Recommendation method for relieving interest mutation based on local interest modeling

Through local interest modeling and adaptive calibration methods, the problem that recommendation system is difficult to cope with changes in user interest is solved, and more accurate interest prediction and personalized recommendation are achieved.

CN118964740AActive Publication Date: 2024-11-15HUNAN UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202411145865.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-15
Estimated Expiration
2044-08-20

AI Technical Summary

Technical Problem

The existing recommendation system is difficult to accurately understand and predict changes in user interests and cannot effectively respond to dynamic changes in user interests, especially when users visit again for a long time, resulting in insufficient personalized recommendation results.

Method used

Using a method based on local interest modeling, the steps of multi-scale local interest modeling, adaptive calibration and candidate item recommendation include interactive subsequence segmentation, evolutionary modeling within and between subsequences, multi-scale local interest generation, Ideal interest generation and adaptive interest mutation calibration, combined with Multi-Layer Perceptron to calculate the probability of user interaction with items, and use attention mechanisms and comparative learning to optimize interest representations.

Benefits of technology

It improves the personalized effect of the recommendation system, can more accurately understand and predict changes in user interests, and provides more personalized recommendation results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118964740A_ABST
    Figure CN118964740A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of information processing, in particular to a recommendation method for relieving interest mutation based on local interest modeling, which comprises multi-scale local interest modeling, self-adaptive calibration and candidate item recommendation. According to the multi-scale local interest modeling, interaction subsequences are segmented, evolution in the subsequences and evolution between the subsequences are modeled, and multi-scale local interests are generated; performing self-adaptive calibration to generate an interest signal related to the article, predicting the sudden change degree of the interest of the user, and performing self-adaptive interest calibration; the interest locality and the interest mutation of the user are comprehensively considered, and the effect of the personalized recommendation system is improved; by combining local modeling and interest mutation modeling, the interest change of the user can be more accurately understood and predicted, and a more personalized recommendation service is provided for the user.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to a recommendation method based on local interest modeling to alleviate interest mutations. Background Art

[0002] Recommendation systems can help people quickly filter out the information they need in the context of information explosion, and have been widely used on various platforms. In order to build an effective recommendation system, it is necessary to build a preference model by analyzing the user's historical behavior.

[0003] Traditional recommendation methods, such as content-based recommendation and collaborative filtering algorithms, often extract static latent features based on all historical behaviors of users as representations of user preferences, and are unable to cope with dynamic changes in user preferences.

[0004] In addition, many current works have proposed deep sequential recommendation (SR) models, which aim to model the dynamic transformations in user-item interactions. However, most existing works are based on the assumption that interests evolve smoothly over time, and model the item transformations within a user's interaction sequence. However, this assumption ignores the locality of user interaction patterns and cannot solve the problem of interest mutations: when a user visits the recommendation system again after a long interval, his or her interests may have changed significantly compared to the past, but this change is difficult to obtain by analyzing historical behavior. Summary of the invention

[0005] The purpose of the present invention is to provide a recommendation method based on local interest modeling to alleviate interest mutations, so as to solve the problem that the recommendation system in the prior art is difficult to accurately understand and predict user interest changes and provide users with accurate and personalized recommendation results.

[0006] To achieve the above object, the present invention provides the following technical solution: a recommendation method for alleviating interest mutation based on local interest modeling, comprising the following steps:

[0007] S1, multi-scale local interest modeling;

[0008] Among them, multi-scale local interest modeling includes three steps: interactive subsequence segmentation, evolution modeling within and between subsequences, and multi-scale local interest generation;

[0009] S2, adaptive calibration;

[0010] Among them, adaptive calibration includes two steps: ideal interest generation and adaptive interest mutation calibration;

[0011] S3, candidate item recommendation;

[0012] In order to predict the interaction probability between the user and the target item, the Multi-Layer Perceptron (MLP) structure is used to calculate the interaction probability between the user u and the target item i at time t+1, as shown in Formula 14:

[0013]

[0014] Preferably, when segmenting the interaction subsequences in step S1, the original user interaction sequence is first hard segmented, and a hyperparameter step is introduced as the segmentation size to generate local sequences at multiple scales. The original sequence S is padded with step-(|S|%step) zeros at the end, as shown in Formula 1:

[0015] S=S+{0,…,0}

[0016] Then, the sequence S is divided into multiple local subsequences {Sub1, Sub2, ..., Sub n-1 ,Sub n},in, is the number of local subsequences obtained by segmentation.

