A Progressive Approach for a Personalized Recommendation System Using Hybrid Deep Learning

E-commerce has grown along with online shopping and now contributes significantly to the global economy. E-commerce platforms use recommendation algorithms to enhance the user experience and assist customers in locating relevant products, with the inclusion of many more areas including food applications and shopping websites including everyday essentials to high-end fashion. Recommendation systems are a ubiquitous and successful paradigm in the retail industry. A title of article should be the fewest possible words E-commerce recommendation systems strive to provide consumers with personalized product recommendations based on their preferences, browsing patterns, previous purchases, demographic information, and other relevant data [1]. These initiatives improve user interaction, boost platform conversion rates, and help users locate the content they are likely to find interesting. These recommendation systems make use of artificial intelligence, algorithms, and data analysis to make the following recommendations. In a typical recommendation system, there are often two sorts of data used: rating or purchase behaviors’ that represent the information that can be linked to the objects and users, and keywords or textural profiles that document the user-item interactions [2]. Overall, users and businesses benefit equally from e-commerce recommendation algorithms. Businesses may increase customer engagement, conversion rates, and revenue creation while users gain from personalized shopping experiences, effective product discovery, and increased satisfaction. E-commerce recommendation systems use concepts from Machine Learning and Deep learning.

Machine learning techniques are used by recommendation systems to analyse user data, spot trends, and produce tailored recommendations. Here are three machine learning methods that are frequently applied in recommendation systems Collaborative Filtering: To generate recommendations, collaborative filtering takes advantage of user preferences and behaviors [3]. It recognizes people with like tastes and suggests products that individuals with comparable profiles have enjoyed or purchased. This method does not explicitly require knowledge of item attributes and can be based on user-user similarity or item-item similarity. Content-Based Filtering [4] is a technique that recommends items based on the similarity of their content. [4]. This method of filtering places an emphasis on the properties of the items themselves. To construct item profiles, it examines item information like descriptions, categories, or tags. The algorithm suggests products that have qualities that the user has previously enjoyed based on the user's preferences. Hybrid Approaches: To increase suggestion accuracy and diversity, hybrid recommendation systems incorporate several strategies. In a hybrid method, for instance, collaborative filtering and content-based filtering may be combined. These systems take advantage of each technique's advantages to produce more precise and varied recommendations. The application of machine learning techniques in recommendation systems enables personalized and dynamic recommendations by continuously adapting to user feedback and learning from historical user data.

Deep learning algorithms are also used by recommendation systems to identify complex linkages and patterns in user data. Three prevalent deep learning approaches in recommendation systems are listed as follows: Deep neural networks can be used to learn non-linear representations of user-item interactions [1]. Examples of these networks include feed forward neural networks and multi-layer perceptron. These networks develop hierarchical representations using input features including user demographics, object qualities, and contextual data to produce precise suggestions. Recurrent neural networks (RNNs): RNNs operate effectively in recommendation systems with sequential data. They can represent time dependencies and the sequential character of user behavior [5]. RNNs can anticipate future user behaviour and offer individualized suggestions by considering the sequence of user interactions.

CNNs (convolutional neural networks), is frequently used in recommendation systems to extract information from item photos or textual data [6]. These networks may learn high-level representations of item features by using convolutional layers, enabling suggestions that are either visual or text-based. Deep learning techniques in recommendation systems enable the extraction of intricate patterns and latent features from user data, leading to more accurate and nuanced recommendations. In order to provide product recommendations, collaborative filtering activities require an autoencoder [7]. An autoencoder is a neural network that learns to encode inputs into a hidden (and typically low-dimensional) representation by copying input to output [8]. Given their ability to approximate any continuous function, neural networks are well suited for overcoming matrix factorization's drawbacks and boosting its expressiveness.

