A RECOMMENDATION SYSTEM FOR GENERATING PERSONALIZED ITEM RECOMMENDATIONS TO USERS AND A METHOD THEREOF

Abstract

ABSTRACT A RECOMMENDATION SYSTEM FOR GENERATING PERSONALIZED ITEM RECOMMENDATIONS TO USERS AND A METHOD THEREOF The present disclosure discloses a recommendation system for generating personalized item recommendations to users and a method thereof. The system(100) comprises a bipartite graph construction module(102) to receive a rating matrix (R) containing user-item interactions to construct a bipartite graph; a community detection module(104) to apply a Louvain community detection model(104a) to partition said bipartite graph into smaller, distinct communities(BG1, BG2, ..., BGc); a content-based filtering module(106) to apply a cosine similarity to generate content-based filtering metrics within each community; a collaborative filtering module(108) to apply matrix factorization techniques to generate collaborative-based filtering metrics; a fusion module(110) to combine said content-based and said collaborative-based filtering metrics thereby producing enhanced predicted rating matrices(R̃1, R̃2, ..., R̃c) that balance said item attribute relevance with latent user preferences, leveraging the strengths of both recommendation techniques; an aggregation module(112) to aggregate said enhanced predicted matrices from all communities into a final predicted rating matrix(R̃).

Specification

Description:FIELD
The present disclosure generally relates to the field of counseling systems. More particularly, the present disclosure relates to a recommendation system for generating personalized item recommendations to users and a method thereof.
BACKGROUND
The background information herein below relates to the present disclosure but is not necessarily prior art.
Existing recommendation systems are essential in helping users discover relevant content or products across industries like e-commerce, media streaming, and social networks. However, these systems face significant technical limitations that impact their performance and effectiveness. One major challenge is data sparsity many users interact with only a small subset of items, creating gaps in the data and making it difficult to generate accurate recommendations. Additionally, these systems struggle with the cold start problem for new users and items, where limited or no interaction data makes it challenging to predict preferences. Scalability is another issue, as handling large datasets with growing user bases and item catalogs can slow down response times and reduce recommendation quality.
Furthermore, existing systems often lack diversity in recommendations, frequently suggesting similar or popular items, which leads to a monotonous experience and limits user discovery of unique content. Many systems also lack the ability to adapt to dynamic user preferences; as user interests change over time, they may continue to receive outdated or irrelevant recommendations based on historical data. Without mechanisms to quickly adjust to evolving user behavior, the accuracy and relevance of recommendations decline, impacting user engagement and satisfaction. These limitations highlight the need for more advanced recommendation systems that can address data sparsity, cold starts, scalability, diversity, and adaptability to provide consistently relevant and engaging recommendations.
There is, therefore felt a need for a recommendation system for generating personalized item recommendations to users and a method thereof that alleviates the aforementioned drawbacks.
OBJECTS
Some of the objects of the present disclosure, which at least one embodiment herein satisfies, are as follows:
It is an object of the present disclosure to ameliorate one or more problems of the prior art or to at least provide a useful alternative.
An object of the present disclosure is to provide a recommendation system for generating personalized item recommendations to users.
Another object of the present disclosure is to provide a system that improves the accuracy of personalized item recommendations.
Still, another object of the present disclosure is to provide a system that scales efficiently with large datasets by using community-based segmentation and applying filtering techniques to smaller community-specific matrices.
Yet another object of the present disclosure is to provide a system that combines content-based and collaborative filtering metrics.
Still another object of the present disclosure is to provide a system allowing real-time or periodic community re-partitioning and recalibration of filtering metrics.
Yet another object of the present disclosure is to provide a system that combines predicted ratings from all communities into a unified recommendation output.
Still another object of the present disclosure is to provide a system that is highly accurate and efficient, providing enhanced personalization and scalability.
Other objects and advantages of the present disclosure will be more apparent from the following description, which is not intended to limit the scope of the present disclosure.
SUMMARY
The present disclosure envisages a recommendation system for generating personalized item recommendations to users and a method thereof. The system comprises a bipartite graph construction module, a community detection module, a content-based filtering module, a collaborative filtering module, a fusion module, and an aggregation module.
The bipartite graph construction module is configured to receive a rating matrix (R) containing user-item interactions to construct a bipartite graph, represent users and items as nodes in the bipartite graph, and connect the user and item nodes with edges based on available ratings.
The community detection module is configured to apply a Louvain community detection model to partition the bipartite graph into smaller, distinct communities (BG1, BG2, ..., BGc), each represented by a community-specific rating matrix (R1, R2, ..., Rc).
The content-based filtering module is configured to apply a cosine similarity within each community-specific rating matrix (R1, R2, ..., Rc) to generate content-based filtering metrics within each community.
The collaborative filtering module is configured to apply matrix factorization techniques, including singular value decomposition (SVD++) or basic matrix factorization, to each community-specific rating matrix (R1, R2, ..., Rc), to generate collaborative-based filtering metrics.
