Hybrid recommendation system using Team Frequency-Inverse Document Frequency (TF-IDF) for film recommendations

Abstract

A hybrid recommendation system for generating personalized film recommendations, the system comprising the following: a processing unit for receiving metadata, configured to receive descriptive data related to films, the descriptive data including plot summaries, genre identifiers, cast lists and director names; a text preprocessing module that is communicatively coupled with the metadata ingestion processing unit and is configured to perform tokenization, stopword removal, stemming, and lemmatization on the descriptive data to obtain a processed corpus; a TF-IDF vectorization unit configured to encode the processed corpus into high-dimensional semantic feature vectors by calculating TF-IDF inverse document frequency values ​​over the entire film dataset, with the semantic feature vectors representing the contextual meaning of terms for individual films; a collaborative filter engine comprising a user-element interaction matrix, wherein the engine is configured to compute latent preference signals using one or more techniques selected from the group consisting of cosine similarity, k-nearest neighbor similarity, and matrix factorization; a score fusion controller coupled to both the TF-IDF vectorization unit and the collaborative filter engine, wherein the module is configured to normalize the semantic feature vectors and the collaborative preference scores and dynamically combine them according to an adaptive weighting coefficient, the coefficient being determined as a function of data density, interaction sparsity, and user history length; and a recommendation output module configured to evaluate candidate films for each user based on the combined score and generate a top-N recommendation list.

Description

Technical field

[0001] The present invention relates to intelligent information retrieval and recommendation systems. In particular, the invention relates to a hybrid recommendation device and a computer architecture that integrates semantic feature extraction based on Term Frequency-Inverse Document Frequency (TF-IDF) with collaborative filter models to improve the accuracy, robustness, and personalization of film recommendations in real-time environments such as online streaming platforms and video-on-demand systems. BACKGROUND OF THE INVENTION

[0002] Given the exponential growth of digital media content, film streaming platforms face significant challenges in providing personalized and accurate recommendations. Existing systems are typically divided into collaborative filtering (CF) and content-based filtering (CBF). CF models use data on user-object interaction to derive preferences, while CBF approaches rely on metadata descriptors such as plot summaries, cast information, and genres. However, both approaches have limitations when used independently.

[0003] For example, CF suffers from the "cold start" problem, where new users or new items cannot be efficiently assigned due to insufficient historical data. Similarly, CBF systems based solely on descriptive metadata may not adequately reflect collective community preferences or capture latent similarities between users. State-of-the-art hybrid systems often use weighted averages of CF and CBF values, but they are not modularly adaptable, do not fully exploit semantic variation at the concept level, and do not dynamically adapt to changing user behavior in real time.

[0004] Accordingly, there is a need for a hybrid recommendation device and method that integrates the weighting of text features via TF-IDF with collaborative models, while simultaneously enabling dynamic weight adjustments, cold start handling, and modular scalability for use in various recommendation infrastructures.

[0005] The rapid spread of digital media has transformed how users consume entertainment, particularly in the film sector, where online streaming platforms offer access to vast content libraries. As these libraries have grown, the need for intelligent recommendation systems that efficiently guide users to relevant content has become critical. Recommendation systems have thus proven to be essential technologies for increasing user engagement, personalizing experiences, and improving platform retention. The technical landscape of recommendation methods is diverse and has evolved through several generations of approaches. Each approach offers specific advantages but also has limitations that can hinder optimal performance in real-world applications.

[0006] The first recommendation systems were based on content-based filtering. These systems analyzed descriptive attributes of objects such as film genre, actors, directors, release year, or freely selectable metadata like synopses and reviews. By comparing the attributes of previously popular objects with the metadata of potential objects, these systems generated recommendations that corresponded to the user's known preferences. Content-based filtering offered the advantage of interpretability, as the rationale for the recommendation could be traced back to certain shared attributes. Furthermore, content-based approaches avoided dependence on data from other users and were therefore well-suited for users with highly specialized or niche preferences. However, content-based systems also had some disadvantages.They often lacked diversity and recommended only objects very similar to those already consumed, leading to overspecialization or "filter bubbles." Furthermore, content-based filtering was heavily dependent on the completeness and quality of the metadata. If the descriptive data was sparse, inaccurate, or distorted, the system's recommendations deteriorated. These approaches also lacked the ability to capture emerging trends in the community or latent patterns beyond explicit metadata capabilities.

