Deep embedded clustering with matrix factorization based user rating prediction for collaborative recommendation

Abstract

Conventional recommendation techniques utilize various methods to compute the similarity among products and customers in order to identify the customer preferences. However, such conventional similarity computation techniques may produce incomplete information influenced by similarity measures in customers’ preferences, which leads to poor accuracy on recommendation. Hence, this paper introduced the novel and effective recommendation technique, namely Deep Embedded Clustering with matrix factorization (DEC with matrix factorization) for the collaborative recommendation. This approach creates the agglomerative matrix for the recommendation using the review data. The customer series matrix, customer series binary matrix, product series matrix, and product series binary matrix make up the agglomerative matrix. The product grouping is carried out to group the similar products using DEC for retrieving the optimal product. Moreover, the bi-level matching generates the best group customer sequence in which the relevant customers are retrieved using tversky index and angular distance. Also, the final product suggestion is made using matrix factorization, with the goal of recommending to clients the product with the highest rating. Also, according to the experimental results, the developed DEC with the matrix factorization approach produced better results with respect to f-measure values of 0.902, precision values of 0.896, and recall values of 0.908, respectively.

Keywords

Deep embedded clustering agglomerative matrix bilevel matching matrix factorization tversky index

1. Introduction

RS has been introduced along with the web for recommending the products to the users. Initially, the RS relies on content-based technique, demographic-based technique and collaborative filtering-based technique. Nowadays, the RS was constructed based on social information. Results are decided using a collaborative recommendation system based on product similarity. Moreover, the effectiveness of collaborative recommendation can be further improved on the basis of local as well as personal information from internet of things. In addition, the RS utilizes various resource of information for making the users with predictions as well as recommendation of products. Moreover, the RS aims to balance the factors, such as stability, novelty, disparity as well as accuracy [10]. Most of the familiar commercial websites use RS for assisting their customers and their respective products. The RS considers the information regarding the recommended products as well as customers for finding the most appropriate products from the available products [11]. Generally, the RS states that the people suggest the recommendation as input, and then the system aggregates and provides the product regarding the requested recommendation to the recipient. RS suggests the product to the customer for purchasing based on the user preference. Moreover, various techniques have been introduced for RS, like knowledge-based, content-based, collaborative-based and so on. These conventional techniques are not effective all time due to the poor design. In such situations, the hybrid recommender provides a better solution for this issue [12].

One of the well-known methods used in RS is CF. In the CF family of algorithms, there are numerous ways to locate comparable users or things, as well as numerous ways to determine ratings based on the ratings of comparable users. It is a system of recommendations that makes predictions about a user’s future behaviour. However, CF technique has the CCS issue and ICS issue. Here, the CCS issue is produced due to the inexistence of rating records, whereas the ICS issue occurs due to the existence of small number of record rating of new product. The RS utilized the huge amount of information based on the user searching activities in the past and it identifies the item based on the user requirement. Moreover, the collaborative RS did not utilize the content information of products for making recommendation. In addition, the RS relies on the relationship among users as well as products, which are generally encoded in a rating feedback matrix with each element indicating a specific user rating on a particular item [13]. CF is the more appropriate and broadly employed recommender systems in which the CF model aims to systematize the process of ‘word-of-mouth’ in which people suggest products to one another. Moreover, the CF scheme utilized the past information of each user for recommending the products to the neighbourhood people who require similar products. In addition, the CF approach estimates the requirement of new items by investigating the neighbourhood people [14].

CF approaches are often divided into two kinds, including memory-based technique and model-based methodology. Based on how well a product is rated, memory-based techniques are used to determine user preferences. Moreover, the memory-based method uses locality-sensitive hashing, which periodically employs the nearest neighbouring approach. On the other hand, model-based approaches are employed to extract the data, and then the machine learning techniques are utilized to identify the pattern using training data. Some of the familiar model-based techniques involve Bayesian methods [16], clustering methods [17], and latent semantic models [18], like probabilistic latent semantic assessment [15] and Singular Value Decomposition. In [19], RBM-based CF algorithm was developed for recommendation. In [20], the hybrid recommendation algorithm was devised by extending the RBM-based techniques with better performance. In [21], a Deep Convolutional Neural Network was employed to identify the latent factors from music audio. In [22], a revised autoencoder scheme was introduced for CF in order to make recommendation. Moreover, the effectiveness of recommendation can be enhanced by including collaborative filtering based on the decomposition of MF strategy. MF strategy utilizes the user’s and product’s latent features, such that the inner multiplication of user and item latent features are equivalent to the user’s rating on that product [23].

This research aims to design and develop the DEC with matrix factorization for performing the collaborative recommendation. The review data is considered as an input for recommendation process, which contains the product ID and the customer ID. Initially, the agglomerative matrix is generated based on the review data. After the generation of agglomerative matrix, product grouping is done using DEC, which provided the grouped product by combining the similar products. The series of best customers is then generated by doing a bilevel matching between the query and the product group using the Tversky index and angular distance. After that, the relevant customers are retrieved from the sequence of best customers, such that the customer preferred product is retrieved. At last, the final recommendation is done using matrix factorization based on the product with maximum rating, and then it is recommended to the customers.

The major contribution of this research are given by,

•
Proposed DEC $+$ matrix factorization for collaborative recommendation: In this research, DEC $+$ matrix factorization method is developed for collaborative recommendation. Here, the DEC method is employed to generate the product grouping and the matrix recommendation is used to recommend the product with maximum rating to the customers.

The other sections of the research paper are as follows: Section 2 describes the various existing techniques based on recommendation techniques, Section 3 illustrates the developed DEC with matrix factorization model, Section 4 illustrates the results of the experiments, and Section 5 illustrates the conclusion.
2. Literature review

There are eight traditional user rating prediction approaches for collaborative recommendation, each with advantages and disadvantages stated. Zhang et al. [1] modelled the DeRec approach for performing the recommendation based on weight loss function. Here, the recommendation was done mainly based on the list of user-item rating. Moreover, this method was attained the optimal accuracy along with minimal computational complexity. However, the memory consumption of this method was high. In order to resolve this, Chen et al. [2] modelled the DDCF method for performing collaborative recommendation. Here, the recommendation system was operated on the basis of user preferences with respect to the dynamic time decay. Moreover, the developed method vigorously utilized the decay functions depending on the behaviour of users. Yet this approach overlooked different datasets. Guo et al. extended .’s degree classification criterion technique for recommendation via collaborative filtering was created in order to make use of different datasets. Here, the sigmoid function was used to determine how closely connected items and pairs of items are to one another. Moreover, this method was employed to resolve the trade-off among accuracy as well as effectiveness of items based on collaborative filtering. However, the developed method did not utilize decision support systems for enhancing the experience of user and productivity of website companies. For increasing the experience and productivity of such companies, Jain et al. [4] modelled the EMUCF algorithm for performing recommendation based on non-linear similarity. Here, the Bhat-sim was used to determine the similarity and the similarities between the users were calculated using the Bhattacharyya coefficient. While the developed method achieved great accuracy, it was challenging to handle the numerous components that grew in size with time. For enhancing the performance with various items, Alhijawi and Kilani [5] designed the Genetic-based recommender system for performing the collaborative recommendation using genetic algorithm. The RS was carried out using two functions in which the first function was employed to utilize the semantic information for estimating the similarity among items, whereas the second function was employed to estimate the similarity among users. Moreover, the developed method has the ability to make more accurate predictions. However, the developed method did not able to handle a small as well as medium sized datasets. In order to resolve the issues of using datasets, Panda et al. [6] devised the Normalization-based collaborative filtering recommender system for performing recommendation. This method had two steps for performing recommendation such that the first step was able to perform the prediction of user ratings, and then the second step was carried out to perform the recommendation. Using a maximum confidence interval, the suggested approach achieved greater accuracy in this case. Yet, it was unable to add new users or goods to the rating matrix. For attaining the effective user rating and product matrix, Chen et al. [7] modelled the Collaborative filtering recommendation algorithm for performing collaborative recommendation with respect to the user correlation as well as evolutionary clustering. Although, the developed method handled the huge volume of data, this method did not include some other effective information of network for further enhancing the performance. For improving the system performance, Li and Han [8] modelled the hybrid collaborative filtering technique for performing the collaborative recommendation. Here, the hybrid collaborating recommendation scheme was attained by integrating collaborative filtering as well as content-based filtering approach. Also, the system that was devised was able to handle vast amounts of information with higher stability. Nevertheless, in order to improve performance, the devised approach was unable to handle the factorization machine.

3. Proposed DEC with matrix factorization for collaborative recommendation

The main goal of this study is to create a new DEC model with matrix factorization for collaborative recommendation utilising a dataset of Netflix movie recommendations. In order to create the matrix based on customer preferences, the review data is initially given to the agglomerative matrix development phase. The produced matrix is then put through the process of product grouping using DEC [25], where comparable groupings are clustered into one group and dissimilar groups are clustered into one group. Using group-product bilevel matching between the query and item group utilising the Tversky index and angular distance, the best product group from the entire product group is then selected. Following the selection of the best product group, the user query is matched with the best product group according to the Tversky index and angular distance in order to retrieve the customer’s favourite product. Additionally, matrix factorization is used to carry out the final recommendation in order to find the product with the highest rating. Figure 1 shows the schematic illustration of developed DEC with matrix factorization for collaborative recommendation.

Figure 1.

Block diagram of proposed method for predicting user ratings based on collaborative recommendation.

3.1 Data acquisition

The review data from Netflix dataset contains set of customer ID and product ID. Let us consider the customer $L$ with customer ID $U_{a}$ and the product $V$ with product $P_{b}$ , and it is expressed as,

$\displaystyle\text{Customer }L=U_{a}$ (1) $\displaystyle\text{Product }V=P_{b}$ (2)

where, $L$ be the customer, $U_{a}$ be the customer ID, $V$ be the product and $P_{b}$ be the product ID. The dataset has $n$ number of customers and $m$ number of products.

3.2 Compute data agglomerative matrix

The gathered data from the dataset is transformed into the matrix format for making the recommendation scheme as effective. The customer series matrix, the customer series binary matrix, the product series matrix, and the product series binary matrix are all computed during this stage.