Traditional systems for recommendation have drawbacks, such as requiring a cold start for new users and items, The cold start problem occurs when new users or products are added into the system. Because matrix factorization relies on previous interactions between users and merchandise to reveal latent variables, it is difficult to create good predictions for fresh users who have not previously conversed with any items, or for novel products that have not yet had any interactions. Without enough data to learn from, the model cannot accurately place new users or things in the latent factor space, resulting in poor suggestions. Which are inefficient for leveraging massive data. The data amount is too vast to examine relationships, test hypotheses, and calculate values. As per current studies various machine learning and deep learning algorithms are used which includes Multi-layer perceptron based recommender system, Auto encoder based recommender system, CNN, GNN, RNN, IA CN, DHMR and many more to optimize the user experience and maximize user satisfaction [8]. This study aims to evaluate recent theoretical and practical notable accomplishments, emphasize limitations, and indicate future research areas in ML and DL for personalized recommender systems using hybrid systems.

The objectives of this work are: (1) To develop an artificial intelligence-based e-commerce recommendation system capable of producing high-quality, relevant recommendations for users. (2) Matrix factorization used to generate suggestions. (3) Cross domain recommendation which produces results from trending items. (4) Result from all the combined sets then undergoes transformer encoding. In section 3 we have explained the machine learning algorithms deeply with its experiments and analysis in section 4. We end with Conclusion and future work in section 5.

The proposed system's purpose is to establish an e-commerce recommendation system based on artificial intelligence to produce high-quality relevant recommendations for users. Since distinct systems are implemented to evaluate exclusive suggestions; let us discuss how all these methodologies collectively produce the desired results. For the first method, user behavior such as users’ interest is used, aided with matrix factorization to generate suggestions. It is a technique for dividing a big user-item interaction matrix into component matrices representing latent variables, or concealed trends in user preferences and item features. By learning these latent variables, the algorithm may anticipate a users’ likelihood of interacting with an item based on comparable individuals or products. This combination enables the recommendation engine to offer more tailored suggestions that closely match the users’ current interests. The second system emphasizes a cross-domain recommendation system, which produces results from trending items. It goes beyond the standard recommendation strategy by generating ideas from numerous unique domains or segments, rather than confining recommendations to a single domain. By recognizing popular items across multiple domains, the system may recommend what is presently popular, even if the user has no prior experience with that type of content.

Now the shown result is the log data which alludes to the recorded behavior of individuals on a platform, it includes precise information on their interactions over the course of time. This data comprises a variety of actions, including clicks, page visits, searches, buying decisions, and even time spent on certain content.

Which is then combined with a decision tree and matrix factorization to produce recommendations. Decision trees are used to categorize users relying on easily comprehensible criteria, and then matrix factorization is applied within each group to produce individualized suggestions. This method enables the system to detect both obvious correlations between parameters and more subtle, latent patterns in user behavior.

Applying context-aware framework to get popular items in social media. Following up, when the recently searched items are taken for profiling and matching items, the system generates a set of recommended items for the user. All the results when compiled together give a set with recommendations from all the different systems mentioned above. To build a highly refined recommendation set as a result for this combined set, it undergoes transformer encoding. Let us now discuss each system in detail.

3.1 User based recommendations with the help of deep neural networks

User Behaviour Analysis: In the suggested approach, the user’s interactions and browser information, i.e., user’s interest is used. The multilayer perceptron (MLP) is non-linearly activated feed-forward neural networks that consist of three or more layers which includes input layer, hidden layers, and an output layer, employing nonlinear activation functions in its neurons to record complex patterns in human behavior and item properties. By evaluating user and item data via these layers, MLPs learn to anticipate the chance of a user interacting with a specific item. This strategy is especially successful in situations where several variables influence user habits in multifaceted ways.

As a result, users' preferences for items are represented as a rating matrix to construct the relationship between users and items to locate users' relevant objects. The matrix factorization can be considered as a one directional model of latent vectors Eq. (1) av and bi.

Figure 1 shows the process starts with customer behavior analysis, which collects and analyzes data on users' interactions with products such as clicks, views, ratings etc. This analysis aids to comprehend user interests and patterns in user-item relationships, which are the foundation for creating correct suggestions.

This behavioural data is then sent into a deep neural network (DNN), which has levels of processing to capture complicated, non-linear interactions between users and things.

The multilayer perceptron (MLP) continues to process the data by passing it through several layers of neurons with non-linear activation functions. The MLP helps to develop deep, latent representations of persons and items, as well as to detect hidden elements that influence user preferences.

