Latent based temporal optimization approach for improving the performance of collaborative filtering

Introduction

Recommendation systems are some of the most powerful methods for suggesting products to customers based on their interests and online purchases (Jonnalagedda et al., 2016; Lin, Li & Lian, 2020; Nilashi, bin Ibrahim & Ithnin, 2014; Nilashi et al., 2015; Zhang et al., 2020b). In terms of personalization of recommendations, one of the most prevalently used methods is collaborative filtering (CF) (Nilashi, bin Ibrahim & Ithnin, 2014; Sardianos, Ballas Papadatos & Varlamis, 2019; Nilashi et al., 2015; Wu et al., 2019). In CF, personalized prediction of products depends on the latent features of users in a rating matrix. However, the CF prediction accuracy decreases if the rating matrix is sparse (Zhang et al., 2020a; Li & Chi, 2018; Idrissi & Zellou, 2020). Several types of factorization techniques such as baseline, singular value decomposition (SVD), matrix factorization (MF), and neighbors-based baseline have been exploited to address the problem of data sparsity (Mirbakhsh & Ling, 2013; Al-Hadi et al., 2017b) by predicting the missing rating scores in the rating matrix. Similarly, various factorization-based techniques including the use of latent (Vo, Hong & Jung, 2020; Nguyen & Do, 2018) and baseline factors (Koenigstein, Dror & Koren, 2011) (such as SVD (Wang et al., 2019)) have been proposed to improve the recommendation accuracy. Nevertheless, an unaddressed problem is that a part of the rating scores is misplaced from its original cells while streaming into the memory. This misplacement decreases the meticulousness of the latent feedback.

A method based on ensemble divide and conquer (Al-Hadi et al., 2016) was adopted to solve the misplacement problem besides addressing the customers’ preferences drift and popularity decay. Integration of temporal preferences with factorization methods to solve the sparsity issue has yielded a better performance compared to basic factorization approaches (Al-Hadi et al., 2017b; Li, Xu & Cao, 2016; Nilashi et al., 2019; Nilashi, bin Ibrahim & Ithnin, 2014). The temporal dynamics approach (Koren, 2009) separates the time period of preferences into static digit of bins and extracts a universal weight according to the stochastic gradient descent method to reduce overfitting. Nonetheless, the learned universal weight using the temporal dynamics approach has limitations with respect to how it personalizes and represents the fluctuating temporal preferences.

The temporal interaction approach (Ye & Eskenazi, 2014) enhanced the effectiveness of CF recommender systems by combining the latent factors, short-term preferences, and long-term preferences. The shrunk neighbor approach is applied to obtain clients’ short-term feedbacks (Koren, 2008). This approach detects overfitting when there is a fluctuating scale in the rating scores. For example, in the rating matrix from the MovieLens dataset, the actual score is smaller than the predicted values. In this theory, the risk of each asset is measured with its beta value (which is the criterion of systematic risk). The fluctuating scale is in the range of 0–5, whereas the anticipated rating scores are [5.75; 6.11; 5.9; 7]. Compared to other temporal approaches (e.g., the short-term based latent technique (Yang et al., 2012)), the temporal interaction approach (Ye & Eskenazi, 2014) efficiently anticipates performance of CF. Nevertheless, problems such as drifting customers’ preferences and popularity decay (e.g., deterioration of marketability of goods) still pose a significant challenge (Ye & Eskenazi, 2014).

The long temporal-based factorization approach addresses the popularity decay issue (Al-Hadi et al., 2018b) while the short temporal-based factorization approach (Al-Hadi et al., 2017a) addresses the drift issue not solved by previous short-term based approaches. These temporal approaches improve the performance of CF but they are characterized by low accuracy. In view of the aforementioned, this paper presents a latent-based temporal optimization (LTO) approach to solve the significant limitations of these temporal approaches. As optimization algorithms have proven successful in various areas such as healthcare (Zainal et al., 2020)and document processing Al-Badarneh Amer (2016), we extend our earlier work (Al-Hadi et al., 2018a) and provide a detailed analysis of the proposed approach. The contributions of this paper are summarised below.

• A comprehensive review of CF-based recommender system techniques.

• A proposed LTO approach that minimizes overfitting and learns by integrating long and short-term features with the baseline and factorization features.

• An LTO approach that learns the drift in the users’ interests through an improved rating score prediction. This is achieved by integrating the long and short-term features of users and items with their baseline and factorization features.

