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

Abstract

The invention provides a short video recommendation method and system based on real-time interest enhancement, and the method comprises the steps: dividing a user historical interaction sequence into a long-term dimension, a short-term dimension and a real-time dimension, and carefully designing different network structures to extract the interest features of each dimension; long-term interest preferences, short-term behavior changes and instant demands of users are effectively integrated through a Gate Recurrent Unit (GRU), interest changes of the users on different time scales are adapted, and the weights of real-time interests are strengthened, so that the real-time interests prominently respond to the current behaviors and demands of the users; experiments are carried out on a plurality of data sets, and compared with a baseline algorithm, the algorithm effectiveness is proved. The technical problem that accurate personalized recommendation is difficult to perform in combination with the current interest of the user is solved.

Description

Technical Field

[0001] The present invention relates to the field of recommendation algorithms, and in particular to a short video recommendation method and system based on real-time interest enhancement. Background Art

[0002] With the continuous development of mobile Internet, short video platforms have become an important entertainment and social tool for the new generation of Internet users. Hundreds of millions of users consume a large amount of short video content every day around the world, and short video platforms have a very high penetration rate among Internet users. Compared with long videos, short videos have the characteristic of short playback time, and users will browse more short videos in one interaction process. Therefore, using the user's historical interaction information for serialization modeling is an effective way to improve user experience.

[0003] For example, the existing invention patent application document "A short video intelligent recommendation system based on data analysis" with publication number CN117150075A, the existing system includes: a management platform, a data acquisition unit, a video division unit, a heat analysis unit, a built-in traffic analysis unit, a display execution unit, a release supervision analysis unit and a supervisory analysis unit; the present invention performs video division supervision analysis on the historical information of the recommended user, so as to analyze the type of short video that the recommended user likes according to the historical information feedback, so as to reasonably recommend interest-type videos, and by collecting the interest-type's own data and circulation data, that is, from two perspectives, evaluation and analysis are performed, namely, the heat brought by the publisher himself and the heat caused by the video after release. And the existing invention patent application document "A short video data recommendation method and system based on tag analysis" with publication number CN114896454A, the existing method includes: by loading each first user behavior activity data into a set knowledge entity network for regular conversion, and generating corresponding second user behavior activity data, the reliability of subsequent mining processing is improved, and the corresponding second user behavior activity data is loaded into the corresponding short video recommendation service from different short video recommendation dimensions, so as to extract all the second user behavior activity data associated with each short video recommendation service, load them into the corresponding interest thermal trend output model to generate the corresponding user interest tag thermal trend, further determine the key user interest tag corresponding to the target user, and recommend short video data to the target user based on this, and then perform personalized recommendation after mining user interests in different short video recommendation dimensions. The existing recommendation algorithm has shortcomings in adjusting and optimizing the recommendation list according to the user's real-time feedback. How to design a recommendation algorithm suitable for short video platforms that takes into account the user's real-time interests and ensures accuracy to recommend the content that the user is currently interested in still requires more in-depth research. In addition, since the short videos browsed by users constitute an interactive sequence, and short videos have the characteristics of short time, this sequence is compact in time and intensive in user feedback. It is somewhat complex to make recommendations based on intensive user feedback information in a short period of time. Finally, while considering users' short-term and real-time interests, retaining some long-term interests is also a direction worthy of further research.

[0004] In summary, the existing technology has the technical problem of being difficult to make accurate personalized recommendations based on the user's current interests. Summary of the invention

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

[0006] The present invention adopts the following technical solutions to solve the above technical problems: 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, and the user's long-term historical interaction sequence, short-term historical interaction sequence and real-time historical interaction sequence are divided to extract long-term interest features, short-term interest features and real-time interest features; wherein, the short video browsing sequence is converted into a directed graph, and the short videos in the short video browsing sequence are used as graph nodes. According to the user's video click information, the edges between the graph nodes are determined, and the standardized weights of the edges are set. The gated graph neural network is used to model the corresponding vectors of the nodes in the directed graph to update the node information, and the sequence features are obtained by learning through mutual communication between the nodes;

[0008] S2. In the multi-interest extraction layer, long-term interest extraction operations, short-term interest extraction operations and real-time interest extraction operations are performed according to sequence features. The user's long-term interest features, short-term interest features and real-time interest features are integrated through the gated recurrent unit GRU, and the real-time interest features are weighted and enhanced to perform operations corresponding 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;

[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. The interest fusion module is designed using a gated recurrent unit. The long-term interest features, short-term interest features and real-time interest features are integrated and processed to obtain the sensitivity data of dynamic changes in user interests.

