Effective movie recommendation based on improved densenet model

Abstract

In recent times, recommendation systems provide suggestions for users by means of songs, products, movies, books, etc. based on a database. Usually, the movie recommendation system predicts the movies liked by the user based on attributes present in the database. The movie recommendation system is one of the widespread, useful and efficient applications for individuals in watching movies with minimal decision time. Several attempts are made by the researchers in resolving these problems like purchasing books, watching movies, etc. through developing a recommendation system. The majority of recommendation systems fail in addressing data sparsity, cold start issues, and malicious attacks. To overcome the above-stated problems, a new movie recommendation system is developed in this manuscript. Initially, the input data is acquired from Movielens 1M, Movielens 100K, Yahoo Y-10-10, and Yahoo Y-20-20 databases. Next, the data are rescaled using a min-max normalization technique that helps in handling the outlier efficiently. At last, the denoised data are fed to the improved DenseNet model for a relevant movie recommendation, where the developed model includes a weighting factor and class-balanced loss function for better handling of overfitting risk. Then, the experimental result indicates that the improved DenseNet model almost reduced by 5 to 10% of error values, and improved by around 2% of f-measure, precision, and recall values related to the conventional models on the Movielens 1M, Movielens 100K, Yahoo Y-10-10, and Yahoo Y-20-20 databases.

Keywords

Deep neural network DenseNet model Min-Max normalization technique movie recommendation sentiment analysis

1. Introduction

Recently, the recommendation system has become one of the emerging research topics, which learns an individual’s preferences for developing effective recommendations [1]. The recommendation systems are implemented in several applications such as electronic-product recommendations, movie recommendations, song recommendations, book recommendations, etc. [2, 3, 4]. The purpose of the recommendation systems is to automatically recommend news, web pages, movies, e-commerce products, songs, etc. for individuals based on their historical preferences [5]. In recent times, the movie recommendation gained more attention among the researcher’s communities, due to the extensive growth of online multimedia platforms [6]. The movie recommendation system helps users to access their preferred movies from large movie libraries. In recent decades, recommendation systems are categorized into 4 types hybrid systems, knowledge-based systems, content-based systems, and collaborative filtering systems [7, 8]. Compared to other recommendation systems, collaborative filtering systems are effective in predicting the users’ ratings based on the same users’ preferences [9, 10]. Most of the prior collaborative filtering systems first find the same users, and further, predicts the movie ratings based on users’ preferences and their prior ratings [11, 12, 13].

The prior collaborating filtering systems are affected by cold-start, data sparsity, and scalability issues [14, 15]. The accuracy of recommendation is reduced, when the users-items interaction matrices become sparse and it creates sparsity and scalability concerns. To highlight the above-stated issues, a novel movie recommendation system is introduced in this manuscript. After acquiring the input data from the Movielens 1M, Movielens 100K, Yahoo Y-10-10, and Yahoo Y-20-20 databases, the Min-Max normalization technique is utilized for rescaling the collected data. The rescaled data have limited outliers and redundant data which helps in improving the accuracy of movie recommendations. Lastly, the rescaled data are fed to the improved DenseNet model for relevant movie recommendations. In the experimental segment, the proposed improved DenseNet models efficiency is analyzed in light of Root Mean Square Error (RMSE), Mean Absolute Error (MAE), f-measure, MSE, precision, and recall values on Movielens 1M, Movielens 100K, Yahoo Y-10-10 and Yahoo Y-20-20 databases. Simulation results confirmed that the improved DenseNet model has shown a 5% to 10% enhancement in error values, and a 2% improvement in the f-measure, precision, and recall values related to the existing models.

The organization of this manuscript is specified as follows: recent papers related to the movie recommendations are given in Section 2. In addition, Section 3 describes the improved DenseNet model with pseudocode, and its results are denoted in Section 4. The conclusion of the present study is represented in Section 5.

2. Literature review

In this section, a few articles related to movie recommendation are efficiently surveyed. Jain et al. [16] implemented an enhanced multi-stage user based Collaborative Filtering (CFs) technique for effective movie recommendation. In this literature study, the developed technique’s performance was validated on two online databases movielens 100K and movielens 1M. The developed technique attained superior recommendation results compared to the existing competing techniques. Generally, the CF technique has a cold start issue in the recommendation systems. In addition, Zou et al. [17] presented a two-stage recommendation system based on multi-objective teaching learning optimizer and improved CF technique. The experiments conducted on the movielens databases show the effectiveness of the developed system in the personalized recommendation system. The presented system obtains a set of recommendation lists for one target user where it spends more running time for all target users.

Singh et al. [18] utilized the Apriori algorithm for creating user profiles by utilizing the categorical attributes and item ratings, and the created user’s profiles comprise the details of categorical properties of the objects. The effectiveness of the developed recommendation system was evaluated on the movielens databases. The comparative results show that the developed system outperformed the traditional CF techniques on the movielens databases in terms of prediction accuracy. The low expandability and the data sparsity were the major problems faced in this literature study. On the other hand, Singh et al. [19] combined K-Nearest Neighbor (KNN) and cosine similarity functions for effective movie recommendation. However, traditional machine learning techniques such as KNN suffer from a cold start problem.

Li et al. [20] implemented a hybrid recommendation system by integrating both users’ interests and movie features for calculating the similarity between users. In this literature study, the user’s rating matrix and the movie’s feature were integrated for generating the user’s interest vectors. Next, the user’s rating matrix and interest vectors were integrated for generating the hybrid user’s interest vectors to calculate the similarity among users. The experimental result confirmed that the presented recommendation system attained higher recommendation results related to the existing systems, and it relieves the issues created due to data sparsity and by changing the user’s interest.

