Multimedia information processing methods and related equipment

By allowing users to select multimedia information recommendation models and combining collaborative filtering, content filtering, and hybrid recommendation algorithms, the system solves the problems of uniform recommendation results and cold start in existing systems, achieving personalized and flexible multimedia information recommendation, and improving user experience and system adaptability.

CN123115462APending Publication Date: 2026-11-17LINGNAN UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510626227.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-15
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing multimedia information recommendation systems struggle to flexibly adjust recommendation algorithms based on user needs, resulting in poor recommendation performance for new users or newly released content. Furthermore, the reliance on a single algorithm leads to homogenized recommendation results and a poor user experience.

Method used

This paper provides a multimedia information processing method that allows users to select a multimedia information recommendation model. It combines collaborative filtering, content filtering, and hybrid recommendation algorithms to generate personalized recommendations based on user behavior and content features, enabling users to choose the most suitable recommendation method in different scenarios.

Benefits of technology

It improves the accuracy and flexibility of recommendations, solves the cold start problem, enhances user experience and trust, meets diverse user needs, and expands the system's business potential and application scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115462A_ABST
    Figure CN123115462A_ABST
Patent Text Reader

Abstract

This disclosure provides a multimedia information processing method and related equipment, belonging to the fields of multimedia information processing and machine learning technology. The method includes: displaying a multimedia information recommendation model selection page, the multimedia information recommendation model selection page including multiple multimedia information recommendation model selection controls, each multimedia information recommendation model corresponding to a different multimedia information recommendation model, the multimedia information recommendation model including a target multimedia information recommendation model; responding to a target object's selection operation on the target multimedia information recommendation model selection control among the multiple multimedia information recommendation model selection controls, determining the target multimedia information recommendation model corresponding to the target multimedia information recommendation model selection control, and generating and displaying target multimedia recommendation information for the target object.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure relates to the fields of multimedia information processing and machine learning technology, and more specifically, to a multimedia information processing method, apparatus, electronic device, computer-readable storage medium, and computer program product. Background Technology

[0002] With the rapid development of network technology and the widespread use of personal computers, the internet has become the primary medium for information dissemination. The problem of information overload is becoming increasingly serious, and coupled with unclear user needs, this makes it increasingly difficult for users to find relevant information. For example, with the development of the internet, users are faced with a massive amount of movie resources, making it increasingly difficult to find movies they like. Summary of the Invention

[0003] This disclosure provides a multimedia information processing method, including: displaying a multimedia information recommendation model selection page, the multimedia information recommendation model selection page including multiple multimedia information recommendation model selection controls, different multimedia information recommendation model selection controls corresponding to different multimedia information recommendation models, the multimedia information recommendation model including a target multimedia information recommendation model; responding to a target object's selection operation on the target multimedia information recommendation model selection control among the multiple multimedia information recommendation model selection controls, determining the target multimedia information recommendation model corresponding to the target multimedia information recommendation model selection control, and generating and displaying target multimedia recommendation information for the target object.

[0004] This disclosure provides a multimedia information processing apparatus, comprising: a display unit for displaying a multimedia information recommendation model selection page, the multimedia information recommendation model selection page including multiple multimedia information recommendation model selection controls, each multimedia information recommendation model corresponding to a different multimedia information recommendation model, the multimedia information recommendation model including a target multimedia information recommendation model; and a processing unit for responding to a target object's selection operation of a target multimedia information recommendation model selection control among the multiple multimedia information recommendation model selection controls, determining the target multimedia information recommendation model corresponding to the target multimedia information recommendation model selection control, and generating and displaying target multimedia recommendation information for the target object.

[0005] This disclosure provides an electronic device, including: one or more processors; and a storage device configured to store one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the method described in any embodiment of this disclosure.

[0006] This disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when run on a computer, causes the computer to execute the methods described in any embodiment of this disclosure.

[0007] This disclosure provides a computer program product, including a computer program that, when executed by a computer, implements the methods described in this disclosure. Attached Figure Description

[0008] Figure 1 This is a flowchart of a multimedia information processing method according to an exemplary embodiment of the present disclosure.

[0009] Figure 2 This is a schematic diagram of a multimedia information recommendation model selection page according to an exemplary embodiment of this disclosure.

[0010] Figure 3 This is a schematic diagram of a multimedia information recommendation model selection page according to another exemplary embodiment of this disclosure.

[0011] Figure 4 This is a schematic diagram of a recommended content display page according to an exemplary embodiment of this disclosure.

[0012] Figure 5 This is a schematic diagram of a movie recommendation system according to an exemplary embodiment of the present disclosure.

[0013] Figure 6 This is a system block diagram of a movie recommendation system according to an exemplary embodiment of the present disclosure.

[0014] Figure 7 This is a flowchart of a movie recommendation method according to an exemplary embodiment of this disclosure.

[0015] Figure 8 This is an architecture diagram of a personalized movie recommendation system according to an exemplary embodiment of this disclosure.

[0016] Figure 9 This is an algorithm structure diagram of a recommendation system according to an exemplary embodiment of the present disclosure.

[0017] Figure 10 This is a schematic diagram illustrating the collection of user preference information for movie recommendations, an exemplary embodiment of this disclosure.

[0018] Figure 11 This is a flowchart of a processing method based on CF (collaborative filtering) according to an exemplary embodiment of this disclosure.

[0019] Figure 12This is a flowchart illustrating a processing method based on CBF (content-based filtering) according to an exemplary embodiment of this disclosure.

[0020] Figure 13 This is a flowchart illustrating the processing of HRA (hybrid recommender algorithm) based on CF and CBF, an exemplary embodiment of this disclosure.

[0021] Figure 14 This is a schematic diagram of the structure of a multimedia information processing apparatus according to an exemplary embodiment of the present disclosure.

[0022] Figure 15 A schematic structural diagram of an electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this disclosure more apparent, exemplary embodiments according to this disclosure will now be described in detail with reference to the accompanying drawings. In the drawings, the same reference numerals denote the same elements throughout. It should be understood that the embodiments described herein are merely illustrative and should not be construed as limiting the scope of this disclosure.

[0024] In this disclosure, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented, wholly or partially, using software, hardware (such as processing circuitry or storage), or a combination thereof. Similarly, a processor (or multiple processors or storage) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0025] In this disclosure, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance; the term "multiple" refers to two or more unless otherwise expressly defined. The terms "install," "connect," "link," and "fix" should be interpreted broadly. For example, "connect" can be a fixed connection, a detachable connection, or an integral connection; "link" can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this disclosure according to the specific circumstances.

[0026] In the description of this specification, the terms "one embodiment," "some embodiments," "specific embodiment," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0027] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

[0028] Figure 1 This is a flowchart of a multimedia information processing method according to an exemplary embodiment of the present disclosure. Figure 1 The multimedia information processing method provided in the embodiments can be executed by any electronic device / computer device, such as a terminal and / or a server, and this disclosure does not limit it. Figure 1 As shown, the method provided in this embodiment includes the following steps.

[0029] In S110, a multimedia information recommendation model selection page is displayed. The multimedia information recommendation model selection page includes multiple multimedia information recommendation model selection controls. Different multimedia information recommendation model selection controls correspond to different multimedia information recommendation models. The multimedia information recommendation model includes a target multimedia information recommendation model.

[0030] In this embodiment, the multimedia information recommendation model selection page refers to a page that allows users to independently choose which multimedia information recommendation model to use for recommending multimedia information. It can be a subpage on a main page, an independent page, or a floating window on a main page; this disclosure does not limit its scope. For example, users can control the touchscreen through clicks, voice input, or gestures to select a multimedia information recommendation model on the selection page; this disclosure does not limit the method by which the user selects the multimedia information recommendation model. For example, the selected multimedia information recommendation model is referred to as the target multimedia information recommendation model.

[0031] In this embodiment of the disclosure, multiple (two or more) multimedia information recommendation model selection controls are displayed on the multimedia information recommendation model selection page. Each multimedia information recommendation model selection control is used to indicate the corresponding multimedia information recommendation model. When a user triggers a multimedia information recommendation model selection control, it is determined that the user has selected the corresponding multimedia information recommendation model as the target multimedia information recommendation model.

[0032] In this embodiment of the disclosure, the multimedia information recommendation model refers to an algorithm, network, etc., used to recommend personalized multimedia information to users. The multimedia information recommendation model can recommend multimedia information that the user is interested in based on the multimedia information to be recommended and the user's personalized characteristics. The multimedia information recommended to the user is called the target multimedia recommendation information.

[0033] The multimedia information in this embodiment may include one or more of the following: text, video, image, and audio.

[0034] In an exemplary embodiment, the method provided by this disclosure further includes: displaying corresponding indication information at a position associated with the multimedia information recommendation model selection control on the multimedia information recommendation model selection page, wherein the indication information is used to indicate the basis for the corresponding multimedia information recommendation model to generate multimedia recommendation information.

[0035] In some embodiments, the multimedia information recommendation model selection page not only displays multiple multimedia information recommendation model selection controls, but also displays corresponding indication information at an associated location for each multimedia information recommendation model selection control. This explicitly informs the user of the logic and basis followed by each multimedia information recommendation model in recommending multimedia information to them. The associated location simply indicates which multimedia information recommendation model's recommendation basis the indication information corresponds to; this disclosure does not limit the specific location.

[0036] In an exemplary embodiment, the plurality of multimedia information recommendation model selection controls include a collaborative filtering model selection control, a content filtering model selection control, and a hybrid recommendation model selection control, wherein the multimedia information recommendation model includes a collaborative filtering model, a content filtering model, and a hybrid recommendation model.

[0037] The indication information displayed at the location associated with the collaborative filtering model selection control is used to instruct the generation of multimedia recommendation information based on the analysis of the similarity between objects based on object behavior data.

[0038] The indication information displayed at the location associated with the content filtering model selection control is used to indicate the generation of multimedia recommendation information based on multimedia content features.

[0039] The indication information displayed at the location associated with the hybrid recommendation model selection control is used to instruct the generation of multimedia recommendation information by combining the collaborative filtering model and the content filtering model.

[0040] In the following embodiments, the multimedia information recommendation model selection page includes three multimedia information recommendation model controls, and the three multimedia information recommendation models corresponding to these three multimedia information recommendation model controls are collaborative filtering model, content filtering model and hybrid recommendation model, respectively, for illustrative purposes. However, this disclosure is not limited to this. The multimedia information recommendation model selection page may include more or fewer multimedia information recommendation model controls, and the corresponding multimedia information recommendation models may be the same as or different from the above-mentioned collaborative filtering model, content filtering model and hybrid recommendation model.

[0041] The collaborative filtering model in this disclosure refers to a recommendation model based on the collaborative filtering algorithm, also known as CF or CF model. For example, the CF model refers to predicting an individual user's (e.g., the target object's) interest preferences for untouched items (e.g., unviewed multimedia information) by analyzing the similarity between users or items (e.g., multimedia information) and utilizing group behavior data (e.g., the object behavior data of the target object's nearest neighbors), thereby achieving personalized recommendations.

[0042] The content filtering model in this disclosure refers to a recommendation model based on a content-based filtering (CBF) algorithm. For example, the CBF model determines whether to recommend multimedia information to the user by detecting and analyzing the characteristics of the multimedia content.

[0043] The CF (Conversion Filtering) model relies on user behavior data regarding multimedia information (such as ratings, clicks, browsing, viewing, and comments), while the CBF (Conversion Filtering) model relies on the features of the multimedia information itself. The hybrid recommendation model is also known as HRA (Hybrid Recommendation Analysis) or the HRA model. For example, the HRA model combines multimedia recommendation information generated by both the CF and CBF models to produce the final multimedia recommendations for the user.

[0044] In this embodiment of the disclosure, object behavior data refers to interactive data and feedback data related to the interaction between the user and multimedia information, such as one or more of the following: user rating, click, view, comment, add to favorites, feedback like or dislike of multimedia information.

[0045] In this embodiment of the disclosure, the multimedia content features refer to the features of the content itself contained in the multimedia information, such as one or more of the following: title, description, director, actors, plot, and genre of the multimedia information.

[0046] In an exemplary embodiment, the method provided in this disclosure further includes: displaying a registration page, the registration page including a basic information input control and a multimedia preference information input control; obtaining target basic information of the target object in response to an input operation of the target object on the basic information input control; and obtaining target multimedia preference information of the target object in response to an input operation of the target object on the multimedia preference information input control.

[0047] The registration page provided in this embodiment includes not only basic information input controls but also multimedia preference information input controls, so that basic or preliminary preference information of the user regarding multimedia information can be obtained during the registration phase. For example, the target multimedia preference information includes the basic or preliminary preference information regarding multimedia information input by the target user during the registration phase. This can be used to provide personalized multimedia information recommendations to new users, solving the cold start problem.