[0010] The present invention is based on real-time interest enhancement to perform short video recommendation algorithm, which divides the user's historical interaction sequence into three dimensions: long-term, short-term, and real-time. Different network structures are carefully designed to extract the interest features of each dimension, and the gated recurrent unit (GRU) effectively integrates the user's long-term interest preferences, short-term behavior changes, and immediate needs, adapting to the user's interest changes on different time scales, and by strengthening the weight of real-time interests, it makes it prominently respond to the user's current behavior and needs. The effectiveness of the algorithm has been proved by conducting experiments on multiple data sets and comparing with the baseline 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 are learnable parameters in the model, A list of short videos representing user interactions, r s，i represents the update gate, z s，i represents the reset gate, ⊙ represents the Hadamard product between matrices, and σ represents the sigmoid activation function.

[0015] In a more specific technical solution, the nodes of S1 communicate and learn with each other, including:

[0016] S11, performing modeling operations according to nodes, standardized weights, and edges;

[0017] S12, processing the short video browsing sequence to obtain a graph structure representation;

[0018] S13. According to the graph structure representation, the output edge weight and the input edge weight of the node are calculated to obtain the out-degree matrix and the in-degree matrix;

[0019] S14. Use the gated graph neural network to learn the hidden representation of items in the short video browsing sequence, and retain the interaction sequence information through the gating mechanism to obtain and output the embedding representation corresponding to each node as the node corresponding vector.

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

[0021] In a more specific technical solution, in the real-time interest extraction module, a real-time interest extraction network based on an attention mechanism is used to calculate the attention score of the 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 feature, wherein 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 Represents the embedding representation of each item in the user's short-term behavior sequence, The embedding representation of the user’s real-time interests, f(q,i) is an attention method that defines the weight of each item in the video in the real-time sequence relative to the short-term behavior sequence. f With wh are all learnable parameter matrices.

[0025] Aiming at the problem that the user recommendation list needs to enhance interest according to the user's real-time feedback, the present invention designs a short video sequence recommendation method based on real-time interest enhancement. The model is mainly divided into four parts: embedding representation layer, multi-interest extraction layer, interest fusion layer, and 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, in the weighted focus operation, a group of shallow neural networks are used to splice the short video embedding representation in the short video browsing sequence and the target short video embedding representation to process and obtain the attention weight; wherein the shallow neural network includes: Dice activation function, linear transformation component; processing the attention weight, and performing Sum Pooling calculation to obtain the long-term interest feature, wherein the long-term interest feature is calculated using the following logic:

[0027]

[0028] In the formula, {e1, e2, ..., e H} is the user behavior sequence, H is the length of the behavior sequence, V A is 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 embedding representation of each video in the long-term viewing video sequence is weighted to obtain the long-term interest feature h long .

[0029] In a more specific technical solution, in the short-term interest extraction module in S2, the interaction features between adjacent items are captured, and the interaction features between adjacent items are captured using a preset convolutional neural network and a sliding window to obtain the long-range dependency between items; the overall sequence pattern of user interests is identified by combining horizontal convolution filters and vertical convolution filters, and the horizontal convolution filters and vertical convolution filters are connected and processed to obtain the interaction features between adjacent items, and the user's short-term interest feature h is obtained. short .

[0030] The multi-interest extraction layer of the present invention extracts long-term, short-term and real-time interests in the interaction sequence by designing different network structures, wherein the long-term interest reflects the user's long-term preference, the short-term interest captures the user's recent behavior changes, and the real-time interest focuses on the user's current immediate needs.