• An LTO approach that solves the sparsity issue by combining the learning output of the overfitting, drift, and decay.

• A comparison of LTO’s performances with other factorization and temporal-based factorization approaches.

In summary, the proposed approach has superior performance as it improves the prediction accuracy in the CF technique by learning accurate latent effects of the temporal preferences of users. The novel features of the LTO approach are as follows:

• It provides a personalized temporal weighting which is incorporated in matrix factorization to reduce data sparsity error.

• It combines time convergence and personalized duration to accommodate consumer’s preferences drifting in the personalized recommendation system.

• It utilizes the bacterial foraging optimization algorithm (BFOA) to accurately learn the personalized temporal weights by regularizing the overfitted predicted scores in the rating matrix and to track the factors of drift and decay.

The rest of this paper is sectioned as follows: ‘Related works’ reviews the past works related to factorization approaches and temporal preferences. In ‘Latent-based temporal optimization approch’, LTO is elaborated, followed by experimental analysis in ‘Experimental Settings’. ‘Experimental results’ discusses the experimental results. The final section (‘Conclusion’) provides a summary and indicates possible future works.

Related Works

Latent-based Temporal Optimization Approach

The LTO approach addresses both long and short temporal preferences by using factorization to solve issues of preference drift and popularity decay (alg1). LTO applies RMSE, Cosine, and Prediction functions to assess the temporal preference representation. The key empirical setting of the temporal-based factorization method and the proposed solution framework are presented in Fig. 1.

BFOA is exploited to capture the preferences of a short duration. By applying k-means, the timestamp convergence deals with short durations in the time matrix. The number of clusters k is recognized based on the number of short durations in the entire period. Generally, bacteria cannot track the drift and the time decay perfectly during the short-term without considering the long-term. Therefore, the integration of the long and short durations represents the accurate solution for solving the limitations of the drift and the time decay.

In Fig. 2, we present an example of how to create the bacteria members by applying the k-means method. In this example, the clusters’ number is assigned 2 for each of the users’ and items’ features. Based on the time convergence between the products (columns) and customers (rows), four bacteria members are shown in Fig. 2.

The standard BFOA is utilized in the LTO process to detect the temporal conducts of customers and products. BFOA members initialize the short-term weights using random values. The LTO approach changes these weights throughout the lifecycle of BFOA dynamically based on the positive effect it has on learning stages. The weights of bacteria members ${\top }_x^u$ and ${\top }_y^i$ are updated dynamically throughout the learning iteration. This provides a novel tracking of users’ interests in the items. The LTO uses Eq. (9) to reduce the overfitted predicted scores throughout the learning iteration. (9)${\displaystyle D_{ui}={\left(\mu +{\top }_x^u{B}_u{\omega }_u+{\top }_y^i{B}_i{\omega }_i+p_u{q}_i^T\right)}^2,}$ where ${\top }_x^u$ and ${\top }_y^i$ indicate the short temporal weights indexed by clusters x and y. User u is indexed by cluster x and item i is indexed by cluster y. The values of ${\top }_x^u$ and ${\top }_y^i$ are updated in each iteration according to the positive effect in developing the accuracy prediction of the CF method. The vectors ω_u and ω_i are the long-temporal independent weights of customer u and product i while B_u, B_i, p_u, $q_i^T$ represent the baseline and factorization variables.

The second contribution of LTO is tracking users’ drifting interests. This is learned by focusing on the time associated with users’ interests as represented by Eq. (10). (10)${\displaystyle F_u={\top }_x^u\left[{\left(B_u{\omega }_u\right)}^2+{∥p_u∥}^2\right],}$ where ${∥p_u∥}^2$ represents the norm value of user’s latent factor and ${\top }_x^u$ is updated according to the positive effects of changing users’ interests throughout the learning process.

The third contribution of LTO is tracking the popularity decay of items throughout the learning process by focusing on the time popularity of items as shown in Eq. (11). (11)${\displaystyle G_i={\top }_y^i\left[{\left(B_i{\omega }_i\right)}^2+{∥q_i^T∥}^2\right],}$ where ${∥q_i^T∥}^2$ is the norm factorization variable of items and ${\top }_y^i$ is updated according to the improvement achieved through the learning iteration which affects the baseline values and norm factorization features of items. Furthermore, BFOA learns the significance of each short-term period by applying the RMSE (which acts as the fitness value).