[0007] With the scaling of online platforms and the increasing amount of user interaction data, collaborative filtering proved to be a powerful alternative. Collaborative filtering harnessed the wisdom of the crowd by recommending items to users based on the preferences of other users with similar interaction histories. The two main paradigms of collaborative filtering were memory-based and model-based methods. Memory-based collaborative filtering relied on similarity calculations and often used cosine similarity or Pearson correlation to identify nearest neighbors among users or items. For example, if user A had similar viewing habits to user B, user A might be recommended movies that user B liked but user A had not yet seen.Model-based collaborative filtering, on the other hand, used statistical or machine learning techniques such as matrix factorization, singular value decomposition, or, more recently, deep neural networks to learn latent factors that captured hidden relationships between users and items. Collaborative filtering systems offered significant advantages, particularly their ability to capture community-driven patterns without requiring detailed item metadata. They could recommend items structurally different from what a user has consumed but popular with similar audiences, thus increasing the discovery rate.

[0008] Despite these strengths, collaborative filtering faced significant challenges. One of the most noticeable was the cold start problem. For new users with little or no interaction history, the system lacked sufficient data to establish similarity relationships or generate accurate recommendations. Newly added movies without prior user interaction were also virtually invisible to the recommendation engine. Another issue was data scarcity, particularly in large catalogs where only a small fraction of items generated any significant interaction, resulting in thin user-item matrices with low predictive power. Scalability was also a concern, as calculating similarity matrices or factoring huge user-item matrices placed considerable computational and memory demands on the servers, especially in real-time recommendation environments.Furthermore, collaborative filter models were prone to popularity biases and often overrepresented blockbusters, while niche or less researched content was underrepresented.

[0009] Given the inherent limitations of both content-based and collaborative filtering approaches, researchers and practitioners have explored hybrid recommendation models. These models sought to combine the strengths of multiple methods to achieve more robust results. One hybrid approach, for example, simply combined the results of content-based and collaborative filtering using weighted averages. Another class of hybrid systems employed sequential or staged designs, where one technique filtered or narrowed the candidate pool, and the other refined the ranking. Such hybrids demonstrated improved performance compared to single-method systems, particularly in handling cold-start situations and improving diversity. However, many of these hybrid systems remained rudimentary, lacking adaptive mechanisms to dynamically balance the influence of each model.Most of them used static weightings that did not account for changing conditions such as the density of user interaction data, the evolution of metadata, or the temporal variation of user preferences. This rigidity led to suboptimal results when deployed on a large scale.

[0010] In the field of text analysis, the introduction of Term Frequency-Inverse Document Frequency (TF-IDF) significantly improved the measurement of content similarities. TF-IDF provided a statistical measure that reflected the importance of a word for a specific document compared to the entire corpus. By emphasizing unique terms and downgrading overly frequent words, TF-IDF enabled systems to extract more meaningful representations of textual data such as film plots, synopses, or cast information. In the context of recommendation systems, TF-IDF, by capturing nuanced text signals, allowed for a more precise matching of user profiles to film descriptions.For example, a user who preferred films with complex narrative structures could be assigned plot descriptions with keywords such as "intrigue", "espionage" or "mystery", which could not be effectively captured by simple keyword matching.

[0011] Nevertheless, TF-IDF-based content matching alone was not a complete solution. While it improved the semantic diversity of content-based filtering, it remained fundamentally dependent on the availability and quality of metadata. Furthermore, TF-IDF could not account for collective user preferences, as it relied solely on text features without incorporating interaction patterns. A TF-IDF-based recommendation mechanism could precisely tailor content to user profiles, but it might not capture the social evidence derived from the community's consumption patterns. Additionally, TF-IDF representations were high-dimensional and sparse, posing computational challenges for real-time similarity comparisons in large-scale streaming environments.

[0012] Other semantic approaches, such as Latent Semantic Indexing (LSI) and Latent Dirichlet Allocation (LDA), went beyond TF-IDF by attempting to reduce dimensionality and uncover hidden themes in textual data. While these approaches improved efficiency and revealed latent associations between terms, they often suffered from interpretability problems and high computational costs. Furthermore, they still largely operated in the content domain, failing to adequately account for the dynamic preferences that emerge from user interactions. Although such models improved representation, they neither solved the fundamental cold-start problems of collaborative filtering nor did they dynamically adapt to evolving patterns of user behavior.

[0013] The shortcomings of existing approaches are becoming increasingly apparent in modern real-time streaming ecosystems. Users expect recommendations to be not only accurate, but also diverse, up-to-date, and responsive to their recent interactions. Content-based systems alone cannot achieve diversity, while collaborative filtering alone cannot effectively address data scarcity. Hybrid systems with static weighting cannot adapt to the dynamic interplay between user history and content semantics. Therefore, there is a growing demand for architectures that are both semantically aware and socially grounded, while also being computationally scalable and adaptable in real time.

