A movie recommendation method based on users' positive and negative profiles
Introduction
Movies have become a daily leisure activity for many people. Since the Lumière Brothers released the first film in 1895, the vigorous development and creation of the film industry in various countries has made films an indispensable leisure activity for the general public. However, nowadays, after more than 100 years of film development, with the rapid development of technology, the traditional screen viewing that originally required audiences to enter a theater has gradually transformed into an online streaming platform that can be watched immediately at home. This online streaming media platform combines novel technologies, focuses on the trends of the big data era, and has improved and established various functions. These functions not only allow users to rate movies, but also allow friends to exchange movie experiences. In addition, the platform uses score data from many users to build a recommendation system to predict the preferred recommendations of individual users.
The recommendation system is widely used in online streaming platforms. The most basic and extensive recommendation methods can be classified as the following categories (He, Parra & Verbert, 2016; Lu, Wu, Mao, Wang & Zhang, 2015; Shah, Gaudani & Balani, 2016; Thorat et al., 2015): (1) Collaborative Filtering, (2) Content-based filtering, and (3) Hybrid Recommendation.
Collaborative filtering aims to use the ratings of similar users or items to predict the user's rating on the next item. On the other hand, content-based filtering (Bergamaschi & Po, 2015) is mainly based on information retrieval and information filtering. It uses the user profile and product content as data, and compares the information contained in the product with the user profile. Content-based methods usually use products that users have purchased in the past to build user profiles. The intuition underpinning this method is that what the users buy represents what they like. Therefore, we can use the features extracted from the purchased products (such as the product's theme, attributes or categories) to represent the user's interests and preferences. In addition, user profiles can also reflect users’ hidden interests through the similarity and relevance of product topics. Content-based filtering will extract product characteristics to represent the target user's preference profile. It then compares the user profile with the characteristics of the product, and recommends the most relevant product set to the user.
When the content-based filtering method is applied to movie recommendation, the method creates a profile representing the user's preference based on the movies previously watched by the user, and then generates the top K movie recommendations from the movies that the user has not yet watched. The criteria for selecting these recommended movies are based on the highest similarity between the content of these movies and the user's interest profile. However, the disadvantage of this method is that it ignores information about movies that the user has never watched before. This missing information has two possible meanings: either the user likes them or the user does not like them. If it is the latter case, the negative information can help the recommendation system make more accurate recommendations by deleting those negative movies. Unfortunately, this negative information has not been used to filter out movies that users do not like. Therefore, in the past, movies recommended by CB to users may hide movies that users do not like. In this paper, we address this issue by first predicting the rating of each user for each item based on the CF algorithm. If the predicted rating value is high, the target user may like the item. On the other hand, if the predicted rating value is low, the target user may not like the item. A positive profile is established based on those items with a higher predicted rating, and a negative profile is established based on those items with a lower predicted rating. If an item has a high degree of similarity with the positive profile of the target user, but a low degree of similarity with the negative profile of the target user, then it may be ideal for recommendation to the target user.
When we are faced with an application situation where we have both product user rating information and product content information, the natural way is to use a hybrid method that combines collaborative filtering methods and content-based methods. This paper uses a hybrid approach to recommend movies that users like. However, there is a big difference between our method and the previous research, that is, we not only use the user's positive preference profile, but we also use the user's negative preference profile to keep favorite movies and delete disliked movies. As a result, our method can improve the accuracy of movie recommendation compared with the traditional CB method.
In order to construct our hybrid method, we first apply the CF method to obtain each user's prediction score for all items. These prediction scores are obtained based on the correlation of the users’ scoring behaviors. However, the main disadvantage of obtaining scores through CF is that user characteristics and movie characteristics are not considered when generating predictions. In order to further improve the prediction results, we divide each user's movie set into two groups. Each user's first movie set contains movies for which the user has a higher predicted score. In other words, it represents the user's positive preference profile. Similarly, each user's second movie set contains movies for which the user has a lower predicted score, which represents the user's negative preference profile. Based on these two profiles, we will recommend movies that are similar to the user's positive profile but different from the user's negative profile.
The advantages of our method include the following: (1) we consider the relevance of the users' scoring behavior; (2) we consider user characteristics and movie characteristics; and (3) we use users' negative profiles to further enhance the recommendation results.
The main contributions of this research are:
- •
We propose a new hybrid method that combines CF and CB.
- •
Our method can build a positive profile and a negative profile for each user.
- •
This is the first paper that considers negative user profiles in CB, so we can delete movies negatively related to user preferences from the recommendation set.
- •
Our experimental results show that considering negative profiles can improve the traditional CB that only considers positive profiles. Based on this finding, all past studies in CB can try to further improve their recommendation results by using negative profiles.
The rest of this paper is organized as follows. In Section 2, we state our research objective. Section 3 is a review of related work. Section 4 introduces a hybrid movie recommendation method based on the user's positive and negative profiles. Section 5 contains a series of experiments to prove the effectiveness of our recommendation algorithm. Section 6 is the conclusion of this paper. Finally, in Section 7 we discuss the implications of this research and future work.