These contributions are combined in Eq. (12) to predict the unknown values within the rating matrix. (12)${\displaystyle {\hat{\Re }}_{ui}=D_{ui}+F_u+G_i.}$ The BFOA operates in three stages: chemotaxis, reproduction, and removal and distribution. The first stage involves seeking the closest nutrition source by the bacteria. This is accomplished by swimming or tumbling or alternating between swimming and tumbling to change direction during the generation. In this process, the flagella of the bacteria make clockwise rotations to choose another path so that rich nutrients in the surrounding can be obtained. The tumbling stage is expressed in Eq. (13). (13)${\displaystyle {\tau }^i\left(j+1,k,l\right)={\tau }^i\left(j,k,l\right)+C_i+\frac{\theta }{\sqrt{{\theta }_i^t{\theta }_i}},}$ where τ is the short-term features of one bacterium, the variables i, j, k and l symbolize i-th bacterium at j-th chemotactic, k-th reproduction and l-th elimination and dispersal steps. C_i is the walk length in an irregular direction, Θ_i is a random value oncterium number i at j-th chemotactic, k-th reproduction and [-1,1], and $\frac{\Theta }{\sqrt{{\Theta }_i^t{\Theta }_i}}$ is the unit walk in the irregular direction.

Swimming follows tumbling whenever the flagella of bacteria make the counterclockwise rotation to move in a particular direction. The bacteria continue swimming in the same direction if the nutrients are rich and the alternation between tumbling and swimming is repeated until the chemotactic stage is complete. The swarming function utilizes a sensor to provide signals in a nutrient-rich environment. When the signal indicates a poor nutrient or a dangerous location, the bacteria shift from the center to the outward direction with a moving ring of the members. If the nutrient has a high level of succinate, the bacteria subsequently neglect aspartate attractant and concentrate in groups.

The bacteria provide an attraction signal for the all members so that they swim together. The bacteria move in a concentric pattern with a high density. The outward movement of the ring and the native releases of attractant constitute the spatial order (Kim & Abraham, 2007). The swarming stage is represented mathematically as shown in Eq. (14). (14)${\displaystyle {\tau }^i\left(j+1,k,l\right)=\sum _{i=1}^S-d_{attr}exp\left(w_{attr}\beta \right)+\sum _{i=1}^S-h_{rep}exp\left(w_{rep}\beta \right),}$ where S denotes the number of bacteria and β is the summation of short-term features that can be learned by the members of bacteria i as shown in Eq. (15). The attractant depth d_attr denotes the magnitude of excretion by a cell, attractant width w_attr denotes how the chemical cohesion signal spreads, repellent height h_rep and width w_rep determine the size of optimization space where the cell is related to the dispersal of chemical signal. (15)${\displaystyle \beta =\sum _{m=1}^P{\left({\tau }_m-{\tau }_m^i\right)}^2,}$ where P denotes the number of short-term features, τ_m denotes the short-term feature number m learned during the chemotaxis process while ${\tau }_m^i$ is the short-term feature number m learned during the chemotaxis process by bacteria i.

In the reproduction stage, the health of bacteria is calculated according to the fitness value of each bacterium. The bacteria values are then sorted in an ascending order in the array. The fitness value provided by RMSE is extracted from the optimization area in the recommendation system (that is based on collaborative filtering). The lower half bacteria with poor foraging die and the upper half bacteria (having better foraging) are copied into two parts. Each part has the same values Al-Hadi et al. (2017a). This procedure keeps the bacterial population constant. The bacterial health can be calculated using Eq. (16). (16)${\displaystyle J_{health}^i=\sum _{j=1}^{N_c+1}J\left(i,j,k,l\right),}$ where $J_{health}^i$ is the healthy score of the short-term preference that can be learned by bacteria i, the number of chemotactic, reproduction and removal steps are j, k and l, respectively.

Provision for the possibility of the ever-changing status of a few bacteria is carried out in the third step (removal and distribution). Here, the rise order involves the arrangement and generation of the random vector. The bacteria are organized based on their health values. Moreover, the randomly generated locations are used to change the locations of the bacteria in the optimization domain. These locations are recognized as the prominent available locations. After the generation, the best result in each repetition is approved as the final (correct) result.

In this work, the BFOA is integrated with the k-means clustering algorithm and matrix factorization approach. The k-means acts as a clustering algorithm used to control the big optimization space based on members’ personal features. It is used to reduce the large number of members to a small number of clusters. These clusters are then controlled using a weight for each cluster to control the optimization domain. Applying the natural choice in the course of repeating generations, the BFOA decreases the number of poor foraging members and increases the number of rich foraging strategies (Al-Hadi, Hashim & Shamsuddin, 2011). After several generations, the poor foraging members are removed or transformed to skilled members.

