System and method for providing personalized recommendations to user queries

Abstract

The present invention relates to a system and method for providing personalized recommendations to user queries. The method includes determining an intent and a context associated with a user query based on an associated user profile. Further, the method includes dynamically determining user-specific privacy attributes associated with the user profile. Thereafter, the method discloses retrieving a first set of recommendations mapping with the intent and the context and the one or more user-specific privacy attributes from a personalized database and filtering the first set of recommendations based on a bias factor by analysing one or more user interaction metrics. Finally, the method discloses extracting a second set of recommendations semantically matching with the filtered first set of recommendations using a multi-modal data processing engine for providing to a user.

Description

TECHNICAL FIELD

[0001] The present disclosure generally relates to the field of natural language processing and data processing. Particularly, but not exclusively, the present disclosure relates to a personalized recommendations assistant system and a method for providing personalized recommendations to user queries.BACKGROUND

[0002] With the widespread adoption of mobile computing devices, consumers are increasingly engaging in online shopping through e-commerce platforms rather than traditional offline stores. The e-commerce platforms have evolved to support mobile-friendly interfaces across varying screen sizes and input modes. However, existing e-commerce systems still experience significant technical limitations that affect the quality of user experience and the relevance of recommendations. Existing search engines embedded within the existing e- commerce platforms are often based on conventional keyword-matching or rule-based architectures. The existing e-commerce platforms tend to produce superficial matches that disregard user intent, leading to irrelevant results and degraded customer satisfaction. The superficial matches include false or misleading information presented as fact also referred as artificial intelligence (Al) hallucinations.

[0003] In many scenarios, the existing e-commerce platforms generate Al hallucinated results that appear contextually disconnected or factually misleading with respect to the user's query. The Al hallucinations arise due to low and / or insufficient semantic understanding, lack of bias correction, and absence of real-time contextual analysis. Furthermore, the search results may be typically generic and do not reflect user-specific preferences. Moreover, the existing e- commerce platforms lack intent and context of the user behind the search query and results in performing repeated searches or manual filtration through large volumes of data / content.

[0004] Furthermore, the existing e-commerce platforms utilize user’s private data without user’s consent. For example, the user do not have the ability to limit access to their purchase history, location, other personal information, etc, while performing search on such platforms. The absence of user-governed data filters not only undermines trust of the user but also contributes to misaligned suggestions. The misaligned suggestions arise from the lack of clarity on which data streams users permit for better results.

[0005] The information disclosed in this background of the disclosure section is only for the enhancement of understanding of the general background of the invention and should not betaken as an acknowledgement or any form of suggestion that this information forms the prior art already known to a person skilled in the art.SUMMARY

[0006] The present disclosure relates to a method for providing personalized recommendations to user queries. The method includes determining an intent and a context associated with the user query based on an associated user profile. Further, the method includes dynamically determining one or more user-specific privacy attributes associated with the user profile, enabled in real time. Further, the method includes retrieving, from a personalized database, a first set of recommendations mapping with the intent and the context and the one or more userspecific privacy attributes. Further, filtering the first set of recommendations based on a bias factor by analysing one or more user interaction metrics. Further, the method includes extracting, using a multi-modal data processing engine, a second set of recommendations by semantically matching with the filtered first set of recommendations for providing to a user.

[0007] In an embodiment, the method further includes accessing information from an external server platform. The information includes at least one of user profiles, listed products and associated descriptions, and listed services and associated descriptions. The method further includes identifying one or more attributes based on metadata extracted from the accessed information. The one or more attributes represent features of the listed products, the listed services, or the user profiles. The method further includes generating a plurality of semantic vectors by associating a plurality of semantic labels with the one or more identified attributes. The method further includes creating the personalized database by storing the plurality of semantic vectors and associated plurality of semantic labels and the one or more identified attributes.

[0008] In an alternate embodiment, prior to accessing the information from the external server platform, the method discloses analysing the information associated with the external server platform. The method further discloses determining an anomaly in the analysed information and rectifying the analysed information based on the determined anomaly.

[0009] In an alternate embodiment, for dynamically determining the one or more user-specific privacy attributes, the method comprises: receiving dynamic selection of one or more predefined user-specific privacy attributes via a user interface and enabling the selected userspecific privacy attributes.

[0010] In an alternate embodiment, the one or more user-specific privacy attributes includes at least one of purchase behavior, historical transaction, demographic details, product metadata, and real-time contextual information created by the multi-modal data processing engine based on the one or more user interactions.

[0011] In an alternate embodiment, for filtering the first set of recommendations, the method discloses: dynamically detecting patterns, indicative of an interaction weightage assigned to a product, a category, and / or a vendor, based on the one or more user interaction metrics, and applying filters to the detected patterns based on the determined intent and the bias factor to derive the first set of recommendations. The one or more user interaction metrics include at least one of user-specific click-through rates, past engagement, external rankings, and third- party interactions. Further, the bias factor being an indicative of a correlation between the determined intent and the one or more user interaction metrics.

[0012] In an alternate embodiment, for extracting the second set of recommendations using the multi-modal data processing engine, the method discloses transforming the filtered first set of recommendations into one or more vectors using natural language processing (NLP) and retrieving a plurality of semantic vectors from the plurality of generated semantic vectors mapping with the intent and the context from the personalized database. The method further discloses comparing the one or more vectors against each of the plurality of semantic vectors. Finally, the method discloses extracting the second set of recommendations based on the comparison and ranking each of the second set of recommendations based on the user profile.

[0013] In an alternate embodiment, for ranking each of the second set of recommendations, the method discloses assigning a relevancy score to each of the second set of recommendations based on the one or more user-specific privacy attributes. Further, the method discloses computing a confidence score for each of the second set of recommendations by aggregating predetermined weights of the second set of recommendations and the relevancy score and ranking each of the second set of recommendations based on the confidence score.