[0014] In this technical context, the need for a hybrid recommendation system that integrates TF-IDF with collaborative filtering becomes clear. By combining the semantic richness of TF-IDF representations with the collective wisdom of collaborative models, such a system can effectively overcome the cold-start problem, address sparsities, balance diversity and relevance, and enable adaptive, real-time personalization. Existing solutions, whether purely content-based, purely collaborative, or naively hybrid, have not achieved the optimal combination of these properties. Their shortcomings underscore the need for a modular, adaptive, and computationally efficient system that dynamically integrates both semantic and collaborative signals to generate recommendations that meet individual preferences while also reflecting community trends. Objectives of the invention

[0015] The invention aims to provide a system and device that uses TF-IDF to extract meaningful semantic features from film metadata, integrates these features into collaborative user-element models, and delivers precise, adaptive, real-time recommendations. A further aim is to solve cold-start problems by leveraging TF-IDF similarity in sparse interaction data. Simultaneously, community-driven collaborative signals are gradually balanced as data density increases. The system is designed to operate with modular computing components and is therefore suitable for OTT platforms, cloud-based recommendation engines, and local deployment on end devices. Summary of the invention

[0016] The invention discloses a hybrid recommendation device configured as a modular computer structure comprising a text processing unit, a TF-IDF vectorization unit, a collaborative filter engine, a score fusion and normalization module, and a recommendation output module. The device processes film metadata such as plot, genre, cast, and directors, performs natural language preprocessing, including tokenization, stopword removal, and stemming, and creates TF-IDF vectors that capture the semantic meaning of the entire corpus. These semantic vectors are then matched against user preference profiles, which are either aggregated from previously consumed content or derived from implicit ratings and playback histories.

[0017] On the collaborative side, the system uses similarity models, including cosine similarity, neighborhood-based techniques, and matrix factorization, to derive recommendations based on historical user-element interactions. Both TF-IDF and collaborative filter values ​​are integrated via a score fusion engine that dynamically adjusts the weighting coefficients based on data availability and density. Under cold-start conditions, the device emphasizes TF-IDF similarity to recommend semantically relevant films, while for active users with extensive history, collaborative weighting is increased.

[0018] The device features adaptive learning mechanisms by continuously receiving real-time user feedback such as playback duration, skip behavior, and explicit ratings, enabling iterative model refinement and accuracy improvement.

[0019] The objectives of the present invention are to overcome the persistent limitations of existing recommendation systems and to provide a technically advanced solution that addresses the challenges of data scarcity, cold start conditions, lack of semantic depth, and static hybridization. A key objective of the invention is to provide a hybrid recommendation framework that meaningfully integrates semantic content analysis using TF-IDF with collaborative filtering of user elements, ensuring that recommendations are both contextually and socially validated. Another objective of the invention is to ensure that the system dynamically adapts its weighting mechanism to the availability and density of user interaction data. This avoids the rigidity of fixed-weight hybrid models, and the engine can balance the contribution of semantic similarity and collaborative signals in real time.Another objective of the invention is the development of a robust recommendation system that can process large amounts of metadata such as text descriptions, cast, genre and user feedback using natural language processing, while simultaneously capturing implicit behavioral signals such as playback duration and skip patterns to improve the accuracy of the recommendations.

[0020] Another objective of the invention is to solve the cold-start problem of new users and new films by prioritizing TF-IDF similarity in situations with minimal interaction data. This ensures that meaningful recommendations can still be generated even under sparse conditions. A further objective is to build a modular architecture that can be seamlessly integrated into streaming platforms, OTT services, or video-on-demand engines without requiring a significant redesign of the existing infrastructure. The invention is also intended to enable continuous learning from real-time user interactions and ensure that recommendations evolve with changing tastes, preferences, and consumption patterns. A further objective is to improve recommendation diversity by combining semantic novelty from TF-IDF analysis with community-driven discovery through collaborative filtering.This avoids overspecialization while maintaining a high degree of personalization.

[0021] Ultimately, an objective of the invention is to develop a computing device and system architecture that integrates dedicated processing units, memory, and communication interfaces to support the large-scale execution of low-latency recommendation techniques. By achieving these objectives, the invention provides an adaptive, scalable, and semantically enriched recommendation solution that surpasses existing models in terms of accuracy, adaptability, and user satisfaction. BRIEF DESCRIPTION OF THE FIGURE

[0022] These and other features, aspects, and advantages of the present invention will be better understood if the following detailed description is read with reference to the accompanying drawing, in which the same symbols consistently represent the same parts. The following applies: Fig. Figure 1 shows a block diagram of a device and system for hybrid movie recommendation using Term Frequency-Inverse Document Frequency (TF-IDF) with collaborative filtering.

