Enhanced context-aware recommendation using topic modeling and particle swarm optimization

Abstract

Context-Aware Recommender System (CARS) suggests more relevant services by adapting them to the user’s specific context situation. Nevertheless, the use of many contextual factors can increase data sparsity while few context parameters fail to introduce the contextual effects in recommendations. Moreover, several CARSs are based on similarity algorithms, such as cosine and Pearson correlation coefficients. These methods are not very effective in the sparse datasets. This paper presents a context-aware model to integrate contextual factors into prediction process when there are insufficient co-rated items. The proposed algorithm uses Latent Dirichlet Allocation (LDA) to learn the latent interests of users from the textual descriptions of items. Then, it integrates both the explicit contextual factors and their degree of importance in the prediction process by introducing a weighting function. Indeed, the PSO algorithm is employed to learn and optimize weights of these features. The results on the Movielens 1 M dataset show that the proposed model can achieve an F-measure of 45.51% with precision as 68.64%. Furthermore, the enhancement in MAE and RMSE can respectively reach 41.63% and 39.69% compared with the state-of-the-art techniques.

Keywords

Collaborative filtering context topic modeling PSO LDA sparsity problem

1 Introduction

With the advent of the World Wide Web and big data, recommender systems (RSs) are becoming more and more popular [1, 2]. The purpose of these systems is to suggest different services to different users. In the existing literature, researchers have presented numerous recommendation approaches such as collaborative filtering (CF), content-based filtering (CBF) and hybrid techniques. Collaborative filtering is one of the most popular methods. It is based on the rating histories to calculate the similarities between users or items [3]. It suggests similar objects to those favored by the user in the past, or items which have been liked by similar users. The most popular technique in CF is the nearest neighbor, where a subset of k suitable users (or items) is chosen based on their similarities to the target one [4]. Then, a weighted aggregate of ratings is used to produce predictions for the active customer. The content-based algorithms are relying on the content of items [5]. They propose objects that are similar to those previously chosen by the user. On the other hand, the hybrid techniques are able to integrate the power of different algorithms and methods to create a more robust recommender system [6].

Furthermore, many researchers have shown that incorporating context in the recommendation process seems one of the promising approaches to improve prediction accuracy [7 –9]. Context-aware recommender system (CARS) has emerged as an efficient tool for enhancing the recommendations by exploiting contextual information in order to recommend items that are relevant to evolving customer needs [10]. However, the notion of context is complex and not easily integrated into recommendation process [11]. Moreover, there are several challenges to be resolved in CARSs, including the sparsity problem, which refers to situations where insufficient data are available for inferring similar users or items [12]. Therefore, the sparsity problem becomes more severe when user preferences are filtered with contextual factors [13]. In fact, using many contextual parameters increases the data sparsity and few context factors in recommendations fail to bring the contextual impact [12].

This study proposes a hybrid method for alleviating the sparsity problem in CARS by considering the latent properties of items to calculate the similarity scores, and by directly incorporating context into the prediction process instead removing items that are irrelevant in a given context. Thus, a context-aware recommender model which is called Topic Modeling and Particle Swarm Optimization based Collaborative Filtering algorithm (TMPSO-CF) is proposed. It combines latent interests and explicit contextual features using Latent Dirichlet Allocation (LDA) and Particle Swarm Optimization (PSO). Firstly, the dimensionality space is reduced by extracting the latent properties from the unstructured texts describing the items. The latent context is modeled as numeric vectors that are automatically learned through the application of LDA, which is probably the most common topic model suitable for low dimensional representation [14]. LDA is an unsupervised technique with proper underlying generative probabilistic semantics that makes sense for the type of data that it models [15]. Then, a weighting function is proposed to integrate explicit contextual factors (numerical and/or categorical data), such as user’s age, user’s occupation, movie’s runtime, etc., with their degree of importance in the prediction scores calculation. Specially, the time evaluation is introduced into the proposed weighting function in order to increase the importance of the most recently accessed items because the user’s interests change over the time [16]. Usually, objects which have recently been consumed by the user play a greater role in generating future recommendations than items with early access [17]. In addition to that, TMPSO-CF uses a particle swarm optimization algorithm for assigning suitable weights to different contextual features. PSO have been commonly applied in many optimization problems having very satisfactory results [18]. In most cases, it outperforms many other approaches applied to the same optimization problems [19]. It has also shown a faster convergence rate than other evolutionary algorithms on some problems [20]. Moreover, PSO has very few parameters to adjust, which makes it particularly easy to implement.

The main contributions of this paper are summarized as follows:

A novel hybrid context-aware recommender system, which is called TMPSO-CF, is proposed. It combines collaborative filtering algorithm and content based method in order to enhance the prediction accuracy, alleviate the sparsity problem in the sparse datasets and handle the contextual sparsity issue in CARS. It considers also the influence of individual contextual features during the recommendation process.

An application of TMPSO-CF for the Movielens 1 M dataset.

The performance of the proposed model on the Top-N Recommendations at the optimal neighborhood size (k) is evaluated.

A comparison of TMPSO-CF with other well-known methods.

The remainder of this paper is organized as follows: Section 2 reviews related works in context-aware recommender system and the use of LDA and PSO in these systems. In Section 3, the proposed TMPSO-FC model with a detailed description of how the system employs the LDA and PSO is provided. Section 4 introduces the dataset and the evaluation metrics. It also discusses the results of the experiments. Section 5 presents the performance of TMPSO-CF at optimal neighborhood. Finally, section 6 provides conclusion and future work.

2 Background and related works

2.1 Context-aware recommender system

Context-aware recommender systems (CARSs) are promising way of generating more relevant services to users. They produce personalized recommendations in accordance with users’ current context [21]. Consequently, the rating function is modeled in a multi-dimensional space as R: Users×Items×Contexts ⟶ Ratings [22]. Several context-aware recommendation algorithms have been proposed and developed [10, 16]. They explore various ways in which contextual factors such as time, location, companion, etc. can be integrated into the recommendation models in order to enhance the prediction quality [23]. The temporal factor has a major impact on people’s desires and allows tracking the evolution of user preferences over time [17, 24]. Lin and Chen [8] proposed a probabilistic method to capture the time dynamics of recommender systems. Essentially, the model differentiates between the recent and early data. In the hidden Markov model, the recent ratings are used to capture the change of the preferences of both the users and items over time. However, the early ratings are used to produce the overall prior distribution of the random variable of both users and items.

Based on the fact that only relevant contextual information should be taken into consideration in a recommender system, Braunhofer and Ricci [25] presented a model which predicts the contextual factors influencing customers when evaluating an item, in order to use only relevant contextual information in the recommendation process. In their research work, Zhou et al. [26] selected the optimal features based on the similarity between a feature and a set of other ones to remove the redundancy. Thus, a content-context interaction graph model that completely fuses and captures the interactions between content and contexts is presented. Tensor Factorization (TF) can also be used to model the context [10, 27]. It extends the traditional two-dimensional Matrix Factorization problem into a n-dimensional version by integrating contextual information [28].