[0014] The present disclosure also relates to a system for providing personalized recommendations to a user query. The system includes one or more processors and a memory that is communicatively coupled to the one or more processors. The one or more processors is configured to determine an intent and a context associated with the user query based on an associated user profile, enabled in real-time. Further, the one or more processors is configured to dynamically determine one or more user-specific privacy attributes, associated with the userprofile, enabled in real time. Further, the one or more processors is configured to retrieve a first set of recommendations, from a personalized database, mapping with the intent and the context and one or more user-specific privacy attributes. Further, the one or more processors is configured to filter the first set of recommendations based on a bias factor by analysing one or more user interaction metrics. Further, the processor is configured to extract, using a multimodal data processing, a second set of recommendations by semantically matching with the filtered first set of recommendations engine for providing to a user. In an embodiment, the one or more processors is further configured to access information associated with an external server platform. The information includes at least one of user profiles, listed products and associated descriptions, and listed services and associated descriptions. The one or more processors is further configured to identify one or more attributes based on metadata extracted from the accessed information. The one or more attributes represent features of the listed products, the listed services, or the user profiles. The one or more processors is further configured to generate a plurality of sematic vectors by associating a plurality of semantic labels with the one or more identified attributes. The one or more processors is further configured to create the personalized database by storing the plurality of semantic vectors and associated plurality of semantic labels and the one or more identified attributes.

[0015] In an embodiment, prior to access of the information from the external server platform, the one or more processors is configured to analyse the information associated with the platform. The one or more processors is further configured to determine an anomaly in the analysed information and rectify the analysed information based on the determined anomaly.

[0016] In an embodiment, to dynamically determine the one or more user-specific privacy attributes, the one or more processors is configured to receive dynamic selection of one or more pre-defined user-specific privacy attributes via a user interface and enable the selected userspecific privacy attributes.

[0017] In an embodiment, the one or more user-specific privacy attributes includes at least one of purchase behavior, historical transaction, demographic details, product metadata, and realtime contextual information created by the multi-modal data processing engine based on the one or more user interactions.

[0018] In an embodiment, to filter the first set of recommendations the one or more processors is further configured to dynamically detect patterns, indicative of an interaction weightage assigned to a product, a category, a vendor, based on the one or more user interaction metrics.The one or more user interaction metrics includes at least one of user-specific click-through rates, past engagement, external rankings, and third-party interactions. The one or more processors is further configured to apply filters to the detected patterns based on the determined intent and the bias factor. The bias factor being an indicative of a correlation between the determined intent and the one or more user interaction metrics.

[0019] In an embodiment, to extract the second set of recommendations, using the multi-modal data processing engine, the one or more processors is further configured to transform the filtered first set of recommendations into one or more vectors using natural language processing (NLP). The one or more processors is further configured to retrieve the plurality of semantic vectors, from the plurality of generated semantic vectors, mapping with the intent and the context from the personalized database. The one or more processors is further configured to compare the one or more vectors against each of the plurality of semantic vectors. Further, the one or more processors is further configured to extract the second set of recommendations based on the comparison and rank each of the second set of recommendations based on the user profile.

[0020] In an embodiment, the one or more processors is further configured to assign a relevancy score to each of the second set of recommendations based on the one or more userspecific privacy attributes. The one or more processors is further configured to compute a confidence score for each of the second set of recommendations by aggregating predetermined weights of the second set of recommendations and the assigned relevancy score. The one or more processors is further configured to rank each of the second set of recommendations based on the confidence score.

[0021] The foregoing summary is illustrative only and is not intended to be in any way limiting. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features will become apparent by reference to the drawings and the following detailed description. For a better understanding of exemplary embodiments of the present disclosure, together with other and further features and advantages thereof, reference is made to the following description, taken in conjunction with the accompanying drawings.BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The embodiments of the disclosure itself, as well as a preferred mode of use, further objectives, and advantages thereof, will best be understood by reference to the following detailed description of an illustrative embodiment when read in conjunction with theaccompanying drawings. One or more embodiments are now described, by way of example only, with reference to the accompanying drawings in which:

[0023] Figure 1 illustrates an environment for providing personalized recommendations to user queries, in accordance with some embodiments of the present disclosure;

[0024] Figure 2 illustrates a block diagram of the personalized recommendations assistant system of Figure 1, for providing personalized recommendations to user queries, in accordance with some embodiments of the present disclosure;

[0025] Figure 3 illustrates a block diagram of the multi-modal data processing engine of Figure 1, in accordance with some embodiments of the present disclosure;

[0026] Figure 4 illustrates a block diagram of a user-specific privacy attributes module, in accordance with some embodiments of the present disclosure;

[0027] Figure 5 illustrates a block diagram of a filtering module of Figure 2, in accordance with some embodiments of the present disclosure;

[0028] Figure 6 illustrates a block diagram of the second set of recommendations module of Figure 2, in accordance with some embodiments of the present disclosure; and

[0029] Figure 7 illustrates a flow diagram of a method for providing personalized recommendations to user queries, in accordance with some embodiments of the present disclosure.

[0030] The figures depict embodiments of the disclosure for purposes of illustration only. A person skilled in the art will readily recognize from the following description that alternative embodiments of the structures and methods illustrated herein may be employed without departing from the principles of the disclosure described herein.DETAILED DESCRIPTION OF THE DISCLOSURE

[0031] In the present document, the word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any embodiment or implementation of the present subject matter described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments.

[0032] While the disclosure is susceptible to various modifications and alternative forms, specific embodiment thereof has been shown by way of example in the drawings and will be described in detail below. It should be understood, however that it is not intended to limit thedisclosure to the particular forms disclosed, but on the contrary, the disclosure is to cover all modifications, equivalents, and alternative falling within the spirit and the scope of the disclosure.

[0033] The terms “comprises”, “comprising”, or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a setup, device or method that comprises a list of components or steps does not include only those components or steps but may include other components or steps not expressly listed or inherent to such setup or device or method. In other words, one or more elements in a device or system or apparatus proceeded by “comprises. . . a” does not, without more constraints, preclude the existence of other elements or additional elements in the device, system, or apparatus.

