Product recommendation method, device and storage medium oriented to surprise degree

Abstract

The present application provides a product recommendation oriented to the degree of surprise and related equipment, which can realize product recommendations with more surprise and smaller granularity for users. The method includes: determining the elasticity factor between the target user and the initial candidate product according to the historical behavior record corresponding to the target user; screening the initial candidate product set based on the elasticity factor and the similarity between the historical product and the candidate product to obtain the target candidate product set; determining the interest influence factor according to the historical behavior record; determining the aspect rating of each aspect corresponding to the target product type by the target user according to the interest influence factor; determining the product score of the target candidate product at the current moment, the embedding vector corresponding to the target candidate product, and the interest cluster corresponding to the target candidate product; recommending the candidate products in the target candidate product set to the target user according to the interest cluster, the interest influence factor, the aspect rating of each aspect in the target product type, and the product score.

Description

[Technical field]

[0001] The present application belongs to the field of recommendation, and in particular, relates to a product recommendation method, device and storage medium oriented to surprise. [Background technology]

[0002] Recommendation systems are designed to recommend products that users may use from a wide variety of products. Currently, most recommendation systems use accuracy as an optimization goal, trying to recommend the most relevant products to users, but this has not led to a sustained increase in user acceptance and satisfaction. On the contrary, excessive accuracy can lead to problems such as information cocoons and filter bubbles, which has led to an increasing focus on surprise recommendations, which are dedicated to providing recommendations that can arouse users' interest but are beyond their ability to discover.

[0003] Online platforms such as e-commerce, music, and movies use recommendation systems to help users select the options that are more in line with their preferences from a variety of goods and cultural products. Existing surprise recommendation technologies recommend products based on their surprise based on the user's historical usage and rating behavior. This recommendation technology has a large granularity and cannot make recommendations based on the evolution of the user's interests. [Summary of the invention]

[0004] The present application provides a product recommendation method, device and storage medium oriented to surprise, which can achieve more surprising and smaller-granular product recommendations for users, thereby improving user experience.

[0005] The first aspect of the present application provides a product recommendation method oriented to surprise, comprising:

[0006] Determine the elasticity factor between the target user and each initial candidate product in the initial candidate product set according to the historical behavior record corresponding to the target user;

[0007] Screening the initial candidate product set based on the elasticity factor and the similarity between historical products and candidate products to obtain a target candidate product set;

[0008] Determine, according to the historical behavior records, an interest influencing factor of each behavior in each aspect corresponding to the target product type, the target product type being the product type corresponding to the target candidate product, and the target candidate product being any candidate product in the target candidate product set;

[0009] Determine a set of important behaviors according to the interest influencing factors of each behavior in each aspect corresponding to the target product type, and then determine the aspect rating of the target user for each aspect corresponding to the target product type;

[0010] Determine the product score of the target candidate product at the current moment, the embedding vector corresponding to the target candidate product, and the interest clustering of all products that meet the target product type based on the optimal aspect;

[0011] The candidate products in the target candidate product set are recommended to the target user according to the interest cluster, the interest influencing factor, the aspect rating of each aspect in the target product type, and the product rating.

[0012] A second aspect of the present application provides a product recommendation device oriented to surprise, comprising:

[0013] An elasticity factor determination unit, configured to determine an elasticity factor between the target user and each initial candidate product in the initial candidate product set according to the historical behavior record corresponding to the target user;

[0014] A screening unit, configured to screen the initial candidate product set according to the elasticity factor and the similarity between the historical products and the candidate products to obtain a target candidate product set;

[0015] an influence factor determination unit, configured to determine, according to the historical behavior record, an interest influence factor of each behavior in each aspect corresponding to the target product type, wherein the target product type is a product type corresponding to the target candidate product, and the target candidate product is any candidate product in the target candidate product set;

[0016] an aspect rating determination unit, configured to determine a set of important behaviors according to an interest influencing factor of each behavior in each aspect corresponding to the target product type, and further determine an aspect rating of the target user for each aspect corresponding to the target product type;

[0017] A comprehensive determination unit, used to determine the product score of the target candidate product at the current moment, the embedding vector corresponding to the target candidate product, and the interest clustering of all products that meet the target product type based on the optimal aspect;

[0018] A recommendation unit is used to recommend the candidate products in the target candidate product set to the target user according to the interest cluster, the interest influencing factor, the aspect rating of each aspect in the target product type, and the product rating.

[0019] A third aspect of an embodiment of the present application provides a computer device comprising at least one connected processor, memory, and transceiver, wherein the memory is used to store program code, and the processor is used to call the program code in the memory to execute the steps of the surprise-oriented product recommendation method described in the first aspect above.