3.2.1 Customer series matrix

The review data acquired from the dataset is utilized to generate the customer series matrix $D_{a}$ , which contains the customer ID, product ID, and the product who visited by the customer. The list of products visited by the customer is illustrated as,

$\displaystyle D_{a}=\{P_{1}^{a},P_{2}^{a},\ldots,P_{z}^{a},\ldots,P_{n}^{a}\}$ (3)

where, $P_{n}^{a}$ depicts the $n^{\text{th}}$ product visited by the $a^{\text{th}}$ customer, $P_{z}^{a}$ depicts the $z^{\text{th}}$ product visited by the $a^{\text{th}}$ customer, and $D_{a}$ depicts the product visited by the $a^{\text{th}}$ customer.

3.2.2 Customer series binary matrix

After constructing the customer series matrix $D_{a}$ , then the customer series binary matrix $BD_{a}^{z}$ is generated with respect to the preferred product, and it is signified as either one or zero. Here, the customer preferred product is represented in binary sequence. For instance, if the customer prefers a particular product, then the corresponding product is represented as 1, otherwise it is denoted as 0. Thus, the customer series binary matrix is portrayed as,

$\displaystyle BD_{a}^{z}=\left\{{\begin{array}[]{ll}1&;P_{z}^{a}\in P_{b}\\ 0&;\text{Otherwise}\\ \end{array}}\right.$ (4)

Here, ‘1’ represents the product searched by the customer and ‘0’ indicates the product not searched by the customer.

3.2.3 Product series matrix

The review data $P_{b}$ taken from the Netflix dataset is employed to construct the product series matrix $N_{b}$ . The product series matrix contains the product ID, customer ID, and the customer searched product. Each product has the particular ID, which is represented as product ID and the customer who searched for the product is termed as customer sequence. The list of products searched by the customer is illustrated as,

$\displaystyle N_{b}=\{{G_{1}^{b},G_{2}^{b},\ldots,G_{r}^{b},\ldots,G_{w}^{b}}\}$ (5)

where, $G_{r}^{b}$ depicts the $r^{\text{th}}$ customer visited $b^{\text{th}}$ product, $G_{w}^{b}$ depicts the $w^{\text{th}}$ customer visited $b^{\text{th}}$ product and $N_{b}$ indicates the customer visited $b^{\text{th}}$ product.

3.2.4 Product series binary matrix

Once the product series matrix is formed, then the product series binary matrix $B^{N_{b}^{r}}$ is generated in binary format, and it is signified as either one or zero. For every product ID, the corresponding binary form of visitor ID is represented in the customer binary sequence. Here, the customer visits a particular product and it is represented as ‘1’, otherwise it is represented as ‘0’. Here, the product series matrix is illustrated as,

$\displaystyle B^{N_{b}^{r}}=\left\{{\begin{array}[]{ll}1&;G_{r}^{b}\in U_{a}\\ 0&;\text{Otherwise}\\ \end{array}}\right.$ (6)

Here, ‘1’ represents the customer searched product, and ‘0’ indicates the customer not searched product.

3.3 Product grouping using DEC algorithm

Deep neural networks are used by DEC to simultaneously learn feature representations and cluster assignments. In order to iteratively optimise a clustering objective, it learns a mapping from the data space to a lower-dimensional feature space. In this work, the product grouping is carried out by the DEC algorithm [25] for finding the best product groups. Here, the input of DEC is $B^{N_{b}^{r}}$ in order to cluster the products. It consists of two various phases, like parameter initialization as well as clustering optimization. Here, the distribution of auxiliary target is calculated and the KL divergence is reduced. The parameter optimization or clustering optimization is demonstrated by supposing a primary estimate of $\varpi$ and $\{{p_{i}}\}_{i}^{q}=1$ .

Clustering with KL convergence

By selecting an initial approximation of cluster centroids $\{{p_{i}}\}_{i}^{q}=1$ , and non-linear mapping $d_{\varpi}$ , an unsupervised algorithm with two phases is introduced for enhancing the clustering process. In the initial phase, soft assignment is computed among the cluster centroids as well as embedded points. In the next phase, deep mapping $d_{\varpi}$ is renewed, and the cluster centroids are developed depending on the current high confidence assignments based on the distribution of auxiliary target. This procedure is continuously done until the convergence condition is fulfilled.

Soft assignment: Here, the student’s t-distribution is employed as a kernel for calculating the similarity between the centroid $p_{i}$ as well as embedded point $z_{j}$ .

$\displaystyle V_{ji}=\frac{\left({{1+\left\|{z_{j}-p_{i}}\right\|^{2}}\mathord% {\left/{\vphantom{{1+\left\|{z_{j}-p_{i}}\right\|^{2}}\rho}}\right.\kern-1.2pt% }\rho}\right)^{-\frac{\rho+1}{2}}}{\sum{\left({{1+\left\|{z_{j}-p_{ij}}\right% \|^{2}}\mathord{\left/{\vphantom{{1+\left\|{z_{j}-p_{ij}}\right\|^{2}}\rho}}% \right.\kern-1.2pt}\rho}\right)^{{}^{-\frac{\rho+1}{2}}}}}$ (7)

where, $p_{i}=d_{\varpi}({y_{j}})\in Z$ corresponds to $p_{i}\in Y$ after the process of embedding, the freedom degree is characterized as $\rho$ , $V_{ji}$ represents the probability of sample $j$ to cluster $i$ .