The fusion module is configured to combine the content-based and the collaborative-based filtering metrics for each community through a convex combination thereby producing enhanced predicted rating matrices (R̃1, R̃2, ..., R̃c) that balance the item attribute relevance with latent user preferences, leveraging the strengths of both recommendation techniques for optimal prediction accuracy.
The aggregation module is configured to aggregate the enhanced predicted matrices from all communities into a final predicted rating matrix (R̃), representing the system's recommendation outputs.
In an embodiment, the system further comprises an evaluation module is configured to quantify recommendation accuracy by calculating the root mean square error (RMSE) between the original rating matrix (R) and the final predicted matrix (R̃), thereby providing an indication of the recommendation system's accuracy.
In an embodiment, the system further comprises a user feedback module configured to gather real-time user feedback on recommended items and feed this feedback back into the content-based and collaborative filtering modules, thereby enhancing future recommendations based on actual user interactions.
In an embodiment, the community detection module further comprises a graph construction sub-module configured to construct the bipartite graph from user-item interactions by assigning edge weights based on the ratings given by users to items.
In an embodiment, the community detection model is configured to dynamically adjust the number of communities (BG1, BG2, ..., BGc) based on the density and clustering of the user-item interaction graph, to ensure optimal partitioning of the data for each specific dataset.
In an embodiment, the content-based filtering module is further configured to use genre information, item descriptions, and other item-specific attributes as input features to calculate the cosine similarity within each community, thereby generating content-based prediction matrices (CBR1, CBR2, ..., CBRc) tailored to specific user preferences within each community.
In an embodiment, the collaborative filtering module applies matrix factorization techniques selected from the group consisting of basic matrix factorization (MF), singular value decomposition (SVD), and SVD++, to each community-specific rating matrix, thereby decomposing user-item interactions into latent factors with minimal computational complexity.
In an embodiment, the collaborative filtering module is further configured to address data sparsity by incorporating implicit feedback data, such as user clicks or browsing history, into the matrix factorization model, enhancing the prediction accuracy for items with sparse ratings.
In an embodiment, the fusion module is configured to adjust the weight of the convex combination dynamically based on the density of the community-specific rating matrices, the content similarity between items, and the strength of the latent factors, to achieve a balance between content-based and collaborative filtering predictions for each community.
The present disclosure also envisages a recommendation method for generating personalized item recommendations to users. The method comprises the following steps:
• receiving, by a bipartite graph construction module, a rating matrix (R) containing user-item interactions to construct a bipartite graph, represent users and items as nodes in the bipartite graph, and connecting the user and item nodes with edges based on available ratings;
• applying, by a community detection module, a Louvain community detection model to partition the bipartite graph into smaller, distinct communities (BG1, BG2, ..., BGc), each represented by a community-specific rating matrix (R1, R2, ..., Rc);
• applying, a content-based filtering module, a cosine similarity within each community-specific rating matrix (R1, R2, ..., Rc) to generate content-based filtering metrics within each community;
• applying, a collaborative filtering module, matrix factorization techniques, including singular value decomposition (SVD++) or basic matrix factorization, to each community-specific rating matrix (R1, R2, ..., Rc), to generate collaborative-based filtering metrics;
• combining, by a fusion module, the content-based and the collaborative-based filtering metrics for each community through a convex combination thereby producing enhanced predicted rating matrices (R̃1, R̃2, ..., R̃c) that balance the item attribute relevance with latent user preferences, leveraging the strengths of both recommendation techniques for optimal prediction accuracy; and
• aggregating, by an aggregation module, the enhanced predicted matrices from all communities into a final predicted rating matrix (R̃), representing the system's recommendation outputs.
BRIEF DESCRIPTION OF THE ACCOMPANYING DRAWING
A recommendation system for generating personalized item recommendations to users and a method thereof of the present disclosure will now be described with the help of the accompanying drawing, in which:
Figure 1 illustrates a block diagram of a recommendation system for generating personalized item recommendations to users in accordance with an embodiment of the present disclosure;
Figures 2A-2B illustrate a flowchart for a method for a recommendation system for generating personalized item recommendations to users in accordance with an embodiment of the present disclosure;
Figure 3 illustrates the architecture of the recommendation system in accordance with an embodiment of the present disclosure;
Figures 4A-4C illustrate the graph depicting the RMSE metrics for the basic matrix factorization in accordance with an embodiment of the present disclosure; and
Figures 5A-5C illustrate the graph depicting the RMSE metrics for the SVD++ method in accordance with an embodiment of the present disclosure.
LIST OF REFERENCE NUMERALS
100 - System
102 - Bipartite Graph Construction Module
104 - Community Detection Module
104a - Community Detection Model
104b - Graph Construction Sub-Module
106 - Content-Based Filtering Module
108 - Collaborative Filtering Module
110 - Fusion Module
110a - Weight Optimization Sub-Module
112 - Aggregation Module
114 - Evaluation Module
116 - User Feedback Module
DETAILED DESCRIPTION
Embodiments, of the present disclosure, will now be described with reference to the accompanying drawing.
