Non-uniformly-spaced sampling time sequence recommendation method based on Horkes process

Abstract

The invention belongs to the field of e-commerce recommendation, and particularly relates to an unequally-spaced sampling time sequence recommendation method based on a Horkes process. The method comprises the steps of obtaining user commodity interaction data and obtaining a non-equal interval sampling time sequence according to the user commodity interaction data; filling missing data of the non-uniformly-spaced sampling time sequence by adopting a cyclic hox method to obtain a uniformly-spaced sampling time sequence; processing the equally-spaced sampling time sequence by adopting a multivariate hox method and a self-attention mechanism, and finally obtaining the score of the commodity by the user; recommending the K commodities with the highest scores to the user; according to the invention, the accuracy of the e-commerce recommendation system can be effectively improved, and an accurate personalized recommendation effect is realized.

Description

Technical Field

[0001] The present invention belongs to the field of e-commerce recommendation, and in particular relates to a non-equally spaced sampling time series recommendation method based on Hawkes process. Background Art

[0002] E-commerce recommendation system is an intelligent information filtering technology in the field of e-commerce. It predicts the products that users may be interested in and recommends them to users by analyzing the user's historical behavior, preferences and other relevant information. In recent years, due to its huge potential in academia and e-commerce, the recommendation system has been hailed as the intelligent engine of e-commerce. Its research results can be directly transformed into commercial applications, which has an important role in promoting the development of the e-commerce industry.

[0003] Obviously, the success of the recommendation system depends on user system modeling, such as time series extraction and user information mining. Although many academic researchers have made many contributions in time series, user time series are generally sampled at non-equal intervals. On the one hand, the interval sampling of non-equally spaced time series data is uneven, and on the other hand, there is a lack of dependency between feature engineering of non-equally spaced time series. Therefore, in order to ensure the accuracy of user modeling, we need to consider the importance of interest point mining and user dynamic interest extraction to solve the uneven interval sampling of non-equally spaced time series data and the lack of dependency between feature engineering.

[0004] At present, many scholars at home and abroad have conducted in-depth research on e-commerce recommendation models, and the research results are mainly reflected in the following aspects: On the one hand, with the development of big data technology, e-commerce recommendation models based on deep learning have gradually replaced traditional e-commerce recommendation models based on collaborative filtering and content. On the other hand, some deep recommendation models focus on feature compression and interaction. In addition, there are also models that focus on the extraction of user behavior sequence features. However, the current e-commerce recommendation models still face the following challenges:

[0005] 1. The non-uniform sampling characteristics of user behavior sequences. Traditional time series recommendation models ignore the impact of non-uniform sampling on user preferences. User behavior is random and data interval sampling is uneven. Non-uniform time intervals also contain important information, which is crucial to capturing user preferences.

[0006] 2. Missing dependencies between feature engineering. Since user time series are sampled at non-equal intervals, the sampling of user data occurs randomly, resulting in missing dependencies between feature engineering. This random and irregular change makes the relationship between user behavior and product interactions extremely complex, making it difficult to accurately capture the dynamic changes in user interests.

[0007] 3. Balance the recommendation of users’ coarse-grained and fine-grained preferences. Users’ preferences are multi-dimensional and multi-layered. Excessive focus on users’ fine-grained preferences may lead to filter bubbles and limit users’ opportunities to explore new things. How to adaptively utilize coarse-grained and fine-grained preferences to achieve personalized and diversified recommendations is a crucial issue. Summary of the invention

[0008] In view of the shortcomings of the prior art, the present invention proposes a non-equally spaced sampling time series recommendation method based on Hawkes process, which includes:

[0009] S1: Obtain user-product interaction data and obtain a non-equally spaced sampling time series based on the user-product interaction data;

[0010] S2: The missing data of the non-equally spaced sampling time series are filled using the cyclic Hawkes method to obtain an equally spaced sampling time series;

[0011] S3: Use the multivariate Hawkes method and self-attention mechanism to process the equally spaced sampling time series to obtain the user's rating of the product;

[0012] S4: Recommend the K products with the highest ratings to the user.

