Multi-scale behavior sequence cascaded user intention analysis and content recommendation method
Through the user intention analysis method of multi-scale behavioral sequence cascade, the graph convolution network and non-sampling loss strategy are used to solve the problem of existing systems ignoring the fine-grained dependence relationship between behavioral sequences, and more accurate user interest modeling and content recommendation effects are achieved.
Patent Information
- Application Number
- CN202411610844.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-12
- Publication Date
- 2025-04-01
AI Technical Summary
The existing multi-behavior sequence recommendation system ignores the fine-grained dependencies between behavior sequences, making it difficult to effectively analyze the sequence and timing characteristics of different types of behaviors, resulting in insufficient user interest modeling.
The user intention analysis method of multi-scale behavior sequence cascade is adopted to extract user, content and behavior feature representations of different scales through graph convolution networks, build user-content interaction subgraphs, learn user-content feature representations, and optimize node representations through non-sampling loss strategy to achieve fine-grained dependency modeling between behavior sequences.
Through the multi-scale behavior sequence cascade method, the fine-grained characteristics of user interests can be captured more accurately, and the accuracy and effectiveness of multi-behavior sequence recommendations can be improved.
Smart Images
Figure CN119739921A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of content recommendation, and in particular relates to a user intention analysis and content recommendation method of multi-scale behavior sequence cascade. Background Art
[0002] Multi-behavior sequence recommendation mainly analyzes the user's multiple behavior sequence data (such as browsing, clicking, collecting, forwarding, etc.), learns the sequence interaction pattern of user's diverse behaviors, models the mutual influence and dependency between different behaviors, and realizes accurate personalized content recommendation. It has been widely used in e-commerce, social media and other fields. Existing multi-behavior sequence recommendation systems capture the contextual dependencies of different user behaviors, realize the enhanced representation of user interests based on the dependencies between behaviors, and improve the accuracy of recommendations. However, different types of behaviors have sequential and temporal characteristics. It is important to conduct multi-scale analysis of different types of behaviors, analyze the interest preferences in a single behavior, and the superposition effect of multiple behaviors, and model the fine-grained interests of users based on the dependency diffusion and propagation effects between behavior sequences for accurate prediction of users' click-through rate on content.
[0003] In the modeling of user interests in multi-behavior sequence recommendation, how to achieve the superposition and diffusion of user interests between multiple behaviors at different scales is crucial for the modeling of sequence features. GCN is used to capture the relationship between users and content at different scales, and learn the feature representation of users and content; different types of behavior sequences are transmitted step by step, and GCN is used to learn the sequence feature representation of users and content at different scales; then the cascade feature representation of behavior sequences at different scales is integrated to generate a unified user and content feature representation; the likelihood of users clicking on content is predicted based on different types of behaviors, and the node representation is optimized through the non-sampling loss strategy to achieve the target behavior click prediction of candidate content, thereby improving the accuracy of target behavior prediction for multi-behavior sequence recommendation. Summary of the invention
[0004] To address the problem that existing systems ignore the contribution of fine-grained dependencies between behavior sequences in behavior representation, the present invention provides a user intention analysis and content recommendation method for multi-scale behavior sequence cascade.
[0005] In order to achieve the above object, the present invention adopts the following technical solutions:
[0006] The method for analyzing user intention and recommending content based on multi-scale behavior sequence cascade includes the following steps:
[0007] Step S1, using graph convolutional networks to extract feature representations of users, content, and behaviors at different scales under any single behavior;
[0008] Furthermore, the specific steps of step S1 are:
[0009] Step 1.1: Under any single behavior k, construct a user-content interaction subgraph G k , according to the idea of LightGCN graph convolutional network, the feature representation of learning user nodes is:
[0010]
[0011] In formula (1), σ is a nonlinear activation function, is the feature representation of user u at scale l+1 propagation layer under single behavior k, is the feature representation of content i in the propagation layer of scale l for a single row k, is the feature representation of a single row k at the propagation layer of scale l, is the number of interactive neighbors of user u under single behavior k, is the number of interactive neighbors of content i under a single line k, ⊙ represents the product of corresponding elements, is the user feature learning parameter of the propagation layer of scale l under a single behavior k, and d is the dimensional space;
[0012] Step 1.2: Under any single behavior k, according to the user-content interaction subgraph G k , based on the idea of LightGCN graph convolutional network, the feature representation of the learning content node is:
[0013]
[0014] In formula (2), σ is a nonlinear activation function, is the feature representation of content i in the propagation layer of scale l+1 at a single line k, is the feature representation of user u in the propagation layer of scale l under single behavior k, Learning parameters for content features of the propagation layer at scale l for a single behavior k;
[0015] Step 1.3: The feature representation of a single row k is propagated from the scale l propagation layer to the scale l+1 layer in the graph convolutional network as follows:
[0016]
[0017] In formula (3), is the learnable parameter of the propagation layer with scale l for the k-feature representation of a single row.
