A non-equidistant sampling time series recommendation method based on the hawkes process

Abstract

The application belongs to the field of e-commerce recommendation, and particularly relates to a non-equidistant sampling time sequence recommendation method based on a Hock process; the method comprises the following steps: obtaining user commodity interaction data and obtaining a non-equidistant sampling time sequence according to the user commodity interaction data; adopting a cyclic Hock method to fill in missing data of the non-equidistant sampling time sequence to obtain an equidistant sampling time sequence; adopting a multivariate Hock method and a self-attention mechanism to process the equidistant sampling time sequence, and finally obtaining a score of a user to a commodity; recommending K commodities with the highest scores to the user; the application can effectively improve the accuracy of an e-commerce recommendation system and realize a precise personalized recommendation effect.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of e-commerce recommendation, and particularly relates to a non-equidistant sampling time series recommendation method based on a Hawkes process. BACKGROUND

[0002] The e-commerce recommendation system is an intelligent information filtering technology in the field of e-commerce, which predicts the goods that the user may be interested in by analyzing the historical behavior, preferences and other related information of the user, and recommends these goods to the user. In recent years, due to the great potential in the academic and e-commerce circles, the recommendation system is hailed as the intelligent engine of e-commerce, and the research results can be directly converted into commercial applications, which plays an important role in promoting the development of the e-commerce industry.

[0003] Obviously, the success of the recommendation system depends on the user system modeling, such as the extraction of time series and the mining of user information. Although many academic researchers have made many contributions in the field of time series, the user time series is generally non-equidistantly sampled. On the one hand, the interval sampling of non-equidistantly sampled time series data is not uniform, and on the other hand, there is a missing dependency between the feature engineering of non-equidistantly sampled 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 non-uniformity of data interval sampling and the missing dependency between feature engineering of non-equidistantly sampled time series.

[0004] At present, many domestic and foreign scholars have carried out in-depth research on e-commerce recommendation models, and the research results mainly reflect in the following aspects: on the one hand, with the development of big data technology, the e-commerce recommendation model based on deep learning gradually replaces the traditional e-commerce recommendation model based on collaborative filtering and content-based. 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 model still has the following challenges:

[0005] 1. Non-equidistant sampling characteristics of user behavior sequence. Traditional time series recommendation models ignore the influence of non-equidistant sampling on user preferences. User behavior is random and data interval sampling is not uniform. Important information is also contained in the non-uniform time interval, which is crucial for capturing user preferences.

[0006] 2. Missing dependency between feature engineering. Since the user time series is non-equidistantly sampled, the sampling of user data occurs randomly, resulting in a missing dependency between feature engineering. This random irregular change makes the user behavior and the relationship between goods extremely complex, making it difficult to accurately capture the dynamic interest changes of users.

[0007] 3. Recommendation balance of user's coarse-grained and fine-grained preferences. The preferences of the user are multi-dimensional and multi-level, and excessive focus on the fine-grained preferences of the user may lead to a filtering bubble phenomenon, limiting the opportunity for the user to explore new things. How to adaptively utilize coarse-grained and fine-grained preferences to achieve personalized and diversified recommendations is a crucial problem. SUMMARY

[0008] In view of the deficiencies of the prior art, the present application proposes a non-equidistant sampling time series recommendation method based on the Hawkes process, which comprises:

[0009] S1: Obtain user commodity interaction data and obtain non-equidistant sampling time series according to the user commodity interaction data;

[0010] S2: Fill in the missing data of the non-equidistant sampling time series by using the cyclic Hawkes method to obtain the equidistant sampling time series;

[0011] S3: Process the equidistant sampling time series by using the multivariate Hawkes method and the self-attention mechanism to obtain the rating of the user to the commodity;

[0012] S4: Recommend the K commodities with the highest rating to the user.

[0013] Preferably, the non-equidistant sampling time series is represented as:

[0014] S u ={(u,i1,t1),(u,i2,t2),...,(u,in,tin)} n n

[0015] Wherein, S u represents the non-equidistant sampling time series of the user u, (u,i n ,t n ) represents the interaction of the user u with the commodity i n at time t n .

[0016] Preferably, the process of filling in the missing data of the non-equidistant sampling time series by using the cyclic Hawkes method comprises:

[0017] S21: Calculate the basic generation rate of the user according to the number of interaction behaviors and the active duration of the user;

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

[0019] S23: Insert the commodity corresponding to the maximum cyclic Hawkes score before the missing data time into the vacant position to obtain the equidistant sampling time series. ​​

[0020] Further, the formula for calculating the basic generation rate of the user is:

[0021]

[0022] wherein μ represents the basic generation rate of the user, n represents the number of interaction behaviors of the user, and T represents the active duration of the user.

