Recommendation method for relieving interest mutation based on local interest modeling
By employing multi-scale local interest modeling and an adaptive calibration mechanism, the problem of sudden changes in user interests in recommendation systems is solved, resulting in more accurate personalized recommendations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUNAN UNIV
- Filing Date
- 2024-08-20
- Publication Date
- 2026-05-29
Smart Images
Figure CN118964740B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of information processing technology, specifically a recommended method for mitigating interest mutations based on local interest modeling. Background Technology
[0002] Recommender systems help people quickly filter out the information they need in the context of information overload and have been widely used on various platforms. To build an effective recommender system, it is necessary to construct a preference model by analyzing users' historical behavior.
[0003] Traditional recommendation methods, such as content-based recommendation and collaborative filtering algorithms, often extract static latent features as user preference representations based on all of a user's historical behavior, which cannot cope with dynamic changes in user preferences.
[0004] Furthermore, many current works propose deep sequence recommendation (SR) models, aiming to model the dynamic transitions in user-item interactions. However, most existing works are based on the assumption that interests evolve smoothly over time, modeling item transitions within a user's interaction sequence. But this assumption ignores the locality of user interaction patterns and cannot adequately address the problem of abrupt interest changes: when a user revisits the recommendation system after a long interval, their interests may have changed significantly compared to the past, but these changes are difficult to obtain by analyzing historical behavior. Summary of the Invention
[0005] The purpose of this invention is to provide a recommendation method based on local interest modeling to mitigate interest mutations, thereby addressing the problem in existing recommendation systems that struggle to accurately understand and predict changes in user interests and provide accurate, personalized recommendation results.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a recommended method for mitigating interest mutations based on local interest modeling, comprising the following steps:
[0007] S1, Multi-scale local interest modeling;
[0008] Multi-scale local interest modeling includes three steps: interactive subsequence segmentation, evolutionary modeling within and between subsequences, and multi-scale local interest generation.
[0009] S2, Adaptive calibration;
[0010] Adaptive calibration includes two steps: ideal interest generation and adaptive interest mutation calibration.
[0011] S3, Recommended Candidate Items;
[0012] To predict the interaction probability between a user and a target item, a Multi-Layer Perceptron (MLP) structure is used to calculate the interaction probability between user u and target item i at time t+1, as shown in Formula 14:
[0013]
[0014] Preferably, in step S1, during the segmentation of the interaction subsequence, 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, sequence S is divided into multiple local subsequences {Sub1, Sub2, ..., Sub...} n-1 Sub n},in, It is the number of local subsequences obtained through segmentation.
[0017] Preferably, the interest evolution model within the subsequence in step S1 is as shown in Formula 2:
[0018]
[0019] In the formula, P intra ∈R |S|×d ,and The local interest representation corresponding to each position in the i-th subsequence, φ(·) represents the evolution within the subsequence, which can be obtained using common structures such as RNN and GRU. In this study, we use the Time4LSTM proposed by Yue et al.
[0020] Preferably, in step S1, for the evolution between subsequences, an evolutionary RNN is designed to model the evolution between subsequences, in order to capture changes in the local interest context and model the dynamic evolution of the interaction pattern. The local representation of the j-th position in the i-th subsequence is defined as... Where i∈[1,n] and j∈[1,step], the representation between subsequences is obtained, as shown in Formula 3:
[0021]
[0022] In the formula, The local representation obtained by selecting the i-th position from each subsequence segment using Equation 2 is then used to model the evolution of interest between subsequences using Time4LSTM, denoted as φ(·).
[0023] Finally, the interest representations within and between subsequences are merged to obtain the final local interest representation, as shown in Equation 4:
[0024] P loc =P inter +LayerNorm(P intra )
[0025] In the formula, LayerNorm(·) is the layer normalization operation.
[0026] Preferably, in step S1, when generating multi-scale local interests, a set of segmentation step sizes is defined, as shown in Formula 5:
[0027] L = {step1, step2, ... step} |L|}
[0028] For the segmentation step size, a common exponential strategy is adopted, setting it as step. i =2 i ;
[0029] For each step size in L i This generates the corresponding local interest representation, as shown in Formula 6:
[0030]
[0031] In the formula, This represents the locality of interest modeling at the i-th layer, where the segmentation step size is L. i ;
[0032] Finally, by combining the different levels of interest, we obtain the final interest representation, as shown in Formula 7:
[0033]
[0034] Preferably, in step S2, when generating the ideal interest, firstly, the ideal interest representation of the item is generated using a lookup embedding table, as shown in Formula 8:
[0035] P ideal =Emd(x i )
[0036] Ideal interest representations are generated for both positive and negative samples, and a sequence encoder is used to learn the overall interest representation. As in formula 9:
[0037]
[0038]
[0039] In the formula, Φ(·) represents an interest learning function that encodes the user's historical behavior sequence to obtain an overall interest representation. Subsequently, to optimize the parameters of the ideal interest representation for items, the interest representation is... The model is compared with one positive sample and four negative samples, and the BPR loss function is used to optimize the model. Treat it as an anchor point, minimize its relationship with the positive sample representation. The distance between them, while maximizing the negative sample representation. The distance between them, and the calculation of the loss function are as shown in Formula 10:
[0040]
[0041] In the formula, <·> represents the inner product of vectors, and σ represents the softplus function.
