender Network. • A real-world Point-Of-Sale dataset of a commercial fruit retailer was used for implementing the system.