[0023] Further, the formula for calculating the cyclic Hawkes score of the user is:

[0024]

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

[0026] Preferably, the process for obtaining the score of the user on the commodity comprises:

[0027] S31: processing the equal-interval sampling time series by using the multi-element Hawkes process to obtain the coarse-grained preference representation and the fine-grained preference representation of the user;

[0028] S32: calculating the coarse-grained preference weight and the fine-grained preference weight by using the self-attention mechanism;

[0029] S33: calculating the multi-element Hawkes score according to the coarse-grained preference representation, the fine-grained preference representation, the coarse-grained preference weight and the fine-grained preference weight of the user and performing normalization processing to obtain the final score of the user on the commodity.

[0030] Further, the formula for obtaining the coarse-grained preference representation and the fine-grained preference representation of the user is:

[0031]

[0032] wherein μ u,i represents the basic love degree of the user u on the commodity i, i.e., the coarse-grained preference representation of the user; g u (t-t h ) represents the time decay function of the interaction time t h of the commodity h with the user u from the occurrence to the current time t; a h，i represents the excitation degree of the commodity h on the current commodity i; is the fine-grained preference representation of the user u, δ u represents the decay rate of the influence of the historical event; H u，t represents the historical equal-interval sampling time series of the user u before time t; and v idenotes the embedding representation of item i, v u denotes the embedding representation of user u, v h denotes the embedding representation of item h.

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

[0034]

[0035] wherein, denotes the coarse-grained preference weight of user u, denotes the fine-grained preference weight of user u, a l , a s denotes the first and second hyperparameters, b l , b s denotes the first and second bias; v u denotes the embedding representation of user u, denotes the aggregated historical embedding, v h denotes the embedding representation of item h, H u，t denotes the historical equidistant sampling time series of user u before time t.

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

[0037]

[0038] wherein, λ i|u (t) denotes the preference intensity of user u for item i at time t, i.e., the multivariate Hawkes score, denotes the coarse-grained preference weight of user u, denotes the fine-grained preference weight of user u, μ u，i denotes the basic love degree of user u for item i, a h，i denotes the initial excitation degree of historical item h for current item i, g u (t-t h ) denotes the time decay function of the interaction time t h of item h and user u to the current time t, t h denotes the interaction time of item h and user u.

[0039] The beneficial effects of the present application are: the present application utilizes the non-equidistant sampling time series data of the users of the e-commerce platform and the target item data, and for the non-equidistant sampling time series of the users, the missing data is filled and the dependence between the features is captured, and then the preference representation is obtained, and the coarse-grained and fine-grained preferences of the users are adaptively utilized to predict the final preferences of the users, the present application can effectively improve the accuracy of the e-commerce recommendation system and realize precise personalized recommendation effect. Attached Figure Description

[0040] Figure 1 This is a flowchart of the non-equal interval sampling time series recommendation method based on Hawkes process in this invention;

[0041] Figure 2 This is a framework diagram of the non-equal interval sampling time series recommendation method based on Hawkes process in this invention;

[0042] Figure 3 A schematic diagram illustrating the prediction and filling of missing values ​​using the cyclic Hawkes model constructed in this invention; Detailed Implementation

[0043] 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.

[0044] This invention proposes a recommendation method for non-equidistant sampling time series based on Hawkes processes, such as... Figure 1 , Figure 2 As shown, the method includes the following:

[0045] S1: Obtain user product interaction data and obtain non-equal interval sampling time series based on user product interaction data.

[0046] Acquire user-product interaction data, including user characteristic data, product characteristic 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 = {u1, u2, ..., u} represents the user set, and each user is represented by u. i Let I = {i1, i2, ..., i} represent the set of items. Therefore, user-item interactions are represented by triples (u, i, t).

[0049] Based on the user-product interaction triplet representation of the non-equal interval sampling time series, it is represented as:

[0050] S u ={(u,i1,t1),(u,i2,t2),...,(u,i n , t n )}

[0051] where S u denotes the non-equidistant sampling time series of user u, (u, i n , t n ) denotes the interaction of user u with item i n at time t n .

[0052] S2: Fill in the missing data of the non-equidistant sampling time series using the circular Hawkes method.

[0053] As shown in Figure 3 , the process of filling in the missing data of the non-equidistant sampling time series using the circular Hawkes method includes:

[0054] S21: Calculate the basic generation rate of the user according to the number of interaction behaviors and the active duration of the user.

[0055] The formula for calculating the basic generation rate of the user is:

[0056]

[0057] where μ denotes the basic generation rate of the user, n denotes the number of interaction behaviors of the user, and T denotes the active duration of the user.

[0058] S22: Calculate the circular Hawkes score of the user at different times according to the trigger kernel and the basic generation rate of the user.

