African University of Science and Technology
https://repository.aust.edu.ng:443/xmlui
The DSpace digital repository system captures, stores, indexes, preserves, and distributes digital research material.Wed, 07 Jun 2023 18:08:36 GMT2023-06-07T18:08:36ZNeural Collaborative Filtering and Autoencoder Enabled Deep Learning Models for Recommender Systems
http://repository.aust.edu.ng/xmlui/handle/123456789/5123
Neural Collaborative Filtering and Autoencoder Enabled Deep Learning Models for Recommender Systems
Arnold, Kwofie
Finding important and useful information is getting harder as much more information is available online. The challenge for content producers is to deliver the appropriate content to the appropriate consumers while making it challenging for users to access that content. The foundation for overcoming these difficulties is provided by recommender systems. Traditional methods like Collaborative Filtering (CF) and Content-Based Recommender Systems have historically been successful in this field of study but are now challenged by problems with data sparsity, cold start, and non-linearity interaction. Evidently, several academic areas, like image detection and natural language processing (El-Bakry, 2008), have shown great interest in deep learning due to outstanding performance and the alluring quality of learning intricate representations. The impact of deep learning is recently showing good advancement when applied to recommender systems research (He, 2008). In this research We dive deep into the Autoencoder and Neural Collaborative Filtering based deep learning models and their implementation on classical collaborative filtering. The research also evaluates the performance of both models and outlines loopholes which can further be improved in future works.
Main Thesis
Tue, 15 Nov 2022 00:00:00 GMThttp://repository.aust.edu.ng/xmlui/handle/123456789/51232022-11-15T00:00:00ZA Lightweight Convolutional Neural Network for Breast Cancer Detection Using Knowledge Distillation Techniques
http://repository.aust.edu.ng/xmlui/handle/123456789/5122
A Lightweight Convolutional Neural Network for Breast Cancer Detection Using Knowledge Distillation Techniques
Modu, Falmata
The second most heterogeneous cancer ever discovered is Breast Cancer (BC). BC is a disease that develops from malignant tumors when the breast cells begin to grow abnor-mally. Although it grows in the breast, it can spread to other body parts or organs.through the lymph and blood vessels of the breast. Globally, more than two million new cases and about 600,000 women died from BC in 2020. Early detection increases the chance of survival by 99%. Deep Learning (DL) models have recorded remarkable achievements in disease diagnosis and treatments. However, it requires powerful computing resources. In this work, we propose a lightweight DL model that can detect BC using the knowledge distillation technique. The knowledge of a pre-trained deep neural network is distilled to a shallow neural work that is easily deployable in a low-power computing environment. We have achieved an accuracy of up to 99%. In addition, we recorded 99% reduction in trainable parameters compared to deploying with a deep neural network.
Main Thesis
Mon, 23 Jan 2023 00:00:00 GMThttp://repository.aust.edu.ng/xmlui/handle/123456789/51222023-01-23T00:00:00ZAlgorithms For Approximation of J-Fixed Points of Nonexpansive - Type Maps, Zeros of Monotone Maps, Solutions of Feasibility and Variational Inequality Problems
http://repository.aust.edu.ng/xmlui/handle/123456789/5121
Algorithms For Approximation of J-Fixed Points of Nonexpansive - Type Maps, Zeros of Monotone Maps, Solutions of Feasibility and Variational Inequality Problems
Nnakwe, Monday Ogudu
It is well known that many physically significant problems in different areas of research can be transformed at equilibrium state into an inclusion problem of the form 0 ∈ Au, where A is either a multi-valued accretive map from a real Banach space into itself or a multi-valued monotone map from a real Banach space into its dual space. In several applications, the solutions of the inclusion problem, when the map A is monotone, corresponds to minimizers of some convex functions. It is known that the sub-differential of any convex function, say g, and denoted by ∂g is monotone, and for any vector, say v, in the domain of g, 0 ∈ ∂g(v) if and only if v is a minimizer of g. Setting ∂g ≡ A, solving the inclusion problem, is equivalent to finding minimizers of g. The method of approximation of solutions of the inclusion problem 0 ∈ Au, when the map A is monotone in real Banach spaces, was not known until in 2016 when Chidume and Idu [52] introduced J-fixed points technique. They proved that the J-fixed points correspond to zerosof monotone maps which are minimizers of some convex functions. In general, finding closed form solutions of the inclusion problem, where A is monotone is
extremely difficult or impossible. Consequently, solutions are sought through the construction of iterative algorithms for approximating J-fixed points of nonlinear maps. In chapter three, four and seven of the thesis, we present a convergence result for approximating zeros of the inclusion problem 0 ∈ Au.
