A service recommendation method based on users' short-term preferences

Abstract

The present invention discloses a service recommendation method based on short-term user preferences. First, the behavioral inertia in service recommendation is formally defined and quantified by the quantity and probability of the goods purchased by users. Secondly, three user purchase behavior patterns are summarized to extract features as the input of the model to predict the set of candidate goods that the user will purchase. Finally, the problem of recommendation is solved by adapting to Newton's first law by defining the user interest retention as the "external force" to change the user's current state. And the user interest retention is calculated through the Ebbinghaus forgetting curve, the inertia and interest retention of the user are compared to determine whether the current purchase state of the user has changed, and the goods that may change the user's purchase state are selected as the final recommendation result. The present invention improves the effectiveness of the recommendation result.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer services, and relates to a service recommendation method for short-term user preferences, and particularly to a service recommendation method based on user behavior pattern inertia. Background Art

[0002] A service recommendation system can recommend information or products that a user may be interested in according to the user's purchase behavior or interest preferences. The service recommendation technology can achieve the function of quickly finding services that meet the user's needs (such as products, movies, music, news, pictures, etc.) in a large amount of data. Among them, the user's interests are divided into the user's long-term preferences and short-term preferences, and the recommendation scenario is the factor that determines the length of the user's interests. The user's short-term interests are more focused on judging the user's next moment preference through the user's behavior. However, the traditional service recommendation method based on short-term user preferences pays more attention to the user's purchase behavior and ignores the influence of other behaviors and the resulting user inertia on the final recommendation result. How to achieve accurate recommendation is still one of the important challenges in service recommendation.

[0003] The traditional service recommendation system lacks attention to the inertial behavior of user continuity and only considers a single user purchase behavior, ignoring the influence of other auxiliary behaviors (such as clicks, collections, add-to-cart) and user behavior inertia on user decision-making, resulting in the inability to accurately recommend services to users.

[0004] The traditional recommendation method focuses on the user's purchase behavior, including repeated purchase behavior, while other purchase behaviors are ignored. However, in real-world applications, such as the scenario of online shopping by users, there are multiple behaviors, and each behavior has different degrees of influence on the final purchase behavior, and there is a very complex relationship between them. Behavioral data can also represent the user's current state. For example, some users continuously browse the same category of products within a period of time, which means they have a clear purchase intention. Others just browse casually, which means they have no clear goal. For these two types of users, the viewing behavior has different meanings and different recommendation results should be given. Therefore, ignoring other behavioral data will result in the recommendation result not matching the current demand. And the user's other behaviors are divided into single-type actions, such as clicks, and such models are relatively easy to learn. However, in a multi-type behavior session, it is necessary to pay attention to more than one type of action. Users usually click on products for comparison, then add them to the shopping cart, and finally decide to purchase. These actions have different degrees of influence on the final purchase behavior. Therefore, it is necessary to use multiple user behavioral data to improve the accuracy of the recommendation result.

[0005] At the same time, these studies also failed to realize that due to the lack of attention to continuous user behavior, there is behavioral inertia. Behavioral inertia refers to the reliability and familiarity accumulated by customers based on their past consumption experiences in the marketing field. It plays an important role in intention generation and behavior occurrence, and is habitual, long-term, and goal-oriented. For example, when users first purchase a new product, they will make multiple comparisons and finally make the best choice. And after users establish trust in an item, when the same scenario occurs again, users will habitually repeat the previous purchase behavior without thinking, which is the manifestation of behavioral inertia. Repeated browsing and the shopping patterns that exist when users purchase goods both belong to a type of behavioral inertia. These inertial behaviors reflect the current needs of users and can affect users' final decisions. Ignoring behavioral inertia will lead to misjudgment of users' recommendation preferences. In current service recommendation research, few studies take into account behavioral inertia. Although there are also studies related to behavioral inertia in other fields, few people quantify behavioral inertia and apply it to service recommendation problems. Summary of the Invention

[0006] To solve the above method for user behavioral inertia existing in the prior art, the present invention provides a service recommendation method based on users' short-term preferences.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A service recommendation method based on users' short-term preferences includes the following steps:

[0009] Step S1, statistically analyze a large-scale real dataset and extract user behavior patterns:

[0010] Step S11, statistically analyze the large-scale real dataset. This dataset includes four behaviors: browsing, adding to cart, favoriting, and purchasing, which are respectively represented as pv, cart, fav, and buy. Each row of the dataset represents a user behavior and consists of a user ID, a product ID, a product category ID, a behavior type, and a timestamp; screen and sort the dataset, and statistically analyze the impact of users' browsing, adding to cart, and favoriting behaviors on users' preferences;

[0011] Step S12, propose three user behavior patterns, including random, reciprocating, and progressive, and distinguish the current different behavior patterns of users by the number of times users browse different types of services per minute;

[0012] Step S2, establish a user behavior inertia model:

[0013] Define the user behavior inertia in the service recommendation problem, quantify the behavior inertia of different users, calculate the user memory retention according to the browsing time interval and the forgetting curve, and finally calculate the user interest retention through the user memory retention. The specific steps are as follows:

[0014] Step S21: Assume that the user has n purchase motivations (x 1 , x 2 ,..., x i ,... x n ). Regard a single commodity as a motivation, and the occurrence probability of x i is p(x i ), and calculate the user behavior inertia B;

[0015] Step S22: Calculate the user memory retention m(T i ), and reflect the user memory retention m(T i ) through the forgetting process presented by the Ebbinghaus curve;

[0016] Step S23: Calculate the user interest retention R i through the user memory retention m(T i );

[0017] Step S3: Establish a user short-term interest recommendation model and complete the final recommendation:

[0018] The user short-term interest recommendation model mainly includes two modules: the short-term preference module based on the user pattern and the probability correction module. Among them, the short-term preference module based on the user pattern learns the user's short-term preferences through various behavior data of the user; the probability correction module is used to judge the influence degree of different commodities on the user's current state and correct the prediction probability output in the short-term preference module of the user pattern. The specific steps are as follows:

[0019] Step S31: Since there is no specific category name in the large-scale real dataset of Taobao, first use FP-tree to mine the association relationship between commodity categories in the dataset, set appropriate support parameters and find the frequent k-item sets;

[0020] Step S32: Input the user behaviors in Step S1, including clicks, adding to the shopping cart, collecting, and the number of commodities browsed by the user per minute, as features into the user short-term interest recommendation model; input the association relationship mined in Step S31 as features into the user short-term interest recommendation model as well;

[0021] Step S33: Use the LSTM algorithm to learn the user's short-term interests and finally output the probability of predicting each commodity;

[0022] Step S34: Map the inertia and external forces in Newton's First Law to the service recommendation problem, and regard the user interest retention calculated in Step S2 as the external force that changes the user's purchase status. At the current moment t, compare the user interest retention of user u for product i with the user's own inertia to determine whether the user's purchase behavior has changed;

[0023] Step S35: Screen the predicted probabilities output in Step S33 in the way of Step S34, retain the products with interest retention greater than the user's inertia magnitude, and finally generate the prediction result.

[0024] Compared with the prior art, the present invention has the following advantages:

[0025] 1. The present invention considers the factor of user behavior inertia in the service recommendation problem, clearly defines user behavior inertia, and provides a method for quantifying user behavior inertia.

[0026] 2. The present invention solves the recommendation problem by adapting to Newton's First Law by defining the user interest retention as the "external force" that changes the user's current state, calculates the user's interest retention through the Ebbinghaus forgetting curve, compares the user's inertia and interest retention, and determines whether the user's current purchase status has changed.

[0027] 3. The present invention proposes a service recommendation method based on user behavior pattern inertia. By using LSTM to learn the user's short-term interest preferences, it well represents the influence of the user's other behaviors on the user's final decision, fuses the user inertia module and the user short-term interest module, corrects the probability output of LSTM, and finally predicts the product that the user will purchase at the next moment, thereby improving the effectiveness of the recommendation result. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is the flowchart of the service recommendation method based on user short-term preferences of the present invention;

[0029] Figure 2 is the flowchart for establishing the user behavior inertia model;

[0030] Figure 3 is the module structure diagram of the short-term preference module based on user patterns;

[0031] Figure 4 is the flowchart of the probability correction module. DETAILED DESCRIPTION OF THE INVENTION

[0032] The technical solutions of the present invention will be further described below in conjunction with the drawings, but are not limited thereto. Any modification or equivalent replacement of the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention shall be covered within the protection scope of the present invention.

