A sequence recommendation method based on attenuated emotion perception

Abstract

The present invention discloses a sequence recommendation method based on attenuated emotion perception, and establishes a attenuated emotion memory network to simulate the dynamic nonlinear attenuation of human attenuated emotions. The memory network adopts an exponential memory loss function to calculate the attenuation rate of emotion according to the time interval between two adjacent items in the sequence, and uses the attenuation rate to adjust the long-term memory unit of the network at each moment to learn the true attenuated emotion characteristics. The present invention proposes a sequence recommendation model based on attenuated emotion perception, which processes two channels, namely emotion attenuation preference and user item preference, and fuses the two channels to obtain fused preferences to explore the user's potential preferences. The present invention can be applied to Chinese application scenarios to simulate the attenuation law of user emotions in the real world, construct a sequence recommendation model based on attenuated emotion perception, and extend its application to real scenarios and platforms containing Chinese text reviews.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recommendation and deep learning technology, and in particular to a sequence recommendation method based on attenuated emotion perception. Background Art

[0002] Recommendation systems are currently widely used on platforms such as e-commerce, short videos, music, and movies. Sequential recommendation systems capture useful information from users' personalized past and recent behaviors, using this information to predict their next actions and make recommendations. Several advanced inventions have been made in this field.

[0003] Markov chains (MCs) are a classic model for sequential recommendation, which assumes that the next action depends only on the previous action (or a few previous actions). Personalized Markov chain methods combine matrix decomposition with Markov chains, which can simultaneously capture temporal information and long-term user preference information. The GRU4Rec model initially applies recurrent neural networks (RNNs) in the field of deep learning to the field of sequential recommendation, reasonably connects the behavior sequences of different users, and solves the problem of uniformly modeling user behavior sequences of different lengths. In addition to the mechanism based on recurrent neural networks, the attention-based mechanism has achieved very significant results in fields such as real-time subtitles and machine translation [5]. This mechanism has been introduced into the field of recommendation systems and has shown better performance. For example, the SASRec model captures users' long-term preferences from sparse and dense data and achieves significant prediction results. Unlike these sequential recommendation systems, the SGS model, which is modeled based on English sentiment, considers the guiding role of user sentiment on user behavior to make recommendations for users.

[0004] The SASRec model, described above, is a recommendation system for user behavior sequences based on a self-attention mechanism. This model can capture the attentional relationships between users' consecutive purchases of items. However, it fails to consider subjective human emotions and overlooks the impact of continuous emotional changes on user behavior. The SGS model models the sentiment of consecutive English review texts and lacks a sequential recommendation modeling approach based on the sentiment of Chinese text. Its application needs to be further expanded. Secondly, the SGS model only models the explicit sentiment of consecutive review texts, ignoring the objective decay of human emotions over time and thus failing to fully tap into users' latent emotional information. Furthermore, the TimeSASRec model considers the temporal relationships between consecutive user behaviors and explores the impact of different time intervals on recommendation results. However, this approach only considers the temporal relationships between items interacted with in user behavior, ignoring the influence of user emotions on shopping behavior or the impact of user emotions on subjective behavior. Summary of the Invention

[0005] The technical problem to be solved by the embodiments of this invention is to provide a sequential recommendation method based on decaying emotion perception. This method can be applied to Chinese application scenarios to simulate the decaying patterns of user emotions in the real world, construct a sequential recommendation model based on decaying emotion perception, and extend its application to real-world scenarios and platforms containing Chinese text reviews.

[0006] In order to solve the above technical problems, an embodiment of the present invention provides a sequence recommendation method based on attenuated emotion perception, comprising the following steps:

[0007] S1: Construct an item embedding matrix for the item set of the dataset , where d represents the hidden dimension and I is the set of items. Adding position embedding to the embedding matrix creates a position matrix P with the same shape as the item embedding matrix E: ;

[0008] S2: Define the attention mechanism as , where Q, K, and V represent the query, key, and value in the attention mechanism, respectively. is the scale factor;

[0009] The attention mechanism is linearly mapped from the input Get Q, K, V and input them into the attention layer:

[0010]

[0011] in is a learnable projection matrix, Both represent the multiplication of two matrices, and the model automatically adjusts during training The element value of A E Representation is the result matrix learned by the model from the user-item sequence;

[0012] S3: Use the self-attention mechanism to calculate the correlation weight between two items, so as to learn the potential relationship P between the items based on historical behavior, and then pass the preference through a feedforward neural network to generate the item preference ;

[0013] S4: Construct a decaying emotion memory network to simulate the nonlinear decay of emotion, and establish a decaying emotion-aware self-attention channel to mine the user's potential preferences and obtain a new user emotion preference matrix ;

[0014] S5: The preferences generated by emotions and the preferences generated by items are integrated, so that the preference channel of item self-guidance generates preferences based on historical behavior, and the emotional preference channel of attenuated emotion perception mines emotional preferences by calculating the attenuated emotional information, which is used to accurately discover potential needs.

[0015] Wherein, the S3 further includes:

[0016] According to the attention mechanism, the item preference is defined as:

[0017]

[0018]

[0019] in and is a learnable weight matrix, and is the learnable bias matrix, It represents matrix multiplication. It is the user's preference matrix for items that the model automatically learns after the item-guided preference channel. yes The output matrix obtained after FNN.

[0020] Wherein, the S4 further includes:

[0021] S41: Steps for calculating personalized emotional time intervals

[0022] For a given user Historical item series The time it takes to interact with each item , if the sequence length is less than , padding on the left until the length is , if the length is greater than , only consider the most recent elements, the time interval between two items Defined as:

[0023] ,make , let the maximum time interval between two items be ,get , get the personalized time interval sequence of user u ;

[0024] S42: Emotional decay rate

[0025] For each user, the sentiment decay rate at each time step is defined as:

[0026]

[0027] in is the timestamp of the user’s i-th personalized time interval, and Personalize time interval sequences for users the maximum and minimum intervals in ;

[0028] S43: Decaying Emotional Memory Network

[0029] For each ASM-Net unit at time t, the long-term memory matrix Changing to a decaying long-term memory matrix , which is defined as:

[0030]

[0031]

[0032] in, is the emotion decay rate at time t, and is calculated by the emotion decay rate. is the long-term memory matrix, Represents the long-term memory matrix Multiply by the decay rate The intermediate matrix obtained later is represents the learnable weight matrix, represents the learnable bias matrix, It means and Matrix multiplication, To obtain the decayed long-term memory matrix;

[0033] ASM-Net uses the original sentiment matrix at time t As input, For the original sentiment embedding matrix obtained after the embedding layer, calculate the attenuated sentiment matrix , the sentiment forgetting gate in ASM-Net is defined as:

[0034]

[0035] in It represents the sigmoid neural network layer. represents the learnable weight matrix, represents the learnable bias matrix, Represents the original sentiment matrix and the decaying emotion matrix of the previous moment Perform splicing operations. It represents the output of the emotion forgetting gate;

[0036] The emotional memory gate is defined as:

[0037]

[0038]

[0039] in is the output of the emotional memory gate, It represents the tanh function The output after the operation. represents the learnable weight matrix, represents the learnable bias matrix. It represents the sigmoid neural network layer, and the function tanh represents the mathematical hyperbolic tangent operation. Represents the original sentiment matrix and the decaying emotion matrix of the previous moment Perform splicing operations.

[0040] Define the sentiment output gate as:

[0041]

[0042] in is the output of the emotion output gate, represents the learnable weight matrix, represents the learnable bias matrix, It represents the sigmoid neural network layer. Represents the original sentiment matrix and the decaying emotion matrix of the previous moment Perform splicing operations;

[0043] The process of updating the network element status is defined as:

[0044]

[0045]

[0046] in, Represents the emotional memory matrix attenuated at the current moment, is the intermediate result matrix;

[0047] ASM-Net is defined as

[0048]

[0049]

[0050] It represents the output of ASM-Net. It represents the output after the FNN network;

[0051] S44: Attentional Mechanisms for Attenuating Emotional Perception

[0052] Will attenuate the emotional matrix Q as an attention mechanism, embedding the item into the matrix As K and V, the user sentiment preference matrix obtained by the sentiment attention mechanism is defined as:

[0053] ,in, , represents the learnable weight matrix, It represents matrix multiplication

[0054] The new user sentiment preference matrix obtained after the FNN layer is:

[0055] .