[0048] In this embodiment of the disclosure, the basic information input control can be used to input the target object's basic information, which refers to personal information related to the target object, such as age, gender, location, name, and mobile phone number. Exemplarily, the target object can input the basic information by operating the basic information input control through any one of typing, voice input, or gesture input; this disclosure does not limit this method.

[0049] In this embodiment of the disclosure, the multimedia preference information input control can be used to input the target multimedia preference information of the target object. Target multimedia preference information refers to the target object's preference information for multimedia information, such as which types of multimedia information it is interested in, which directors it directs, and which actors it stars in. During the registration phase, only basic or preliminary preference information can be input. The system can continuously adjust the target multimedia preference information based on the increasing amount of collected behavioral data about the target object.

[0050] In an exemplary embodiment, the method provided by this disclosure further includes: displaying a login page, the login page including a login control; obtaining the identity information of the target object in response to the login operation of the target object on the login control; verifying the identity information of the target object; after the identity information of the target object is verified, obtaining page operation behavior data of the target object within a predetermined time after login; and obtaining target multimedia preference information of the target object based on the page operation behavior data of the target object within the predetermined time after login.

[0051] For example, this disclosure also provides a login page. Through the login control on the login page, a target object that has successfully registered through the aforementioned registration page can enter its identity information, such as the username and password used during registration, to verify its identity. After successful identity verification, the target object's short-term page behavior after login (i.e., page operation behavior data within a predetermined time period), such as browsing multimedia information shortly after login, can be obtained to predict the target object's target multimedia preference information.

[0052] In an exemplary embodiment, the method provided in this disclosure further includes: displaying a login page, the login page including a login control; obtaining the identity information of the target object in response to the login operation of the target object on the login control; verifying the identity information of the target object; displaying different types of multimedia information after the identity information of the target object is verified; and obtaining the target multimedia preference information of the target object based on the target object's selection operation on the displayed different types of multimedia information.

[0053] For example, after a target user successfully logs in, the system can automatically display various types of multimedia information to that user. This allows the system to obtain the user's feedback data or selection actions on these different types of multimedia information, thereby determining the user's target multimedia preferences. For instance, if the system simultaneously displays martial arts films and comedies, and the user chooses to watch the comedy, it can predict that the user is interested in comedies. Alternatively, feedback controls, such as "like" or "dislike," can be displayed in the relevant locations of the recommended multimedia information. The user can click on different feedback controls to indicate whether they like or dislike the recommended multimedia information.

[0054] In some embodiments, the target multimedia preference information of a new user can be obtained through the registration page. In other embodiments, the target multimedia preference information of a new user can be predicted based on their short-term page behavior after successful login. In still other embodiments, different types of multimedia information can be displayed to the new user after successful login, and feedback data on these different types of multimedia information can be obtained to predict the new user's target multimedia preference information. In yet another embodiment, two or three of the aforementioned methods can be combined to obtain the new user's target multimedia preference information.

[0055] In S120, in response to the target object's selection operation of the target multimedia information recommendation model selection control among the multiple multimedia information recommendation model selection controls, the target multimedia information recommendation model corresponding to the target multimedia information recommendation model selection control is determined, and target multimedia recommendation information for the target object is generated and displayed.

[0056] In this embodiment of the disclosure, the target object can be any user.

[0057] The method provided in this disclosure provides multiple multimedia information recommendation model selection controls on a multimedia information recommendation model selection page, allowing users to independently select a target multimedia information recommendation model to generate and display target multimedia recommendation information, thereby enabling more targeted multimedia information recommendations.

[0058] In an exemplary embodiment, determining the target multimedia information recommendation model corresponding to the selection control using the target multimedia information recommendation model, and generating and displaying target multimedia recommendation information for the target object, includes: when the target multimedia information recommendation model includes a content filtering model, extracting multimedia content features of the multimedia information; converting the multimedia content features into multimedia feature vectors; obtaining multimedia similarity between different multimedia information based on the multimedia feature vectors; and generating and displaying target multimedia recommendation information for the target object based on the target multimedia preference information of the target object and the multimedia similarity between different multimedia information.

[0059] In this embodiment of the disclosure, multimedia content features refer to features related to the content of multimedia information, such as one or more related to the title, description, author, plot, and genre of the multimedia information. When a user selects the CBF model for multimedia information recommendation, the system can construct multimedia similarity between different multimedia information based on multimedia content features, and then use the target object's target multimedia preference information and the multimedia similarity between different multimedia information to recommend personalized multimedia information to the target object.

[0060] In an exemplary embodiment, determining the target multimedia information recommendation model corresponding to the selection control using the target multimedia information recommendation model, and generating and displaying target multimedia recommendation information for the target object, includes: when the target multimedia information recommendation model includes a collaborative filtering model, establishing an object multimedia matrix based on object behavior data; performing matrix decomposition on the object multimedia matrix to obtain an object feature matrix; performing local sensitive hashing on the object feature vectors in the object feature matrix to obtain the nearest neighbor objects of the target object; and generating and displaying target multimedia recommendation information for the target object based on the object behavior data of the nearest neighbor objects.

[0061] In this embodiment of the disclosure, object behavior data refers to interaction data between an object and multimedia information, feedback data of the object to the multimedia information, etc. An object multimedia matrix can be established based on object behavior data. The elements in this object multimedia matrix reflect the object behavior data between the object and the multimedia information. For example, each element in the same row indicates the rating data of a certain object for a certain multimedia information. This rating data can include not only the object's direct rating of the multimedia information, such as a score of 4 out of 5, but also other interaction data and feedback data between the object and the multimedia information, such as the object's likes, favorites, adding to favorites list, rating, and viewing time.

[0062] When a user selects the CBF model for multimedia information recommendation, the system can use object behavior data to obtain an object feature matrix, and based on the object feature matrix, obtain the nearest neighbor objects of the target object. The nearest neighbor objects are those that have similar multimedia preference information to the target object. Then, based on the object behavior data of the nearest neighbor objects, the system can predict the target multimedia information that the target object is interested in and recommend it to the target object.

[0063] Most recommendation systems in related technologies focus on a specific domain and are difficult to apply widely to other multimedia content recommendation fields, thus limiting their commercial potential and development space. The methods, systems, and apparatus provided in this disclosure are suitable for flexible application scenarios. They are not only applicable to online movie platforms but can also be widely applied to other multimedia content recommendation fields, such as music, books, and television programs. By flexibly configuring recommendation algorithms / multimedia information recommendation models, the system can adapt to different content types and user needs, providing more comprehensive and personalized recommendation services. The system is not limited to movie recommendations but can be extended to other multimedia content recommendation fields, possessing broad application prospects and commercial potential. Its wide applicability and powerful recommendation functions give the system a significant competitive advantage in the market, attracting more users and partners, further expanding its commercial value and development space. The following examples illustrate the application of the solutions provided in this disclosure to online movie platforms for personalized movie recommendations to users, but this disclosure is not limited to this.

[0064] Figure 2 This is a schematic diagram of a multimedia information recommendation model selection page according to an exemplary embodiment of this disclosure. Figure 2 As shown, the multimedia information recommendation model selection page 200 of this embodiment provides user-selectable multi-mode recommendations, such as multimedia information recommendation model selection controls 201 to 20n, where n is a positive integer greater than 1.

[0065] Recommender systems in related technologies typically rely on a single algorithm, making it difficult to meet the diverse needs of different users. This results in overly concentrated recommendation results and a poor user experience. The solution provided in this disclosure allows users to select the most suitable recommendation algorithm / multimedia information recommendation model (e.g., content-based filtering algorithms, collaborative filtering, or hybrid recommendation) according to their personal preferences. This flexibility not only improves the accuracy of recommendations but also enhances user engagement and satisfaction.

[0066] In some embodiments, this disclosure provides user-selectable multi-mode recommendations, allowing users to choose the most suitable recommendation algorithm (e.g., content-based filtering, collaborative filtering, or hybrid recommendation) according to their personal preferences. Users can freely select the recommendation algorithm type / multimedia information recommendation model in the settings page (which may be an example of a multimedia information recommendation model selection page, or a settings page may include a multimedia information recommendation model selection page), and change their selection at any time. The system will respond instantly to the user's preference adjustments and update the recommendation results (i.e., the target multimedia recommendation information). This flexibility is not found in recommendation systems of related technologies. Most recommendation systems in related technologies rely on only a single algorithm, failing to meet the diverse needs of different users. By allowing users to choose their own recommendation algorithm / multimedia information recommendation model, the system not only improves the accuracy of recommendations but also enhances user engagement and satisfaction. Supporting multiple recommendation algorithms enables the system to better adapt to the personalized needs of different users, providing more flexible recommendation services. This method significantly improves the user experience, solves the problem of monotonous recommendation results caused by a single algorithm in recommendation systems of related technologies, and enhances user stickiness, engagement, satisfaction, and loyalty.

[0067] With the development of the internet, users are faced with a massive amount of movie resources, making it increasingly difficult to find movies they like. Recommendation systems in related technologies often rely on user behavior data (such as ratings and reviews) while neglecting the content characteristics of the movies themselves (such as actors, directors, and genres). Therefore, recommendation mechanisms in these technologies often perform poorly when dealing with new users or newly released movies because they lack sufficient user interaction data to generate effective recommendations. Furthermore, recommendation systems in these technologies typically employ only one recommendation algorithm, failing to meet the diverse needs of different users. Specifically, for new users or newly released movies, the lack of sufficient user behavior data makes it difficult for traditional recommendation systems to provide effective recommendations, leading to the cold start problem. At the same time, a single recommendation algorithm may result in overly concentrated recommendation results, failing to fully meet the diverse needs of users and impacting user experience. While complex models such as deep learning may be more accurate, their internal workings are opaque, making it difficult for users to understand the basis for recommendations and reducing trust.

[0068] Figure 3This is a schematic diagram of a multimedia information recommendation model selection page according to another exemplary embodiment of this disclosure. Figure 3 As shown, the multimedia information recommendation model selection page 300 may include a collaborative filtering model selection control 301, a content filtering model selection control 302, and a hybrid recommendation model selection control 303, for users to select the collaborative filtering model, the content filtering model, and the hybrid recommendation model respectively.

[0069] Continue to refer to Figure 3 The collaborative filtering model selection control 301 displays the prompt message "Recommend movies based on the similarity between users by analyzing user behavior data" 3011. The content filtering model selection control 302 displays the prompt message "Recommend movies based on the characteristics of movie content (such as genre, director, actors, etc.)" 3021. The hybrid recommendation model selection control 303 displays the prompt message "Based on the advantages of CBF and CF, the recommendation results are combined" 3031.

[0070] This disclosure combines three main recommendation algorithms—content-based filtering, collaborative filtering, and hybrid recommendation—and allows users to choose the most suitable recommendation method based on their preferences. This solves the cold start problem, the uniformity of recommendation results, and the poor interpretability of related recommendation systems, improving the flexibility and accuracy of the recommendation system and enhancing the user experience. This disclosure also provides a multi-modal personalized movie recommendation system and its implementation method that supports user selection of recommendation algorithms.

[0071] In related technologies, recommendation systems often suffer from a cold start problem when dealing with new users or newly released movies due to a lack of sufficient user behavior data, making it difficult to provide effective recommendations. Furthermore, a single recommendation algorithm may lead to overly concentrated recommendation results, failing to fully meet the diverse needs of users. This disclosure combines multiple recommendation algorithms: content-based filtering, collaborative filtering, and hybrid recommendation. Content-based filtering generates recommendations by analyzing movie text features; collaborative filtering algorithms / models construct recommendations based on user behavior data; and hybrid recommendation algorithms combine the advantages of both, ensuring that recommendations are both relevant to user interests and diverse. By combining the strengths of different algorithms / models, the system can more accurately capture user interests and provide high-quality recommendation results. The hybrid recommendation algorithm avoids the result concentration problem that may result from a single algorithm, providing more diverse recommendations to meet the diverse needs of users. For new users or newly released movies, the system can quickly provide reasonable recommendations by directly analyzing movie content features, reducing the impact of the cold start and ensuring the timeliness and relevance of the recommendation results. This not only solves the cold start problem but also provides more diverse recommendations and avoids the result concentration problem that may result from a single algorithm.

[0072] While complex models like deep learning may be more accurate, their internal workings are opaque, making it difficult for users to understand the recommendation criteria and reducing trust. This disclosure provides a transparent and explainable recommendation process. This disclosure maintains a high degree of transparency in the recommendation process, allowing users to clearly understand the recommendation criteria. The recommendation logic is intuitive and easy to understand, increasing user trust in the system. Users can freely choose content-based filtering algorithms, collaborative filtering algorithms, or hybrid recommendation algorithms. Users can clearly understand the recommendation criteria; for example, under a content-based filtering algorithm, it might state that the recommendation is based on the user's movie preferences; under a collaborative filtering algorithm, it might state that the recommendation is based on the user's web browsing behavior; and under a hybrid recommendation algorithm, it might state that it combines the two factors for a comprehensive recommendation. The transparent recommendation process and intuitive, easy-to-understand logic increase user trust and willingness to use the system. By understanding the logic behind the recommendations, users can better understand their interests and preferences, improving their cognitive level. The transparent recommendation process makes users more accepting of the recommendation results, improving overall user satisfaction and experience quality. This transparency not only enhances the user experience but also provides users with a more reliable reference, increasing their trust in the system.