[0042] Preferably, in step S2, during adaptive interest mutation calibration, firstly, the user's interest representation is obtained, then calibration is performed using the ideal interest signal, and an attention mechanism is used to adaptively fuse the short-term and long-term interests, as shown in Equation 11:
[0043] α = sigmoid(MLP([x targ ,P loc ,P long ,T itv ])),
[0044]
[0045] In the formula, α is a weight used to control the fusion of long-term and local interests, and T itv P represents the time interval between the last interaction and the test sample. long It represents long-term interest and is a weighted sum of historically interacted items;
[0046] Finally, the ideal interest representation obtained from contrastive learning is combined with the overall interest to calibrate interest mutations. Since the degree of mutation may differ among different users and different subsequences, an attention-based adaptive fusion method is used to determine the adjustment strength. A parameter β is introduced as the calibration strength of interest mutations. This parameter is determined by comprehensively considering the overall interest, the ideal interest signal, and the time interval, and is specifically calculated as shown in Equation 12:
[0047]
[0048] In the formula, ⊙ and + represent element-wise multiplication and element-wise addition, respectively. Then, the final calibrated interest representation is calculated as shown in Formula 13. An attention-based adaptive fusion is used to perform abrupt calibration of user interests. Formula 13:
[0049]
[0050] Preferably, in step S3, a negative log-likelihood loss function is used to optimize the sequence recommendation task.
[0051] Equation 15 is:
[0052]
[0053] Compared with the prior art, the beneficial effects of the present invention are:
[0054] The recommendation method based on local interest modeling that relates to this invention comprehensively considers the locality and mutation of user interests, thereby improving the effectiveness of personalized recommendation systems. By combining locality modeling and interest mutation modeling, it is possible to more accurately understand and predict changes in user interests, and provide users with more personalized recommendation services. Attached Figure Description
[0055] Figure 1 This is a flowchart of the recommended method for mitigating interest mutations based on interest locality modeling according to the present invention;
[0056] Figure 2 This is a flowchart of the interactive subsequence segmentation step of the present invention;
[0057] Figure 3 This is a flowchart of the evolution steps within a subsequence in this invention;
[0058] Figure 4 This is a flowchart of the evolution steps between subsequences in this invention. Detailed Implementation
[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0060] Please see Figure 1-4 This invention provides a technical solution for a recommendation method that mitigates interest mutations based on local interest modeling, comprising multi-scale local interest modeling, adaptive calibration, and candidate item recommendation. Multi-scale local interest modeling segments interactive 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 abrupt changes in 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 evolution of interest within each segment and between adjacent segments is captured. There are two possible approaches to determining the appropriate subsequence segmentation: (1) flexible segmentation, which may result in non-uniform subsequences; and (2) fixed segmentation, which yields uniform subsequences. While the former may be closer to reality, the latter facilitates further processing (e.g., using advanced neural network models for interest locality modeling). To fully leverage the advantages of both approaches, a compromise solution is proposed: first, the sequence is segmented into subsequences of different fixed lengths, and then a multi-scale approach is employed to capture interest locality at different scales. This allows the user's subsequences to have different lengths while utilizing advanced neural network models.
[0062] Therefore, interest locality modeling is performed in three steps: interactive subsequence segmentation, evolutionary modeling within and between subsequences, and multi-scale local interest generation.
[0063] Step S1.1: Interaction Subsequence Segmentation. First, the original user interaction sequence is hard-segmented. Considering that the local context length may be personalized for different users, we introduce a hyperparameter `step` as the segmentation size, which helps generate local sequences at multiple scales. Furthermore, 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 Equation 1:
[0064] S = S + {0,…,0}
[0065] Then, sequence S is divided into multiple local subsequences {Sub1, Sub2, ..., Sub...} n-1 Sub n},in It is the number of local subsequences obtained through segmentation.
[0066] Step S1.2: Evolutionary modeling within and between subsequences. User interests exhibit strong locality, and the degree of difference in interests within and between local contexts (i.e., subsequences) also differs. Therefore, modeling the evolutionary patterns of interests within and between subsequences is necessary to explain local contextual information in interactions and to help capture the dynamic changes in user interaction patterns.
