Internet content intelligent recommendation method

Abstract

The invention provides an internet content intelligent recommendation method, which belongs to the technical field of information processing, and comprises the following steps: step 1, collecting multi-modal data of a user in an internet environment, and constructing a user portrait based on the multi-modal data; 2, behavior data and interest data of a user are obtained to update a user portrait; 3, acquiring a main recommendation content table and an alternative recommendation content table from a preset content recommendation library based on the user portrait; 4, generating a comprehensive content recommendation list based on the main recommendation content list and the alternative recommendation content list; and 5, obtaining real-time operation data of the user and updating the comprehensive content recommendation list. And the operation data of the user is responded in real time for updating, so that the recommendation effect and the user experience are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of information processing, and particularly to an intelligent recommendation method for Internet content. Background Art

[0002] With the development of the Internet and the explosive growth of information, it is difficult for users to quickly obtain content that highly matches their own needs. Therefore, intelligent recommendation technology has gradually become an important means to solve the problem of information overload.

[0003] Existing Internet content recommendation methods usually generate recommended content based on single behavioral data of users, combined with users' historical preferences and simple recommendation algorithms. However, there are significant technical defects, that is, they only rely on single behavioral data of users, ignore the comprehensiveness of multi-modal data, and it is difficult to update user portraits in real time, resulting in a low matching degree between the recommended content and the current interests of users, affecting the user experience.

[0004] Therefore, the present invention provides an intelligent recommendation method for Internet content. Summary of the Invention

[0005] The present invention provides an intelligent recommendation method for Internet content, which constructs and dynamically updates user portraits by combining multi-modal data, improves the comprehensive understanding of users' interests by the recommendation system, adopts a comprehensive generation mechanism of main recommendations and alternative recommendation lists, makes the recommended content more accurate and diversified, and can be updated in real time in response to users' operation data, thereby significantly improving the recommendation effect and user experience, and solving the defects in the prior art that multi-modal data of users cannot be fully utilized, user portraits cannot be updated in real time, and personalized recommended content cannot be generated.

[0006] The present invention provides an intelligent recommendation method for Internet content, including:

[0007] Step 1: Collect multi-modal data of users in the Internet environment and construct user portraits based on the multi-modal data;

[0008] Step 2: Obtain the behavioral data and interest data of users to update the user portraits;

[0009] Step 3: Obtain a main recommendation content list and an alternative recommendation content list from a preset content recommendation library based on the user portraits;

[0010] Step 4: Generate a comprehensive content recommendation list based on the main recommendation content list and the alternative recommendation content list;

[0011] Step 5: Obtain the real-time operation data of users and update the comprehensive content recommendation list.

[0012] The present invention provides an intelligent recommendation method for Internet content, which collects multi-modal data of users and constructs user portraits based on the multi-modal data, including:

[0013] Collect multi-modal data of users in the Internet environment and preprocess the collected multi-modal data;

[0014] Based on a preset deep learning model, perform feature fusion on the multi-modal data, and then construct a user portrait.

[0015] The present invention provides an intelligent recommendation method for Internet content, which obtains the behavior data and interest data of users to update the user portrait, including:

[0016] Obtain the historical behavior data of users to perform a first update on the user portrait;

[0017] Based on the interest data of users and the real-time behavior data of users within a specific recent time period, perform a second update on the first update result of the user portrait.

[0018] The present invention provides an intelligent recommendation method for Internet content. The first update is a static update, and the second update includes long-term interest update and short-term interest update.

[0019] The present invention provides an intelligent recommendation method for Internet content, which obtains the historical behavior data of users to perform a first update on the user portrait, including:

[0020] Collect the historical behavior data of users on the Internet platform, and based on the user historical behavior data, extract the long-term interest characteristics of users, and then form the interest characteristic values of historical behaviors;

[0021] Perform a first update on the long-term interest part in the user portrait based on the interest characteristic values of historical behaviors:

[0022]

[0023] where L i is the first update weight of the long-term interest in the user portrait for the i-th category of content, L 1i is the long-term interest weight in the user portrait for the i-th category of content, α is the influence factor of historical behaviors, w ij is the historical behavior weight of the j-th behavior performed by the user on the i-th category of content, f ij is the interest characteristic value of the j-th historical behavior performed by the user on the i-th category of content, T is the static update time, t ij is the time when the j-th behavior is performed by the user on the i-th category of content, γ is a control parameter for presetting the control decay rate, c i is the importance weight of the i-th category of content, m iThe number of types of behaviors performed by the user on the content of the i-th category.