2.2 Latent Dirichlet allocation in CARS

Latent Dirichlet allocation (LDA) is a generative probabilistic and statistical modeling method to extract topics from a text corpus [15]. The basic idea behind LDA is that the groups of terms which are contextually similar constitute different topics. LDA models each document as a random mixture of latent topics, and it defines each topic by a word distributions. It generates the probability distributions of topics p (t|d) for each document d. Each topic t (t ∈{1,..., T}) is composed of the word probabilities p(w_j |t) for words w_j, j = 1,...,V. Where V is the size of the vocabulary and T is the pre-defined number of topics. The topic density in the model is controlled by Dirichlet hyperparameters α and β, which play an important role in learning accurate topics. In fact, LDA’s popularity is due to its simplicity and modularity [29].

Given N items with a textual description of the item i (i ∈{1,..., n}) having V_i words, the generative process is as follow [15]:

Step 1: Choose a multinomial distribution ϕt for topic t (t∈{1,...,T}) from a Dirichlet distribution with parameter β.

Step 2: Choose a multinomial distribution θi for item i (i∈{1,..., n})from a Dirichlet distribution with parameter α.

Step 3: For a word w_j (j∈{1,...,V_i}) in the item i:

Step 3.1: Select a topic zj from θi.

Step 3.2: Select a word wj from ϕzj.

The variables θi are item-level variables, sampled when per item. zj, wj are word-level variables. They are sampled when for each word in each text description item.

In fact, latent context is very useful in increasing the efficiency of recommender systems. Hence, unsupervised tools such as LDA have been used to learn the latent characteristics in RS [29, 30]. Yan et al. [31] presented a unified video recommendation solution for YouTube by transferring and incorporating the social and content information of users through Twitter network. So, LDA model is applied by considering users as documents and their hashtags as words. Lin et al. [30] used user reviews from the Airbnb platform for deriving the features of the products and creating customer preferences. The authors implemented the LDA technique to infer both the features and preferences.

2.3 Particle swarm optimization in CARS

Particle swarm optimization (PSO) is a population based technique firstly developed in 1995 by Kennedy and Eberhart [32]. In the PSO algorithm, a swarm is a group of particles which represent candidate solutions. For a d-dimensional space, the particles correspond to d-dimensional vectors. The position and velocity of the ith particle are represented by X_i = (x_i1, x_i2,..., x_id) and V_i = (v_i1, v_i2,..., v_id), respectively. At each time step, each particle i keeps the information about its personal best position (pbest P_i(t)) and the swarm best position (gbest g_i(t)). The particles move throughout the search space by updating their velocities and their positions according to Equations (1) and (2), respectively. For every particle, i∈{1, 2, ... ., N}.

$\begin{matrix} v_{j} (t + 1) = {wv}_{j} (t) + c_{1} r_{1} (P_{j} (t) - x_{j} (t)) \\ + c_{2} r_{2} (g_{j} (t) - x_{j} (t)) \end{matrix}$ (1) $x_{j} (t + 1) = x_{j} (t) + v_{j} (t + 1)$ (2)

Where N is the number of particles in the swarm. c₁ and c₂ are positive acceleration constants. c₁ expresses the confidence of a particle in itself, while c₂ expresses the confidence of a particle in its neighbors. r₁ and r₂ are random numbers within the range [0, 1]. The inertia weight w is an important parameter to strike a better balance between global and local exploitation. The linear decreasing inertia weight has been widely used to update the value of w according to Equation (3) [33]. $w (tr) = w_{start} - tr \times (w_{start} - w_{end}) / {Nb}_{\max}$ (3)

Where Nb_max is the maximum number of iterations, tr is the iteration index, w_start and w_end are the maximum and minimum values of the inertia weight, respectively.

PSO has been used by many researchers to select the features or assign weights to different factors in the recommender systems [34, 35]. The work of Ujjin and Bentley [36] applied the PSO to learn users’ personal preferences and provide the recommendations based on the adjusted Euclidean distance. Katarya and Verma [37] combined k-means clustering algorithm with Fuzzy c-means and PSO to classify the types of movies according to users. Firstly, k-means is employed to provide initial parameters to the PSO algorithm, and then PSO is used to optimize Fuzzy c-means clustering. Sumathi et al. [32] combined particle swarm optimization with user access based ranking algorithm for the new user recommendation. The authors referred each user as a particle and applied PSO to find similar users for an active one.

It is observed by surveying the existing literature that collaborative filtering approaches and content-based recommendations were tried numerously in the context-aware recommender system area. However, most studies have mainly focused on improving the measure of similarity; with relatively little research on prediction score models, although it is extremely important to improve the accuracy of the recommendations [38]. Moreover, collaborative methods are among the most widely used and successful techniques [17]. Unfortunately, their success is based on the availability of a sufficient number of available ratings. In fact, when the data are too sparse, the information that can be used in exploring similar users or items is too little to make a recommendation. Thus, several works in CARS focused on using the unitary context because multidimensional contexts will aggravate the sparsity problem. There is also a lot of research that claims to improve the quality of recommendations through the combination of content based methods and collaborative information. The greatest advantage of the hybrid approach is that it can overcome the drawbacks of both methods and benefit from their advantages. However, it will also make the final model more complicated. Therefore, the method of mixing needs to be analyzed according to the actual situation. This is the setting for the proposal presented in this study, i.e. the design of a hybrid recommendation system with the aim of solving the sparsity problem in CARS, especially when more than one kind of context is used. This paper proposes a mixed model combining the item-based k nearest neighbor algorithm with content based enhancements via LDA for improving the accuracy of the recommendations. The proposed approach also takes advantage of the PSO algorithm to estimate the contribution of contextual information instead of context selection.

3 Proposed methodology

The proposed system aims to enhance the prediction accuracy, alleviate the sparsity problem in the sparse datasets and handle the contextual sparsity issue in CARS. It exploits the hidden properties of items, from their textual description, and directly integrates context in the prediction process by introducing a weighting function, which combines contextual information with their degree of importance.

3.1 TMPSO-CF algorithm

Figure 1 illustrates the complete system flow. The textual descriptions of items are passed to the preprocessing unit before the LDA training, in order to filter out noisy data and non-informative words that do not add any distinctive character in the objects (removal of stop words and punctuations, Lemmatization tokenization, stemming, etc.). Then, the LDA algorithm is applied to determine the optimal number of topics and represent each item in the latent topic space using the item-topic distributions as its feature vector. In the other hand, the PSO algorithm is employed to learn and optimize weights of different contextual features in order to control their contribution in the recommendation process. In fact, the contextual data in numeric form are directly normalized. However, the categorical features must be converted into numeric values before the normalization. Finally, when the predefined number of iterations is reached, we get the optimal weights of features, which will be used to calculate the predictions.