[0056] Wherein, the S5 further includes:

[0057] Define the combined user behavior-sentiment preference matrix:

[0058]

[0059] in, , is the dimension size of the output;

[0060] An LSTM network is used to integrate sentiment preferences and item preferences in long-term contexts, and layer normalization is used to stabilize and accelerate network training.

[0061] The implementation of the embodiments of the present invention has the following beneficial effects:

[0062] 1. Sequential Recommendation Model Based on Attenuated Sentiment Perception

[0063] Unlike traditional sequential recommendation models guided by emotion perception, this paper specifically designs a decaying emotion memory network (ASM-Net) to simulate the complex changes in user emotions in real-world situations. This network can exploit the patterns of emotion decay in complex real-world situations, facilitating realistic simulation of user emotions and capturing their underlying preferences. This is the first sequential recommendation model to employ decaying emotion perception.

[0064] 2. Good performance on Chinese industrial datasets

[0065] Unlike traditional recommendation systems, this approach uses a real-world Chinese dataset from industry for experiments. Experimental results demonstrate that this approach performs well on large, sparse datasets. Compared to baseline models for sequential recommendation, the ASAR model significantly improves recommendation performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a diagram showing the connection between user emotional changes and shopping behavior in real scenarios;

[0067] Figure 2 Schematic diagram of the ASAR model in the present invention;

[0068] Figure 3 It is the data set details table;

[0069] Figure 4 It is the result of comparative test with existing advanced models;

[0070] Figure 5 This is the ablation experiment result on dataset 1;

[0071] Figure 6 It is the ablation experiment result on dataset 2. DETAILED DESCRIPTION

[0072] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be described in further detail below with reference to the accompanying drawings.

[0073] A sequence recommendation method based on attenuated emotion perception in an embodiment of the present invention is implemented through the following process.

[0074] like Figure 1 This paper establishes an Attenuated Affective Memory Network (ASM-Net) to simulate the dynamic nonlinear decay of human emotions. This memory network uses an exponential memory loss function to calculate the decay rate of emotion based on the time interval between two adjacent items in the sequence. The decay rate is then used to adjust the network's long-term memory units at each moment to learn the true decayed emotional characteristics.

[0075] On this basis, the present invention proposes an Attenuated Sentiment-Aware Sequential Recommendation (ASAR) model based on attenuated sentiment perception. The model processes from two channels, namely sentiment attenuation preference and user item preference, and fuses the two channels to obtain fused preference to mine and recommend items.

[0076] 1. Setting up the Embedding layer

[0077] First, construct an item embedding matrix for the item set in the dataset , where d represents the hidden dimension (a hyperparameter that needs to be set manually) and I is the set of items. Since the self-attention mechanism cannot learn the sequential information of the user behavior sequence, a position embedding matrix is ​​added to the embedding matrix to be consistent with the item embedding matrix. The position matrix P has the same shape, that is, .

[0078]

[0079] Similarly, for the sentiment dimension, for the original sentiment input with N dimensions Where N represents the dimension of sentiment. This paper specifically models Chinese sentiment, extracting three sentiment dimensions from user comments: positive sentiment (pos), negative sentiment (neg), and sentiment confidence (con), with N = 3. The values ​​of these three sentiment dimensions range from [0, 100]. In an actual recommendation system, a specific tuple (item, pos, neg, con) is associated with user u at a given moment.

[0080] Similar to the item embedding matrix, a specific sentiment value needs to be mapped into a sentiment embedding matrix. That is, similar to the operation of formula (1), the sentiment sequence Transformed into the corresponding embedding matrix ,i=1,2,...,N. It represents the i-th original emotion.

