1. Introduction
Recommender systems typically analyze user ratings and choice histories to provide tailored suggestions. Recommendation algorithms include content-based filtering, collaborative filtering, deep learning, and hybrid recommendation techniques.
Content-based recommendation techniques analyze the properties of an item and recommend similar items. They recommend content similar to the items that users have used, preferred, or selected in the past. For example, in the context of movie recommendations, similar films are recommended to the user based on the genre, director, and actors of the movies that the user selected in the past. Content-based recommender systems directly reflect individual preferences. However, a drawback is that the accuracy of recommendations may decline in the absence of sufficient information about new items, or if the user lacks a diverse range of past experiences.
On the other hand, collaborative filtering recommendation techniques are based on the choices of other users who have similar preferences to the user. These techniques fundamentally ground their recommendations in measures of user similarity. Specifically, user-based collaborative filtering employs user similarity metrics to generate recommendations, while item-based collaborative filtering relies on item similarity for the same purpose. Deep learning-based recommendation techniques employ deep learning models of neural networks, to predict interactions between users and items for the purpose of making recommendations. Lastly, hybrid recommendation techniques combine the above recommendation methods to offset the limitations inherent in each individual recommender system.
Figure 1 illustrates the process of recommending movies with a conventional recommender system. After categorizing the features of each user’s favored movies, the recommender system computes similarities based on users’ rating matrices.
For example, in
Figure 1, Users 1 and 3 have high user similarity. Therefore, the other movie that User 3 favored are recommended to User 1. When a new user appears, the similarities between the existing users and the new are measured for the recommendation to the new user.
Pearson correlation coefficient and the Cosine similarity are often used to measure such similarities. The Pearson correlation coefficient is a numerical measure indicating the relationship between two compared variables. It is used based on the ratings of items commonly rated by two users, with its formula being as follows:
Here, X and Y represent the ratings of the main user, and are the average ratings of other users, with +1 indicating a positive linear relationship, 0 indicating no linear relationship, and −1 indicating a negative linear relationship. The cosine similarity for two users is defined by the cosine of the angle between two vectors of users, and the formula is,
Here, X and Y represent the rating vectors of two users. The cosine similarity is between −1 and 1, where −1 indicates a completely different direction (i.e., an angle of 180 degrees), 1 indicates the same direction, and 0 corresponds to an angle of 90 degrees. The closer the cosine similarity value is to 1, the higher the similarity is judged to be.
Existing recommendation algorithms have to repeat similar calculations for new and rapidly created information (user or movie) using the Pearson correlation coefficient and Cosine similarity. The reason for repeatedly calculating the similarity for new users is to compare the preferences of the new users with those of the existing users to optimally recommend. This enables fast recommendations difficult due to the time and resources consumed by the iterative calculations required to compare the preferences of new users or items with those of existing users [
1]. The need to calculate user similarities for a recommender system to reflect user preferences and provide personalized recommendations contributes to this issue. Such similarity calculations are time and resource-intensive tasks, with computational complexity increasing as the number of users or items increases. The need to recalculate these similarities each time a new user or item is added results in increased data storage and learning and computation time [
2].
However, the existing techniques has several challenges. First, as the size of the computational matrix expands, computational time quickly grows. This makes it difficult to quickly make recommendations based on new data and results in increased costs. Second, the lack of initial interactions between users and items results in the sparsity, leading to decreased accuracy in recommendations. Third, the Cold Start emerges when new users or items are added to the recommender system. This problem results from the lack of initial information about users or items, making it difficult to make recommendations. Lastly, it is difficult to generate recommendations for inactive users or those not previously included in the recommender system.
This study proposes a user persona-based recommendation technique to address these issues. The user persona-based technique was developed to clearly understand and explain users’ goals and behavioral patterns [
3]. A persona is a virtual character that consumes items and represents actual users, reflecting their characteristics. This technique helps to understand the target user and design products or services that suit that user from a user-centric perspective.
The recommender system implemented in this study used the movie rating data in MovieLens [
4]. The ratings written by users were topic modeled through Latent Dirichlet Allocation (LDA). Then, each user was assigned topics, and these topics were clustered using K-Means Clustering to complete the final user Persona. Collaborative filtering (CF) is divided into memory-based CF and model-based CF [
5]. Memory-based CF calculates the similarity of surrounding users to predict ratings. When the ratings data evaluated by the user is small compared to the total ratings data, the number of common evaluation items among users decreases. This makes it difficult to accurately calculate the similarity between users, leading to a decline in the accuracy of recommendations. Additionally, when new users or rating information is registered in the recommender system, situations in which there are few or no common evaluation items occur. Finally, when there are a small number of common evaluation items among users, the task of calculating the similarity between all users is very inefficient, and the complexity of calculations increases as the number of users and items increases with system expansion. These negatively affect the performances when there are few common evaluation items among users [
6].