Fig. 1

General workflow of TMPSO-CF model.

The following pseudo-code will describe the proposed model in details.

Input:

The users set: U = {u₁, ... ., u_m}, where m is the number of users.

The items set: I = {i₁, ... ., i_n}, where n is the number of items.

The rating histories: R = {R_u,i| u ∈ U; i ∈ I}.

The rating time matrix.

The content information of items.

The topic-word distributions.

The item-topic distributions.

The contextual features set: P = Pⁿ ∪ P^c = {Pⁿ₁, ... ., Pⁿ_ln} ∪ {P^c₁, ... ., P^c_lc}, where ln is the number of numerical features and lc is the number of categorical features.

The neighborhood size (k).

The optimal weights of contextual parameters: w_t and w_i ( $i = \bar{1, l}$ ), where l = ln + lc is the total number of contextual features.

Output: - Prediction P_u,i for any (u, i) ∈ (U, I)

List of TopN items.

TMPSO-CF (u, i)

Step 1: Represent the profile of the item i by the content information (such as title, abstract, etc.).

Step 2: Applied the data preprocessing on the textual description of the item i.

Step 3: Generate the item-topic distributions of the item i using the LDA.

Step 4: Calculate the similarity degrees between the items i and other ones using the Hellinger distance (HD) according to Equation (4): $sim (i, j) = 1 - HD (i, j) = 1 - \frac{1}{\sqrt{2}} \sqrt{\sum_{h = 1}^{T} {(\sqrt{i_{h}} - \sqrt{j_{h}})}^{2}}$ (4)

Where, T is the number of topics, i_h and j_h are the probability distributions of the items i and j, respectively, over the topic h. The Hellinger distance varies from 0 to 1. The upper bound 1 (i.e. sim(i, j) = 0) means the distance is maximal and thus the two items are very different. The lower bound 0 means the two items are very similar. So, they share the same topic distributions.

Step 4: Select the k nearest neighbors of the item i.

Step 5: For each neighbour j,

Step 5.1: Calculate the time coefficient τ, which is a monotonic decreasing coefficient with an exponential form in order to express the changes of the user’s interests over time by making his most recent evaluations more significant than the older ones. The parameter τ can be expressed by Equation (5): $τ = e^{- λ}$ (5) $λ = \frac{t_{B} - t_{A}}{Δ t}$ (6)

Where, w_t represents the weight of the time evaluation. t_A is the time of the latest evaluation given by the user u to an item sharing at least one genre or one topic with the item j. t_B represents the time of the latest evaluated item by the user u. Δt corresponds to the average duration between each two successive ratings of user u.

Step 5.2: Each categorical feature p Î P^c (such as, music and movie genres, country names, etc.) must be converted to a numerical form. It is represented by a number from 1 to n_c, where n_c is the number of distinct categories of the feature p.

Step 5.3: Before calculating the proposed weighting function, each feature value must be normalized using Equation (7) in order to bring all the contextual parameters on the same scale. $h (x_{p}) = \frac{x_{p} - \min (p)}{Δ p}$ (7)

x_p is the value of the feature p (the value of a numerical feature or the numerical value of a categorical feature). Δp is the interval of variation of the feature p. min(p) is the minimum value of the feature p.

Step 5.4: Calculate the proposed weighting function f(u, j) according to Equation (8): $f (u, j) = \sum_{p = 1}^{l} w_{p} [h (x_{p})] + w_{t} [τ]$ (8)

x_p represents the value of the contextual feature p about the user u or the item j. w_t and w_k represent the weight of the time evaluation and the weight of the feature p, respectively, and satisfying the condition, $\sum_{p = 1}^{l} w_{p} + w_{t} = 1$ (9)

Step 6: Calculate the rating prediction by the sum of the ratings given by the user u on the items similar to i. Each rating is weighted by the corresponding similarity sim(i, j) between items i and j, and by the value of the proposed function f(u, j), which is used to model the context of the user u for the item j. Formally, the predicted rating P_u,i is given by Equation (10): $P_{u, i} = \bar{R_{u}} + \frac{\sum_{j = 1}^{k} (R_{u, j} \times sim (i, j) \times f (u, j))}{\sum_{j = 1}^{k} (sim (i, j) \times f (u, j))}$ (10)

Where, R_u,j is the rating of user u on item j. k is the neighborhood size of the item i.

Step 7: Generate a list of the top N items that the user u will appreciate.

Illustrative example

Illustrative examples of the data normalization and the parameter λ calculation are presented respectively in Table 1 and Fig. 2. Given five users u₁, u₂, u₃, u₄ and u₅ to whom we want to recommend movies 7, 5, 3, 1 and 2, respectively. The data in Table 1 include the duration of movies (r), occupation (o) and age (g) of the users, with age Î [7, 73] and runtime Î [24, 238]. Firstly, each occupation is represented by a unique integer number between 1 and 21, such as 1 for “other”, 2 for “academic/educator”, 3 for “artist”, etc. Then, the values of age, occupation and runtime are normalized according to Equation (7) in order to calculate the function f(u, j). In the other hand, Fig. 2 describes the movies evaluated by the target user in the past. According to Equation (6):

Table 1

An illustrative example of data normalization.

User ID	Movie ID	Input data			Normalized data
		a	o	r	h_a	h_o	h_r
u₁	7	24	Technician	127	0.2575	0.8	0.4813
u₂	5	53	Programmer	106	0.6969	0.1	0.3831
u₃	3	23	Writer	25	0.2424	0.95	0.0046
u₄	1	24	Technician	81	0.2575	0.8	0.2663
u₆	2	42	Executive	104	0.5303	0.35	0.3738

Fig. 2

An illustrative example of parameter λ calculation.

$\begin{matrix} Δ t = \frac{[(t_{5} - t_{4}) + (t_{4} - t_{3}) + (t_{3} - t_{2}) + (t_{2} - t_{1})]}{4} \\ λ = \frac{1820687}{481853.25} = 3.77850933 \end{matrix}$

3.2 PSO in learning weights of the features

In this paper, the position of each particle represents the weight vector of l contextual features (w₁, w₂, ... .., w_p, ... .., w_l), where w_p (0 < w_p≤1) is the weight of the feature p and $\sum_{p = 1}^{l} w_{p} = 1$ . At the initialization stage, the velocities and the positions of the particles are initialized randomly. Moreover, the RMSE measure is treated as the fitness function. It is given by Equation (11). Thus, to achieve the lowest fitness, it is necessary to search the optimal position. $fitness = \sqrt{\frac{\sum_{(u, i) \in Q} {(P_{u, i} - P_{u, i}^{'})}^{2}}{| Q |}}$ (11)

|Q| is the cardinality of the training set. P’_u,i and P_u,i are real rating and predicted rating, respectively. The predictions calculation is based on the proposed weighting function defined in Equation (8).