Embodiments are provided so as to thoroughly and fully convey the scope of the present disclosure to the person skilled in the art. Numerous details, are set forth, relating to specific components, and methods, to provide a complete understanding of embodiments of the present disclosure. It will be apparent to the person skilled in the art that the details provided in the embodiments should not be construed to limit the scope of the present disclosure. In some embodiments, well-known processes, well-known apparatus structures, and well-known techniques are not described in detail.
The terminology used, in the present disclosure, is only for the purpose of explaining a particular embodiment and such terminology shall not be considered to limit the scope of the present disclosure. As used in the present disclosure, the forms "a," "an," and "the" may be intended to include the plural forms as well, unless the context clearly suggests otherwise. The terms "including," and "having," are open ended transitional phrases and therefore specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not forbid the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. The particular order of steps disclosed in the method and process of the present disclosure is not to be construed as necessarily requiring their performance as described or illustrated. It is also to be understood that additional or alternative steps may be employed.
When an element is referred to as being "engaged to," "connected to," or "coupled to" another element, it may be directly engaged, connected, or coupled to the other element. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed elements.
Existing recommendation systems are essential for delivering personalized content across industries like e-commerce and media streaming. However, they face several technical limitations. Data sparsity is a primary issue, as users typically interact with only a small subset of items, making it difficult to generate accurate predictions, especially for less popular items. The cold start problem further complicates recommendations for new users and items due to a lack of initial interaction data. Scalability is another challenge, as large user bases and item catalogs demand significant computational resources, potentially slowing down real-time recommendations. Additionally, these systems often lack diversity in recommendations, frequently suggesting popular or similar items, which limits content discovery. Finally, many systems are not adaptive to changing user preferences, resulting in outdated or irrelevant recommendations over time. These limitations highlight the need for more advanced systems that can handle sparse data, scale efficiently, and adapt dynamically to user behaviors.
To address the issues of the existing systems and methods, the present disclosure envisages a recommendation system (hereinafter referred to as "system 100") for generating personalized item recommendations to users and a recommendation method (hereinafter referred to as "method 200") for generating personalized item recommendations to users. The system 100 will now be described with reference to Figure 1 and the method 200 will be described with reference to Figures 2A-2B.
Figure 1 shows the recommendation system 100 recommendation system for generating personalized item recommendations to users. The system 100 comprises a bipartite graph construction module 102, a community detection module 104, a content-based filtering module 106, a collaborative filtering module 108, a fusion module 110, and an aggregation module 112.
The bipartite graph construction module 102 is configured to receive a rating matrix (R) containing user-item interactions to construct a bipartite graph, represent users and items as nodes in the bipartite graph, and connect the user and item nodes with edges based on available ratings.
The community detection module 104 is configured to apply a Louvain community detection model 104a to partition the bipartite graph into smaller, distinct communities (BG1, BG2, ..., BGc), each represented by a community-specific rating matrix (R1, R2, ..., Rc).
The community detection module 104 includes a community detection model 104a and a graph construction sub-module 104b.
In an embodiment, the community detection model 104a is configured to dynamically adjust the number of communities (BG1, BG2, ..., BGc) based on the density and clustering of the user-item interaction graph, to ensure optimal partitioning of the data for each specific dataset.
In an embodiment, the graph construction sub-module 104b is configured to construct the bipartite graph from user-item interactions by assigning edge weights based on the ratings given by users to items.
The content-based filtering module 106 is configured to apply a cosine similarity within each community-specific rating matrix (R1, R2, ..., Rc) to generate content-based filtering metrics within each community.
In an embodiment, the content-based filtering module 106 is further configured to use genre information, item descriptions, and other item-specific attributes as input features to calculate the cosine similarity within each community, thereby generating content-based prediction matrices (CBR1, CBR2, ..., CBRc) tailored to specific user preferences within each community.
In an embodiment, the collaborative filtering module 106 is further configured to address data sparsity by incorporating implicit feedback data, such as user clicks or browsing history, into the matrix factorization model, enhancing the prediction accuracy for items with sparse ratings.
The collaborative filtering module 108 is configured to apply matrix factorization techniques, including singular value decomposition (SVD++) or basic matrix factorization, to each community-specific rating matrix (R1, R2, ..., Rc), to generate collaborative-based filtering metrics.
In an embodiment, the collaborative filtering module 108 applies matrix factorization techniques selected from the group consisting of basic matrix factorization (MF), singular value decomposition (SVD), and SVD++, to each community-specific rating matrix, thereby decomposing user-item interactions into latent factors with minimal computational complexity.