[0018] Step S2, under any single behavior, calculate the feature representation of users and contents of the cascade of behavior sequences of different scales;
[0019] Furthermore, the specific steps of step S2 are:
[0020] Step 2.1: fuse the user node feature representation of the scale l propagation layer under a single behavior k with the user node feature representation of the scale l propagation layer under the previous behavior k-1, and calculate the user feature representation of the scale l propagation layer behavior sequence cascade for:
[0021]
[0022] Step 2.2: fuse the content node feature representation of the scale l propagation layer under a single behavior k with the content node feature representation of the scale l propagation layer under the previous behavior k-1, and calculate the content feature representation of the cascade of the scale l propagation layer behavior sequence for:
[0023]
[0024] Step S3, under any single behavior, the feature representations of the user and content of the cascaded behavior sequence under multiple scales are integrated, and the final feature representations of the user and content under different behaviors are calculated respectively;
[0025] Furthermore, the specific steps of step S3 are:
[0026] Step 3.1, based on the user node feature representation of the cascaded behavior sequence of scale l propagation layer under single behavior k and the content node feature representation of the cascaded behavior sequence, the user node feature representation and the content node feature representation are updated respectively. The process is expressed as:
[0027]
[0028] Step 3.2, repeat step 1.1, average pool the user node feature representation of the behavior sequence cascade of each scale propagation layer under single behavior k, and calculate the final feature representation of user u as
[0029]
[0030] In formula (8), is the feature representation of the Lth layer of user u under single behavior k; meanpooling means average pooling;
[0031] Step 3.3, repeat step 1.2, average pool the content node feature representation of the behavior sequence cascade of each scale propagation layer under single behavior k, and calculate the final feature representation of content i as
[0032]
[0033] In formula (9), is the feature representation of the Lth layer of content i in a single line k.
[0034] Step S4, average pooling is performed on the final feature representations of users and contents under different behaviors to obtain user feature representations and content feature representations;
[0035] Furthermore, the specific steps of step S4 are:
[0036] Step 4.1: Average pooling is performed on the final feature representation of user u under all behaviors, and the user feature representation is calculated as e u :
[0037]
[0038] Step 4.2: Average pooling is performed on the final feature representation of content i under all behaviors, and the content feature representation is calculated as e i :
[0039]
[0040] In the formula, meanpooling means average pooling.
[0041] Step S5, based on the user feature representation and the content feature representation, the likelihood of the user clicking on the content when performing different behaviors is calculated, the likelihood of the user clicking on the content under different behaviors is jointly optimized through the non-sampling transfer loss function strategy, and the model parameters are updated to learn the final feature representation of the user and the content;
[0042] Furthermore, the specific steps of step S5 are:
[0043] Step 5.1, based on the user feature representation and content feature representation, calculate the likelihood of a user clicking on content based on the embedding dimension d under a single behavior k:
[0044]
[0045] In formula (12), n is the loop counter; r k (n) is the n-th dimension feature representation value of a single row k;
[0046] Step 5.2: Calculate the non-sampled migration loss function for a single action k as L k :
[0047]
[0048] In formula (13), U b is a batch of user sets during training, I is the content set, is the set of content that users interact with under k in a single behavior, Respectively The positive example weight and negative example weight of m and n are both loop counters; is the m-th dimension value of the final feature representation of user u under single behavior k; r k (m) is the value of the mth dimension of the embedding representation of a single row k; is the m-th dimension value of the final feature representation of content i under single row k;
[0049] Step 5.3, the sum of the non-sampled transfer loss functions of multiple behaviors is:
[0050]
[0051] In formula (14), λ k is the weighting coefficient of the non-sampling loss for a single behavior k, and K represents the target behavior K.