Following the MLP, matrix factorization is used to break down the user-item interface matrix into lower-dimensional matrices. This stage aids in identifying latent elements that are not readily apparent, such as a users’ preference for specific genres or product attributes.

Finally, the system analyses this data to provide recommendations based on each users’ specific tastes. The flow chart describes a complete method of recommendation that incorporates user behavior analysis, deep learning, and matrix factorization, leveraging both explicit and implicit signals to improve the relevance and efficacy of suggestions presented to users.

1.jpg

Figure 1. Recommendations based on user behavior

$x_{v i}=g\left(v, i \| a_v, b_i\right)=a_v^t b_i=\sum k$          (1)

where,

x_vi: Real valued latent vector for user

bi: Real valued latent vector for item

k: Latent space

The bi-directional user and item latent factors’ communication considers that each direction of the latent space is linearly added with the same load which results in a filtered list of items produced from user behavior analysis for certain items in the e- commerce system. This model is further formulated as:

$x v i=\sigma\left(f^t(\phi o u t)\right)$           (2)

In order to describe user-item latent structures, Eq. (2) suggests that the model combines linearity from matrix factorization and non-linearity from neural networks.

3.2 Recommendations based on log data using active learning techniques

Every user-item pairing in a recommender system is essential for determining a user's preferences and has a big influence on how well the system works. The problem of data scarcity demonstrates that the more user ratings a system collects, the better it will perform in making recommendations.

Audit Trail: The web server keeps a log file called the audit trail. This online access server log describes how, and which pages are accessed by various web users. It is an orderly record that contains extensive details on the order in which operations or transactions occur within a system. This record contains information such as the algorithms utilized, the parameters implemented, the user behaviors assessed, and the recommendations issued. It is vital for transparency, accountability, and compliance.

Recommender systems are now more accurate and efficient because of the implementation of active learning, which helps them choose the most representative things to give to consumers. Rapid profiling of users and ratings is often achieved through the rating impact analysis and bootstrapping process.

Figure 2 shows the method commences with user audit trails, which provide a thorough record of previous interactions, judgments, and suggestions, providing insight into how the system operates.

This data is paired with user ratings, which represent user feedback on things, to determine the following stages.

The exemplary item stage entails choosing crucial items that embody the user's preferences or are necessary in interpreting their behavior. These things are utilized in active learning, a procedure in which the system intelligently chooses the information points to prioritize in order to increase its forecast accuracy.

Following that, rating effect analysis/bootstrapping is used to determine how modifications in user ratings or the addition of new data affect the system's suggestions. Bootstrapping techniques can be used to evaluate the dependability of these adjustments, guaranteeing that the recommendations are strong and not unduly sensitive to minor differences in data.

This results in an improved comprehension of user behavior, enabling the system to tailor its recommendations based on more precise and complete information. Finally, the procedure generates revised recommendations based on the audit trail, to guarantee the recommendations are not only individualized but also consistent with the system's overall aims of openness and accountability.

2.jpg

Figure 2. Recommendations using active learning with log data as a dataset

We employ decision trees for the model shown in Figure 3 because they produce a list of suggested items at their lead nodes. The recommendation system may scale by reducing the amount of searching necessary while creating a user's personalized suggestion list utilizing the tree. Similar to a tree probability formulation, a max heap is employed. For each user u, each non-leaf node x in level j fulfils Eq. (3) as follows:

$x_0\left\{x^{\prime}  schildrennodeninlevel  j+1\right\}^{r(j+1)}(u) / a(j)$            (3)

where,

r^(j)(x|u)-probability that user u prefers

a^(j)-normalization for j

3.jpg

Figure 3. Context aware recommendation using deep learning

3.3 Recommendation based on social media popularity based on context aware framework using deep learning

A context-aware recommendation system establishes linkages among context, things, and users. Here, we start by creating a numerical vector of all prominent instances, to which we then add compressed latent contextual embeddings. Finally, we make systematic and implicit use of hierarchical contextual information. Figure 4 represents the framework.