[0033] The present invention provides a service recommendation method based on short-term user preferences. First, the behavioral inertia in service recommendation is formally defined and quantified by the number and probability of products purchased by users. Secondly, three user purchase behavior patterns are summarized to extract features as the input of the model to predict the set of candidate products that the user will purchase. Finally, by defining the user interest retention as the "external force" to change the user's current state to adapt to Newton's first law to solve the recommendation problem. And the user's interest retention is calculated through the Ebbinghaus forgetting curve, the user's inertia and interest retention are compared to determine whether the user's current purchase state has changed, and the products that may change the user's purchase state are selected as the final recommendation result. As shown in Figure 1, the method includes the following steps:

[0034] Step S1: Statistically analyze a large-scale real dataset and extract user behavior patterns.

[0035] In this step, the publicly available large-scale Taobao dataset is screened and statistically analyzed. The dataset contains all behaviors (behaviors include clicks, purchases, add-to-carts, and favorites) of approximately one million random users with behaviors between November 25, 2017 and December 3, 2017. Through structured query statements, data with missing values and outliers in the timestamp attribute (not in the time period defined in the dataset description) are found. The proportion of each category of behavior is counted, and the number of user clicks and purchases is counted by the time period of each day. The trends of click volume, add-to-cart volume, favorite volume, and purchase volume are counted, and the impact of different behaviors on the user's final decision is analyzed based on the real behavior data and purchase data left by the user on the network platform. The relationship between the number of shopping items and the number of users is counted, with a time span of 9 days, and users with too few or too many purchase times are screened to ensure the accuracy of the data results. At the same time, three user behavior patterns in online shopping are proposed, including random, reciprocating, and progressive, and the current different behavior patterns of users are distinguished by the number of times the user browses different types of services per minute.

[0036] Step S2: Establish a user behavior inertia model.

[0037] This step needs to process the data screened in step S1, and the specific process is as Figure 2 shown. First, user behavior inertia is defined in the service recommendation problem. Behavioral inertia refers to the reliability and familiarity accumulated by users based on past consumption experiences, which can be understood as a person's behavior habit and is quantified by calculating the number and probability of products purchased by the user.

[0038] Suppose the user has n purchase motivations (x 1 , x 2 ,..., x i ,...x n) Consider a single item as a motivation, x i The probability of occurrence is p(x i ), and calculate the user's behavioral inertia When p(x 1 ) = p(x 2 ) =... = p(x i ) =... = p(x n ) = p = 1 / n, it means that the probability of the user purchasing all items is the same, and the behavioral inertia value is the smallest. In this case, the user's preference for purchasing items is not very clear, and the purchase status is easily changed. On the other hand, if the user repeatedly purchases a certain type of item, the inertia value will increase. The greater the behavioral inertia value, the more difficult it is to change the user's behavior.

[0039] Secondly, calculate the user's memory retention degree where T is the independent variable, representing the time interval between the current time and the time of viewing item i, e is the natural number base, b, c, and δ are constants, and t 0 = 0.00255. Although forgetting is a complex psychological phenomenon, its process is regular. It can be reflected by the forgetting process presented by the Ebbinghaus curve. The forgetting curve function reflects that the user's interest in the most recently viewed content is more important than the interest in the content viewed a long time ago, and its importance gradually decays with time t.

[0040] Finally, calculate the user's interest retention amount R i , the number of times a product or service is viewed will also affect the measurement of the user's interest retention amount. The user's interest retention amount is jointly determined by time and the user's memory retention degree m(T i ). where V i is the number of times the user views item i, V is the total number of times the user views and uses the item, and T is the time interval from the time stamp of the last view of i to the current time stamp.

[0041] Step S3: Establish a user short-term interest recommendation model and complete the final recommendation.