[0081] 2. Self-Attention Mechanism

[0082] This embodiment defines the attention mechanism:

[0083]

[0084] in Represent the query, key and value in the attention mechanism respectively, is the scaling factor. In the calculation, Q, K, and V represent tensors, Q and K perform matrix multiplication, and T represents the matrix transpose operation.

[0085] In the input matrix As input, the self-attention mechanism [2] needs to go through a linear mapping, from the input get , and input into the attention layer:

[0086]

[0087] in is a learnable projection matrix, Both represent the multiplication of two matrices, and the model automatically adjusts during training The input matrix E can be a user-defined input matrix, which is usually obtained through the embedding layer; Representation is the result matrix learned by the model from the user-item sequence.

[0088] 3. Item-guided preference channel

[0089] This paper adopts the self-attention mechanism and designs an item-guided preference channel to capture the long-term semantics in historical behavior. The item self-attention channel uses the self-attention mechanism to calculate the correlation weight between two items, thereby learning the potential relationship P between the items based on historical behavior, and then passes the preference through a feedforward neural network (FNN) to generate the item preference. According to the self-attention mechanism in the second point, item preference can be defined as:

[0090]

[0091]

[0092] in and is a learnable weight matrix, and is the learnable bias matrix, It represents matrix multiplication. In deep learning, both the weight matrix and the bias matrix need to be learned automatically by the model. This is the item embedding matrix obtained through the embedding layer in point 1. It is the user's preference matrix for items that the model automatically learns after the item-guided preference channel. that is The output matrix obtained after FNN.

[0093] 4. Attenuate the preference channel of emotional perception

[0094] This embodiment of the present invention designs a method for calculating decaying emotions using time intervals and constructs a decaying emotion memory network to simulate the nonlinear decay of emotions. Based on this, a self-attention channel for decaying emotion perception is established to explore users' potential preferences.

[0095] 4.1 Personalized Emotional Time Interval Calculation

[0096] For a given user Historical item series The time it takes to interact with each item , the present invention adopts the following operation: if the sequence length is less than , the present invention will choose to fill on the left until the length is If the length is greater than , this invention only considers the most recent The time interval between two items. Defined as:

[0097]

[0098] Generally, the present invention , let the maximum time interval between two items be .Right now:

[0099]

[0100] At this point, the personalized time interval sequence of user u can be obtained .

[0101] 4.2 Emotional Decay Rate

[0102] The present invention uses the emotion decay rate to calculate the nonlinear decay of human emotions. For each user, the present invention defines its emotion decay rate at each time step as

[0103]

[0104] in is the timestamp of the user’s i-th personalized time interval, and Personalize time interval sequences for users The maximum and minimum intervals in .

[0105] 4.3 Attenuated Sentiment Memory Network (ASM-Net)

[0106] The present invention provides an attenuated emotional memory network (ASM-Net) that can attenuate emotional memory information. ASM-Net takes into account the impact of historical emotions on the current situation under the relationship of time. The long-term memory output at the previous moment will be affected by the attenuation rate when it is input at the next moment. The specific structure of ASM-Net is as follows: Figure 2 shown.

[0107] For each ASM-Net unit at time t, the long-term memory matrix Changing to a decaying long-term memory matrix , and this process can be defined as:

[0108]

[0109]

[0110] in, is the emotion decay rate at time t, and is calculated by the emotion decay rate at 4.2 points. is the long-term memory matrix, It means Multiply by the decay rate The intermediate matrix obtained later. represents the learnable weight matrix, Represents the learnable bias matrix. The function tanh represents the mathematical hyperbolic tangent operation (the same below). It means and Matrix multiplication, To obtain the decayed long-term memory matrix.

[0111] ASM-Net uses the original sentiment matrix at time t ( The original sentiment embedding matrix obtained by the embedding layer in Section 1 is used as input to calculate the attenuated sentiment matrix , the sentiment forgetting gate in ASM-Net is defined as:

[0112]

[0113] in It represents the sigmoid neural network layer (the same below). represents the learnable weight matrix, represents the learnable bias matrix. Represents the original sentiment matrix and the decaying emotion matrix of the previous moment Perform splicing operations (same below), It represents the output of the emotion forgetting gate.

[0114] The emotional memory gate is defined as:

[0115]

[0116]

[0117] in is the output of the emotional memory gate, It represents the tanh function The output after the operation. represents the learnable weight matrix, represents the learnable bias matrix. It represents the sigmoid neural network layer, and the function tanh represents the mathematical hyperbolic tangent operation. Represents the original sentiment matrix and the decaying emotion matrix of the previous moment Perform splicing operations.

[0118] The sentiment output gate is defined as:

[0119]

[0120] in is the output of the emotion output gate, represents the learnable weight matrix, Represents the learnable bias matrix. The function tanh represents the mathematical hyperbolic tangent operation. It represents the sigmoid neural network layer. Represents the original sentiment matrix and the decaying emotion matrix of the previous moment Perform splicing operations.

[0121] The process of updating the network element status is defined as:

[0122]

[0123]

[0124] in, Represents the emotional memory matrix attenuated at the current moment, is the intermediate result matrix. The decayed long-term memory matrix is ​​obtained for formula (10), It represents the output of the emotional forgetting gate obtained by formula (11), is the output of the emotional memory gate obtained by formula (12), It represents the output obtained by formula (13), The output of the sentiment output gate is obtained from formula (14). The function tanh represents the mathematical hyperbolic tangent operation.

[0125] So ASM-Net can be defined as:

[0126]

[0127]

[0128] in It represents the original sentiment embedding matrix (N=1,2,3,…) obtained by the embedding layer in Section 1, where N represents the dimension of sentiment. It represents the output of ASM-Net. The FNN network is also added after the output of ASM-Net to increase the model's ability to fit nonlinearities. It represents the output after the FNN network (obtained by formula (5)).

[0129] 4.4 Attentional Mechanism for Attenuating Emotional Perception

[0130] Studies have shown that emotional factors can effectively mine user preferences in sequential recommendation models. On this basis, the user's historical emotions and current emotional characteristics are used to predict the user's future behavior. In particular, the present invention will attenuate the emotional matrix Q as an attention mechanism, embedding the item into the matrix As K and V.

[0131]

[0132]

[0133] in This is the attenuated sentiment matrix obtained after 4.3 points, It is the item embedding matrix obtained by the embedding layer in the first section. , represents the learnable weight matrix, It represents matrix multiplication. It represents the network of formula (5). It represents the user sentiment preference matrix obtained through the sentiment attention mechanism. It represents the new user sentiment preference matrix obtained through the FNN layer.

[0134] 5. Emotional Perception Preference Fusion

[0135] After the above operations, the item self-guided preference channel generates preferences based on historical behavior, and the emotional preference channel of attenuated emotional perception mines emotional preferences by calculating attenuated emotional information, which can be used to accurately discover potential needs. The present invention integrates the preferences generated by emotions and the preferences generated by items:

[0136]

[0137] in, , is the dimension size of the output, which are all learnable weight matrices. is the user's item preference matrix obtained by formula (5), It represents the user sentiment preference matrix obtained in 4.4 points, where N=1,2,3,…. Represents a merge operation. is the combined user behavior-emotion preference matrix.

[0138] Finally, the present invention uses an LSTM network to integrate sentiment preferences and item preferences over a long period of time. To avoid overfitting and instability during training, the present invention uses layer normalization to stabilize and accelerate network training. In addition, a fully connected layer is used as the final output layer to enhance the nonlinearity of the model. This process can be defined as:

[0139]

[0140]

[0141] in, is a learnable weight matrix, is the learnable bias matrix. Represents layer normalization operation, LSTM represents long short-term recurrent neural network, is the combined user behavior-emotion preference matrix obtained above. is the intermediate result matrix obtained by LSTM, This is the result matrix predicted by the final model.

[0142] Layer Normalization: Assuming x is a vector containing sample features, it is defined as:

[0143]

[0144] in, represents the vector dot product, and are the mean and variance of x, denotes the learning rate and bias.

