یک چارچوب کاری برای تولیدA Framework for Generating Explainable Recommendations via Multitask Learning پیشنهادهای قابل توضیح با استفاده از یادگیری چند وظیفه‌ای

نوع: Type: thesis

مقطع: Segment: masters

عنوان: Title: یک چارچوب کاری برای تولیدA Framework for Generating Explainable Recommendations via Multitask Learning پیشنهادهای قابل توضیح با استفاده از یادگیری چند وظیفه‌ای

ارائه دهنده: Provider: Mohammad Mobin Shahidi

اساتید راهنما: Supervisors: Dr. Muharram Mansoorizadeh

اساتید مشاور: Advisory Professors:

اساتید ممتحن یا داور: Examining professors or referees: Dr. Mir Hossein Dezfoulian - Dr. Hassan Bashiri

زمان و تاریخ ارائه: Time and date of presentation: 2023/02/05

مکان ارائه: Place of presentation: Amphitheater

چکیده: Abstract: Recommender System is an intelligent information retrieval system that predicts a list of recommendations the user is likely to be interested in based on past user feedback like ratings or likes. In the field of recommender systems, the user's ratings of the items are used as a data source to extract the preferences of the users on the characteristics of the items. Generally, users' ratings of items are stored in a two-dimensional matrix called the Utility Matrix. On the other hand, when a user comes across a recommendation list without explanation, the question arises in his subconscious why a particular item is recommended to him. In other words, there are no explanations to clarify the results of the recommender system. The sparsity of the Utility Matrix (lack of data) and the lack of transparency of the recommended list are known as two main challenges in recommender systems. In recent years, to overcome these challenges, the idea of using user reviews has been proposed. First, reviews like ratings include users' preferences for items. Second, reviews can be used as a data source to build textual explanations. So, by using reviews along with ratings as a supplementary data source, the problem of data deficiency can be solved to some extent. They can also be used to create textual explanations for recommendations. In short, the aim of this research is to develop an explainable recommender system, in which ratings and reviews consist of its data sources. This system is introduced in the form of a framework that is responsible for doing two tasks: predicting the user-item ratings and making related explanations for recommendations. In order to perform the first task, the Context-Aware Probabilistic Matrix Factorization method has been used. In this method, users' preferences and item attributes are learned in four low-rank matrices from two sources of ratings and reviews with the help of multi-task learning. In order to perform the second task, a multi-layer filtering algorithm has been used. The purpose of this algorithm is to select a few suitable sentences from the set of reviews that can be used as explanations for the recommendations provided for the target user. This algorithm refines sentences that are close to the interest of the target user; furthermore, these sentences contain useful and effective information about the item and also have positive emotionality. The explanations provided for the recommendations are actually sentences written by different users, so these sentences have different writing styles. In order to improve the level of personalization, this research tried to make the explanations obtained from the filtering algorithm closer to the writing style of the target user. Therefore, for each user, a specific model of writing style reconstruction was developed. These models learn the writing style of each user based on his reviews. After completing the training process of these models, it was expected that by injecting the explanations obtained from the filtering algorithm into the reconstruction model of the user's writing style, the explanations would be closer to the target user's writing style. Although we achieved acceptable results in recommending items for users and generating related explanations for recommendations, we did not achieve good results in reconstructing the writing style section due to the data deficiency and exposure bias problem, which is a common problem in sequence-to-sequence models

فایل: ّFile: