International Journal of Electrical and Computer Engineering (IJECE) Vol.
No.
October 2025, pp.
ISSN: 2088-8708.
DOI: 10.
11591/ijece.
A novel approach for recommendation using optimized bidirectional gated recurrent unit Prakash Pandharinath Rokade1.
Swati Babasaheb Bhonde2.
Prashant Laxmanrao Paikrao3.
Umesh Baburao Pawar4 Department of Information Technology.
College of Engineering and Research Center Yeola.
Savitribai Phule Pune University.
Maharashtra.
India Department of Computer Engineering.
Amrutvahini College of Engineering.
Maharashtra.
India Department of Electronics and Telecommunication.
Government COE Amravati.
Maharashtra.
India School of Computing.
Sandip University Nashik.
Maharashtra.
India
Article Info
ABSTRACT
Article history:
In today's world, every one of us refreshes our mood and gets energy through entertainment and enjoyment.
Human nature is to provide feedback through ratings or comments for products used, services received, or films The recommendation system serves the user with recommendations based on historical stored information of user preferences.
These systems amass information about the user in order to provide personalized These systems put efforts into delivering personalized experiences by accumulating information about the user.
Hybrid algorithms are necessary to address the issues recommendation systems confront, which include low prediction accuracy, output that exceeds range, and inadequate convergence speed.
This study suggests building a movie recommendation system using the remora optimization algorithm (ROA) and the bidirectional gated recurrent unit (BiGRU), the most recent version of the recursive neural network (RNN).
The proposed method's results are compared with those of the genetic algorithm (GA), feed forward neural network (FFNN), and multimodal deep learning (MMDL).
In terms of movie recommendation.
BiGRU with ROA performs better than GA.
MMDL, and FFNN.
Received Sep 26, 2024 Revised Jul 3, 2024 Accepted Jul 12, 2025 Keywords:
Bidirectional gated recurrent Feedforward neural network Gated recurrent unit Genetic algorithm Multimodal deep learning This is an open access article under the CC BY-SA license.
Corresponding Author:
Prakash Pandharinath Rokade Department of Information Technology.
College of Engineering and Research Center Yeola.
Savitribai Phule Pune University Nashik.
Maharashtra.
India - 423401 Email: prakashrokade2005@gmail.
INTRODUCTION
Social media is heavily used by people to share their opinions and feelings on many websites.
These feelings manifest as opinions or ratings for a good or service.
The greatest difficulty facing someone who has to make an online purchase in the information era is not only finding enough options or information but also deciding how to use the wealth of available data to their advantage .
Ae.
This leads to the generation and analysis of massive amounts of data in order to forecast and suggest a user any product or service based on his interests.
Movie rating databases comprising user ratings and various movie parameters may be found on a number of well-known websites, such as Kaggle.
In order to address a problem, decisions supported by several stronger past impressions are always preferable to those produced by a single user impression .
, .
Only those users whose ratings are more strongly relevant to one another are gathered, as opposed to gathering all reviews and ratings .
Using various machine learning techniques, such as support vector Journal homepage: http://ijece.
ISSN: 2088-8708
machines (SVM), neural networks (NN), and genetic algorithms (GA), numerous academics have worked to improve recommendation systems .
Ae.
In this study, we suggested a bidirectional gated recurrent unit (BiGRU) optimized remora optimization algorithm (ROA) based movie recommendation system.
The 100 kB IMDB dataset is processed using the BiGRU method.
Userid, movieid, and rating are the only parameters taken into account prior to preprocessing .
Similarity between users' ratings of the same movies and their ratings of their differences from one another are obtained.
Finally, a user receives recommendations for the top 10 movies based on his interest pattern after the weights used to discover user similarity are optimized using ROA.
The outcomes of GA, multimodal deep learning (MMDL), and feed forward neural networks (FFNN) are contrasted with the results.
BiGRU appears to have produced better outcomes for every testing parameter we compared taken into account prior to preprocessing.
Similarity between users' ratings of the same movies and their ratings of their differences from one another are obtained .
, .
K-nearest neighbor algorithm with collaborative filtering is used for movie rating prediction .
The long short-term memory using cyclic learning rate (LSTM-CLR) framework is used to identify cyberbullying on social media, improving classification accuracy without numerous trials and adjustments .
Finally, a user receives recommendations for the top 10 movies based on his interest pattern after the weights used to discover user similarity are optimized using ROA.
The outcomes of GA.
MMDL, and FFNN are contrasted with the results.
BiGRU appears to have produced better outcomes for every testing parameter we compared .
Traditional recommendation systems, such as collaborative filtering and content-based filtering, often suffer from limitations including low prediction accuracy, slow convergence speed, and ineffective weight While deep learning models like FFNN and MMDL have shown improvements, they still struggle with optimal weight assignment and real-time adaptability.
This paper introduces an optimized recommendation system using BiGRU with ROA to enhance accuracy and convergence speed.
The main contributions of this paper are to propose of a BiGRU-based recommendation model optimized using ROA to enhance convergence speed and reduce training loss, to compare BiGRU with FFNN.
MMDL, and GA to demonstrate superior performance and to implement proposed system on the IMDB movie dataset, achieving a significant improvement in prediction accuracy.
The proposed system is particularly useful in movie recommendation systems, personalized content filtering, and e-commerce platforms, where real-time and high-accuracy predictions are crucial for user engagement.
METHOD
Proposed system architecture The system architecture comprises input file to various models and then applying Remora and BiGRU for recommendation.
As shown in Figure 1, the system architecture contains input file, the matrix representation of input file, co-occurrence matrix representation, constrain model, rating independent model.
Bidirectional GRU, remora optimization and finally the recommendations.
Figure 1.
Architecture of proposed system .
Int J Elec & Comp Eng.
Vol.
No.
October 2025: 5019-5030
Int J Elec & Comp Eng
ISSN: 2088-8708
Bidirectional gated recurrent unit Bidirectional gated recurrent unit (BiGRU), a more recent RNN version than LSTM, is becoming increasingly popular these days.
Because RNN performs the same operation at every time point, its calculation graph is extraordinarily deep .
A neural network's long- and short term memory strategy is suggested to address RNN problems, although it has a more complex structure and struggles to converge more quickly .
, .
The BiGRU outperforms the LSTM in terms of speed.
The network configuration of BiGRU is the following: there are 2 GRU units, 2 dense layers and 2 dropout layers .
6 layer.
Inside each GRU unit 1 reset and 1 update unit will be there .
As shown in Figure 2.
GRU1 is a forward GRU, and Figure 2 displays its internal features, while Figure 3 displays the internal details of GRU2, a reverse GRU.
The forward calculation shown in Figure 3 is carried out as follows.
Suppose at time yc.
EEycyc is the reset gate of the positive input GRU.
Here is the formula:
ycEEyc = yua.
cO ycuyc e ycOyc e EaycOe1 In the formula.
E is the sigmoid function, e ycuyc and e EaycOe1 are the values of the most recent activation and the current input, correspondingly.
ycOyc is the input weight matrix.
e ycOyc is the weight matrix for cyclic Similarly, suppose e ycyc is the update gate of the forward GRU at time yc.
the formula is as:
ycyc = yua.
cO e ycOyc e EaycOe1 Figure 2.
BiGRU working architecture Suppose at time yc, e Eayc is the activation value of the positive input GRU, which is a compromise between the candidate activation value e EaycOe1 and the last activation value e EaycOe .
e Eayc = .
Oe e ycyc )A e EaycOe1 e ycyc .
EaycOe .
The formula for EaycOe is as .
e ycuyc e EaE ycyc .
e ycOEa e EaycOe1 ) ycOe = tanh.
cOEa .
In the formula .
, is the Hadamard product.
For reset gate, if e ycyc is closed means its value approaches 0, the GRU eliminates the previous activation value e EaycOe1 and the current input e ycuyc is the only factor affecting it.
This allows hEt to deny irrelevant information, thereby more effectively communicating pertinent facts .
On other side, the update gate e ycyc controls how much information in e EaycOe1 .
can be delivered to the current e Eayc .
This is the key to designing the results for this unit.
It functions as a memory unit akin to an LSTM, aiding GRU in remembering long-term data .
Similarly, formula .
Ae.
provide the computation method for the reverse GRU shown in Figure 3 and Figure 4.
Eaeyc Eae ycEae=yua.
cO ycuyc Eae ycOyc Eae Eayc 1 ) yc A novel approach for recommendation using optimized bidirectional A (Prakash P.
Rokad.
A ISSN: 2088-8708 Eae Eae ycuyc ycO Eaeyc Ea ycEae yc = yua.
cOyc Eae yc 1 ) .
Eae Eae Ea Ea.
AEa ycyc Eae yc = .
Oeyc yc yc 1 Eaea yc 1 .
Eae ycuyc ycEae.
Eae Eae EaE yc = tanh.
cOEa Eae yc ycOEa Eayc 1 ) .
The results of two directions are average to obtain final output Eayc .
Figure 3.
Inner structure of a forward GRU neuron Figure 4.
Inner structure of a backward GRU neuron Remora optimization algorithm ROA is an optimization algorithm that uses metaheuristics and is inspired by the foraging behavior of remora species.
ROA, inspired by remora fish behavior, dynamically adjusts weight values in the BiGRU network, reducing training errors and accelerating convergence.
It ensures optimal weight selection to maximize the accuracy of recommendations.
The main intention of using ROA is to optimize the weight values in BiGRU and to result in accurate outcomes.
ROA plays an important role to optimize input weights if the output at neuron exceeds the range 0 to 1.
The ROA optimization algorithm is used to fine-tune the Bi-GRU's settings.
This is accomplished by repeatedly looking for the ideal weight values while minimizing inaccuracy or loss.
Therefore, the optimization algorithm's fitness function is the reduction of loss/error in Bi-GRU.
As shown in Figure 5, the memory optimization algorithm consists of the following steps .
, .
, .
Flowchart of remora Remora optimization technique is inspired by symbiotic relationships in nature of remoras and To balance exploration and exploitation It combines global and local search strategies.
Agents .
follow and adapt to leaders .
in the population to find optimal solutions.
It is commonly used in solving complex, nonlinear optimization problems.
The method is lightweight, converges fast, and suits real-world engineering applications.
Remora optimization algorithm flowchart is designed as follows.
Create the first population In this case, the search agent's parameter is population.
We must initialize the number of remora, or search agents, in our system.
The search agent is remora.
Define network weight The weight values of neural networks will be defined here.
These weight values are nearly zero and are created at random.
Modification of search agents We first define a solution space in all optimization techniques.
We also specify the number of search agents and the search area.
We change the value of search agents if their number in a search space surpasses a threshold.
in a search space, only search agents below the threshold will be permitted entry.
Error reduction In this case, the mistake is reduced by the error or loss function.
Int J Elec & Comp Eng.
Vol.
No.
October 2025: 5019-5030
Int J Elec & Comp Eng
ISSN: 2088-8708
The current search agent's position and the store's fitness Several network weights are available in the search space.
We shall maintain a record of the optimal network weight.
Determine the nearest optimal weight We can get the next weight value, or optimal weight, by computing the fitness value of each weight value we initially chose.
The first weight value is known.
Every iteration's fitness function is being Analyze fitness .
rror minimizatio.
For every weight in the network, we find its fitness value.
A comparison of the new and old weight values is presented.
Error minimization Here, the mistake is minimized by using the error function or loss function.
Store fitness and position of current search agent:
There are several network weights in the search space.
The ideal network weight will be kept on file.
Find near optimal weight The first weight value is known, and by calculating the fitness value of each weight value we initially selected, we can get the next weight value, or optimal weight.
We are assessing the fitness function for each iteration.
Evaluate fitness .
rror minimizatio.
Every network weight is assessed for its fitness value.
The new weight value and the old weight value are being compared.
Stopping criteria First, we will define the number of iterations that will be used as our cutoff point.
The goal is to increase the correctness of the network model by choosing the weight values for each iteration of the remora algorithm in an optimal manner.
Start Generate initial Evaluate fitness .
rror Find near optimal weight Store fitness and position of current search agents Evaluate fitness .
rror Local update If attempt < Yes Define network weights Current iteration= current iteration 1 Check if search agent goes beyond the search space and amend it Yes Position update Stopping End Figure 5.
Remora optimization algorithm flowchart .
Input file uploading MovieLens dataset of 150.
35 kB is taken as an input from Kaggle.
The dataset comprises 100,000 ratings captured from 671 users for a total of 1,682 movies.
The input data contains the features userid, movieid, ratings.
Input file contains 100,000 entries.
It comprises the movie ratings along with the title, genre, time stamp for movies.
About 671 users have rated with different ratings between 1 to 5.
The A novel approach for recommendation using optimized bidirectional A (Prakash P.
Rokad.
A ISSN: 2088-8708 overall size of the dataset is 150.
35 kB.
The input data contains the features userid, movieid, ratings, title, genre, timestamp, tag, imdbid, tmdbid.
Only four features userid, movieid, moviename, ratings are considered in proposed model as an input.
As shown in Table 1, 671 users have rated 1,682 movies with ratings in between 1 to 5.
It is not the case that all users have rated all the movies.
Some users have not rated few movies.
Either they have not seen those movies or they are not interested in providing the ratings to those movies.
Consider user 1 has rated movie number 1343 with ratings 2 and this movie belongs to the genre action/adventure/thriller.
Table 1.
Input file
Userid
Movieid
Rating Timestamp
26E 09
26E 09
26E 09
26E 09
26E 09
26E 09
26E 09
26E 09
26E 09
06E 09
07E 09
06E 09
Title Toy Story Jumanji Grumpier Old Men Waiting to Exhale Father of Bride-II Heat Sabrina Tom and Huck Sudden Death Dracula Balto Nixon Genre Adventure | Animation Adventure Comedy | Romance Comedy | Drama | Romance Comedy Action | Crime | Thriller Comedy | Romance Adventure Children Action Comedy Horror Adventure Animation Drama Reduction of sparsity issue Dimensionality reduction is one of the good solutions for sparsity reduction.
In this step, the sparsity issue in the input file is resolved by considering only the parameters userid, movieid, ratings.
considering only these limited parameters, the complex calculations are minimized and sparsity issue is As shown in Table 2, ratings for movies by different users are shown in matrix format showing that user ratings for movie rated.
There is no rating provided for unrated movies.
User id 12 has provided ratings 2, 2, 4.
5, 4, 3, 2.
5 for the movies 1, 2, 5, 6, 10, 11 respectively.
There is empty space in the matrix by the user id 12 for the movies 3, 4, 7, 8, 9, 12 as it has not rated those movies.
This experiment is evaluated for all users and movies but only for 12 users and movies matrix is designed.
Table 2.
Matrix representation of ratings for different movies by users Userid/Movieid Co-occurrence model There are few movies which are either rated by a single user, some movies are rated by all users.
Co-occurrence model is obtained by intersection operator.
It shows number of similar movies rated by two A co-occurrence matrix is generated to capture relationships or patterns between items .
r user.
, which helps in understanding associations.
Table 3 shows the co-occurrence model in which if two users have rated same movies then they are included into the matrix.
Out of 671 users for first 12 users only the matrix is prepared.
Here user 7 has rated 5, 9, 11, 41, 10 same movies with user 1, 2, 3, 4, 5, 6 and so on.
This shows co-occurrence model of similar movies rated by two users.
Co-occurrence model is the part of input to the BiGRU algorithm for through calculation for predicting movies to the user.
Int J Elec & Comp Eng.
Vol.
No.
October 2025: 5019-5030
Int J Elec & Comp Eng
ISSN: 2088-8708
Table 3.
Co-occurrence model of similar movies rated by two users Userid/Userid Constraint model Users who have given similar movies the same ratings are taken into consideration to create this This model is the restricted version of co-occurrence model.