[0042] The user short-term interest recommendation model mainly includes two modules: the short-term preference module based on the user pattern and the probability correction module. Among them: the short-term preference module based on the user pattern learns the user's short-term preferences through various behavioral data of the user; the probability correction module is used to judge the influence degree of different items on the user's current state and correct the prediction probability output by the short-term preference module of the user pattern. The user short-term interest recommendation model mainly includes:

[0043] Step S31: Since there are no specific category names in the large-scale Taobao real dataset, first, mine the association relationships between product categories in the dataset, set appropriate support parameters, and find frequent k-itemsets.

[0044] Step S32: Use the user behaviors in Step S1, including clicks, adds to cart, favorites, and the number of products browsed per minute by the user, as features and input them into the user short-term interest recommendation model; also use the association relationships mined in Step S31 as features and input them into the user short-term interest recommendation model.

[0045] Step S33: Learn the user short-term interest through the LSTM algorithm and finally output the probability of predicting each product. The specific steps are as follows:

[0046] Represent the user behaviors and frequent k-itemsets in Step S32 as the input of the user short-term interest recommendation model, denoted as x t = [u b ; f t , where: u b represents various types of user behaviors and behavior patterns, and f t represents the mined frequent k-itemsets; there are three gates in the user short-term interest recommendation model, namely the input gate i t = σ(W i [h t-1 , x t +b f ), the forget gate f t = σ(W f [h t-1 , x t +b i ), and the output gate o t = σ(W o [h t-1 , x t +b o ), which respectively determine what information to store, forget, and output. Here: σ is a sigmoid function that outputs a number between 0 and 1, W i , W f , and W o represent weights respectively, and b i , b f , and b o represent the corresponding biases respectively; g t = tanh(W c [h t-1 , x t +b g ) represents the new candidate state vector at step t; ⊙ is the element-wise product of two vectors, f t ⊙ ct-1 represents the old state c that we decided to forget t-1 After the information is obtained, the retained information is obtained from the old state; i t ⊙g t-1 represents the new state g that we decided to store t New information obtained from t =f t ⊙c t-1 +i t ⊙g t-1 is combined with the old state c t-1 and the new state g t The final state vector of the information; h t The hidden output vector representing the user's preferences is denoted as h t =o t ⊙tanh(c t ); Through the above process, the user's short-term interests are learned, and the probability of predicting each product is finally output. LSTM is a long short-term memory network, a time recursive neural network, suitable for processing and predicting important events in time series. Specific structure of the short-term preference module based on user patterns Figure 3 First, various user behaviors, behavior patterns, and frequent item sets are input as features into the LSTM model. The user's short-term preferences are learned through LSTM, and the final output is the predicted purchase probability of each item by the user. The probabilities are sorted from high to low to form a top-k set.

[0047] Step S34, map the inertia and external force in Newton's first law to the service recommendation problem, and regard the user interest retention amount calculated in step S2 as the external force that changes the user's purchasing status. At the current time t, compare the user u's interest retention amount in product i with the user's own inertia to determine whether the user's purchasing behavior has changed.

[0048] Step S35: filter the predicted probability output in step S33 by the method of step S34, retain the products whose interest retention is greater than the user's inertia, and finally generate the prediction result. Figure 4 As shown in the figure, first, the user interest retention and inertia are compared. If the user inertia is greater than the user interest retention, it means that the user's preference for the item is not enough to change the user's current purchase status, so the item is deleted from the item set, otherwise the item is retained. After screening, a new recommendation candidate set is formed and recommended to the user.