[0020] A fourth aspect of an embodiment of the present application provides a computer storage medium, which includes instructions, which, when executed on a computer, enables the computer to execute the steps of the product recommendation method oriented to surprise level as described in any of the above aspects.

[0021] Compared with the related art, in the embodiments provided by the present application, when recommending products to users, the products that users are interested in are initially selected as candidate products in combination with the elasticity factor and the similarity between historical products and candidate products, and then the basic scores are predicted based on the aspect features. In addition, the evolution of different types of aspect preferences is taken into account, and the user's surprise for the candidate products is determined through interest influence factors, interest clustering, user's aspect ratings of product types, and attenuation functions. This can achieve more surprising and smaller-granular surprise recommendations for users, thereby improving user experience.

Brief Description of the Drawings

[0022] Figure 1 A flowchart of a product recommendation method for surprise level provided in an embodiment of the present application;

[0023] Figure 2 A schematic diagram of a gated recurrent unit provided in an embodiment of the present application;

[0024] Figure 3 A schematic diagram of behavior clustering in a latent space provided by an embodiment of the present application;

[0025] Figure 4 A virtual structural diagram of a product recommendation device oriented to surprise degree provided in an embodiment of the present application;

[0026] Figure 5 A schematic diagram of the hardware structure of the server provided in an embodiment of the present application. [Specific implementation method]

[0027] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0028] The following describes a method for recommending products based on surprise from the perspective of a product recommendation device based on surprise. The product recommendation device based on surprise may be a server or a service unit in the server, and is not specifically limited.

[0029] Please refer to Figure 1 , Figure 1 A flowchart of a product recommendation method for surprise level provided in an embodiment of the present application includes:

[0030] 101. Determine the elasticity factor between the target user and each initial candidate product in the initial candidate product set and the similarity between the historical products in the historical behavior record and each initial candidate product according to the historical behavior record corresponding to the target user.

[0031] In this embodiment, the product recommendation device oriented to surprise can obtain the historical behavior record corresponding to the target user, and determine the elasticity factor between the target user and each initial candidate product in the initial candidate product set based on the historical behavior record corresponding to the target user, wherein the historical behavior record includes the user's product review set and the overall score corresponding to each review. At the same time, the product recommendation device oriented to surprise can also determine the similarity between the historical products in the historical behavior record and each initial candidate product, and the similarity here refers to the cosine similarity, and the method of determining the similarity is not specifically limited.

[0032] In one embodiment, the surprise-oriented product recommendation device determines the elasticity factor between the target user and each initial candidate product in the initial candidate product set according to the historical behavior record corresponding to the target user, including:

[0033] Determine the diversity of product categories and aspects corresponding to target users based on historical behavior records;

[0034] Determine the user elasticity corresponding to the target users based on product category diversity and aspect diversity;

[0035] Determine the user group diversity corresponding to each initial candidate product in the initial candidate product set and the aspect popularity corresponding to each initial candidate product according to the historical behavior records;

[0036] Determine the product elasticity corresponding to the candidate product based on the diversity of the user group and the popularity of the aspect;

[0037] The elasticity factor is determined according to the user elasticity of the target users and the product elasticity corresponding to the candidate products.

[0038] In this embodiment, the degree of surprise brought by the recommended product to the user is related to the user's acceptance level and the audience level of the product. Users will lack interest in products that are beyond their acceptance level or are too familiar to them. On the contrary, they will be more interested in products that are within their acceptance level but have a certain novelty and will bring them surprises. After obtaining the historical behavior records corresponding to the target user, the product recommendation device oriented to surprise level can first determine the diversity of product categories and the diversity of aspects corresponding to the target user based on the historical behavior records, and determine the user elasticity corresponding to the target user based on the diversity of product categories and the diversity of aspects. User elasticity represents the user's ability to accept different products. The richer the product categories related to the user, the stronger the user's ability to accept different products. The number of aspects that users pay attention to for a certain type of product (i.e., aspect diversity) will also affect the user's acceptance of different products of the same type. Here, the normalized D(G) and D(A) are used. m ) to represent the diversity of product categories and aspects of target users. Aspect diversity refers to the diversity of product categories of users in historical behavior records. m The number of aspects that the product focuses on, then the user elasticity corresponding to the target user From D(G) and D(A m ) jointly decide on the specific user flexibility It can be calculated using either of the following two formulas:

[0039]

[0040] It should be noted that the above formula is only for illustration, and other formulas may also be used for calculation without limitation.