[0017] Preferably, the interest evolution modeling within the subsequence in step S1 is as shown in Formula 2:

[0018]

[0019] Where P intra ∈R |S|×d ,and Corresponding to the local interest representation of each position in the i-th subsequence, φ(·) represents the evolution within the subsequence and can be obtained using common structures such as RNN and GRU. This time, the Time4LSTM proposed by Yu et al. is used.

[0020] Preferably, in step S1, for the evolution between subsequences, an evolutionary RNN is designed for modeling between subsequences to capture the changes in the local interest context and model the dynamic evolution of the interaction pattern. The local representation of the jth position in the i-th subsequence is defined as Where i∈[1,n] and j∈[1,step], we get the representation between subsequences, as shown in Formula 3:

[0021]

[0022] In the formula, represents the local representation obtained by selecting the i-th position from each subsequence fragment using Formula 2, and then using Time4LSTM to model the evolution of interest between subsequences, denoted as φ(·);

[0023] Finally, the interest representations within and between subsequences are merged to obtain the final local interest representation, as shown in Formula 4:

[0024] P loc =P inter +LayerNorm(P intra )

[0025] Where LayerNorm(·) is the layer normalization operation.

[0026] Preferably, when generating multi-scale local interests in step S1, a set of segmentation steps is defined, such as formula 5:

[0027] L={step1,step2,…step |L|}

[0028] For the segmentation step, a common exponential strategy is used, setting it to step i =2 i ;

[0029] For each step in L i , generate the corresponding local interest representation, as shown in Formula 6:

[0030]

[0031] In the formula, Represents the local interest modeling of the i-th layer, where the segmentation step is L i ;

[0032] Finally, the interests at different levels are combined to obtain the final interest representation, as shown in Formula 7:

[0033]

[0034] Preferably, when generating the ideal interest in step S2, first, the ideal interest representation of the item is generated by using the lookup embedding table, as shown in Formula 8:

[0035] P ideal =Emd(x i )

[0036] Generate ideal interest representations for positive and negative samples respectively, and use a sequence encoder to learn the overall interest representation As shown in formula 9:

[0037]

[0038]

[0039] In the formula, Φ(·) represents an interest learning function, which encodes the user's historical behavior sequence to obtain the overall interest representation. Then, in order to optimize the parameters of the ideal interest representation of the item, the interest representation is transformed into Compare with one positive sample and four negative samples, and use the BPR loss function to optimize the model. Consider it as an anchor point and minimize its difference with the positive sample representation The distance between them is maximized while maximizing the negative sample representation The distance between them, the loss function is calculated as shown in Formula 10:

[0040]

[0041] Where <·> represents the inner product of the vectors, and σ represents the softplus function.

[0042] Preferably, during the adaptive interest mutation calibration in step S2, first, the user's interest representation is obtained, and then the ideal interest signal is used for calibration, and the attention mechanism is used to adaptively fuse the short-term interest and the long-term interest, as shown in Formula 11:

[0043] α=sigmoid(MLP([x targ ,P loc ,P long ,T itv ])),

[0044]

[0045] Where α is a weight used to control the fusion of long-term interest and local interest, T itv represents the time interval between the last interaction and the test sample, P long represents long-term interest, which is the weighted sum of historical interaction items;

[0046] Finally, the ideal interest representation obtained by contrastive learning is combined with the overall interest to calibrate interest mutations. Since the mutation degree of different users and different subsequences may be different, an attention-based adaptive fusion method is used to determine the strength of adjustment. A parameter β is introduced as the calibration strength of interest mutation. This parameter is determined by comprehensively considering the overall interest, the ideal interest signal and the time interval. The specific calculation is shown in Formula 12:

[0047]

[0048] In the formula, ⊙ and + represent element-by-element multiplication and element-by-element addition, respectively. Then, the final calibrated interest representation is calculated as shown in Formula 13. The user interest is calibrated by mutation using attention-based adaptive fusion. Formula 13:

[0049]

[0050] Preferably, in step S3, a negative log-likelihood loss function is used to optimize the sequence recommendation task.