We extracted context representation from apparent, implicit unstructured, and implicit structured data. To create a context-dependent vector with the form context = [x1, x2,..., xT], We initially normalize the context-relevant feature values to a 0–1 scale and turn conventional qualities into binary traits.

4.jpg

Figure 4. Recommendations from the current search dataset

In addition to unstructured latent contextual data, which is a compact representation including latent L values (LT) derived from hidden layers, explicit contextual representation extraction involves employing all of the available (raw) T contextual characteristics. This latent contextual information is utilized for categorization, top-recommendation, along with rating prediction. A compact feature representation is produced by the auto-encoder approach by passing the initial input values through a low-dimensional layer once again. In order to understand the structure of latent variables and the significance of relationships between them, we deploy a structured latent contextual representation.

For the recommendation process, we use a structured latent contextual model that consists of a set of contextual factors (cluster Identification) at various granularity levels that are generated via a hierarchical tree. A hierarchical model is created by automatically classifying an extensive set of unstructured implicit contextual vectors into a limited number of clusters, each representing an implicit contextual event.

To acquire structured latent contextual vectors, we primarily extract compressed latent contextual vectors from an AE's bottleneck layer, and then use the unstructured latent contextual vectors to construct a hierarchical tree at the resolution level. In order to combine contextual vectors with similar latent qualities that are pertinent to a certain context or organisational latent, we use automated agglomerative hierarchical clustering (AHC) to apply the k- means algorithm approach and assess how many ambient scenarios (clusters) there may be. It is a bottom-up clustering technique that can be applied in recommendation systems for combining comparable individuals or objects based on their features or interactions. In AHC, each client or item is assigned to its own cluster, which is recursively integrated based on resemblance until a certain number of clusters or a stopping threshold is met.

The path taken by raw information, or latent context-dependent vector, from a tree's top to a leaf. Then, we compile user reviews or ratings for each item and standardize their representation as one-hot vectors by index ID. The model is then trained using the data we have gathered about items, users, and contexts. We make advantage of ECAM, which gathers all latent unstructured contexts and feeds them into a neural CF model. In order to add unstructured latent contexts, we secondly employ the UCAM model. The HCAM model, which encompasses the hierarchical latent contexts observed in neural-CF models, is the last one we use. We amalgamate the contextual vector c to the client u and product i integration to develop a new nonlinear function spanning all three elements-objects, people, and contexts. In order to automatically assess how context impacts the predicted target as a whole, the neural framework considers context. In order to forecast which class would have the highest likelihood in a classification task, we first calculate the likelihood of each class label (categorical value) using the SoftMax activation function.

U: Set of Users(U=u1,u2,…,um)

I: Set of Items(I=i1,i2,…,in)

R: Matrix of user item ratings

C: Matrix of contextual information, where each data point cui represents the contextual information

We can modify the factorization equation as in Eq. (4):

$R \approx U^* I^T+C^* H$             (4)

The latent elements connected to the contextual information are captured by the matrix H. Through an optimisation procedure that reduces the difference between the estimated ratings derived from the factorization and the real ratings, H may be discovered.

3.4 Recommendations based on current search data using fuzzy techniques

Fuzzy collection and fuzzy link concepts can be helpful for addressing information uncertainty and can be used to recommender systems [24]. As shown in Figure 4, the two main categories of fuzzy recommendation approaches are memory-based CF recommender systems with fuzzy methods and content-based recommender systems with fuzzy techniques. This system can be enhanced through the application of fuzzy methods, which improve the handling of uncertainty and imprecision in user preferences. The dataset being utilized is the most current data that was searched.

These strategies can increase accuracy in specific areas by matching consumer desires with the services offered and reducing the inevitable noise of uncertainty and producing refined recommendation sets.

In Figure 4, it starts with the existing search data that gathers the user's present interests depending on their most recent queries or interactions with the system.

Next, fuzzy approaches are used to deal with the inherent unpredictability and complexity in client search activity. Fuzzy logic enables the system to make more versatile and subtle decisions by taking into account degrees of significance rather than strictly labelling items as relevant or irrelevant.