[0034] In the following detailed description of the embodiments of the disclosure, reference is made to the accompanying drawings that form a part hereof, and in which are shown by way of illustration specific embodiments in which the disclosure may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the disclosure, and it is to be understood that other embodiments may be utilized and that changes may be made without departing from the scope of the present disclosure. The following description is, therefore, not to be taken in a limiting sense.

[0035] The present disclosure relates to a system and method for providing personalized recommendations to a user query using natural language processing (NLP), context analysis, and bias-aware filtering. With the growing volume of e-commerce and content platforms, users are often presented with an overwhelming number of product, services, or content options. Traditional search and recommendation engines heavily rely on keyword matching or historical popularity, which is prone to introduce biasness in output and can lead to irrelevant results and overlook user-specific needs.

[0036] To address these limitations, the disclosed system utilizes an intelligent recommendation engine that combines user intent detection, contextual understanding, to remove artificial intelligence (Al) hallucination, and mitigate real-time bias from search results. For example, a user searching for "affordable smartwatches with sleep tracking" may expect results tailored to their budget and feature preference, rather than generic high-ranking products driven by search engine optimization (SEO). The disclosed system captures such expectations by identifying relevant entities and features from the user query, enriching the context related to the user query, using prior behavior and demographic data, and filtering the results toeliminate influence from externally biased metrics like third-party rankings or promotional exposure. The disclosed architecture integrates various modules for privacy-aware feature refinement, semantic vector matching, Extract, Transform, And Load (ETL) data processing, and multi-layered re-ranking based on personalized scoring metrics. As a result, users receive highly relevant, fair, and explainable recommendations aligned with corresponding individual intent and context free from bias and Al hallucination.

[0037] The present disclosure provides elimination of Al hallucinations from the extracted results, provides improvement in data privacy, allows faster function calling accuracy, and provides high scalability with non-reliance on external databases. The present disclosure may achieve improved recommendation precision, resulting, for example, in more relevant, diverse, and context-aware suggestions. The present disclosure reduces reliance on generic searchbased models to decrease the need for manual curation or post-processing, making the system suitable for real-time integration in personalized digital assistants, e-commerce platforms, and content delivery services, but not limited thereto.

[0038] Figure 1 illustrates an environment 100 for providing personalized recommendations to user queries, in accordance with some embodiments of the present disclosure.

[0039] As shown in Figure 1, the environment 100 includes one or more data sources 102A- N communicatively coupled to a personalized recommendations assistant system (PRAS) 104 and an external server platform 108 via a communication network 112 (interchangeably referred to as network 112). In an example, the data sources 102A-N may include, but not limited to, external platforms, third-party tools, user interaction databases, or metadata repositories that provide product descriptions, service information, user profiles, and behavior logs. Moving on, the external server platform 108 includes a user interface 110. The user interface 110 enables users to input queries, manage user profiles, and optionally select privacy settings. In an embodiment, the external server platform 108 may be implemented on a computing device. Examples of the computing device includes, but are not limited to, a cellular phone or smartphone, a pager, a laptop computer, a desktop computer, a wireless handset, a portable communication device, a portable computing device (e.g., a personal data assistant), or any other suitable computing device or other equipment / sensors including a wired or wireless communications interface. In one example, the communication network 112 may be a wireless network or a wired network. The communication network 112 is a combination of a wired network and a wireless network. In some embodiments, the communication network 112 is a direct interconnection, Local Area Network (LAN), Wide Area Network (WAN),Controller Area Network (CAN). The wireless network may be Bluetooth Low Energy (BLE), Near Field Communication (NFC), Bluetooth, Wi-Fi, the Internet, and the like.

[0040] Further, as shown in Figure 1, the PRAS 104 includes a multi-modal data processing engine 106. The PRAS 104 is configured to provide personalized recommendations to user queries, as disclosed in the present disclosure. In some embodiments, the PRAS 104 may include a processor (not shown). The PRAS 104 may include an application that may be installed in the computing device or any mobile device to provide personalized recommendations to the user queries. In an exemplary embodiment, the external server platform 108 is installed via a virtual assistant (not shown) that may be stored in a memory (not shown) which is communicatively coupled with the PRAS 104. The virtual assistant may be Artificial-Intelligent (Al)-enabled. In an example, the external server platform 108 may be an application or a browser implemented external server platform 108 to host an online store to sell one or more products and / or one or more services. Without any limitation, the external server platform 108 may be any similar e-commerce platform. Those skilled in the art will appreciate, t

[0041] The PRAS 104 may be deployed onto the external server platform 108 to enhance the customer experience of performing online shopping. In a non -limiting example, the PRAS 104 may be configured to seamlessly integrate or deploy with any existing e-commerce platforms. The PRAS 104 may further provide the users, such as platform operators or customers, to have complete control over their data by providing options to customize their shopping preferences and may implement data security protocols to protect personal information to enhance trust and user engagement. Examples of the data security protocols data encryption protocols may include, but not limited to, advanced encryption standard (AES) -256 encryption technique for data storage, secure socket layer (SSL) / transport layer security (TLS) for data transfer, etc.

[0042] In some embodiments, the user interface 110 may include a graphical interface rendered on a web browser, a native mobile application interface, or a voice-enabled assistant interface. For example, the user interface 110 may be installed as a mobile application allowing users to input queries, adjust privacy preferences through toggles or sliders, and view personalized recommendations in a card or list format. In another example, the user interface 110 may be voice-driven, where the user speaks the query and receives auditory responses or visual suggestions on a smart display. The user interface 110 may further enable selection or customization of recommendation categories, privacy attribute toggling, or feedback provision.

[0043] Further, the multi-modal data processing engine 106 of Figure 1 is further configured to access information associated with the external server platform 108. The information includes user profiles, listed products and associated descriptions, and listed services and associated descriptions. Precisely, the multi-modal data processing engine 106 is further configured to identify one or more attributes based on metadata extracted from the accessed information. In an exemplary aspect, the one or more attributes represent features of the listed products, the listed services, or the user profiles. The metadata may be extracted using data extraction methods. Examples of the data extraction methods include, but not limited to, named entity recognition (NER), rule-based parsing, dependency parsing, regular expression extraction, etc.

[0044] Further, the multi-modal data processing engine 106 is further configured to generate a plurality of semantic vectors by associating a plurality of semantic labels with the one or more extracted attributes. The plurality of semantic vectors is generated by mapping the plurality of semantic labels using natural language processing (NLP) techniques. Examples of NLP techniques may include, but not limited to, word to vector (Word2Vec), bidirectional encoder representations (BERT), etc. Moving on, the multi-modal data processing engine 106 is further configured to create the personalized database by storing the one or more semantic vectors and associated plurality of semantic labels and the one or more identified attributes.

[0045] In another exemplary aspect, the multi-modal data processing engine 106 is further configured to analyse the information associated with the external server platform 108 before collecting the information associated with the external server platform 108. The multi-modal data processing engine 106 is further configured to determine an anomaly in the analysed information. The multi-modal data processing engine 106 is further configured to rectify the analysed information based on the determined anomaly. For example, if there is a product, such as a toy bike is registered on the external server platform 108 which is an automobile spare parts-related online store, the PRAS 104 determines that the toy bike is not relevant to the spare part category for the vehicle. Accordingly, the PRAS 104 may identify the mismatch of the category as an anomaly and suitably inform the operator regarding the identified anomaly. The PRAS 104 may aggregate data from various data sources 102A-N beyond basic product descriptions, which may include, without any limitation, user reviews, social media trends, purchase history, and individual browsing behaviours. The PRAS 104 may allow to offer dynamic product comparisons that are highly relevant and tailored to the user's specific needs and preferences. By doing so, the multi-modal data processing engine 106 allows any type ofanomaly associated with the external server platform to be rectified even before the personalized database is created. Thus, ensuring that only relevant information is passed to the personalized database.

[0046] The PRAS 104 is further configured to determine an intent and a context associated with the user query by processing the user query and an associated user profile. Without any limitation, the PRAS 104 may receive the user query via text or voice. In another non-limiting embodiment, the PRAS 104 may request for additional information related to the user query to suggest an appropriate product or service. For example, the user query may be segmented into one or more entities using entity recognition. The one or more entities may be further analyzed using semantic analysis to determine the intent of the user query. In an exemplary aspect, the context of the user query may be determined using long short-term memory (LSTM) networks and a context history layer.

[0047] Further, the PRAS 104 is configured to identify one or more user-specific privacy attributes associated with the user profile. Examples of the one or more user-specific privacy attributes include, but are not limited to, at least one of purchase behavior, historical transaction, demographic details, product metadata, and real-time contextual information. In an embodiment, the PRAS 104 is configured to allow selection of one or more pre-defined user-specific privacy attributes via the user interface 110. The one or more user-specific privacy attributes includes purchase behavior, historical transaction, demographic details, product metadata, and real-time contextual information created by the multi-modal data processing engine 106 based on the one or more user interactions. For example, the user may choose granular profile settings, or toggle categories or may enable and disable specific data streams. The PRAS 104 may be configured to identify and enable the selected user-specific privacy attributes.

[0048] Further, the PRAS 104 is configured to process the user query based on the intent, the context, and the one or more user-specific privacy attributes to determine a first set of recommendations that matches the intent and the context and aligns with the one or more userspecific privacy attributes.

[0049] Additionally, the PRAS 104 is configured to filter the first set of recommendations based on a bias factor by analysing one or more user interaction metrics to remove any type of bias from the filtered first set of recommendations. The one or more user interaction metrics includes user-specific click-through rates, past engagement, external rankings, and third-partyinteractions. The PRAS 104 may be configured to understand the user interaction metrics using natural language processing (NLP) techniques. Further, the PRAS 104 is configured to dynamically detect patterns in user behavior using clustering. The detected patterns is an indicative of an interaction weightage assigned to a product, a category, a vendor, based on the one or more user interaction metrics. The PRAS 104 is further configured to apply filters to the detected patterns based on the determined intent and the bias factor. Those skilled in the art, will appreciate that the bias factor is indicative of a correlation between the determined intent and the one or more user interaction metrics.

[0050] Finally, the PRAS 104 is configured to extract a second set of recommendations by applying semantic matching to the filtered first set of recommendations using a multi-modal data processing engine 106. Examples of semantic matching may include, but not limited to, conceptual mapping, embedding based mapping, vector similarity search, ontology-driven matching, neural semantic retrieval, etc.

[0051] In an exemplary aspect, the PRAS 104 may be configured to transform the filtered first set of recommendations into one or more vectors using natural language processing (NLP) techniques. The NLP techniques may include, but not limited to, tokenization, embedding generation, or transformer models, for example, term frequency-inverse document frequency (TF-IDF), bidirectional encoder representations (BERT), etc, for vector generation. Alternatively in an embodiment, if the semantic matching between the filtered first set of recommendations and the intended context of the query falls below a predefined threshold, the processor may be configured to generate a fallback response. The fallback response may include prompting the user to refine the query or displaying a notification indicating the unavailability of meaningful matches. The failed or semantically mismatched queries may optionally be logged into a training dataset and used in the future to enhance the vector mapping logic or retrain classification algorithms in the multi-modal data processing engine 106.