[0051] Formula 15 is:

[0052]

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] The recommendation method for alleviating interest mutation based on local interest modeling involved in the present invention comprehensively considers the user's interest locality and interest mutation, thereby facilitating improving the effect of the personalized recommendation system; by combining local interest modeling and interest mutation modeling, it is possible to more accurately understand and predict the user's interest changes, and provide users with more personalized recommendation services. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A flowchart of a recommended method for alleviating interest mutation based on interest locality modeling of the present invention;

[0056] Figure 2 A flowchart of the interactive subsequence segmentation steps of the present invention;

[0057] Figure 3 is a flow chart of the evolution steps within the subsequence of the present invention;

[0058] Figure 4 Flow chart of the steps of evolution between subsequences of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0060] See also Figure 1-4 The present invention provides a technical solution, a recommendation method for alleviating interest mutation based on local interest modeling, including multi-scale local interest modeling, adaptive calibration and candidate item recommendation. Among them, multi-scale local interest modeling divides the interaction subsequences, models the evolution within and between subsequences, and generates multi-scale local interests. Adaptive calibration generates item-related interest signals, predicts the degree of mutation of user interests, and performs adaptive interest calibration.

[0061] To model the locality of interest, the user's interaction sequence is segmented into appropriate subsequences and the interest evolution within each part and between two adjacent parts is captured. There are two possible ways to determine the appropriate subsequence segmentation: (1) flexible segmentation, which may lead to non-uniform subsequences; (2) fixed segmentation, which obtains uniform subsequences. Although the former may be closer to the actual situation, the latter is convenient for further processing (e.g., using advanced neural network models for interest locality modeling). In order to take full advantage of the advantages of both methods, a compromise solution is proposed: first, the sequence is segmented into subsequences with different fixed lengths, and then a multi-scale method is adopted to capture interest locality at different scales. In this way, the user's subsequences can have different lengths while taking advantage of advanced neural network models.

[0062] Therefore, three steps are used to model interest locality: interactive subsequence segmentation, intra-subsequence and inter-subsequence evolution modeling, and multi-scale local interest generation.

[0063] Step S1.1, interaction subsequence segmentation, first hard segmentation of the original user interaction sequence, considering that the local context length of different users may be personalized, we introduce a hyperparameter step as the segmentation size, which helps to generate local sequences at multiple scales. In addition, in order to ensure uniform segmentation to meet the input requirements of advanced neural network models, the original sequence S is padded with step-(|S|%step) zeros at the end, as shown in Formula 1:

[0064] S=S+{0,…,0}

[0065] Then, the sequence S is divided into multiple local subsequences {Sub1, Sub2, ..., Sub n-1 ,Sub n},in is the number of local subsequences obtained by segmentation.

[0066] Step S1.2, modeling the evolution within and between subsequences. The user's interest shows strong locality, and the degree of interest difference within the local context and between contexts (i.e., subsequences) is also different. Therefore, the interest evolution pattern within and between subsequences is modeled to explain the local context information in the interaction and help capture the dynamic changes of user interaction patterns.

[0067] The modeling process of interest evolution within a subsequence is shown in Formula 2:

[0068]

[0069] Where P intra ∈R |S|×d ,and Corresponding to the local interest representation of each position in the i-th subsequence, φ(·) represents the evolution within the subsequence and can be obtained using common structures such as RNN and GRU. This time, the Time4LSTM proposed by Yu et al. is used.

[0070] For the evolution between subsequences, an evolutionary RNN is designed for modeling between subsequences to capture the changes in local interest context and model the dynamic evolution of interaction patterns. For the sake of convenience, we define the local representation of the jth position in the i-th subsequence as Where i∈[1,n] and j∈[1,step]. Then we can get the representation between subsequences, as shown in Formula 3:

[0071]

[0072] In the formula, represents the local representation obtained by selecting the i-th position from each subsequence fragment using Formula 2. Then, Time4LSTM is used to model the evolution of interests between subsequences, denoted as φ(·).

[0073] Finally, the interest representations within and between subsequences are merged to obtain the final local interest representation, as shown in Formula 4:

[0074] P loc =P inter +LayerNorm(P intra ),

[0075] Where LayerNorm(·) is the layer normalization operation, which is used to improve the generalization ability and training stability of the model.

[0076] Step S1.3, multi-scale local interest generation, considering that the user's interaction mode may vary in different recommendation scenarios, in order to improve the generalization ability of the model, a multi-scale method is designed to generate local interests. In other words, subsequences of different lengths can be generated. First, we define a set of segmentation steps, as shown in Formula 5:

[0077] L={step1,step2,…step |L|}.

[0078] The split step size plays a crucial role in model performance and needs to be carefully selected according to different recommendation scenarios. A commonly used exponential strategy is adopted, which is set to step i =2 i .

[0079] Then, for each step in L i, generate the corresponding local interest representation, as shown in Formula 6:

[0080]

[0081] In the formula, Represents the local interest modeling of the i-th layer, and the segmentation step is L i .