[0013] Preferably, the non-equally spaced sampling time series is expressed as:

[0014] S u ={(u,i 1 ,t 1 ),(u,i 2 ,t 2 ),...,(u,i n ,t n )}

[0015] Among them, S u represents the non-uniformly spaced sampling time series of user u, (u,i n ,t n ) indicates that user u is in t n Moments and Products n to interact.

[0016] Preferably, the process of filling missing data of non-equally spaced sampling time series using the cyclic Hawkes method includes:

[0017] S21: Calculate the basic generation rate of users based on the number of user interaction behaviors and active time;

[0018] S22: Calculate the cyclic Hawkes score of the user at the time when the time series is missing value according to the trigger core and the basic generation rate of the user;

[0019] S23: Insert the product corresponding to the maximum cyclic Hawkes score before the missing data moment into the vacant position to obtain an equally spaced sampling time series.

[0020] Furthermore, the formula for calculating the basic generation rate of users is:

[0021]

[0022] Among them, μ represents the basic generation rate of users, n represents the number of user interaction behaviors, and T represents the user's active time.

[0023] Furthermore, the formula for calculating the user's Cyclic Hawkes score is:

[0024]

[0025] Among them, λ(t) represents the occurrence rate of the user at time t, i.e., the cyclic Hawkes score, α represents the incentive intensity coefficient, and g(tt i ) represents the trigger kernel function, and N(s) represents the sum of events that have occurred before time t.

[0026] Preferably, obtaining the user's rating of the product includes:

[0027] S31: Use multivariate Hawkes process to process the equally spaced sampling time series to obtain the user's coarse-grained preference representation and fine-grained preference representation;

[0028] S32: Use the self-attention mechanism to calculate the coarse-grained preference weights and fine-grained preference weights;

[0029] S33: Calculate the multivariate Hawkes score based on the user's coarse-grained preference representation, fine-grained preference representation, coarse-grained preference weight, and fine-grained preference weight, and perform normalization processing to obtain the final user's score for the product.

[0030] Furthermore, the formulas for obtaining the user's coarse-grained preference representation and fine-grained preference representation are:

[0031]

[0032] Among them, μ u,i represents the basic preference of user u for product i, i.e. the user's coarse-grained preference representation; g u (tt h ) represents the interaction time t between product h and user u h The time decay function from the occurrence to the current time t; α h，i Indicates the degree of stimulation of product h on the current product i; is the fine-grained preference representation of user u, δ uRepresents the decay rate of the influence of historical events; H u，t represents the historical equally spaced sampling time series of user u before time t; v i represents the embedding representation of product i, v u represents the embedding representation of user u, v h Represents the embedding representation of product h.

[0033] Furthermore, the formula for calculating the coarse-grained preference weight and the fine-grained preference weight is:

[0034]

[0035] in, represents the coarse-grained preference weight of user u, represents the fine-grained preference weight of user u, a l 、a s represents the first and second hyperparameters, b l 、b s Indicates the first and second bias; v u represents the embedding representation of user u, represents the aggregated historical embedding, v h represents the embedding representation of product h, H u，t represents the historical equally spaced sampling time series of user u before time t.

[0036] Furthermore, the formula for calculating the user's multivariate Hawkes score based on the self-attention weight is:

[0037]

[0038] Among them, λ i|u (t) represents the preference intensity of user u for product i at time t, i.e., the multivariate Hawkes score. represents the coarse-grained preference weight of user u, represents the fine-grained preference weight of user u, μ u，i Indicates the basic preference of user u for product i, α h，i represents the initial incentive degree of historical commodity h to current commodity i, g u (tt h ) represents the interaction time t between product h and user u h The time decay function that occurs to the current time t, t h Represents the interaction moment between product h and user u.

[0039] The beneficial effects of the present invention are as follows: the present invention utilizes the non-equally spaced sampling time series data and target product data of users of the e-commerce platform, fills in the missing data of the non-equally spaced sampling time series of users and captures the dependencies between features, and then obtains the preference representation, and adaptively utilizes the coarse-grained and fine-grained preferences of users to predict the final preferences of users. The present invention can effectively improve the accuracy of the e-commerce recommendation system and achieve accurate personalized recommendation effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flowchart of a method for recommending non-equally spaced sampling time series based on Hawkes process in the present invention;

[0041] Figure 2 It is a framework diagram of the method for recommending non-equally spaced sampling time series based on Hawkes process in the present invention;

[0042] Figure 3 A schematic diagram of the cyclic Hawkes model constructed for the present invention to predict and fill in missing values; DETAILED DESCRIPTION

[0043] 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.