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

[0060]

[0061] where β is the decay rate, which determines the decay speed of the incentive effect over time. The larger β indicates the faster the decay of the incentive effect.

[0062] The formula for calculating the circular Hawkes score of the user is:

[0063]

[0064] where λ(t) denotes the occurrence rate of the user at time t, i.e., the circular Hawkes score, α denotes the incentive intensity coefficient, and the larger α indicates the stronger the incentive effect of past events; g(t-t i ) denotes the trigger kernel function, and N(s) is a counting process, indicating the total number of events that have occurred before time t.

[0065] S23: Insert the item corresponding to the maximum circular Hawkes score before the missing data time into the empty position to obtain a complete non-equidistant sampling time series.

[0066] t m 、t q 、t n is the current data missing moment before the multiple moments, its cycle hawks score:

[0067] λ(t m ) > λ(t q ) > … > λ(t n )

[0068] The score λ(t m ) of this sequence before the current data missing moment is the maximum value, then λ(t m ) corresponding to the goods i m is inserted into the current data missing position, the user's interaction behavior with the goods is determined according to the final hawks score size, for example: if the score is greater than 0.9, it is purchase, the score is between 0.8-0.9, it is added to the shopping cart, the score is between 0.7-0.8, it is collected, the score is between 0.6-0.7, it is like, and the score is between 0.5-0.6, it is considered to be searched.

[0069] S3: adopt multi-hawks method and self-attention mechanism to process non-equidistant sampling time series to obtain the score of the user to the goods.

[0070] The process of obtaining the score of the user to the goods includes:

[0071] S31: adopt multi-hawks process to process equidistant sampling time series to obtain coarse-grained preference representation and fine-grained preference representation of the user.

[0072] The coarse-grained preference representation of the user is:

[0073]

[0074] According to the decay rate δ u ≥0 of the influence of historical events, the hawks score of the user is affected by the exponential kernel function g u (.), the specific formula is as follows:

[0075]

[0076] According to the vector representation of the current goods i and the vector representation of the historical goods h, the excitation degree αh,i of the goods h to the current goods i is calculated, and the specific formula is as follows:

[0077]

[0078] Where μ u，i represents the basic love degree of the user u to the goods i, that is, the coarse-grained preference representation of the user, g u (t-th ) represents the interaction time t of commodity h and user u h Time decay function from the occurrence to the current time t; a h，i Represents the excitation degree of commodity h to the current commodity i; the sequence is summed after multiplication Is the fine-grained preference representation of the user; wherein, v i Represents the embedding representation of commodity i, v h Represents the embedding representation of commodity h, v u Represents the embedding representation of user u, H u，t Represents a historical non-equidistant sampling time sequence of user u before time t; δ u ≥0 represents the decay rate of historical event influence, because the preference of each user may decay at different rates, so this parameter is related to the user.

[0079] S32: Calculate coarse-grained preference weight and fine-grained preference weight by using self-attention mechanism.

[0080] The present application designs an adaptive attention mechanism structure for obtaining preference weight, which is defined as follows:

[0081]

[0082] Among them, 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 s Represents the first and second bias; relu represents the relu activation function; v u Represents the embedding representation of user u; Represents the aggregated historical embedding, which is obtained by average pooling strategy:

[0083]

[0084] S33: Calculate the multivariate Huxley score according to the coarse-grained preference representation and fine-grained preference representation of the user, the coarse-grained preference weight and the fine-grained preference weight, and perform normalization processing to obtain the final user score of the commodity.

[0085] The multi-Hawkes process can be summarized as the embedding of all historical interaction commodities, and the formula for calculating the multi-Hawkes score of the user is:

[0086]

[0087] Wherein, λ i|u (t) represents the preference intensity of the user u to the commodity i at the time t, that is, the multi-Hawkes score, represents the coarse-grained preference weight of the user u, represents the fine-grained preference weight of the user u, μ u，i represents the basic love degree of the user u to the commodity i, α h，i represents the excitation degree of the commodity h to the current commodity i, g u (t-t h ) represents the time decay function from the interaction time t h of the commodity h and the user u to the current time t, t h represents the interaction time of the commodity h and the user u, H u，t represents a historical equal-interval sampling time sequence of the user u before the time t.

[0088] The application uses an adaptive method of attention and fuses the coarse-grained preference and fine-grained preference of the user, and realizes accurate user sequence personalized recommendation.

[0089] The multi-Hawkes score of the user is normalized to obtain the score of the user to the commodity.

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

[0091]

[0092] Wherein, P i|u (t) represents the score of the user u to the commodity i.

[0093] S4: recommend the K commodities with the highest scores to the user.

[0094] The score of the user to the commodity is sorted, and the K commodities with the highest scores are selected as the recommended commodities of the user and are recommended to the user.