[0082] Finally, the interests at different levels are combined to obtain the final interest representation, as shown in Formula 77:

[0083]

[0084] In order to calibrate mutations, some additional information needs to be introduced. Since similar interests are likely to drive users to have similar interactive behaviors, a contrastive learning method is used to learn the ideal interest distribution of items as additional information for mutation calibration.

[0085] The adaptive calibration module consists of two steps: ideal interest generation and adaptive interest mutation calibration.

[0086] S2.1, ideal interest generation;

[0087] First, the ideal interest representation of the item is generated using the lookup embedding table, as shown in Formula 8:

[0088] P ideal =Emd(x i ).

[0089] Generate ideal interest representations for positive and negative samples respectively, and use a sequence encoder to learn the overall interest representation As shown in Formula 9:

[0090]

[0091]

[0092] In the formula, Φ(·) represents an interest learning function, which encodes the user's historical behavior sequence to obtain the overall interest representation. Then, in order to optimize the parameters of the ideal interest representation of the item, the interest representation Compare and learn with one positive sample and four negative samples. Use the commonly used BPR loss function to optimize the model. Consider it as an anchor point and minimize its difference with the positive sample representation The distance between them is maximized while maximizing the negative sample representation The loss function is calculated as shown in formula 10:

[0093]

[0094] Where <·> represents the inner product of the vectors, and σ represents the softplus function.

[0095] S2.2, Adaptive calibration of mutations of interest;

[0096] First, the user's interest representation is obtained and then calibrated with the ideal interest signal. Because the user's long-term and short-term interests can enhance the robustness of interest modeling. Therefore, the user's long-term interests and multi-scale short-term local interests are considered at the same time. The attention mechanism is used to adaptively fuse these two parts, as shown in Formula 11:

[0097] α=sigmoid(MLP([x targ ,P loc ,P long ,T itv ])).

[0098]

[0099] Where α is a weight used to control the fusion of long-term interest and local interest, T itv represents the time interval between the last interaction and the test sample, P long represents the long-term interest, which is the weighted sum of historical interaction items.

[0100] Finally, the ideal interest representation obtained by contrastive learning is combined with the overall interest to calibrate interest mutations. Since the degree of mutation may be different for different users and different subsequences, an attention-based adaptive fusion method is used to determine the strength of adjustment. A parameter β is introduced as the calibration strength of interest mutation. This parameter is determined by comprehensively considering the overall interest, the ideal interest signal, and the time interval. The specific calculation is shown in Formula 12:

[0101]

[0102] In the formula, ⊙ and + represent element-by-element multiplication and element-by-element addition, respectively. Then, the final calibrated interest representation is calculated as shown in Formula 13, and the user interest is calibrated by attention-based adaptive fusion.

[0103]

[0104] In order to predict the interaction probability between the user and the target item, the widely used Multi-Layer Perceptron (MLP) structure is used to calculate the interaction probability between user u and target item i at time t+1. The specific calculation is shown in Formula 14:

[0105]

[0106] The negative log-likelihood loss function is used to optimize the sequence recommendation task, as shown in Formula 15:

[0107]

[0108] Ultimately, targeted item recommendations are provided to users.

[0109] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0110] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A recommendation method based on local interest modeling to alleviate interest mutations, characterized in that: The following steps are involved: S1, multi-scale local interest modeling; Among them, multi-scale local interest modeling includes three steps: interactive subsequence segmentation, evolution modeling within and between subsequences, and multi-scale local interest generation; S2, adaptive calibration; Among them, adaptive calibration includes two steps: ideal interest generation and adaptive interest mutation calibration; S3, candidate item recommendation; In order to predict the interaction probability between the user and the target item, the Multi-Layer Perceptron (MLP) structure is used to calculate the interaction probability between the user u and the target item i at time t+1, as shown in Formula 14:

2. The recommendation method based on local interest modeling to alleviate interest mutation according to claim 1 is characterized by: When segmenting the interaction subsequences in step S1, the original user interaction sequence is first hard segmented, and a hyperparameter step is introduced as the segmentation size to generate local sequences at multiple scales. The original sequence S is padded with step-(|S|%step) zeros at the end, as shown in Formula 1: S=S+{0,…,0} Then, the sequence S is divided into multiple local subsequences {Sub1, Sub2, ..., Sub n-1 ,Sub n },in, is the number of local subsequences obtained by segmentation.