4 Experimental evaluation

To demonstrate the effectiveness of the proposed model, several experiments have been conducted on the Movielens 1 dataset and CMU Movie Summary 2 Corpus. The experiments use a five-fold cross validation scheme to randomly splitting Movielens dataset into training and testing, where 80% of the data are used for training and 20% are used for testing. All the experiments are repeated five times, and the evaluation outcome is the aggregated result from all folds.

4.1 Dataset

Movielens is widely used to evaluate recommender systems. This dataset is provided by the University of American Minnesota GroupLens project group. The ML-1M dataset consists of 6,040 users, 3,900 movies and 1,000,209 ratings [39]. The sparsity of the ML-1M dataset is 95.754%. The CMU Movie Summary Corpus is also used in this paper. It contains plot summaries for 42,306 movies and metadata for 81,741 movies such as genre, release date, runtime, languages, etc. Each movie is indexed by a Wikipedia Movie ID.

4.2 Methods compared

The following algorithms are compared in our experiments:

Item-CF denotes the traditional item-based collaborative filtering algorithm using Pearson correlation coefficient to measure the similarity between the movies.

LDA-CF denotes the item-based collaborative filtering using the LDA algorithm to identify the latent properties of the movies from their textual descriptions (plot summaries, titles and genres). The similarity scores are computed in the latent topic space using the Hellinger distance to estimate the neighborhood of movies.

Context-CF represents the item-based collaborative filtering algorithm using the proposed weighting function defined in Equation (8), with w_t = 1 and w_p = 1 for each p ∈{1, ... , l}. The proposed function considers the user’s age, user’s gender, user’s occupation, the movie’s runtime and the evaluation time. Furthermore, the movie recommendations are provided based on the Pearson correlation coefficient.

PSO-CF is the item-based collaborative filtering using the PSO algorithm to assign suitable weights to contextual features (user’s age, user’s gender, user’s occupation, movie’s runtime and evaluation time), which are considered by the proposed weighting function defined in Equation (8). The Pearson correlation coefficient is used to measure the similarity scores between the movies.

TMPSO-CF is our new hybrid method applied for the movie recommendations. It combines latent properties and explicit contextual information with their degree of importance. The first step is based on the LDA to identify latent properties of the movies from their textual descriptions (plot summaries, titles and genres), and then calculate the similarity scores using the Hellinger distance. The second step uses the PSO algorithm to optimize the proposed weighting function defined in Equation (8), which combines five explicit features (user’s age, user’s gender, user’s occupation, movie’s runtime and evaluation time).

4.3 Parameter setting

For all experiments with the LDA model, the number of topics T is fixed to 25, the parameters α and β are set to 0.01 and 0.02, respectively. The training of the LDA is based on the Gibbs sampling method to estimate the latent parameters θ and ϕ. For the PSO algorithm, the inertia weight w is linearly decreasing from 0.9 to 0.4. c₁ and c₂ are selected as 1.4945, and the maximum number of iterations is 80.

4.4 Evaluation Metrics

The proposed model is evaluated from two differ-ent perspectives: predictive accuracy and classifica-tion accuracy. Predictive accuracy measures the accuracy of a prediction value on a target item. The Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) are the frequently used metrics in predictive accuracy [40]. MAE represents the average of the absolute difference between the predicted ratings and real ratings. The formulas for RMSE and MAE are defined in Equations (11) and (12), respectively. $MAE = \frac{\sum_{(u, i) \in Q} | P_{u, i} - P_{u, i}^{'} |}{| Q |}$ (12)

|Q| is the cardinality of a training set. P’_u,i and P_u,i are real rating and predicted rating, respectively.

Classification accuracy evaluates the rate of cor-rect predictions. Precision, recall and F-measure are the three popular metrics for classification accuracy. Precision is a percentage of recommended items that are relevant while recall is a percentage of the rele-vant item that is recommended [40]. Precision, recall and F-measure are defined in equations Equation (13), Equations (14) and (15), respectively: $Precision = \frac{Number of correct recommendations}{Total number of recommendations}$ (13) $Recall = \frac{Number of correct recommendations}{Total number of relevant recommendations}$ (14) $F - measure = \frac{2 \times Precision \times Recall}{Precision + Recall}$ (15)

4.5 Experimental results and analysis

In this section, each user is considered to be an active user and the item ratings for each of them are predicted. The recommendation algorithms are evaluated by comparing the predicted ratings with the actual ratings. Since a different neighborhood size k leads to different recommendation results, the impact of k is also considered in this experiment. So, recommendation results are calculated and compared under the condition of different k.

Figures 3 and 4 illustrate the evolution of the MAE and RMSE through the neighborhood size. It can be seen that the errors in metrics MAE and RMSE decrease as the number of neighbors increases for Item-CF approach. However, for the other methods, the error starts with a relatively high value and then it reaches the lowest points at k = 40. The MAE and the RMSE of TMPSO-CF are 0.7043 and 0.9071, respectively. It is also observed that the values of MAE and RMSE increase with the increase of the neighborhood size in the range of 40–100. This can be explained by the fact that these methods achieve a better prediction quality when more similar items are taken into account. Among all the methods, TMPSO-CF always has the best MAE and RMSE no matter how the number of neighbors changes, which means that the predicted scores by TMPSO-CF are closer to the original scores. It is also obvious from Figs. 3 and 4 that the methods which consider the degree of importance of contextual information (i.e. TMPSO-CF and PSO-CF) perform better than the other methods. Furthermore, compared with the closest competitor (PSO-CF), TMPSO-CF has a 3.55% advantage for MAE and 5.55% for RMSE. The main reason is that TMPSO-CF also considers hidden context, which is conducive to the improvement of prediction performance.

Fig. 3

MAE with different neighborhood size.

Fig. 4

RMSE with different neighborhood size.

Overall, the results of these experiments confirm that the model which combines the latent interests and explicit features with their degree of importance performs better than the other CF algorithms. On one hand, latent factors have a real meaning for recommendations. So, getting only several interest information about the target user allows discovering the possible items that this user may like, through the item–interest relationship [41]. On the other hand, the ratings produced at different times and in different context situations have different impacts on the prediction process [22, 42].

Figure 5 outlines the evolution of precision through the neighborhood size. The results indicate that the precision of all methods decrease when the number of neighbors increases. This means that the number of the relevant objects increases when more similar items are considered. Moreover, TMPSO-CF mimics PSO-CF, under different neighborhood size, and both perform significantly better compared with the remaining methods. The proposed TMPSO-CF enhances precision by 4.49% and 22.50% compared with PSO-CF and item-CF, respectively.

Fig. 5

Precision with different neighborhood size.

Figure 6 illustrates the evolution of recall through the neighborhood size. It can be seen that, the recall values of all methods become better as the number of neighbors increases. The results also confirm that TMPSO-CF exceeds the Item-CF, Context-CF and LDA-CF in recall by more than 17%. To sum up, the experiments verify that the TMPSO-CF enhance the prediction performance. This is beneficial to find the proper trade-off between efficiency and performance.