[0067] The modeling process of interest evolution within a subsequence is shown in Equation 2:
[0068]
[0069] In the formula, P intra ∈R |S|×d ,and The local interest representation corresponding to each position in the i-th subsequence, φ(·), represents the evolution within the subsequence. It can be obtained using common structures such as RNN and GRU. In this study, Time4LSTM proposed by Yu et al. is used.
[0070] For the evolution between subsequences, an evolutionary RNN is designed to model the changes in local interest context and model the dynamic evolution of interaction patterns. For ease of explanation, we define the local representation of the j-th position in the i-th subsequence as... Where i∈[1,n] and j∈[1,step]. Then we can obtain the representation between subsequences, as shown in Equation 3:
[0071]
[0072] In the formula, This represents the local representation obtained by selecting the i-th position from each subsequence segment using Equation 2. Then, Time4LSTM is used to model the evolution of interest among 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 Equation 4:
[0074] P loc =P inter +LayerNorm(P intra ),
[0075] In the formula, LayerNorm(·) is a 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 user interaction patterns may change in different recommendation scenarios, a multi-scale method is designed to generate local interests to improve the model's generalization ability. That is, it can generate subsequences of different lengths. First, we define a set of segmentation step sizes, as shown in Equation 5:
[0077] L = {step1, step2, ... step} |L|}
[0078] The segmentation step size plays a crucial role in model performance and needs to be carefully selected based on different recommendation scenarios. A common exponential strategy is adopted, setting it as the step size. i =2 i .
[0079] Then, for each step in L iThis generates the corresponding local interest representation, as shown in Formula 6:
[0080]
[0081] In the formula, This represents the locality of interest modeling at the i-th layer, with a segmentation step size of L. i .
[0082] Finally, the different levels of interest are combined to obtain the final interest representation, as shown in Formula 7:
[0083]
[0084] To calibrate mutations, additional information needs to be introduced. Since similar interests are likely to drive users to produce similar interactive behaviors, a contrastive learning approach is used to learn the ideal interest distribution of items, providing 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, Generation of Ideal Interests;
[0087] First, the ideal interest representation of the item is generated using a lookup embedding table, as shown in Formula 8:
[0088] P ideal =Emd(x i ).
[0089] Ideal interest representations are generated for both positive and negative samples, and a sequence encoder is used to learn the overall interest representation. As shown in Formula 9:
[0090]
[0091]
[0092] In the formula, Φ(·) represents an interest learning function that encodes the user's historical behavior sequence to obtain an overall interest representation. Subsequently, to optimize the parameters of the ideal interest representation for an item, the interest representation is... The model is compared with one positive sample and four negative samples for learning. The commonly used BPR loss function is used to optimize the model. Treat it as an anchor point, minimize its relationship with the positive sample representation. The distance between them, while maximizing the negative sample representation. The distance between them. The loss function is calculated as shown in Formula 10:
[0093]
[0094] In the formula, <·> represents the inner product of vectors, and σ represents the softplus function.
[0095] S2.2 Adaptive interest mutation calibration;
[0096] First, the user's interest representation is obtained, and then calibrated using an ideal interest signal. Since both long-term and short-term user interests can enhance the robustness of interest modeling, both long-term and multi-scale short-term local interests are considered simultaneously. An attention mechanism is used to adaptively fuse these two parts, as shown in Equation 11:
[0097] α = sigmoid(MLP([x targ ,P loc ,P long ,T itv ])).
[0098]
[0099] In the formula, α is a weight used to control the fusion of long-term and local interests, and T itv P represents the time interval between the last interaction and the test sample. long It represents long-term interest and is a weighted sum of historically interacted items.
[0100] Finally, the ideal interest representation obtained through contrastive learning is combined with the overall interest to calibrate interest mutations. Since the degree of mutation may differ among different users and subsequences, an attention-based adaptive fusion method is used to determine the adjustment strength. A parameter β is introduced as the calibration strength for interest mutations. This parameter is determined by comprehensively considering the overall interest, the ideal interest signal, and the time interval, and its specific calculation is shown in Equation 12.
[0101]
[0102] In the formula, ⊙ and + represent element-wise multiplication and element-wise addition, respectively. Then, the final calibrated interest representation is calculated as shown in Equation 13, using attention-based adaptive fusion to perform abrupt calibration of user interests.
[0103]
[0104] To predict the probability of interaction between a user and a target item, a widely used multi-layer perceptron (MLP) structure is employed to calculate the interaction probability between user u and target item i at time t+1, as shown in Equation 14.
[0105]
[0106] The negative log-likelihood loss function is used to optimize this sequence recommendation task, as shown in Equation 15:
[0107]
[0108] Ultimately, this will enable users to receive targeted item recommendations.