Experimental Settings

The CF technique is used to predict the interest of an active user. This takes into account the calculated similarity values between the rating scores of common users (neighbours) and the active user. However, sparse rating scores in the rating matrix negatively affect the prediction accuracy of the CF technique. Thus, this research work aims to improve the prediction of sparse rating scores in the rating matrix of each active user. The data sparsity is an important issue which this research aims. The factorization approaches (including temporal-based factorization) are used to predict missing rating scores in the rating matrix which improves the prediction accuracy of CF. The percentage accuracy can be measured using RMSE function where the lower values of RMSE refers to the accurate predicted values for the missing rating score in the rating matrix and also refers to the heights accuracy prediction of the recommendation list of items that can provided to the active user.

Experimental Results

This section discusses the performances of the benchmark and the proposed approaches for improving the CF prediction performance technique under two scoring scales which are [0–5] and [0–1]. The efficacy of LTO in resolving the decay and drift issues by reducing the RMSE values are also discussed. Table 7 shows the personal vectors of the Test-set matrices that can impact the experimental results. There are 31 rating matrices for MovieLens which are selected by the sequence 30, 60, 90, ..…, 930 for the Test-set. Each matrix in the Test-set has different numbers of rows and columns. The sparsity levels are compared with other matrices to provide unique results for each rating matrix. The average numbers of these matrices and their factors are used for performance evaluation in the experiments.

LTO approach under the scoring [0–5]

The LTO approach is applied to five short-term periods (which are a week, two weeks, a month, and a year) according to the tested dataset under the rating scale [0–5]. Figures 3–5 demonstrate the prediction accuracy while performing the iterations on the datasets. Figure 3 shows that the two weeks period in MovieLens has a higher prediction accuracy during 3–20 iterations compared to one week and one month. Users’ activities during the long and short duration preferences are best learned within the two weeks period. This makes it an accurate short-term preference.

Table 6:

An Example of how LTO reduces RMSE values in Movielens under scoring [0-5].

User	Iteration number										Min RMSE
	1	2	3	4	5	6	7	8	9	10
1	0.827	0.814	0.801	0.797	0.793	0.786	0.782	0.782	0.780	0.777	0.777
2	0.796	0.758	0.758	0.755	0.751	0.751	0.752	0.750	0.751	0.754	0.750
3	1.274	1.255	1.245	1.231	1.231	1.229	1.243	1.226	1.229	1.222	1.222
4	0.704	0.645	0.640	0.635	0.632	0.629	0.628	0.631	0.627	0.628	0.627
5	1.128	1.116	1.058	1.036	1.018	1.018	1.009	1.005	1.002	1.000	1.000
Avg. RMSE	0.946	0.918	0.900	0.891	0.885	0.882	0.883	0.879	0.878	0.876	0.875

DOI: 10.7717/peerjcs.331/table-6

Table 7:

Average personal vectors of the Test-set matrices.

Test-set	MovieLens	Netflix	Epinions
Number of the rating matrices	31	20	5
Avg. No. of customers	915	953	107
Avg. No. of products	105	128	103
Avg. No. of rating scores >0	17491	11662	517
Avg. No. of total rating scores > = 0	97855	125183	10481
Avg. percentage of SparsityLevel	79.18 %	89.12 %	94.59 %

DOI: 10.7717/peerjcs.331/table-7

In Fig. 4, the period of one month in the Epinions dataset has a higher prediction accuracy during iterations 4 to 20 compared to the period of one season. Here, one month is an accurate short-term period. This period has the best short and long-term performances compared to one season. One year period in Netflix provides a greater prediction performance from iterations 1 to 5 in Fig. 5. Across iterations 5 to 20, the period of one year is more precise than one month but less than the period of one season.

The temporal attitudes are learned by the LTO approach over 20 iterations. This helps to achieve precise predictions by optimizing the temporal weights of each duration. The periods of Netflix are 6 years, 42 seasons or 73 months. Experimental results show that LTO, in one season, achieved the highest prediction performance compared to its predicted performance using the year and month periods. This is because the duration of a season is an intermediate between a month and a year. Moreover, various customers’ activities are performed therein.