To overcome the lack of common evaluation items, we used the Non-negative Matrix Factorization (NMF), a model-based Collaborative Filtering (CF) method. NMF works by decomposing the rating matrix into two lower-dimension submatrices and then recombining them to predict missing ratings. Furthermore, we implemented our recommendation algorithm using a Deep Learning algorithm, one of the machine learning algorithms known for its outstanding performance. This method uses a multi-layer artificial neural network to simultaneously perform feature extraction and recommendation modeling. It learns the information of interactions between the user and the item, thereby providing accurate recommendations. We compared the movie recommendation information using the recommender system implemented through these two methods. The results showed that the system using the Deep Learning algorithm demonstrated a higher accuracy and performance [
Table 1].
The NMF recommendation algorithm works by extracting latent factors from the user and item matrix. It represents each user and item as an n-dimensional vector and factorizes it into two lower-dimensional vectors. This factorization allows each element of the given user-item matrix to be represented as the dot product of the user vector and the item vector. These factorized vectors are used to recommend new items or analyze user preferences [
7]. The NMF method is especially effective when there are few features or factors within the user-item matrix. Furthermore, NMF can be easily updated even when new items or users are added, without recalculating the existing data since it extracts latent factors within the data. NMF has a constraint that the factorized matrices are non-negative (0 or above). This prevents predicting user-item preferences below 0, providing more accurate recommendations. Secondly, NMF shows that the factorized matrices represent the latent features of users and items. This allows easy interpretation of what each feature means, enhancing the performance of the recommender system and increasing understanding of user-item interaction. Lastly, it reduces the dimensionality of the data. Since NMF factorizes into a low-rank matrix, it reduces the original high-dimensional data to a lower dimension. This simplifies complex data for faster processing during analysis.
Recommender systems using Deep Learning have various features. Firstly, they can utilize not only rating data but also various other data types. Conventional recommender systems primarily used item rating data evaluated by the user. However, recommender systems using Deep Learning can use not only rating data but also various data such as click history, search history, purchase history, review content, and activities on social media. Secondly, they can be used when the data contains many abstract and complex features such as the movie’s genre, director, and actors, which are difficult to express in simple numbers or categorical data in a movie recommender system. Deep Learning can easily learn such abstract information. Thirdly, recommender systems using Deep Learning can provide more precise recommendations using various information about users and items. For example, it can recommend movies based on the information about the user’s preferred genre, actors, and directors.
In programs that do not apply topic modeling and personas, the approximate recommendation time is about 10 s with an accuracy of 10%. However, with our modeling, the time is reduced to about 3 s, and the accuracy increases by 40%. Additionally, our model enables responsive movie recommendations based on changes in user personas.
All in all, recommender systems using Deep Learning can use more diverse data, easily learn data with abstract and complex features, and provide more precise and better recommendations, compared to traditional recommender systems.
4. Discussion
Through topic modeling using LDA, the recommender system condensed the complex preferences, interests, and favored characteristics of individual users to a lower-dimensional representation; then, the system captured the essence and tendencies of users’ inherent characteristics by completing user personas through k-means clustering based on the results of user topic modeling.
We elucidate the methodologies implemented to mitigate the cold start issue.
Intentional data input through multi-stage recommendations or tournament-style selections, encouraging users to input their preferences. This allows the system to discern user preferences and construct a persona, which is then used to recommend items corresponding to that persona.
Prompting users to create an initial profile at sign-up, which serves as a foundational persona for personalized recommendations.
Generation of provisional user data by assigning popular items or highly rated items as initial data for new users.
Creation of hybrid user data by combining strategies (1) and (3), thus forming a more robust initial user data set.
The essence of the cold start problem lies in the lack of data. To tackle this, we have devised a system where new users are encouraged to fill out initial profiles upon entry, which are then made to be persona. This approach fundamentally resolves the data scarcity issue inherent to the cold start problem, as it allows the system to generate relevant recommendations even for users without historical data.
With user personas, NMF and DL models were implemented to create a movie recommendation service. A comparison of the movie recommendation results was made between NMF and DL models, with and without the use of user personas, using NMF models without personas as the baseline. The results showed that the use of user personas enabled faster recommendations even with an increasing number of users but with a relatively lower precision. A performance improvement in the NMF and DL models with user personas was required. To evaluate the models, we used MSE as a metric, and the NMF model with personas demonstrated lower performance compared to the NMF model without personas in terms of @Precision Top 10, with scores of 94.16% and 92.27%, respectively. As N increased (e.g., Top 50), the performance gap between the two models widened (N-Persona 90.99%, Persona 67.36%). This change is from transforming the relationship between users and ratings to the relationship between personas and ratings. For example, assuming 100 users and that each user has a persona value of 20, the data for 100 users can be compressed to 20. However, obtaining precise recommendations for all users becomes infeasible due to the creation of a matrix with overlapping user and movie information. Regarding the DL model, the one with personas outperformed the model without personas, albeit marginally. With 70 topics and 35 personas, the NMF model achieved a precision@K performance of 65.11% for 100 users, while the DNN (Deep Neural Network) model achieved 80.52% [
Figure 20].