The fusion module 110 is configured to combine the content-based and the collaborative-based filtering metrics for each community through a convex combination thereby producing enhanced predicted rating matrices (R̃1, R̃2, ..., R̃c) that balance the item attribute relevance with latent user preferences, leveraging the strengths of both recommendation techniques for optimal prediction accuracy.
In an embodiment, the fusion module 110 is configured to adjust the weight of the convex combination dynamically based on the density of the community-specific rating matrices, the content similarity between items, and the strength of the latent factors, to achieve a balance between content-based and collaborative filtering predictions for each community.
In an embodiment, the fusion module 110 further includes a weight optimization sub-module 110a that learns an optimal weighting parameter for each community, maximizing the predictive accuracy of the enhanced predicted rating matrices (R̃1, R̃2, ..., R̃c).
The aggregation module 112 is configured to aggregate the enhanced predicted matrices from all communities into a final predicted rating matrix (R̃), representing the system's recommendation outputs.
In an embodiment, the system 100 further comprises an evaluation module 114 configured to quantify recommendation accuracy by calculating the root mean square error (RMSE) between the original rating matrix (R) and the final predicted matrix (R̃), thereby providing an indication of the recommendation system's accuracy.
In an embodiment, the system 100 further comprises a user feedback module 116 configured to gather real-time user feedback on recommended items and feed this feedback back into the content-based and collaborative filtering modules, thereby enhancing future recommendations based on actual user interactions.
In an embodiment, the bipartite graph construction module 102 receives a rating matrix 'R', capturing user-item interactions, and constructs a bipartite graph. Users and items are represented as nodes connected by edges according to ratings. The community detection module 104 applies the Louvain community detection model 104a to partition this bipartite graph into distinct communities, represented by community-specific rating matrices (R1, R2, ..., Rc). This approach segments users and items into groups of similar interests, enhancing recommendation relevance.
In an embodiment, each community-specific rating matrix (R1, R2, ..., Rc) undergoes content-based filtering within a content-based filtering module 106. Cosine similarity is calculated for item comparisons within each community, generating content-based prediction matrices that reflect user preferences based on item attributes, such as genre, description, and other item-specific features.
In another embodiment, for each community-specific rating matrix, the collaborative filtering module 108 applies matrix factorization techniques, such as singular value decomposition (SVD++), to identify latent factors in user-item interactions. This allows the system to detect hidden patterns and preferences, even for sparsely rated items, and provide collaborative-based filtering metrics.
In an embodiment, the fusion module 110 combines content-based and collaborative-based filtering metrics for each community through a convex combination, creating enhanced predicted rating matrices (R̃1, R̃2, ..., R̃c). This approach balances item attribute relevance with latent preferences, allowing both recommendation techniques to complement one another and maximize prediction accuracy.
In an embodiment, the aggregation module 112 consolidates the enhanced predicted matrices from all communities into a final predicted rating matrix (R̃), which represents the system's overall recommendation output. This final matrix provides users with recommendations based on both community-level preferences and individual user interactions.
In an embodiment, the evaluation module 114 measures recommendation accuracy by calculating the root mean square error (RMSE) between the original rating matrix R and the final predicted matrix (R̃). This metric allows for continuous assessment of recommendation precision, enabling system tuning and validation.
In an embodiment, the user feedback module 116 gathers real-time feedback from users on recommended items, feeding this information back into the content-based and collaborative filtering modules. This module enables the system to incorporate actual user interactions into future predictions, adapting to evolving user preferences for greater personalization.
In an embodiment, the community detection model 104a adjusts the number of communities (BG1, BG2, ..., BGc) dynamically based on graph density and clustering tendencies in user-item interactions. This adjustment ensures optimal partitioning of data for each dataset, improving segmentation accuracy and adaptability across various data distributions.
In an embodiment, the collaborative filtering module 108 can incorporate implicit feedback data, such as clicks or browsing history, in addition to explicit ratings. This addition addresses data sparsity issues, particularly for less popular items, by including indirect indicators of user interest, which enhances the recommendation accuracy for sparsely rated items.
In an embodiment, the fusion module 110 dynamically adjusts the weights in the convex combination based on community-specific data, such as matrix density, content similarity, and strength of latent factors. A weight optimization sub-module 110a can further refine these weights to maximize predictive accuracy, allowing the system to adaptively prioritize content-based or collaborative predictions based on community characteristics.
In an embodiment, the recommender system is an information filtering tool that predicts user preferences or ratings for various items. Its primary aim is to deliver accurate recommendations tailored to user interests. These systems are essential for providing personalized content and improving user experiences. The recommender system is classified as content-based, collaborative filtering, and hybrid recommendations. Content-based recommendations offer tailored, explainable suggestions, while collaborative filtering plays a crucial role in recommendation systems by predicting user preferences by analyzing common behavior patterns. It provides personalized recommendations without needing content-specific data, works across various fields, and scales well with larger datasets. By utilizing user interactions, it uncovers hidden preferences and boosts recommendation accuracy. Matrix factorization strengthens collaborative filtering by decomposing user-item interactions into latent features, boosting recommendation precision while addressing challenges like scalability, data sparsity, and the cold-start issue efficiently.