Vilakone et al. [21] used an improved k-clique algorithm for effective movie recommendation. The presented improved k-clique algorithm’s effectiveness was evaluated on the Movielens databases, where the developed algorithm attained maximum performance compared to the existing CF techniques, k-clique algorithm, KNN etc. Whereas, the improved k-clique algorithm suffered from scalability and sparsity issues. Further, Widiyaningtyas et al. [22] developed a similarity technique: user profile correlation-based similarity for relevant movie recommendations. The extensive experimental evaluations showed that the developed similarity technique outperformed the existing techniques in light of MAE, RMSE, and recommendation accuracy, but it has a system scalability issue that needs to be addressed as a future extension.

Kharita et al. [23] implemented a new item based CF for movie recommendation in real time. As denoted earlier, in the recommendation systems, the CF technique has a cold-start issue. On the other hand, Ali et al. [24] created a hybrid movie recommendation framework based on a content based filtering technique and genomic movie tags. In this literature study, the Pearson correlation method and principal component analysis were utilized for reducing the redundant tags which show low variance proportion. The removal of redundant movie tags helps in decreasing computational complexity. The experimental results prove that the hybrid movie recommendation system identifies similar types of movies compared to the existing systems but it suffers from inherent problems: data sparsity, poor scalability, and cold start.

Lin and Chi [25] combined the CF technique and neural network for developing an effective movie recommendation system. Zhang and Mao [26] introduced a model named Markovian factorization for relevant movie recommendations. The experiments conducted on the Movielens databases confirmed that the presented model achieved better recommendation performance than the standard factorization models. Hence, the textual information related to the rating leads to early-voter and cold-start problems. Vilakone et al. [27] implemented an effective personalized movie recommendation system based on k-clique algorithm, which attained better recommendation results on the movielens databases. As specified earlier, the k-clique algorithm suffered from scalability and sparsity concerns.

Kumar and Prabhu [28] integrated fuzzy C means clustering algorithm and fire-fly optimizer for relevant movie recommendations as per user’s request. In this study, the fire-fly optimizer selects and initializes the cluster’s position and then, the fuzzy C means classifies the similarity of the users’ rating. The evaluation measures such as recall, precision, and MAE confirmed the effectiveness of the developed model on the movielens databases. In addition, the presented model was more sensitive during the selection of the initial cluster center.

Vilakone et al. [29] integrated normalized discounted cumulative gain and k-clique to recommend relevant movies. This model works based on the movie rating and user’s personal information. The implemented model’s performance was compared with the prior successful models to assess its effectiveness but the implemented model was computationally expensive. Behera et al. [30] implemented a hybrid movie recommendation system based on restricted Boltzmann machine and KNN. The developed system’s effectiveness was validated on the movielens databases. As mentioned above, the machine learning techniques like KNN suffer from a cold start problem.

Zarzour et al. [31] integrated hypergraph partition technique and expectation maximization scheme with the CF technique for movie recommendation. The presented hybrid recommendation system was tested on the real-world movielens databases in terms of precision, RMSE, F1-score, accuracy, and recall. The presented model has a higher computational time and it was the main concern in this literature. Alhijawi and Kilani [32] implemented a genetic algorithm-based movie recommendation system, which works based on historical data rating and semantic information. In future work, the implemented system will be tested on larger sized imbalanced databases and more features are considered and tested apart from genre features.

Gupta and Kant [33] created a multi-criteria recommendation system based on genetic algorithms. The experiments conducted on the movielens and Yahoo databases demonstrated the effectiveness of the developed recommendation system. Whereas, the developed system’s performance was validated in light of precision, f-measure, coverage, and accuracy. The presented recommendation system does not deal with sparsity issues, and it was the main issue in this study.

In order to address the above-mentioned issues, a new deep learning based movie recommendation system is implemented in the present research manuscript.

3. Improved DenseNet model for movie recommendation

In recent decades, web expansion has brought huge user’s convenience, but it leads to an issue of information overload. A recommendation system is an effective tool that helps in solving the overload issue by suggesting relevant information to the users. In recent times, the recommendation system is deployed in a variety of online systems like hotels, news articles, online social networks, books, songs, movies, online videos, etc., due to the intense growth of the internet. Usually, the recommendation systems are implemented based on numerous filtering techniques like CF, Content based Filtering (CBF), Hybrid-Filtering, etc. [34, 35]. However, the filtering techniques create effective recommendations only for similar users. Therefore, a new deep learning (improved DenseNet model) based movie recommendation system is implemented in this manuscript, and the flow diagram of the proposed system is depicted in Fig. 1.

Figure 1.

Flow diagram of the proposed system.

3.1 Database description

The proposed improved DenseNet model’s effectiveness is analysed on four benchmark databases as movielens 1M, movielens 100K, Yahoo Y-10-10, and Yahoo Y-20-20. The Yahoo! movies database comprises 976 movies and 6078 users, and every movie is classified based on five criteria like visuals, direction, acting, overall rating and story. The movie rating is varied from A $+$ to F, where it has a 13 rating scale and then the overall rating is normalized into a five point rating scale which ranges from 1 to 5. The users who have given ratings to at least 10 movies are in the Y-10-10 database. In addition, the users who have given ratings to at least 20 movies are in the Y-20-20 database. After performing the extraction, a total of 491 movies, 429 users, and 18,504 ratings are present in the Y-20-20 database.