[0052] Further, in a non-limiting example, if the external server platform 108 caters user needs regarding computer accessories, the PRAS 104 may receive a query, from the customer, relating to a data transfer cable for connecting a recently purchased laptop with a smartphone. The processor may analyse the search query to determine the intent of the customer, that, as both the devices are recently purchased, the devices may be compatible with the state-of-the- art communication standards. For example, the PRAS 104 may determine that the subject devices may support a specific type of connector. Accordingly, the PRAS 104 may generate search results for the user which includes communication cables supporting the specific typeof connector, using the personalized database, which would be free from Al hallucinations. In yet another non-limiting example, the PRAS 104 may suggest a budget-based products, based on a pervious search session of the customer. Accordingly, the PRAS 104 may analyse the budget preferences of the customer for dynamic product comparison. Thereby, the PRAS 104 improves the accuracy of the search results and enhances the customer satisfaction and efficiency of store operation. The present disclosure may also facilitate e-commerce platforms to improve user engagement, overall customer satisfaction without disrupting existing business operations, real-time store management integration, and enhanced privacy controls. If no strong semantic match is found, the PRAS 104 suggests refining the user query.

[0053] Figure 2 illustrates a detailed block diagram of the personalized recommendations assistant system (PRAS) 104 of Figure 1, for providing personalized recommendations to user queries, in accordance with some embodiments of the present disclosure. FIG. 2 is explained in conjunction with the PRAS 104 of FIG. 1. With reference to FIG. 2, there is shown a block representation of the PRAS 104.

[0054] In an embodiment, the PRAS 104 comprises a Central Processing Units 200 (also referred as “CPUs” or “one or more processors 200”), a memory 202, and Input / Output (VO) interface 204 (also referred as “user interfaces”).

[0055] In an embodiment, the memory 202 may include data 206. Further, as shown in Figure 2, PRAS 104 may include one or more modules 216-224 . The one or more modules 216-224 may be configured to perform the steps of the present disclosure using the data 206, for providing personalized recommendations to user queries. In an embodiment, each of the one or more modules 216-224 may be a hardware unit or a combination of hardware and software components, which may be configured external to the memory 202 and communicatively coupled with the processor 200. As used herein, the term modules refers to processors, application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), digital signal processors (DSPs) and / or other suitable components that provide described functionality. The one or more modules 216-224, when configured with the described functionality defined in the present disclosure, collectively constitute a recommendation system operable to generate personalized, context-aware, and bias-filtered recommendations based on user interactions.

[0056] In one implementation, the data 206 may include, for example, user query 208, information 210, user profiles 212, and user interaction metrics 214. In one implementation,the one or more modules 216-224 may include, for example, an intent and context determining module 216, a privacy attributes identifying module 218, a user query processing module 226, a bias filtering module 222 and a second set of recommendations module 224. It will be appreciated that such modules may be represented as a single module or a combination of different modules.

[0057] In an embodiment, the intent and context determining module 216 may be configured to determine an intent and a context associated with the user query 208 by processing the user query and an associated user profile 212. For example, the user query may be segmented into one or more entities using entity recognition. The one or more entities may be further analyzed using semantic analysis to determine the intent of the user query. The context of the user query may be determined using long short-term memory (LSTM) networks and a context history layer.

[0058] Further, in an embodiment, the privacy attributes identifying module 218 may be configured to identify one or more user-specific privacy attributes associated with the user profile. In an example, the processor 200 may be configured to allow selection of one or more pre-defined user-specific privacy attributes via a user interface and enable the selected userspecific privacy attributes during the search. The one or more user-specific privacy attributes includes purchase behavior, historical transaction, demographic details, product metadata, and real-time contextual information created by the multi-modal data processing engine 106 based on the one or more user interactions.

[0059] The user query processing module 226 comprising first set of recommendations module 220 that may be configured to process the user query based on the intent, the context, and the one or more user-specific privacy attributes to determine a first set of recommendations that matches with the intent and the context and aligns with the one or more user-specific privacy attributes. Moving on, the user query processing module 226 may include the bias filtering module 222 that may be configured to filter the first set of recommendations based on a bias factor by analyzing one or more user interaction metrics. The user interaction metrics may be analyzed using natural language processing (NLP) techniques and clustering methods.

[0060] Further, a second set of recommendations module 224 may be configured to extract a second set of recommendations by semantically matching to the filtered first set of recommendations. The second set of recommendations may be ranked and displayed through the user interface 110.

[0061] In an embodiment, the multi-modal data processing engine 106 may be configured to extract metadata from the collected information to identify one or more attributes using data extraction. The one or more attributes represent features of the listed products, the listed services, or the user profiles. The multi-modal data processing engine 106 is further configured to associate a plurality of semantic labels with the one or more extracted attributes to generate a plurality of semantic vectors. The semantic vectors may be generated using semantic vector generation methods. The multi-modal data processing engine 106 is further configured to store the plurality of semantic vectors in the personalized database.

[0062] A person skilled in the art will appreciate that the processor 200 may be configured to perform the steps of the present disclosure using the data 206 instead of the one or more modules 216-224, to provide personalized recommendations to user queries.

[0063] A person skilled in the art will appreciate that any techniques other than the above- mentioned technique may be used to perform the steps performed by the intent and context determining module 216, the privacy attributes identifying module 218, the first set of recommendations module 220, the bias filtering module 222 and the second set of recommendations module 224, which are configured to provide personalized recommendations to user queries.

[0064] Figure 3 illustrates a block diagram of the multi-modal data processing engine 106 of Figure 1, in accordance with some embodiments of the present disclosure. The multi-modal data processing engine 106 includes an information accessing module 302, an attribute identification module 304, a semantic vectors generation module 306, and a personalized database creation module 308. The information accessing module 302 access information associated with external server platform 108. The information comprises at least one of user profiles, listed products and associated descriptions, and listed services and associated descriptions.

[0065] The attribute identification module 304 may be configured to extract metadata from the collected information to identify one or more attributes using data extraction. The one or more attributes represent features of the listed products, the listed services, or the user profiles. For example, the one or more attributes may be product features, category classifications, user engagement descriptors, or service tags.