Following that, the system employs either content-based or memory-based collaborative filtering (CF). Content-based filtering suggests things like the ones with which the user previously communicated with respect to item qualities. Memory-based collaborative filtering, on the other hand, recommends products based on the actions of users with similar tastes, utilizing the expertise of the community to forecast outcomes.

The recommendations are further refined by profiling/matching applicable items, in which the algorithm fits possible ideas with the client's already defined profile or inclinations.

Finally, the procedure ends with a filtered set from the currently selected search, which displays a carefully selected group of objects that pertain to the clients’ current search context and general preferences.

Rt maps from the rewards/scores ri. We now use the previous calculation to determine the end result and focal point of the entire list of recommendations:

$p t=x \sum \gamma^{k-1} u^k$               (5)

k=1

k-Order of each item

K-total items

$\gamma \in(0,1]$

Let us find the Pearson correlation coefficient, which computes the similarity between two users, vx and vy, based only on their common ratings.

$corr\left(v_x, v_y\right)=\frac{\sum sg x, y\left(p_{(x, y)}-m_x\right)\left(p_{(x, y)}-m_y\right)}{\sqrt{\sum\left(p_{(x, k)}-m\right)^2\left(p_{(y, k)}-m\right)^2}}$              (6)

where,

px - Ratings of users vx for k^th item

pbi - Ratings of users vy for k^th item m-average ratings of user v

my - Average rating of user vy

In Figure 4, Current search data-based recommendation algorithm using fuzzy techniques

The cosine of angle will be:

$cos \left(v_x, v_y\right)=\frac{\sum_p x, k p_y, k}{\sqrt{\sum_{p_{x, k}^2} p_k^2}}$              (7)

The model is forced to arrange things that the user may order in the top of the suggested list because the idea from the aforementioned Eqs. (5)-(7) is that the final result and focus point on the top of the recommended list has a bigger contribution to the overall outcomes.

These sets are further augmented to form a quality recommendation set for the users.

3.5 Recommendations based on a collective set of generated recommendations using a personalized Re-Ranking model

A re-ranking model in a recommendation system is a sophisticated strategy for refining the sequence of the first set of recommendations, to make sure the most relevant things are given preference for the user. After the system generates an initial group of recommendations using typical approaches such as collaborative filtering or content-based filtering, the re-ranking model reorganizes the list depending on further factors that were not completely assessed in the original ranking [25].

The input layer receives a combined list of things made up of the outcomes of the earlier approaches covered in this work, which it uses to create comprehensive representations of all the items in the original list and send them to the encoding layer. Raw feature matrix X serves as a representation for the first sequential list of fixed lengths. For each item i in S, a row in X stands in for the raw feature vector xi.

Input Layer: The raw matrix and the customized matrix PV are combined to create an intermediate matrix E' for a user- specific encoding function. Additionally, add a position embedding PE to E' to take advantage of the sequential information in the original list.

Encoding Layer: In order to fulfil the combination of mutual influences, a transformer encoder is used, because, it directly represents the reciprocal influences of any two items, regardless of their distances. Every element in this encoding module's N-block structure has an attention module and a Feed-Forward Network (FNN). Eq. (8) describes how the attention layer works.

$Attention(A, B, C)=softmax\left(\frac{A B^T}{\sqrt{d}}\right) C$          (8)

where,

A-queries matrix

B-keys matrix

C-values matrix

D-dimensionality of matrix K table

Proposed algorithm

Output Layer: This layer aims to produce a score for each item I = i1, ……, in. Here, a linear layer and a SoftMax layer are utilized to determine how an item-pair's impacts interact. Eq. (9) contains the score formulation.

$Attention(i)=P(X, P V ; \theta)=softmax\left(A^{\left(o_x\right)} Q^F b^F\right), i \in S$               (9)

where,

A-the output of N_x blocks

Q-projection matrix

B^F-bias term

o-total items

This will generate a probability of clicking for each item to generate a score. The set items with the relevant scores will be selected and recommended to the user. Based on the analysis performed so far, we can claim that the proposed personalized hybrid e-commerce recommendation system outperformed existing approaches.

The proposed e-commerce recommendation system overcomes the drawbacks and performs multiple filtering to produce rectified recommendations.