KL divergence optimization: By accurately realising their assignments with the greatest degree of confidence based on the auxiliary goal function, KL divergence optimization is intended to iteratively optimise the clusters. It calculates the convergence loss $J_{j}$ between the auxiliary distribution as well as soft assignment $K_{j}$ .

$\displaystyle L=KL({S||T})=\sum\limits_{j}{\sum\limits_{i}{J_{ji}}}\log\frac{J% _{ji}}{K_{ji}}$ (8)

Additionally, the computation $J_{j}$ is carried out by initially raising $K_{j}$ to the second power and afterward normalizing the result using frequency per cluster.

$\displaystyle J_{ji}=\frac{\raise 3.01pt\hbox{${K_{ji}^{2}}$}\!\mathord{\left/% {\vphantom{{K_{ji}^{2}}{d_{i}}}}\right.\kern-1.2pt}\!\lower 3.01pt\hbox{${d_{i% }}$}}{\sum\limits_{i}{\raise 3.01pt\hbox{${K_{ji}^{2}}$}\!\mathord{\left/{% \vphantom{{K_{ji}^{2}}{d_{i}^{\prime}}}}\right.\kern-1.2pt}\!\lower 3.01pt% \hbox{${d_{i}^{\prime}}$}}}$ (9)

where, $d_{i}=\sum\limits_{i}{K_{ji}}$ characterizes the soft cluster frequency. Hence, the DEC algorithm effectively progresses low confidence prediction outcomes.

To gather together related products into a single group, products are grouped. Let us consider the group attained by the DFC is represented as,

$\displaystyle Q_{q}=\{Q_{1},Q_{2},\ldots,Q_{r}\}$ (10)

where, $r$ depicts the total group-count and $Q$ represents the grouped product. Thus, the output attained by the product grouping using DEC is depicted as $Q$ .

3.4 Bilevel product matching

The group-product bilevel matching is carried out using Tversky index and angular distance in order to select the optimal product from the grouped product $Q$ . The operation of matching is done between the product group and the query. Also, the customer query is transformed into a binary query before matching is done in order to get the best group. The steps for group-product bilevel matching are described as below.

3.4.1 Query

The matrix form of number of queries is illustrated as,

$\displaystyle H=\{{h_{1},h_{2},\ldots,h_{f},\ldots,h_{p}}\}$ (11)

where, $h_{p}$ depicts the $p^{\text{th}}$ number of products in query, $h_{f}$ be the $f^{\text{th}}$ product in query, $f$ and $p$ represents the number of queries.

3.4.2 Binary query

The mathematical sequence of number of queries are converted into binary form, and the binary series of query is termed as,

$\displaystyle BH=\left\{{\begin{array}[]{ll}1&;h_{f}\in P_{b}\\ 0&;\text{Otherwise}\\ \end{array}}\right.$ (12)

where, $h_{f}$ denotes the product count in query, $f$ and $B Q$ represents the binary query sequence. Moreover, the group-product bilevel matching is done between the binary query and grouped product based on using Tversky index and angular distance.

Tversky index: The Tversky index, which is represented as, is a statistic used to determine the similarity between a binary query and a grouped product.

$\displaystyle TI({Q,BQ})=\frac{|{Q\cap BQ}|}{|{Q\cap BQ}|+\mu|{Q\backslash BQ}% |+\lambda|{BQ\backslash Q}|}$ (13)

where, $|{Q\backslash BQ}|$ depicts the relative complement of $B Q$ in $Q$ .

Angular distance: Angular distance is a distant metric, which is used to find the angular similarity among $Q$ and $B Q$ , and it is mathematically expressed as,

$\displaystyle\delta({Q,BQ})=\frac{Q.BQ}{\|Q\|\|{BQ}\|}$ (14)

Thus, the group product bilevel matching is obtained by combining the angular distance and Tversky index, and it is mathematically expressed as,

$\displaystyle\text{Bilevel matching }\mu({Q,BQ})=\alpha TI({Q,BQ})+({1-\alpha}% )AD({Q,BQ})$ (15)

where, $T I$ denotes the using Tversky index, $A D$ be the angular distance, $\alpha$ be the constant, $G$ denotes the group item, and $B Q$ indicates the binary sequence theory. Moreover, the output obtained from the group product bilevel matching is the retrieved optimal group, which is then applied to the relevant customer retrieval phase.

3.5 Relevant customer retrieval

To choose the best group user series in binary format after obtaining the ideal group, the pertinent customer retrieval phase is executed. Binary best group is used to identify the customer ID, and the best product is chosen depending on the consumers. The consumer who is in the best groups searches the product lists here. Thus, the relevant customer retrieval series is indicated as $R^{C}$ , and is expressed as,

$\displaystyle R^{c}=\{{c_{1}^{x},c_{2}^{x},\ldots,c_{u}^{x},\ldots,c_{w}^{x}}\}$ (16)

where, $R^{C}$ indicates the retrieved best group user series.

Binary best group

After that, the binary best group is obtained by transforming the series of best user group into the binary form, and is portrayed as,

$\displaystyle BR_{e}^{c}=\left\{{\begin{array}[]{ll}1&;c_{u}^{x}\in P_{b}\\ 0&;\text{Otherwise}\\ \end{array}}\right.$ (17)

where, $BR_{e}^{c}$ indicates the binary series of best group.