[0023] Experts will also recognize that the elements in the drawing are shown for the sake of simplicity and are not necessarily to scale. For example, the flowcharts illustrate the process by highlighting the main steps to enhance understanding of the aspects of this disclosure. Furthermore, with regard to the design of the device, one or more components of the device may be represented in the drawing by conventional symbols, and the drawing may show only the specific details relevant to understanding the embodiments of this disclosure, so as not to clutter the drawing with details that are readily apparent to those skilled in the art after reading this description. Detailed description of the invention

[0024] For a better understanding of the inventive principles, reference is made below to the embodiment shown in the drawing, which is described in specific terminology. However, this does not limit the scope of the invention. Changes and further modifications of the illustrated system, as well as further applications of the inventive principles, are possible, as would normally occur to a person skilled in the art in the field of invention.

[0025] It is clear to the person skilled in the art that the preceding general description and the following detailed description are exemplary and explanatory of the invention and are not intended as a limitation of it.

[0026] References in this specification to “an aspect”, “another aspect”, or similar expressions mean that a particular feature, structure, or property described in connection with the embodiment is included in at least one embodiment of the present disclosure. Therefore, occurrences of the expressions “in one embodiment”, “in another embodiment”, and similar expressions in this specification may all refer to the same embodiment, but need not.

[0027] The terms "includes," "include," or other variations thereof are intended to cover non-exclusive inclusion, such that a process or method that includes a list of steps may not only contain those steps but may also include other steps not expressly listed or inherent in such process or method. Likewise, the statement "includes..." in the case of one or more devices, subsystems, elements, structures, or components does not, without further limitations, preclude the existence of other devices, subsystems, elements, structures, components, or additional devices, subsystems, elements, structures, or components.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by a person skilled in the art in the field of the invention. The system, methods, and examples provided here serve only for illustration and are not to be construed as a limitation.

[0029] Embodiments of the present disclosure are described in detail below with reference to the attached drawing.

[0030] In Fig.Figure 1 shows a block diagram of the device and system for hybrid film recommendations using Term Frequency-Inverse Document Frequency (TF-IDF) with collaborative filtering. The system 100 comprises: a metadata ingestion processing unit (102) configured to receive descriptive data related to films, including plot summaries, genre identifiers, cast lists, and director names; a text preprocessing module (104) communicatively connected to the metadata ingestion processing unit and configured to perform tokenization, stopword removal, stemming, and lemmatization of the descriptive data to create a processed corpus;a TF-IDF vectorization unit (106) configured to encode the processed corpus into high-dimensional semantic feature vectors by calculating Term Frequency-Inverse Document Frequency values ​​across the entire film dataset, the semantic feature vectors representing the contextual meaning of terms for individual films; a collaborative filter engine (108) comprising a user-element interaction matrix, the engine being configured to compute latent preference signals using one or more techniques selected from the group consisting of cosine similarity, k-nearest neighbor similarity, and matrix factorization;a score fusion module (110) coupled to both the TF-IDF vectorization unit and the collaborative filter engine, wherein the module is configured to normalize the semantic feature vectors and the collaborative preference scores and dynamically combine them according to an adaptive weighting coefficient, the coefficient being determined as a function of data density, interaction sparsity, and user history length; and a recommendation output module (112) configured to evaluate candidate films for each user based on the combined score and generate a top-N recommendation list.

[0031] In one embodiment, the TF-IDF vectorization unit (106) is further configured to create a user profile vector by aggregating TF-IDF representations of movies previously consumed by a particular user, comparing the profile vector with candidate movie vectors using cosine similarity to derive a semantic relevance value, and transferring this semantic relevance value to the rating fusion controller for integration with collaborative preference values.

[0032] In one embodiment, the collaborative filter engine (108) is also configured to apply dimensionality reduction by singular value decomposition or probabilistic matrix factorization to the user-element interaction matrix, thereby generating latent factor vectors for both users and elements, wherein the engine calculates the collaborative scores by measuring the similarity of the inner product between latent factor vectors corresponding to a target user and a candidate film.

[0033] In one embodiment, the score fusion controller (110) is configured to dynamically adjust weighting coefficients in real time, such that in the case of a sparse user interaction history, the weighting coefficient favors the semantic similarity value derived from the TF-IDF vectorization unit, and in the case of a dense user interaction history, the weighting coefficient favors the collaborative value, thereby mitigating cold start problems for both new users and newly introduced movies.