LTO approach under the scoring [0–1]

This subsection demonstrates the normalizing effect on the performance of LTO for reducing the RMSE under the scoring [0-1]. Figures 6, 7 and 8 track the effects of the temporal vectors in improving the prediction accuracy of the CF using the LTO approach. Figure 6 indicates that the RMSE of MovieLens for a week is better than that of a month under the rating scale [0–1]. Additionally, the RMSE during the period of two weeks is the best compared to those of a week and a month. This emphasizes the significance of the two-week period in learning the drift of customers’ interests and products’ popularity time decay.

Figure 7 shows the prediction accuracy using Epinions, Here, the period of one month has a significant effect on reducing the RMSE in iterations 3 to 20 compared to the effect of one season.

In Figure 8, the effect of a season using Netflix is equivalent to the effect of a year but more than the effect of a month during iterations 1 to 13. Iterations 13 to 20 has the sharpest accuracy prediction in the season compared to the period of one year and one month.

Figures 3–8 show the potential of BFOA in learning the temporal features by swarming in the dimensional time-space. The effects of the equivalent time periods show that the customers’ interests and the products’ popularity changed during these periods. The proposed approach will be evaluated by the current factorization and temporal approaches in the next subsection.

Comparison of the performances of CF, MF, and Temporal-based approaches

In this subsection, the LTO approach is evaluated by comparing its effectiveness in reducing RMSE values with other benchmark approaches. Both LTO and the benchmarks are used to predict the sparse scores as they all lower the RMSE values. Note that the lower the RMSE value, the higher the prediction accuracy of the CF approach. In Table 8, seven approaches are implemented to benchmark the prediction performance of LTO. The improvement in the prediction performance of the CF technique is represented by two scoring scales: [0–5] and [0–1]. First, the scores from 0 to 5 are provided by the users of the three experimental datasets, and the benchmarks are categorized into three parts. The first part contains the prediction accuracy of the CF technique by RMSE for the rating matrix of the active user without predicting the sparse rating scores. The second part contains the prediction accuracy of the CF by two factorization approaches which are Neighbour-based Baseline (Bell & Koren, 2007) and the Ensemble Divide and Conquer (Al-Hadi et al., 2016). These approaches are used to solve the sparsity issue and as well learn the accurate factorization features. From the evaluations, it is observed that the prediction performance of Ensemble Divide and Conquer is better than that of the CF and the Neighbours-based Baseline approach. However, the approaches in the second category have weaknesses in learning the overfitted predicted scores and temporal issues (such as drift and decay). For the third category, five temporal approaches are used to solve five issues i.e., sparsity, accurately learning latent features, overfitting, drift, and decay. The Temporal Dynamics (Koren, 2009) has a good prediction performance in solving these issues but it has a weakness in learning the personalized features using the equaled time slices. Temporal Integration using Netflix performs better compared to Temporal Dynamics and Short-Term based Latent approach. However, Temporal Integration has a weakness with respect to drift and decay. Short Temporal-based Factorization (Al-Hadi et al., 2017a) addresses all issues except popularity decay. It improves the prediction performance of the CF technique when compared to the above approaches using MovieLens and Netflix. The performance of the Short Temporal-based Factorization approach is lower than that of the Ensemble Divide and conquer approach using the Epinions dataset because the recorded timestamp factors are registered using the number of months only (which represent weak temporal features). Distinctively, the LTO approach addresses all issues including the limitations in the benchmark schemes. It improves the prediction performance of the CF technique through a perfect combination of various factorization and temporal features. It also tracks the drift of users and the decay of items throughout the learning process. Table 8 shows that LTO exhibits superior prediction performance when compared to all benchmarks.

Table 8:

The RMSE of several prediction approaches using three datasets.

Approach	Scoring [0–5]			Scoring [0–1]
	MovieLens	Epinions	Netflix	MovieLens	Epinions	Netflix
CF	0.9573	1.0536	0.9983	0.1915	0.2107	0.1997
Neighbors based Baseline (Bell & Koren, 2007)	0.9613	1.0562	0.9982	0.1923	0.2112	0.1996
Ensemble Divide and Conquer (Al-Hadi et al., 2016)	0.9481	1.0351	0.973	0.1896	0.2069	0.1948
Temporal Dynamics (Koren, 2009)	0.9514	1.0486	1.0173	0.1903	0.2097	0.2035
Short-Term based Latent (Yang et al., 2012)	0.9613	1.0562	0.9982	0.1923	0.2110	0.1996
Temporal Integration (Ye & Eskenazi, 2014)	0.9557	1.0563	0.9982	0.1912	0.2112	0.1996
Short Termporal based Factorization (Al-Hadi et al., 2017a)	0.8716	1.0492	0.9704	0.1771	0.2088	0.1900
LTO	0.7933	0.9887	0.8136	0.1642	0.199	0.1564