[0024] The present invention provides an intelligent recommendation method for Internet content, which performs a second update on the first update result of the user portrait based on the user's interest data and the user's real-time behavior data, including:

[0025] Obtain the data input explicitly by the user, and dynamically update the long-term interest weight in combination with the behaviors within a preset recent time period:

[0026]

[0027] Among them, L1 i is the dynamic update weight of the long-term interest in the content of the i-th category in the user portrait, β is a preset dynamic adjustment coefficient, and β ∈ (0, 1), L i is the first update weight of the long-term interest in the content of the i-th category in the user portrait, d ij is the eigenvalue of the j-th behavior performed by the user on the content of the i-th category recently, h ij is the behavior weight of the j-th behavior performed by the user on the content of the i-th category recently, T1 is the second static update time, t1 ij is the time of the j-th behavior performed by the user on the content of the i-th category recently, r i is the content weight of the data input explicitly by the user, ∈ is a preset adjustment coefficient;

[0028] At the same time, in the process of dynamically updating the long-term interest weight, adjust the weights of the recent behaviors and historical behaviors within the preset time period through a preset time weighting model, so as to realize the dynamic update of the long-term interest of the user portrait;

[0029] Real-time monitor the operation behaviors of the user on the Internet platform, and construct the short-term interest eigenvalue of the user based on the real-time behavior data;

[0030] Update the short-term interest part in the user portrait based on the short-term interest eigenvalue:

[0031]

[0032] Among them, S i (t) is the short-term interest weight of the i-th category of content at time t, S i (0) is the short-term interest weight of the i-th category of content at the initial time, γ0 is the initial time decay coefficient, δ controls the change rate of the decay coefficient with time, q ij is the real-time behavior weight of the j-th behavior performed by the user on the content of the i-th category, t is the current time, e ij is the j-th behavior performed by the user on the content of the i-th category in real time, c iwhere \(w_i\) is the importance weight of the content of the \(i\)-th category, and \(\rho\) is the preset personalized decay parameter of the user;

[0033] Implement the second update based on the user's long-term dynamic interest update and short-term interest update.

[0034] The present invention provides an intelligent recommendation method for Internet content. The explicitly input data includes: interest tags actively selected by the user, subscribed content categories, and user preference settings.

[0035] The present invention provides an intelligent recommendation method for Internet content. Based on the user profile, obtain the main recommendation content table and the alternative recommendation content table from the preset content recommendation library, including:

[0036] Obtain several interest tags of the user based on the user profile, match the interest tags of the user with the ordinary content tags in the preset content recommendation library, select the ordinary content with high tag matching degree, and then determine several interest recommendation contents in the preset content recommendation library;

[0037] Match the interest tags of the user with the current hot content tags in the preset content recommendation library, select the hot content with high tag matching degree, and then obtain several hot contents;

[0038] Respectively obtain a specific number of interest recommendation contents and hot contents with tag matching degrees greater than the preset matching degree as the main recommendation contents, and the rest as the alternative recommendation contents;

[0039] Assign initial weights to the interest recommendation contents and hot contents based on a preset method;

[0040] Respectively obtain the behavior characteristic values of the user's historical click on interest recommendation contents and the behavior characteristic values of hot contents based on the user's historical behavior data;

[0041] Adjust the initial characteristic values of the interest recommendation contents and hot contents based on the behavior characteristic values of the user's historical click on interest recommendation contents, the behavior characteristic values of hot contents, and the preset weight adjustment method, and then obtain the recommendation weights of the interest recommendation contents and hot contents;

[0042] Generate the main recommendation content table and the alternative recommendation content table based on the recommendation weights of the interest recommendation contents and hot contents, the main recommendation contents, and the alternative recommendation contents.

[0043] Compared with the prior art, the beneficial effects of the present application are as follows:

[0044] By constructing and dynamically updating user profiles by integrating multimodal data, the comprehensive understanding of user interests by the recommendation system is improved. An integrated generation mechanism of primary recommendations and alternative recommendation lists is adopted to make the recommended content more accurate and diverse, and it can be updated in real time in response to users' operation data, thus significantly enhancing the recommendation effect and user experience, and solving the deficiencies in the prior art of being unable to fully utilize users' multimodal data, update user profiles in real time, and generate personalized recommended content. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0046] Figure 1 It is a schematic flowchart of an intelligent recommendation method for Internet content provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention with reference to the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0048] Embodiment 1

[0049] An embodiment of the present invention provides an intelligent recommendation method for Internet content, as Figure 1 shown, including:

[0050] Step 1: Collect multimodal data of a user in an Internet environment and construct a user profile based on the multimodal data;

[0051] Step 2: Obtain the behavior data and interest data of the user to update the user profile;

[0052] Step 3: Obtain a primary recommendation content list and an alternative recommendation content list from a preset content recommendation library based on the user profile;

[0053] Step 4: Generate an integrated content recommendation list based on the primary recommendation content list and the alternative recommendation content list;

[0054] Step 5: Obtain the real-time operation data of the user and update the integrated content recommendation list.

[0055] In this embodiment, the multimodal data includes, but is not limited to, behavioral data such as the user's click behavior, search history, social media interactions, purchase records, viewing habits, etc. The multimodal data also includes text data (comments, search keywords), image data (pictures browsed by the user), video data (video content watched by the user), audio data (listening records), and geographical location information, etc.;

[0056] In this embodiment, the user profile includes information such as the user's basic attributes (gender, age, region), long-term interest tags, recent behavioral characteristics, content preferences, etc. The user profile is the core of the entire recommendation system, reflecting the user's personalized characteristics and interests.