[0034] In one embodiment, the recommendation output module (112) is also coupled with a feedback capture interface configured to record user behavior in real time, including playback start, playback completion rate, fast-forward or skip activity, dwell time, and explicit rating results, and wherein the feedback data is iteratively passed to both the TF-IDF vectorization unit for recalibrating term weights and the collaborative filter engine for updating user-element interactions.

[0035] In one embodiment, the computing device comprises a modular hardware structure with a multi-core central processing unit for performing collaborative filter calculations, a graphics processing unit or tensor processing unit optimized for vector and matrix operations in the TF-IDF vectorization unit, volatile memory for storing sparse TF-IDF matrices and latent factor representations, persistent memory for historical user-element data, and at least one communication interface, selected from Ethernet, Wi-Fi, or application programming interfaces, for integration with streaming platforms.

[0036] In one embodiment, the text preprocessing module (104) also includes enhancements in natural language processing, including named entity recognition for extracting proper names corresponding to the names of actors or directors, part-of-speech identification for refining the contextual meaning of terms, and bigram or trigram recognition for capturing multi-part expressions in plot summaries, thereby improving the discrimination capability of the TF-IDF vectors.

[0037] In one embodiment, the recommendation output module (112) is also configured to use a re-evaluation level based on diversity optimization, wherein the re-evaluation level penalizes elements that are semantically too similar to films already consumed by the user, thereby ensuring that the final recommendation list balances both personalization accuracy and content novelty.

[0038] In one embodiment of the score fusion controller (110), it is additionally configured to maintain temporal adjustment by incorporating a time decay function for the collaborative scores, so that older interactions are attenuated compared to newer interactions, thereby ensuring that the recommendation list reflects evolving user preferences in real time.

[0039] In one embodiment, the feedback capture interface is also able to transmit user interaction signals to an adaptive learning controller that modifies the weighting function of the score fusion module, with the adaptive learning controller employing reinforcement learning techniques to maximize long-term user satisfaction, measured by metrics such as recommendation acceptance rate, engagement duration, and retention frequency.

[0040] The system is hardware-based and implemented through structurally defined components that interact in a coordinated manner to ensure the desired functionality. The system comprises sensor modules in the form of transducers, optical detectors, or strain gauges, each mounted in housings with electrical leads and generating analog or digital signals representing physical parameters. These sensors are electrically connected to signal conditioning circuits, including amplifiers, filters, and analog-to-digital converters, arranged on a printed circuit board to ensure noise reduction and signal integrity. The processed output signals are routed to a processing unit, implemented as a microcontroller, digital signal processor, or FPGA chip on a hardware substrate. The embedded firmware is stored in non-volatile flash memory for execution.The memory elements, including RAM modules and EEPROM, are structurally arranged to enable real-time data buffering and persistent parameter storage. The actuator assemblies are implemented as servo motors, solenoids, or piezoelectric components, each mechanically connected to the system frame and electrically linked via driver circuits that receive control signals directly from the processing unit. Communication interfaces are structurally implemented via RF transceivers, on-board Bluetooth or WLAN modules, or Ethernet ports with dedicated controllers, enabling data exchange with external devices. The power management circuitry, consisting of voltage regulators, capacitors, and battery connections, is physically integrated to provide stable operating voltages to each subsystem.The connection of these structural elements via conductor tracks, connectors and buses ensures that the invention functions as a fully hardware-implemented device, with each element having a concrete physical embodiment that enables practical implementation and avoids abstraction.

[0041] The present invention describes a hybrid recommendation system implemented on a computer device that combines the semantic representation performance of Term Frequency-Inverse Document Frequency (TF-IDF) with the community-based insights of collaborative filtering. The system is structured as a modular pipeline, with each module specializing in a specific computational function. The modules are interconnected to ensure seamless data flow and adaptive feedback. The central purpose of the invention is to provide accurate, diverse, and personalized movie recommendations in real time by dynamically balancing semantic similarity scores and collaborative preference signals. This overcomes limitations of existing methods such as cold-start scenarios, data sparseness, and a lack of adaptability to changing user behavior over time.

[0042] The system begins with the metadata ingestion processing unit, which serves as the entry point for film-related descriptions. Metadata such as text summaries, genre identifiers, cast lists, and director names are continuously captured from the content catalog and prepared for further processing. The ingestion module ensures that heterogeneous data sources are normalized into a structured format that can be technically analyzed. For example, free-form plot descriptions can be encoded into a structured string format, while categorical terms such as genre tags are standardized in controlled vocabularies.