Section snippets
Research objective
Content-based filtering uses user profiles and product content as data, and compares the information contained in the product with the user profile. Content-based methods usually use products purchased in the past to build user profiles. The intuition underpinning the previous approach is that what users buy represents what they like. However, for a certain product, the user may like or dislike it. If we can discern the relationship between users and unrated products, then this negative
Related work
The recommendation system is an information filtering mechanism, mainly used to reduce the extra cost involved in the process of searching for information. Based on the user's preferences, interests, behaviors, or needs, it recommends information, services, or products that the user potentially needs. The most popular recommendation algorithms can be classified as three main categories: content-based, collaborative filtering and hybrid approaches.
The process of content-based filtering is
Recommendation algorithm
In this section, we will outline our recommendation system. Fig. 1 shows the architecture of the entire system. In the system architecture, we have designed two main modules, namely the Collaborative-Filtering module (CF) and the Profile-Based module (PB). The CF module first collects actual rating data from the MovieLens dataset. Since the entire user-movie rating matrix is very sparse, we apply the user-based collaborative filtering algorithm (Herlocker, Konstan, Borchers & Riedl, 2015) to
Experiments
In this section, we first describe how we collected the dataset, and discuss the measurement metric. Then we conducted a series of experiments. Finally, we discuss the experimental results.
Conclusion
This paper proposes a new approach that is different from traditional recommendations, that is, including positive/negative user profiles to obtain more accurate recommendations. In previous studies, recommendation algorithms usually generated recommendation results based on items that users like, items that similar users like, or items that the user previously purchased. In other words, previous research mainly focused on the positive profile to start the research on recommendation. No
Implications of the study and future work
The results of this study indicate that negative profiles (users’ negative preferences) should be included when designing recommendation algorithms. However, it is not easy to identify the user's negative preferences because almost all datasets only store the user's positive feedback, such as click, purchase or browsing. In this paper, we apply the CF algorithm to obtain the user's predicted score for all items, so we can divide the user's preferences into a positive profile and a negative
CRediT authorship contribution statement
Yen-Liang Chen: Conceptualization, Methodology, Supervision, Validation, Writing - original draft, Writing - review & editing. Yi-Hsin Yeh: Data curation, Software, Validation. Man-Rong Ma: Conceptualization, Methodology.
References (35)
- et al.
A survey of topic modeling in text mining
International Journal of Advanced Computer Science and Applications
(2015) - et al.
User tweets based genre prediction and movie recommendation using LSI and SVD
- Beel, J. (2019). And the winner is MovieLens - On the popularity of recommender system datasets....
- et al.
Comparing LDA and LSA topic models for content-based movie recommendation systems
- et al.
Enhanced movie content similarity based on textual, auditory and visual information
Expert Systems with Applications
(2018) - et al.
An improved hybrid recommender system by combining predictions
- et al.
Item-network-based collaborative filtering: A personalized recommendation method based on a user's item network
Information Processing & Management
(2017) - et al.
The movielens datasets: History and context
ACM Transactions on Interactive Intelligent Systems
(2015) - et al.
Interactive recommender systems: A survey of the state of the art and future research challenges and opportunities
Expert Systems with Applications
(2016) - et al.
An algorithmic framework for performing collaborative filtering
ACM SIGIR Forum
(2015)
A novel recommendation model with google similarity
Decision Support Systems
Efficient machine learning model for movie recommender systems using multi-cloud environment
Mobile Networks and Applications
Recommendation systems: Principles methods and evaluation
Egyptian Informatics Journal
Latent Dirichlet allocation (LDA) and topic modeling: Models, applications, a survey
Multimedia Tools and Applications
Hybrid collaborative filtering for high-involvement products: A solution to opinion sparsity and dynamics
Decision Support Systems
Matrix factorization techniques for recommender systems
Computer
Recommender system application developments: A survey
Decision Support Systems
Cited by (30)
Multi-objective optimization and integrated indicator-driven two-stage project recommendation in time-dependent software ecosystem
2024, Information and Software TechnologyNeural TV program recommendation based on dynamic long-short term interest
2023, Applied Soft ComputingNegative Can Be Positive: Signed Graph Neural Networks for Recommendation
2023, Information Processing and ManagementSelection bias mitigation in recommender system using uninteresting items based on temporal visibility
2023, Expert Systems with ApplicationsCitation Excerpt :Secondly, the key is how to identify the items in the missing data that the user is not interested in. Unrated items can be caused by the following two reasons (Chen et al., 2021): The user has not seen the item and does not know its existence so that there is no rating.
Ensemble transfer learning-based multimodal sentiment analysis using weighted convolutional neural networks
2022, Information Processing and ManagementCitation Excerpt :The text2vec model and CNN network were used to extract the features of the texts in this paper, and The SVM was performed for sentiment analysis. Some related studies in the field of scene segmentation (Chen et al., 2020), movie genre classification (Yadav & Vishwakarma, 2020), video activity recognition (Mliki, Bouhlel, & Hammami, 2020) and movie recommendation (Chen, Yeh, & Ma, 2021) were also carried out in this area. Reinforcement learning is one of the important categories of machine learning, which has been utilized in various areas of NLP, including sentiment analysis.