[0057] In this embodiment, the user's behavioral data is the operation or interaction behavior of the user in the Internet environment, such as clicking, browsing, searching, liking, commenting, etc., which can reflect the user's actual preference for a certain type of content. For example, if the user frequently clicks on technology news, these click records are the user's behavioral data;

[0058] In this embodiment, the user's interest data is the user's long-term preferences and interest points analyzed based on the user's historical behavioral data and multimodal data (such as pictures uploaded by the user, videos watched, music listened to). For example, if the user repeatedly watches football game-related content on multiple platforms, it is inferred that their interest lies in sports, especially football.

[0059] In this embodiment, the preset content recommendation library contains a large amount of content to be recommended. These contents may be pre-classified and stored according to different dimensions such as themes, types, popularity, etc. for use by the recommendation system. For example, the preset content recommendation library may include different types of content resources such as technology news, entertainment videos, learning materials, etc.;

[0060] In this embodiment, the main recommended content list is a list of contents that the recommendation system selects from the preset content library as the most matching ones for the user's current interests and behaviors. For example, for a user who loves technology, the main recommended content list may include the latest technology news, product review videos, etc.;

[0061] In this embodiment, the alternative recommended content list is a secondary content list generated by the recommendation system to supplement the main recommendation list, usually including contents related to the user's secondary interests or exploratory recommended contents. For example, although the user mainly focuses on technology, the alternative recommended content list may provide contents related to travel and music to enrich the user experience.

[0062] In this embodiment, the comprehensive content recommendation list combines the main recommendation content list and the alternative recommendation content list to form a personalized recommendation content list finally provided to the user. For example, the comprehensive content recommendation list includes both technology news and may also include some music videos or travel recommendation articles, providing diverse choices.

[0063] In this embodiment, the real-time operation data is the operation data generated during the user's current usage. For example, the user clicks on a certain video, exits a certain web page, or stays on a certain piece of content for a long time. For example, when the user clicks on a technology article in the recommendation list, the system will immediately record this operation and give feedback;

[0064] In this embodiment, updating the comprehensive content recommendation list is based on the user's real-time operation data to dynamically adjust the recommendation list, making the subsequent recommended content more matched with the user's latest interests. For example, after the user clicks on multiple music videos, the system may give priority to recommending more music-related content and reduce the recommendation of technology-related content.

[0065] The beneficial effects of the above technical solution are as follows: By combining multi-modal data construction and dynamically updating the user portrait, the recommendation system's comprehensive understanding of the user's interests is improved. Adopting the comprehensive generation mechanism of the main recommendation and alternative recommendation lists makes the recommended content more accurate and diverse, and can be updated in real time in response to the user's operation data, thus significantly improving the recommendation effect and user experience, and solving the defects in the prior art that cannot fully utilize the user's multi-modal data, update the user portrait in real time, and generate personalized recommendation content.

[0066] Embodiment 2

[0067] The embodiment of the present invention provides an intelligent recommendation method for Internet content, which collects the multi-modal data of the user and constructs a user portrait based on the multi-modal data, including:

[0068] Collect the multi-modal data of the user in the Internet environment and preprocess the collected multi-modal data;

[0069] Based on a preset deep learning model, perform feature fusion on the multi-modal data, and then construct a user portrait.

[0070] In this embodiment, preprocessing the collected multi-modal data is to perform word segmentation and word vectorization processing on the text data, extract features from the image and video data, and perform feature encoding on the audio data. Standardize the metadata such as geographical location and time.

[0071] In this embodiment, the preset deep learning model is a pre-trained neural network model for processing and analyzing multimodal data. It can extract useful information and features from complex data. After being trained with a large amount of sample data, it has powerful data processing and reasoning capabilities and is suitable for content recommendation. For example, a preset deep learning model can process text, image, and audio data simultaneously. By analyzing the articles the user browses, the pictures the user views, and the music the user listens to, it can understand the user's preferences and make recommendations.

[0072] In this embodiment, feature fusion refers to combining different features from multimodal data (such as text features, image features, audio features) to generate a more comprehensive and accurate user profile. By integrating information from different data sources, it can better capture the user's complex interests and behavior patterns. For example, visual features (such as color and style) are extracted from the user's pictures, keywords are extracted from the text, and emotional features are extracted from the music the user listens to. Finally, through feature fusion, a comprehensive user preference is generated. For example, the user may be interested in pop music and art-related content.

[0073] The beneficial effects of the above technical solution are as follows: By collecting the user's multimodal data and using the deep learning model for feature fusion, a more comprehensive user profile is constructed. This method breaks through the limitations of a single data source, combines multi-dimensional data (such as text, audio, video, etc.), and improves the accuracy and diversity of the user profile. Through the feature fusion of the deep learning model, the intelligent level of the recommendation system and the effect of personalized recommendation are enhanced, thereby improving the user experience.

[0074] Embodiment 3

[0075] An embodiment of the present invention provides an intelligent recommendation method for Internet content, which updates the user profile by obtaining the user's behavior data and interest data, including:

[0076] Obtaining the user's historical behavior data to perform a first update on the user profile;

[0077] Based on the user's interest data and the user's real-time behavior data within a recent specific time period, a second update is performed on the result of the first update of the user profile.