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

[0032]

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

[0034] In a more specific technical solution, a vertical convolution 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; the output of the vertical convolution filter is processed using a maximum pooling operation and connected to obtain the interaction features between adjacent items, wherein the vertical convolution filter is used to perform a sliding extraction operation according to the following logic:

[0035]

[0036] In the formula, represents the vertical convolution filter numbered k, represents the set of all features extracted by the vertical convolution filter numbered k during the sliding process, represents the jth column of the item sequence matrix, o ver It is the final output obtained by connecting vertical convolution filters with different numbers after the maximum pooling operation. represents the vertical convolution filter, n v is the number of vertical convolution filters.

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

[0038]

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

[0040] In the formula, FFN represents feedforward neural network, GRU represents gated recurrent unit, Embedding representation that represents the user's long-term behavior sequence, Embedding representations that represent user short-term behavior sequences.

[0041] The feature fusion layer of the present invention fuses the features extracted by the interest extraction layer based on the gated recurrent unit, and gives more weight to the real-time interest features through the gating mechanism, so that the prediction results are more in line with the user's current interests. In the prediction layer, the features obtained by the previous interest fusion are transformed and activated through multiple layers of linear transformation to obtain the final prediction score. Finally, the top K predicted scores are taken as the recommendation list. The present invention can effectively obtain the user's β interests, improve the recommendation performance and improve the user experience.

[0042] In a more specific technical solution, the short video recommendation system based on real-time interest enhancement includes:

[0043] The embedding representation layer is used to embed the short video browsing sequence and divide it into 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 converted into a directed graph, and the short videos in the short video browsing sequence are used as graph nodes. According to the user's video click information, the edges between the graph nodes are determined, and the standardized weights of the edges are set. The gated graph neural network is used to model the corresponding vectors of the nodes in the directed graph to update the node information, and the sequence features are obtained by learning through mutual communication between the nodes;

[0044] The multi-interest extraction layer is used to perform long-term interest extraction operations, short-term interest extraction operations and real-time interest extraction operations according to sequence features, integrate the user's long-term interest features, short-term interest features and real-time interest features through the gated recurrent unit GRU, and perform weighted reinforcement processing on the real-time interest features to perform operations corresponding to the user's current behavior needs, wherein 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, and the multi-interest extraction layer is connected to the embedding representation layer;

[0045] The interest fusion layer adopts a real-time interest enhancement fusion network based on a gating mechanism, adjusts the proportion of different interests according to the weight coefficient, designs an interest fusion module using a gated recurrent unit, integrates long-term interest features, short-term interest features and real-time interest features, and obtains the user's interest dynamic change sensitivity data. The interest fusion layer is connected to the multi-interest extraction layer.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] The present invention is based on real-time interest enhancement to perform short video recommendation algorithm, which divides the user's historical interaction sequence into three dimensions: long-term, short-term, and real-time. Different network structures are carefully designed to extract the interest features of each dimension, and the gated recurrent unit (GRU) effectively integrates the user's long-term interest preferences, short-term behavior changes, and immediate needs, adapting to the user's interest changes on different time scales, and by strengthening the weight of real-time interests, it makes it prominently respond to the user's current behavior and needs. The effectiveness of the algorithm has been proved by conducting experiments on multiple data sets and comparing with the baseline algorithm.

[0048] Aiming at the problem that the user recommendation list needs to enhance interest according to the user's real-time feedback, the present invention designs a short video sequence recommendation method based on real-time interest enhancement. The model is mainly divided into four parts: embedding representation layer, multi-interest extraction layer, interest fusion layer, and 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 the present invention extracts long-term, short-term and real-time interests in the interaction sequence by designing different network structures, wherein the long-term interest reflects the user's long-term preference, the short-term interest captures the user's recent behavior changes, and the real-time interest focuses on the user's current immediate needs.

[0050] The feature fusion layer of the present invention fuses the features extracted by the interest extraction layer based on the gated recurrent unit, and gives more weight to the real-time interest features through the gating mechanism, so that the prediction results are more in line with the user's current interests. In the prediction layer, the features obtained by the previous interest fusion are transformed and activated through multiple layers of linear transformation to obtain the final prediction score. Finally, the top K predicted scores are taken as the recommendation list. The present invention can effectively obtain the user's β interests, improve the recommendation performance and improve the user experience.