The movielens 1M database has 6040 users, 3900 movies and 1,000,209 ratings, and also it has demographic information such as user’s occupations (executive/managerial, craftsman, artist, customer service, 12 ${}^{\text{th}}$ student, farmers, home-maker, self-employed, lawyer, academic/educator, college/grade student, scientist, programmer, doctor/health care, retirement, marketing, unemployed, clerical, writer, and engineer/technician), movie genres (thriller, animation, Sci-fi, fantasy, children, comedy, horror, action, documentary, mystery, adventure, western, war, romance, crime, drama, film-noir, and musical) and users’ age (56 $>$ , 50–55, 45–49, 35–44, 25–34, 18–24, and $<$ 18). On the other hand, the movielens 100K database is recorded between the time periods of 19 ${}^{\text{th}}$ September 1997 to 22 ${}^{\text{nd}}$ April 1998 and the movielens 100K database has 943 users, 1682 movies and 100000 ratings. In addition, the movielens 100K database has the user’s demographic information such as user’s gender, occupation and user’s age. Here, each user should rate at-least 20 movies and need to complete their demographic information, or-else, the respective user will eliminate from the database.

3.2 Data denoising

After the acquisition of Movielens 1M, Movielens 100K, Yahoo Y-10-10, and Yahoo Y-20-20 databases, the data normalization is accomplished by using Min-Max normalization technique. In this research, the Min-Max normalization technique performs an effective linear transformation on the acquired data and it rescales the data between the ranges of zero to one. This procedure helps in preserving the linear relationships among the collected data variables. The mathematical presentation of the Min-Max normalization technique is stated in Eq. (1) [36].

$\displaystyle X_{\textit{norm}}=\frac{X-X_{\min}}{X_{\max}-X_{\min}}$ (1)

Where, $X_{\textit{norm}}$ indicates normalized data, $X_{\min}$ represents minimum rescaling value and $X_{\max}$ indicates maximum rescaling value.

3.3 Movie recommendation

After rescaling the acquired data, the rating criteria (0 to 5) are calculated for every user and movie, and the overall rating criteria are computed for all users and movies. In this manuscript, the improved DenseNet is used to predict the rating of target users for a recommendation. The DenseNet is an effective deep-learning model, where every layer is directly-connected with each other for achieving better information flow. In the DenseNet model, each layer collects inputs from the previous layers, and then, transfers them as the feature maps to the subsequent layers. The feature map of the current layer is combined with prior layers by performing a concatenation process. In a network, each layer linked with all the prior layers is named as DenseNet, which needs lower parameters compared to the traditional models. Additionally, the DenseNet decreases the overfitting issue, which occurs in the training sets [37].

Initially, the denoised data are passed to the improved DenseNet model, which includes $N$ layers where each layer executes with a non-linear transformation function $F_{n}(.)$ . The input feature maps 0 to $n-1$ are concatenated to generate $x_{0},\ldots,x_{n-1}$ values, which are initially fed to the embedding layer to embed feature maps and learned during training. The improved DenseNet model has $N({N+1})/2$ connections, and $n^{\text{th}}$ layer output is stated in Eq. (2).

$\displaystyle x_{n}=F_{n}({[{x_{0},\ldots,x_{n-1}}]})$ (2)

Where, $x_{n}$ indicates present $n^{\text{th}}$ layer, $([{x_{0},\ldots,x_{n-1}}]$ represents concatenated feature maps, and $F_{n}(.)$ indicates a batch normalization with Rectified Linear Units (ReLU) function. If the feature map size is changed, the concatenation operation is not possible in DenseNet, and the layers with dissimilar feature map sizes are down-sampled. The transition layers contain $2\times 2$ average pooling operations and $1\times 1$ convolution layer, where the initial convolution layer has $7\times 7$ convolution blocks with a maximum stride of 2 and a pooling layer size of $3\times 3$ . After the Dense-convolution block, the classification layer consists of a softmax classifier and global average pooling.

In the DenseNet model, the convolution operation learns the data features with filtering techniques. After performing the convolution operation, a ReLU activation function is applied to the output feature maps, and it is mathematically depicted in Eq. (3). In addition to this, the pooling operation is accomplished for decreasing the output feature maps dimensionality, and it is performed either by utilizing average pooling or max-pooling operation. The average pooling operation partitions the input into different pooling areas and then estimates the mean value. Similarly, the max-pooling operation takes the largest elements from the feature maps. The global average pooling function calculates the mean of every feature map and further, the resultant feature vectors are fed to the softmax layer.

$\displaystyle f({x_{0}})=\max({0,x_{0}})$ (3)

In the proposed model, the classification layer has a fully-connected softmax layer, and it sets neurons based on the attributes in the acquired databases like movielens 1M, movielens 100K, Yahoo Y-10-10 and Yahoo Y-20-20. The softmax function is used to categorize multi-class classification issues by computing the probability distributions of every class $i$ , as stated in Eq. (4).

$\displaystyle S({y_{i}})=\frac{e^{y_{i}}}{\mathop{\sum}\nolimits_{j}e^{y_{j}^{% \prime}}}$ (4)

Where, $y_{j}$ represents all input values of $I$ and $y_{i}$ indicates input value. Though, the sum of the exponential input data values and input element is computed using Eq. (4). The class imbalance is the main concern, while processing large databases like movielens 1M, movielens 100K, yahoo Y-10-10 and yahoo Y-20-20. These databases have enormous samples for some classes (romance, crime, horror, action etc.) and few samples in some classes (fantasy, Sci-fi, war drama, musical, film-noir, documentary etc.). To address a class imbalance issue, a weighting factor $W_{i}$ and class-balanced loss functions are included with the DenseNet model. The weighting factor $W_{i}$ is mathematically specified in Eq. (5), and the efficient number of samples in each class is indicated as $S_{n_{i}}$ and it is depicted in Eq. (6).