[0078] In this embodiment, the first update refers to the preliminary adjustment and improvement of the user profile according to the user's historical behavior data. This process aims to reflect the user's past behavior patterns and preferences, thereby laying a foundation for subsequent recommendations. If the user has frequently browsed technology articles in the past period, the system will increase the weight of the technology interest in the user profile during the first update.

[0079] In this embodiment, the second update is to further adjust the user profile based on the first update, combining the user's interest data and recent real-time behavior data. This update can more accurately reflect the user's current interest changes and behavior trends. Suppose the user starts to pay attention to healthy eating in the last week, the second update will add interests related to healthy eating to the user profile, thus adjusting the subsequent recommended content;

[0080] In this embodiment, the recent specific time period refers to the user's behavior data or activities within a relatively short time range (such as the past day, week or month). This time period is usually used to observe the user's immediate changes and new interests. If the system observes that the user has frequently watched fitness videos in the past three days, this period is the "recent specific time period", and the system will use this data for the second update.

[0081] The beneficial effects of the above technical solutions are as follows: By updating the user profile in stages, first performing the first update based on historical behavior data, and then performing the second update according to the user's interest data and recent and real-time behavior data, effectively combining long-term and short-term behaviors, dynamically optimizing the user profile, enabling the recommendation system to more accurately reflect the user's current interests and preferences. Through real-time adjustment, the relevance and timeliness of the recommended content are enhanced, improving the user's satisfaction and the intelligent recommendation effect of the system.

[0082] Embodiment 4

[0083] An embodiment of the present invention provides an intelligent recommendation method for Internet content. The first update is a static update, and the second update includes long-term interest update and short-term interest update.

[0084] In this embodiment, the static update refers to an update based on the user's long-term and relatively unchanged historical behavior data. The static update is usually one-time and does not change frequently, aiming to reflect the user's stable interests and long-term preferences. If the user has been reading technology news or purchasing electronic products for a long time, these behavior data will be used for static update, thus reflecting the label of "technology enthusiast" in the user profile;

[0085] In this embodiment, the long-term interest update refers to the continuous update of the user profile according to the user's behavior and interest changes over a long period of time. The long-term interest update aims to capture the user's deep interests and preference trends over time. For example, if the user has gradually shifted from reading technology news to focusing on health and fitness content in the past six months, through long-term interest update, the system will shift the focus of the user profile from technology to health-related content recommendations;

[0086] In this embodiment, short-term interest update refers to quickly adjusting the user profile according to the user's immediate behaviors and temporary interests within a short period of time. Short-term interest update captures the user's temporary interests and ensures that the recommended content meets the user's current needs or mood. If the user has searched for and browsed a lot of travel guides in recent days, the system will, through short-term interest update, increase the recommendation of travel-related content within a short time, even if this is not the user's long-term interest.

[0087] The beneficial effects of the above technical solution are as follows: The first update of the user profile is performed through static update, and the second update is achieved by combining the dynamic adjustment of long-term and short-term interests. Long-term interest update captures the user's stable preferences, while short-term interest update reflects the user's immediate needs, ensuring the timeliness and accuracy of the recommended content. By integrating static and dynamic interest features, the flexibility, personalization of the recommendation system, and its adaptability to user behaviors are effectively improved, enhancing the relevance of the recommendation results and user satisfaction.

[0088] Embodiment 5

[0089] The embodiment of the present invention provides an intelligent recommendation method for Internet content. The first update of the user profile is obtained by acquiring the user's historical behavior data, including:

[0090] Collect the user's historical behavior data on the Internet platform. Based on the user's historical behavior data, extract the user's long-term interest features, and then form the interest feature values of the historical behaviors.

[0091] Perform the first update on the long-term interest part in the user profile based on the interest feature values of the historical behaviors:

[0092]

[0093] where L i is the first update weight of the long-term interest in the user profile for the i-th category of content, L 1i is the long-term interest weight in the user profile for the i-th category of content, α is the influence factor of the historical behavior, w ij is the historical behavior weight of the j-th behavior performed by the user for the i-th category of content, f ij is the interest feature value of the j-th historical behavior performed by the user for the i-th category of content, T is the static update time, t ij is the time when the j-th behavior is performed by the user for the i-th category of content, γ is the control parameter for presetting the control decay rate, c i is the importance weight of the i-th category of content, m i is the number of types of behaviors performed by the user for the i-th category of content.