Fig. 6

Recall with different neighborhood size.

Figure 7 shows the evolution of F-measure through the neighborhood size. It is clear that the F-measure of all methods changes slightly with different numbers of neighbours. This means that these models perform stably in F-measure. Furthermore, the presented TMPSO-CF has better F-measure values than the other CF models. It achieves an F-measure value close to 0.45. Compared with the closest competitor PSO-CF, the TMPSO-CF has 12.54% advantage. However, the advantage of TMPSO-CF is more apparent compared to the other CFs discussed here. It exceeds 37%.

Fig. 7

F-Measure with different neighborhood size.

Overall, the results of this experiment confirm that TMPSO-CF has better performance in both successful and perfect predictions. It can identify the indirect correlations between the movies. Thus, it considers movies in the latent space to discover the hidden interests and utilizes the contextual factors with their degree of importance when computing predictions. This was an important reason why the TMPSO-CF performed better than the other models.

4.6 Statistical significance of TMPSO-CF

In order to verify the difference between the results of TMPSO-CF and the four competing methods over the five evaluation metrics, a paired t-test is computed. This statistical analysis was performed for a confidence level of 95%. The results of paired t-tests for the TMPSO-CF model are summarized in Table 2. For all the paired t-tests results, the p-values are less than the value of the significance level (i.e. less than 0.05), which proof the robust performance of the TMPSO-CF method. Consequently, the null hypothesis of mean equality is rejected and significant differences in prediction and classification accuracy of the proposed model are proven against the competing methods.

Table 2
Statistical significance of TMPSO-CF results using paired t-test

Metrics Measures Competing Methods TMPSO-FC

Item-CF LDA-CF Context-CF PSO-CF

MAE Mean 0.89126 0.87018 0.82256 0.75062 0.73213

SE Mean 1.18409E-02 6.5979E-03 7.86732E-03 4.53202E-03 6.53519E-03

SD 3.7444E-02 2.0864E-02 2.48786E-02 1.43315E-02 2.06661E-02

t-value –9.4278596 –20.769544 –58.402008 –5.3916024 –

p-value 2.9150E-06 3.2531E-09 3.19 E-13 2.189 E-04 –

RMSE Mean 1.14310 1.12626 1.05983 0.98395 0.94141

SE Mean 1.63848E-02 6.5427E-03 7.42307E-03 4.51554E-03 7.98435E-03

SD 5.18132E-02 2.0690E-02 2.34738E-02 1.42794E-02 2.52487E-02

t-value –8.86368468 –23.837373 –125.122579 –8.81319965 –

p-value 4.8359E-06 9.6071E-10 3.3791E-16 5.0664E-06 –

Precision Mean 0.52878 0.53573 0.58017 0.63135 0.65073

SE Mean 6.01372E-03 4.8961E-03 9.51202E-03 4.71603E-03 6.01166E-03

SD 1.90170E-02 1.54829E-02 3.00796E-02 1.49134E-02 1.90105E-02

t-value 62.870013 31.6521754 16.7153322 16.674434 –

p-value 1.6436E-13 7.697E-11 2.1952E-08 2.2427E-08 –

Recall Mean 0.09908 0.14730 0.21333 0.26086 0.32326

SE Mean 8.03597E-03 7.5052E-03 7.97252E-03 1.09140E-02 9.10115E-03

SD 2.54120E-02 2.3733E-02 2.52113E-02 3.45130E-02 2.87804E-02

t-value 100.194465 59.4384087 56.4142657 23.186152 –

p-value 2.4926E-15 2.7209E-13 4.3483E-13 1.2282E-09 –

F-measure Mean 0.16047 0.23213 0.31009 0.36362 0.43054

SE Mean 1.12494E-02 9.49009E-03 7.65781E-03 1.02516E-02 7.05610E-03

SD 3.55736E-02 3.00103E-02 2.42161E-02 3.24185E-02 2.23134E-02

t-value 55.0259932 53.379035 61.3673012 17.7355834 –

p-value 5.4382E-13 7.1433E-13 2.0425E-13 1.3065E-08 –

Metrics	Measures	Competing Methods	TMPSO-FC
MAE	Mean	0.89126	0.87018	0.82256	0.75062	0.73213
	SE Mean	1.18409E-02	6.5979E-03	7.86732E-03	4.53202E-03	6.53519E-03
	SD	3.7444E-02	2.0864E-02	2.48786E-02	1.43315E-02	2.06661E-02
	t-value	–9.4278596	–20.769544	–58.402008	–5.3916024	–
	p-value	2.9150E-06	3.2531E-09	3.19 E-13	2.189 E-04	–
RMSE	Mean	1.14310	1.12626	1.05983	0.98395	0.94141
	SE Mean	1.63848E-02	6.5427E-03	7.42307E-03	4.51554E-03	7.98435E-03
	SD	5.18132E-02	2.0690E-02	2.34738E-02	1.42794E-02	2.52487E-02
	t-value	–8.86368468	–23.837373	–125.122579	–8.81319965	–
	p-value	4.8359E-06	9.6071E-10	3.3791E-16	5.0664E-06	–
Precision	Mean	0.52878	0.53573	0.58017	0.63135	0.65073
	SE Mean	6.01372E-03	4.8961E-03	9.51202E-03	4.71603E-03	6.01166E-03
	SD	1.90170E-02	1.54829E-02	3.00796E-02	1.49134E-02	1.90105E-02
	t-value	62.870013	31.6521754	16.7153322	16.674434	–
	p-value	1.6436E-13	7.697E-11	2.1952E-08	2.2427E-08	–
Recall	Mean	0.09908	0.14730	0.21333	0.26086	0.32326
	SE Mean	8.03597E-03	7.5052E-03	7.97252E-03	1.09140E-02	9.10115E-03
	SD	2.54120E-02	2.3733E-02	2.52113E-02	3.45130E-02	2.87804E-02
	t-value	100.194465	59.4384087	56.4142657	23.186152	–
	p-value	2.4926E-15	2.7209E-13	4.3483E-13	1.2282E-09	–
F-measure	Mean	0.16047	0.23213	0.31009	0.36362	0.43054
	SE Mean	1.12494E-02	9.49009E-03	7.65781E-03	1.02516E-02	7.05610E-03
	SD	3.55736E-02	3.00103E-02	2.42161E-02	3.24185E-02	2.23134E-02
	t-value	55.0259932	53.379035	61.3673012	17.7355834	–
	p-value	5.4382E-13	7.1433E-13	2.0425E-13	1.3065E-08	–

SE Mean: Standard error mean; SD: Standard deviation.

In brief, the results are very promising, showing the advantage of combining the hidden and explicit contexts for movie recommendation system. However, the performance of the proposed model strongly depends on the neighborhood size. Herlocker et al. [4] studied the size of neighborhood in detail, and reached a conclusion that the size of the neighborhood have a significant impact on recommender system’s efficiency. So, it is crucial to select the optimal value of the neighborhood size in order to obtain the best trade-off between the efficiency and performance of the TMPSO-CF proposed model on Movielens 1 M dataset.