[0052] Step S6, based on the optimized user and content feature representation, calculate the likelihood of users clicking on new content under the target behavior, sort the new content according to the click likelihood, and generate a top-t content recommendation list;
[0053] Furthermore, the specific steps of step S6 are:
[0054] Step 6.1: Based on the embedding representation r under the target behavior K K , User feature representation u and content feature representation v , calculate the likelihood of a user clicking on content based on the embedding dimension d under the target behavior K as
[0055]
[0056] In formula (15), n is the loop counter; r K (n) is the n-th dimension feature representation value of the target behavior K;
[0057] Step 6.2, for multiple candidate contents, calculate the likelihood of user u clicking on the content under the target behavior K, and generate a top-t content recommendation list with high scores.
[0058] Compared with the prior art, the present invention has the following advantages:
[0059] The method provided in the present invention is different from the existing methods in that it designs a user intention analysis and content recommendation method of multi-scale behavior sequence cascade, makes full use of the interest dependency of user behavior sequence, realizes the user interest modeling representation of the layer-by-layer diffusion fusion of behavior sequence, and constructs the user and content representation method of behavior sequence cascade at different scales, and realizes the prediction of user interest in content by using multi-scale behavior sequence cascade interest, and produces a recommended content list. The present invention models the interest dependency between different behaviors by designing a behavior sequence cascade diffusion method, and integrates the characteristics of users and content learned by behavior sequence cascade interests at different scales, so as to realize the capture of user multi-scale interests, which can be effectively promoted in the content recommendation application of multi-behavior sequence. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 It is a schematic diagram of the overall model architecture of the present invention. DETAILED DESCRIPTION
[0061] In order to gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description of the present invention. However, the present invention has multiple implementations and is not limited to the specific examples listed herein. The presentation of these examples is intended to deepen the comprehensive understanding of the disclosure of the present invention.
[0062] The user intention analysis and content recommendation method of multi-scale behavior sequence cascade described in the present invention is implemented by a computer program. The specific implementation method of the technical solution proposed by the present invention will be described in detail below according to the process.
[0063] like Figure 1 As shown, the user intention analysis and content recommendation method of the multi-scale behavior sequence cascade of the present invention includes the following steps:
[0064] Step S1, using graph convolutional networks to extract feature representations of users, content, and behaviors at different scales under any single behavior;
[0065] Furthermore, the specific steps of step S1 are:
[0066] Step 1.1: Under any single behavior k, construct a user-content interaction subgraph G k , according to the idea of LightGCN graph convolutional network, the feature representation of learning user nodes is:
[0067]
[0068] In formula (1), σ is a nonlinear activation function, is the feature representation of user u at scale l+1 propagation layer under single behavior k, is the feature representation of content i in the propagation layer of scale l for a single row k, is the feature representation of a single row k at the propagation layer of scale l, is the number of interactive neighbors of user u under single behavior k, is the number of interactive neighbors of content i under a single line k, ⊙ represents the product of corresponding elements, is the user feature learning parameter of the propagation layer of scale l under a single behavior k, and d is the dimensional space;
[0069] Step 1.2: Under any single behavior k, according to the user-content interaction subgraph G k , based on the idea of LightGCN graph convolutional network, the feature representation of the learning content node is:
[0070]
[0071] In formula (2), σ is a nonlinear activation function, is the feature representation of content i in the propagation layer of scale l+1 at a single line k, is the feature representation of user u in the propagation layer of scale l under single behavior k, Learning parameters for content features of the propagation layer at scale l for a single behavior k;
[0072] Step 1.3: The feature representation of a single row k is propagated from the scale l propagation layer to the scale l+1 layer in the graph convolutional network as follows:
[0073]
[0074] In formula (3), is the learnable parameter of the propagation layer with scale l for the k-feature representation of a single row.
