Short video recommendation method and system based on real-time interest enhancement

Abstract

This invention provides a short video recommendation method and system based on real-time interest enhancement. The method includes: dividing the user's historical interaction sequence into three dimensions: long-term, short-term, and real-time; carefully designing different network structures to extract interest features for each dimension; effectively integrating the user's long-term interest preferences, short-term behavioral changes, and immediate needs through a Gate Recurrent Unit (GRU) to adapt to changes in user interests at different time scales; and strengthening the weight of real-time interests to highlight responses to the user's current behavior and needs. Experiments on multiple datasets and comparisons with baseline algorithms demonstrate the effectiveness of the algorithm. This invention solves the technical problem of accurately personalized recommendations based on the user's current interests.

Description

Technical Field

[0001] This invention relates to the field of recommendation algorithms, specifically to a short video recommendation method and system based on real-time interest enhancement. Background Technology

[0002] With the continuous development of the mobile internet, short video platforms have become an important entertainment and social tool for the new generation of internet users. Globally, hundreds of millions of users consume a large amount of short video content every day, demonstrating the extremely high penetration rate of short video platforms among internet users. Compared to long videos, short videos are characterized by their shorter playback duration, leading users to browse more short videos during a single interaction. Therefore, using users' historical interaction information for serialization modeling is an effective way to improve user experience.

[0003] For example, the existing invention patent application document CN117150075A, entitled "A Short Video Intelligent Recommendation System Based on Data Analysis," includes: a management platform, a data acquisition unit, a video segmentation unit, a popularity analysis unit, a built-in traffic analysis unit, a display execution unit, a publishing supervision and analysis unit, and a monitoring analysis unit. This invention performs video segmentation and supervision analysis on the historical information of the users to be recommended, in order to analyze the types of short videos liked by the users based on historical information feedback, so as to reasonably recommend videos based on their interests. Furthermore, it evaluates and analyzes the interest types by collecting their own data and dissemination data, that is, from the perspectives of the popularity brought by the publisher and the popularity generated by the video after its publication. The existing invention patent application document CN114896454A, entitled "A Short Video Data Recommendation Method and System Based on Tag Analysis," describes a method that includes: loading each first user behavior activity data into a defined knowledge entity network for rule-based transformation to generate corresponding second user behavior activity data, thereby improving the reliability of subsequent mining processing; dividing short video recommendation data into multiple short video recommendation services based on different short video recommendation dimensions; loading the corresponding second user behavior activity data into the corresponding short video recommendation services; extracting all second user behavior activity data associated with each short video recommendation service; loading this data into the corresponding interest heatmap output model to generate corresponding user interest tag heatmap trends; further determining the key user interest tags corresponding to the target user; and then recommending short videos to the target user based on these tags. This process involves further personalized recommendations after mining user interests across different short video recommendation dimensions. However, existing recommendation algorithms have shortcomings in adjusting and optimizing the recommendation list based on real-time user feedback. Designing a recommendation algorithm suitable for short video platforms that considers real-time user interests and ensures accuracy in recommending content that users are currently interested in still requires further in-depth research. Furthermore, since the short videos viewed by users constitute an interactive sequence, and these videos are short in duration, this sequence is characterized by its compact timeframe and dense user feedback. Therefore, utilizing this concentrated user feedback within a short period for recommendations presents a degree of complexity. Finally, considering users' short-term and real-time interests while still retaining some long-term interests is a direction worthy of further research.

[0004] In summary, existing technologies have the technical problem of failing to provide accurate personalized recommendations based on users' current interests. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to solve the technical problem in the prior art that it is difficult to make accurate personalized recommendations based on the user's current interests.

[0006] This invention solves the above-mentioned technical problems by employing the following technical solution: A short video recommendation method based on real-time interest enhancement includes:

[0007] S1. In the embedding representation layer, the short video browsing sequence is embedded and represented to obtain the user's long-term historical interaction sequence, short-term historical interaction sequence, and real-time historical interaction sequence, which are used to extract long-term interest features, short-term interest features, and real-time interest features. Among them, the short video browsing sequence is transformed into a directed graph, and the short videos in the short video browsing sequence are used as graph nodes. Based on the user's video click information, the edges between graph nodes are determined, the standardized weights of the edges are set, and a gated graph neural network is used to model and obtain the corresponding vectors of the nodes in the directed graph to update the node information. Sequence features are obtained by learning through mutual communication between nodes.