3.6 Matching of query and best group customer sequence

After the retrieval of relevant customer, the matching among relevant customer $BR_{e}^{c}$ and binary query $B Q$ is carried out using Tversky index and angular distance. The Tversky index is expressed as,

$\displaystyle TI({BQ,BR_{e}^{c}})=\frac{|{BQ\cap BR_{e}^{c}}|}{B|{Q\cap BR_{e}% ^{c}}|+\mu|{BQ\backslash BR_{e}^{c}}|+\lambda|{BR_{e}^{c}\backslash BQ}|}$ (18)

Moreover, the angular distance among relevant customer $BR_{e}^{c}$ and binary query $B Q$ is expressed as,

$\displaystyle\delta({BQ,BR_{e}^{c}})=\frac{BQ,BR_{e}^{c}}{\|{BQ}\|\|{BR_{e}^{c% }}\|}$ (19)

Thus, the group product matching is achieved by adding the angular distance and Tversky index, and it is mathematically expressed as,

$\displaystyle\mu 2({BQ,BR_{e}^{c}})=\beta TI({BQ,BR_{e}^{c}})+({1-\beta})AD({% BQ,BR_{e}^{c}})$ (20)

where, $TI({BQ,BR_{e}^{c}})$ denotes the Tversky index among binary query and relevant customer, $AD\linebreak({BQ,BR_{e}^{c}})$ be the angular distance among $B Q$ and $BR_{e}^{c}$ , $\beta$ be the constant, $BR_{e}^{c}$ be the binary best group and $B Q$ indicates the binary sequence query. The output produced from the matching customer preferred product, and it is represented as,

$\displaystyle B_{m}=\{{B_{1},B_{2},\ldots,B_{s}}\}$ (21)

where, $s$ indicates the number of products selected by the customer.

3.7 Recommendation using matrix factorization

The matrix factorization [24] is an effective approach, which has implicit feedback. Here, the user information is not directly provided by the user where it has been collected by analysing the past behaviour of user about the product through the user-item interaction matrix. This technique is more effective to predict the user rating of any product. The user rating matrix is calculated by multiplying the two matrices, such as $U_{a}$ and $P_{b}$ , and it is expressed as,

$\displaystyle M_{f}=U_{a}\times P_{b}$ (22)

where, $U_{a}$ signifies set of customer and $P_{b}$ signifies set of products. Here, $U_{a}\in\Re^{W\times\kappa}$ be user latent matrix, and $P_{b}\in\Re^{v\times\kappa}$ indicates item latent matrix, and $\kappa$ signifies total number of latent factors, and $M_{f}\in\Re^{U_{a}\times P_{b}}$ denotes user-item rating matrix. Here, each product $P_{b}$ is linked with vector $p_{b}^{v}\in\Re^{v\times\kappa}$ and each customer $U_{a}$ is linked with vector $u_{a}^{w}\in\Re^{w\times\kappa}$ , then the interaction among product and customer is determined by,

$\displaystyle\Pr ed_{\text{int}\textit{eraction}}=({p_{b}^{v}})^{T}u_{a}^{w}$ (23)

From the user preferred items, the items having highest rating will be recommended, which can be denoted as $R_{p}$ .

Algorithm 1: Pseudocode of course review framework
Inputs: $U_{a}\rightarrow$ Customer ID, $P_{b}\rightarrow$ Product ID, $R\rightarrow$ Rating, $H\rightarrow$ Query, Cluster size $=$ 3
Parameter: $D_{a}\rightarrow$ Customer preference matrix, $Q_{q}\rightarrow$ optimal clustered product group, $R^{c}\rightarrow$ relevant retrieved customer, $BD_{a}^{z}\rightarrow$ customer preference binary matrix
Output: Recommended product
1. Begin
2. Read input $(U_{a},P_{b},R)$ ;
3. $BD_{a}^{z},B^{N_{b}^{r}}=D_{a}({U_{a},P_{b}})$
4. $Q_{q}=$ DEC $({B^{N_{b}^{r}},\textit{clustersize}=3})$
5. Find $Q_{q}$
6. $Q_{q}=\textit{product Matching phase}(BH,Q_{q})$
7. Calculate $R^{c}=\text{Relevant customer phase}(n,BD_{a}^{z})$ ;
8. ${B}_{m}=\text{Matched customer phase}(H,R^{c})$
9. $R_{p}=\textit{Recommend Phase}(\Pr ed_{\text{int}\textit{eraction}},BD_{a}^{z}% ,R)$
10. //Product preference matrix phase
11. $BD_{a}^{z}=({U_{a},P_{b}})$
12. If customer visit the product
13. Print 1
14. else
15. Print 0
16. $B^{UL_{j}}=({D_{s},D_{c}})$
17. If ( $b$ product is visited by the customer)
18. Print 1
19. else
20. Print 0
21. $H$ generation based on $B^{N_{b}^{r}}$
22. //Course matching phase
23. $\textit{grp}=[]$
24. for $j=$ 1 to $Q$
25. Sum val $=$ 0
26. For $j=$ 1 to $k$ (products in that group)
27. $\text{Sum val}+=\mu({Q,BQ})=\alpha TI({Q,BQ})+({1-\alpha})AD({Q,BQ})$
28. End for
29. grp.append (Sum val)
30. End for
31. Q $=$ max (grp)
32. //Relevant scholar phase
33. $R^{c}=[]$
34. for $j=$ 1 to $s$ (customer in that group)
35. C $=$ got customers who searched the products
36. $R^{c}.\textit{append}(BD_{a}^{z}(C))$
37. End for
38. Return $R^{c}$
39. //Matched scholar phase
40. $B_{m}=[]$
41. for $j=$ 1 to 10 ( $R^{c}$ )
42. $B_{m}.\textit{append}(\mu 2({BQ,BR_{e}^{c}})=\beta TI({BQ,BR_{e}^{c}})+({1-% \beta})AD({BQ,BR_{e}^{c}}))$
43. End for
44. Sort by min ( $B_{m}$ )
45. Return $B_{m}$

