A. Dataset and Features
An analysis of the recommended approach, its performance measures, and results is presented in this section. The hybrid system that is being presented here was created especially for movie recommendations using Tensorflow in Python. The Movielens databases provided a wealth of information that was used in the inquiry. The collection comprises ratings for 62,000 films and 1 million tags from 162,000 individuals, for a total of 25 million entries. It also includes tag genome data with 15 million relevance ratings and 1,129 tags. This dataset, dubbed "ml-25m," combines free-text labeling tasks and 5-star ratings taken from the MovieLens recommendation engine. From January 9, 1995, to November 21, 2019, 162,541 users contributed 1,093,360 tag applications, 62,423 movies, and 25,000,095 ratings to the collection. On November 21, 2019, a random selection of individuals who had each rated at least 20 different movies was used to build the dataset. Nevertheless, demographic information is missing, and users are only identified by their ID. The MovieLens databases provide extensive information about users, movies, ratings, and tags. The proposed model was trained on 75% of the training data and tested on the remaining 25% after the dataset was divided into training and testing sets. The hybrid system-based machine learning approach uses the dataset's information in both stages to recommend movies. User and movie ratings are graphically represented in Figs. 1 and 2, which are important for recommendation generation. The results pertaining to user and movie ratings are presented in Figs. 1 and 2 below, along with suggestions.
November, the eleventh month of the year, has the largest average number of ratings (about 70,000), followed by August (around 50,000) and December (almost 30,000), as can be seen in Fig. 2. However, it's crucial to remember that these numbers are estimates because there isn't much data for 2003—just two months' worth—available. Let's now examine the thorough evaluation of the overall number of ratings received every month for the entire year.
There is a discernible increase in ratings, as seen in Fig. 3, especially in November 2000. As the graph illustrates, this surge—which accounts for nearly 90% of the total ratings—has a major impact on the high average ratings that are seen in November, August, and December. That being said, a negative tendency becomes evident in 2001 and 2002. With only two months' worth of data available for 2003, it is difficult to draw firm conclusions from this pattern.
From here on out, we'll be looking at how different rating levels are distributed in order to learn more about how users generally rate things.
Approximately 350,000 instances of the rating value 4 have been seen, making it the most often observed rating, according to the statistics shown in Fig. 4. In our one million rating dataset, around 35% of the ratings were 4 stars, while approximately 26% and 21% of the ratings were 3 and 5 stars, respectively. It is significant to note that these estimations may include some mistakes due to the limited data available for 2003. However, it is evident that a considerable proportion of users typically provide ratings of four or above.
Every year in the dataset can be subjected to comparable analysis.
Comparable distributions are shown for each year period in Fig. 5. The analysis of the monthly distribution of rating values may then be undertaken.
A distribution resembling that seen in the annual and comprehensive graphs may be seen in Fig. 6. All categories combined have an average rating of about 3.6. After that, we'll examine the ratings' temporal fluctuations and use standard deviation-based techniques to depict the lower and upper bounds.
It is clear that throughout time, the average rating has continuously fluctuated between 3 and 4.
Figure 7 shows that in the late 1990s, there was a lot of interest in the comedy and film noir genres.
Figure 8 illustrates that the Film-Noir and Horror categories consistently demonstrate high and low average ratings, respectively, with occasional extreme values. Following this observation, a density plot depicting ratings by genre will be generated.
In Fig. 9, it is evident that all genres exhibit a left-skewed distribution (with an approximate mean of 3.5), except for the Horror genre, which is characterized by low ratings.
The matrix output generated for movie recommendations is displayed in Fig. 10. In this case, each column represents a different movie ID, and each row represents a different user ID. This matrix, which was created using the planned hybrid system, shows suggestions for the users' preferred films. For each user, a certain movie's recommendation score is represented as a cell in the matrix.
The results of the movie recommendations are shown in Fig. 11. These suggestions are developed by combining conclusions or approximations obtained by combining the hybrid approach with K-means + + and IKSOM. The final product could include reviews, release years, movie names, and other pertinent details.
A variety of performance criteria were employed in order to assess the efficacy of the suggested methodology. These measurements include precision, accuracy, recall, mean absolute error, F1-score, and root mean square error. A detailed description of every performance metric is given below.
The performance metrics for Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are shown in Fig. 12. Using content-driven K-nearest neighbors (KNN) with cosine similarity, the hybrid approach shown in this work addresses the cold start issue and yields an RMSE of 0.410 and an MAE of 0.256. This technique makes rating predictions possible even in situations with limited data. Consequently, the suggested hybrid strategy has the ability to lower errors in RMSE and MAE by resolving issues with data sparsity and cold start.
The approaches outlined below were used to evaluate the suggested hybrid technique's accuracy, recall, and F1-score.
$$Precision=\left(relevent movies recommended \right)/\left(all movie recommended\right)$$
1
$$recall=\left(relevent movies recommended\right)/\left(all possible movies\right)$$
2
$$F1-score=2.(precision.recall)/(precision+recall)\dots ...\left(3\right)$$
The performance metrics for the suggested method are shown in Fig. 13. The proposed hybrid technique, which makes use of content-driven K-nearest neighbors (KNN), IKSOM, and K Means + + clustering, produced impressive results: 91.41% accuracy, 93.09% precision, 93.82% recall, and a 92.44% F1-score. This approach successfully addressed issues with data sparsity and scalability that are frequently present in movie recommendations. A review of the movie recommendations' quality included determining the accuracy, recall, and Root Mean Squared Error (RMSE). The hybrid system, as shown in the picture, attained an RMSE of 0.4140, compared to the prior system's RMSE values of 1.151 for user KNN, 1.044 for item KNN, 1.157 for Slope One, 1.143 for co-clustering, and 1.136 for NMF.A detailed comparison between the suggested approach and some of the techniques currently in use is shown in Table 2.
Table 2
COMPARISON OF PROPOSED METHOD WITH OTHER EXISTING METHOD.
Methods
|
RMSE
|
MAE
|
User based KNN [2]
|
1.151
|
0.889
|
Item based KNN [6]
|
1.044
|
0.819
|
Slope one [8]
|
1.157
|
0.893
|
CO-Clustering [10]
|
1.143
|
0.893
|
Non-Negative Matrix-Factorization-based approach (NMF) [12]
|
1.136
|
0.894
|
Proposed Hybrid method
|
0.410
|
0.256
|
The aforementioned study comes to the conclusion that current content-based filtering algorithms are unsuitable since they are laborious and inefficient when used on a variety of datasets. As a result, a hybrid strategy that combines content screening and collaborative techniques is suggested as a fix. It has been shown that the best approach for rating prediction combines matrix factorization with nearest neighbor selection. Cosine similarity is used to evaluate user similarity based on their rated films and recommend movies with similar users, hence resolving the cold-start issue. In the task of grading recommender systems, the hybrid system provided performs better than previous approaches in both the overall and cold start scenarios across significant evaluation criteria such as RMSE and MAE. High anticipated accuracy is the outcome of the hybrid system's capacity to learn complicated knowledge successfully. Therefore, the suggested hybrid system maintains excellent accuracy, precision, recall, and F1-Score while achieving low RMSE and MAE error rates. The suggested hybrid model offers recommendations for the top N films in the recommendation system that are more accurate than the present method.