$\displaystyle W_{i}\propto\frac{1}{S_{n_{i}}}$ (5) $\displaystyle S_{n_{i}}=({1-B_{i}^{n_{i}}})/({1-B_{i}})$ (6)

Where, Bias $B=({I-1})/I$ and $I$ indicates possible instances in each class and it is mathematically specified in Eq. (7).

$\displaystyle I=\mathop{\lim}\limits_{n\to\infty}\mathop{\sum}\limits_{i=1}^{n% }B^{i-1}=1/({1-B})$ (7)

In the proposed model, an Adaptive Learning Rate Optimization Algorithm (ADAM) is utilized for updating weights based on acquired data. The ADAM optimizer identifies a learning rate for individual parameters. The ADAM optimizer utilizes first and second-order-gradient moments for adjusting the learning rate of individual weight in the DenseNet model, which is known as adaptive moment estimation. By utilizing increased moving averages, the ADAM optimizer evaluates the gradient moments, and the moving averages are computed using the present mini-batch, as mentioned in Eqs (8) and (9).

$\displaystyle a_{t}=\beta_{1}a_{t-1}+({1-\beta_{1}})g_{t}$ (8) $\displaystyle b_{t}=\beta_{2}b_{t-1}+({1-\beta_{2}})g_{t}^{2}$ (9)

Where, $\beta_{1}$ and $\beta_{2}$ represents decay rates, $a$ and $b$ indicate moving averages, and $g$ states gradients in the present mini-batch. The log loss or cross entropy loss function is used for assessing the effectiveness of the prediction model. Hence, the cross entropy loss function increases, while the predicted probability deviates from the real class labels. The cross-entropy loss function is mathematically specified in Eq. (10).

$\displaystyle\textit{Cross entropy}=-\mathop{\sum}\limits_{i}^{C}t_{i}\log({s_% {i}})$ (10)

Where, $t_{i}$ indicates ground-truth, $s_{i}$ denotes the score of every class $i$ in $C$ , and $C$ indicates all classes in each database. Hence, the Categorical Cross Entropy loss function (CCE) and Class Balanced Cross Entropy loss function (CBCE) are mathematically indicated in Eqs (11) and (12).

$\displaystyle\textit{CCE}({f,y})=-\log\left({\frac{\exp({f_{y}})}{\mathop{\sum% }\nolimits_{i=1}^{C}\exp({f_{i}})}}\right)$ (11) $\displaystyle\textit{CBCE}({f,y})=\frac{1-B}{1-B^{n_{y}}}\log\left({\frac{\exp% ({f_{y}})}{\mathop{\sum}\nolimits_{i=1}^{C}\exp({f_{i}})}}\right)$ (12)

Figure 2.

Architecture of the improved DenseNet model.

Where, $n_{y}$ denotes training samples, and $y$ represents class labels (movie ratings 0 to 5). The hyper-parameters of the DenseNet model are: learning rate is 0.001, batch size is 64, epoch is 100, grow rate is 12 and depth is 100. Pseudocode of the proposed recommendation system is given below, and its architecture is given in Fig. 2.

Pseudocode of the proposed recommendation system

Input: User data, items data and ratings data with timestamp, cold start user, and time-stamp of the user is $t$

Output: The list of recommended movies of user is $R L$

Pre-process the input data

Read pre-processed data and set parameters for the neural network: improved DenseNet model

Read user features and movie features from data frame

Construct similarity calculation for users and construct a graph

K fold split of data is done and the train

Save the trained model and parameters

Load the model and recommend for the cold start user according to timestamp $t$

10.

Recommend list of the movies based on multi-criteria

11.

Return $R L$

4. Experimental results and discussion

In this research manuscript, the proposed improved DenseNet model is simulated using a python software environment on a computer with Intel core i5 processor, windows 10 (64-bit) operating system, 8 GB random access memory and NVIDIA GeForce GT 730 graphics card. The proposed improved DenseNet model’s efficacy is tested on the movielens 1M, movielens 100K, Yahoo Y-10-10, and Yahoo Y-20-20 databases in light of RMSE, MAE, f-measure, MSE, precision, and recall. The mathematical formula of the RMSE, MAE and MSE is stated in Eqs (13–(15). Where, $n$ represents some observations, $Y_{i}$ denotes actual variables, and $\hat{Y}_{i}$ represents predicted variables.

$\displaystyle\textit{MAE}=\frac{1}{n}\mathop{\sum}\limits_{i=1}^{n}|{Y_{i}-% \hat{Y}_{i}}|$ (13) $\displaystyle\textit{MSE}=\frac{1}{n}\mathop{\sum}\limits_{i=1}^{n}({Y_{i}-% \hat{Y}_{i}})^{2}$ (14) $\displaystyle\textit{RMSE}=\sqrt{\frac{1}{n}\mathop{\sum}\limits_{i=1}^{n}({Y_% {i}-\hat{Y}_{i}})^{2}}$ (15)

The mathematical formula of recall, precision, and f-measure is represented in Eqs (16)–(18). Where, TP, TN, FP and FN indicate true positives, true negatives, false positives and false negatives.

$\displaystyle\textit{Recall}=\frac{TP}{TP+FN}\times 100$ (16) $\displaystyle\textit{Precision}=\frac{TP}{TP+FP}\times 100$ (17) $\displaystyle\textit{F-measure}=\frac{2TP}{2TP+FP+FN}\times 100$ (18)

4.1 Quantitative evaluation

The experimental results of the improved DenseNet model on the movielens 1M and 100K databases in light of RMSE, MAE, and MSE are specified in Table 1. By investigating Table 1, the experimental outcomes are analyzed with two different cross fold validations ( $K=$ 2 and 5) in both single and multi-criteria. The single criteria recommendation systems models a user’s utility for an item as a single preference rating. In multi-criteria systems, the users provide subjective preference ratings on multiple attributes of an item. Further, the results are indicated for different epochs such as 50, 100, 150, 200, and 250. As represented in Table 1, the improved DenseNet model achieved a minimum error rate in both scenarios: single and multi-criteria on the movielens 1M and 100K databases. The graphical representation of the improved DenseNet model’s results on the movielens 1M and 100K databases are represented in Figs 3 and 4, correspondingly.