Claims

1. A service recommendation method based on users' short-term preferences, characterized in that the method comprises the following steps: Step S1, statistically analyze a large-scale real dataset and extract user behavior patterns: Step S11, statistically analyze the large-scale real dataset, which includes four behaviors: browsing, adding to cart, favoriting, and purchasing, represented as pv, cart, fav, and buy respectively. Each row of the dataset represents a user behavior, which consists of user ID, product ID, product category ID, behavior type, and timestamp; screen and sort the dataset, and statistically analyze the influence of users' browsing, adding to cart, and favoriting behaviors on user preferences; Step S12, propose three user behavior patterns, including random, reciprocating, and progressive, and distinguish the current different behavior patterns of users by the number of times users browse different types of services per minute; Step S2, establish a user behavior inertia model: Define the user behavior inertia in the service recommendation problem, quantify the magnitude of behavior inertia for different users, and calculate the user memory retention m(T i ) according to the browsing time interval and the forgetting curve. Finally, calculate the user interest retention R i ) based on the user memory retention m(T i ); The specific steps are as follows: Step S21. Assume that the user has n purchase motivations (x 1 , x 2 ,..., x i ,... x n ). Consider a single commodity as one motivation. The occurrence probability of x i is p(x i ), and calculate the user behavior inertia B; Step S22, calculate the user's memory retention m(T i ), and reflect the user's memory retention m(T i ) through the forgetting process presented by the Ebbinghaus curve; Step S23: Calculate the user interest retention amount R based on the user memory retention degree m(T i ) i ; Step S3, establish a user short-term interest recommendation model and complete the final recommendation: The user short-term interest recommendation model mainly includes two modules: a short-term preference module based on user patterns and a probability correction module. Among them: the short-term preference module based on user patterns learns users' short-term preferences through various behavior data of users; the probability correction module is used to judge the influence degree of different products on the user's current state and correct the prediction probability output by the short-term preference module of user patterns. The specific steps are as follows: Step S31, since there is no specific category name in the large-scale real dataset, first use FP-tree to mine the association relationship between product categories in the dataset, set appropriate support parameters, and find frequent k-item sets; Step S32, input the user behaviors in Step S1, including clicks, adding to cart, favoriting, and the number of times users browse products per minute, as features into the user short-term interest recommendation model; input the association relationship mined in Step S31 as features into the user short-term interest recommendation model as well; Step S33, learn the user short-term interest through the LSTM algorithm and finally output the probability of predicting each product; Step S34, map the inertia and external force in Newton's first law to the service recommendation problem, and regard the user interest retention calculated in Step S2 as the external force to change the user's purchase state. At the current moment a, compare the user interest retention of user u for product i with the user's own inertia to judge whether the user's purchase behavior has changed; Step S35, screen the prediction probability output in Step S33 in the way of Step S34, retain the products with the interest retention greater than the user's inertia, and finally generate the prediction result.

2. The service recommendation method based on users' short-term preferences according to claim 1, characterized in that The user memory retention m(T i ) is calculated as follows: where T is the independent variable representing the time interval between the current time and viewing product i, e is the natural number base, b, c, and δ are constants, and t 0 = 0.00255.

3. The service recommendation method based on users' short-term preferences according to claim 1, characterized in that The user interest retention amount R i has the following calculation formula: where V i is the number of times the user views product i, V is the total number of times the user views and uses products, T is the independent variable representing the time interval between the current time and the viewing of product i, e is the natural number base, b, c, and δ are constants, and t 0 = 0.00255.

4. The service recommendation method based on users' short-term preferences according to claim 1, characterized in that The specific steps of Step S33 are as follows: The user behavior and frequent k-item sets in step S32 are used as the input of the user short-term interest recommendation model, denoted as x t = [u b ; f t , where: u b represents various types of user behaviors and behavior patterns, and f t represents the mined frequent k-item sets; there are three gates in the user short-term interest recommendation model, namely the input gate i t = σ(W i [h t-1 , x t + b f ), the forget gate f t = σ(W f [h t-1 , x t + b i ), and the output gate o t = σ(W o [h t-1 , x t + b o ), which respectively determine what information to store, forget, and output. Among them: σ is a sigmoid function that outputs a number between 0 and 1, W i , W f , and W o represent weights respectively, and b i , b f , and b o represent the corresponding biases respectively; g t = tanh(W c [h t-1 , x t + b g ) represents the new candidate state vector at step t; ⊙ is the element-wise product of two vectors, and f t ⊙ c t-1 represents the retained information obtained from the old state c t-1 after we decide to forget the information in the old state; i t ⊙ g t-1 represents the new information obtained from the new state g t that we decide to store; c t = f t ⊙ c t-1 + i t ⊙ g t-1 is the final state vector that combines the information of the old state c t-1 and the new state g t ; h t represents the hidden output vector of user preferences, denoted as h t = o t ⊙tanh(c t ); By learning the user's short-term interest through the above process, the probability of each product is finally predicted and output.

Images