[0066] The semantic vectors generation module 306 may be configured to generate a plurality of semantic vectors by associating a plurality of semantic labels with the one or more extractedattributes. For example, the semantic labels may include intent-aligned descriptors (e.g., “budget fitness,” “sleep tracking enabled,” “eco-friendly”) or structural groupings (e.g., “category: wearables,” “attribute: waterproof’). The personalized database creation module 310 may be configured to create a personalized database by storing the plurality of semantic vectors and associated plurality of semantic labels and the one or more identified attributes to enable fast similarity computation and contextual matching during query resolution.

[0067] In an embodiment, whenever there is an update in the data, for example, newly listed products, updated service descriptions, or modified user profile entries, then the multi-modal data processing engine 106 processes the updated data. The information accessing module 302 captures the updated entries, and the attribute identification module 304 extracts relevant features, such as product category, specifications, and tags, from the updated entries. The features are classified and enriched by the semantic vectors generation module 306 using existing label sets or dynamically created semantic tags based on the contextual meaning. The semantic vectors are indexed and stored by the personalized database creation module 308 in the semantic vector database. This update may occur incrementally or in batch, depending on system configuration, ensuring that the multi-modal data processing engine 106 remains responsive to the most current platform content.

[0068] Figure 4 illustrates a block diagram of the privacy attributes identifying module 218 of Figure 2, in accordance with some embodiments of the present disclosure. The privacy attributes identifying module 218 includes a privacy attributes selection module 402 and a privacy attributes enabling module 404. The privacy attributes selection module 402 allows selection of one or more pre-defined user-specific privacy attributes via the user interface. The privacy attributes enabling module 404 enables the selected user-specific privacy attributes.

[0069] In an example scenario, a user may search for "smartwatches under ?3000” via the user interface. Subsequently, using the privacy attributes selection module 402 through the user interface, the user may select user-specific privacy attributes. For example, the user may selectively enable or disable attributes, such as choosing to share browsing history and purchase history, but opting out of sharing the current location and demographic details like age and gender. The selections are registered through the privacy attributes selection module 402. The privacy attributes enabling module 404 enables only the selected attributes, for example, browsing history and purchase history, while ensuring that the opted-out attributes (e.g., location and gender) are excluded from downstream processing.

[0070] Figure 5 illustrates a block diagram of the bias filtering module 222 of Figure 2, in accordance with some embodiments of the present disclosure. The bias filtering module 222 may include a dynamic pattern detection module 502 and filters module 504. The dynamic pattern detection module 502 dynamically detects patterns, which are an indicative of an interaction weightage assigned to a product, a category, and / or a vendor, based on the one or more user interaction metrics. The filters module 508 may apply filters to the detected patterns based on the determined intent and the bias factor. For example, the bias factor is an indicative of a correlation between the determined intent and the one or more user interaction metrics.

[0071] In one scenario, a user submits a query, for example, “affordable fitness watch with sleep tracking”. The first set of recommendations module 220 analyses the user query to determine both the intent, for example, the user’s interest in health-focused wearable devices and the context, such as affordability. For example, the user’s previously selected privacy attributes may be purchasing history and budget preferences. Based on the analysis, the first set of recommendations module 220 retrieves a first set of recommendations that align with the intent and context while complying with the privacy constraints set by the user. Once the first set of recommendations is generated, the bias filtering module 222 initiates a bias filtering process. To do so, the dynamic pattern detection module 502 detects a range of user interaction metrics, including system-wide click-through rates, engagement trends, third-party popularity scores, and external product rankings. The user interaction metrics are used to dynamically detect recurring patterns, for example, a specific vendor’s products consistently appearing at the top of lists due to artificially elevated SEO scores or sponsored visibility. The patterns reflect interaction weightage linked to product exposure that may not correspond to the user’s intent. The filters module 504 applies bias correction filters on top of the detected patterns by evaluating how strongly they correlate with the user’s original intent. For example, if a particular product appears frequently across users but offers limited relevance to the user’s current query for e.g., a high-end smartwatch with no sleep tracking, then the high-end smartwatch is downranked or excluded. The bias factor represents the mismatch between the influence of system-wide popularity metrics and the user’s personal search context.

[0072] Figure 6 illustrates a block diagram of the second set of recommendations module 224 of Figure 2, in accordance with some embodiments of the present disclosure. The second set of recommendations module 224 includes a first set of recommendations transformation module 602, a semantic vectors retrieval module 604, a semantic vectors comparison module 606, a second set of recommendations extraction module 608 and a ranking module 610. Thefirst set of recommendations transformation module 602 converts the filtered first set of recommendations into one or more vectors using natural language processing (NLP) techniques. The semantic vectors retrieval module 604 retrieves a plurality of pre-defined vectors that match the intent and the context. The semantic vectors comparison module 606 compares the one or more vectors against each of the plurality of pre-defined vectors. The second set of recommendations extraction module 608 generates the second set of recommendations based on the comparison. The ranking module 610 ranks each of the second set of recommendations based on the user profile.

[0073] In an embodiment, each of the second set of recommendations is assigned a relevancy score based on the comparison of the product / service metadata with the user’s selected privacy attributes. A confidence score for each recommendation is computed by aggregating the relevancy score and pre-assigned weight factors (e.g., user affinity, item freshness, or trend alignment). The second set of recommendations are then ranked in descending order of the confidence score. This ensures that the highest-ranking results not only match the user’s intent and context semantically but also reflect user-specific relevance preferences, such as price sensitivity, brand affinity, or topical interests.

[0074] In an example scenario, after generating and filtering the first set of recommendations for a user query such as “budget fitness tracker with sleep monitoring,” the first set of recommendations transformation module 602 processes the filtered product list comprising items like budget smartwatches with varying specifications. The first set of recommendations transformation module 602 converts the textual metadata and descriptions of the filtered product list into vector representations using natural language processing (NLP) techniques. The vectors include semantic and structural attributes of the products, for example, core functions (e.g., step tracking, sleep monitoring), price range, and design profile. The semantic vectors retrieval module 604 retrieves a set of vectors from an indexed store that were previously generated based on aggregated product descriptions and aligned with various intents and contexts. For example, pre-defined vectors may represent vectors like “budget-friendly health wearables,” “fitness-focused smart bands,” or “female-oriented trackers with sleep features.” The semantic vectors comparison module 606 then compares each vector derived from the first set of recommendations with the pre-defined vectors using similarity metrics such as cosine similarity. Hence, the PRAS 104 identifies which items in the first set are semantically closest to the user’ s intended context. Finally, the second set of recommendations extraction module 608 selects and outputs only those products that exhibit high semanticalignment with the retrieved pre-defined vectors. The ranking module 610 ranks the second set of recommendations based on the user profile. Thereby producing a refined, contextually enriched second set of recommendations that are better tailored to the user’s original query and intent.