Table 1
Experimental results of the improved DenseNet model on the movielens 1M and 100K databases by using different evaluation metrics

Movielens 1M database
Epochs	$K=$ 2						$K=$ 5
	Single criteria			Multi criteria			Single criteria			Multi criteria
	MAE	MSE	RMSE	MAE	MSE	RMSE	MAE	MSE	RMSE	MAE	MSE	RMSE
50	0.69	0.95	0.97	0.68	0.89	0.94	0.68	0.93	0.96	0.67	0.88	0.93
100	0.73	1.04	1.02	0.69	0.92	0.96	0.71	1.03	1.02	0.67	0.90	0.94
150	0.71	1.02	1.04	0.69	0.94	0.96	0.69	0.99	0.98	0.68	0.92	0.95
200	2.53	7.70	2.77	0.70	0.94	0.97	2.51	7.69	2.76	0.68	0.93	0.95
250	2.53	7.67	2.77	0.68	0.94	0.97	2.52	7.66	2.75	0.67	0.93	0.95
Movielens 100K database
50	0.68	0.94	0.96	0.68	0.89	0.93	0.67	0.93	0.96	0.66	0.87	0.92
100	0.72	1.04	1.01	0.68	0.91	0.95	0.70	1.02	1.01	0.67	0.90	0.93
150	0.70	1.02	0.99	0.69	0.93	0.96	0.69	0.99	0.98	0.67	0.91	0.95
200	2.53	7.69	2.76	0.69	0.93	0.96	2.51	7.68	2.75	0.67	0.92	0.95
250	2.52	7.67	2.76	0.68	0.93	0.96	2.51	7.66	2.74	0.66	0.92	0.94

Figure 3.

Graphical comparison of the improved DenseNet model on movielens 1M database.

Similarly, the experimental results of the improved DenseNet model on the Yahoo Y-10-10 and Yahoo Y-20-20 databases in light of RMSE, MAE, and MSE are denoted in Table 2. By viewing Table 2, the improved DenseNet model achieved a minimum error rate in the $K=$ 5 fold cross-validation (80:20% training and testing) compared to $K=$ 2 fold cross-validation. By performing different cross-fold validation, the proposed models’ computational time, variance and bias are reduced. The improved DenseNet model consumes 34.50, 35.22, 21.06 and 30.2 seconds of computational time on the movielens 1M, movielens 100K, Yahoo Y-10-10, and Yahoo Y-20-20 databases, which are better related to the traditional models. The graphical comparisons of the proposed improved DenseNet model on the Yahoo Y-10-10, and Yahoo Y-20-20 databases are denoted respectively, in Figs 5 and 6.

Table 2

Experimental results of the improved DenseNet model on the Yahoo Y-10-10 and Yahoo Y-20-20 databases by using different evaluation metrics

Yahoo Y-10-10 database
Epochs	$K=$ 2						$K=$ 5
	Single criteria			Multi criteria			Single criteria			Multi criteria
	MAE	MSE	RMSE	MAE	MSE	RMSE	MAE	MSE	RMSE	MAE	MSE	RMSE
50	0.68	0.94	0.96	0.67	0.88	0.92	0.67	0.93	0.95	0.66	0.87	0.92
100	0.71	1.03	1.01	0.67	0.90	0.94	0.70	1.02	0.99	0.66	0.89	0.94
150	0.69	0.99	0.98	0.68	0.92	0.95	0.69	0.98	0.98	0.67	0.91	0.95
200	2.52	7.69	2.75	0.69	0.92	0.95	2.51	7.68	2.75	0.67	0.92	0.95
250	2.51	7.66	2.75	0.67	0.92	0.95	2.51	7.65	2.75	0.66	0.92	0.94
Yahoo Y-20-20 database
50	0.67	0.93	0.95	0.66	0.87	0.92	0.67	0.93	0.95	0.66	0.87	0.92
100	0.71	1.02	1.01	0.66	0.90	0.93	0.70	1.02	1.01	0.67	0.90	0.93
150	0.68	0.99	0.97	0.67	0.92	0.94	0.69	0.98	0.98	0.67	0.91	0.94
200	2.51	7.68	2.75	0.68	0.92	0.95	2.51	7.68	2.75	0.67	0.92	0.95
250	2.51	7.65	2.75	0.66	0.91	0.94	2.51	7.65	2.75	0.66	0.92	0.94

Figure 4.

Graphical comparison of the improved DenseNet model on movielens 100K database.

4.2 Comparative evaluation

In this phase, the proposed improved DenseNet model’s effectiveness is compared with the existing works developed by Behera et al. [30] and Gupta and Kant [33]. Behera et al. [30] developed a hybrid movie recommendation system based on KNN and Boltzmann machine. The developed model’s performance is investigated on the movielens 1M and 100K databases by means of RMSE and MAE, and the comparative results are depicted in Table 3. In addition, Gupta and Kant [33] created a Multi-Criteria Recommendation System (MCRS) based on a genetic algorithm and its Credibility Score (CS). The developed MCRS-CS model has achieved higher recommendation results by using the evaluation metrics like f-measure, precision and recall on the Y20-20 and Y10-10 databases, and the results are depicted in Table 4. By investigating Tables 3 and 4, the improved DenseNet model achieved a minimum error rate and maximum recommendation results related to the existing models.