[0094] In this embodiment, historical behavior data refers to the records of a user's past operations, browsing, clicks, searches, purchases, etc. on the Internet platform. It is the activity traces accumulated by the user over a relatively long period of time and is used to analyze the user's long-term interests and behavior patterns. For example, all the article browsing records, video viewing records, click history of shopping websites, etc. of the user in the past year are historical behavior data;

[0095] In this embodiment, long-term interest features are the features extracted from the user's historical behavior data that reflect the user's stable interests within a relatively long time range. These features are used to describe the fields or categories that the user continuously pays attention to. For example, if the user has browsed blogs related to programming and purchased programming books many times in the past year, the system will mark "programming" as the long-term interest feature of this user;

[0096] In this embodiment, the historical behavior interest feature value is a numerical value calculated based on the user's historical behavior data and is used to quantify the user's long-term interest intensity in a certain type of content. Through these numerical values, the system can more specifically understand the user's preference degree for different types of content. Example: If the user's browsing times and durations of "technology" content are significantly higher than other categories, the system will assign a higher historical behavior interest feature value to this category. For example, the feature value of the "technology" category is 0.8, while the feature value of the "entertainment" category is 0.3;

[0097] In this embodiment, the long-term interest part refers to the part of the user profile that reflects the user's long-term interests. It contains the long-term interest weights of the user for different content categories, and these weights will affect the content direction of the system's long-term recommendations. For example, in the long-term interest part of the user profile, the user's weight for "finance" is 0.7, and the weight for "sports" is 0.4, which means that the system is more inclined to recommend finance-related content.

[0098] In this embodiment, the influence factor of historical behavior is a coefficient used to measure the influence degree of historical behavior data in the user profile. This factor determines the weight of historical behavior data when updating the user profile, that is, the contribution size of historical behavior to long-term interests. If the historical behavior of a certain user has a high influence factor (such as 0.9) set in the system, the system will rely more on past behavior data to infer its long-term interests; on the contrary, if the influence factor is low (such as 0.4), the role of historical behavior will be relatively weakened;

[0099] In this embodiment, the control parameter for presetting the attenuation rate refers to the parameter used to control the attenuation speed of historical behavior data over time. As time goes by, the influence of the user's past behavior gradually weakens, and the attenuation rate determines the speed of this weakening. If a user has frequently browsed travel guides in the past year but has not paid attention to travel content in recent months, the system uses the attenuation rate parameter to gradually reduce the influence of past behavior, making the content recommended by the system more in line with the user's recent interest changes.

[0100] The beneficial effects of the above technical solution are as follows: By analyzing the user's historical behavior data, extracting the user's long-term interest characteristics and updating the user profile, using the influence factor and behavior weight of historical behavior, combining the static update time and the attenuation rate parameter, dynamically adjusting the long-term interest part in the user profile, effectively capturing the changes in the user's long-term interests, making the user profile more accurate and dynamic, enhancing the personalization and long-term recommendation accuracy of the recommendation system, improving the user experience, and having strong adaptability and intelligence.

[0101] Embodiment 6

[0102] The embodiment of the present invention provides an intelligent recommendation method for Internet content, which performs a second update on the first update result of the user profile based on the user's interest data and the user's real-time behavior data, including:

[0103] Obtain the data input explicitly by the user, and dynamically update the long-term interest weight in combination with the behavior within a preset recent time period:

[0104]

[0105] Among them, L1 i is the dynamic update weight of the long-term interest in the user profile for the content of the i-th category, β is the preset dynamic adjustment coefficient, and β ∈ (0, 1), L i is the first update weight of the long-term interest in the user profile for the content of the i-th category, d ij is the eigenvalue of the user's execution of the j-th behavior for the content of the i-th category recently, h ij is the behavior weight of the user's execution of the j-th behavior for the content of the i-th category recently, T1 is the second static update time, t1 ij is the time when the user recently executed the j-th behavior for the content of the i-th category, r i is the content weight of the data input explicitly by the user; ∈ is the preset adjustment coefficient;

[0106] At the same time, during the process of dynamically updating the long-term interest weight, adjust the weights of the recent behavior and the historical behavior within the preset time period through the preset time weighting model, so as to realize the dynamic update of the long-term interest of the user profile;

[0107] Monitor the operation behavior of users on the Internet platform in real time, and construct short-term interest eigenvalue of users based on real-time behavior data;

[0108] Update the short-term interest part in the user portrait based on the short-term interest eigenvalue:

[0109]

[0110] where S i (t) is the short-term interest weight of the i-th type of content at time t, S i (0) is the short-term interest weight of the i-th type of content at the initial time, γ0 is the initial time decay coefficient, δ controls the change rate of the decay coefficient over time, q ij is the real-time behavior weight of the j-th behavior executed by the user for the content of the i-th category, t is the current time, e ij is the j-th behavior executed by the user in real time for the content of the i-th category, c i is the importance weight of the content of the i-th category, and ρ is the preset personalized decay parameter of the user;

[0111] Implement the second update based on the long-term dynamic interest update and short-term interest update of the user.

[0112] In this embodiment, the preset dynamic adjustment coefficient refers to the coefficient used to adjust the influence degree of the user's recent behavior on the long-term interest weight when dynamically updating the user's long-term interest weight. This coefficient determines the influence of the user's recent behavior data in the user portrait. Compared with historical data, the preset dynamic adjustment coefficient is used to balance the weights of recent behavior and long-term interest. If the preset dynamic adjustment coefficient is set to 0.7, it means that when the system considers the user's recent behavior for a certain category of content, it will make a large adjustment to its dynamic influence to reflect the importance of recent behavior;

[0113] In this embodiment, the preset adjustment coefficient is a parameter that adjusts various weights or behavior eigenvalues during the entire user portrait update process. This coefficient is used to control the adjustment degree of the system's interest in the user in different situations, making the recommendation algorithm more flexible. For example, when adjusting the weight of the user's explicit input data, the system may use the preset adjustment coefficient (such as 0.5) to ensure that the explicit input has an appropriate impact on the recommendation result without completely dominating the recommendation result.

