Application of Constrained Optimization Models to Recommender Systems

Seymen, Sinan

doi:https://doi.org/10.21985/n2-swfx-9f88

Work

Application of Constrained Optimization Models to Recommender Systems

Public

Download PDF

Download All Files (.zip)

Recommender systems (RSs) have become essential tools that provide personalized recommendations to their users. These systems may consider user, item provider, and system requirements simultaneously. With the inclusion of possibly clashing considerations, there is a growing focus on solving multiple-objective recommender system (MORS) problems as efficiently as possible. The constrained optimization models can be applied to MORS problems and obtain the optimal solution that can beat myopic heuristics approaches. In this dissertation, we investigate the applications of constrained optimization models to tackle MORS problems. In Chapter 1, we give an introduction and overview of the application of constrained optimization models to RS problems. In Chapter 2, we unify different considerations into a constrained optimization framework where different sets of metrics can be improved by simply using different sets of constraints. Rather than focusing solely on user needs, we tackle some of the most frequently investigated considerations in RSs, such as novelty, diversity, calibration, and fairness of the recommendations. We offer models that can handle multiple considerations simultaneously. Our scalable constrained optimization model tackling the calibration problem is the first in the RS literature. Also, our models are simple and easy to generalize with other considerations. Our experimental results show that the optimization models we offer can outperform state-of-the-art heuristics. We illustrate reasons why the heuristics might struggle to find the optimal solution using a small example. In Chapter 3, we offer a novel constrained optimization model that combines the RS ideas with inventory management. We consider both the preferences of the customers and retailer considerations while direct customer demand by item recommendations. These recommendations will consider perishability and inventory in an online retailing setting, in which we aim to minimize the number of wasted and stockout products. Our model can solve problems with stochastic supply and demand, where the demand, perishability, and inventory are considered not deterministic. We reformulate this model to be able to handle large data and stochasticity. We also note that creating recommendation lists only considering user needs or retailer needs can be counterproductive to the quality of the solution. If the user needs are the exclusive focus, this can lead to stockouts and a large number of perished items. Similarly, if the retailer's needs are the exclusive focus, this can lead to low utility recommendations to the users. Our model tackles this MORS problem by reducing waste by recommending soon-to-perish items, reducing stockouts by considering inventory, and making retailer and user-relevant recommendations simultaneously. We propose heuristic methods to improve the scalability issue the constrained optimization models may face. We compare the performance of these heuristics and note that the optimal solution quality does not decrease significantly. In Chapter 4, we focus on ways to alleviate feasibility and scalability issues that can arise using constrained optimization models in MORS problems. We propose a Dantzig-Wolfe (DW) decomposition-inspired optimization model that overcomes these limitations. We define within-list and across-list constraints, and how our model handles scalability considering these constraints. We compare our model with a recently proposed constrained optimization model; and state-of-the-art heuristics that specialize in one objective at a time using the MovieLens 20M dataset. We claim that our model can scale, find near-optimal solutions, and solve MORS problems with the flexibility to incorporate different considerations. We discuss the future research possibilities of applying constrained optimization models to RS problems in Chapter 5.

Creator