[0145] Residual layer (Dropout Layer): In order to solve the overfitting problem, the network uses the Dropout technology:

[0146]

[0147] in Represents the self-attention mechanism layer or feedforward neural network FNN, and x is the input sample. The present invention uses binary cross entropy loss function and Adam optimizer to train ASAR network

[0148] The embodiments of the present invention have the following advantages:

[0149] 1. Sequential Recommendation Model Based on Attenuated Sentiment Perception

[0150] Unlike traditional sequential recommendation models guided by emotion perception, this paper specifically designs a decaying emotion memory network (ASM-Net) to simulate the complex changes in user emotions in real-world situations. This network can exploit the patterns of emotion decay in complex real-world situations, facilitating realistic simulation of user emotions and capturing their underlying preferences. This is the first sequential recommendation model to employ decaying emotion perception.

[0151] 2. Good performance on Chinese industrial datasets

[0152] Unlike traditional recommendation systems, this approach uses a real-world Chinese dataset from industry for experiments. Experimental results demonstrate that this approach performs well on large, sparse datasets. Compared to baseline models for sequential recommendation, the ASAR model significantly improves recommendation performance.

[0153] The following experimental results are provided to demonstrate the present invention.

[0154] We used crawling technology to obtain a dataset of real user reviews from BiliBili, China's largest entertainment video website. After data processing, we used the Baidu NLP API to perform sentiment analysis on the reviews. The analysis results show that each review record includes three sentiment factors: positivity, negativity, and confidence, reflecting the current emotional state. Figure 3 Lists the statistics of the dataset

[0155] The dataset is divided into three parts: (1) the most recently interacted items as the test set, (2) the second to last interacted items as the validation set, and (3) the rest of the interacted items as the training set.

[0156] (1) Experiment 1: Comparing ASAR with the following related sequence recommendation models

[0157] • SASRec: A sequential recommendation model based on self-attention mechanism.

[0158] • GRU4Rec: GRU-based sequence recommendation model.

[0159] • SGS: A sentiment-aware sequential recommendation model based on self-attention mechanism

[0160] like Figure 4 The experimental results of the comparison with the existing advanced models are shown in the figure. The experiments show that ASAR has better performance than the existing excellent sequence recommendation models.

[0161] (2) Experiment 2: Demonstration of the effectiveness of the emotion attenuation mechanism

[0162] In order to study and prove the effectiveness of the emotion attenuation mechanism in ASAR, the present invention designed an ablation experiment. In the experiment, two variants of ASAR were designed and compared with ASAR.

[0163] • Remove ASM-Net of ASAR: The ASM-Net version of the model in ASAR has been deleted. This means that the model does not include the emotion attenuation mechanism and only considers the impact of the original emotion.

[0164] • LSTM-ASAR: ASAR version model that uses LSTM instead of ASM-Net.

[0165] from Figure 5 and Figure 6 As can be seen, ASAR using ASM-Net outperforms other models in all evaluation metrics. This proves that the emotion attenuation mechanism in this invention is effective and cannot be replaced by ordinary LSTM layers.