[0095] In summary, the present application predicts and fills in the missing data points of the non-equidistant sampling time series of the user by using a model based on the cyclic hawks process; by using a multivariate hawks process model, the change of user preference over time is simulated; then combined with an adaptive attention mechanism, the attention to the target item in the user's historical behavior is strengthened, and the user's preference is adaptively used to predict the final recommended item. The present application can effectively improve the accuracy of the e-commerce recommendation system and realize precise personalized recommendation effect by predicting the user's preference through the preference representation captured by the dependence between the missing data filling and feature engineering of the non-equidistant sampling time series of the user.

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

Claims

1. A non-equally-spaced sampling time series recommendation method based on the Hox process, characterized by, The method comprises the following steps: S1: obtaining user commodity interaction data and obtaining non-equidistant sampling time series according to the user commodity interaction data; S2: filling in the missing data of the non-equidistant sampling time series by using a cyclic hawkes method to obtain equidistant sampling time series; S3: processing the equidistant sampling time series by using a multivariate hawkes method and a self-attention mechanism to obtain a score of the user on the commodity; the process comprises: S31: processing the equidistant sampling time series by using a multivariate hawkes process to obtain coarse-grained preference representation and fine-grained preference representation of the user; the formula for obtaining the coarse-grained preference representation and the fine-grained preference representation of the user is: ； ； wherein, represents a base level of like for item i by user u, i.e., a coarse-grained preference representation for the user; represents the time of interaction of item h with user u a time decay function that occurs to the current time t; represents the level of excitement of item h for current item i; is a fine-grained preference representation for user u, represents a decay rate for the influence of historical events; represents a history of equally spaced sampling time series for user u up to time t; represents an embedding representation for item i, represents an embedding representation for user u, represents an embedding representation for item h; S32: calculating coarse-grained preference weights and fine-grained preference weights by using a self-attention mechanism; S33: calculating multivariate hawkes scores according to the coarse-grained preference representation, the fine-grained preference representation, the coarse-grained preference weights and the fine-grained preference weights of the user and performing normalization processing to obtain the final score of the user on the commodity; S4: recommending K commodities with the highest scores to the user.

2. The non-equally-spaced sampling time series recommendation method based on the Hocke process according to claim 1, characterized in that, The non-equidistant sampling time series is represented as: ； in, This represents the non-equal interval sampling time series of user u. Indicates user u in Time and Goods Interact with each other.

3. The non-equally spaced sampling time series recommendation method based on the Hocke process according to claim 1, characterized in that, The process of filling in the missing data of the non-equidistant sampling time series by using a cyclic hawkes method comprises: S21: calculating a basic generation rate of the user according to the number of interaction behaviors and the active duration of the user; S22: calculating a cyclic hawkes score of the user at the time of missing values in the time series according to a trigger core and the basic generation rate of the user; S23: inserting the commodity corresponding to the maximum cyclic hawkes score before the time of missing data into the vacant position to obtain equidistant sampling time series.

4. The non-equally-spaced sampling time series recommendation method based on the Hocke process according to claim 3, characterized in that, The formula for calculating the basic generation rate of the user is: ； wherein, represents a base generation rate of the user, represents a number of interaction behaviors of the user, represents an active duration of the user.

5. The non-equally spaced sampling time series recommendation method based on the Hocke process according to claim 3, characterized in that, The formula for calculating the cyclic hawkes score of the user is: ； wherein, represents the incidence of the user at time t, i.e. the cycle hawks score, represents the excitation intensity coefficient, represents the trigger kernel function, represents the sum of events that have occurred prior to time t.

6. The non-equally-spaced sampling time series recommendation method based on the Hocke process according to claim 1, characterized in that, The formula for calculating the coarse-grained preference weights and the fine-grained preference weights is: ； ； ； wherein, denotes a coarse-grained preference weight for user u, denotes a fine-grained preference weight for user u, , denotes first, second hyperparameters, , denotes first, second bias terms; denotes an embedding representation of user u, denotes an aggregated history embedding, denotes an embedding representation of item h, denotes a history of equally-spaced sampling time series of user u up to time t.

7. The non-equally-spaced sampling time series recommendation method based on the Hocke process according to claim 1, characterized in that, The formula for calculating the multivariate hawkes score of the user based on the self-attention weights is: ； wherein, represents the preference intensity of user u for item i at time t, i.e. the Multivariate Hodge Score, represents the coarse-grained preference weight of user u, represents the fine-grained preference weight of user u, represents the base level of love of user u for item i, represents the initial excitement level of historical item h for current item i, represents the time of interaction of item h with user u. a time decay function from the time of interaction of item h with user u to the current time t, represents the time of interaction of item h with user u.

Images