[0044] The present invention proposes a non-uniformly spaced sampling time series recommendation method based on Hawkes process, such as Figure 1 , Figure 2 As shown, the method includes the following contents:

[0045] S1: Obtain user-product interaction data and obtain a non-equally spaced sampling time series based on the user-product interaction data.

[0046] Obtain user-product interaction data, including user feature data, product feature data, user-product interaction behavior data, interaction time, etc. The user-product interaction set is represented as:

[0047] {R u，i，t =(u,i,t)|(u,i,t)∈(U,I,T)}

[0048] Where T is a time domain; U = {u 1 ,u 2 ,...,u} represents the user set, each user is represented by u i I={i 1 ,i 2,...,i} to represent the item set. Therefore, the user-item interaction is represented by a triple (u, i, t).

[0049] According to the user-item interaction triplet, the non-uniformly spaced sampling time series is represented as:

[0050] S u ={(u,i 1 , t 1 ), (u, i 2 , t 2 ),...,(u,i n , t n )}

[0051] Among them, S u represents the non-uniformly spaced sampling time series of user u, (u, i n , t n ) indicates that user u is in t n Moments and Products n to interact.

[0052] S2: The cyclic Hawkes method is used to fill the missing data of the non-equally spaced sampling time series.

[0053] like Figure 3 As shown in Figure 1, the process of using the cyclic Hawkes method to fill the missing data of non-equally spaced sampling time series includes:

[0054] S21: Calculate the basic generation rate of users based on the number of user interaction behaviors and active time.

[0055] The formula for calculating the base generation rate of users is:

[0056]

[0057] Among them, μ represents the basic generation rate of users, n represents the number of user interaction behaviors, and T represents the user's active time.

[0058] S22: Calculate the user's cyclic Hawkes score at different times according to the trigger core and the user's basic generation rate.

[0059] The trigger core is calculated from the decay amplitude:

[0060]

[0061] Among them, β is the decay rate. This parameter determines the decay rate of the incentive effect over time. The larger β is, the faster the incentive effect decays.

[0062] The formula for calculating a user's Cyclic Hawkes score is:

[0063]

[0064] Among them, λ(t) represents the occurrence rate of the user at time t, that is, the cyclic Hawkes score, α represents the incentive intensity coefficient, and the larger α is, the stronger the incentive effect of the past event is; g(tt i ) represents the trigger kernel function, N(s) is a counting process, which represents the sum of events that have occurred before time t.

[0065] S23: Insert the product corresponding to the maximum cyclic Hawkes score before the missing data moment into the vacant position to obtain a complete non-equally spaced sampling time series.

[0066] t m ,t q ,t n The cyclic Hawkes score is the multiple moments before the current data missing moment:

[0067] λ(t m )>λ(t q )>…>λ(t n )

[0068] The score λ(t m ) is the maximum value, then λ(t m ) corresponding to the product i m Inserted into the current data missing position, the user's interactive behavior on the product is determined by the final Hawkes score. For example: if the score is greater than 0.9, it is purchased; if the score is between 0.8-0.9, it is added to the shopping cart; if it is between 0.7-0.8, it is collected; if it is between 0.6-0.7, it is liked; if it is between 0.5-0.6, it is searched.

[0069] S3: The multivariate Hawkes method and self-attention mechanism are used to process the non-equally spaced sampling time series to obtain the user's rating of the product.

[0070] The process of obtaining user ratings of products includes:

[0071] S31: The multivariate Hawkes process is used to process the equally spaced sampling time series to obtain the user's coarse-grained preference representation and fine-grained preference representation.