Table 3
Comparative results by means of RMSE and MAE

Models	Database	RMSE	MAE
KNN with Boltzmann [30]	Movielens 1M	1.06	0.76
	Movielens 100K	1.11	0.93
Improved DenseNet	Movielens 1M	0.97	0.72
	Movielens 100K	0.96	0.68

Table 4

Comparative results by means of f-measure, recall, and precision

Folds	Measures (%)	Y20-20 database		Y10-10 database
		MCRS-CS [33]	Improved DenseNet	MCRS-CS [33]	Improved DenseNet
$K=$ 2	F-measure	95.72	95.90	92.50	93.98
	Precision	97.27	98.02	90.24	94.90
	Recall	94.71	95.62	95.92	96.92
$K=$ 5	F-measure	94.99	95.02	93.99	94.90
	Precision	96.37	97.66	91.84	94.32
	Recall	95.23	96.05	96.84	97.04

Figure 5.

Graphical comparison of the improved DenseNet model on Yahoo Y-10-10 database.

Figure 6.

Graphical comparison of the improved DenseNet model on Yahoo Y-20-20 database.

4.3 Discussion

The primary aim of the movie recommendation system is to predict and filter the movies as per user request. In the present scenario, the machine learning and deep learning models are effective in the movie recommendation by handling an enormous amount of data. In this manuscript, an improved DenseNet model is proposed for movie recommendation. The improved DenseNet model has a class-balanced loss function and a weighting factor for effectively managing the overfitting risks. In addition, the improved DenseNet model superiorly understands the movie content and resolves the problems like sparsity, scalability, and cold start. These are the major benefits of employing an improved DenseNet model in movie recommendation. The performance of the improved DenseNet model is validated on the movielens 1M, movielens 100K, Yahoo Y-10-10, and Yahoo Y-20-20 databases in light of RMSE, MAE, f-measure, MSE, precision, and recall. Further, the improved DenseNet model’s effectiveness is shown in the Tables 1–4.

5. Conclusion and future work

In recent decades, the number of online movies is rapidly increasing, due to the growth of multimedia networks. The recommendation systems are effective in dealing with this issue and a movie recommendation system provides users with highly ranked lists of movies based on an individual’s constraint and preference. A user rating on every movie is impossible, and the same ratings on similar movies from different users become highly overlap, and it leads to data sparsity issues that limit the accuracy of the recommendation model. So, a new recommendation model is introduced in this article for relevant movie recommendations. In the initial phase, the input data is acquired from Movielens 1M, Movielens 100K, Yahoo Y-10-10, and Yahoo Y-20-20 databases. The collected high dimensional data are rescaled using Min-Max normalization technique, which decreases redundant data and provides better data representation with a limited outlier. The rescaled data are given to the improved DenseNet for a relevant movie recommendation, where it incorporates a weighting factor and class-balanced loss function for managing class imbalance issues and overfitting risk. The evaluation metrics: MAE, MSE, and RMSE demonstrated that the improved DenseNet model reduced by 5 to 10% of error values. Further, the evaluation metrics: f-measure, precision and recall showed that the proposed model improved 2% of recommendation results related to the traditional models on the Movielens 1M, Movielens 100K, Yahoo Y-10-10 and Yahoo Y-20-20 databases.

As a future extension, a novel hybrid clustering technique is developed to overcome scalability issues and computational time. Additionally, big data technology and parallel processing are considered in the large scaled databases.

Funding

This research received no external funding.

Data availability

The datasets generated during and/or analysed during the current study are available in the [Movielens 1M], [Movielens 100K] and [Yahoo Y-10-10 and Yahoo Y-20-20] repositories.

Footnotes

Conflict of interest

The authors declare that they have no conflict of interest.

Author’s Bios

	V. Lakshmi Chetana received her M. Tech in Computer Science and Engineering from JNTUK, Kakinada in 2014 and Masters in Computer Applications from ANU, Guntur in 2009. She is currently pursuing her PhD in Computer Science and Engineering at VIT-AP University, Amaravathi, Near Vijayawada, Andhra Pradesh. She had an academic experience of 13 years. She published papers in various international journals and presented papers in national and international conferences. Her research interests include Recommender Systems, Machine Learning, Deep Learning, Big Data Analytics, and Data Science.
	Rajkumar Batchu received his Ph.D. from VIT-AP University, Andhra Pradesh. M.Tech degree in Computer Science and Engineering from KL University in 2013 and B.E in Computer Science and Engineering from Anna University, Chennai in 2011. He is currently working as Assistant professor in the department of Computer Science and Engineering at School of computing, Amrita Vishwa Vidyapeetham, Amaravati campus, Andhra Pradesh. He had an academic and industrial experience of more than 5 years. His-research interests include Machine Learning, Deep Learning, Intrusion Detection, Big Data Analytics, and Data Science.
	Prasad Devarasetty obtained his Ph.D form Andhra University, Visakhapatnam in the year 2023. He obtained his M Tech degree from JNTU University, Hyderabad in the year 2007. He had an academic experience of 19 Years. He published various papers in reputed international journals and conferences. His research interests include Software Engineering, Cloud Computing, and Internet of Things. He is currently working as Associate Professor & heading the department of Computer Science and Engineering.
	V. Srilakshmi obtained her Ph.D degree from Acharya Nagarjuna University, Namburu, Guntur district, Andhra Pradesh. Received her M. Tech in Computer Science and Engineering from JNTU, Kakinada in 2012 and Masters in Computer Applications from ANU, Guntur in 2007. She had an academic experience of 15 years. She published papers in various international journals and presented papers in national and international conferences. Her research interests include Machine Learning, Deep Learning, Big Data Analytics, Privacy and security and Data Science.
	Interested in the areas of Computer science & engineering particularly Cloud Computing and its applications. Has been author & coauthor for papers published in International Journals & conferences of IEEE, Springer, Elsevier.

References

Zhang

Zhao

Cheng

and Wang

, Three-way recommendation model based on shadowed set with uncertainty invariance, International Journal of Approximate Reasoning 135 (2021), 53–70. doi: 10.1016/j.ijar.2021.04.009.

Das

Majumder

Gupta

and Datta

, Scalable recommendations using decomposition techniques based on Voronoi diagrams, Information Processing & Management 58(4) (2021), 102566. doi: 10.1016/j.ipm.2021.102566.

Reddy

S.R.S.

Nalluri

Kunisetti

Ashok

and Venkatesh

, Content-based movie recommendation system using genre correlation, in: Smart Intelligent Computing and Applications, Proceedings of the Second International Conference on SCI 2018 Satapathy

Bhateja

Das

, eds., Vijayawada, Andhra Pradesh, India, 2, Springer, Singapore, 2019, pp. 391–397. doi: 10.1007/978-981-13-1927-3_42.

Choudhury

S.S.

Mohanty

S.N.

and Jagadev

A.K.

, Multimodal trust based recommender system with machine learning approaches for movie recommendation, International Journal of Information Technology 13(2) (2021), 475–482. doi: 10.1007/s41870-020-00553-2.

Kumar

and Roy

P.P.

, Movie recommendation system using sentiment analysis from microblogging data, IEEE Transactions on Computational Social Systems 7(4) (2020), 915–923. doi: 10.1109/TCSS.2020.2993585.

Wang

and Xu

, A sentiment-enhanced hybrid recommender system for movie recommendation: A big data analytics framework, Wireless Communications and Mobile Computing 2018 (2018), 8263704. doi: 10.1155/2018/8263704.

Balakrishnan

Bouneffouf

Mattei

and Rossi

, Using Contextual Bandits with Behavioral Constraints for Constrained Online Movie Recommendation, in: Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18) Lang

, ed., July 13–19, Stockholm, Sweden, IJCAI Organization, 2018, pp. 5802–5804. doi: 10.24963/ijcai.2018/843.

Zhao

Wang

Yang

Zhao

Chen

and Shen

, Leveraging long and short-term information in content-aware movie recommendation via adversarial training, IEEE Transactions on Cybernetics 50(11) (2019), 4680–4693. doi: 10.1109/TCYB.2019.2896766.

Wang

and Du

, HI2Rec: Exploring knowledge in heterogeneous information for movie recommendation, IEEE Access 7 (2019), 30276–30284. doi: 10.1109/ACCESS.2019.2902398.

10.

Vimala

S.V.

and Vivekanandan

, A Kullback-Leibler divergence-based fuzzy C-means clustering for enhancing the potential of an movie recommendation system, SN Applied Sciences 1(7) (2019), 698. doi: 10.1007/s42452-019-0708-9.

11.

Chen

Zhu

Niu

and Zuo

, Knowledge discovery and recommendation with linear mixed model, IEEE Access 8 (2020), 38304–38317. doi: 10.1109/ACCESS.2020.2973170.

12.

Wang

Lou

and Chao

, A Personalized Movie Recommendation System based on LSTM-CNN, in: 2nd International Conference on Machine Learning, Big Data and Business Intelligence (MLBDBI), 23–25 October, Taiyuan, China, IEEE, 2020, pp. 485–490. doi: 10.1109/MLBDBI51377.2020.00102.

13.

Chen

Y.-L.

Yeh

Y.-H.

and Ma

M.-R.

, A movie recommendation method based on users’ positive and negative profiles, Information Processing & Management 58(3) (2021), 102531. doi: 10.1016/j.ipm.2021.102531.

14.

Thakker

Patel

and Shah

, A comprehensive analysis on movie recommendation system employing collaborative filtering, Multimedia Tools and Applications 80(19) (2021), 28647–28672. doi: 10.1007/s11042-021-10965-2.

15.

Kanmani

R.S.A.

Surendiran

and Ibrahim

S.P.S.

, Recency augmented hybrid collaborative movie recommendation system, International Journal of Information Technology 13(5) (2021), 1829–1836. doi: 10.1007/s41870-021-00769-w.

16.

Jain

Nagar

Singh

P.K.

and Dhar

, EMUCF: Enhanced multistage user-based collaborative filtering through non-linear similarity for recommendation systems, Expert Systems with Applications 161 (2020), 113724. doi: 10.1016/j.eswa.2020.113724.

17.

Zou

Chen

Jiang

and Kang

, A two-stage personalized recommendation based on multi-objective teaching-learning-based optimization with decomposition, Neurocomputing 452 (2021), 716–727. doi: 10.1016/j.neucom.2020.08.080.

18.

Singh

P.K.

Othman

Ahmed

Mahmood

Dhahri

and Choudhury

, Optimized recommendations by user profiling using apriori algorithm, Applied Soft Computing 106 (2021), 107272. doi: 10.1016/j.asoc.2021.107272.

19.

Singh

R.H.

Maurya

Tripathi

Narula

and Srivastav

, Movie recommendation system using cosine similarity and KNN, International Journal of Engineering and Advanced Technology 9(5) (2020), 556–559. doi: 10.35940/ijeat.E9666.069520.

20.

Wan

and Sun

, Movie recommendation based on bridging movie feature and user interest, Journal of Computational Science 26 (2018), 128–134. doi: 10.1016/j.jocs.2018.03.009.

21.

Vilakone

Park

D.-S.