[0075] Figure 7 illustrates a flow diagram of a method 700 for providing personalized recommendations to user queries, in accordance with some embodiments of the present disclosure.

[0076] As illustrated in Figure 7, the method may comprise one or more steps. The method 700 may be described in the general context of computer executable instructions. Generally, computer executable instructions can include routines, programs, objects, components, data structures, procedures, modules, and functions, which perform particular functions or implement particular abstract data types.

[0077] The order in which the method is described is not intended to be construed as a limitation, and any number of the described method blocks can be combined in any order to implement the method. Additionally, individual blocks may be deleted from the methods without departing from the scope of the subject matter described herein. Furthermore, the method 700 can be implemented in any suitable hardware, software, firmware, or combination thereof.

[0078] At a block 702, the method includes determining an intent and a context associated with the user query based on an associated user profile.

[0079] At a block 704, the method includes dynamically determining one or more user-specific privacy attributes associated with the user profile. In an embodiment, for dynamically determining the one or more user-specific privacy attributes, the method further includes receiving dynamic selection of one or more pre-defined user-specific privacy attributes via a user interface and enabling the selected user-specific privacy attributes. The one or more userspecific privacy attributes includes at least one of purchase behavior, historical transaction, demographic details, product metadata, and real-time contextual information created by the multi-modal data processing engine based on the one or more user interactions.

[0080] At a block 706, the method includes retrieving a first set of recommendations mapping with the intent, the context and the one or more user-specific privacy attributes.

[0081] At a block 708, the method includes filtering the first set of recommendations and applying a bias factor by analysing one or more user interaction metrics. In an embodiment,for filtering the first set of recommendations, the method further includes dynamically detecting patterns, indicative of an interaction weightage assigned to a product, a category, and / or a vendor, based on the one or more user interaction metrics, and applying filters to the detected patterns based on the determined intent and the bias factor to derive the first set of recommendations. The one or more user interaction metrics include at least one of user-specific click-through rates, past engagement, external rankings, and third-party interactions. Further, the bias factor being an indicative of a correlation between the determined intent and the one or more user interaction metrics.

[0082] At a block 710, the method includes extracting a second set of recommendations by semantically matching with the filtered first set of recommendations using a multi-modal data processing engine for providing to a user. In an embodiment, for extracting the second set of recommendations using the multi-modal data processing engine, the method discloses transforming the filtered first set of recommendations into one or more vectors using natural language processing (NLP) and retrieving a plurality of semantic vectors from the plurality of generated semantic vectors mapping with the intent and the context from the personalized database. The method further discloses comparing the one or more vectors against each of the plurality of semantic vectors. Finally, the method discloses extracting the second set of recommendations based on the comparison and ranking each of the second set of recommendations based on the user profile.

[0083] In an embodiment, for ranking each of the second set of recommendations, the method discloses assigning a relevancy score to each of the second set of recommendations based on the one or more user-specific privacy attributes. Further, the method discloses computing a confidence score for each of the second set of recommendations by aggregating predetermined weights of the second set of recommendations and the relevancy score and ranking each of the second set of recommendations based on the confidence score. For example, a user searches “comfortable formal shoes under ?2000” via a mobile application. The method identifies the intent (“formal shoes”) and context (“budget-conscious”). The user’s profile shows a preference for lightweight brands and avoids leather. Privacy settings exclude location tracking. A first set of results is retrieved, but several top items show inflated rank due to popularity signals. After bias filtering, the method uses semantic vectors to find alternative shoes that match the budget, style, and material preference. The method assigns relevancy and confidence scores, ranks them, and displays the results in the user interface. If no strong semantic match is found, the method suggests refining the user query.

[0084] In an alternate embodiment, the method includes accessing information associated from an external server platform. The information comprises at least one of user profiles, listed products and associated descriptions, and listed services and associated descriptions. The method further includes identifying one or more attributes based on metadata extracted from the accessed information. In an example, the one or more attributes represent features of the listed products, the listed services, or the user profiles. The method further includes generating a plurality of sematic vectors by associating a plurality of semantic labels with the one or more identified attributes. The method further includes creating the personalized database by storing the plurality of semantic vectors and associated plurality of semantic labels and the one or more identified attributes.

[0085] In an alternate embodiment, prior to accessing the information from the external server platform, the method includes analysing the information associated with the platform. The method includes determining an anomaly in the analysed information. The method includes rectifying the analysed information based on the determined anomaly.

[0086] In a non-limiting embodiment of the present disclosure, one or more non-transitory computer-readable media may be utilized for implementing the embodiments consistent with the present disclosure.

[0087] The various illustrative logical blocks, modules, and operations described in connection with the present disclosure may be implemented or performed with a general- purpose processor, discrete gate or transistor logic, discrete hardware components or any combination thereof designed to perform the functions described herein. A general -purpose processor may include a microprocessor, but in the alternative, the processor may include any commercially available processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a plurality of microprocessors, or any other such configuration.

Claims

We claim:

1. A method for providing personalized recommendations to a user query, the method comprising: determining an intent and a context associated with the user query based on an associated user profile; dynamically determining one or more user-specific privacy attributes, associated with the user profile, enabled in real time; retrieving a first set of recommendations, from a personalized database, mapping with the intent, the context and the one or more user-specific privacy attributes; filtering the first set of recommendations based on a bias factor by analysing one or more user interaction metrics; and extracting, using a multi-modal data processing engine, a second set of recommendations semantically matching with the filtered first set of recommendations for providing to a user.