[0109] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0110] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A recommendation method for mitigating interest mutations based on local interest modeling, characterized in that, Includes the following steps: S1, Multi-scale local interest modeling; Multi-scale local interest modeling includes three steps: interactive subsequence segmentation, evolutionary modeling within and between subsequences, and multi-scale local interest generation. S2, Adaptive calibration; Adaptive calibration includes two steps: ideal interest generation and adaptive interest mutation calibration. S3, Recommended Candidate Items; To predict the probability of user interaction with target items, a Multi-Layer Perceptron (MLP) structure is used to calculate the user's... exist Time and target item The interaction probability between them is shown in Formula 14: ， in, This refers to the softplus function. In step S1, during the multi-scale local interest generation, a set of segmentation step sizes is defined, as shown in Formula 5: For the segmentation step size, a common exponential strategy is adopted, and it is set to... ; for Each step size This generates the corresponding local interest representation, as shown in Formula 6: In the formula, Indicates the first Interest locality modeling of layers, where the segmentation step size is ; Finally, by combining the different levels of interest, we obtain the final interest representation, as shown in Formula 7: ， In step S2, during adaptive interest mutation calibration, firstly, the user's interest representation is obtained, and then calibration is performed using the ideal interest signal. An attention mechanism is used to adaptively fuse the short-term and long-term interests, as shown in Equation 11: ， In the formula, It is a weight used to control the fusion of long-term interests and local interests. This indicates the time interval between the last interaction and the test sample. It represents long-term interest and is a weighted sum of historically interacted items; Finally, the ideal interest representation obtained through contrastive learning is combined with the overall interest to calibrate interest mutations. Since the degree of mutation may differ among different users and subsequences, an attention-based adaptive fusion method is used to determine the strength of the adjustment, introducing a parameter. As the calibration strength for interest mutations, this parameter is determined by comprehensively considering the overall interest, the ideal interest signal, and the time interval, and is specifically calculated as shown in Formula 12: In the formula, and These represent element-wise multiplication and element-wise addition, respectively. The final calibrated interest representation is calculated as shown in Equation 13. An attention-based adaptive fusion is used to perform abrupt calibration of user interests. Equation 13: 。 2. The recommendation method for mitigating interest mutations based on local interest modeling according to claim 1, characterized in that: In step S1, during the segmentation of the interaction subsequence, the original user interaction sequence is first hard-segmented, and a hyperparameter is introduced. As the size of the segmentation, to generate local sequences at multiple scales, the original sequence Padding at the end A zero, as in Formula 1: Then, the sequence Divided into multiple local subsequences ,in, It is the number of local subsequences obtained through segmentation.
3. The recommendation method for mitigating interest mutations based on local interest modeling according to claim 1, characterized in that: The interest evolution model within the subsequence in step S1 is shown in Equation 2: In the formula, ,and Corresponding to the Local interest representation at each position in each subsequence The evolution within a subsequence can be represented using common structures such as RNN and GRU. In this study, we use the Time4LSTM proposed by Yue et al.
4. The recommendation method for mitigating interest mutations based on local interest modeling according to claim 3, characterized in that: In step S1, for the evolution between subsequences, an evolutionary RNN is designed to model the evolution between subsequences, capture changes in the local interest context, and model the dynamic evolution of the interaction pattern. The th subsequence The local representation of each position is defined as follows: ,in as well as This yields the representation between subsequences, as shown in Formula 3: In the formula, This represents selecting the first segment from each subsequence using Formula 2. The local representations obtained at each position are then used to model the evolution of interest among subsequences using Time4LSTM, represented as follows: ; Finally, the interest representations within and between subsequences are merged to obtain the final local interest representation, as shown in Equation 4: In the formula, It is a layer normalization operation.
5. The recommendation method for mitigating interest mutations based on local interest modeling according to claim 1, characterized in that: In step S2, when generating the ideal interest, firstly, the ideal interest representation of the item is generated using a lookup embedding table, as shown in Formula 8: Ideal interest representations are generated for both positive and negative samples, and a sequence encoder is used to learn the overall interest representation. For example, in formula 9: In the formula, This represents an interest learning function that encodes a user's historical behavior sequence to obtain an overall interest representation. Subsequently, to optimize the parameters of the ideal interest representation for items, the interest representation is... The model is compared with one positive sample and four negative samples, and the BPR loss function is used to optimize the model. Treat it as an anchor point, minimize its relationship with the positive sample representation. The distance between them, while maximizing the negative sample representation. The distance between them, and the calculation of the loss function are as shown in Formula 10: In the formula, Represents the dot product of vectors.
6. The recommendation method for mitigating interest mutations based on local interest modeling according to claim 1, characterized in that, In step S3, the negative log-likelihood loss function is used to optimize the sequence recommendation task, as shown in Equation 15: 。