4.7 Computational complexity

The computational complexity of the proposed algorithm depends on the time complexity of model construction, and the time required to generate recommendations using this model. The former is performed offline, whereas the latter is accomplished online. The time complexity of training phase consists of two parts: including the time required to build both the LDA and PSO algorithms. The overall time complexity of LDA is O(I₁TnS), where I₁ is the iteration number of Gibbs sampling, T is the number of latent topics, n denotes the number of items and S is the average number of words.

In the next phase, the PSO is applied to find the weights for the contextual features. The time complexity of the PSO algorithm is O(I₂Cf_k), with I₂ is the number of iterations, C denotes the number of particles and f_k represents the fitness calculation complexity. For the fitness function, the similarity scores between items should be computed. The complexity of this step is O(Tn²). Furthermore, the time of O(nlogn) is spent for choosing the k nearest neighbors. Additionally, the time required to compute the proposed weighting function defined in Equation (8) is O(klm), where l is the size of the particles (i.e., the number of contextual features) and m is the number of users. Finally, the computational cost of the error function is about O(km). Thus, the complexity of training PSO is O(I₂ C(Tn² + nlogn + klm + km)) ≅ O(I₂ C(Tn² + m)). Therefore, the total computational complexity of the proposed model during the offline phase becomes approximately O(I₁TnS + I₂ C(Tn² + m)). In the online phase, the time complexity of the proposed model is O(k) for one rating prediction and O(n) for ranking the items and making recommendations.

5 Performance of TMPSO-CF at optimal neighborhood

5.1 Optimization of neighborhood size

The desirability function approach has been used for multiple evaluation metrics optimization of TMPSO-CF using the Design Expert Software. This approach involves specification of the individual desirability function (d_i) for each evaluation metric (Y_i) by assigning to the estimated values a score ranging from 0 (very undesirable) to 1 (very desirable). The optimization module searches for a neighborhood size (k) in the range of 10–100 that simultaneously satisfies the requirements placed on the evaluation metrics. During the optimization process, the aim was to find the optimal value of neighborhood size (k) in order to produce the lowest values of MAE and RMSE with the maximal values of precision, recall and F-measure. The constraints used during the optimization process are summarized in Table 3 and the optimal solution is reported in Table 4.

Table 3
Constraints for optimization of neighborhood size (k)

Condition Goal Lower limit Upper limit

k In range 10 100

MAE Minimize 0.70432 0.75795

RMSE Minimize 0.90713 0.97482

Precision Maximize 0.63425 0.68635

Recall Maximize 0.27048 0.35521

F- Measure Maximize 0.38794 0.45512

Condition	Goal	Lower limit	Upper limit
k	In range	10	100
MAE	Minimize	0.70432	0.75795
RMSE	Minimize	0.90713	0.97482
Precision	Maximize	0.63425	0.68635
Recall	Maximize	0.27048	0.35521
F- Measure	Maximize	0.38794	0.45512

Table 4

Optimal solution

k	MAE	RMSE	Precision	Recall	F-Measure
37.40	0.7054	0.9103	0.6606	0.3113	0.4231

For clear assessment, desirability value of each individual factor and responses associated with the modeling are shown in Fig. 8. The MAE and RMSE are found to be more desirable with 0.97109 and 0.97056, respectively. It can be said that the TMPSO-CF achieves the best compromise between the efficiency and performance with a neighborhood size equal to 37. Similar results were reported in [4], where Herlocker et al. suggest that, in the real-world situations, a neighborhood of 20–60 neighbors is reasonable to be used for making predictions.

Fig. 8

Bar chart of the desirability optimization.

5.2 Performance of TMPSO-CF

To further check the performance of the proposed algorithm, Figs. 9 and 10 illustrate the comparisons between the used models in different number of recommendations (TopN). In these experiments, the neighborhood size is fixed to 37. As can be seen from Fig. 9, the precision results of TMPSO-CF significantly outperform the others in returning relevant movies for all TopN values. In fact, compared with the Item-CF algorithm, the TMPSO-CF has a remarkable improvement; it is about 28% when the TopN is 5. The average improvement of TMPSO-CF compared with the Item-CF, LDA-CF, Context-CF and PSO-CF is 29.67%, 27.10%, 17.28% and 5.7%, respectively. This is because the TMPSO-CF is able to remove movies that are less related to the target one. Furthermore, it is established that the precision decreases as the number of recommendations increases. Figure 10 compares the recall rates of user interesting movies. Contrary to the precision results, the recall values increase smoothly as the TopN increases. It’s also apparently that TMPSO-CF still provides greater recall rates with each value in the number of recommendations. Compared with PSO-CF algorithm, the TMPSO-CF can improve 6.86% when the TopN is 25. Moreover, the average improvement of TMPSO-CF compared with Context-CF and PSO-CF is 43.37% and 11.83%, respectively.

Fig. 9

Precision performance with different TopN.

Fig. 10

Recall performance with different TopN.

Furthermore, Table 5 reports comparison results of the proposed model with the best results of the state-of-the-art RS’s methods [43 –45] in MovieLens 1 M. It is clear that TMPSO-CF produces significant improvement in terms of MAE by 14.36%, 4.48% and 41.63% for GA + GELS, BiUCF and Random method, respectively. Moreover, the enhancement of TMPSO-CF in RMSE compared with GA + GELS, BiUCF and Random method is 13.87%, 12.7% and 39.69%, respectively. However, the percentage error between TMPSO-CF and NMF, SlopeOne, K-NN, Centered K-NN, Co-Clustering and Baseline is insignificant for both MAE and RMSE. In fact, the performances of these methods are close, and they all have good application potential.

Table 5

Performance of TMPSO-CF compared with state-of-the-art methods

Authors	Algorithms	MAE	RMSE
Mohammadpour et al. [43]	GA + GELS	0.822	1.053
Kant and Mahara [44]	BiUCF	0.737	1.039
Hug [45]	NMF	0.724	0.916
	SlopeOne	0.715	0.907
	K-NN	0.727	0.923
	Centered K-NN	0.738	0.929
	Co-Clustering	0.717	0.915
	Baseline	0.719	0.909
	Random	1.206	1.504
Proposed method	TMPSO-CF	0.704	0.907

To sum up, this study provides a new way to combine the collaborative filtering algorithm and the content based method to alleviate the data sparsity in CARS. The proposed model is tested on MovieLens 1 M dataset, which contains sparse ratings. It can be concluded from different tests conducted in this work that TMPSO-CF performs better with respect to five metrics.

Consequently, it can identify the users’ needs and suggest relevant resources for the customers. This is due to the topic analysis method, which can analyze the item themes those users are interested in, and the weighting function which consider the change of user’s interest over context. These encouraging results clearly confirm that the context is an important criterion in recommendation systems. Moreover, the combination of hidden and explicit features with their degree of importance is a promising way to improve the quality of prediction and recommendation.