[0075] Step S2, under any single behavior, calculate the feature representation of users and contents of the cascade of behavior sequences of different scales;
[0076] Furthermore, the specific steps of step S2 are:
[0077] Step 2.1: fuse the user node feature representation of the scale l propagation layer under a single behavior k with the user node feature representation of the scale l propagation layer under the previous behavior k-1, and calculate the user feature representation of the scale l propagation layer behavior sequence cascade for:
[0078]
[0079] Step 2.2: fuse the content node feature representation of the scale l propagation layer under a single behavior k with the content node feature representation of the scale l propagation layer under the previous behavior k-1, and calculate the content feature representation of the cascade of the scale l propagation layer behavior sequence for:
[0080]
[0081] Step S3, under any single behavior, the feature representations of the user and content of the cascaded behavior sequence under multiple scales are integrated, and the final feature representations of the user and content under different behaviors are calculated respectively;
[0082] Furthermore, the specific steps of step S3 are:
[0083] Step 3.1, based on the user node feature representation of the cascaded behavior sequence of scale l propagation layer under single behavior k and the content node feature representation of the cascaded behavior sequence, the user node feature representation and the content node feature representation are updated respectively. The process is expressed as:
[0084]
[0085]
[0086] Step 3.2, repeat step 1.1, average pool the user node feature representation of the behavior sequence cascade of each scale propagation layer under single behavior k, and calculate the final feature representation of user u as
[0087]
[0088] In formula (8), is the feature representation of the Lth layer of user u under single behavior k; meanpooling means average pooling;
[0089] Step 3.3, repeat step 1.2, average pool the content node feature representation of the behavior sequence cascade of each scale propagation layer under single behavior k, and calculate the final feature representation of content i as
[0090]
[0091] In formula (9), is the feature representation of the Lth layer of content i in a single line k.
[0092] Step S4, average pooling is performed on the final feature representations of users and contents under different behaviors to obtain user feature representations and content feature representations;
[0093] Furthermore, the specific steps of step S4 are:
[0094] Step 4.1: Average pooling is performed on the final feature representation of user u under all behaviors, and the user feature representation is calculated as e u :
[0095]
[0096] Step 4.2: Average pooling is performed on the final feature representation of content i under all behaviors, and the content feature representation is calculated as e i :
[0097]
[0098] In the formula, meanpooling means average pooling.
[0099] Step S5, based on the user feature representation and the content feature representation, the likelihood of the user clicking on the content when performing different behaviors is calculated, the likelihood of the user clicking on the content under different behaviors is jointly optimized through the non-sampling transfer loss function strategy, and the model parameters are updated to learn the final feature representation of the user and the content;
[0100] Furthermore, the specific steps of step S5 are:
[0101] Step 5.1, based on the user feature representation and content feature representation, calculate the likelihood of a user clicking on content based on the embedding dimension d under a single behavior k:
[0102]
[0103] In formula (12), n is the loop counter; r k (n) is the n-th dimension feature representation value of a single row k;
[0104] Step 5.2: Calculate the non-sampled migration loss function for a single action k as L k :
[0105]
[0106] In formula (13), U b is a batch of user sets during training, I is the content set, is the set of content that users interact with under k in a single behavior, Respectively The positive example weight and negative example weight of m and n are both loop counters; is the m-th dimension value of the final feature representation of user u under single behavior k; r k (m) is the value of the mth dimension of the embedding representation of a single row k; is the m-th dimension value of the final feature representation of content i under single row k;
[0107] Step 5.3, the sum of the non-sampled transfer loss functions of multiple behaviors is:
[0108]
[0109] In formula (14), λ k is the weighting coefficient of the non-sampling loss for a single behavior k, and K represents the target behavior K.
[0110] Step S6, based on the optimized user and content feature representation, calculate the likelihood of users clicking on new content under the target behavior, sort the new content according to the click likelihood, and generate a top-t content recommendation list;
[0111] Furthermore, the specific steps of step S6 are:
[0112] Step 6.1: Based on the embedding representation r under the target behavior K K , User feature representation u and content feature representation v , calculate the likelihood of a user clicking on content based on the embedding dimension d under the target behavior K as
[0113]
[0114] In formula (15), n is the loop counter; r K (n) is the n-th dimension feature representation value of the target behavior K;
[0115] Step 6.2, for multiple candidate contents, calculate the likelihood of user u clicking on the content under the target behavior K, and generate a top-t content recommendation list with high scores.