2. The method of claim 1, further comprising: accessing information from an external server platform, wherein the information comprises at least one of user profiles, listed products and associated descriptions, and listed services and associated descriptions; identifying one or more attributes based on metadata extracted from the accessed information, wherein the one or more attributes represent features of the listed products, the listed services, or the user profiles; generating a plurality of sematic vectors by associating a plurality of semantic labels with the one or more identified attributes; and creating the personalized database by storing the plurality of semantic vectors and associated plurality of semantic labels and the one or more identified attributes.

3. The method of claim 2, wherein prior to accessing the information from the external server platform, the method comprises: analyzing the information associated with the external server platform; determining an anomaly in the analyzed information; and rectifying the analyzed information based on the determined anomaly.

4. The method of claim 1, wherein dynamically determining the one or more user-specific privacy attributes comprises:receiving dynamic selection of one or more pre-defined user-specific privacy attributes via a user interface; and enabling the selected user-specific privacy attributes.

5. The method of claim 1, wherein the one or more user-specific privacy attributes comprises at least one of purchase behavior, historical transaction, demographic details, product metadata, and real-time contextual information created by the multi-modal data processing engine based on the one or more user interactions.

6. The method of claim 1, wherein the filtering the first set of recommendations comprises steps of: dynamically detecting patterns, indicative of an interaction weightage assigned to a product, a category, a vendor, based on the one or more user interaction metrics, wherein the one or more user interaction metrics comprises at least one of user-specific click-through rates, past engagement, external rankings, and third-party interactions; and applying filters to the detected patterns based on the determined intent and the bias factor to derive the first set of recommendations, wherein the bias factor is indicative of a correlation between the determined intent and the one or more user interaction metrics.

7. The method of claim 2, wherein extracting the second set of recommendations using the multi-modal data processing engine comprises: transforming, using natural language processing (NLP), the filtered first set of recommendations into one or more vectors; retrieving, from the personalized database, the plurality of semantic vectors, from the plurality of generated semantic vectors mapping with the intent and the context; comparing the one or more vectors against each of the plurality of semantic vectors; based on the comparison, extracting the second set of recommendations; and ranking each of the second set of recommendations based on the user profile.

8. The method of claim 7, wherein ranking each of the second set of recommendations comprises: assigning a relevancy score to each of the second set of recommendations based on the one or more user-specific privacy attributes;computing a confidence score for each of the second set of recommendations by aggregating predetermined weights of the second set of recommendations and the relevancy score; and ranking, based on the confidence score, each of the second set of recommendations.

9. A system for providing personalized recommendations to a user query, the system comprising: one or more processors; and a memory communicatively coupled to the one or more processors, the one or more processors is configured to: determine an intent and a context associated with the user query based on an associated user profile; dynamically determine one or more user-specific privacy attributes associated with the user profile, enabled in real time; retrieve a first set of recommendations, from a personalized database, mapping with the intent and the context and the one or more user-specific privacy attributes; filter the first set of recommendations based on a bias factor by analyzing one or more user interaction metrics; and extract, using a multi-modal data processing engine, a second set of recommendations by semantically matching with the filtered first set of recommendations for providing to a user.

10. The system of claim 9, wherein the one or more processors is further configured to: access information associated with an external server platform, wherein the information comprises at least one of user profiles, listed products and associated descriptions, and listed services and associated descriptions; identify one or more attributes based on metadata extracted from the accessed information, wherein the one or more attributes represent features of the listed products, the listed services, or the user profiles; generate a plurality of semantic vectors by associating a plurality of semantic labels with the one or more extracted attributes; andcreate the personalized database by storing the one or more semantic vectors and associated plurality of semantic labels and the one or more identified attributes.

11. The system of claim 10, wherein prior to access of the information from the external server platform, the one or more processors is configured to: analyse the information associated with the platform; determine an anomaly in the analyzed information; and rectify the analyzed information based on the determined anomaly.

12. The system of claim 9, wherein to dynamically determine the one or more userspecific privacy attributes, the one or more processors is further configured to: receive dynamic selection of one or more pre-defined user-specific privacy attributes via a user interface; and enable the selected user-specific privacy attributes.

13. The system of claim 9, wherein the one or more user-specific privacy attributes comprises at least one of purchase behavior, historical transaction, demographic details, product metadata, and real-time contextual information created by the multi-modal data processing engine based on the one or more user interactions.

14. The system of claim 9, wherein to filter the first set of recommendations the one or more processors is further configured to: dynamically detect patterns, indicative of an interaction weightage assigned to a product, a category, a vendor, based on the one or more user interaction metrics, wherein the one or more user interaction metrics comprises at least one of user-specific click-through rates, past engagement, external rankings, and third-party interactions; and apply filters to the detected patterns based on the determined intent and the bias factor to derive the first set of recommendations, wherein the bias factor is indicative of a correlation between the determined intent and the one or more user interaction metrics.

15. The system of claim 10, wherein to extract the second set of recommendations, using the multi-modal data processing engine, the one or more processors is further configured to:transform, using natural language processing (NLP), the filtered first set of recommendations into one or more vectors; retrieve, from the personalized database, the plurality of semantic vectors, from the plurality of generated semantic vectors, mapping with the intent and the context; compare the one or more vectors against each of the plurality of semantic vectors; based on the comparison, extract the second set of recommendations; and rank each of the second set of recommendations based on the user profile.

16. The system of claim 15, wherein to rank each of the second set of recommendations, the one or more processors is further configured to: assigning a relevancy score to each of the second set of recommendations based on the one or more user-specific privacy attributes; computing a confidence score for each of the second set of recommendations by aggregating predetermined weights of the second set of recommendations and the assigned relevancy score; and ranking, based on the confidence score, each of the second set of recommendations.

Images