3.8 Final recommendation

Based on the highest possible customer rating, the user-preferred product is chosen in the final recommendation process. In this instance, the product that received the highest rating from the consumer is recommended in the final recommendation stage. Algorithm 1 depicts the pseudo-code for the aforementioned procedure.

4. Results and discussion

The findings and analysis of the developed DEC with matrix factorization technique based on evaluation metrics are described in this part.

4.1 Experimental setup

Using a Python tool, Windows 10 OS, and an Intel i3 core CPU, the experimentation of the created DEC matrix factorization approach is performed.

4.2 Dataset description

The dataset employed for the developed technique is Netflix-movie recommendation Dataset [9]. The reviews that were used in this case were obtained directly from Netflix. The collection also includes 4 text data files, each of which has more than 20M rows, or more than 4K movies and 400K users. Consequently, the Netflix-movie recommendation dataset currently has 17K movies and 500K $+$ users. Every data file also includes the movie ID, customer ID, rating, and rating date.

4.3 Performance metrics

Performance indicators like F-measure, precision, and recall are used to analyse the effectiveness of the developed DEC with the matrix factorization approach.

(i) F-measure

It is a measurement that determines the recall and precision weighted harmonic means.

$\displaystyle l=\frac{2\ast go}{g+o}\ast 100$ (24)

where, $g$ shows the precision, and $o$ shows the recall.

(ii) Precision

It is a measure that is calculated by adding the true positive ratio and the false positive ratio together.

$\displaystyle\Pr\textit{ecision }g=\frac{y}{y+t}$ (25)

where, $y$ specifies the TPR, and $t$ designates the FPR.

(iii) Recall

Recall value refers to the proportion of TPR, and the summation of true positive and FNR.

$\displaystyle o=\frac{o}{o+v}$ (26)

where, $o$ shows the TPR, and $v$ specifies the FNR.

4.4 Performance assessment

By altering the amount of queries with different iterations depending on evaluation metrics and using the Netflix-movie recommendation dataset, the performance of the constructed DEC $+$ matrix factorization is evaluated.

4.4.1 Performance assessment based on queries

Figure 2 displays the performance evaluation of the created approach using various queries and iterations. The performance evaluation of the developed technique with respect to the f-measure value is shown in Fig. 2a. When the number of the query is 2, the created DEC $+$ matrix factorization method yields the following f-measure values: 0.801, 0.809, 0.819, 0.836, and 0.856 for iterations of 20, 40, 60, 80, and 100. The performance evaluation of the devised DEC $+$ matrix factorization method based on precision value is shown in Fig. 2b. When there are 4 queries, the created DEC $+$ matrix factorization method yielded the following precision values: 0.835 for iterations of 20, 40, 60, 80, and 100; 0.841 for iterations of 40, 80, and 100; 0.854 for iterations of 60; 0.865 for iterations of 80; and 0.896 for iterations of 100. The performance evaluation of the devised technique with respect to recall value is shown in Fig. 2c. The created DEC $+$ matrix factorization method produced recall values of 0.825, 0.854, 0.865, 0.874, and 0.885 for a number of queries of 3, respectively, for iterations of 20, 40, 60, 80, and 100.

Figure 2.

Performance assessment of developed technique by adjusting the queries.

4.5 Comparative assessment

The comparative assessment is done by the inclusion of various existing techniques along with the proposed DEC $+$ matrix factorization technique by changing the cluster size in terms of evaluation metrics.

4.6 Comparative techniques

The effectiveness of developed technique is investigated by comparing the developed DEC $+$ matrix factorization method with the existing techniques, such as data-driven RS [1], DDCF [2], EMUCF [4] and Genetic-based RS [5].

Figure 3.

Comparative assessment of developed technique by adjusting the query using cluster size 3.

4.6.1 Comparative assessment using cluster size 3

The comparative evaluation of the developed technique with regard to performance criteria utilising cluster size 3 is explained in Fig. 3. The comparative evaluation of created technique based on f-measure value is shown in Fig. 3a. When the number of queries is 1, the newly created method yielded an f-measure value of 0.768, while the corresponding values for the data-driven RS, DDCF, EMUCF, and Genetic-based RS were 0.613, 0.639, 0.686, and 0.739, respectively, for the current methodologies. The comparative evaluation of created technique with respect to precision value is shown in Fig. 3b. The created technique in this case acquired a precision value of 0.814 for the query of 4, while current techniques like data-driven RS, DDCF, EMUCF, and genetic-based RS attained f-measure values of 0.685, 0.704, 0.736, and 0.785 for the same query. The comparison of created approach evaluations based on recall value is shown in Fig. 3c. When there are three queries, the created and existing approaches each achieved recall values of 0.678, 0.704, 0.732, 0.798, and 0.814, respectively.