3. The recommendation method based on local interest modeling to alleviate interest mutation according to claim 1 is characterized by: The interest evolution modeling within the subsequence in step S1 is as shown in Formula 2: Where P intra ∈R |S|×d ,and Corresponding to the local interest representation of each position in the i-th subsequence, φ(·) represents the evolution within the subsequence and can be obtained using common structures such as RNN and GRU. This time, the Time4LSTM proposed by Yu et al. is used.

4. The recommendation method based on local interest modeling to alleviate interest mutation according to claim 3 is characterized by: In step S1, for the evolution between subsequences, an evolutionary RNN is designed for modeling between subsequences to capture the changes in local interest contexts and model the dynamic evolution of interaction patterns. The local representation of the jth position in the i-th subsequence is defined as Where i∈[1,n] and j∈[1,step], we get the representation between subsequences, as shown in Formula 3: In the formula, represents the local representation obtained by selecting the i-th position from each subsequence fragment using Formula 2, and then using Time4LSTM to model the evolution of interest between subsequences, denoted as φ(·); Finally, the interest representations within and between subsequences are merged to obtain the final local interest representation, as shown in Formula 4: P loc =P inter +LayerNorm(P intra ) Where LayerNorm(·) is the layer normalization operation.

5. The recommendation method based on local interest modeling to alleviate interest mutation according to claim 1 is characterized by: When generating multi-scale local interests in step S1, a set of segmentation steps is defined, as shown in Formula 5: L={step1,step2,…step |L| } For the segmentation step, a common exponential strategy is used, setting it to step i =2 i ; For each step in L i , generate the corresponding local interest representation, as shown in Formula 6: In the formula, Represents the local interest modeling of the i-th layer, where the segmentation step is L i ; Finally, the interests at different levels are combined to obtain the final interest representation, as shown in Formula 7:

6. The recommendation method based on local interest modeling to alleviate interest mutation according to claim 1 is characterized by: When generating the ideal interest in step S2, first, the ideal interest representation of the item is generated using the lookup embedding table, as shown in Formula 8: P ideal =Emd(x i ) Generate ideal interest representations for positive and negative samples respectively, and use a sequence encoder to learn the overall interest representation As shown in formula 9: In the formula, Φ(·) represents an interest learning function, which encodes the user's historical behavior sequence to obtain the overall interest representation. Then, in order to optimize the parameters of the ideal interest representation of the item, the interest representation is transformed into Compare and learn with one positive sample and four negative samples, and use BPR loss function to optimize the model. Consider it as an anchor point and minimize its difference with the positive sample representation The distance between them is maximized while representing the negative sample The distance between them, the loss function is calculated as shown in Formula 10: Where <·> represents the inner product of the vectors, and σ represents the softplus function.

7. The recommendation method for alleviating interest mutation based on local interest modeling according to claim 1 is characterized in that: When calibrating the adaptive interest mutation in step S2, first, the user's interest representation is obtained, and then the ideal interest signal is used for calibration, and the attention mechanism is used to adaptively integrate the short-term interest and the long-term interest, as shown in Formula 11: α=sigmoid(MLP([x targ ,P loc ,P long ,T itv ])), Where α is a weight used to control the fusion of long-term interest and local interest, T itv represents the time interval between the last interaction and the test sample, P long represents long-term interest, which is the weighted sum of historical interaction items; Finally, the ideal interest representation obtained by contrastive learning is combined with the overall interest to calibrate interest mutations. Since the mutation degree of different users and different subsequences may be different, an attention-based adaptive fusion method is used to determine the strength of adjustment. A parameter β is introduced as the calibration strength of interest mutation. This parameter is determined by comprehensively considering the overall interest, the ideal interest signal and the time interval. The specific calculation is shown in Formula 12: In the formula, ⊙ and + represent element-by-element multiplication and element-by-element addition, respectively. Then, the final calibrated interest representation is calculated as shown in Formula 13. The user interest is calibrated by mutation using attention-based adaptive fusion. Formula 13:

8. The recommendation method for alleviating interest mutation based on local interest modeling according to claim 1 is characterized in that: In step S3, a negative log-likelihood loss function is used to optimize the sequence recommendation task. Formula 15 is:

Citation Information

Patent Citations

  • Sequence recommendation method based on long-term and short-term preferences of user

    CN114969533A

  • User dynamic interest recommendation method and system based on recurrent neural network

    CN117312668A

  • De-polarization sequence recommendation method and system based on long-term and short-term interest separation

    CN117951382A