However, as datasets grow, scalability becomes a challenge. This paper introduces a hybrid approach that combines content-based and collaborative filtering with community detection to tackle scalability. By modeling the rating matrix as a bipartite network, generating communities, and applying filtering techniques in parallel, the method improves recommendation accuracy. Matrix factorization techniques such as basic MF and SVD++, used with the Louvain algorithm, are tested on three datasets. The approach achieves 94% effectiveness by reducing RMSE, and providing accurate, user-centered recommendations.
Figures 2A-2B illustrate a flowchart for a recommendation method for generating personalized item recommendations to users in accordance with an embodiment of the present disclosure. The order in which method 200 is described is not intended to be construed as a limitation, and any number of the described method steps may be combined in any order to implement method 200, or an alternative method. Furthermore, method 200 may be implemented by processing resource or computing device(s) through any suitable hardware, non-transitory machine-readable medium/instructions, or a combination thereof. The method 200 comprises the following steps:
At step 202, the method 200 includes receiving, by a bipartite graph construction module 102, a rating matrix (R) containing user-item interactions to construct a bipartite graph, represent users and items as nodes in the bipartite graph, and connecting the user and item nodes with edges based on available ratings.
At step 204, the method 200 includes applying, by a community detection module 104, a Louvain community detection model 104a to partition the bipartite graph into smaller, distinct communities (BG1, BG2, ..., BGc), each represented by a community-specific rating matrix (R1, R2, ..., Rc).
At step 206, the method 200 includes applying, a content-based filtering module 106, a cosine similarity within each community-specific rating matrix (R1, R2, ..., Rc) to generate content-based filtering metrics within each community.
At step 208, the method 200 includes applying, a collaborative filtering module 108, matrix factorization techniques, including singular value decomposition (SVD++) or basic matrix factorization, to each community-specific rating matrix (R1, R2, ..., Rc), to generate collaborative-based filtering metrics.
At step 210, the method 200 includes combining, by a fusion module 110, the content-based and the collaborative-based filtering metrics for each community through a convex combination thereby producing enhanced predicted rating matrices (R̃1, R̃2, ..., R̃c) that balance the item attribute relevance with latent user preferences, leveraging the strengths of both recommendation techniques for optimal prediction accuracy.
At step 212, the method 200 includes aggregating, by an aggregation module 112, the enhanced predicted matrices from all communities into a final predicted rating matrix (R̃), representing the system's recommendation outputs.
Figure 3 illustrates the architecture of the recommendation system in accordance with an embodiment of the present disclosure. The system constructs a bipartite graph (BG) using user-item interaction data from a rating matrix R. The edges of this graph are weighted based on the ratings provided by users for items. This architecture representation allows the system to capture user-item interactions visually and structurally. Using the Louvain community detection algorithm 104a, the bipartite graph is partitioned into distinct communities, denoted as (BG1, BG2, ..., BGc). Each community contains a subset of users and items with higher similarity in preferences, represented by community-specific rating matrices (R1, R2, ..., Rc). This segmentation allows the system to process smaller, more focused groups for personalized recommendations.
Within each community-specific rating matrix, the system applies a content-based filtering process. It utilizes cosine similarity to calculate content-based metrics (CBR1, CBR2, ..., CBRc) based on item attributes, such as genre information. This ensures that recommendations within each community consider item-specific features relevant to users' preferences in that community.
In parallel, the collaborative filtering module performs matrix factorization (e.g., SVD++ or basic matrix factorization) on each community-specific rating matrix to generate collaborative-based filtering metrics (CFR1, CFR2, ..., CFRc). These metrics represent latent user preferences derived from patterns in user-item interactions.
The content-based (CBRi) and collaborative-based (CFRi) filtering metrics for each community are then combined through a convex combination, represented by R̃i=[W∗(CBRi)+(1−W)∗(CFRi)], where 'W' is a weight factor. This fusion balances item attribute relevance with latent preferences, creating an enhanced predicted rating matrix R̃i for each community.
The predicted rating matrices (R̃1, R̃2, ..., R̃c). from each community are then aggregated to form a final predicted rating matrix (R̃). This consolidated matrix represents the recommendation system's outputs, combining insights from each community for comprehensive recommendations.
Finally, an RMSE (Root Mean Square Error) calculation compares the original rating matrix 'R' with the final predicted rating matrix (R̃), to assess the accuracy of the recommendations. Lower RMSE values indicate higher prediction accuracy, enabling continuous performance assessment and system refinement.