[0051] The present invention solves the technical problem in the prior art that it is difficult to make accurate personalized recommendations based on the current interests of users. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

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

[0056] Figure 4 This is a schematic diagram of the short-term interest extraction network structure of Example 1 of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work 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 the present invention includes the following basic steps:

[0060] Step S1: In the embedding representation layer, the sequence is embedded and represented, and 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 interesting features of each dimension.

[0061] In this embodiment, the short video browsing sequence is converted into a directed graph. Assuming that 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 the user first clicks v1 and then clicks v2, there will be an edge from node v1 to node v2 in the graph. The weight of the edge is set to the standardized weight, that is, the number of times each edge appears divided by the out-degree of the starting node of the edge. After the short video browsing sequence is converted into a graph structure representation, the 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 node information with the help of the gating mechanism, so that it retains valid information and forgets redundant information, and fully learns the sequence features through mutual communication between nodes.

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

[0063]

[0064] in, and are learnable parameters in the model, A list of short videos representing user interactions, rs，i represents the update gate, z s，i represents the reset gate, ⊙ represents the Hadamard product between matrices, and σ represents the sigmoid activation function. These parameters and variables implement the screening and transmission of information through the gating mechanism during the node update process, which helps to learn the representation of the node corresponding vector.

[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 by the graph structure representation to obtain the out-degree matrix and the in-degree matrix; the hidden representation of the items in the sequence can be learned through the graph neural network, and the important information in the interaction sequence is retained through the gating mechanism, and finally the embedded representation corresponding to each node is output.

[0066] Step S2: In the multi-interest extraction layer, long-term interest extraction operations, short-term interest extraction operations and real-time interest extraction operations are performed. The user's long-term interest preferences, short-term behavior changes and immediate needs are effectively integrated through the gated recurrent unit (GRU), adapting to the user's interest changes on different time scales, and by strengthening the weight of real-time interests, it prominently responds to the user's current behavior and needs;

[0067] In this embodiment, the multiple interest extraction layers include but are 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, the user's long-term interest is the user's interest that remains stable over a long period of time, reflecting the user's overall preference. The user's behavior sequence is the most informative source of user interest, and long-term historical data can be used to model user behavior and mine the user's long-term interest.

[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 not of interest to the user. At the same time, user interest will continue to decline over time. Therefore, the long-term sequence should be used dynamically to select the most useful information.

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

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

[0072] In this embodiment, the user's browsing behavior is often based on previous historical behavior, so the user's interest can only be activated locally. In this embodiment, in order to solve the problem of fixed interest vector dimensions, an attention mechanism is introduced to optimize the user feature expression. By setting weights reasonably, the proportion of behavioral features related to the recommended products is strengthened, and the proportion of irrelevant behavioral features is weakened. In this way, different user interest vectors are constructed for different recommended items, which fully captures the diversity of user interests and achieves optimization under fixed feature vector dimensions. The calculation formula for the user's long-term interest is as follows:

[0073]

[0074] Among them, {e1, e2, ..., e H} is the user behavior sequence, H is the length of the behavior sequence, V A is 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 summing the embedding representation of each video in the long-term viewing video sequence. long .

[0075] In the short-term interest extraction module of this embodiment, the user's short-term interest refers to the interest change trend in the past week. Short-term interests can better reflect the user's recent preferences compared to long-term interests. Therefore, fully exploring the user's short-term interests can effectively improve the accuracy of the recommendation algorithm, thereby improving important indicators such as user stay time, likes, and collections on the short video platform. Since most users have a certain degree of coherence between short videos when browsing, when extracting users' short-term interests, the interaction characteristics between adjacent items are focused on to obtain users' short-term interests.

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

[0077] In this embodiment, a method combining horizontal convolution filters and vertical convolution filters is used to extract user short-term interest features. The aforementioned combined method can identify the overall sequence pattern of user interests, reduce the probability of interest drift and wrong clicks generated during the interaction process, so as to capture the user's short-term interests while extracting higher-level and more abstract sequence features.