Xinchang

and Hao

, An efficient movie recommendation algorithm based on improved k-clique, Human-centric Computing and Information Sciences 8 (2018), 38. doi: 10.1186/s13673-018-0161-6.

22.

Widiyaningtyas

Hidayah

and Adji

T.B.

, User profile correlation-based similarity (UPCSim) algorithm in movie recommendation system, Journal of Big Data 8 (2021), 52. doi: 10.1186/s40537-021-00425-x.

23.

Kharita

M.K.

Kumar

and Singh

, Item-based collaborative filtering in movie recommendation in real time, in: First International Conference on Secure Cyber Computing and Communication (ICSCCC), 15–17 December, Jalandhar, India, IEEE, 2018, pp. 340–342. doi: 10.1109/ICSCCC.2018.8703362.

24.

Ali

S.M.

Nayak

G.K.

Lenka

R.K.

and Barik

R.K.

, Movie recommendation system using genome tags and content-based filtering, in: Advances in Data and Information Sciences, Proceedings of ICDIS-2017 Kolhe

Trivedi

Tiwari

Singh

, eds., 3 to 4 November, Madhya Pradesh, Springer, Singapore, Vol. 1, 2018, pp. 85–94. doi: 10.1007/978-981-10-8360-0_8.

25.

Lin

C.-H.

and Chi

, A novel movie recommendation system based on collaborative filtering and neural networks, in: Proceedings of the 33rd International Conference on Advanced Information Networking and Applications, AINA-2019 Barolli

Takizawa

Xhafa

Enokido

, eds., 27 to 29 March 2019, Matsue, Japan, Springer, Cham, 2020, pp. 895–903. doi: 10.1007/978-3-030-15032-7_75.

26.

Zhang

and Mao

, Movie recommendation via markovian factorization of matrix processes, IEEE Access 7 (2019), 13189–13199. doi: 10.1109/ACCESS.2019.2892289.

27.

Vilakone

Xinchang

and Park

D.-S.

, Personalized movie recommendation system combining data mining with the k-clique method, Journal of Information Processing Systems 15(5) (2019), 1141–1155. doi: 10.3745/JIPS.04.0138.

28.

Kumar

M.S.

and Prabhu

, Hybrid model for movie recommendation system using fireflies and fuzzy c-means, International Journal of Web Portals (IJWP) 11(2) (2019), 1–13. doi: 10.4018/IJWP.2019070101.

29.

Vilakone

Xinchang

and Park

D.-S.

, Movie recommendation system based on users’ personal information and movies rated using the method of k-clique and normalized discounted cumulative gain, Journal of Information Processing Systems 16(2) (2020), 494–507. doi: 10.3745/JIPS.04.0169.

30.

Behera

D.K.

Das

Swetanisha

and Sethy

P.K.

, Hybrid model for movie recommendation system using content K-nearest neighbors and restricted Boltzmann machine, Indonesian Journal of Electrical Engineering and Computer Science 23(1) (2021), 445–452. doi: 10.11591/ijeecs.v23.i1.pp445-452.

31.

Zarzour

Maazouzi

Al-Zinati

Jararweh

and Baker

, An Efficient Recommender System Based on Collaborative Filtering Recommendation and Cluster Ensemble, in: Eighth International Conference on Social Network Analysis, Management and Security (SNAMS), 06–09 December, Gandia, Spain, IEEE, 2021, pp. 01–06. doi: 10.1109/SNAMS53716.2021.9732118.

32.

Alhijawi

and Kilani

, A collaborative filtering recommender system using genetic algorithm, Information Processing & Management 57(6) (2020), 102310. doi: 10.1016/j.ipm.2020.102310.

33.

Gupta

and Kant

, Credibility score based multi-criteria recommender system, Knowledge-Based Systems 196 (2020), 105756. doi: 10.1016/j.knosys.2020.105756.

34.

Deldjoo

Dacrema

M.F.

Constantin

M.G.

Eghbal-Zadeh

Cereda

Schedl

Ionescu

and Cremonesi

, Movie genome: Alleviating new item cold start in movie recommendation, User Modeling and User-Adapted Interaction 29(2) (2019), 291–343. doi: 10.1007/s11257-019-09221-y.

35.

Awan

M.J.

Khan

R.A.

Nobanee

Yasin

Anwar

S.M.

Naseem

and Singh

V.P.

, A recommendation engine for predicting movie ratings using a big data approach, Electronics 10(10) (2021), 1215. doi: 10.3390/electronics10101215.

36.

Jain

Shukla

and Wadhvani

, Dynamic selection of normalization techniques using data complexity measures, Expert Systems with Applications 106 (2018), 252–262. doi: 10.1016/j.eswa.2018.04.008.

37.

Zhong

Zheng

Mai

Zhao

and Liu

, Cancer image classification based on DenseNet model, in: 2nd International Conference on Artificial Intelligence Technologies and Application (ICAITA), 21–23 August, Dalian, China, IOP Publishing, 2020, Journal of Physics: Conference Series 1651(1) (2020), 012143. doi: 10.1088/1742-6596/1651/1/012143.

Effective movie recommendation based on improved densenet model

Abstract

Keywords

1. Introduction

2. Literature review

3. Improved DenseNet model for movie recommendation

3.2 Data denoising

Table 1 Experimental results of the improved DenseNet model on the movielens 1M and 100K databases by using different evaluation metrics

Table 3 Comparative results by means of RMSE and MAE

5. Conclusion and future work

Funding

Data availability

Footnotes

Conflict of interest

Author’s Bios

References

Table 1
Experimental results of the improved DenseNet model on the movielens 1M and 100K databases by using different evaluation metrics

Table 3
Comparative results by means of RMSE and MAE