From the input matrix representation, it is clear that there are some users who have rated similar movies by similar ratings with the other users and few users are there who have rated similar movies by different ratings with the other users.
As shown in Table 4, user 7 has rated 2,3,5,21,5,0 similar movies by the same ratings as users 1, 2, 3, 4, 5, 6, respectively.
The model is obtained for all users, but for the first 12 users.
Table 4 is prepared.
This model can be termed as a dependent model, as similar movies with the same ratings by users are considered Table 4.
Constraint model of similar movies with same ratings by two users Userid/Userid This module applies certain predefined constraints .
, user behavior rules, diversity constraints, etc.
to guide the learning or filtering process.
Rating independent model This approach takes into consideration viewers who have given similar movies varied ratings.
shown in Table 5, user 7 has rated 3, 9, 6, 20, 5, 0 similar movies by different ratings with the users 1, 2, 3, 4, 5, 6 respectively.
This model is evaluated for all 671 users out of which for first 12 userAos values are shown in Table 5.
Table 5.
Rating independent model of similar movies with different ratings by two users Userid/Userid 80 60 14
143 100 34
A model that operates without directly depending on user-provided ratingsAipossibly using implicit feedback, content similarity, or contextual signals A novel approach for recommendation using optimized bidirectional A (Prakash P.
Rokad.
A ISSN: 2088-8708
RESULT AND DISCUSSION
Testing for recommendation Co-occurrence model, constraint model, rating independent model are the inputs to BiGRU By considering random input multiplied with weights followed by addition of bias value at each neuron in hidden layer, the output is calculated.
If the output at these neurons is between the ranges 0 to 1, these weights will be forwarded to next layer neurons.
If the output at hidden layer neuron is beyond the range 0 to 1, remora optimization algorithm will be active to optimize these weights.
Finally, by testing the proposed model for any random user, the recommended movie list will be suggested to the user.
The output as recommended movies are shown in Figure 6.
Figure 6 shows that when we select any user for recommending him some movies, the interest pattern of user is already studied and based on that patter he will be recommended some movies.
Here we can provide the threshold for getting first top x number of movies as result.
Figure 6.
Recommendation of movies for random user Result analysis of proposed system The performance of the suggested system is assessed using testing measures like precision, recall, f measure, accuracy, mean absolute error, and root mean square error.
These values are ascertained by obtaining a confusion matrix.
Proposed system comparison with existing system To compare the performance of proposed model for the input dataset, the performance criteria are applied to MMDL.
GA and FFNN for the same input and results are observed.
As shown in Figure 7, proposed optimized BiGRU model for movie recommendation has the 98% precision which is the highest as compare to MMDL with 92%.
GA with 96% and FFNN with 85%.
As shown in Figure 8, proposed optimized BiGRU model for movie recommendation has the 97.
recall which is the highest as compare to MMDL with 96%.
GA with 97% and FFNN with 88%.
As shown in Figure 9, proposed optimized BiGRU model for movie recommendation has the 97% f-measure which is the highest as compare to MMDL with 95%.
GA with 96% and FFNN with 87%.
As shown in Figure 10, proposed optimized BiGRU model for movie recommendation has the 98.
5% accuracy which is the highest as compare to MMDL with 95%.
GA with 96% and FFNN with 86%.
As shown in Figure 11, proposed optimized BiGRU model for movie recommendation has the 0.
03 MAE which is the lowest as compare to MMDL with 0.
GA with 0.
05 and FFNN with 0.
As shown in Figure 12, proposed optimized BiGRU model for movie recommendation has the 0.
17 RMSE which is the lowest as compare to MMDL with 0.
GA with 0.
22 and FFNN with 0.
Int J Elec & Comp Eng.
Vol.
No.
October 2025: 5019-5030
Int J Elec & Comp Eng
Recall in % Precision % Proposed A ISSN: 2088-8708
MMDL
BL_GA
FFNN
Proposed Proposed MMDL
BL_GA
FFNN
Proposed Proposed MMDL
BL_GA
BL_GA
FFNN
Figure 10.