[0166] The above disclosure is only a preferred embodiment of the present invention and certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A sequence recommendation method based on attenuated emotion perception, characterized in that: The following steps are involved: S1: Construct an item embedding matrix for the item set of the dataset , where d represents the hidden dimension and I is the set of items. Adding position embedding to the embedding matrix creates a position matrix P with the same shape as the item embedding matrix E: ; S2: Define the attention mechanism as , where Q, K, and V represent the query, key, and value in the attention mechanism, respectively. is the scale factor; The attention mechanism is linearly mapped from the input Get Q, K, V and input them into the attention layer: in is a learnable projection matrix, Both represent the multiplication of two matrices, and the model automatically adjusts during training The element value of A E Representation is the result matrix learned by the model from the user-item sequence; S3: Use the self-attention mechanism to calculate the correlation weight between two items, so as to learn the potential relationship P between the items based on historical behavior, and then pass the preference through a feedforward neural network to generate the item preference ; S4: Construct a decaying emotion memory network to simulate the nonlinear decay of emotion, and establish a decaying emotion-aware self-attention channel to mine the user's potential preferences and obtain a new user emotion preference matrix ; For each ASM-Net unit at time t, the long-term memory matrix Changing to a decaying long-term memory matrix , which is defined as: in, is the emotion decay rate at time t, and is calculated by the emotion decay rate. is the long-term memory matrix, Represents the long-term memory matrix Multiply by the decay rate The intermediate matrix obtained later is represents the learnable weight matrix, represents the learnable bias matrix, It means and Matrix multiplication, To obtain the decayed long-term memory matrix; S5: The preferences generated by emotions and those generated by items are integrated, so that the preference channel of item self-guidance generates preferences based on historical behavior, and the emotion preference channel of attenuated emotion perception mines emotion preferences by calculating the attenuated emotion information, which is used to accurately discover the user's potential needs.

2. The sequence recommendation method based on attenuated emotion perception according to claim 1 is characterized in that Said S3 further comprises: According to the attention mechanism, the item preference is defined as: in and is a learnable weight matrix, and is the learnable bias matrix, It represents matrix multiplication. It is the user's preference matrix for items that the model automatically learns after the item-guided preference channel. yes The output matrix obtained after FNN.

3. The sequence recommendation method based on attenuated emotion perception according to claim 2, characterized in that: Said S4 further comprises: S41: Steps for calculating personalized emotional time intervals For a given user Historical item series The time it takes to interact with each item , if the sequence length is less than , padding on the left until the length is If the length is greater than , only considering the most recent elements, the time interval between two items Defined as: ,make , let the maximum time interval between two items be ,get , get the personalized time interval sequence of user u ; S42: Emotional decay rate For each user, the sentiment decay rate at each time step is defined as: in For the user's personalized time interval timestamps, and Personalize time interval sequences for users the maximum and minimum intervals in ; S43: Decaying Emotional Memory Network ASM-Net uses the original sentiment matrix at time t As input, For the original sentiment embedding matrix obtained after the embedding layer, calculate the attenuated sentiment matrix , the sentiment forgetting gate in ASM-Net is defined as: in It represents the sigmoid neural network layer. represents the learnable weight matrix, represents the learnable bias matrix, Represents the original sentiment matrix and the decaying emotion matrix of the previous moment Perform splicing operations. It represents the output of the emotion forgetting gate; The emotional memory gate is defined as: in is the output of the emotional memory gate, It represents the tanh function The output after the operation is, represents the learnable weight matrix, represents the learnable bias matrix, It represents the sigmoid neural network layer, and the function tanh represents the mathematical hyperbolic tangent operation. Represents the original sentiment matrix and the decaying emotion matrix of the previous moment Perform splicing operations. Define the sentiment output gate as: in is the output of the emotion output gate, represents the learnable weight matrix, represents the learnable bias matrix, It represents the sigmoid neural network layer. Represents the original sentiment matrix and the decaying emotion matrix of the previous moment Perform splicing operations; The process of updating the network element status is defined as: in, Represents the emotional memory matrix attenuated at the current moment, is the intermediate result matrix; ASM-Net is defined as It represents the output of ASM-Net. It represents the output after the FNN network; S44: Attentional Mechanisms for Attenuating Emotional Perception Will attenuate the emotional matrix Q as an attention mechanism, embedding the item into the matrix As K and V, the user sentiment preference matrix obtained by the sentiment attention mechanism is defined as: ,in, , represents the learnable weight matrix, It represents matrix multiplication The new user sentiment preference matrix obtained after the FNN layer is: 。 4. The sequence recommendation method based on attenuated emotion perception according to claim 3 is characterized in that: The S5 further includes: Define the combined user behavior-sentiment preference matrix: in, , is the dimension size of the output; An LSTM network is used to integrate sentiment preferences and item preferences in long-term contexts, and layer normalization is used to stabilize and accelerate network training.

Images