Let H1 and H2 be real Hilbert spaces and K1, K2, · · · , KN , and Q1, Q2, · · · , QP , be nonempty, closed and convex subsets of H1 and H2, respectively, with nonempty intersections K and Q, respectively, that is,
K = K1 ∩ K2 ∩ · · · ∩ KN ̸= ∅ and Q = Q1 ∩ Q2 ∩ · · · ∩ QP ̸= ∅. Let B : H1 → H2 be a bounded linear map, Gi : H1 → H1, i = 1, · · · , N and Aj : H2 → H2,
j = 1, · · · , P be given maps. The common split variational inequality problem introduced by vi Censor et al. [32] in 2005, and denoted by (CSVIP), is the problem of finding an element u ∗ ∈ K for which ( ⟨u − u∗ , Gi(u∗)⟩ ≥ 0, ∀ u ∈ Ki, i = 1, 2, · · · , N, such that ∗ = Bu∗ ∈ Q solves ⟨v − v ∗ , Aj (v∗ )⟩ ≥ 0, ∀ v ∈ Qj
, j = 1, 2, · · · , P. The motivation for studying this class of problems with N > 1 stems from a simple observation that if we choose Gi ≡ 0, the problem reduces to finding u ∗ ∈ ∩N i=1Ki , which is the known convex feasibility problem (CFP) such that Bu∗ ∈ ∩P j=1V I(Qj , Aj ). If the sets Ki are the fixed
point sets of maps Si : H1 → H1, then, the convex feasibility problems (CFP) is the common fixed points problem(CFPP) whose image under B is a common solution to variational inequality problems (CSVIP). If we choose Gi ≡ 0 and Aj ≡ 0, the problem reduces to finding u ∗ ∈ ∩N i=1Ki such that the point
Bu∗ ∈ ∩P j=1Qj which is the well known multiple-sets split feasibility problem or common split feasibility problem which serves as a model for many inverse problems where the constraints are imposed on the solutions in the domain of a linear operator as well as in the range of the operator. A lot of research interest is now devoted to split variational inequality problem and its gener-alizations.In chapter five and six of the thesis, we present convergence theorems for approximating solu-tions of variational inequalities and a convex feasibility problem; and solutions of split varia-tional inequalities and generalized split feasibility problems.
Main Thesis
Fri, 05 Jul 2019 00:00:00 GMThttp://repository.aust.edu.ng/xmlui/handle/123456789/51212019-07-05T00:00:00ZA Performant Predict Analytics Approach to Recommender Systems Using Deep Learning Methods
http://repository.aust.edu.ng/xmlui/handle/123456789/5120
A Performant Predict Analytics Approach to Recommender Systems Using Deep Learning Methods
Obaje, Williams Usman
Recently, Massive amounts of data have been generated as a result of the frequency at which the amount of information available digitally is advancing. To enable users to effectively utilise the huge amount of information, recommendation system has been implemented to effectively manipulate a large amount of data in other to communicate necessary output to the user. The reliability of the final recommendations is a common metric used to determine if a recommendation system is effective or not. The RMSE, MSE, and MAE were used as recommendation-based metrics. The recommender system's performance is typically measured using the metrics RMSE, MSE, and MAE. This statistic demonstrates how effectively the Recommender system performs. The performance of the recommendations improves with decreasing RMSE, MAE, and MSE. It offers an erroneous value that illustrates how far off from the real data our model was. It assesses how closely the projections supplied correlate to the quantities that were observed. The final result of evaluating RMSE, MAE, and MSE on 1M Movies Datasets. Taking the mean result of the output, MAE outperformed other matric because it has the lowest mean value of 0.5843. Also, evaluating results of algorithm SVD on the 100k movies dataset MAE outperformed other matric having the lowest output of 0.6593. Furthermore, evaluating the RMSE, MAE, and MSE of algorithm SVD on 5k movies data set, MAE still outperformed other matric having the lowest mean value of 2.8898. Ultimately, it was discovered that the movie data sets with the most customers and reviews performed better than the others with fewer datasets obtainable. Additionally, we suggest a deep learning approach for creating efficient and accurate deep learning collaborative filtering systems (DLCFS). The proposed method and the currently used methods were compared.
Main Theses
Mon, 05 Sep 2022 00:00:00 GMThttp://repository.aust.edu.ng/xmlui/handle/123456789/51202022-09-05T00:00:00Z