[0073] Figure 4 This is a schematic diagram of a recommended content display page according to an exemplary embodiment of this disclosure.

[0074] like Figure 4 As shown, the recommended content display page 400 can display target multimedia recommendation information 404 for the target audience. For example, the recommended content display page 400 may also include a CF model selection control 401, a CBF model selection control 402, and a hybrid recommendation model selection control 403, so that when viewing the target multimedia recommendation information 404, the selected target multimedia information recommendation model can be changed at any time according to personal preference. For example, assuming the CF model is initially selected to generate target multimedia recommendation information, the CBF model selection control 402 can be clicked to switch to the CBF model to generate updated target multimedia recommendation information.

[0075] Most contemporary people enjoy watching movies, which can help them relax. However, there are many different types of movies available today, with hundreds of new films released domestically and internationally every year. Therefore, finding a favorite movie among these dazzling choices requires a significant amount of time and effort. Recommendation systems in related technologies cannot adjust their recommendation algorithms based on user preferences. Therefore, embodiments of this disclosure propose a movie recommendation system and method that allows users to choose according to their own needs. The system utilizes user data and preferences, combined with movie characteristics, and employs recommendation algorithm techniques. Exemplarily, the system is primarily based on collaborative filtering (CF) and content-based filtering (CBF), utilizing the MovieLens 1M dataset. By calculating user similarity (i.e., similarity between objects) and movie similarity (i.e., multimedia similarity between different multimedia information), the system generates recommendations for users. This helps recommend movies that users may be interested in, thereby improving efficiency and reducing the time spent searching for movies. Embodiments of this disclosure provide a novel, efficient, and personalized movie recommendation system based on user selection algorithms and dynamic hybrid optimization.

[0076] Figure 5 This is a schematic diagram of a movie recommendation system according to an exemplary embodiment of this disclosure. Figure 5 As shown, the system provided in this embodiment includes a user registration interface 500 (i.e., a registration page), which can be used to allow users to fill in basic information 501 to obtain the target basic information of the target object; exemplarily, it may also include a user questionnaire 502 to obtain the target multimedia preference information of the target object; exemplarily, it may also include a registration button 503; exemplarily, it may also include an existing account link 504.

[0077] refer to Figure 5 The system may also include a user login interface 600 (i.e., a login page), which can be used by the user to log in by entering the username and password used during registration 601; exemplarily, it may also include a login button 602; exemplarily, it may also include a forget password link 603.

[0078] Continue to refer to Figure 5 The system may also include a recommendation model selection page 700 (i.e., a multimedia information recommendation model selection page), which may include a collaborative filtering model selection control 701, a content filtering model selection control 702, and a hybrid recommendation model selection control 703; for example, it may also include a user favorites and likes list 704.

[0079] Continue to refer to Figure 5The system may also include a recommended content display page 800, which can be used to display a personalized list of recommended movies as the target multimedia information recommendation model based on the recommendation model selected by the user. For example, each recommended movie includes basic information such as title, rating, genre, and synopsis. For example, clicking on the movie displays its detailed information.

[0080] Figure 5 In this embodiment, the system can also collect and provide feedback on user behavior. For example, the system continuously records user behavior data as object behavior data. For example, when watching a movie, a user can provide feedback such as liking or disliking, or give a rating or review. This feedback is collected in real time and used to update the user's preference model, further optimizing recommendations.

[0081] like Figure 5 As shown, the system supports 1000 interactions and feedback. On each page, users can see operation prompts and feedback information to ensure clarity of operations. The system dynamically adjusts its recommendation strategy based on feedback behavior data to optimize recommendation results.

[0082] This disclosure provides a multi-modal personalized movie recommendation system that supports user-selected recommendation algorithms. By combining three main recommendation algorithms—content-based filtering, collaborative filtering, and hybrid recommendation—it addresses the cold start problem, the uniformity of recommendation results, and the poor interpretability of related technologies. The following provides illustrative examples of the system architecture, workflow, algorithm implementation, user interface design, and overall flowchart.

[0083] The system provided in this disclosure is built using the Django framework and integrates various machine learning and data processing libraries from the Python ecosystem (such as scikit-learn, Pandas, NumPy, etc.) to achieve efficient data processing and recommendation algorithm integration. Figure 6 As shown, the system includes a front-end user interface module 1, a back-end recommendation engine module 2, a data management module 3, a continuous learning mechanism module 4, an electronic device 5, and a computer-readable storage medium 6.

[0084] For example, the front-end user interface module 1 is the user interface for interacting with the system. Its main functions include, but are not limited to, viewing recommendations, registering / login, watching, rating, and browsing. New users fill in basic information (i.e., target basic information) and their movie preferences (as target multimedia preference information) during registration. After registration, they can perform normal rating and browsing operations. As user interaction with the system increases, the system continuously records user behavior information, gradually understanding the user's interests. Users can rate their favorite movies, and the system stores these ratings and user browsing behavior in logs and calls the data analysis engine for analysis to update the user model, ensuring that recommendations are more aligned with user preferences.

[0085] For example, the backend recommendation engine module 2, as the core of the entire recommendation system, employs advanced content-based filtering algorithms, collaborative filtering, and hybrid recommendation methods to calculate the similarity between users and movies, thereby identifying nearest neighbor data (in collaborative filtering algorithms, the similarity between users is calculated; in content-based filtering algorithms, the similarity between movies is calculated), and storing it in the database. This nearest neighbor data (including nearest objects and / or nearest movies) provides a direct source of information for personalized recommendations. When a change in a user's interests is detected—for example, a user who previously liked horror movies suddenly starts liking, saving, and commenting on comedy or other types of movies; or when their page browsing behavior shows a decrease in attention to horror movies and an increase in attention to other types of movies—the system can promptly capture and adjust the recommendation results.

[0086] Specifically, by monitoring user behavior data in real time (such as likes, favorites, comments, and views), the system captures changes in user interests. For example, if a user shifts from a long-standing preference for horror movies to comedies or other genres, the system can identify this change in user interest. The system then makes corresponding adjustments to the CF and CBF algorithms to address these changes. In the collaborative filtering algorithm, the system updates the user-item matrix (i.e., the object multimedia matrix, where items refer to multimedia information such as movies), adjusting the similarity relationship between the user and movies based on the user's latest behavioral data (such as ratings and viewing history), thereby optimizing subsequent recommendation results. In the content-based filtering algorithm, the system updates the user's movie preference information, updating the user's preference model (i.e., target multimedia preference information) based on the user's interactions with specific genres, directors, or actors, ensuring that the recommendation results match the user's latest interests. In this way, the system can flexibly adjust its recommendation strategy, ensuring that the recommendation results always keep pace with changes in user interests. For example, considering that this module requires significant computing resources, it can be run offline in the background to improve efficiency.

[0087] For example, the data management module 3 is responsible for managing and maintaining movie data, including functions such as category management, ranking lists, basic information management, and core recommendations. Category management and ranking lists are designed to help users find movies of interest more easily, serving as a supplement to personalized recommendations. Furthermore, the movie information management module constructs a detailed movie information model and creates various types of ranking lists, such as popular movie rankings, high-scoring blockbuster rankings, and box office rankings. These comprehensively analyze online data and the statistical data of this webpage (approximately 70% and 30% respectively, ensuring that recommended movies are as high-quality as possible with high overall ratings). For example, popular movies are calculated based on the number of views and views, while high-scoring blockbusters are selected from films with high average ratings. The ranking lists in this module are generated based on a combination of statistical data from all users and behavioral data from individual users. Specifically, the ranking lists first generate general ranking lists based on the behavioral data of all users, such as popular movie rankings and high-scoring blockbuster rankings. These ranking lists reflect the overall trends in big data and provide universally applicable recommendations. Then, the system combines the historical behavioral data of individual users to weight and adjust the ranking lists, thereby providing personalized recommendations for each user. While the leaderboard is based on data from all users, the leaderboard that each user ultimately sees will differ based on their individual interests and behavioral data, ensuring that the recommended content better matches the user's preferences.

[0088] For example, the system has a continuous learning mechanism that can continuously update the user model and movie information based on real-time user feedback. The system regularly conducts in-depth analysis of user behavior data (i.e., object behavior data) such as click-through rate, viewing time, rating, comments, favorites, and sharing, and evaluates the recommendation effect in combination with real-time feedback. Based on the analysis results, the system adjusts the recommendation strategy in a targeted manner and optimizes the user model and recommendation algorithm to ensure that the system can provide optimal performance at different times.

[0089] For example, if the system detects that a user has recently been frequently watching and rating comedies, it can adjust the user model to increase the user's preference weight for comedies, thereby optimizing the recommendation algorithm and recommending more comedies. Similarly, if the system finds that a user is gradually losing interest in a certain type of movie, such as horror, it reduces the proportion of that type of movie in recommendations and increases the recommendations of new types of movies that the user might be interested in. In addition, the system will regularly or in real-time update the user-item matrix, similarity matrix (a matrix generated based on the multimedia similarity between different multimedia information), and other key indicators. Simultaneously, during the update process, the system utilizes incremental learning and mini-batch retraining techniques to not only quickly capture changes in user interests but also promptly incorporate information on newly released movies. By regularly or in real-time updating the user-item matrix, similarity matrix, and other key indicators, the system further improves the accuracy and diversity of recommendation results, thereby enhancing the overall user experience and platform activity.

[0090] Recommendation systems in related technologies lack dynamic optimization mechanisms, making it difficult for recommendation results to keep up with the latest trends and changes in user interests, leading to a gradual decline in recommendation effectiveness. This disclosure addresses this issue by employing a continuous learning mechanism to periodically analyze metrics such as user click-through rates and viewing duration to evaluate recommendation results and adjust recommendation strategies accordingly, ensuring optimal system performance. Recommendation results consistently reflect the latest trends and changes in user interests, providing more intelligent and personalized services. Through continuous optimization of recommendation strategies, the system improves recommendation accuracy, ensuring a high degree of match between recommendation results and actual user interests. The continuous learning mechanism encourages greater user interaction, promoting higher user engagement and platform activity. This dynamic optimization not only enhances user experience but also attracts a more active user base, driving the platform's long-term development.

[0091] This disclosure also provides an electronic device, including a storage device and a processor, and a computer program stored on the storage device and executable on the processor, wherein the processor executes the computer program to implement the method described above.

[0092] This disclosure also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the above-described method, such as a movie recommendation method.

[0093] To address the issue of homogeneous recommendation results, the solution provided in this disclosure integrates multiple recommendation algorithms / models, improving the system's flexibility and adaptability, and ensuring that recommendation results are both aligned with user interests and diverse. Furthermore, the recommendation process is transparent, allowing users to clearly understand the basis for recommendations, thus enhancing their trust in the system. This intelligent and personalized recommendation service not only saves users time and enhances user loyalty but also promotes higher user engagement, bringing a more active user base and greater commercial value to multimedia content providers. Simultaneously, the system possesses a continuous learning mechanism, constantly updating user models and movie information based on real-time user feedback, ensuring that recommendation results always keep pace with the latest trends and changes in user interests, improving the dynamism and timeliness of recommendation effectiveness. This not only solves the cold start problem, homogeneous recommendation results, and poor interpretability in recommendation systems but also improves the system's flexibility and accuracy by integrating multiple recommendation algorithms, significantly enhancing the user experience.

[0094] Figure 7 This is a flowchart of an exemplary embodiment of a movie recommendation method disclosed herein. Figure 7 As shown, the method provided in this embodiment includes the following steps.

[0095] S701, User Login / Registration.

[0096] For example, users can register or log in to the system, fill in basic information such as age, gender, occupation, etc., and select movie preferences to obtain target multimedia preference information such as movie type, language, director, actors, etc., and complete the initial setup.

[0097] S702, the user selects the recommended method.

[0098] For example, the system provides an intuitive user interface where users can select the recommendation algorithm type on the settings page, choosing their current preferred recommendation algorithm type as the target multimedia information recommendation model: content-based filtering algorithm, collaborative filtering algorithm, or hybrid recommendation algorithm.