Figures 4A-4C illustrate the graph depicting the RMSE metrics for the basic matrix factorization in accordance with an embodiment of the present disclosure. Figure 4A shows the examination of the RMSE metrics for the basic matrix factorization method across k=10 latent features and 25 communities for movie lens-100K. Figure 4B shows the examination of the RMSE metrics for the basic matrix factorization method across k=10 latent features and 25 communities for movie lens-1M. Figure 4C shows the examination of the RMSE metrics for the basic matrix factorization method across k=10 latent features and 25 communities for anime recommendation datasets. The datasets were evaluated with varying convex combinations of weights for content-based and collaborative filtering methods. In the MovieLens-100K dataset, the lowest RMSE was achieved at community 11 with a weight of 0.4 for content-based filtering and the remaining assigned to collaborative filtering. Similarly, for the MovieLens-1M dataset, the lowest RMSE occurred at community 7 with weights of 0 for content and 0.6 for collaborative filtering. For the anime recommendation dataset, the lowest RMSE was obtained at community 22 when using content information weights of 0.6, 0.8, and complete content information.
These findings highlight that combining content and collaborative filtering information in a hybrid model significantly reduces prediction error across various datasets. Moreover, the optimal balance between content-based and collaborative filtering varies depending on the dataset, underscoring the importance of dataset-specific tuning for achieving the best performance. The overall results demonstrate that leveraging both filtering approaches in a hybrid model leads to more accurate recommendations with reduced error.
Figures 5A-5C illustrate the graph depicting the RMSE metrics for the SVD++ method in accordance with an embodiment of the present disclosure. Figure 5A shows the examination of the RMSE metrics for the SVD++ method across k=10 latent features and 25 communities for movie lens-100K. Figure 5A shows the examination of the RMSE metrics for the SVD++ method across k=10 latent features and 25 communities for movie lens-1M. Figure 5A shows the examination of the RMSE metrics for the SVD++ method across k=10 latent features and 25 communities for anime recommendation datasets. The datasets were evaluated using various convex combinations of content-based and collaborative filtering methods. In the MovieLens-100K dataset, the lowest RMSE value was achieved at community 8. As the convex combination for content-based filtering increased, the RMSE also increased, whereas increasing the weight for collaborative filtering resulted in a lower RMSE. A similar trend was observed in the MovieLens-1M dataset, where the lowest RMSE occurred at community 13. The figure shows that raising the content-based weight leads to higher RMSE values, while increasing the collaborative filtering weight consistently reduces RMSE. In the anime recommendation dataset, the lowest RMSE was recorded at community 2. Notably, as the number of communities increased, the RMSE also tended to rise. Additionally, when the content-based convex combination was increased, RMSE values grew, while higher collaborative filtering combinations reduced them.
These findings highlight the effectiveness of using a hybrid approach that combines content-based and collaborative filtering. The results clearly show that the right balance between these techniques varies across datasets and adjusting the convex combination for collaborative filtering can significantly lower RMSE. Furthermore, the hybrid approach helps pinpoint the optimal community where users can receive the most accurate predictions. This reinforces the value of using hybrid models for personalized recommendations, especially when dealing with large-scale datasets, where the integration of both filtering techniques provides greater accuracy and efficiency.
In an operative configuration, the recommendation system 100 functions by sequentially utilizing a series of specialized modules to process user-item interaction data, thereby generating personalized and accurate item recommendations. The process begins with the bipartite graph construction module 102, which receives a rating matrix R containing user-item interactions and constructs a bipartite graph (BG) where user and item nodes are connected by weighted edges based on user ratings. Next, the community detection module 104 applies the Louvain community detection algorithm to partition the graph into smaller, distinct communities, denoted as BG1, BG2,…, BGc, with each community represented by a community-specific rating matrix R1, R2,…,Rc. This segmentation enables the system to process more focused groups of users with similar preferences, enhancing the relevance of recommendations.
Following community partitioning, the system employs a content-based filtering module 106 to calculate cosine similarity within each community-specific matrix. By using genre information and other item attributes, this module generates content-based predicted matrices CBR1, CBR2,…, CBRc, which capture item similarities relevant to each community's preferences. Concurrently, the collaborative filtering module 108 applies matrix factorization techniques, such as SVD++ or basic matrix factorization, to each community-specific rating matrix, producing collaborative-based filtering metrics CFR1, CFR2,…,CFRc that reveal latent patterns in user-item interactions. These two sets of metrics are then combined in the fusion module 110, which performs a convex combination to balance content-based and collaborative filtering predictions based on a weighting factor WWW. This fusion yields enhanced predicted matrices (R̃1, R̃2, ..., R̃c) for each community, integrating both explicit item attributes and latent user preferences.
The aggregation module 112 then consolidates these enhanced predicted matrices from each community into a final predicted rating matrix R~R̃R~, which serves as the system's comprehensive recommendation output. This matrix reflects both community-specific and individual-level insights, ensuring that recommendations are highly relevant to users' preferences. Optionally, the evaluation module 114 can calculate the Root Mean Square Error (RMSE) between the original and predicted rating matrices to assess recommendation accuracy. Additionally, a user feedback module 116 gathers real-time feedback on recommended items, integrating this data back into the filtering modules to refine future recommendations based on evolving user preferences. Overall, the system 100 ensures optimized accuracy, adaptability, and scalability, making the recommendation system capable of delivering highly personalized and dynamic recommendations tailored to individual user interests.