[0114] In this embodiment, the preset time-weighted model refers to the weighted processing of a user's behavior through a weight allocation model within a set time range. This model is used to balance the user's historical behavior and recent behavior, ensuring that the user's recent behavior does not completely overwrite historical behavior, or vice versa, historical behavior does not completely suppress recent behavior. For example, the system can set a preset time-weighted model of three months. Within this time period, the user's behavior will receive a higher weight, while the behavior data beyond three months will gradually decay, thereby weakening their influence.

[0115] In this embodiment, the personalized decay parameter is a personalized setting used to control the system's decay of the user's interests according to time. Different users have different behavior patterns and interest change rhythms. The personalized decay parameter helps the system formulate a unique interest decay strategy for each user. If a user's interest in a certain type of content changes rapidly, the system can accelerate the decay of outdated behavior by setting a larger personalized decay parameter (e.g., 0.9) to ensure that the recommended content is consistent with the user's latest interests; while for another user with stable interests, a smaller decay parameter (e.g., 0.2) is set to extend the influence of historical behavior.

[0116] The beneficial effects of the above technical solutions are as follows: By combining the user's explicit input and real-time behavior data, dynamically adjusting the long-term and short-term interests in the user profile, realizing the second update of the user profile, accurately capturing recent and historical behaviors through the time-weighted model, optimizing the dynamic update of long-term interests, monitoring the user's behavior in real time, updating the user profile based on short-term interest eigenvalues, more accurately reflecting the user's interest changes, enhancing the personalization and accuracy of recommendations, significantly enhancing the intelligence and user experience of the recommendation system, and effectively improving the relevance and satisfaction of content recommendations.

[0117] Embodiment 7

[0118] The embodiment of the present invention provides an intelligent recommendation method for Internet content. The explicitly input data includes: interest tags actively selected by the user, subscribed content categories, and user preference settings.

[0119] In this embodiment, the interest tags actively selected by the user are tags that the user can actively select on the platform related to their interests. These tags can help the system better understand the user's preferences and provide customized content recommendations. Since the interest tags are actively input by the user, they reflect the user's preferences most accurately. For example, on a news platform, the user can select tags such as "technology", "finance", "sports", etc., indicating that the user is more interested in content in these fields. The system can recommend corresponding articles or videos based on these tags;

[0120] In this embodiment, the subscribed content categories are those that the user actively pays attention to certain specific types of content sources, which means that the user hopes to continuously obtain updated content of these categories. This is usually reflected in the user subscribing to certain types of news, bloggers, or channels. Based on this, the system pushes relevant updates to the user. For example, if the user subscribes to channels or playlists of "classical music" and "popular music" on a music platform, the system will recommend newly released classical or popular music according to the user's subscription habits;

[0121] In this embodiment, user preference settings refer to the specific content consumption preferences clearly expressed by the user through system settings. These preference settings can affect how the system displays content, including even the frequency and type of content recommendations. For example, the user may filter out content categories that they are not interested in through preference settings, or choose the frequency of receiving recommendations. On a video website, the user can set video categories that they are "not interested in", such as "beauty" or "games", then the system will reduce or stop recommending such videos.

[0122] The beneficial effects of the above technical solutions are: content recommendations are made based on the data explicitly input by the user (interest tags, subscribed categories, preference settings), which improves the recommendation accuracy, strengthens the user's sense of control, combines explicit data with recommendation algorithms, provides personalized and diverse content recommendations, and optimizes the user experience.

[0123] Embodiment 8

[0124] An embodiment of the present invention provides an intelligent recommendation method for Internet content. Based on a user profile, a main recommended content table and an alternative recommended content table are obtained from a preset content recommendation library, including:

[0125] Obtain several interest tags of the user based on the user profile, match the interest tags of the user with the ordinary content tags in the preset content recommendation library, select the ordinary content with a high tag matching degree, and then determine several interest recommended contents in the preset content recommendation library;

[0126] Match the interest tags of the user with the current hot content tags in the preset content recommendation library, select the hot content with a high tag matching degree, and then obtain several hot contents;

[0127] Respectively obtain a specific number of interest recommended contents and hot contents with a tag matching degree greater than the preset matching degree as the main recommended contents, and the rest as the alternative recommended contents;

[0128] Assign initial weights to the interest recommended contents and hot contents respectively based on a preset method;

[0129] Respectively obtain the behavioral characteristic values of the user's historical click on interest recommended contents and the behavioral characteristic values of hot contents based on the user's historical behavior data;

[0130] Based on the behavioral feature values of content recommended according to the user's historical click interests, the behavioral feature values of hot content, and a preset weight adjustment method, adjust the initial feature values of the interest-recommended content and the hot content, and then obtain the recommendation weights of the interest-recommended content and the hot content;

[0131] Generate a main recommendation content table and an alternative recommendation content table based on the recommendation weights, main recommended content, and alternative recommended content of the interest-recommended content and the hot content.