[0043] After capture, the descriptive data is passed to the text preprocessing module, which performs a comprehensive set of natural language processing (NLP) operations. Tokenization segments sentences into individual words or part words. Stop word removal eliminates common but semantically uninformative terms such as articles and conjunctions. Stemming and lemmatization reduce words to their base forms to unify variations of the same root term. In an extended implementation, the preprocessing module also includes named entity recognition for extracting proper names such as actor or director names, part-of-speech tagging for capturing the syntactic roles of terms, and phrase recognition for identifying significant multi-part expressions.Preprocessing ensures that the subsequent TF-IDF calculation uses clean, semantically meaningful terms that maximize the discrimination value of the resulting vectors.

[0044] The TF-IDF vectorization unit receives the processed corpus and converts the metadata of each film into a high-dimensional feature vector. For each term in the corpus, a term frequency (TF) value is calculated, indicating its frequency in a given film description. This value is combined with an inverse document frequency (IDF), which measures the rarity of the term in the entire film dataset. Multiplying TF by IDF yields a value that highlights terms that are both frequent in a particular film and relatively rare in the dataset as a whole. This emphasizes features that uniquely characterize a film. The final output of this unit is a sparse matrix of TF-IDF vectors, with each row representing a film and each column representing a weighted term.

[0045] User profiles are created by aggregating the TF-IDF vectors of previously viewed or positively rated films. The aggregated profile vector is compared to the candidate film vectors using cosine similarity, resulting in a semantic relevance score. This score indicates how well the content of a new film semantically matches the user's previous preferences, independent of general community trends.

[0046] In parallel, the collaborative filtering engine processes a user-element interaction matrix, in which each entry reflects either explicit feedback, such as a rating, or implicit feedback, such as replay frequency or completion rate. Depending on the system configuration, the engine can use memory-based methods, such as k-nearest neighbor similarity, or model-based methods, such as matrix factorization. With the memory-based approach, similarity values ​​between users or elements are calculated using cosine similarity or Pearson correlation. With the model-based approach, the system decomposes the user-element matrix into latent factor representations for users and elements using singular value decomposition or probabilistic matrix factorization.The similarity between a user's latent factor vector and that of a candidate film yields a collaborative value that captures the community-driven affinity that may not be directly apparent in the metadata.

[0047] The semantic relevance score of the TF-IDF unit and the collaborative score of the filter engine are transferred to the scoring fusion module. This module performs a normalization process to bring both values ​​to a comparable scale and then integrates them via a dynamically adjustable weighting function. In contrast to static hybrids with fixed weights, the present invention uses an adaptive mechanism in which the weighting coefficients are determined in real time based on contextual factors. For example, for a user with a limited interaction history, the coefficient favors the semantic similarity between TF and IDF, thus ensuring recommendations even with sparse data. Conversely, for a user with an extensive viewing history, the coefficient increasingly emphasizes the collaborative value, thereby leveraging the wealth of crowdsourcing interaction patterns.The weighting function can also take into account the sparse catalog structure, the timing of a film's introduction, and the rate of recorded new interactions, making the system extremely flexible.

[0048] The recommendation output module receives the combined ratings and creates a ranking of the top N films for each user. In certain implementations, the recommendation output module includes a re-ranking mechanism to optimize variety and novelty. This re-ranking level penalizes items that are semantically too similar to previously recommended or viewed items, thus ensuring that the user is not trapped in a narrow recommendation cycle. The re-ranking therefore balances personalization and exploration, allowing users access to both familiar and new content.

[0049] To ensure the system adapts to changing user preferences, the recommendation output module is coupled with a feedback capture interface. This interface continuously monitors user behavior in real time, including playback start time, viewing duration, fast-forwarding, skipping, and completion rates, as well as, where available, explicit rating signals. The captured feedback is forwarded to the TF-IDF and collaborative components. Within the TF-IDF unit, the feedback is used to recalibrate the contribution of terms in the user profile vector. This increases the weighting of terms associated with positively rated films and decreases the weighting of terms associated with abandoned or negatively rated films.In the collaborative engine, feedback updates the user-element interaction matrix or retrains latent factor models to ensure alignment with the most up-to-date user activity patterns.

[0050] An enhanced embodiment of the invention integrates reinforcement learning into the adaptive weighting controller. The system formulates recommendation generation as a sequential decision-making process aimed at maximizing long-term user satisfaction. The reinforcement learning technique adjusts the weighting function of the score fusion controller by observing the reward signal derived from user engagement metrics such as recommendation acceptance rate, average dwell time, and long-term retention. Through iterative updates of its guidelines, the system learns to optimize the balance between semantic similarity and collaborative influence to optimally meet both individual user needs and platform engagement goals.