[0008] S2. In the multi-interest extraction layer, long-term interest extraction, short-term interest extraction, and real-time interest extraction operations are performed based on sequence features. The user's long-term interest features, short-term interest features, and real-time interest features are integrated through a gated recurrent unit (GRU). The real-time interest features are weighted and strengthened to perform corresponding operations based on the user's current behavior needs. The multi-interest extraction layer includes, but is not limited to, a long-term interest extraction module, a short-term interest extraction module, and a real-time interest extraction module.

[0009] S3. In the interest fusion layer, a real-time interest enhancement fusion network based on a gating mechanism is adopted. The proportion of differential interests is adjusted according to the weight coefficient. An interest fusion module is designed using a gating loop unit to fuse long-term interest features, short-term interest features and real-time interest features to obtain sensitive data on the dynamic changes of user interests.

[0010] This invention presents a short video recommendation algorithm based on real-time interest enhancement. It divides user historical interaction sequences into three dimensions: long-term, short-term, and real-time. Different network structures are carefully designed to extract interest features for each dimension. A Gate Recurrent Unit (GRU) effectively integrates users' long-term interest preferences, short-term behavioral changes, and immediate needs, adapting to changes in user interests across different time scales. Furthermore, by strengthening the weight of real-time interests, the algorithm prominently responds to users' current behaviors and needs. Experiments on multiple datasets, compared with baseline algorithms, demonstrate the effectiveness of this algorithm.

[0011] In a more specific technical solution, in S1, the following logic is used to learn and obtain sequence features:

[0012]

[0013]

[0014] In the formula, and These are the learnable parameters in the model. A list of short videos representing user interaction, r s，i Indicates the update gate, z s，i denoted as the reset gate, ⊙ as the Hadamard product between matrices, and σ as the sigmoid activation function.

[0015] In a more specific technical solution, the inter-node communication and learning in S1 includes:

[0016] S11. Perform modeling operations based on nodes, standardized weights, and edges;

[0017] S12. Based on the short video browsing sequence, a graph structure representation is obtained;

[0018] S13. Based on the graph structure representation, calculate the output edge weights and input edge weights of the nodes to obtain the out-degree matrix and in-degree matrix;

[0019] S14. Using a gated graph neural network, learn the hidden representation of items in the short video browsing sequence. The interaction sequence information is preserved through a gating mechanism to obtain and output the embedding representation corresponding to each node, which is used as the node's corresponding vector.

[0020] In a more specific technical solution, the long-term interest extraction module in S2 uses long-term historical data to model user behavior, calculate attention weights, and perform weighted focusing operations on different parts of the short video browsing sequence to preserve long-term interest features.

[0021] In a more specific technical solution, within the real-time interest extraction module, an attention-based real-time interest extraction network is used to calculate the attention score of real-time interest relative to the sequence behavior. The importance information in the attention score is normalized and weighted multiplied using the softmax function to obtain the real-time interest features. The normalization and weighted multiplication operations are performed using the following logic:

[0022]

[0023] f(q,i)=(i·tanh(W f ·q+b f ))·W h

[0024] In the formula, e i This represents the embedding representation of each item in a short-term user behavior sequence. An embedding representation of a user's real-time interests is given, where f(q, i) is an attention method that defines the weights of videos in a real-time sequence relative to each item in a short-term behavior sequence. f With wh These are all learnable parameter matrices.

[0025] This invention addresses the problem of user recommendation lists needing to be enhanced with real-time user feedback, and designs a short video sequence recommendation method based on real-time interest enhancement. The model mainly consists of four parts: an embedding representation layer, a multi-interest extraction layer, an interest fusion layer, and a prediction layer. The embedding representation layer is used to obtain the embedding representation of the short video sequence features.

[0026] In a more specific technical solution, during the weighted focusing operation, a set of shallow neural networks is used to concatenate the short video embedding representations from the short video browsing sequence with the target short video embedding representation to obtain attention weights. The shallow neural network includes a Dice activation function and a linear transformation component. The attention weights are processed, and Sum Pooling is performed to obtain long-term interest features. The long-term interest features are calculated using the following logic:

[0027]

[0028] In the formula, {e1, e2, ..., e H Let} represent the user behavior sequence, H be the length of the behavior sequence, and V be the user behavior sequence. A For the embedding representation of the target video to be predicted, a(·) is a small feedforward neural network, and the output of the small feedforward neural network is the weight of video j in the user's long-term viewing sequence; the long-term interest feature h is obtained by weighting the embedding representation of each video in the long-term viewing video sequence. long .