[0132] In this embodiment, the interest tags are tags generated based on the user interest characteristics recorded in the user profile, indicating the user's preference for certain types of content. Through these tags, the system can identify the user's interests in certain fields and thus make content recommendations. If a user is interested in technology news, photography, and tourism, then the user's interest tags may include "technology", "photography", "tourism";

[0133] In this embodiment, the general content tags are content classification tags in the preset content recommendation library, used to mark the category to which the content belongs. By matching the user's interest tags with these general content tags, the content that the user may be interested in can be found. For example, a technology news article may be tagged with "technology", "AI", "innovation", and these are the general content tags of this content;

[0134] In this embodiment, the interest-recommended content is a series of content highly relevant to the user's interests screened out after matching the user's interest tags with the general content tags. These contents conform to the long-term interests and historical behaviors in the user profile. If the user's interest tag is "photography", then the recommended interest content may include photography technique articles, photography equipment reviews, etc.

[0135] In this embodiment, the hot content refers to the content that is very popular or has a high degree of discussion on the Internet platform currently. By matching the user's interest tags with the hot content tags, the popular content that the user is interested in can be recommended. For example, when a technology company releases a new smartphone, the related news and comments are hot content.

[0136] In this embodiment, the tag matching degree measures the similarity or relevance between the user's interest tags and the general content tags or hot content tags of the content. The higher the matching degree, the more likely the content is to conform to the user's interests. If the user's interest tag is "photography", and the general content tags of an article include "photography", "equipment", then the tag matching degree may be 0.9, while the matching degree with the content related to "sports" may be 0.2;

[0137] In this embodiment, the preset matching degree is a threshold set by the system for screening recommended content. Only the content with a matching degree between the user's interest tags and the content tags higher than this threshold will be selected into the recommendation list. For example, if the preset matching degree set by the system is 0.7, only the content with a matching degree higher than 0.7 will be recommended to the user;

[0138] In this embodiment, the specific number refers to the number of recommended content selected from the content with a matching degree higher than the preset matching degree. The system restricts the number of recommended content according to these numbers. For example, if the specific number is set to 5, that is, 5 pieces of content meeting the conditions are selected and recommended to the user.

[0139] In this embodiment, the preset method is the rule or calculation method used by the system when assigning an initial weight to the content. Usually, the initial weight is set according to the type of the content, the user's interest weight, or the timeliness of the content. For example, the content is weighted according to the release time of the content and the user's recent behavior. For example, the newly released content may get a higher initial weight;

[0140] In this embodiment, the initial weight is the basic weight set for each piece of recommended content or hot content, which determines the initial sorting position of the content in the recommendation list. For example, a newly released hot news may be given an initial weight of 0.8, while an older article may only have 0.5.

[0141] In this embodiment, the behavior feature value is a numerical value calculated based on the user's past click, browse, interaction and other behaviors, which is used to reflect the user's interest intensity in a certain type of content. For example, if the user often clicks on articles related to "photography", then the behavior feature value of this type of content will be relatively high, perhaps 0.9, while the user rarely clicks on content related to "sports", and its feature value may be 0.3;

[0142] In this embodiment, the preset weight adjustment method refers to the method of adjusting the initial weight according to the user's historical behavior data and interest tags. This method is used to dynamically update the weight of the recommended content to more accurately reflect the user's current interest. If the user has frequently clicked on "technology" content in the past, the system increases the weight of the relevant content through the weight adjustment method, making it rank higher in the recommendation list;

[0143] In this embodiment, the recommended weight is the weight finally generated by the system for a certain piece of content, which combines factors such as the initial weight, the user behavior feature value, and the preset weight adjustment method, and determines the final sorting of the content in the recommendation list. For example, the initial weight of a certain piece of content is 0.6, and after the user's click behavior and weight adjustment, the recommended weight becomes 0.85, making its position in the recommendation list rise.

[0144] The beneficial effects of the above technical solution are as follows: By matching the interest tags with the ordinary content and hot content tags through the user portrait, and dynamically adjusting the recommendation weight in combination with the user's historical behavior data, the accuracy and diversity of the recommended content are improved. Adjusting the recommendation weight based on the user's behavior ensures the continuous optimization of personalized recommendations. The combination of the main recommendation and the alternative recommendation enhances the flexibility and diversity of the user experience.

[0145] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

[0146] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus the necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. And these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent recommendation of Internet content, characterized in that: include: Step 1: Collect multimodal data of users in the Internet environment and build user portraits based on the multimodal data; Step 2: Obtain the user's behavior data and interest data to update the user portrait; Step 3: Obtain a main recommended content table and an alternative recommended content table from a preset content recommendation library based on the user portrait; Step 4: Generate a comprehensive content recommendation list based on the main recommended content table and the alternative recommended content table; Step 5: Obtain the user's real-time operation data and update the comprehensive content recommendation list.