6 Conclusion and future work

This study proposes a context-aware recommender system based on Latent Dirichlet Allocation (LDA) and Particles Swarm Optimization (PSO) to make recommendations that consist not only of items that are popular in the community, but also those that are similar in content and in context. The main aim is to alleviate the sparsity problem and improve the quality of recommendations. The LDA algorithm is used to learn latent properties of objects in order to compute the similarity scores between items through the hidden context. Furthermore, a weighting function is introduced to directly integrate explicit context in the recommendation process. It exploits contextual factors and their degree of importance, which are computed based on the PSO algorithm in order to detect user’s interest drifts. The effectiveness of the proposed method was evaluated experimentally using data from the Movielens 1 M dataset and CMU Movie Summary Corpus. The experiments indicated that the developed model alleviated the sparsity problem and achieved significantly better recommendation quality than the baseline methods. It also achieves the best trade-off between the efficiency and performance. In a future work, we plan to incorporate some other useful information obtained from social networking sites. We will also propose a dynamic context weighting mechanism for further performance enhancement.

Footnotes

References

Moradi

and Ahmadian

, A reliability-based recommendation method to improve trust-aware recommender systems, Expert Systems with Applications 42 (2015), 7386–7398 10.1016/j.eswa.2015.05.027.

Shao

, Li

and Bian

, A survey of research hotspots and frontier trends of recommendation systems from the perspective of knowledge graph, Expert Systems with Applications 165 (2021), 113764–10.1016/j.eswa.2020.113764.

Patro

S.G.K.

, Mishra

B.K.

, Panda

S.K.

, Kumar

, Long

H.V.

and Tuan

T.M.

, Knowledge-based preference learning model for recommender system using adaptive neuro-fuzzy inference system, Journal of Intelligent & Fuzzy Systems 39 (2020), 4651–4665 10.3233/JIFS-200595.

Herlocker

, Konstan

J.A.

and Riedl

, An Empirical Analysis of Design Choices in Neighborhood-Based Collaborative Filtering Algorithms, Information Retrieval 5 (2002), 287–310 10.1023/A:1020443909834.

Sánchez

and Bellogín

, Building user profiles based on sequences for content and collaborative filtering, Information Processing & Management 56 (2019), 192–211 10.1016/j.ipm.2018.10.003.

Burke

, Hybrid Recommender Systems: Survey and Experiments, User Model User- Adapted Interaction, 12 (2002), 331–370 10.1023/A:1021240730564.

Zheng

, Context-Aware Mobile Recommendation By A Novel Post-Filtering Approach, The Thirty-First International Flairs Conference (2018).

Lin

and Chen

, Recommendation over time: a probabilistic model of time-aware recommender systems, Science China Information Sciences (2019), 62–10.1007/s11432-018-9915-8.

Jain

and Dixit

V.S.

, Recommendations with context aware framework using particle swarm optimization and unsupervised learning, Journal of Intelligent & Fuzzy Systems 36 (2019), 4479–4490 10.3233/JIFS-179001.

10.

Abdi

, Okeyo

and Mwangi

, Matrix Factorization Techniques for Context-Aware Collaborative Filtering Recommender Systems: A Survey, Computer and Information Science (2018), 11–10.5539/cis.v11n2p1.

11.

Tamine-Lechani

, Boughanem

and Daoud

, Evaluation of contextual information retrieval effectiveness: Overview of issues and research, Knowledge and Information Systems 24 (2010), 1–34 10.1007/s10115-009-0231-1.

12.

Dixit

V.S.

, Jain

Computational Science and Its Applications – (2018), 289–305 Recommendations with Sparsity Based Weighted Context Framework, 10.1007/978-3-319-95171-3_23O. Gervasi, B. Murgante, S. Misra, E. Stankova, C.M. Torre, A.M.A.C. Rocha, et al., editors,ICCSA Cham: Sringer International Publishing;.

13.

Zheng

, A Revisit to The Identification of Contexts in Recommender Systems, 20th International Conference on Intelligent Users Interfaces (2015), 109–115 10.1145/2732158.2732167.

14.

Sinoara

R.A.

, Scheicher

R.B.

and Rezende

S.O.

, Evaluation of latent dirichlet allocation for document organization in different levels of semantic complexity, 2017 IEEE Symposium Series on Computational Intelligence (SSCI) (2017), 1–8 10.1109/SSCI.2017.8280939.

15.

Blei

D.M.

, Ng

A.Y.

and Jordan

M.I.

, Latent dirichlet allocation, J Mach Learn Res 3 (2003), 993–1022.

16.

Gasmi

, Anguel

, Seridi-Bouchelaghem

, Azizi

Cham: Springer International Publishing (2021), 217–232 Context-Aware Based Evolutionary Collaborative Filtering Algorithm, 10.1007/978-3-030-58861-8_16S. Chikhi, A. Amine, A. Chaoui, D.E. Saidouni, M.K. Kholladi, editors. Modelling and Imlementation of Complex Systems,;.

17.

, Tang

, Ma

, Liu

and Daneshmand

, A collaborative filtering recommendation algorithm based on user confidence and time context, Journal of Electrical and Computer Engineering 2019 (2019), 1–12 10.1155/2019/7070487.

18.

Fan

S.-.K.S.

and Jen

C.-H.

, An Enhanced Partial Search to Particle Swarm Optimization for Unconstrained Optimization, Mathematics 7 (2019), 357–10.3390/math7040357.

19.

Tassopoulos

I.X.

and Beligiannis

G.N.

, Solving effectively the school timetabling problem using particle swarm optimization, Expert Systems with Applications 39 (2012), 6029–6040 10.1016/j.eswa.2011.12.013.

20.

Sahu

, Panigrahi

S.K.

and Pattnaik

, Fast Convergence Particle Swarm Optimization for Functions Optimization, Procedia Technology 4 (2012), 319–324 10.1016/j.protcy.2012.05.048.

21.

Adomavicius

, Mobasher

, Ricci

and Tuzhilin

, Context-Aware Recommender Systems, AI Magazine 32 (2011), 67–80.

22.

Zheng

, Interpreting Contextual Effects By Contextual Modeling In Recommender Systems, IDM Workshop @ CIKM’17 (2017).

23.

Ren

, Tomko

, Salim

F.D.

, Chan

, Clarke

C.L.A.

and Sanderson

, A Location-Query-Browse Graph for Contextual Recommendation, IEEE Transactions on Knowledge and Data Engineering 30 (2018), 204–218 10.1109/TKDE.2017.2766059.

24.

Gasmi

, Seridi-Bouchelaghem

, Hocine

and Abdelkarim

, Collaborative filtering recommendation based on dynamic changes of user interest, Intelligent Decision Technologies 9 (2015), 271–281 10.3233/IDT-140221.