[0078] In this embodiment, the horizontal convolution filter extracts the interaction features between different items. The horizontal convolution filter can slide from top to bottom on the item matrix to extract the interaction features between sequence items during the sliding process; the output of the horizontal convolution filter is maximum pooled, and the output results of all convolution filters are connected to obtain the final output vector o hor ,in The final output vector is calculated as follows:

[0079]

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

[0081] like Figure 4 As shown in FIG. 1 , in this embodiment, the process of extracting some short-term interest features in a sequence by a horizontal convolution filter is shown. Figure 4 , the length of the user sequence is 4, the length of the target sequence is 2, first the user sequence is converted into an embedded representation, then a horizontal convolution filter with a height of 2 is used to extract the features between the user interaction sequences, and the features extracted by all filters are combined together to obtain the user's short-term interest features, thereby achieving the prediction of the target items. For some items generated by wrong clicks and interest drift, after feature extraction and maximum pooling operations using multiple convolution filters of different heights, these small number of misoperated items in the sequence are well ignored, thereby more accurately obtaining the user's short-term interest features.

[0082] In this embodiment, the vertical convolution filter extracts the interaction between features of different dimensions of the same item, using represents the vertical convolution filter, n vis the number of vertical convolution filters. Like the horizontal convolution filter, the vertical convolution filter slides from left to right along the vector dimension of the item sequence matrix to extract the interaction features between different dimensions of the same item. Finally, the output of the vertical convolution filter is processed and concatenated using the maximum pooling operation 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 represents the vertical convolution filter numbered k, represents the set of all features extracted by the vertical convolution filter numbered k during the sliding process, represents the jth column of the item sequence matrix, o ver It is the final output obtained by connecting vertical convolution filters with different numbers after the maximum pooling operation.

[0085] In this embodiment, the feature vector o hor and ver They represent the interaction features extracted by the horizontal convolution filter from different items in the sequence and the interaction features extracted by the vertical convolution filter from different dimensions of the same item in the sequence. Connecting the two can get the final user short-term interest h short .

[0086] In the real-time interest extraction module of this embodiment, when processing real-time data, it is usually faced with the challenge of data scarcity. In order to accurately extract the user's current interest in this case, it is necessary to design a simple and effective network structure. Therefore, a real-time interest extraction network based on the 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; the softmax function is used to normalize these importances, and they are weighted and multiplied with the corresponding behavior items 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 Represents the embedding representation of each item in the user's short-term behavior sequence, The embedding representation of the user’s real-time interests, f(q,i) is an attention method that defines the weight of each item in the video in the real-time sequence relative to the short-term behavior sequence. f With w h are all learnable parameter matrices.

[0090] Step S3, performing 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 adopt, for example, a connection operation. However, simply performing a connection operation cannot effectively distinguish the weight proportion of different interests in the overall representation. Therefore, an interest fusion module is designed based on a gated recurrent unit. The interest fusion module integrates the representation of long-term interests, short-term interests and real-time interests in a more refined way. The weight coefficient is introduced through the gated recurrent unit to effectively adjust the proportion of different interests in the final interest representation, so that the interests at different stages complement each other and obtain a more robust interest representation.

[0092] In this embodiment, the interest fusion module uses a series of weight coefficients, which are used to quantify the weight of each weighted interest in the comprehensive interest representation. In this way, we can consider the user's inclination in long-term, short-term, and real-time interests in a more detailed manner, and give different interests different influences during the fusion process. In this way, we can capture the user's multi-level interests more comprehensively, thereby improving the model's sensitivity to dynamic changes in user interests. The calculation formula for the sensitivity to dynamic changes in user interests is as follows:

[0093]

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

[0095] Among them, FFN represents feedforward neural network, GRU represents gated recurrent unit, Embedding representation that represents the user's long-term behavior sequence, Embedding representations that represent user short-term behavior sequences.

[0096] In summary, the present invention is based on real-time interest enhancement to perform short video recommendation algorithm, divides the user's historical interaction sequence into three dimensions: long-term, short-term, and real-time, carefully designs different network structures to extract interest features of each dimension, and effectively integrates the user's long-term interest preferences, short-term behavior changes, and immediate needs through the gated recurrent unit (GRU), adapts to the user's interest changes on different time scales, and strengthens the weight of real-time interests to make it prominently respond to the user's current behavior and needs. The effectiveness of the algorithm was proved by conducting experiments on multiple data sets and comparing with the baseline algorithm.

[0097] Aiming at the problem that the user recommendation list needs to enhance interest according to the user's real-time feedback, the present invention designs a short video sequence recommendation method based on real-time interest enhancement. The model is mainly divided into four parts: embedding representation layer, multi-interest extraction layer, interest fusion layer, and 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 the present invention extracts long-term, short-term and real-time interests in the interaction sequence by designing different network structures, wherein the long-term interest reflects the user's long-term preference, the short-term interest captures the user's recent behavior changes, and the real-time interest focuses on the user's current immediate needs.

[0099] The feature fusion layer of the present invention fuses the features extracted by the interest extraction layer based on the gated recurrent unit, and gives more weight to the real-time interest features through the gating mechanism, so that the prediction results are more in line with the user's current interests. In the prediction layer, the features obtained by the previous interest fusion are transformed and activated through multiple layers of linear transformation to obtain the final prediction score. Finally, the top K predicted scores are taken as the recommendation list. The present invention can effectively obtain the user's β interests, improve the recommendation performance and improve the user experience.

[0100] The present invention solves the technical problem in the prior art that it is difficult to make accurate personalized recommendations based on the current interests of users.

[0101] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions 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 comprises: S1. In the embedding representation layer, the short video browsing sequence is embedded and represented, and the user's long-term historical interaction sequence, short-term historical interaction sequence and real-time historical interaction sequence are divided to extract long-term interest features, short-term interest features and real-time interest features; wherein, the short video browsing sequence is converted into a directed graph, and the short videos in the short video browsing sequence are used as graph nodes. According to the user's video click information, the edges between the graph nodes are determined, and the standardized weights of the edges are set. The gated graph neural network is used to model the corresponding vectors of the nodes in the directed graph to update the node information, and the sequence features are obtained by learning through mutual communication between the nodes; S2. In the multi-interest extraction layer, long-term interest extraction operations, short-term interest extraction operations and real-time interest extraction operations are performed according to the sequence features, and the user's long-term interest features, short-term interest features and real-time interest features are integrated through the gated recurrent unit GRU, and the real-time interest features are weighted and strengthened to perform operations corresponding to the user's current behavior needs, wherein 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; 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, and an interest fusion module is designed using a gated recurrent 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.

2. The short video recommendation method based on real-time interest enhancement according to claim 1, characterized in that: In S1, the following logic is used to learn and obtain the sequence features: In the formula, and are learnable parameters in the model, A list of short videos representing user interactions, r s，i represents the update gate, z s，i represents the reset gate, ⊙ represents the Hadamard product between matrices, and σ represents the sigmoid activation function.

3. The short video recommendation method based on real-time interest enhancement according to claim 1, characterized in that: The nodes in S1 communicate and learn with each other, including: S11, performing a modeling operation according to the nodes, the standardized weights, and the edges; S12, processing the short video browsing sequence to obtain a graph structure representation; S13, according to the graph structure representation, calculating the output edge weight and the input edge weight of the node to obtain the out-degree matrix and the in-degree matrix; S14. Using a gated graph neural network, learn the hidden representations of items in the short video browsing sequence, retain the interaction sequence information through a gating mechanism, so as to obtain and output the embedding representation corresponding to each of the nodes as the node corresponding vector.

4. The short video recommendation method based on real-time interest enhancement according to claim 1, characterized in that: In the long-term interest extraction module in S2, long-term historical data is used to model user behavior, calculate attention weights, and perform weighted focusing operations on different parts of the short video browsing sequence to retain the long-term interest features. In the real-time interest extraction module, a real-time interest extraction network based on an attention mechanism is used to calculate the attention score of the 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 feature, wherein the normalization and weighted multiplication operations are performed using the following logic: f(q,i)=(i·tanh(W f ·q+b f ))·W h In the formula, e i Represents the embedding representation of each item in the user's short-term behavior sequence, The embedding representation of the user’s real-time interests, f(q,i) is an attention method that defines the weight of each item in the video in the real-time sequence relative to the short-term behavior sequence. f With w h are all learnable parameter matrices.