[0072] The user's coarse-grained preference is expressed as:

[0073]

[0074] According to the decay rate δ of the influence of historical events u ≥0, calculate the influence of historical events on the user's Hawkes score, which is an exponential kernel function g u(.), the specific formula is as follows:

[0075]

[0076] According to the vector representation of the current product i and the vector representation of the historical product h, the degree of stimulation αh,i of product h on the current product i is calculated. The specific formula is as follows:

[0077]

[0078] Among them, μ u，i represents the basic preference of user u for product i, i.e. the user’s coarse-grained preference representation, g u (tt h ) represents the interaction time t between product h and user u h The time decay function from the occurrence to the current time t; α h，i Indicates the degree of stimulation of product h on the current product i; the sum of the two after multiplication is the user’s fine-grained preference representation; where v i represents the embedding representation of product i, v h represents the embedding representation of product h, v u represents the embedding representation of user u, H u，t represents a historical non-uniformly spaced sampling time series of user u before time t; δ u ≥0 indicates the decay rate of the influence of historical events. Since each user’s preference may decay at a different rate, this parameter is user-related.

[0079] S32: Use the self-attention mechanism to calculate the coarse-grained preference weights and fine-grained preference weights.

[0080] The present invention designs an adaptive attention mechanism structure to obtain preference weights, which are defined as follows:

[0081]

[0082] in, represents the coarse-grained preference weight of user u, represents the fine-grained preference weight of user u; a l ∈R d , a s ∈R d , b l ∈R,b s ∈R are all model parameters; a l 、a s represents the first and second hyperparameters, b l 、b sRepresents the first and second biases; relu represents the relu activation function; v u represents the embedding representation of user u; represents the aggregated historical embedding, which is obtained by the average pooling strategy:

[0083]

[0084] S33: Calculate the multivariate Hawkes score based on the user's coarse-grained preference representation, the fine-grained preference representation coarse-grained preference weight and the fine-grained preference weight, and perform normalization processing to obtain the final user's score for the product.

[0085] The multivariate Hawkes process can be summarized as the embedding of all historical interaction items, and the impact of historical interaction items on the prediction and recommendation of the target item is calculated. The formula for calculating the user's multivariate Hawkes score is:

[0086]

[0087] Among them, λ i|u (t) represents the preference intensity of user u for product i at time t, i.e., the multivariate Hawkes score. represents the coarse-grained preference weight of user u, represents the fine-grained preference weight of user u, μ u，i Indicates the basic preference of user u for product i, α h，i Indicates the degree of stimulation of product h on the current product i, g u (tt h ) represents the interaction time t between product h and user u h The time decay function that occurs to the current time t, t h represents the interaction time between product h and user u, H u，t Represents a historical equally spaced sampling time series of user u before time t.

[0088] The present invention uses an adaptive method of attention and integrates the coarse-grained preferences and fine-grained preferences of users to achieve accurate personalized recommendations for user sequences.

[0089] The user's multivariate Hawkes rating is normalized to obtain the user's rating of the product.

[0090] For each product i, the present invention defines the conditional distribution P on the entire product set I i|u (t):

[0091]

[0092] Among them, P i|u (t) represents the rating of user u on product i.

[0093] S4: Recommend the K products with the highest ratings to the user.

[0094] Sort the user's ratings of the products, select the K products with the highest ratings as the user's recommended products and recommend them to them.

[0095] In summary, the present invention predicts and fills the missing data points of the user's non-equally spaced sampling time series by using a model based on a cyclic Hawkes process; simulates the change of user preferences over time by using a model based on a multivariate Hawkes process; and then combines an adaptive attention mechanism to strengthen the focus on target items in the user's historical behavior, and adaptively utilizes the user's preferences to make the final recommendation item prediction. The present invention predicts user preferences by capturing the dependency between missing data filling and feature engineering in the user's non-equally spaced sampling time series. The present invention can effectively improve the accuracy of the e-commerce recommendation system and achieve accurate personalized recommendation effects.

[0096] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A non-equally spaced sampling time series recommendation method based on Hawkes process, characterized in that: include: S1: Obtain user-product interaction data and obtain a non-equally spaced sampling time series based on the user-product interaction data; S2: The missing data of the non-equally spaced sampling time series are filled using the cyclic Hawkes method to obtain an equally spaced sampling time series; S3: Use the multivariate Hawkes method and self-attention mechanism to process the equally spaced sampling time series to obtain the user's rating of the product; S4: Recommend the K products with the highest ratings to the user.