[0116] In order to verify the effectiveness of the method, experiments were conducted on the user multi-behavior interaction dataset of Tmall product item content (https: / / tianchi.aliyun.com / dataset / 140281). The dataset contains multiple types of user-item interaction behaviors, namely: browsing, collection, shopping cart, and forwarding. The dataset information is shown in Table 1:
[0117] Table 1 Dataset
[0118]
[0119] The evaluation indicators used are HR@10, HR@50, NDCG@10, and NDCG@50. In order to verify the effectiveness and advancement of the technical solution proposed in this invention, several existing recommended prediction model methods are selected for comparison: MBGMN, GHCF, CML, and MBSSL. The experimental results are shown in Table 2:
[0120] Table 2 Experimental results
[0121]
[0122] It can be seen from the results in Table 2 that the technical solution of the present invention can obtain a ranking index of recommendation results that is better than the detection results of existing methods in the process of predicting the user's interest in project content.
[0123] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.
Claims
1. A user intention analysis and content recommendation method based on multi-scale behavior sequence cascade, characterized in that: The following steps are involved: Step S1, using graph convolutional networks to extract feature representations of users, content, and behaviors at different scales under any single behavior; Step S2, under any single behavior, calculate the feature representation of users and contents of the cascade of behavior sequences of different scales; Step S3, under any single behavior, the feature representations of the user and content of the cascaded behavior sequence under multiple scales are integrated, and the final feature representations of the user and content under different behaviors are calculated respectively; Step S4, average pooling is performed on the final feature representations of users and contents under different behaviors to obtain user feature representations and content feature representations; Step S5, based on the user feature representation and the content feature representation, the likelihood of the user clicking on the content when performing different behaviors is calculated, the likelihood of the user clicking on the content under different behaviors is jointly optimized through the non-sampling transfer loss function strategy, and the model parameters are updated to learn the final feature representation of the user and the content; Step S6, based on the optimized user and content feature representation, calculate the likelihood of users clicking on new content under the target behavior, sort the new content according to the click likelihood, and generate a top-t content recommendation list.
2. The method for user intention analysis and content recommendation based on multi-scale behavior sequence cascade according to claim 1, characterized in that: The step S1 uses a graph convolutional network to extract feature representations of users, content, and behaviors of different scales under any single behavior; The specific steps are: Step 1.1: Under any single behavior k, construct a user-content interaction subgraph G k , according to the idea of LightGCN graph convolutional network, the feature representation of learning user nodes is: In formula (1), σ is a nonlinear activation function, is the feature representation of user u at scale l+1 propagation layer under single behavior k, is the feature representation of content i in the propagation layer of scale l for a single row k, is the feature representation of a single row k at the propagation layer of scale l, is the number of interactive neighbors of user u under single behavior k, is the number of interactive neighbors of content i under a single line k, ⊙ represents the product of corresponding elements, is the user feature learning parameter of the propagation layer of scale l under a single behavior k, and d is the dimensional space; Step 1.2: Under any single behavior k, according to the user-content interaction subgraph G k , based on the idea of LightGCN graph convolutional network, the feature representation of the learning content node is: In formula (2), σ is a nonlinear activation function, is the feature representation of content i in the propagation layer of scale l+1 at a single line k, is the feature representation of user u in the propagation layer of scale l under single behavior k, Learning parameters for content features of the propagation layer at scale l for a single behavior k; Step 1.3: The feature representation of a single row k is propagated from the scale l propagation layer to the scale l+1 layer in the graph convolutional network as follows: In formula (3), is the learnable parameter of the propagation layer with scale l for the k-feature representation of a single row.