Comparison of accuracy for proposed system with existing systems RMAE
MMDL
System for Comparison Figure 9.
Comparison of f-measure for proposed system with existing systems MAE
FFNN
System for Comparison
BL_GA
Figure 8.
Comparison of recall for proposed system with existing systems Accuracy F-Measure in % MMDL System for Comparison Figure 7.
Comparison of precision for proposed system with existing systems System for comparison
FFNN
System for Comparison Figure 11.
Comparison of MAE for proposed system with existing systems Proposed MMDL
BL_GA
FFNN
System for Comparison Figure 12.
Comparison of RMSE for proposed system with existing systems Experimental results and comparisons To evaluate the effectiveness of the proposed BiGRU-ROA model, we compared its performance with three existing approaches Ai GA.
MMDL, and FFNN Ai across four key metrics: precision, recall.
F-measure, and accuracy.
These metrics offer a comprehensive assessment of model performance in terms of both predictive accuracy and robustness.
The following bar chart clearly illustrates that the BiGRU-ROA model consistently outperforms other models in each category, showcasing its superior precision .
%), recall .
5%).
F-measure .
%), and accuracy .
5%).
As shown in Table 6, the values of precision, recall, f-measure, accuracy.
MAE.
RMSE for existing models MMDL.
GA.
FFNN and proposed optimized BiGRU model are calculated.
The results of the proposed BiGRU-ROA model show an accuracy improvement of 12.
3% over FFNN and 8.
5% over GA, 32% faster convergence rate compared to traditional neural networks, reduction in training error to 0.
enhancing predictive performance.
Figure 13 shows the bar chart comparing the performance of BiGRUROA.
GA.
MMDL, and FFNN in terms of Precision.
Recall.
F-Measure, and Accuracy.
A novel approach for recommendation using optimized bidirectional A (Prakash P.
Rokad.
A ISSN: 2088-8708 Table 6.
Proposed system performance comparison with existing models Model vs criteria
Proposed system
MMDL
FFNN
Precision 98 98 97
92 94 94
96 95 95
85 85 85
Recall F-Measure 97 97 98
95 94 95
96 95 95
87 88 89
Accuracy 98 97 98
95 92 93
96 94 95
86 87 86
MAE
RMSE
Figure 13.
Performance comparison of model As proposed model came out with comparatively better results than existing approaches, the objectives of research are attended.
To further validate the robustness and learning efficiency of the proposed BiGRU-ROA model, we analyzed its training loss behavior over multiple epochs as shown in Figure 14.
The training loss curve provides insights into how quickly and effectively the model learns from the input data.
Faster convergence with minimal fluctuations indicates stable learning and effective weight The ROA plays a key role here by dynamically adjusting weights to minimize error during each iteration.
Figure 14.
Training loss vs epochs BiGRU-ROA Int J Elec & Comp Eng.
Vol.
No.
October 2025: 5019-5030
Int J Elec & Comp Eng
ISSN: 2088-8708
As shown in Figure 14, the training loss for BiGRU-ROA decreases rapidly within the first 10 epochs and continues to improve steadily, reaching a minimal loss value of approximately 0.
025 by the 20 th This demonstrates the modelAos capability to converge faster compared to traditional deep learning The smooth and consistent decline in the loss curve is a clear indication of ROA's contribution to weight optimization and error reduction, leading to improved generalization and predictive accuracy of the recommendation system.
CONCLUSION
The findings are compared to those produced using GA.
MMDL, and FFNN and examined for characteristics like precision, recall, accuracy.
F-measure.
MAE, and RMSE.
It is discovered that BGRU with ROA has better results than the others, with 97% accuracy, 97.
5% F-measure, 97% precision, and 98% recall.
The lowest values among all the remaining methods for comparison are MAE 0.
RMSE
Therefore, it can be said that BiGRU with ROA performs better when it comes to movie selection.
The novel approach in the future, the newest machine learning algorithms for movie recommendation may employ ROA.
Additionally.
BiRGU can be combined with any recent optimization algorithm for improved REFERENCES