[0099] S703, the specific implementation of the recommendation algorithm includes the following three types: content-based filtering algorithm, collaborative filtering algorithm, and hybrid recommendation algorithm. Content-based filtering algorithms mainly extract content features, calculate similarity, and generate a recommendation list based on these features. Its core idea is to find items similar to content that the user has previously been interested in, thereby recommending similar content. Collaborative filtering algorithms construct a matrix of users and items, calculate the similarity between users or items, and generate a recommendation list based on these similarities. Its basic principle is to use the behavior or preferences of other users to provide recommendations to the target user. Hybrid recommendation algorithms combine the strategies of content filtering and collaborative filtering, integrating their advantages and optimizing the recommendation results through weight adjustments. This fully leverages the strengths of both algorithms, generating more accurate recommendation results to be displayed to the user. For example, based on the selected recommendation algorithm, the system generates a recommendation list as target multimedia recommendation information and displays it to the user through the user interface.

[0100] For example, content filtering algorithms include feature extraction: extracting content features from the movie preferences filled in by the user during registration and from similar movies the user likes (such as science fiction movies, comedy movies, etc.) as multimedia content features, such as movie genre, title, director, actors, language, synopsis, tags, user ratings, and review content. For example, TF-IDF (Term Frequency–Inverse Document Frequency) vectorization technology is used to convert text into numerical feature vectors as multimedia feature vectors. For example, the multimedia feature vector TF-IDF(t,d) can be obtained through the following formula:

[0101] TF-IDF(t,d)=TF(t,d)×IDF(t)(1)

[0102] In the above formula, TF(t,d) represents the term frequency of term t in document d (such as movie text, which is text used to describe the content features of a movie). IDF(t) represents the inverse document frequency of term t.

[0103] For example, TF(t,d) can be calculated using the following formula:

[0104]

[0105] In the above formula, n t,d This represents the number of times term t appears in document d; ∑ k n k,dThis represents the total number of occurrences of all terms in document d (i.e., the total number of words in the document).

[0106] For example, IDF(t) can be calculated using the following formula:

[0107]

[0108] In the above formula, N represents the total number of documents in the corpus (or document set), and N is a positive integer greater than or equal to 1; DF(t) represents the number of documents containing term t.

[0109] Because some implementations add 1 to the logarithm, or add a smoothing factor to the numerator or denominator of the logarithm. The formulas listed here are illustrative and are not limited thereto.

[0110] This disclosure also includes large-scale data processing and computational optimization. With the exponential growth of the number of users and movie data, the computational efficiency of recommendation algorithms faces significant challenges. In Content Filtering (CF), the computational complexity of the user similarity matrix (indicating the similarity between different objects) is proportional to the square of the number of users. Simultaneously, TF-IDF feature extraction based on content filtering also faces performance bottlenecks in high-dimensional text data. Therefore, the system provided in this disclosure optimizes computational efficiency through a multi-dimensional strategy.

[0111] Regarding feature reduction and approximate calculation, for high-dimensional feature vectors generated by TF-IDF, PCA (Principal Component Analysis) retains more than 90% of the variance information. PCA projects the original features to a low-dimensional space through orthogonal transformation, aiming to maximize the variance of the projection. Assume the original feature matrix (i.e., a matrix composed of at least one multimedia feature vector) is X∈R. k×m (Where k is the number of movies, m is the feature dimension of the original multimedia feature vector, and both k and m are positive integers greater than or equal to 1). PCA solves the eigenvalue decomposition of the covariance matrix. (X T (representing the transpose of X), and select the first K eigenvectors corresponding to the largest eigenvalue to form the projection matrix W∈R. m×K This reduces the feature dimension to K (where K is a positive integer greater than or equal to 1, e.g., K = 300), significantly reducing the computational cost of subsequent multimedia similarity calculations (e.g., cosine similarity). After dimensionality reduction, the movie feature vector (i.e., the multimedia feature vector) X reduce for:

[0112] X reduce =XW(4)

[0113] In some embodiments, cosine similarity is used to calculate the similarity between movies (i.e., multimedia similarity), generating a similarity matrix (i.e., movie similarity matrix) that reflects the characteristics of movie content. For example, this can be expressed by the following formula:

[0114]

[0115] In the above formula, and For example, the TF-IDF vectors of movie A and movie B. Representing vectors and The dot product. Representing vectors The Euclidean norm (L2 norm) is calculated as follows: i represents a vector The i-th element in the array, where i is a positive integer greater than or equal to 1; Representing vectors The Euclidean norm.

[0116] In some embodiments, in the implementation of the content-based filtering algorithm, the movies that the user has selected in the "I like" list can be obtained to determine their target multimedia preference information.

[0117] For example, when making personalized recommendations, based on the user's favorite movies (such as movie genres, actors, and other preferences filled in by the user during registration), the system calculates the similarity between the user's favorite movies and candidate movies in a movie similarity matrix, and accumulates the similarity scores. The top N movies (those with the highest combined similarity scores to the user's favorite movies, ranked from highest to lowest using different strategies such as accumulation, averaging, and weighted averaging) that the user has not yet watched are selected as the recommendation list.

[0118] For example, assuming a user likes movies A, B, and C, the system calculates the similarity between these movies and candidate movies (such as movies D, E, and F), then sums these similarity values ​​to obtain a score for each candidate movie. Finally, the system selects the top N movies (N is a positive integer greater than or equal to 1) with the highest scores as recommendations, excluding movies that the user has already watched, forming a personalized recommendation list.

[0119] For example, in the collaborative filtering algorithm, this disclosure embodiment splits the user-item matrix into a user feature matrix (i.e., an object feature matrix) and a movie feature matrix through matrix factorization technology, and uses Locality-Sensitive Hashing (LSH) technology to process the user feature matrix, thereby accelerating the calculation of similarity between users.

[0120] For example, the system first constructs a user-item matrix using user behavior data (such as ratings, viewing history, clicks, etc.). Specifically, the user-item matrix is ​​established based on user ratings of movies and other interactive behaviors (such as page browsing behavior). Let U = {u1, u2, ... u...} M Let I represent the set of users, and M represent the number of users, where M is a positive integer greater than or equal to 1; let I = {I1, I2, ..., I...} k} represents the set of items (movies). This user-item matrix can be denoted as R∈R M×k , where r u,i This represents the rating (or quantified score of other interactive behaviors) of user u for movie i.

[0121] Next, the system performs matrix decomposition to obtain a user feature matrix and a movie feature matrix. Then, it uses LSH to process the user feature matrix, quickly finding other users (nearest neighbors) similar to the target user (i.e., the target object). In this way, the system can efficiently calculate the similarity between users in a large-scale user-item matrix and predict movies that the target user might be interested in based on the behavior of similar users, generating a personalized recommendation list.

[0122] In this embodiment, the similarity calculation between users is achieved through a low-dimensional user feature matrix after matrix factorization, or the similarity between movies is achieved through a low-dimensional movie feature matrix after matrix factorization. Specifically, by processing the user feature matrix using LSH technology, the system can quickly find the user most similar to the target user, thereby optimizing the efficiency of similarity calculation. LSH technology can perform fast hash mapping on feature vectors, ensuring that similar users are mapped to the same bucket, avoiding the high computational cost of comparing the entire dataset one by one.

[0123] For example, the system utilizes matrix factorization techniques such as SVD (Singular Value Decomposition) to decompose the user-item matrix R into a user feature matrix U and a movie feature matrix V, and optimizes it by minimizing the following loss function L:

[0124]

[0125] In formula (6) above, rij This represents the rating of the i-th user for the j-th movie; u i v represents the feature vector of the i-th user; j Let represent the feature vector of the j-th movie; U and V are the feature matrices of the user and the movie, respectively; λ is the regularization coefficient, which is a number greater than or equal to 0 and can be set according to the actual scenario; Ω is the set of known ratings in the user-movie rating matrix (i.e., the user-item matrix); F is the Frobenius norm, used to calculate the sum of squares of the matrix.

[0126] For example, distributed computing can be implemented using Apache Spark for distributed storage and parallel computation of the user-movie rating matrix. Let the user-movie rating matrix (also called the user movie rating matrix) be R∈R M×k (where M represents the number of users and k is the number of movies). The matrix factorization model of CF decomposes it into a latent factor matrix by minimizing the loss function described in formula (6) above. and Matrix. k1 is a positive integer greater than or equal to 1. For example, Alternating Least Squares (ALS) in Spark MLlib accelerates the matrix factorization process by distributing the optimization task to multiple nodes in the cluster for parallel execution, thereby parallelizing U and V.

[0127] In this embodiment, the user feature matrix U, assuming M represents the number of users and k1 represents the latent factor dimension, means that each row in the user feature matrix represents a user's k1-dimensional latent preference vector. For example, if k1 = 3, a user's feature vector is [0.8, 0.2, -0.5], which may represent: Dimension 1 (Science Fiction Preference): 0.8 (highly like); Dimension 2 (Comedy Preference): 0.2 (moderately like); Dimension 3 (Horror Movie Dislike): -0.5 (strongly dislike). The item feature matrix V, assuming k represents the number of movies, means that each column represents a movie's k1-dimensional attribute vector. For example, if k1 = 3, a movie's feature vector is [0.9, 0.1, 0.3], which may represent: Dimension 1 (Science Fiction Attribute): 0.9 (strong science fiction elements); Dimension 2 (Comedy Attribute): 0.1 (almost no comedy elements); Dimension 3 (Horror Attribute): 0.3 (few horror elements).

[0128] For example, the predicted rating of the i-th user for the j-th movie can be obtained based on the user feature matrix U and the item feature matrix V. That is, the inner product of the feature vector representing user preference and the feature vector representing item attribute can be used as its predicted rating. Taking movie recommendation as an example, the data input is: user-movie rating matrix (e.g., a 5-point scale, where 0 represents no rating). Matrix decomposition: Let k1 = 50, decompose to obtain the user feature matrix U and the movie feature matrix V. Calculate the predicted rating of movies that the user has not rated (e.g., the predicted rating of user A for "XXX" = user A's science fiction preference dimension value × "XXX"'s science fiction attribute dimension value + ...). Sort by predicted score to generate a movie recommendation list. In this embodiment of the disclosure, the user-item relationship is mapped to a low-dimensional latent space through matrix decomposition. The user feature matrix U captures user preference patterns, and the item feature matrix V describes the distribution of item attributes. The inner product of the two achieves accurate prediction of unknown ratings.

[0129] For example, similarity between users can be obtained based on the user feature matrix U, for instance, by measuring the cosine of the angle between the feature vectors representing user preferences (e.g., user A's vector [0.8, 0.2] and user B's vector [0.7, 0.3] have a small angle, indicating similar preferences). Similarly, similarity between movies can be calculated based on the item feature matrix, for instance, by measuring the cosine of the angle between the feature vectors representing item attributes (e.g., science fiction movie vector [0.9, 0.1] is similar to another science fiction movie vector [0.85, 0.15]).

[0130] In this embodiment, Locality Sensitive Hashing (LSH) is a technique for efficiently processing Approximate Nearest Neighbor (ANN) searches in high-dimensional data. It designs specific hash functions so that similar data points (such as vectors) in high-dimensional space have a high probability of mapping to the same hash value (i.e., falling into the same "hash bucket") after hashing, while dissimilar data points collide with a low probability. This method is suitable for efficient similarity searches on large-scale datasets. A single hash might result in missed detections due to probability; therefore, multiple independent hash functions or hash tables are used. For example, by hashing multiple times or concatenating multiple hash functions, the probability of similar points colliding can be significantly increased, while the probability of dissimilar points colliding can be reduced. LSH designs different hash functions based on specific distance metrics (such as Euclidean distance, cosine similarity, and Jaccard similarity). For example: Euclidean distance: using random projection (such as p-stable LSH). Cosine similarity: using symbolic random projection (SimHash). Jaccard similarity: using MinHash. In the user-item matrix, LSH is used to quickly find other users or items with similar interests to the target user, thus achieving collaborative filtering.

[0131] In this embodiment of the disclosure, LSH is introduced into CF to accelerate similarity search. LSH uses a family of hash functions H to process the feature vector u of a high-dimensional user. i ∈R d (where d represents the dimension of a user's feature vector, and d is a positive integer greater than or equal to 1) is mapped to a low-dimensional bucket, thus increasing the probability of hash collisions between similar users. For cosine similarity, LSH uses a randomized hyperplane hash function:

[0132] h(u i ) = sign(γ T u i (7)

[0133] where γ∈R d It is a random Gaussian vector. The system can reduce the time complexity of searching for similar users by constructing multiple hash tables and performing approximate nearest neighbor searches on millions of users.

[0134] In the recommender system architecture provided in this disclosure, after matrix factorization (such as ALS) generates the user feature matrix U, it can be combined with Local Sensitive Hashing (LSH) to achieve efficient Approximate Nearest Neighbor (ANN) search. After matrix factorization, each row of the user feature matrix U is a low-dimensional dense vector of the user (e.g., k1 = 50 dimensions), suitable as input to LSH. These vectors encode the user's latent preference patterns (e.g., a preference for science fiction or comedy), and spatial distance (e.g., cosine similarity) directly reflects the similarity of user interests. LSH maps the feature vectors of similar users to the same hash bucket through a hash function, achieving similar user retrieval with sublinear time complexity. It is particularly suitable for scenarios with massive numbers of users (e.g., millions of users), avoiding the high computational overhead of exact nearest neighbor search (e.g., KNN).

[0135] For example, for a target user, the system uses the user feature matrix after matrix decomposition and LSH technology to find the Q most similar neighbors (the Q value is dynamically determined based on the target user's activity level, the number of ratings overlapping with others, etc., rather than a fixed value). Based on the behavior of these neighbors (such as ratings, viewing, etc.), the system predicts the movies that the target user may be interested in and forms a personalized recommendation list.

[0136] In an exemplary embodiment, determining the target multimedia information recommendation model corresponding to the selection control using the target multimedia information recommendation model, and generating and displaying target multimedia recommendation information for the target object, includes: when the target multimedia information recommendation model includes a hybrid recommendation model, and the hybrid recommendation model includes a collaborative filtering model and a content filtering model, obtaining object behavior data of the target object; determining weight exploration parameters based on the object behavior data of the target object; when it is determined to be in the exploration stage according to the weight exploration parameters, randomly selecting a first weight of the collaborative filtering model and a second weight of the content filtering model according to the probability of the weight exploration parameters; when it is determined to be in the utilization stage according to the weight exploration parameters, determining the first weight of the collaborative filtering model and the second weight of the content filtering model based on the feedback data of the target object to the recommended multimedia information; determining first multimedia recommendation information based on the collaborative filtering model, and determining second multimedia recommendation information based on the content filtering model; generating and displaying target multimedia recommendation information for the target object based on the first multimedia recommendation information and the first weight, and the second multimedia recommendation information and the second weight.

[0137] This disclosure also provides a combined strategy of content filtering and collaborative filtering algorithms, automatically adjusting weights through reinforcement learning to address the cold start problem and achieve personalized recommendations. During the cold start phase, the system initially employs a high degree of exploration (e.g., an ε-greedy strategy and a large learning rate) to quickly test different weight combinations (i.e., randomly combining the first and second weights), ensuring the system can quickly find a suitable recommendation strategy based on limited initial user feedback. Content filtering algorithms dominate this phase (e.g., the second weight is greater than the first weight), making recommendations based on the user's initial interest tags or behaviors, while collaborative filtering algorithms provide movie recommendations related to other similar users. As user behavior data accumulates, the system gradually reduces exploration and increases the first weight of collaborative filtering to more accurately reflect the user's personalized preferences. When user interaction data reaches a certain amount, the system enters a stable phase, and the weight adjustments of the content filtering and collaborative filtering algorithms tend to balance, ensuring that the recommendation results have both personalized accuracy and high diversity.

[0138] In this embodiment of the disclosure, the weight adjustment function of the ε-greedy strategy is set as follows: during the exploration phase (cold start), ε (i.e., the weight exploration parameter) is relatively large, for example, 0.8 is selected. As the number of interactions P between the user and the movie (P is an integer greater than or equal to 0) increases, ε gradually decreases, thereby reducing the exploration ratio and increasing the utilization ratio.

[0139] For example, the adjustment of ε is shown in the following formula:

[0140]

[0141] Where, ε min It is the smallest ε value (e.g., 0.1) to avoid exploration from completely stalling.

[0142] For example, regarding weight selection:

[0143] 1) If the system explores (with probability ε), it randomly selects a weight combination of the content-based filtering algorithm and the collaborative filtering algorithm (e.g., selecting the second weight W). CBF =0.7, first weight W CF =0.3 or other random combinations) to determine the most suitable recommendation strategy for new users.

[0144] 2) If the system utilizes the algorithm (with a probability of 1-ε), it uses the current optimal weight combination (determined based on user feedback and the current model to select the optimal weights for content-based filtering and collaborative filtering algorithms). For example, during the utilization phase (with a probability of 1-ε), the system will use the optimal weight combination derived from historical user feedback and behavioral data to generate recommendation results. Specifically, the system will determine the optimal weight combination for CBF and CF algorithms based on previous interaction data. For example, after a period of operation and user feedback, the system may discover W... CBF =0.6 and W CF A weighting of ε = 0.4 is considered optimal. Based on this, the system uses this weight combination to calculate the final recommendation score for movies, ensuring more accurate recommendations that align with user interests, rather than random exploration. As user behavior data accumulates, the system regularly updates and optimizes these weight combinations to provide more personalized and efficient recommendations. With increasing user behavior data, the ε value gradually decreases, and the system reduces exploration, relying instead on the optimal weight combination from collaborative filtering. At this point, the system makes greater use of existing historical data for recommendations, rather than engaging in excessive weight exploration.

[0145] In calculating the final recommendation score, the system calculates recommendation scores for movie A based on both content-based filtering (CBF) and collaborative filtering (CBD) algorithms. Then, the final recommendation score for movie A is obtained by weighted summing of the recommendation scores from these two algorithms. Specifically, if the CBF algorithm assigns a recommendation score of 1 to movie A... CBF (A), the recommendation score given to movie A by the CF algorithm is score. CF (A), then the final recommendation score for movie A is:

[0146] score final (A) = W CBF ×score CBF (A)+WCF ×score CF (A)(9)

[0147] Among them, W CBF and W CF These are the weights of content-based filtering algorithms and collaborative filtering algorithms, respectively. The system will dynamically adjust these weights based on the exploration and utilization phases.

[0148] S704, the recommended results are displayed to new users.

[0149] Based on the selected recommendation algorithm, the system generates a recommendation list and displays it to the user through the user interface.

[0150] S705 collects user feedback and behavioral data.

[0151] The system records user behavior data such as click-through rate and viewing time, as well as user ratings and comments on the recommendation results.

[0152] S706, update user models and movie information.

[0153] The system continuously updates the user model based on real-time user feedback and incorporates information on newly released movies in a timely manner to ensure that the recommendation results always keep up with the latest trends and changes in user interests.

[0154] For example, during operation, the system continuously collects real-time user feedback, such as click-through rates, viewing duration, ratings, comments, favorites, and shares, and then cleans, preprocesses, and stores this data uniformly. Subsequently, the system dynamically adjusts the user model using incremental updates or mini-batch retraining, i.e., updating the user's feature representation or rating matrix in the content-based filtering algorithm and collaborative filtering algorithm. In the content-based filtering algorithm, feature representation refers to the numerical characteristics of the movie content (such as the movie's genre, director, actors, etc.), which helps the system understand the movie's content features and recommend similar movies to the user. Although these content features are not directly related to the user's real-time feedback, they improve the accuracy of recommendations when combined with the user's preferences and behavioral data, ensuring that the user model can reflect their latest preferences in a timely manner.

[0155] For example, for newly released movies, the system retrieves their basic information (title, genre, director, actors, synopsis, language, release date, tags, etc.) from a database or external data source, and generates content feature vectors (i.e., multimedia feature vectors) through processing methods such as word segmentation, stop word removal, and TF-IDF. These vectors are then incorporated into a "movie-movie" similarity matrix (which contains multimedia similarities between different multimedia information) or a user-item matrix. After a user rates or interacts with a new movie, the system further updates the similarity or predicted rating model to ensure the new movie is quickly integrated into the overall recommendation system.

[0156] The predictive rating model in this embodiment refers to the system predicting the rating a user might give a new movie based on the user's historical behavior and movie content characteristics, and then using these predicted ratings to update the recommendation results. This allows new movies to be quickly integrated into the overall recommendation system.

[0157] To balance scalability and real-time performance, the system can employ either periodic batch updates (daily or weekly) or real-time incremental updates (processing immediately after a user generates new behavior) as needed, while leveraging a distributed computing framework in the background to improve the efficiency of large-scale data processing. Through this series of continuously iterative mechanisms, the system can keep pace with changes in user interests and the dynamics of new films in the market, constantly optimizing the accuracy and timeliness of recommendation results, thereby providing users with more personalized movie recommendation services that better meet their needs.

[0158] S707, Return to User Interface. The system returns the latest recommendations to the user interface for viewing and interaction.

[0159] Figure 8 This is an architecture diagram of a personalized movie recommendation system according to an exemplary embodiment of this disclosure. Figure 8 As shown, this personalized movie recommendation system can include a front-end subsystem and a back-end subsystem. The front-end subsystem can be used for users to search for movies, display movie lists, handle user login and logout, display join lists, allow users to rate movies, and may also include a personal center and provide three recommendation options (e.g., CBF, CF, and hybrid recommendation). The back-end subsystem can include user management, movie management, and system settings.

[0160] Adding a movie to a list refers to a user adding movies they are interested in to a specific list. On movie recommendation systems or video platforms, when a user browses for a movie and clicks on it, the movie is added to a custom list, such as "Movies I Want to Watch," "Favorites," or "To Watch List." This allows users to easily manage their viewing schedule and later find and watch these movies in their profile or on the list page.

[0161] Movie ratings refer to the ratings users give to movies they've watched. In movie recommendation systems or video platforms, users can typically rate a movie after watching it. Ratings can be star ratings (e.g., 1-5 stars), numerical ratings (e.g., 1-10 points), or simply like / dislike. This rating data is a crucial input for recommendation algorithms, used to calculate the similarity between movies and user preferences, thereby providing users with more personalized recommendations.

[0162] The personal center refers to a user's account page or settings page. On movie recommendation systems or video platforms, users can access their personal center after logging in. In the personal center, users can manage personal information (such as username, password, email, etc.), view viewing history, adjust recommendation preferences (such as favorite movie genres, directors, actors, etc.), manage added lists, and view rating records. The personal center is an important interface for user interaction with the system, providing rich personalization settings and management functions.

[0163] For example, the overall system design adopts a Model-View-Template (MVT) model based on the Django framework, combined with Python's powerful data processing ecosystem, to build a high-efficiency and scalable movie recommendation platform. The backend uses the Django framework to provide application programming interface (API) services; the frontend renders dynamic pages based on a template engine. The data layer uses PostgreSQL to manage user, movie, and rating (scoring) information, and integrates Redis caching and the Apache Spark distributed computing framework. The overall architecture is as follows: Figure 8 As shown. Through modular design and flexible scaling strategies, the system ensures stable operation in scenarios with large-scale data and high concurrency. Users can select recommendations based on their personal preferences using a recommendation algorithm, ensuring that the recommended results closely match the user's movie preferences.

[0164] The solutions provided in this disclosure address the cold start problem, improve recommendation accuracy and diversity, and optimize business value. On one hand, for new users or newly released movies, the system can quickly provide reasonable recommendations even without a large amount of user behavior data by directly analyzing movie content features (such as title, genre, and synopsis), reducing the impact of the cold start. On the other hand, the system combines movie content features and user behavior data, using a combination of multiple recommendation algorithms to improve the diversity and accuracy of recommendation results, ensuring that recommendations are both in line with user interests and diverse. Furthermore, the system has a continuous learning mechanism, periodically analyzing metrics such as user click-through rates and viewing duration to evaluate recommendation effectiveness and adjust recommendation strategies accordingly to ensure optimal performance. Specifically, by periodically analyzing user click-through rates, viewing duration, ratings, and comments, the system evaluates recommendation effectiveness and optimizes recommendation strategies.

[0165] In the solution provided by this disclosure, users can independently choose from content-based filtering algorithms, collaborative filtering algorithms, or hybrid recommendation algorithms for movie recommendations. Each user can select the recommendation algorithm that best suits their needs, thereby increasing the system's flexibility and personalization. The system generates recommendation results based on the user's selection and further optimizes the recommended content through real-time feedback (such as user viewing behavior and ratings) to better align with changes in user interests. Furthermore, the system continuously adjusts and improves the recommendation results through periodic data analysis to ensure that the recommendations always meet the user's latest needs and preferences. This mechanism significantly improves user satisfaction and engagement, provides more accurate recommendation services, and thus enhances the platform's user stickiness and commercial value. The practical effects of these functions not only enhance user satisfaction and engagement but also bring a more active user base to the platform, promoting precise advertising and improved user retention rates, thereby translating into significant commercial advantages.

[0166] Figure 9 This is a structural diagram of a recommendation system according to an exemplary embodiment of this disclosure. Figure 9 As shown, the recommendation system allows users to choose from three recommendation models: collaborative filtering, content filtering, and hybrid recommendation. The system also implements the logic of collaborative filtering (CF) and content filtering (CBF) to provide users with personalized recommendation lists. Users can select different recommendation algorithms based on their personal preferences, and the system generates corresponding recommendations based on these user choices, ensuring the accuracy and personalization of the suggestions, thereby enhancing the user experience.

[0167] The solutions provided in this disclosure offer a multi-modal personalized movie recommendation system that supports user selection of recommendation algorithms. By integrating three main recommendation algorithms—content-based filtering, collaborative filtering, and hybrid recommendation algorithms—it significantly improves the flexibility and accuracy of the recommendation system. This system not only solves the cold start problem, the uniformity of recommendation results, and poor interpretability in recommendation systems, but also enhances the user experience. By allowing users to choose the most suitable recommendation method based on their personal preferences, the system can more accurately capture user interests and provide highly personalized recommendation services. Furthermore, the system has a continuous learning mechanism that can continuously optimize recommendation results based on real-time user feedback, ensuring that recommendations always keep up with the latest trends and changes in user interests. Ultimately, this intelligent and personalized recommendation service not only saves users time and enhances user loyalty, but also promotes higher user engagement, bringing a more active user base and higher commercial value to multimedia content providers.

[0168] Meanwhile, the user interaction module supports user login and logout functions, allowing users to add movies to their favorites list to indicate which movies they are interested in, and to browse, delete, or tag collected movies. This behavioral data provides the necessary input algorithms, which can be used to further optimize recommendation results.

[0169] For example, on the input side, the system receives user search requests, click behavior, rating feedback, and data sources from external movie data contexts (such as movie information APIs). As for the output, based on the input data, the system returns a personalized list of movie recommendations and a list of search results, and displays the user's movie collection status, helping the user manage and view their interests and preferences.

[0170] Figure 10 This is a schematic diagram illustrating the collection of user preference information for movie recommendations, an exemplary embodiment of this disclosure. Figure 10 As shown, it is possible to collect user preferences regarding movie genres, directors, actors, movie viewing history, user feedback, emotional preferences, cultural and language preferences, recommendation goals, and social circles. For example, the collected user preference information can be used to determine the user's target multimedia preferences.

[0171] A limitation of Cloud Computing (CF) is the cold start problem. When new users or movies are introduced to the system, calculating similarity becomes challenging due to a lack of sufficient rating data, thus accurately impacting the quality of recommendations. To address this, the system guides new users through initial preference settings during registration, such as selecting favorite movie genres or rating a representative set of movies, to build a basic user profile (e.g., ...). Figure 10(As shown). For a new movie, the system uses metadata such as genre, director, and actors, combined with CBF to supplement CF, improving recommendation accuracy and coverage in cold start scenarios.

[0172] For example, recommendations based on plot attractiveness refer to a recommendation system recommending works such as movies, books, and games to users based on the plot attractiveness of the content. Plot attractiveness refers to elements that can attract viewers or readers, such as the story's interest, tension, emotional depth, innovation, or suspense. Natural Language Processing (NLP) techniques are used to analyze the plot synopsis and extract plot features (such as conflict type, theme, and pacing). Plot attractiveness is quantified by combining user ratings, reviews, or viewing behaviors (such as fast-forwarding or rewatching). For instance, if a user likes the multi-layered dream setting of "XXXX," the system might recommend other movies with complex narrative structures (such as "YYYY").

[0173] For example, recommendations based on actor performance refer to a recommendation system recommending works based on an actor's performance quality, popularity, or user preferences. Actor performance may include factors such as acting skills, character portrayal ability, audience reviews, or box office appeal. Actor feature extraction includes: quantitative indicators: awards (e.g., Oscars), box office performance, social media influence; subjective evaluation: analyzing an actor's performance style through user comments (e.g., "suitable for villainous roles"). User preference modeling includes: explicit feedback: users actively following actors or marking them as "liked." Implicit feedback: the frequency and duration of users watching a particular actor's works. For example, if a user frequently watches actor B's movies, the system will recommend their new works.

[0174] For example, the system can combine both approaches, such as: "Users like action movies starring actor B and prefer fast-paced plots → Recommend 'xxx'". For new actors or new IPs, recommendations can be made based on metadata (such as director, genre) or cross-domain information (such as "a new work by the same screenwriter as 'zzz'"). These two recommendation logics satisfy user needs from the dimensions of "content appeal" and "creator influence" respectively, and are components of the diversified strategy of the recommendation system.

[0175] For example, recommendations based on common interests with friends or people on social media refer to a strategy where recommendation systems provide personalized content, product, or service suggestions by analyzing the overlap of interests between a user and their friends, followers, or similar users in their social networks. This leverages trust and shared interests within social relationships to enhance recommendation acceptance and is found on social media, e-commerce platforms, and content platforms. The process involves: building a social graph: constructing a social network through users' friend relationships, following lists, or interactive behaviors (such as likes and comments); interest mining and alignment, including: explicit data: user-generated interest tags (e.g., "science fiction fan") and joined groups (e.g., "mountain climbing club"); implicit data: inferring interests through user behavior (e.g., watching, purchasing, searching) and comparing them with the interests of friends in social relationships; and combining collaborative filtering techniques to find "shadow friends" who share similar interests but have no direct social relationship with the user, thus expanding the recommendation scope.

[0176] The solution provided in this disclosure addresses the cold start problem. For new users or newly released movies, recommendation systems in related technologies often struggle to provide effective recommendations due to a lack of sufficient user behavior data, leading to the cold start problem. This disclosure analyzes four basic attributes (age, region, gender, occupation) filled in by new users during registration, their movie preference choices (such as movie genre, language, director, actors, etc.), their page browsing behavior in the short term after login, and the system's selection of different types of popular movies based on the new user's choices (to infer their preferred movie preferences). It provides three different recommendation models, enabling the system to quickly provide reasonable recommendations even without a large amount of user behavior data, thus reducing the cold start problem. Personalized recommendations can be obtained immediately based on the movie preference information filled in by new users during registration, without waiting for sufficient user behavior data to accumulate. New users or newly released movies can receive personalized recommendations in a short time, reducing the time spent accumulating sufficient user behavior data. The recommendation results are highly correlated with the movie content characteristics, ensuring the timeliness and relevance of the recommendations and improving the user experience. This method is particularly suitable for scenarios involving new users or newly released movies, expanding the system's application scope and practicality, and solving the long-standing problem of cold start that has plagued recommendation systems.

[0177] Figure 11 This is a flowchart illustrating a CF-based processing embodiment of this disclosure. (As follows...) Figure 11As shown, user rating data is acquired to create a user movie rating matrix as the object multimedia matrix; based on the user movie rating matrix, a user feature matrix U and a movie feature matrix V are obtained. A movie similarity matrix is ​​generated based on the similarity between movies in the movie feature matrix V. Finally, the movie similarity matrix is ​​weighted according to user preferences (i.e., target multimedia preference information) to generate a recommended movie list.

[0178] For example, CF's data source uses the Myrating model, which stores user ratings for different movies. The system creates a user rating matrix (user movie rating matrix) to implement the algorithm, specifically a pivot table where users are rows, movie IDs are columns, and the corresponding ratings are cell values. For missing ratings, the system fills in zeros to ensure the matrix's integrity and availability, thus providing stable input data for the recommendation algorithm. This is illustrated in Table 1, for example.

[0179] Table 1. Overview of the CF Algorithm and User Rating Matrix

[0180]

[0181]

[0182] For example, a Pearson correlation matrix based on user ratings is used to measure movie similarity. If users tend to give two movies similar ratings, then the two movies are considered the same.

[0183] The flowchart of the user interaction module is as follows: Figure 11 As shown, the system first receives user rating data (or score data) through Django-based view functions and creates a user movie rating matrix as an object multimedia matrix (e.g., rows represent users, columns represent movies, and values ​​represent ratings). Next, it calculates a movie similarity matrix (e.g., measuring similarity between movies using cosine similarity or Pearson correlation coefficient) and weights the similarities according to user preferences (e.g., preference for science fiction or comedy) to improve recommendation relevance. Finally, the system generates a personalized movie recommendation list based on these weighted results. For example, if user A's favorite movie X is highly similar to movie Y, then movie Y might be recommended to user A.

[0184] For example, before querying movie details in the database, the system verifies the user's identity (e.g., login status) and activity status (e.g., whether the account is disabled). If a record exists in the user list (MyList), it indicates that the user has viewing history, and the system retrieves the movie's viewing status (e.g., watched / not watched). Furthermore, the system supports real-time updates to user data via POST requests (e.g., modifying ratings or marking movies as "watched") to update user ratings or viewing status data, ensuring that recommendation results promptly reflect changes in user behavior.

[0185] The method provided in this disclosure uses a user rating matrix as the core input and combines movie similarity with user preferences to achieve accurate recommendations. Identity verification and status checks ensure the legitimacy of data access. It supports dynamically updating user behavior data, improving the timeliness and personalization of the recommendation system.

[0186] For example, movie similarity is weighted according to user personalized preferences, which strengthens the weight of the movie genre preferred by the user in the similarity calculation, thereby improving the matching degree between the recommendation results and the user's interests. In some embodiments, user preference modeling is performed. Specifically, data collection is carried out. Explicit feedback includes user-submitted preferences (e.g., "I like science fiction") and rating data (e.g., giving "XXX" 5 stars). Implicit feedback includes viewing history (e.g., watching science fiction movies multiple times), search history (e.g., frequently searching for "space-themed"), and dwell time (e.g., watching the entire "XX" series). Then, preferences are quantified. User preferences are converted into computable vectors or weights. For example, TF-IDF is used: if a user has watched 10 movies, 5 of which are science fiction, then the TF-IDF value of the science fiction genre is log(1+5)×log(10 / 5), reflecting its importance. Another example is using embedding vectors, mapping user preferences to low-dimensional vectors through Word2Vec or deep learning models. Then, similarity weighting is performed. Basic similarity calculation is performed first. For example, initial similarity between movies is calculated using cosine similarity, Pearson correlation coefficient, or matrix factorization (such as SVD). For instance, the cosine similarity between movie A and movie B is cos(θ) = ||A||||B||A×B, where vectors A and B represent movie features (such as genre, director, and actors). Then, preference weights are introduced. The similarity is dynamically adjusted based on user preferences. For example, genre weighting: if a user prefers science fiction, the weight of the science fiction genre in the similarity calculation is increased. For example: WeightedSim(A,B) = Sim(A,B)×(1+α×GenreWeight science fiction), where α is an adjustment coefficient, set according to the actual situation; GenreWeight science fiction is the user's preference strength for the science fiction genre. Personalized similarity is generated by taking the dot product of the user preference vector and the movie feature vector. For example: PersonalizedSim(u,A,B) = Sim(A,B)×(u preference × A feature), where u preference is the user preference vector, and A feature is the feature vector of movie A.

[0187] For example, user A prefers science fiction (weight = 1.5) and has no preference for comedies (weight = 1.0). The original similarity between movie X (science fiction) and movie Y (science fiction + comedy) is 0.8. The weighted similarity is: WeightedSim = 0.8 × 1.5 = 1.2. If another pair of movies, Z (comedy) and W (comedy), has a similarity of 0.7, the weighted similarity remains 0.7 (because the user has no preference for comedies). Ultimately, this results in movie X having a higher priority than movie Z in the recommendation list, despite its lower original similarity.

[0188] For example, movies are sorted according to weighted similarity, and a preset number of the top-ranked movies are selected as the recommendation list.

[0189] This embodiment of the disclosure generates a recommendation list that better matches user interests by quantifying user preferences into weights and dynamically adjusting movie similarity. This process combines user behavior analysis, similarity calculation, and machine learning techniques, serving as a means for recommendation systems to achieve personalization. In practical applications, it is necessary to balance accuracy and diversity, and continuously optimize model performance through feedback.

[0190] With the continuous maturation of data mining and machine learning techniques, CF (Conversion Theory) has gradually become a popular method in recommender systems. By analyzing historical rating correlations between users or movies, CF performs well in predicting a target user's preference for unknown movies. For example, in a study using the MovieLens dataset, CF significantly improved prediction accuracy. However, CF performs poorly in cold-start scenarios because there is a lack of sufficient system rating data for new users or new movies to quickly provide accurate recommendations. At the same time, CF's strong reliance on historical ratings makes adapting to rapid changes in user preferences challenging.

[0191] Figure 12 This is a flowchart illustrating a CBF-based processing embodiment of this disclosure. (As follows...) Figure 12 As shown, movie descriptions are obtained as multimedia content features, and then vectorized (e.g., using TF-IDF) to obtain multimedia feature vectors. Cosine similarity is calculated based on these feature vectors to represent the multimedia similarity between different multimedia information. Similar movies are identified based on the user's favorites. The movies saved in the user's favorites represent the user's target multimedia preferences. Movies of interest to the user are then identified based on these target multimedia preferences to generate a recommendation list.

[0192] Unlike Content-Based Filtering (CBF), which focuses on user behavior data, Content-Based Filtering (CBF) focuses on the characteristics of the movies themselves, such as genre, director, actors, and plot elements. When a user's favorite movie contains specific elements (e.g., science fiction theme, a particular lead actor), the system matches movies with similar characteristics in the movie feature space and recommends them to the user. Metadata-based case studies, such as the Internet Movie Database (IMDb) and the Movie Database (TMDb), show that CBF performs better and is very robust when new content is added. However, Content-Based Filtering (CBF) can easily limit users to pre-existing preferences, leading to overly homogenized recommendations that lack originality and diversity.

[0193] For example, CBF primarily calculates similarity based on textual features such as the movie's title, genre, and synopsis. As shown in Table 2, the movie's title and genre are combined into feature text (or multimedia content features).

[0194] Table 2. CBF: Film Characteristics Table

[0195]

[0196]

[0197] After tagging the feature text of each movie and applying TF-IDF, the system converts each movie into a vector containing keyword weights. This vector represents the importance and relevance of each keyword in the movie, providing a basis for calculating the similarity between movies.

[0198] After obtaining the TF-IDF vectors of all movies, the system calculates the cosine similarity between each pair of movies, resulting in the similarity matrix shown in Table 3. For example, the similarity between movie B and movie C is 0.65, indicating a relatively high content-based similarity, which may be due to shared science fiction elements. Meanwhile, the similarity between movie E and movie D is 0.00, indicating that these two movies differ in content.

[0199] Table 3. Examples of Cosine Similarity Matrix

[0200]

[0201] Workflows based on recommender systems (such as...) Figure 12 (As shown). First, the system retrieves movie descriptions and converts them into TF-IDF vectors. Next, it calculates the cosine similarity between these movie vectors to identify other movies similar to those in the user's collection. When interaction data is limited in the early stages of new user engagement, the system can leverage preference tags selected during registration and apply CBF to filter candidate movies based on metadata such as genre and theme. This approach further improves the effectiveness of recommendations during the cold start phase. Finally, the system generates a personalized recommendation list based on these similar movies and presents it to the user.

[0202] To address the cold start problem, the solution provided in this disclosure directly analyzes the characteristics of movie content, enabling it to quickly provide reasonable recommendations even without a large amount of user behavior data, thus reducing the impact of the cold start.

[0203] Figure 13 This is a flowchart illustrating the processing of HRA based on CF and CBF according to an exemplary embodiment of this disclosure. Figure 13 As shown, recommendation list A is generated based on user historical ratings using a CF (Conversion Flow) model, and recommendation list B is generated based on a CBF (Conversion Flow Flow) model. Then, the final recommendation list displayed to the user is generated based on recommendation lists A and B.

[0204] HRA combines the results of CF (based on other users' rating preferences) and CBF (based on the user's previous behavior and preferences to recommend content that matches the user's interests) to obtain a more comprehensive recommendation list. Figure 12 As shown, for example, both CF and CBF use users' historical rating data to generate two recommendation lists. These two lists are then merged, deweighted, and sorted to form a hybrid recommendation list.

[0205] As shown in Table 4, CF recommends movies A, C, F (and more) based on the preferences of other users with similar interests. In contrast, CBF recommends movies B, C, D, etc., based on content features such as the genre, theme, and other characteristics of movies the user already likes. Both algorithms recommended movie C, so based on the HRA model, movie C is considered worthy of recommendation.

[0206] Table 4. Recommended lists comparing CF and CBF

[0207]

[0208] During the blending process, CF and CBF results can be handled through weighted or alternating selection. In a simple scenario, movie C has high CF and CBF scores and is therefore prioritized for recommendation. In another scenario, movie A scores highly in CF, while movie B only scores highly in CBF. After merging, the final recommendation list is obtained by sorting the combined scores. The resulting blended recommendation list is shown in Table 5. Movie C ranks first with 1.5 points, movie A ranks second with 0.9 points, movie B ranks third with 0.85 points, and so on. The end user can see a personalized recommendation list that combines the best CF and CBF results.

[0209] Table 5. Overview of Merging and Weighting

[0210]

[0211] To address the limitations of single methods, HRA combines the advantages of collaborative filtering (CF) and content-based filtering (CBF). For example, practical experience shows that combining CF, CBF, and other advanced algorithms can achieve a better balance between accuracy and diversity. HRA can mitigate the cold-start problem by leveraging content features, while simultaneously improving the accuracy and richness of recommendations by utilizing user group preference patterns.

[0212] The system provided in this disclosure allows users to independently choose their preferred recommendation algorithm, breaking the limitation of related technologies where recommendation systems rely on fixed algorithms. Unlike traditional systems that automatically select algorithms based on user behavior, this system enables users to freely choose different recommendation algorithms (such as CF, CBF, and HRA) according to their preferences and needs. This design makes the recommendation system more flexible and provides more accurate recommendations based on user needs. During use, users can choose the recommendation algorithm that best suits their needs, and the system will generate a personalized recommendation list based on their selection. Collaborative filtering algorithms recommend movies based on similar interests derived from user movie rating data; content-based filtering algorithms recommend movies based on metadata such as genre, director, and actors; hybrid recommendation algorithms combine the advantages of both, providing diverse and accurate recommendations. This innovative design enhances users' sense of control over the recommendation system and solves the cold start problem. For new users, the system can guide them to choose a suitable recommendation algorithm based on their initial preferences, thereby improving the accuracy and diversity of recommendations. By giving users more control, the system can provide the best recommendation experience in different scenarios.

[0213] In recent years, emerging technologies such as deep learning, including Graph Neural Networks (GNNs), have been injected into recommender systems, giving them greater intelligence and generalization capabilities. Utilizing neural networks to extract latent interaction features between users and movies offers flexibility and achieves better prediction accuracy. Furthermore, GNNs treat users and movies as nodes in a graph, mapping interactions to edges, and refine the latent structure through multi-layer graph convolutions to realize patterns and aggregation mechanisms in the data. These methods can effectively explore latent relationships in more dimensions and address data sparsity issues, providing more diverse and dynamic optimization paths for recommender systems. However, it should be noted that parameter tuning and model training for deep and graphical models are significantly more difficult, placing higher demands on computational resources and engineering capabilities.

[0214] To address the cold start problem, the uniformity of recommendation results, and poor interpretability in related technologies, this disclosure solves the cold start problem for new users or newly released movies by directly analyzing movie content features; it improves the diversity and accuracy of recommendation results through the combination of multiple recommendation algorithms; and because the recommendation process is transparent, users can clearly understand the basis for the recommendations. Furthermore, the system has a continuous learning mechanism that can continuously update the user model and movie information based on real-time user feedback, including proactive user feedback on recommended movies, such as liking or disliking, skipping, clicking to watch, or giving a rating. This feedback is used to dynamically update the user model, specifically by updating the user's rating matrix and feature vector, to adjust the recommendation results according to the latest interests and preferences. Simultaneously, the system updates movie information based on user behavior, specifically by updating the movie's rating matrix and feature matrix, to ensure that the recommendation results keep pace with the latest changes in user interests. This ensures that the recommendation results always keep up with the latest trends and changes in user interests. Ultimately, this intelligent and personalized recommendation service not only saves users time and enhances user loyalty but also promotes higher user engagement, bringing a more active user base and higher commercial value to multimedia content providers.

[0215] Figure 14 This is a schematic diagram of the structure of a multimedia information processing apparatus according to an exemplary embodiment of this disclosure. Figure 14 As shown, the multimedia information processing apparatus 1400 provided in this embodiment includes a display unit 1410 and a processing unit 1420. The display unit 1410 displays a multimedia information recommendation model selection page, which includes multiple multimedia information recommendation model selection controls. Different multimedia information recommendation model selection controls correspond to different multimedia information recommendation models, and the multimedia information recommendation model includes a target multimedia information recommendation model. The processing unit 1420, in response to a target object's selection operation of the target multimedia information recommendation model selection control among the multiple multimedia information recommendation model selection controls, determines the target multimedia information recommendation model corresponding to the target multimedia information recommendation model selection control, and generates and displays target multimedia recommendation information for the target object.

[0216] In an exemplary embodiment, the display unit 1410 is further configured to: display corresponding indication information at a position associated with the multimedia information recommendation model selection control on the multimedia information recommendation model selection page, wherein the indication information is used to indicate the basis for the corresponding multimedia information recommendation model to generate multimedia recommendation information.

[0217] In an exemplary embodiment, the plurality of multimedia information recommendation model selection controls include a collaborative filtering model selection control, a content filtering model selection control, and a hybrid recommendation model selection control. The multimedia information recommendation models include a collaborative filtering model, a content filtering model, and a hybrid recommendation model. Indication information displayed at a location associated with the collaborative filtering model selection control is used to indicate that multimedia recommendation information is generated by analyzing the similarity between objects based on object behavior data. Indication information displayed at a location associated with the content filtering model selection control is used to indicate that multimedia recommendation information is generated based on multimedia content features. Indication information displayed at a location associated with the hybrid recommendation model selection control is used to indicate that multimedia recommendation information is generated by combining the collaborative filtering model and the content filtering model.

[0218] In an exemplary embodiment, the display unit 1410 is further configured to: display a registration page, the registration page including a basic information input control and a multimedia preference information input control. The processing unit 1420 is further configured to: obtain target basic information of the target object in response to an input operation of the target object on the basic information input control; and obtain target multimedia preference information of the target object in response to an input operation of the target object on the multimedia preference information input control.

[0219] In an exemplary embodiment, the display unit 1410 is further configured to: display a login page, the login page including a login control. The processing unit 1420 is further configured to: obtain the identity information of the target object in response to the login operation of the target object on the login control; verify the identity information of the target object; after the identity information of the target object is verified, obtain page operation behavior data of the target object within a predetermined time after login; and obtain the target multimedia preference information of the target object based on the page operation behavior data of the target object within the predetermined time after login.

[0220] In an exemplary embodiment, the display unit 1410 is further configured to: display a login page, the login page including a login control. The processing unit 1420 is further configured to: obtain the identity information of the target object in response to the login operation of the target object on the login control; verify the identity information of the target object; display different types of multimedia information after the identity information of the target object is verified; and obtain the target multimedia preference information of the target object based on the target object's selection operation on the displayed different types of multimedia information.

[0221] In an exemplary embodiment, the processing unit 1420 is further configured to: extract multimedia content features of multimedia information when the target multimedia information recommendation model includes a content filtering model; convert the multimedia content features into multimedia feature vectors; obtain multimedia similarity between different multimedia information based on the multimedia feature vectors; and generate and display target multimedia recommendation information for the target object based on the target multimedia preference information of the target object and the multimedia similarity between different multimedia information.

[0222] In an exemplary embodiment, the processing unit 1420 is further configured to: when the target multimedia information recommendation model includes a collaborative filtering model, establish an object multimedia matrix based on object behavior data; perform matrix decomposition on the object multimedia matrix to obtain an object feature matrix; perform local sensitive hashing on the object feature vectors in the object feature matrix to obtain the nearest neighbor objects of the target object; and generate and display target multimedia recommendation information for the target object based on the object behavior data of the nearest neighbor objects.

[0223] In an exemplary embodiment, the processing unit 1420 is further configured to: obtain object behavior data of the target object when the target multimedia information recommendation model includes a hybrid recommendation model and the hybrid recommendation model includes a collaborative filtering model and a content filtering model; determine weight exploration parameters based on the object behavior data of the target object; when it is determined to be in the exploration stage according to the weight exploration parameters, randomly select a first weight of the collaborative filtering model and a second weight of the content filtering model according to the probability of the weight exploration parameters; when it is determined to be in the utilization stage according to the weight exploration parameters, determine the first weight of the collaborative filtering model and the second weight of the content filtering model based on the feedback data of the target object to the recommended multimedia information; determine first multimedia recommendation information based on the collaborative filtering model and determine second multimedia recommendation information based on the content filtering model; and generate and display target multimedia recommendation information for the target object based on the first multimedia recommendation information and the first weight, and the second multimedia recommendation information and the second weight.

[0224] Figure 14 Other aspects of the embodiments can be found in the other embodiments described above.

[0225] This disclosure also provides a personalized movie recommendation method, system, and device that supports user selection. A personalized movie recommendation system combining content-based filtering algorithms, collaborative filtering algorithms, and hybrid recommendation algorithms is disclosed, belonging to the field of multimedia information processing. This system allows users to choose their own recommendation algorithm (content-based filtering, collaborative filtering, or hybrid recommendation algorithm) and provides personalized movie recommendation services by analyzing the content features of movies (such as title, genre, synopsis, etc.) and the user's historical behavioral data. Compared with single recommendation methods, this disclosure provides greater flexibility and accuracy, enhancing the user experience.