[0051] The computing device on which the invention is based is designed as a modular hardware architecture. The system utilizes a multi-core processor for executing collaborative filtering procedures, a graphics processing unit (GPU) or tensor processor (TPU) for accelerating TF-IDF vector calculations and matrix multiplications, volatile memory for storing high-dimensional, sparse matrices and latent vectors during ongoing computations, and persistent memory for storing historical user-element data across sessions. Communication interfaces such as Ethernet, WLAN, and API endpoints enable seamless integration with streaming platforms and content management systems, thus ensuring real-time operation of the recommendation system with minimal latency.

[0052] Through this design, the invention creates a robust, scalable, and adaptive recommendation framework that overcomes the limitations of existing systems. By combining semantically rich TF-IDF representations with collaborative, community-developed models, dynamically merging their results with adaptive weights, and continuously refining predictions through real-time user feedback, the system offers a technically superior solution for generating precise, diverse, contextually relevant, and evolving film recommendations that respond to changing user preferences.

[0053] The device of the present invention comprises a central processing module containing both TF-IDF and collaborative filter submodules. The text processing pipeline begins by ingesting raw metadata, performs tokenization, stemming, and stopword removal, and then encodes the processed terms into a sparse TF-IDF matrix representation. This enables the system to highlight unique descriptive terms that distinguish one film from another, thus providing semantically rich embeddings.

[0054] A user profile generator aggregates TF-IDF vectors from previously watched or positively rated films and forms a composite user vector. The semantic similarity between this profile and the vectors of candidate films is calculated using cosine similarity.

[0055] The device's collaborative filtering module uses a user-element matrix representation, where each element denotes a rating, implicit feedback, or interaction strength. To reduce sparsity and detect latent preferences, the system uses matrix factorization or k-nearest neighbor similarity to generate collaborative ratings.

[0056] The score fusion engine then receives both content-based and collaborative scores. A dynamic weighting technique is used, where the system adaptively assigns a higher weight to TF-IDF scores when interaction is limited and gradually strengthens collaborative filtering as user activity increases. Normalization is performed before fusion to ensure comparability between the modalities.

[0057] The recommendation output module evaluates the candidate films and generates a top-N recommendation list. A feedback capture interface is embedded in the device architecture to record user behavior such as viewing time, playback completion rate, and explicit ratings. This data is then iteratively fed back to update both the TF-IDF weights and the collaborative similarity parameters.

[0058] In one embodiment, the device is built as a standalone computing machine and includes a multi-core processor, a dedicated accelerator for natural language processing (GPU or TPU), fast volatile memory for vector computations, persistent memory for user-element matrices, and communication interfaces (Ethernet, WLAN, or API endpoints) for integration with streaming platforms. The machine is housed in a modular enclosure containing a recommendation board, a power supply, and secure data exchange ports. This hardware-based design ensures low-latency processing for real-time use in large streaming environments.

[0059] The present invention relates generally to the field of intelligent information retrieval and recommendation technologies. In particular, the invention relates to a hybrid film recommendation system with a computer device that combines semantic feature extraction using Term Frequency-Inverse Document Frequency (TF-IDF) with collaborative filtering techniques. The invention lies at the interface between natural language processing, machine learning, and the development of recommendation systems and is particularly applicable to online streaming platforms, video-on-demand engines, and infrastructures for the personalization of digital media that require precise and adaptive film recommendations in real time.

[0060] The drawing and the preceding description show examples of embodiments. Those skilled in the art will recognize that one or more of the described elements can be combined to form a single functional element. Alternatively, certain elements can be divided into several functional elements. Elements of one embodiment can be added to another embodiment. For example, the sequence of the processes described here can be changed and is not limited to the manner described here. Furthermore, the actions of a flowchart need not be implemented in the sequence shown; nor does it necessarily have to be performed by all actions. Actions that are not dependent on other actions can also be performed in parallel with the other actions. The scope of the embodiments is in no way limited by these specific examples.Numerous variations are possible, whether explicitly stated in the specification or not, such as differences in structure, dimensions, and material use. The range of embodiments is at least as broad as specified in the following claims.