25.

Braunhofer

and Ricci

, Selective contextual information acquisition in travel recommender systems, Information Technology & Tourism 17 (2017), 5–29 10.1007/s40558-017-0075-6.

26.

Zhou

, Qin

, Chen

and Zhang

, Real-time context-aware social media recommendation, The VLDB Journal 28 (2019), 197–219 10.1007/s00778-018-0524-7.

27.

Hildebrandt

, Sunder

S.S.

, Mogoreanu

, Joblin

, Mehta

, Thon

The Semantic Web. Cham: Springer International Publishing (2019), 179–193 et al., A Recommender System for Complex Real-World Applications with Nonlinear Dependencies and Knowledge Graph Context, 10.1007/978-3-030-21348-0_12Hitzler, M. Fernández, K. Janowicz, A. Zaveri, A.J.G. Gray, V. Lopez, et al., editors,;.

28.

Fang

and Guo

, A context-aware matrix factorization recommender algorithm, 2013 IEEE 4th International Conference on Software Engineering and Service Science (2013), 914–918 10.1109/ICSESS.2013.6615454.

29.

Al-Ghossein

, Murena

P.-A.

, Abdessalem

, Barré

and Cornuéjols

, Adaptive collaborative topic modeling for online recommendation,;, Proceedings of the 12th ACM Conference on Recommender Systems. British Columbia, Canada: Association for Computing Machinery (2018), 338–346 10.1145/3240323.3240363.

30.

Lin

K.-P.

, Shen

C.-Y.

, Chang

T.-L.

and Chang

T.-M.

, A Consumer Review-Driven Recommender Service for Web E-Commerce, 2017 IEEE 10th Conference on Service-Oriented Computing and Applications (SOCA) (2017), 206–210 10.1109/SOCA.2017.35.

31.

Yan

, Sang

, Xu

and Hossain

M.S.

, A Unified Video Recommendation by Cross-Network User Modeling, ACM Transactions Multimedia Computing, Communications, and Applications 12 (2016), 1–24 10.1145/2957755.

32.

Sumathi

, Sendhilkumar

and Mahalakshmi

G.S.

, Hybrid Recommendation System using Particle Swarm Optimization and User Access Based Ranking,;, Proceedings of the International Conference on Informatics and Analytics. Pondicherry, India: Association for Computing Machinery (2016), 1–9 10.1145/2980258.2980405.

33.

Wang

, Luo

, Sang

, Zeng

and Hirokawa

, Personalized movie recommendation system based on support vector machine and improved particle swarm optimization,E100D:, IEICE Transactions on Information and Systems (2017), 285–293 10.1587/transinf.2016EDP7054.

34.

Peška

, Tashu

T.M.

and Horváth

, Swarm intelligence techniques in recommender systems - A review of recent research, Swarm and Evolutionary Computation 48 (2019), 201–219 10.1016/j.swevo.2019.04.003.

35.

Yang

, Peng

, Chen

Computational Intelligence and Intelligent Systems. Singapore: Springer (2016), 435–444 The BPSO Based Complex Splitting of Context-Aware Recommendation, 10.1007/978-981-10-0356-1_46K. Li, J. Li, Y. Liu, A. Castiglione, editors,;.

36.

Ujjin

and Bentley

P.J.

, Particle swarm optimization recommender system, Proceedings of the 2003 IEEE Swarm Intelligence Symposium SIS’03 (Cat No03EX706) (2003), 124–131 10.1109/SIS.2003.1202257.

37.

Katarya

and Verma

O.P.

, A collaborative recommender system enhanced with particle swarm optimization technique, Multimedia Tools and Applications 75 (2016), 9225–9239 10.1007/s11042-016-3481-4.

38.

Al-bashiri

, Abdulgabber

M.A.

, Romli

and Kahtan

, An improved memory-based collaborative filtering method based on the TOPSIS technique,e, PLOS ONE 13 (2018), 0204434–10.1371/journal.pone.0204434.

39.

Harper

F.M.

and Konstan

J.A.

, The MovieLens Datasets: History and Context,19:1–19:, ACM Transactions on Interactive Intelligent Systems 5 (2015), 19–10.1145/2827872.

40.

Shani

, Gunawardana

Recommender Systems Handbook. Boston, MA: Springer US (2011), 257–297 Evaluating Recommendation Systems, 10.1007/978-0-387-85820-3_8F. Ricci, L. Rokach, B. Shaira, P.B. Kantor, editors,;.

41.

Liu

, Chen

, Xiong

, Ding

C.H.Q.

and Chen

, Enhancing Collaborative Filtering by User Interest Expansion via Personalized Ranking, IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 42 (2012), 218–233 10.1109/TSMCB.2011.2163711.

42.

Adomavicius

, Tuzhilin

Recommender Systems Handbook. Boston, MA: Springer US (2011), 217–253 Context-Aware Recommender Systems, 10.1007/978-0-387-85820-3_7F. Ricci, L. Rokach, B. Shaira, P.B. Kantor, editors,;.

43.

Mohammadpour

, Bidgoli

A.M.

, Enayatifar

and Javadi

H.H.S.

, Efficient clustering in collaborative filtering recommender system: Hybrid method based on genetic algorithm and gravitational emulation local search algorithm, Genomics 111 (2019), 1902–1912 10.1016/j.ygeno.2019.01.001.

44.

Kant

and Mahara

, Merging user and item based collaborative filtering to alleviate data sparsity, International Journal of System Assurance Engineering and Management 9 (2018), 173–179 10.1007/s13198-016-0500-9.

45.

Hug

, Surprise: A Python library for recommender systems, JOSS 5 (2020), 2174–10.21105/joss.02174.

Enhanced context-aware recommendation using topic modeling and particle swarm optimization

Abstract

Keywords

1 Introduction

2 Background and related works

2.1 Context-aware recommender system

2.2 Latent Dirichlet allocation in CARS

2.3 Particle swarm optimization in CARS

3.1 TMPSO-CF algorithm

4.1 Dataset

4.2 Methods compared

4.3 Parameter setting

4.4 Evaluation Metrics

5 Performance of TMPSO-CF at optimal neighborhood

5.1 Optimization of neighborhood size

Table 3 Constraints for optimization of neighborhood size (k) Condition Goal Lower limit Upper limit k In range 10 100 MAE Minimize 0.70432 0.75795 RMSE Minimize 0.90713 0.97482 Precision Maximize 0.63425 0.68635 Recall Maximize 0.27048 0.35521 F- Measure Maximize 0.38794 0.45512

Footnotes

References

Table 3
Constraints for optimization of neighborhood size (k)

Condition Goal Lower limit Upper limit

k In range 10 100

MAE Minimize 0.70432 0.75795

RMSE Minimize 0.90713 0.97482

Precision Maximize 0.63425 0.68635

Recall Maximize 0.27048 0.35521

F- Measure Maximize 0.38794 0.45512