DOI: 10.7717/peerjcs.331/table-8

The normalization process for rating scores has reduced the RMSE values by almost 80 % due to the percentage difference between the rating scores [0–5] and [0–1]. For example, the percentage difference of LTO approach in MovieLens is calculated using Eq. (18). (18)${\displaystyle PercentageDifference=1-\frac{Scorin{g}^{scal{e}_1}-Scorin{g}^{scal{e}_2}}{Scorin{g}^{scal{e}_1}},}$ where Scoring^scale₁ is from 1 to 5 and Scoring^scale₂ is from 0 to 1. The percentage difference between 0.7933 and 0.1642 is 79.3%. Figure 9 indicates the high prediction accuracy achieved by the LTO approach for the three datasets when compared with the benchmark methods. Additionally, the graphs show the positive impact of the normalization in reducing the RMSE by around 80%.

Comparison of the output prediction scores of CF and LTO

The CF technique utilizes the similarity function to calculate the similarity between the active user and the common users or neighbors. In the second stage, CF utilizes the prediction function according to the similarity values to recommend items to the active user. However, CF’s predicted scores are not accurate because of the sparsity values in the rating matrix. This is solved using the LTO approach. Table 9 is an example showing the improved accuracy of the predicted scores achieved by the LTO approach. In this example, the short duration is one year. Active users rate items from 1 to 5, where 1 and 2 indicate unlike items while 3, 4, and 5 indicate the liked items.

Table 9:

Feedback prediction scores by CF and LTO approaches.

items	i1	i2	i3	i4	i5	i6	i7	i8	i9	i10	i11	i12	i13	i14
Active user	1	4	1	3	1	2	2	1	5	2	5	2	5	2
CF	2.9	2.9	2.9	3.1	3.3	2.7	2.8	2.5	2.9	2.4	2.9	2.4	2.9	2.9
	R	R	R	R	R	R	R	R	R	N	R	N	R	R
LTO	0.4	4.3	2.4	2.3	2.2	2.5	1.8	2.4	3.9	2.3	5	3	3.7	3.3
	N	R	N	N	N	R	N	N	R	N	R	R	R	R

DOI: 10.7717/peerjcs.331/table-9

The CF predicts rating scores from 2.4 to 3.3. This provides the active users with recommended and unrecommended items (denoted by R and N, respectively in Table 9). On the other hand, the LTO approach predicts rating scores from 0.4 to 5.0 which provides more accurate prediction compared to the CF. Figure 10 visualizes the output in Table 9 and indicates the high prediction performance of the LTO approach.

Conclusion

The CF performance is affected by several factors including changes in the customer’s taste, time decay in the popularity of products, data sparsity, and the overfitting in the predicted rating scores. Prior research has attempted to enhance the CF’s prediction function by integrating the long and short-term preferences via the temporal interaction method (Ye & Eskenazi, 2014) with the factorization factors. However, the achievement is low. The goal of the long temporal-based factorization approach (Al-Hadi et al., 2018b) is to solve the popularity decay problem and understand the drifting taste of clients over the long-term. On the other hand, the main focus of the short temporal-based factorization approach is to understand the behaviors of customers and solve the drift issue in the short-term. Nonetheless, there are several limitations associated with predicting popularity decay issues as well as the drift in customers’ preferences over time.

To address these problems, the LTO approach presented in this paper integrates both short and long-term preferences. It utilizes the k-means and BFOA method which derived the fitness value by combining the signal and RMSE values. The swarming function represents the preferences of the short-term based on the sensitivity of bacteria to rich nutrients or dangerous signals. According to the empirical findings, a higher prediction precision is achieved by the LTO approach compared to the benchmark approaches. This is attributed to the temporal-based factorization approach and its ability to enhance the accuracy of the CF technique by understanding the temporal behaviors in both long and short preferences.

Possible extensions of this work include integrating the LTO approach with other factorization features such as neighbors’ latent feedbacks. This would contribute to addressing issues such as the cold start when recommending new items to active users. Besides, the genre’s features of movies can be integrated with the factors that are utilized by LTO approach for the purpose of addressing challenges of new items by the MovieLens and Yahoo! Music datasets.

Supplemental Information