Advantageously, the system 100 utilizes a sophisticated hybrid approach that combines community detection, content-based filtering, and collaborative filtering techniques to enhance personalization and accuracy. By employing a bipartite graph structure and the Louvain community detection model, the system identifies natural clusters of users with similar preferences, enabling tailored recommendations that are more relevant and precise. The fusion of content-based and collaborative filtering metrics through a convex combination allows the system to balance explicit item attributes with latent user preferences, leveraging the strengths of both techniques for a well-rounded recommendation. Additionally, the system's modular design, with components like dynamic community adjustment and real-time user feedback integration, ensures adaptability to evolving user interests. This flexibility is further supported by an evaluation module that quantifies accuracy, allowing continuous refinement of recommendations. By addressing data sparsity through implicit feedback and dynamically adjusting fusion weights, the system can cater to both popular and niche items, resulting in high-quality, scalable recommendations with optimal computational efficiency.
The functions described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Other examples and implementations are within the scope and spirit of the disclosure and appended claims. For example, due to the nature of software, functions described above can be implemented using software executed by a processor, hardware, firmware, hardwiring, or combinations of any of these. Features implementing functions may also be physically located at various positions, including being distributed such that portions of functions are implemented at different physical locations.
The foregoing description of the embodiments has been provided for purposes of illustration and is not intended to limit the scope of the present disclosure. Individual components of a particular embodiment are generally not limited to that particular embodiment, but are interchangeable. Such variations are not to be regarded as a departure from the present disclosure, and all such modifications are considered to be within the scope of the present disclosure.
TECHNICAL ADVANCEMENTS
The present disclosure described herein above has several technical advantages including, but not limited to, the realization of a recommendation system for generating personalized item recommendations to users and a method thereof that:
• ensures more accurate, contextually relevant recommendations for each user by grouping them with others who have similar preferences;
• improves the accuracy of personalized item recommendations;
• optimize computational efficiency and scalability;
• provide enhanced recommendation quality via hybrid filtering;
• scalable and robust recommendation outputs;
• increased prediction accuracy and reduced computational complexity;
• provide an efficient fusion of filtering metrics for balanced predictions;
• balance content-based and collaborative filtering strengths;
• dynamic adaptation to user preferences;
• allow dynamic adaptation of recommendations; and
• generate a final aggregated rating matrix for comprehensive recommendations.
The embodiments herein and the various features and advantageous details thereof are explained with reference to the non-limiting embodiments in the following description. Descriptions of well-known components and processing techniques are omitted so as to not unnecessarily obscure the embodiments herein. The examples used herein are intended merely to facilitate an understanding of ways in which the embodiments herein may be practiced and to further enable those of skill in the art to practice the embodiments herein. Accordingly, the examples should not be construed as limiting the scope of the embodiments herein.
The foregoing description of the specific embodiments so fully reveals the general nature of the embodiments herein that others can, by applying current knowledge, readily modify and/or adapt for various applications such specific embodiments without departing from the generic concept, and, therefore, such adaptations and modifications should and are intended to be comprehended within the meaning and range of equivalents of the disclosed embodiments. It is to be understood that the phraseology or terminology employed herein is for the purpose of description and not of limitation. Therefore, while the embodiments herein have been described in terms of preferred embodiments, those skilled in the art will recognize that the embodiments herein can be practiced with modification within the spirit and scope of the embodiments as described herein.
The use of the expression "at least" or "at least one" suggests the use of one or more elements or ingredients or quantities, as the use may be in the embodiment of the disclosure to achieve one or more of the desired objects or results.
While considerable emphasis has been placed herein on the components and component parts of the preferred embodiments, it will be appreciated that many embodiments can be made and that many changes can be made in the preferred embodiments without departing from the principles of the disclosure. These and other changes in the preferred embodiment as well as other embodiments of the disclosure will be apparent to those skilled in the art from the disclosure herein, whereby it is to be distinctly understood that the foregoing descriptive matter is to be interpreted merely as illustrative of the disclosure and not as a limitation. , Claims:WE CLAIM:
1. A recommendation system (100) for generating personalized item recommendations to users, said system (100) comprising:
• a bipartite graph construction module (102) configured to receive a rating matrix (R) containing user-item interactions to construct a bipartite graph, represent users and items as nodes in the bipartite graph, and connect the user and item nodes with edges based on available ratings;
• a community detection module (104) configured to apply a Louvain community detection model (104a) to partition said bipartite graph into smaller, distinct communities (BG1, BG2, ..., BGc), each represented by a community-specific rating matrix (R1, R2, ..., Rc);
• a content-based filtering module (106) configured to apply a cosine similarity within each community-specific rating matrix (R1, R2, ..., Rc) to generate content-based filtering metrics within each community;
• a collaborative filtering module (108) configured to apply matrix factorization techniques, including singular value decomposition (SVD++) or basic matrix factorization, to each community-specific rating matrix (R1, R2, ..., Rc), to generate collaborative-based filtering metrics;
• a fusion module (110) configured to combine said content-based and said collaborative-based filtering metrics for each community through a convex combination thereby producing enhanced predicted rating matrices (R̃1, R̃2, ..., R̃c) that balance said item attribute relevance with latent user preferences, leveraging the strengths of both recommendation techniques for optimal prediction accuracy; and
• an aggregation module (112) configured to aggregate said enhanced predicted matrices from all communities into a final predicted rating matrix (R̃), representing the system's recommendation outputs.