5. The short video recommendation method based on real-time interest enhancement according to claim 4, characterized in that: In the weighted focusing operation, a group of shallow neural networks are used to splice the short video embedding representation in the short video browsing sequence and the target short video embedding representation to obtain the attention weight; wherein the shallow neural network includes: a Dice activation function and a linear transformation component; the attention weight is processed and a Sum Pooling calculation is performed to obtain the long-term interest feature, wherein the long-term interest feature is calculated using the following logic: In the formula, {e1, e2, ..., e H } is the user behavior sequence, H is the length of the behavior sequence, V A is 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 embedding representation of each video in the long-term viewing video sequence is weighted to obtain the long-term interest feature h long .

6. The short video recommendation method based on real-time interest enhancement according to claim 4, characterized in that: In the short-term interest extraction module in S2, the interaction features between adjacent items are captured, and the interaction features between adjacent items are captured using a preset convolutional neural network and a sliding window to obtain the long-range dependency between items; the overall sequence pattern of user interests is identified by combining the horizontal convolution filter and the vertical convolution filter, and the horizontal convolution filter and the vertical convolution filter are connected and processed to obtain the interaction features between adjacent items, and the user short-term interest feature h is obtained. short .

7. The short video recommendation method based on real-time interest enhancement according to claim 6, characterized in that: The horizontal convolution filter is used to perform a sliding extraction operation on the item matrix, and the following logic is used to perform the East China extraction operation to obtain the interaction features between adjacent items: In the formula, S i:i+h-1 represents the i-th row to the i+h-1th row in the sequence matrix, represents the horizontal convolution filter, φ is the nonlinear activation function, is the output value of the i-th convolution, represents the set of all feature vectors extracted by the horizontal filter numbered k during the sliding process, o hor Represents the final output obtained by connecting the maximum pooling results of filters with different number levels.

8. The short video recommendation method based on real-time interest enhancement according to claim 7, characterized in that: Using the vertical convolution 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; using the maximum pooling operation to process the output of the vertical convolution filter and perform a connection operation to obtain the interaction features between adjacent items, wherein the sliding extraction operation is performed using the vertical convolution filter with the following logic: In the formula, represents the vertical convolution filter numbered k, represents the set of all features extracted by the vertical convolution filter numbered k during the sliding process, represents the jth column of the item sequence matrix, o ver It is the final output obtained by connecting vertical convolution filters with different numbers after the maximum pooling operation. represents the vertical convolution filter, n v is the number of vertical convolution filters.

9. The short video recommendation method based on real-time interest enhancement according to claim 1, characterized in that: In S3, the following logic is used to obtain the interest dynamic change sensitivity data: u=α·[u short ::in long ]+(1-α)u real In the formula, FFN represents feedforward neural network, GRU represents gated recurrent unit, Embedding representation that represents the user's long-term behavior sequence, Embedding representations that represent user short-term behavior sequences.

10. A short video recommendation system based on real-time interest enhancement, characterized in that: The system comprises: The embedding representation layer is used to embed the short video browsing sequence, divide it into the user's long-term historical interaction sequence, short-term historical interaction sequence and real-time historical interaction sequence, so as to extract long-term interest features, short-term interest features and real-time interest features; wherein, the short video browsing sequence is converted into a directed graph, the short videos in the short video browsing sequence are used as graph nodes, the edges between the graph nodes are determined according to the user's video click information, the standardized weights of the edges are set, and the gated graph neural network is used to model the corresponding vectors of the nodes in the directed graph to update the node information, and the sequence features are obtained by learning through mutual communication between the nodes; A multi-interest extraction layer, used to perform long-term interest extraction operations, short-term interest extraction operations and real-time interest extraction operations according to the sequence features, integrate the user's long-term interest features, short-term interest features and real-time interest features through the gated recurrent unit GRU, and perform weight enhancement processing on the real-time interest features to perform operations corresponding to the user's current behavior needs, wherein 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, and the multi-interest extraction layer is connected to the embedding representation layer; The interest fusion layer is used to adopt a real-time interest enhancement fusion network based on a gating mechanism, adjust the proportion of different interests according to the weight coefficient, design an interest fusion module using a gated recurrent unit, fuse and process the long-term interest features, the short-term interest features and the real-time interest features, and obtain the user interest dynamic change sensitivity data. The interest fusion layer is connected to the multi-interest extraction layer.

Images