2. According to the method for recommending time series with non-equal interval sampling based on Hawkes process in claim 1, it is characterized in that: The non-equally spaced sampling time series is expressed as: WITH u ={(u,i1,t1),(u,i2,t2),...,(u,i n ,t n )} Among them, S u represents the non-uniformly spaced sampling time series of user u, (u,i n ,t n ) indicates that user u is in t n Moments and Products n to interact.

3. The method for recommending time series with non-equal interval sampling based on Hawkes process according to claim 1, characterized in that: The process of using the cyclic Hawkes method to fill missing data of non-equally spaced sampling time series includes: S21: Calculate the basic generation rate of users based on the number of user interaction behaviors and active time; S22: Calculate the cyclic Hawkes score of the user at the time when the time series is missing value according to the trigger core and the basic generation rate of the user; S23: Insert the product corresponding to the maximum cyclic Hawkes score before the missing data moment into the vacant position to obtain an equally spaced sampling time series.

4. The method for recommending time series with non-equally spaced sampling based on Hawkes process according to claim 3, characterized in that: The formula for calculating the base generation rate of users is: Among them, μ represents the basic generation rate of users, n represents the number of user interaction behaviors, and T represents the user's active time.

5. The method for recommending time series with non-equally spaced sampling based on Hawkes process according to claim 3, characterized in that: The formula for calculating a user's Cyclic Hawkes score is: Among them, λ(t) represents the occurrence rate of the user at time t, i.e., the cyclic Hawkes score, α represents the incentive intensity coefficient, and g(tt i ) represents the trigger kernel function, and N(s) represents the sum of events that have occurred before time t.

6. The method for recommending time series with non-equally spaced sampling based on Hawkes process according to claim 1, characterized in that: Get the user's rating of the product. The process includes: S31: Use multivariate Hawkes process to process the equally spaced sampling time series to obtain the user's coarse-grained preference representation and fine-grained preference representation; S32: Use the self-attention mechanism to calculate the coarse-grained preference weights and fine-grained preference weights; S33: Calculate the multivariate Hawkes score based on the user's coarse-grained preference representation, fine-grained preference representation, coarse-grained preference weight, and fine-grained preference weight, and perform normalization processing to obtain the final user's score for the product.

7. The method for recommending time series with non-equally spaced sampling based on Hawkes process according to claim 6, characterized in that: The formula for obtaining the user's coarse-grained preference representation and fine-grained preference representation is: Among them, μ u,i represents the basic preference of user u for product i, i.e. the user's coarse-grained preference representation; g u (tt h ) represents the interaction time t between product h and user u h The time decay function that occurs to the current time t; α h,i Indicates the degree of stimulation of product h on the current product i; is the fine-grained preference representation of user u, δ u Represents the decay rate of the influence of historical events; H u,t represents the historical equally spaced sampling time series of user u before time t; v i represents the embedding representation of product i, v u represents the embedding representation of user u, v h Represents the embedding representation of product h.

8. The method for recommending time series with non-equally spaced sampling based on Hawkes process according to claim 6, characterized in that: The formula for calculating coarse-grained preference weights and fine-grained preference weights is: in, represents the coarse-grained preference weight of user u, represents the fine-grained preference weight of user u, a l 、a s represents the first and second hyperparameters, b l 、b s Indicates the first and second bias; v u represents the embedding representation of user u, represents the aggregated historical embedding, v h represents the embedding representation of product h, H u,t represents the historical equally spaced sampling time series of user u before time t.

9. The method for recommending time series with non-equally spaced sampling based on Hawkes process according to claim 6, characterized in that: The formula for calculating the user's multivariate Hawkes score based on the self-attention weight is: Among them, λ i|u (t) represents the preference intensity of user u for product i at time t, i.e., the multivariate Hawkes score. represents the coarse-grained preference weight of user u, represents the fine-grained preference weight of user u, μ u,i Indicates the basic preference of user u for product i, α h,i represents the initial incentive degree of historical commodity h to current commodity i, g u (tt h ) represents the interaction time t between product h and user u h The time decay function that occurs to the current time t, t h Represents the interaction moment between product h and user u.

Images