3. The method for user intention analysis and content recommendation based on multi-scale behavior sequence cascade according to claim 2, characterized in that: The step S2, in any single behavior, calculates the feature representation of the user and content of the behavior sequence cascade of different scales, and the specific steps are: Step 2.1: fuse the user node feature representation of the scale l propagation layer under a single behavior k with the user node feature representation of the scale l propagation layer under the previous behavior k-1, and calculate the user feature representation of the scale l propagation layer behavior sequence cascade for: Step 2.2: fuse the content node feature representation of the scale l propagation layer under a single behavior k with the content node feature representation of the scale l propagation layer under the previous behavior k-1, and calculate the content feature representation of the cascade of the scale l propagation layer behavior sequence for:
4. The method for user intention analysis and content recommendation based on multi-scale behavior sequence cascade according to claim 3 is characterized in that: The step S3, in any single behavior, fuses the feature representations of users and contents in the cascade of behavior sequences at multiple scales, and respectively calculates the final feature representations of users and contents under different behaviors. The specific steps are: Step 3.1, based on the user node feature representation of the cascaded behavior sequence of scale l propagation layer under single behavior k and the content node feature representation of the cascaded behavior sequence, the user node feature representation and the content node feature representation are updated respectively. The process is expressed as: Step 3.2, repeat step 1.1, average pool the user node feature representation of the behavior sequence cascade of each scale propagation layer under single behavior k, and calculate the final feature representation of user u as In formula (8), is the feature representation of the Lth layer of user u under single behavior k; meanpooling means average pooling; Step 3.3, repeat step 1.2, average pool the content node feature representation of the behavior sequence cascade of each scale propagation layer under single behavior k, and calculate the final feature representation of content i as In formula (9), is the feature representation of the Lth layer of content i in a single line k.
5. The method for user intention analysis and content recommendation based on multi-scale behavior sequence cascade according to claim 4, characterized in that: The step S4 is to average pool the final feature representations of users and contents under different behaviors to obtain user feature representations and content feature representations. The specific steps are as follows: Step 4.1: Average pooling is performed on the final feature representation of user u under all behaviors, and the user feature representation is calculated as e u : Step 4.2, average pool the final feature representation of content i under all behaviors and calculate the content feature representation as e i : In the formula, meanpooling means average pooling.
6. The method for user intention analysis and content recommendation based on multi-scale behavior sequence cascade according to claim 5, characterized in that: The step S5 calculates the likelihood of a user clicking on content when performing different behaviors based on the user feature representation and the content feature representation, jointly optimizes the likelihood of a user clicking on content under different behaviors through a non-sampling transfer loss function strategy, and updates the model parameters to learn the final feature representation of the user and content. The specific steps are: Step 5.1, based on the user feature representation and content feature representation, calculate the likelihood of a user clicking on content based on the embedding dimension d under a single behavior k: In formula (12), n is the loop counter; r k (n) is the n-th dimension feature representation value of a single row k; Step 5.2: Calculate the non-sampled migration loss function for a single action k as L k : In formula (13), U b is a batch of user sets during training, I is the content set, is the set of content that users interact with under k in a single behavior, Respectively The positive example weight and negative example weight of m and n are both loop counters; is the m-th dimension value of the final feature representation of user u under single behavior k; r k (m) is the value of the mth dimension of the embedding representation of a single row k; is the m-th dimension value of the final feature representation of content i under single row k; Step 5.3, the sum of the non-sampled transfer loss functions of multiple behaviors is: In formula (14), λ k is the weighting coefficient of the non-sampling loss for a single behavior k, and K represents the target behavior K.
7. The method for user intention analysis and content recommendation based on multi-scale behavior sequence cascade according to claim 6, characterized in that: The step S6, based on the optimized user and content feature representation, calculates the likelihood of users clicking on new content under the target behavior, sorts the new content according to the click likelihood, and generates a top-t content recommendation list. The specific steps are: Step 6.1: Based on the embedding representation r under the target behavior K K , User feature representation u and content feature representation v , calculate the likelihood of a user clicking on content based on the embedding dimension d under the target behavior K as In formula (15), n is the loop counter; r K (n) is the n-th dimension feature representation value of the target behavior K; Step 6.2, for multiple candidate contents, calculate the likelihood of user u clicking on the content under the target behavior K, and generate a top-t content recommendation list with high scores.