[0029] In a more specific technical solution, the short-term interest extraction module in S2 captures the interaction features between adjacent items. Using a pre-built convolutional neural network and a sliding window, it captures these interaction features to obtain long-range dependencies between items. Then, it combines horizontal and vertical convolutional filters to identify the overall sequence pattern of user interests. Finally, it connects and processes these filters to obtain the interaction features between adjacent items, thus yielding the user's short-term interest feature h. short .

[0030] The multi-interest extraction layer of this invention extracts long-term, short-term, and real-time interests from the interaction sequence by designing different network structures. Long-term interests reflect the user's long-term preferences, short-term interests capture the user's recent behavioral changes, and real-time interests focus on the user's current immediate needs.

[0031] In a more specific technical solution, a horizontal convolutional filter is used to perform a sliding extraction operation on the item matrix. The following logic is used to perform the East China extraction operation to obtain the interaction features between adjacent items:

[0032]

[0033] In the formula, S i:i+h-1 This represents rows i to i+h-1 in the sequence matrix. This represents a horizontal convolutional filter, where φ is a nonlinear activation function. This is the output value of the i-th convolution. This represents the set of all feature vectors extracted by the horizontal filter numbered k during the sliding process. hor This represents the final output obtained by concatenating the max-pooling results of filters with different numbering levels.

[0034] In a more specific technical solution, a vertical convolutional filter is used to perform a sliding feature extraction operation along the vector dimension of the item sequence matrix to extract the interaction features between different dimensions of the same item. Max pooling is then used to process and concatenate the output of the vertical convolutional filter to obtain the interaction features between adjacent items. Specifically, the sliding extraction operation using the vertical convolutional filter follows the logic below:

[0035]

[0036] In the formula, This represents the vertical convolution filter numbered k. This represents the set of all features extracted by the vertical convolution filter numbered k during the sliding process. Represents the j-th column of the item sequence matrix, o ver The final output is obtained by concatenating vertical convolutional filters with different numbers after performing a max pooling operation. Represents a vertical convolution filter, n v This represents the number of vertical convolution filters.

[0037] In a more specific technical solution, within S3, the following logic is used to obtain the sensitivity data for dynamic changes in interest:

[0038]

[0039] u=α·[u short ::u long ]+(1-α)u real

[0040] In the formula, FFN represents a feedforward neural network, and GRU represents a gated recurrent unit. An embedding representation of a user's long-term behavior sequence. An embedded representation of a short-term sequence of user behaviors.

[0041] The feature fusion layer of this invention fuses features extracted from the interest extraction layer based on a gated recurrent unit, and assigns more weight to real-time interest features through a gating mechanism, thereby making the prediction results more consistent with the user's current interests. In the prediction layer, the features obtained from the previous interest fusion are subjected to multi-layer linear transformations and activations to obtain the final predicted score. Finally, the top K predicted scores are selected as the recommendation list. This invention can effectively obtain the user's β interest, improve recommendation performance, and enhance user experience.

[0042] In more specific technical solutions, short video recommendation systems based on real-time interest enhancement include:

[0043] The embedding representation layer is used to embed the short video browsing sequence into a representation, dividing it into long-term historical interaction sequences, short-term historical interaction sequences, and real-time historical interaction sequences for extracting long-term interest features, short-term interest features, and real-time interest features. Specifically, the short video browsing sequence is transformed into a directed graph, with the short videos in the sequence as graph nodes. Based on the user's video click information, the edges between graph nodes are determined, and standardized weights are set for the edges. A gated graph neural network is used to model the corresponding vectors of the nodes in the directed graph to update the node information. Sequence features are obtained through inter-node communication and learning.

[0044] The multi-interest extraction layer is used to perform long-term interest extraction, short-term interest extraction, and real-time interest extraction operations based on sequence features. Through a gated recurrent unit (GRU), it integrates the user's long-term interest features, short-term interest features, and real-time interest features, and performs weight enhancement processing on the real-time interest features to respond to the user's current behavior needs. The multi-interest extraction layer includes, but is not limited to, a long-term interest extraction module, a short-term interest extraction module, and a real-time interest extraction module. The multi-interest extraction layer is connected to the embedded representation layer.

[0045] The interest fusion layer employs a real-time interest enhancement fusion network based on a gating mechanism. It adjusts the proportion of differential interests according to the weight coefficients and uses a gating recurrent unit to design an interest fusion module. It integrates and processes long-term interest features, short-term interest features, and real-time interest features to obtain sensitive data on the dynamic changes of user interests. The interest fusion layer is connected to the multi-interest extraction layer.

[0046] The present invention has the following advantages over the prior art:

[0047] This invention presents a short video recommendation algorithm based on real-time interest enhancement. It divides user historical interaction sequences into three dimensions: long-term, short-term, and real-time. Different network structures are carefully designed to extract interest features for each dimension. A Gate Recurrent Unit (GRU) effectively integrates users' long-term interest preferences, short-term behavioral changes, and immediate needs, adapting to changes in user interests across different time scales. Furthermore, by strengthening the weight of real-time interests, the algorithm prominently responds to users' current behaviors and needs. Experiments on multiple datasets, compared with baseline algorithms, demonstrate the effectiveness of this algorithm.

[0048] This invention addresses the problem of user recommendation lists needing to be enhanced with real-time user feedback, and designs a short video sequence recommendation method based on real-time interest enhancement. The model mainly consists of four parts: an embedding representation layer, a multi-interest extraction layer, an interest fusion layer, and a prediction layer. The embedding representation layer is used to obtain the embedding representation of the short video sequence features.

[0049] The multi-interest extraction layer of this invention extracts long-term, short-term, and real-time interests from the interaction sequence by designing different network structures. Long-term interests reflect the user's long-term preferences, short-term interests capture the user's recent behavioral changes, and real-time interests focus on the user's current immediate needs.

[0050] The feature fusion layer of this invention fuses features extracted from the interest extraction layer based on a gated recurrent unit, and assigns more weight to real-time interest features through a gating mechanism, thereby making the prediction results more consistent with the user's current interests. In the prediction layer, the features obtained from the previous interest fusion are subjected to multi-layer linear transformations and activations to obtain the final predicted score. Finally, the top K predicted scores are selected as the recommendation list. This invention can effectively obtain the user's β interest, improve recommendation performance, and enhance user experience.

[0051] This invention solves the technical problem in the prior art that it is difficult to make accurate personalized recommendations based on the user's current interests. Attached Figure Description

[0052] Figure 1a This is a schematic diagram of the basic steps of the short video recommendation method based on real-time interest enhancement according to Embodiment 1 of the present invention;

[0053] Figure 1b This is a schematic diagram of the model structure of the short video recommendation system based on real-time interest enhancement in Embodiment 1 of the present invention;

[0054] Figure 2 This is a schematic diagram of the adjacency matrix of the gated graph neural network in Embodiment 1 of the present invention;

[0055] Figure 3This is a schematic diagram of the long-term interest extraction network structure in Embodiment 1 of the present invention;

[0056] Figure 4 This is a schematic diagram of the short-term interest extraction network structure in Embodiment 1 of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1

[0059] like Figure 1a , Figure 1b As shown, the short video recommendation method based on real-time interest enhancement provided by this invention includes the following basic steps:

[0060] Step S1: In the embedding representation layer, the sequence is embedded and represented. The user's historical interaction sequence is divided into three dimensions: long-term, short-term, and real-time. Different network structures are carefully designed to extract the interest features of each dimension.

[0061] In this embodiment, the short video browsing sequence is transformed into a directed graph. Assuming the short videos are nodes in the graph, an interaction sequence can be modeled as a directed graph. For example, in the sequence v = v1, v2, ..., v... n In the graph, v1 and v2 are two nodes. When a user clicks v1 first and then v2, an edge is created from node v1 to node v2. The weight of the edge is set to a standardized weight, which is the number of times each edge appears divided by the out-degree of the node from which the edge originates. Once the short video browsing sequence is converted into a graph structure representation, a gated graph neural network can be used to model the vectors corresponding to each node in the graph. The gated graph neural network can update the information of the nodes through a gating mechanism, retaining valid information and discarding redundant information, and fully learning the sequence features through communication between nodes.

[0062] In this embodiment, the formula for calculating the learned sequence features is as follows:

[0063]

[0064] in, and These are the learnable parameters in the model. A list of short videos representing user interaction, rs，i Indicates the update gate, z s，i The symbol represents the reset gate, ⊙ represents the Hadamard product between matrices, and σ represents the sigmoid activation function. These parameters and variables, through a gating mechanism, filter and transmit information during node updates, which helps in learning the vector representation of the corresponding node.

[0065] like Figure 2 As shown, in this embodiment, modeling operations are performed based on the nodes, weight information, and edges in the short video sequence. In this embodiment, a graph structure representation is obtained based on a user interaction sequence; the output edge weights and input edge weights of the nodes in the graph are calculated from the graph structure representation to obtain the out-degree matrix and in-degree matrix; the hidden representations of the items in the sequence can be learned through a graph neural network, and important information in the interaction sequence is retained through a gating mechanism, finally outputting the embedding representation corresponding to each node.

[0066] Step S2: In the multi-interest extraction layer, long-term interest extraction, short-term interest extraction, and real-time interest extraction operations are performed. The gated recurrent unit (GRU) effectively integrates the user's long-term interest preferences, short-term behavioral changes, and immediate needs, adapting to the user's interest changes at different time scales. By strengthening the weight of real-time interests, it makes them stand out in response to the user's current behavior and needs.

[0067] In this embodiment, the multi-interest extraction layer includes, but is not limited to, a long-term interest extraction module, a short-term interest extraction module, and a real-time interest extraction module.

[0068] like Figure 3 As shown in this embodiment, in the long-term interest extraction module, a user's long-term interest refers to interests that remain stable over a long period, reflecting the user's overall preferences. User behavior sequences are the most revealing source of information about user interests; long-term historical data can be used to model user behavior and uncover long-term interests.

[0069] In this embodiment, there is redundant information in the long-term user interaction sequence. For example, a long-term short video viewing sequence may contain videos recommended by others but which the user is not interested in. At the same time, the user's interest will decline over time. Therefore, the most useful information should be selected dynamically from the long-term sequence.

[0070] In this embodiment, a network structure for long-term interest extraction is designed. This network structure calculates a set of attention weights to focus on different parts of the sequence in order to retain important features in the short video sequence and remove redundant parts.

[0071] In this embodiment, a set of shallow neural networks is used to concatenate the embedding representations of short videos in the sequence with the embedding representations of the target short video; after passing through a shallow network composed of the Dice activation function and linear transformation, attention weights are output; after obtaining the attention weights of each short video, the final long-term user interest representation can be calculated using the attention weights and Sum Pooling.

[0072] In this embodiment, user browsing behavior is often based on previous historical behavior, therefore user interests can only be locally activated. To address the issue of fixed interest vector dimensions, an attention mechanism is introduced to optimize the representation of user features. By appropriately setting weights, the proportion of behavioral features related to the recommended items is strengthened, while the proportion of irrelevant behavioral features is weakened. In this way, different user interest vectors are constructed for different recommended items, fully capturing the diversity of user interests and achieving optimization under a fixed feature vector dimension. The formula for calculating long-term user interests is as follows:

[0073]

[0074] Where, {e1, e2, ..., e H Let} represent the user behavior sequence, H be the length of the behavior sequence, and V be the user behavior sequence. A Let be the embedding representation of the target video to be predicted. a(·) is a small feedforward neural network whose output is the weight of video j in the user's long-term viewing sequence. Finally, the user's long-term interest h is obtained by weighted summation of the embedding representations of each video in the long-term viewing sequence. long .

[0075] In the short-term interest extraction module of this embodiment, a user's short-term interest refers to the trend of interest changes over the past week. Compared with long-term interest, short-term interest reflects a user's recent preferences more effectively. Therefore, fully exploring a user's short-term interest can effectively improve the accuracy of the recommendation algorithm, thereby increasing important metrics such as user dwell time, likes, and favorites on the short video platform. Since most users find a certain continuity between short videos when browsing, the extraction of a user's short-term interest focuses on the interaction features between adjacent items.

[0076] In this embodiment, the convolutional neural network uses a sliding window to capture user features. This method can capture feature interactions between adjacent items and obtain long-range dependencies between items.

[0077] In this embodiment, a combination of horizontal and vertical convolutional filters is used to extract short-term user interest features. This combined approach can identify the overall sequence pattern of user interests, reducing the probability of interest drift and erroneous clicks during interaction. This allows for the capture of short-term user interests while simultaneously extracting higher-level and more abstract sequence features.

[0078] In this embodiment, a horizontal convolutional filter extracts the interaction features between different items. The horizontal convolutional filter slides from top to bottom on the item matrix, extracting the interaction features between sequential items during the sliding process. The output of the horizontal convolutional filter is then max-pooled, and the outputs of all convolutional filters are concatenated to obtain the final output vector o. hor ,in The final output vector is calculated using the following formula:

[0079]

[0080] Among them, S i:i+h-1 This represents rows i to i+h-1 in the sequence matrix. This represents a horizontal convolutional filter, where φ is a nonlinear activation function. This is the output value of the i-th convolution. This represents the set of all feature vectors extracted by the horizontal filter numbered k during the sliding process. hor This represents the final output obtained by concatenating the max-pooling results of filters with different numbering levels.

[0081] like Figure 4 As shown in this embodiment, the process of extracting some short-term interest features from a sequence using a horizontal convolutional filter is illustrated. See also... Figure 4 The user sequence has a length of 4, and the target sequence has a length of 2. First, the user sequence is converted into an embedded representation. Then, a horizontal convolutional filter with a height of 2 is used to extract features between user interaction sequences. The features extracted by all filters are combined to obtain the user's short-term interest features, thereby enabling the prediction of the target item. For items generated due to erroneous clicks or interest drift, feature extraction and max pooling operations using multiple convolutional filters of different heights effectively ignore these few erroneous items in the sequence, thus more accurately obtaining the user's short-term interest features.

[0082] In this embodiment, the vertical convolutional filter extracts the interaction between features of different dimensions of the same item, using... Represents a vertical convolution filter, n vThe vertical convolution filter is the number of filters. Similar to the horizontal convolution filter, the vertical convolution filter slides from right to left along the vector dimension of the item sequence matrix, extracting the interaction features between different dimensions of the same item. Finally, max pooling is used to process the output of the vertical convolution filter and a concatenation operation is performed to obtain the final output result. The calculation process of the final output result of the vertical convolution filter is shown in the following formula:

[0083]

[0084] in This represents the vertical convolution filter numbered k. This represents the set of all features extracted by the vertical convolution filter numbered k during the sliding process. Represents the j-th column of the item sequence matrix, o ver It is the final output obtained by connecting vertical convolutional filters with different numbers after performing a max pooling operation.

[0085] In this embodiment, the feature vector o hor and o ver Let represent the interaction features extracted by the horizontal convolutional filter from different items in the sequence, and and represent the interaction features extracted by the vertical convolutional filter from different dimensions of the same item in the sequence, respectively. Concatenating these two features yields the final user short-term interest h. short .

[0086] In the real-time interest extraction module of this embodiment, the challenge of data scarcity is often faced when processing real-time data. To accurately extract the user's current interest under such circumstances, a simple yet effective network structure needs to be designed. Therefore, a real-time interest extraction network based on an attention mechanism is proposed to accurately extract real-time interests from the user's behavior sequence. The real-time interest extraction network calculates the attention score of the real-time interest relative to each behavior in the behavior sequence; it normalizes these importance scores using the softmax function and then multiplies them by the corresponding behavior items in a weighted manner to obtain the final real-time interest representation.

[0087]

[0088] f(q,i)=(i·tanh(W f ·q+b f ))·W h

[0089] Among them, e i This represents the embedding representation of each item in a short-term user behavior sequence. An embedding representation of a user's real-time interests is given, where f(q, i) is an attention method that defines the weights of videos in a real-time sequence relative to each item in a short-term behavior sequence. f With w h These are all learnable parameter matrices.

[0090] Step S3: Perform interest fusion at the interest fusion layer;

[0091] In this embodiment, a real-time interest enhancement fusion network based on a gating mechanism is used to fuse user interest features at different stages. After extracting the user's long-term interests, short-term interests, and real-time interests, feature fusion is performed. In this embodiment, the feature fusion method can employ, for example, a join operation. However, simply performing a join operation cannot effectively distinguish the weight ratio of different interests in the overall representation. Therefore, an interest fusion module is designed based on a gating recurrent unit. The interest fusion module integrates the representations of long-term interests, short-term interests, and real-time interests in a more refined manner. By introducing weight coefficients through the gating recurrent unit, the proportion of different interests in the final interest representation is effectively adjusted, allowing interests at different stages to complement each other and resulting in a more robust interest representation.

[0092] In this embodiment, the interest fusion module employs a series of weighting coefficients, which quantify the weight of each interest in the overall interest representation. This allows for a more nuanced consideration of users' preferences in long-term, short-term, and real-time interests, and assigns different influences to different interests during the fusion process. Consequently, we can more comprehensively capture users' multi-layered interests, thereby improving the model's sensitivity to dynamic changes in user interests. The formula for calculating the sensitivity to dynamic changes in user interests is shown below:

[0093]

[0094] u=α·[u short ::u long ]+(1-α)u real

[0095] Here, FFN represents a feedforward neural network, and GRU represents a gated recurrent unit. An embedding representation of a user's long-term behavior sequence. An embedded representation of a short-term sequence of user behaviors.

[0096] In summary, this invention presents a short video recommendation algorithm based on real-time interest enhancement. It divides user historical interaction sequences into three dimensions: long-term, short-term, and real-time. Different network structures are carefully designed to extract interest features for each dimension. A Gate Recurrent Unit (GRU) effectively integrates users' long-term interest preferences, short-term behavioral changes, and immediate needs, adapting to changes in user interests across different time scales. Furthermore, by strengthening the weight of real-time interests, it emphasizes responses to users' current behaviors and needs. Experiments on multiple datasets, compared with baseline algorithms, demonstrate the effectiveness of this algorithm.

[0097] This invention addresses the problem of user recommendation lists needing to be enhanced with real-time user feedback, and designs a short video sequence recommendation method based on real-time interest enhancement. The model mainly consists of four parts: an embedding representation layer, a multi-interest extraction layer, an interest fusion layer, and a prediction layer. The embedding representation layer is used to obtain the embedding representation of the short video sequence features.

[0098] The multi-interest extraction layer of this invention extracts long-term, short-term, and real-time interests from the interaction sequence by designing different network structures. Long-term interests reflect the user's long-term preferences, short-term interests capture the user's recent behavioral changes, and real-time interests focus on the user's current immediate needs.

[0099] The feature fusion layer of this invention fuses features extracted from the interest extraction layer based on a gated recurrent unit, and assigns more weight to real-time interest features through a gating mechanism, thereby making the prediction results more consistent with the user's current interests. In the prediction layer, the features obtained from the previous interest fusion are subjected to multi-layer linear transformations and activations to obtain the final predicted score. Finally, the top K predicted scores are selected as the recommendation list. This invention can effectively obtain the user's β interest, improve recommendation performance, and enhance user experience.

[0100] This invention solves the technical problem in the prior art that it is difficult to make accurate personalized recommendations based on the user's current interests.

[0101] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A short video recommendation method based on real-time interest enhancement, characterized in that, The method includes: S1. In the embedding representation layer, the short video browsing sequence is embedded and represented to obtain the user's long-term historical interaction sequence, short-term historical interaction sequence, and real-time historical interaction sequence, for extracting long-term interest features, short-term interest features, and real-time interest features; wherein, the short video browsing sequence is transformed into a directed graph, and the short videos in the short video browsing sequence are used as graph nodes. Based on the user's video click information, the edges between the graph nodes are determined, the standardized weights of the edges are set, and a gated graph neural network is used to model and obtain the corresponding vectors of the nodes in the directed graph to update the node information. Sequence features are obtained by learning through mutual communication between nodes. In S1, the sequence features are learned using the following logic: In the formula, and These are the learnable parameters in the model. A list of short videos representing user interaction. Indicates an update to the door. This indicates that the door is being reset. Represents the Hadamard product between matrices. This represents the sigmoid activation function; S2. In the multi-interest extraction layer, long-term interest extraction, short-term interest extraction, and real-time interest extraction operations are performed based on the sequence features. The user's long-term interest features, short-term interest features, and real-time interest features are integrated through a gated recurrent unit (GRU). The real-time interest features are weighted and strengthened to perform corresponding operations based on the user's current behavior needs. The multi-interest extraction layer includes a long-term interest extraction module, a short-term interest extraction module, and a real-time interest extraction module. In the long-term interest extraction module, user behavior is modeled using long-term historical data, attention weights are calculated, and weighted focusing operations are performed on different parts of the short video browsing sequence to preserve the long-term interest features. In the real-time interest extraction module, an attention-based real-time interest extraction network is used to calculate the attention score of real-time interest relative to the sequence behavior. The importance information in the attention score is normalized and weighted multiplied using a softmax function to obtain the real-time interest features. The normalization and weighted multiplication operations are performed using the following logic: In the formula, This represents the embedding representation of each item in a short-term user behavior sequence. An embedded representation of a user's real-time interests. It is an attention method that defines the weights of videos in a real-time sequence relative to each item in a short-term action sequence. and All of them are learnable parameter matrices; In the weighted focusing operation, a set of shallow neural networks is used to concatenate the short video embedding representations in the short video browsing sequence with the target short video embedding representation to obtain attention weights. The shallow neural network includes a Dice activation function and a linear transformation component. The attention weights are processed, and Sum Pooling is performed to obtain the long-term interest features. The long-term interest features are calculated using the following logic: In the formula, Let H be the user behavior sequence, and H be the length of the behavior sequence. The embedding representation of the target video to be predicted. A small feedforward neural network is used, and the output of the small feedforward neural network is the weight of video j in the user's long-term viewing sequence; the long-term interest feature is obtained by weighted summation of the embedding representation of each video in the long-term viewing sequence. ; S3. In the interest fusion layer, a real-time interest enhancement fusion network based on a gating mechanism is adopted. The proportion of differential interests is adjusted according to the weight coefficient. An interest fusion module is designed using a gating loop unit to fuse and process the long-term interest features, the short-term interest features, and the real-time interest features to obtain the sensitivity data of dynamic changes in user interests. In S3, the following logic is used to obtain the sensitivity data of the dynamic changes in interest: In the formula, FFN represents a feedforward neural network, and GRU represents a gated recurrent unit. An embedding representation of a user's long-term behavior sequence. An embedded representation of a short-term sequence of user behaviors.

2. The short video recommendation method based on real-time interest enhancement according to claim 1, characterized in that, The inter-node communication learning in S1 includes: S11. Perform modeling operations based on the nodes, the standardized weights, and the edges; S12. Based on the short video browsing sequence, a graph structure representation is obtained; S13. Based on the graph structure representation, calculate the output edge weight and input edge weight of the node to obtain the out-degree matrix and in-degree matrix; S14. Using a gated graph neural network, learn the hidden representation of items in the short video browsing sequence, retain the interaction sequence information through a gating mechanism, and obtain and output the embedding representation corresponding to each node as the corresponding vector of the node.

3. The short video recommendation method based on real-time interest enhancement according to claim 1, characterized in that, In the short-term interest extraction module of S2, interaction features between adjacent items are captured. A pre-set convolutional neural network is used, and a sliding window is employed to capture these interaction features, obtaining long-range dependencies between items. Combined with horizontal and vertical convolutional filters, the overall sequence pattern of user interests is identified. The horizontal and vertical convolutional filters are then processed to obtain the interaction features between adjacent items, thus yielding the user's short-term interest features. .

4. The short video recommendation method based on real-time interest enhancement according to claim 3, characterized in that, Using the horizontal convolutional filter, a sliding extraction operation is performed on the item matrix. This sliding extraction operation is performed using the following logic to obtain the interaction features between adjacent items: In the formula, This represents rows i to i+h-1 in the sequence matrix. This represents a horizontal convolution filter. It is a non-linear activation function. Let i be the output value of the i-th convolution. Let k represent the set of all feature vectors extracted by the horizontal filter numbered k during the sliding process. This represents the final output obtained by concatenating the max-pooling results of filters with different numbering levels.

5. The short video recommendation method based on real-time interest enhancement according to claim 4, characterized in that, Using the vertical convolutional filter, a sliding feature extraction operation is performed along the vector dimension of the item sequence matrix to extract the interaction features between different dimensions of the same item. Max pooling is then used to process and concatenate the output of the vertical convolutional filter to obtain the interaction features between adjacent items. The sliding extraction operation is performed using the vertical convolutional filter with the following logic: , ..., In the formula, This represents the vertical convolution filter numbered k. This represents the set of all features extracted by the vertical convolution filter numbered k during the sliding process. This represents the j-th column of the item sequence matrix. The final output is obtained by concatenating vertical convolutional filters with different numbers after performing a max pooling operation. This represents a vertical convolution filter. This represents the number of vertical convolution filters.

6. A short video recommendation system based on real-time interest enhancement, used to execute the short video recommendation system method based on real-time interest enhancement as described in any one of claims 1 to 5, characterized in that, The system includes: An embedding representation layer is used to embed the short video browsing sequence into a representation, dividing it into long-term historical interaction sequences, short-term historical interaction sequences, and real-time historical interaction sequences for extracting long-term interest features, short-term interest features, and real-time interest features. Specifically, the short video browsing sequence is transformed into a directed graph, with the short videos in the sequence as graph nodes. Based on user video click information, edges between the graph nodes are determined, and standardized weights are set for these edges. A gated graph neural network is used to model the corresponding vectors of the nodes in the directed graph to update node information. Sequence features are obtained through inter-node communication and learning. A multi-interest extraction layer is used to perform long-term interest extraction, short-term interest extraction, and real-time interest extraction operations based on the sequence features. Through a gated recurrent unit (GRU), it integrates the user's long-term interest features, short-term interest features, and real-time interest features, and performs weight enhancement processing on the real-time interest features to respond to the user's current behavior needs. The multi-interest extraction layer includes a long-term interest extraction module, a short-term interest extraction module, and a real-time interest extraction module, and the multi-interest extraction layer is connected to the embedded representation layer. The interest fusion layer employs a real-time interest enhancement fusion network based on a gating mechanism. It adjusts the proportion of differential interests according to the weight coefficients, and uses a gating recurrent unit to design an interest fusion module. It fuses and processes the long-term interest features, the short-term interest features, and the real-time interest features to obtain sensitive data on the dynamic changes of user interests. The interest fusion layer is connected to the multi-interest extraction layer.

Images