[0226] This disclosure provides a multi-modal personalized movie recommendation system that supports user-selected recommendation algorithms, possessing broad application potential and market prospects. This system is not only applicable to multimedia content platforms (such as online movies, music streaming, and book recommendations), but also to e-commerce platforms (general e-commerce and vertical e-commerce), social media and social networks (social platforms and professional social networks), and the advertising and marketing field, helping these industries improve user experience and business value through precise recommendations. For B2C consumers, including movie lovers, music enthusiasts, bookworms, and shopping aficionados, personalized recommendations allow them to discover more films of interest, find new songs that suit their tastes, enrich their reading life, or save time on online shopping. For B2B enterprises, content providers (such as film production companies, music record companies, and publishers) hope to increase the exposure of their works, e-commerce platforms (general e-commerce and vertical e-commerce) hope to improve user purchase conversion rates and satisfaction through personalized recommendations, and advertising agencies hope to help advertisers achieve higher returns on investment through precise ad placement.

[0227] The solution provided in this disclosure allows users to select the most suitable recommendation algorithm according to their personal preferences, improving the system's adaptability and user experience. Furthermore, by combining three algorithms—content-based filtering, collaborative filtering, and hybrid recommendation—it can provide more accurate and diverse recommendation results, while also solving the cold start problem and increasing the transparency of the recommendation process.

[0228] For example, using the Django framework, an intelligent movie recommendation system was designed and implemented based on theoretical analysis and practical application. This system integrates CF and CBF, and improvements were made in requirements analysis, feasibility assessment, database design, and security strategies. It helps users discover movies on movie detail pages and provides personalized recommendations. On one hand, a flexible and personalized movie recommendation system is proposed, allowing users to choose their preferred recommendation algorithm, solving the problem of algorithm selection based on user needs. On the other hand, the cold start problem is solved by collecting initial user preferences during registration, ensuring that the system can provide relevant suggestions even for new users with limited data. Furthermore, computational techniques such as LSH, PCA, and distributed computing are utilized to optimize system performance and scalability while reducing computational and storage costs, ensuring stable operation under high concurrency conditions. Caching and pre-computation mechanisms further balance performance and real-time requirements. Static data (such as the movie content similarity matrix) is pre-computed offline in batches, updated daily, and stored in a Redis cache. Real-time user requests are read directly from the cache to avoid redundant computation. In addition, the Least Recently Used (LRU) strategy is used to cache recent user behavior data (e.g., ratings and click records), thereby reducing database query frequency and I / O pressure.

[0229] In high-concurrency scenarios, the system employs a microservice architecture and Docker container deployment, with Kubernetes dynamically managing resources and Nginx handling request distribution. Asynchronous task processing is implemented using Celery queues, reducing the impact of long tasks on real-time interaction. The frontend displays task status in real time, and pushes the results to the client via WebSocket upon completion. Resource monitoring is performed using Prometheus and Grafana; when resource usage exceeds thresholds, the system automatically scales nodes or uses degradation algorithms to ensure stable operation.

[0230] Comparative experiments validated the performance of the proposed recommendation algorithm and benchmarked it against the widely used MovieLens baseline model. Experiments were conducted on the MovieLens 1M dataset, using precision, recall, F1 score, and root mean square error (RMSE) as evaluation metrics. These metrics comprehensively assess the performance of the recommendation algorithm, reflecting its accuracy, coverage, recommendation balance, and prediction error.

[0231] In the experiments, the performance of three individual algorithms—CF, CBF, and HRA—was tested, and each algorithm was evaluated independently. Furthermore, considering that users can manually select algorithms according to their preferences, different user choices were simulated, and the system's performance was evaluated when users selected different algorithms. This method enables the evaluation of the effectiveness of each algorithm in real-world use and verifies whether the recommendation algorithm selected by the user can provide the optimal recommendation experience. The system provided by the embodiments of this disclosure can provide flexible and efficient recommendation services.

[0232] The personalized movie recommendation system provided in this disclosure combines CF and CBF to offer accurate movie recommendations based on user preferences. Experimental results show that the system outperforms the traditional MovieLens baseline model on multiple evaluation metrics, demonstrating its advantages in personalized recommendations and recommendation diversity. Furthermore, the system solves the cold start problem. It optimizes efficiency and accuracy through techniques such as LSH, PCA, and distributed computing, effectively handling large-scale data and high-concurrency users.

[0233] As user data accumulates, to protect user privacy, system design should comply with data protection regulations and ensure that all user personal information, including search history, viewing history, and other sensitive data, is encrypted and protected. Users should be able to see how their data is used at any time and have the right to manage it, such as granting authorization and deleting data.

[0234] Meanwhile, recommender systems can lead to the "filter bubble" problem, where the system may restrict users' access to diverse content, resulting in information silos. Therefore, in algorithm design, in addition to personalized recommendations, measures should be taken to increase the diversity of recommendations. For example, diversity control mechanisms can be introduced to prevent the system from over-relying on past user behavior data and to recommend more content that matches user interests but has not been previously explored. Furthermore, the transparency of recommendations is crucial; the system should be able to explain the reasons behind its recommendations to increase user trust in the system.

[0235] During system use, special attention must be paid to bias issues. Since both CF and CBF methods rely on historical user data, they may lead to suboptimal recommendations tailored to specific groups. Therefore, this system implements bias correction measures, such as diversity control and adding diverse user group characteristics, to mitigate algorithmic bias in the recommendation results.

[0236] Figure 15 A schematic structural diagram of an electronic device according to an embodiment of the present disclosure is shown. Figure 15 The illustrated electronic device includes a processor 10, which can call and run computer programs from storage to implement the methods in the embodiments of this disclosure.

[0237] Optionally, such as Figure 15 As shown, the electronic device may also include a storage unit 20. The processor 10 can retrieve and run computer programs from the storage unit 20 to implement the methods in any embodiment of this disclosure.

[0238] The memory 20 can be a separate device independent of the processor 10, or it can be integrated into the processor 10.

[0239] Optionally, the electronic device may also include a transceiver, which the processor 10 can control to communicate with other devices. Specifically, it can send information or data to other devices or receive information or data sent by other devices.

[0240] Optionally, the processor 10, the memory 20, and the transceiver can communicate bidirectionally with each other via a communication bus.

[0241] For example, the electronic device can be a movie recommendation device. The movie recommendation device includes a storage unit 20, a processor 10, and a computer program 30 stored on the storage unit and executable on the processor. When the processor 10 executes the computer program 30, it implements the movie recommendation method described above. Specifically, the movie recommendation device can be a computer, server, host computer, etc. In some embodiments, the processor 10 can be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip, used to run program code stored in the storage unit 20 or process data, such as executing access restriction programs. The storage unit 20 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type storage (e.g., SD or DX storage), magnetic storage, disk, optical disk, etc. In some embodiments, the storage unit 20 can be an internal storage unit of the movie recommendation device, such as the hard disk of the movie recommendation device. In other embodiments, storage device 20 may also be an external storage device of the movie recommendation device, such as a plug-in hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc., provided on the movie recommendation device.

[0242] Furthermore, the storage device 20 may include both internal storage units and external storage devices. The storage device 20 can be used not only to store application software and various types of data installed on the movie recommendation device, but also to temporarily store data that has been output or will be output. It should be noted that... Figure 15The structure shown does not constitute a limitation on the movie recommendation device. In other embodiments, the movie recommendation device may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0243] This disclosure also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the movie recommendation method described above. The above embodiments are merely illustrative of several implementations of this disclosure, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this disclosure.

[0244] It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this disclosure, and these all fall within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the appended claims.

Claims

1. A multimedia information processing method, characterized by, include: The page displays a multimedia information recommendation model selection page, which includes multiple multimedia information recommendation model selection controls. Different multimedia information recommendation model selection controls correspond to different multimedia information recommendation models, and the multimedia information recommendation model includes a target multimedia information recommendation model. In response to the target object's selection operation of the target multimedia information recommendation model selection control among the multiple multimedia information recommendation model selection controls, the target multimedia information recommendation model corresponding to the target multimedia information recommendation model selection control is determined, and target multimedia recommendation information for the target object is generated and displayed.

2. The method of claim 1, wherein, Also includes: On the multimedia information recommendation model selection page, corresponding instruction information is displayed at the position associated with the multimedia information recommendation model selection control. The instruction information is used to indicate the basis for the corresponding multimedia information recommendation model to generate multimedia recommendation information.

3. The method of claim 2, wherein, The multimedia information recommendation model selection controls include collaborative filtering model selection controls, content filtering model selection controls, and hybrid recommendation model selection controls. The multimedia information recommendation models include collaborative filtering models, content filtering models, and hybrid recommendation models. The indication information displayed at the location associated with the collaborative filtering model selection control is used to indicate the generation of multimedia recommendation information based on the analysis of the similarity between objects based on object behavior data; The indication information displayed at the location associated with the content filtering model selection control is used to indicate the generation of multimedia recommendation information based on multimedia content features; The indication information displayed at the location associated with the hybrid recommendation model selection control is used to instruct the generation of multimedia recommendation information by combining the collaborative filtering model and the content filtering model.

4. The method of claim 1, wherein, Also includes: Display a registration page, which includes basic information input controls and multimedia preference information input controls; In response to the input operation of the target object on the basic information input control, the target basic information of the target object is obtained; In response to the target object's input operation on the multimedia preference information input control, the target object's target multimedia preference information is obtained.

5. The method of claim 1, wherein, Also includes: Display a login page, which includes login controls; In response to the login operation of the target object on the login control, the identity information of the target object is obtained; Verify the identity information of the target object; After the identity information of the target object is verified, the page operation behavior data of the target object within a predetermined time after login is obtained; Based on the page operation behavior data of the target object within the predetermined time after login, the target multimedia preference information of the target object is obtained.

6. The method of claim 1, wherein, Also includes: Display a login page, which includes login controls; In response to the login operation of the target object on the login control, the identity information of the target object is obtained; Verify the identity information of the target object; After the identity information of the target object is verified, different types of multimedia information are displayed; Based on the target object's selection operation of different types of multimedia information displayed, the target object's target multimedia preference information is obtained.

7. The method according to any one of claims 4 to 6, characterized in that, Determine the target multimedia information recommendation model corresponding to the selection control using the target multimedia information recommendation model, and generate and display target multimedia recommendation information for the target object, including: When the target multimedia information recommendation model includes a content filtering model, the multimedia content features of the multimedia information are extracted. Convert the multimedia content features into multimedia feature vectors; Based on the multimedia feature vectors, the multimedia similarity between different multimedia information is obtained; Based on the target multimedia preference information of the target object and the multimedia similarity between different multimedia information, target multimedia recommendation information is generated and displayed for the target object.

8. The method as described in claim 1, characterized in that, Determine the target multimedia information recommendation model corresponding to the selection control using the target multimedia information recommendation model, and generate and display target multimedia recommendation information for the target object, including: When the target multimedia information recommendation model includes a collaborative filtering model, an object multimedia matrix is ​​established based on object behavior data; Perform matrix decomposition on the multimedia matrix of the object to obtain the object feature matrix; Locality-sensitive hashing is performed on the object feature vectors in the object feature matrix to obtain the nearest neighbor objects of the target object. Based on the object behavior data of the neighboring objects, target multimedia recommendation information is generated and displayed for the target object.

9. The method as described in claim 1, characterized in that, Determine the target multimedia information recommendation model corresponding to the selection control using the target multimedia information recommendation model, and generate and display target multimedia recommendation information for the target object, including: When the target multimedia information recommendation model includes a hybrid recommendation model, and the hybrid recommendation model includes a collaborative filtering model and a content filtering model, the object behavior data of the target object is obtained; Based on the object behavior data of the target object, determine the weight exploration parameters; When it is determined that the system is in the exploration phase based on the weight exploration parameters, the first weight of the collaborative filtering model and the second weight of the content filtering model are randomly selected according to the probability of the weight exploration parameters. When it is determined that the system is in the utilization stage based on the weight exploration parameters, the first weight of the collaborative filtering model and the second weight of the content filtering model are determined based on the feedback data of the target object to the recommended multimedia information. The first multimedia recommendation information is determined based on the collaborative filtering model, and the second multimedia recommendation information is determined based on the content filtering model. Based on the first multimedia recommendation information and the first weight, and the second multimedia recommendation information and the second weight, target multimedia recommendation information for the target object is generated and displayed.

10. A multimedia information processing device, characterized in that, include: The display unit is used to display a multimedia information recommendation model selection page. The multimedia information recommendation model selection page includes multiple multimedia information recommendation model selection controls. Different multimedia information recommendation model selection controls correspond to different multimedia information recommendation models. The multimedia information recommendation model includes a target multimedia information recommendation model. The processing unit is configured to respond to a target object's selection operation of a target multimedia information recommendation model selection control among a plurality of multimedia information recommendation model selection controls, determine the target multimedia information recommendation model corresponding to the target multimedia information recommendation model selection control, and generate and display target multimedia recommendation information for the target object.

11. An electronic device, characterized in that, include: One or more processors; A storage device configured to store one or more programs that, when executed by one or more processors, cause the electronic device to perform the method of any one of claims 1 to 9.

12. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is run on a computer, it causes the computer to perform the method of any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 9.