With 200 users, under the same conditions, the NMF model’s precision@K performance improved to 80.01%, and the DNN model’s performance improved to 92.83%, with both showing improvements of approximately 15% and 10%, respectively. [
Figure 21].
The experiment with 50 topics and 35 identical Personas showed that the NMF’s precision@K performance was measured at 86.01% and DNN’s precision@K performance was measured at 92.67% with 500 user data [
Figure 22].
When measured with 50 topics and 45 Persona under the identical conditions and 900 users, NMF’s precision@K performance was 97.04%, and DNN’s precision@K performance was 95.55%, showing performance improvements of 10% and 5% or more, respectively [
Figure 23].
These results indicate that the performance improves as the amount of movie rating data and user data increases, allowing it to adapt to changes in user personas. By managing data according to the number of user personas, storage space can be saved, and learning time can be reduced. Additionally, computation costs and time decrease since model updates only occur when user personas change.
The process of creating user personas involves analyzing user behaviors, preferences, habits, and more to form groups with common characteristics. This approach defines groups of users showing similar behavioral patterns or preferences as personas and utilizes them as input for the recommender system. User personas thus created are employed to recommend items that the corresponding user groups are likely to prefer.
Furthermore, merging user personas with other user data enables more sophisticated user modeling. For instance, by jointly modeling user demographic information (age, gender, occupation, etc.), interaction data (movie ratings, click history, etc.), and user behavior patterns that change over time, more personalized recommendations become possible. Integrating user personas with other user data allows for providing recommendations that reflect both individual user characteristics and common characteristics within groups. This surpasses conventional generic recommender systems, enabling highly personalized recommendations and enhancing user satisfaction. It is expected that user personas created in this manner can be utilized in other recommendation services and analyses as well.
Typically, the performance of models improves with the increase in data volume. Our model is no exception to this. For instance, in programs that do not apply topic modeling and personas, the recommendation time is approximately 10 s with an accuracy of 10%. However, when our modeling approach is applied, the recommendation time is significantly reduced to about 3 s, and accuracy is enhanced to 40%.
We believe that these enhancements in our model will allow for more dynamic and adaptable user profiling, effectively addressing the broad spectrum of user interests and their evolution over time.
5. Conclusions
We demonstrated the movies recommendation system based on user personas when a new user is added. However, it is essential to consider the situation when new movies are added. In addition to user personas, creating movie personas presents a novel approach. Movie personas are based on metadata, such as genre, director, cast, and release year. Each movie possesses a unique persona, enabling more sophisticated analysis for movie recommendations. Establishing a model that connects movie personas with user personas makes it possible to identify the movie personas preferred by users and recommend new movies accordingly. For instance, if a specific user persona exhibits a high preference for a particular movie persona, recommending other movies with similar personas becomes feasible. This approach significantly enhances the adaptability to newly added movies. Such an approach finds applications in various domains beyond movie recommendation services. For instance, creating personas for exhibitions or tourist attractions and linking them with user personas is achievable. This enables recommending the most suitable exhibitions or tourist attractions to users by aligning their preferences with the characteristics of the attractions, leading to increased user satisfaction and improved service quality.
When the input format remains the same, our method can be applied to various models. For example, if the previous model used was NMF, it can be replaced with a deep learning model using the same input data. This allows leveraging the strengths of each model and integrating results from different algorithms to achieve more precise predictive capabilities. This approach reminds one of the principles of ensemble learning, which combines multiple models to achieve better performance by allowing each model to approach the given problem from different perspectives. Consequently, using a combination of different models through replacement or ensemble methods can enhance insights and accuracy that were not attainable with a single model alone. This approach proves to be highly useful in improving the performance of recommender systems and can also be applied in the research and development of new recommendation algorithms. In this study, a recommendation algorithm was implemented, generating user personas based on user-rating-movie data to recommend movies to users.
Figure 24 illustrates the overall configuration of a user persona and movie persona recommender system. In the future, a movie-centric user recommendation algorithm can be implemented by creating movie personas based on movie-rating-user data. Moreover, we can employ ensemble models such as NMF, DL, Graph Convolutional Network (GCN), among others, to link these two algorithms to constitute the recommendation algorithm.
The topics of Artificial Neural Networks (AN) and Deep Reinforcement Learning (DRL) are not covered within the scope of this paper. Future research endeavors are planned to extend into the realm of Generative Adversarial Networks (GANs), aimed at processing unstructured data such as images and videos, not just text. Moreover, advancements in accuracy and adaptability will be explored through the application of DRL, which will be a subject of subsequent investigation