[0061] Advantages, further benefits, and problem solutions have been described above with reference to specific embodiments. However, the advantages, benefits, problem solutions, and all components that can lead to an advantage, benefit, or solution occurring or becoming more apparent are not to be construed as critical, necessary, or essential features or components of individual or all claims. REFERENCES 100 Device and System for Hybrid Film Recommendations Using the Term Frequency-Inverse Document Frequency (TF-IDF) with Collaborative Filtering. 102 Processing Unit for Metadata Acquisition 104 Text preprocessing module 106 TF-IDF vectorization unit 108 Collaborative Filter Engine 110 Score Fusion Controller 112 Recommendation output module

Claims

[1] A hybrid recommendation system for generating personalized film recommendations, wherein the system includes the following: a processing unit for receiving metadata, configured to receive descriptive data related to films, the descriptive data including plot summaries, genre identifiers, cast lists and director names; a text preprocessing module that is communicatively coupled with the metadata ingestion processing unit and is configured to perform tokenization, stopword removal, stemming, and lemmatization on the descriptive data to obtain a processed corpus; a TF-IDF vectorization unit configured to encode the processed corpus into high-dimensional semantic feature vectors by calculating TF-IDF inverse document frequency values ​​over the entire film dataset, with the semantic feature vectors representing the contextual meaning of terms for individual films; a collaborative filter engine comprising a user-element interaction matrix, wherein the engine is configured to compute latent preference signals using one or more techniques selected from the group consisting of cosine similarity, k-nearest neighbor similarity, and matrix factorization; a score fusion controller coupled to both the TF-IDF vectorization unit and the collaborative filter engine, wherein the module is configured to normalize the semantic feature vectors and the collaborative preference scores and dynamically combine them according to an adaptive weighting coefficient, the coefficient being determined as a function of data density, interaction sparsity, and user history length; and a recommendation output module configured to evaluate candidate films for each user based on the combined score and generate a top-N recommendation list. [2] System according to claim 1, wherein the TF-IDF vectorization unit is further configured to create a user profile vector by aggregating TF-IDF representations of films previously consumed by a particular user, wherein the profile vector is compared with candidate film vectors using cosine similarity to derive a semantic relevance value, and wherein this semantic relevance value is transferred to the rating fusion controller for integration with collaborative preference values. [3] System according to claim 1, wherein the collaborative filter engine is further configured to apply dimensionality reduction by singular value decomposition or probabilistic matrix factorization to the user-element interaction matrix, thereby generating latent factor vectors for both users and elements, wherein the engine calculates the collaborative scores by measuring the similarity of the inner product between latent factor vectors corresponding to a target user and a candidate film. [4] System according to claim 1, wherein the score fusion controller is configured to dynamically adjust weighting coefficients in real time, such that in the case of a sparse user interaction history the weighting coefficient prefers the semantic similarity value derived from the TF-IDF vectorization unit and in the case of a dense user interaction history the weighting coefficient prefers the collaborative value, thereby mitigating cold start problems for both new users and newly introduced films. [5] System according to claim 1, wherein the recommendation output module is further coupled with a feedback capture interface configured to record user behavior in real time, including playback start, playback completion rate, fast forward or skip activity, dwell time and explicit rating results, and wherein the feedback data is iteratively passed to both the TF-IDF vectorization unit for recalibrating term weights and the collaborative filter engine for updating user-element interactions. [6] System according to claim 1, wherein the computing device comprises a modular hardware structure comprising a multi-core central processing unit for performing collaborative filter calculations, a graphics processing unit or tensor processing unit optimized for vector and matrix operations in the TF-IDF vectorization unit, volatile memory for storing sparse TF-IDF matrices and latent factor representations, persistent memory for historical user element data and at least one communication interface, selected from Ethernet, Wi-Fi or application programming interfaces, for integration with streaming platforms. [7] System according to claim 1, wherein the text preprocessing module further includes improvements in natural language processing, including named entity recognition for extracting proper names corresponding to the names of actors or directors, part-of-speech identification for refining the contextual meaning of terms, and bigram or trigram recognition for capturing multi-part expressions in plot summaries, thereby improving the discrimination capability of the TF-IDF vectors. [8] System according to claim 1, wherein the recommendation output module is further configured to use a re-evaluation level based on diversity optimization, the re-evaluation level penalizing elements that are semantically too similar to films already consumed by the user, thereby ensuring that the final recommendation list balances both personalization accuracy and content novelty. [9] System according to claim 1, wherein the score fusion controller is additionally configured to maintain temporal adjustment by incorporating a time decay function for collaborative scores, so that older interactions are attenuated compared to newer interactions, thereby ensuring that the recommendation list reflects evolving user preferences in real time. [10] System according to claim 1, wherein the feedback capture interface also serves to transmit user interaction signals to an adaptive learning controller which modifies the weighting function of the score fusion module, wherein the adaptive learning controller uses reinforcement learning techniques to maximize long-term user satisfaction, which is measured by metrics such as recommendation acceptance rate, engagement duration and retention frequency.

Images