2. The method for intelligently recommending Internet content according to claim 1, characterized in that: Collect multimodal data of users and build user portraits based on multimodal data, including: Collect multimodal data of users in the Internet environment and pre-process the collected multimodal data; Based on the preset deep learning model, the features of multimodal data are fused to build user portraits.

3. The method for intelligently recommending Internet content according to claim 1, characterized in that: Obtain user behavior data and interest data to update user profiles, including: Obtain the user's historical behavior data to perform the first update on the user profile; A second update is performed on the first update result of the user portrait based on the user's interest data and the user's recent specific time period and real-time behavior data.

4. The method for intelligently recommending Internet content according to claim 1, characterized in that: The first update is a static update, and the second update includes a long-term interest update and a short-term interest update.

5. The method for intelligently recommending Internet content according to claim 3, characterized in that: Obtain the user's historical behavior data to perform the first update of the user profile, including: Collect historical behavior data of users on the Internet platform, extract the long-term interest characteristics of users based on the historical behavior data, and then form the interest characteristic value of historical behavior; The first update of the long-term interest part in the user portrait is performed based on the interest feature value of historical behavior: Among them, L i is the first update weight of the user's long-term interest in the content of the i-th category, L 1i is the long-term interest weight of the content in the i-th category in the user portrait, α is the influencing factor of historical behavior, and w ij is the historical behavior weight of the jth behavior performed by the user on the i-th category of content, f ij is the interest feature value of the user's jth historical behavior on the i-th category of content, T is the static update time, t ij is the time for the user to perform the jth behavior on the i-th category of content, γ is the control parameter for the preset decay rate, c i is the importance weight of the content of the i-th category, m i is the number of actions performed by the user on the content of the i-th category.

6. The method for intelligently recommending Internet content according to claim 3, characterized in that: Performing a second update on the first update result of the user portrait based on the user's interest data and the user's real-time behavior data includes: Obtain the data explicitly input by the user, and dynamically update the long-term interest weight based on the recent behavior in the preset time period: Among them, L1 i is the dynamic update weight of the user's long-term interest in the content of the i-th category in the user portrait, β is the preset dynamic adjustment coefficient, and β∈(0,1), L i is the first update weight of the user's long-term interest in the content of the i-th category, d ij is the characteristic value of the user's recent j-th behavior on the i-th category of content, h ij is the behavior weight of the jth behavior recently performed by the user on the i-th category of content, T1 is the second static update time, t1 ij is the time when the user recently performed the jth behavior on the i-th category of content, r i is the content weight of the data explicitly input by the user, ∈ is the preset adjustment coefficient; At the same time, in the process of dynamically updating the long-term interest weight, the weights of recent behaviors and historical behaviors within a preset time period are adjusted through a preset time weighting model, thereby realizing the dynamic update of the long-term interest of the user portrait; Monitor users' operation behaviors on the Internet platform in real time, and build users' short-term interest feature values ​​based on real-time behavior data; Update the short-term interest part of the user portrait based on the short-term interest feature value: Among them, S i (t) is the short-term interest weight of the i-th content at time t, S i (0) is the short-term interest weight of the i-th category content at the initial time, γ0 is the initial time attenuation coefficient, δ controls the rate of change of the attenuation coefficient over time, and q ij is the real-time behavior weight of the jth behavior performed by the user on the i-th category of content, t is the current time, e ij is the jth behavior performed by the user on the content of the i-th category in real time, c i is the importance weight of the content of the i-th category, ρ is the preset user's personalized attenuation parameter; The second update is achieved based on the user's long-term dynamic interest update and short-term interest update.

7. The method for intelligently recommending Internet content according to claim 1, characterized in that: Explicitly input data includes: interest tags actively selected by users, subscribed content categories, and user preferences.

8. The method for intelligently recommending Internet content according to claim 1, characterized in that: Based on the user profile, the main recommended content table and the alternative recommended content table are obtained from the preset content recommendation library, including: Based on the user portrait, several interest tags of the user are obtained, and the user's interest tags are matched with common content tags in the preset content recommendation library, and common content with high tag matching degree is selected, and then several interest recommendation contents are determined in the preset content recommendation library; Match the user's interest tags with the current hot content tags in the preset content recommendation library, select hot content with high tag matching, and then obtain several hot content; Obtain a specific number of interest recommendation contents and hot topics whose tag matching degree is greater than a preset matching degree as the main recommendation contents, and the rest as the alternative recommendation contents; Based on the preset method, initial weights are assigned to interest recommendation content and hot content respectively; Based on the user's historical behavior data, the behavior feature values ​​of the user's historical click interest recommendation content and the behavior feature values ​​of the hot content are obtained respectively; Based on the behavior characteristic values ​​of the user's historical clicks on the interest-recommended content, the behavior characteristic values ​​of the hot content, and a preset weight adjustment method, the initial characteristic values ​​of the interest-recommended content and the hot content are adjusted to obtain the recommendation weights of the interest-recommended content and the hot content; Based on the recommendation weights of the interest recommended content and the hot content, the main recommended content and the alternative recommended content, a main recommended content table and an alternative recommended content table are generated.

Images