4.6.2 Comparative assessment using cluster size 4

Figure 4 explains the comparison of developed techniques employing cluster size 4 with regard to performance parameters. Figure 4a shows a comparison of developed techniques based on F-measure value. In this case, the created technique yielded an f-measure value of 0.875 for the number of queries of 3, while the existing techniques, such as data-driven RS, DDCF, EMUCF, and genetic-based RS, yielded f-measure values of 0.679, 0.716, 0.754, and 0.791 for the same number of queries. The comparison of produced approach evaluations based on accuracy value is shown in Fig. 4b. The created method acquired the recall value of 0.847 for the number of question is 2, while the current strategies achieved the recall values of 0.648, 0.665, 0.705, and 0.754. The comparative evaluation of proposed technique based on recall value is demonstrated in Fig. 4c. When there are 4 queries, the created method has a recall value of 0.908, compared to the recall values of 0.725 for the data-driven RS, 0.785 for the DDCF, 0.801 for the EMUCF, and 0.836 for the genetic-based RS attained by the existing methodologies.

Figure 4.

Comparative assessment of developed technique by adjusting the query using cluster size 4.

4.7 Comparative discussion

The developed DEC $+$ matrix factorization technique for collaborative recommendation is compared in Table 1 for discussion. Here, the cluster sizes 3 and 4 are changed to conduct the testing. During evaluation, it was determined that the devised DEC $+$ matrix factorization algorithm performed better utilising cluster size 4. The devised DEC $+$ matrix factorization algorithm achieved maximum F-measure, precision, and recall values of 0.902, 0.896, and 0.908, respectively, for the cluster size 4. The F-measure values for the existing approaches, such as data-driven RS, DDCF, EMUCF, and Genetic-based RS, were 0.710, 0.751, 0.770, and 0.820, respectively. The precision values were 0.695, 0.719, 0.741, and 0.805, while the recall values were 0.725, 0.785, 0.801, and 0.836. It is obvious from the table that the created technique’s improved performance was related to the efficiency of clustering, specifically DEC and the matrix factorization approach. The created method achieved the least FPR and FNR values; as a result, the precision and recall value increased, which automatically enhanced the f-measure value.

Table 1
Comparative discussion

Variations	Metrics	Data-driven RS	DDCF	EMUCF	Genetic-based RS	Proposed DEC $+$ matrix factorization
Cluster size 3	F-measure	0.685	0.704	0.736	0.785	0.814
	Precision	0.685	0.704	0.736	0.785	0.814
	Recall	0.698	0.725	0.754	0.814	0.835
Cluster size 4	F-measure	0.710	0.751	0.770	0.820	0.902
	Precision	0.695	0.719	0.741	0.805	0.896
	Recall	0.725	0.785	0.801	0.836	0.908

5. Conclusion

This paper presents the developed DEC $+$ matrix factorization technique for collaborative recommendation. The DEC clustering is employed to perform the product grouping based on RS. The developed model involves two major processes, such as user rating prediction and product retrieval. The recommendation of developed method is done based on the agglomerative matrix, which comprises of customer series binary matrix and product rating binary matrix. Following the acquisition of the product rating binary matrix from the product series matrix, the customer series binary matrix is obtained from the customer series matrix. By conducting the bilevel matching between the user query and the item group, followed by the query and best group customer sequence, the best recommendation based on the user query is obtained. Moreover, the final recommendation is carried out using matrix factorization based on the retrieved product, such that the product with maximum rating is achieved. Also, based on the f-measure value of 0.902, accuracy value of 0.896, and recall value of 0.908, respectively, the created DEC $+$ matrix factorization technique for collaborative recommendation achieved the higher performance. Future collaborative recommendation performance can be enhanced by using additional efficient clustering methods and review datasets.

Footnotes

Abbreviations

Author’s Bios

Jagannath E. Nalavade has completed Ph.D. in Computer Engineering from Vel Tech Rangrajan Dr. Sagunthala R&D Institute of Science and Technology, Chennai. Currently, he is working as a Associate Professor at MIT Art Design and Technology University, School of Computing, Pune, Maharashtra (IN). His research areas are Data Mining, Machine Learning, Artificial Intelligence. He has 16 years of teaching experience. He worked as various capacities as Head of Department of Information Technology, National Level Hackathon coordinator and NISP coordinator. He has published more than 28 research papers in peer reviewed International journals like Springer, Elsevier Procedia. He is serving as reviewer for Springer and Elsevier Journals (Information Sciences, Journal of Intelligent Systems, Computers & Electrical Engineering, International Journal of Bioinformatics Research and Applications).

Chandra Sekhar Kolli is pursuing (Submitted Thesis) PhD in Computer Science from GITAM (Deemed to be University), Visakhapatnam, India and working as a Senior Assistant Professor in Aditya College of Engineering and Technology, Surampalem, Andhra Pradesh, India. His research interests include Data Science and Data Analytics, Optimization and Classification Algorithms, Security & Privacy, Cyber Security.