2. The recommendation system (100) as claimed in claim 1, wherein said system (100) further comprises an evaluation module (114) configured to quantify recommendation accuracy by calculating the root mean square error (RMSE) between the original rating matrix (R) and the final predicted matrix (R̃), thereby providing an indication of the recommendation system's accuracy.
3. The recommendation system (100) as claimed in claim 1, wherein said system (100) further comprises a user feedback module (116) configured to gather real-time user feedback on recommended items, and feed this feedback back into the content-based and collaborative filtering modules, thereby enhancing future recommendations based on actual user interactions.
4. The recommendation system (100) as claimed in claim 1, wherein said community detection module (104) further comprises a graph construction sub-module (104b) configured to construct the bipartite graph from user-item interactions by assigning edge weights based on the ratings given by users to items.
5. The recommendation system (100) as claimed in claim 1, wherein said community detection model (104a) is configured to dynamically adjust the number of communities (BG1, BG2, ..., BGc) based on the density and clustering of the user-item interaction graph, to ensure optimal partitioning of the data for each specific dataset.
6. The recommendation system (100) as claimed in claim 1, wherein said content-based filtering module (106) is further configured to use genre information, item descriptions, and other item-specific attributes as input features to calculate the cosine similarity within each community, thereby generating content-based prediction matrices (CBR1, CBR2, ..., CBRc) tailored to specific user preferences within each community.
7. The recommendation system (100) as claimed in claim 1, wherein said collaborative filtering module (108) applies matrix factorization techniques selected from the group consisting of basic matrix factorization (MF), singular value decomposition (SVD), and SVD++, to each community-specific rating matrix, thereby decomposing user-item interactions into latent factors with minimal computational complexity.
8. The recommendation system (100) as claimed in claim 1, wherein said collaborative filtering module (106) is further configured to address data sparsity by incorporating implicit feedback data, such as user clicks or browsing history, into the matrix factorization model, enhancing the prediction accuracy for items with sparse ratings.
9. The recommendation system (100) as claimed in claim 1, wherein said fusion module (110) is configured to adjust the weight of the convex combination dynamically based on the density of the community-specific rating matrices, the content similarity between items, and the strength of the latent factors, to achieve a balance between content-based and collaborative filtering predictions for each community.
10. A recommendation method (200) for generating personalized item recommendations to users, said method (200) comprises the following steps:
• receiving, by a bipartite graph construction module (102), a rating matrix (R) containing user-item interactions to construct a bipartite graph, represent users and items as nodes in the bipartite graph, and connecting the user and item nodes with edges based on available ratings;
• applying, by a community detection module (104), a Louvain community detection model (104a) to partition said bipartite graph into smaller, distinct communities (BG1, BG2, ..., BGc), each represented by a community-specific rating matrix (R1, R2, ..., Rc);
• applying, a content-based filtering module (106), a cosine similarity within each community-specific rating matrix (R1, R2, ..., Rc) to generate content-based filtering metrics within each community;
• applying, a collaborative filtering module (108), matrix factorization techniques, including singular value decomposition (SVD++) or basic matrix factorization, to each community-specific rating matrix (R1, R2, ..., Rc), to generate collaborative-based filtering metrics;
• combining, by a fusion module (110), said content-based and said collaborative-based filtering metrics for each community through a convex combination thereby producing enhanced predicted rating matrices (R̃1, R̃2, ..., R̃c) that balance said item attribute relevance with latent user preferences, leveraging the strengths of both recommendation techniques for optimal prediction accuracy; and
• aggregating, by an aggregation module (112), said enhanced predicted matrices from all communities into a final predicted rating matrix (R̃), representing the system's recommendation outputs.

Dated this 14th Day of November, 2024

_______________________________
MOHAN RAJKUMAR DEWAN, IN/PA - 25
OF R. K. DEWAN & CO.
AUTHORIZED AGENT OF APPLICANT

TO,
THE CONTROLLER OF PATENTS
THE PATENT OFFICE, AT CHENNAI

Documents