Sanjay Nakharu Prasad Kumar: Studying for a Doctor of Engineering at George Washington University with a concentration in Engineering Management, and I have nine years of experience as a professional data scientist in deep learning, machine learning, and analytics. The expertise includes data mining of large structured and unstructured data sets, performing data acquisition, data pre-processing, exploratory data analysis (EDA), statistical analysis, data validation, predictive modelling, and Visualization.

References

Zhang

Yang

and Yoshida

, DeRec: A data-driven approach to accurate recommendation with deep learning and weighted loss function, Electronic Commerce Research and Applications 31 (2018), 12–23.

Chen

Y.C.

Hui

and Thaipisutikul

, A collaborative filtering recommendation system with dynamic time decay, The Journal of Supercomputing 77(1) (2021), 244–262.

Guo

Deng

Ran

Wang

and Jin

, An efficient and accurate recommendation strategy using degree classification criteria for item-based collaborative filtering, Expert Systems with Applications 164 (2021), 113756.

Jain

Nagar

Singh

P.K.

and Dhar

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

Alhijawi

and Kilani

, A collaborative filtering recommender system using genetic algorithm, Information Processing & Management 57(6) (2020), 102310.

Panda

S.K.

Bhoi

S.K.

and Singh

, A collaborative filtering recommendation algorithm based on normalization approach, Journal of Ambient Intelligence and Humanized Computing (2020), 1–23, 2020.

Chen

Zhao

and Chen

, Collaborative filtering recommendation algorithm based on user correlation and evolutionary clustering, Complex & Intelligent Systems 6(1) (2020), 147–156.

and Han

, A Novel Time-Aware Hybrid Recommendation Scheme Combining User Feedback and Collaborative Filtering, IEEE Systems Journal, 2020.

Netflix-Movie recommendation, https://www.kaggle.com/laowingkin/netflix-movie-recommendation, Accesed on Novemeber 2021.

10.

Bobadilla

Ortega

Hernando

and Gutiérrez

, Recommender systems survey, Knowledge-Based Systems 46 (2013), 109–132.

11.

Schafer

J.B.

Konstan

and Riedl

, Recommender systems in e-commerce, in: Proceedings of the 1st ACM conference on Electronic commerce, 1999, pp. 158–166.

12.

Burke

, Hybrid recommender systems: Survey and experiments, User Modeling and User-Adapted Interaction 12(4) (2002), 331–370.

13.

Wei

Chen

Zhou

and Tang

, Collaborative filtering and deep learning based recommendation system for cold start items, Expert Systems with Applications 69 (2017), 29–39.

14.

Hwang

C.S.

and Chen

Y.P.

, Using trust in collaborative filtering recommendation, in: Proceedings of International Conference on Industrial, Engineering and Other Applications of Applied Intelligent Systems, 2007, pp. 1052–1060.

15.

Lee

and Lee

, Scalable deep learning-based recommendation systems, ICT Express 5(2) (2019), 84–88.

16.

Chein

Y.H.

and George

, A Bayesian model for collaborative filtering, in: Proceedings of the Seventh International Workshop on Artificial Intelligence and Statistics, 1999.

17.

Ungar

L.H.

and Foster

D.P.

, Clustering methods for collaborative filtering, in: Proceedings of the Workshop on Recommendation Systems, 1998.

18.

Hofmann

, Latent semantic models for collaborative filtering, ACM Transactions Information System 22(1) (2004), 89–115.

19.

Salakhutdinov

Mnih

and Hinton

, Restricted Boltzmann machines for collaborative filtering, in: Proceedings of the 24th International Conference on Machine Learning, 2007, pp. 791–798.

20.

Georgiev

and Nakov

, A non-IID framework for collaborative filtering with restricted Boltzmann machines, in: Proceedings of the 30th International Conference on Machine Learning, 2013, pp. 1148–1156.

21.

Oord

Dieleman

and Schrauwen

, Deep content-based music recommendation, Adv. Neural Inf. Process. Syst, 2013, 2643–2651.

22.

Yuanxin

Wenqi

Wewnge

and Zhang

, Autoencoder-based colaborative filtering, in: International Conference On Neural Information Processing, 2014, pp. 284–291.

23.

Mohammadian

Forghani

and Torshiz

M.N.

, An initialization method to improve the training time of matrix factorization algorithm for fast recommendation, Soft Computing 25(5) (2021), 3975–3987.

24.

Sanwal

and Çalişkan

, A hybrid movie recommender system and rating prediction model, International Journal of Information Technology and Applied Sciences (IJITAS) 3(3) (2021), 161–168.

25.

Xie

Girshick

and Farhadi

, Unsupervised deep embedding for clustering analysis, in: International Conference on Machine Learning, 2016, pp. 478–487.

Deep embedded clustering with matrix factorization based user rating prediction for collaborative recommendation

Abstract

Keywords

1. Introduction

3. Proposed DEC with matrix factorization for collaborative recommendation

3.2.1 Customer series matrix

Clustering with KL convergence

3.4.1 Query

Binary best group

4. Results and discussion

4.1 Experimental setup

4.2 Dataset description

4.3 Performance metrics

(i) F-measure

(ii) Precision

(iii) Recall

4.4.1 Performance assessment based on queries

4.6 Comparative techniques

4.6.2 Comparative assessment using cluster size 4

Table 1 Comparative discussion

Footnotes

Abbreviations

Author’s Bios

References

Table 1
Comparative discussion