[0041] Afterwards, the surprise-oriented product recommendation device can determine the user group diversity and aspect popularity of each initial candidate product in the initial candidate product set based on historical behavior records, and determine the product elasticity corresponding to the candidate product based on the user group diversity and aspect popularity. Product elasticity represents the probability of a product being accepted by different users, where the diversity of the user group is related to the user's age, occupation, and the size of the group itself, and aspect popularity is related to the number of aspects with high product ratings. The user group diversity and aspect popularity of each initial candidate product are represented by the normalized D(U) and D(A), respectively. Then the product elasticity E(p j) is jointly determined by D(U) and D(A). D(U) is obtained by weighting the size of the age group set of users who have used the initial candidate products, the size of the occupational set, and the size of the group itself. D(A) is determined based on the number of aspects of each initial candidate product with an average score above a certain threshold. In addition, there are many ways to normalize, such as the number of aspects with an average score above a certain threshold / the total number of aspects of products in this category.

[0042] Finally, the surprise-oriented product recommendation device determines the elasticity factor according to the user elasticity corresponding to the target user and the product elasticity corresponding to each initial candidate product. The elasticity factor E(ui,pj) represents the user's ability to accept the candidate product under the premise of considering only the elasticity without considering the correlation between the user and the product. E(ui,pj) is given by Specifically, the elasticity factor can be determined by the following formula:

[0043] Among them, δ is the weight of the product elasticity E(pj) of each candidate product.

[0044] It should be noted that the above formula is only for illustration, and other formulas may also be used for calculation without limitation.

[0045] 102. Screen the initial candidate product set according to the elasticity factor and similarity to obtain the target candidate product set.

[0046] In this embodiment, after determining the elasticity factor, the product recommendation device for surprise can filter the initial candidate product set according to the elasticity factor and the similarity between the historical products and the candidate products to obtain the target candidate product set. The following is a detailed description of how to filter and obtain the target candidate product set:

[0047] In one embodiment, the product recommendation device oriented to surprise degree screens the initial candidate product set according to the elasticity factor and similarity to obtain the target candidate product set including:

[0048] Determine the update factor based on the similarity and elasticity factor;

[0049] Determine the target user's interest status for each initial candidate product based on the update factor;

[0050] A target candidate product set is determined according to the target user's interest status in each initial candidate product.

[0051] In this embodiment, the product recommendation device oriented to surprise can determine the update factor of the update gate of the GRU model according to the similarity and elasticity factor, and then determine the interest state of the target user for each initial candidate product based on the update factor.

[0052] The target user's interest in each initial candidate product is an important factor to be considered first when recommending to the target user. Only when the target user's interest in the initial candidate product reaches a certain level can it be further measured and recommended as a target candidate product. In real life, user preferences change dynamically over time. Therefore, in order to make better product recommendations, capturing and exploring the dynamic changes in these user preferences will more accurately predict the user's current interest in a product.

[0053] First, the surprise-oriented product recommendation device sorts the historical behavior records of the target user by time, and generates an embedding vector through an embedding layer for the sorted historical behavior sequence. Each historical behavior corresponds to an embedding vector, which is the input of the gated recurrent unit (GRU), where GRU is a commonly used gated recurrent neural network.

[0054] Then use GRU to process each historical behavior, that is, use the gating mechanism to control the input (the input is the embedding vector corresponding to each behavior in the historical behavior sequence, such as Figure 2 e(i) in ) and memory (the memory is the hidden state processed by the GRU before the current time step, such as Figure 2 h′(T-1) in , thereby predicting the current target user’s interest status for each initial candidate product. GRU has two gates, a reset gate and an update gate. Intuitively, the reset gate determines how to combine the embedding vector corresponding to the historical behavior input to the GRU at the current time step with the hidden state processed by the GRU before the current time step, and the update gate defines the amount of previous memory saved to the current time step.

[0055] Considering that the elasticity of users will change over time, and the size of the elasticity factor will affect the hidden state of the user at the current time step, the same product is recommended to two users. The larger the elasticity factor, the stronger their acceptance of the product, and the greater their interest. Otherwise, the interest is small. Therefore, the elasticity factor is introduced in the update gate to control the impact of the interest state generated by the behavior at the current time step on the prediction of the final interest state. At the same time, the target user's final interest state for each initial candidate product is only related to the historical behavior of products similar to the candidate product in the historical behavior record. Therefore, the influence of related historical behavior should be enhanced, while the influence of irrelevant behavior on the current interest state should be weakened. In the update gate, each candidate product is combined with the product at the current time step (the product at the current time step refers to the product in the historical behavior processed at the current time step, such as Figure 2As shown in the figure, there are T time steps, each time step has a corresponding historical product) to achieve this effect, that is, introduce the attention mechanism in the update gate, use the elasticity factor and Sim as the attention weight to recalculate the update factor, and thus generate the target user's final interest state for each initial candidate product. The formula is as follows:

[0056] u t =σ(W u i t +U u h t-1 +b u )

[0057] r t =σ(W r i t +U r h t-1 +b r )

[0058]

[0059]

[0060]

[0061] Among them, a = μE(ui,pj)+(1-μ)Sim is the attention weight, E(ui,pj) is the elasticity factor, Sim is the similarity, μ is the weight of the elasticity factor, ut is the update gate, σ is the sigmoid function, through which the data can be transformed into a value in the range of 0-1, thus acting as a gating signal, i t is the input of the current time step, i t =e b (t) represents the embedding vector of the t-th behavior in the historical behavior sequence, r t is the reset gate, h t-1 is the hidden state passed down from the previous time step, is the candidate hidden layer state, which mainly includes the i input at time t t The information and the hidden layer state h at time t-1 are targeted t-1 reservations, is the update gate after attention weight processing, ht is the final hidden state of the tth GRU, W u , U u , U r , W r , W h and U h They are all parameters or weight matrices of the GRU model.

[0062] Given an interest threshold h, if and only if the target user currently has interest in the initial candidate product p j Interest status When the initial candidate product p j Incorporate the target candidate product set. Thus, the initial candidate product set can be screened to obtain the target candidate product set.

[0063] It should be noted that in order to predict the target user's interest state for each initial candidate product in the initial candidate product set based on the target user's historical behavior sequence, it can only be recommended to the target user when the interest state reaches a certain level. Since the elasticity of the target user will change after each historical behavior, and the degree of acceptance of the candidate product will change accordingly, the elasticity factor (user-product elasticity) between the target user and the candidate product after each historical behavior is introduced into the update gate. Furthermore, since the target user's interest in the initial candidate product is often only related to part of the historical behavior, the use of similar products will have a greater impact on the prediction of the target user's interest state, so the impact of the use of irrelevant products on the prediction should be weakened, so the cosine similarity between the product of the historical behavior and the candidate product is introduced to enhance or weaken this impact. Therefore, the elasticity factor and the similarity weight are used as the attention weight to recalculate the value of the update gate, so that the user's interest state for the candidate product can be predicted more dynamically.

[0064] 103. Determine the interest influencing factor of each behavior corresponding to the target aspect based on historical behavior records.

[0065] In this embodiment, the surprise-oriented product recommendation device can determine the interest influence factor of each behavior corresponding to the target aspect based on the historical behavior record, wherein the target aspect is any aspect corresponding to the target product type, the target product type is the product type corresponding to the target candidate product, and the target candidate product is any candidate product in the target candidate product set. The following is a specific description of how to determine the interest influence factor of each behavior corresponding to the target aspect:

[0066] User comments can provide a more fine-grained explanation of the product ratings given by users. Therefore, by deeply mining user comments, users' aspect preferences for products can be obtained from the comments. When recommending candidate products to the target user, products with better attributes in this aspect can be recommended to the target user based on the target user's historical aspect preference information. Target users have different aspect preferences for different types of products. Therefore, to classify and process the target user's comments and rating data, first classify the target user's historical comments by product category and sort them by time. Then, they will be sorted by the target product category. mAs a unit, the aspect feature sequence of the target user is determined and the type of aspect preference of the target user is determined through the following steps:

[0067] Step 1: Extract the historical aspect score and historical aspect weight of each comment in the historical behavior record; use the LARA algorithm to extract the historical aspect score and historical aspect weight of each comment of the target user, and each comment will get a triple set is called the aspect feature of the review, where are the normalized historical aspect scores and historical aspect weights respectively.

[0068] Step 2: determining the feature sequence of each aspect corresponding to the target product type according to the historical aspect score and the historical aspect weight, and calculating the interest influence factor of each behavior of each aspect corresponding to the target product type in the historical record;

[0069] After obtaining the target user's characteristic sequences in several aspects of the products corresponding to the target product type, the target user's interest rating for each aspect of the products corresponding to the target product type is calculated based on each aspect characteristic sequence. The higher the frequency of a user's mention within a period of time, the lower the probability of the user forgetting it. At the same time, the user's rating and weight on a certain aspect can also bring positive or negative effects to the user. Based on this, the user's aspect preferences are divided into momentary preferences, short-term preferences, and long-term preferences.

[0070] First, quantify the impact of the target user's usage behavior at each point in time on the target user's future interest in this aspect. This application proposes an interest influence factor to measure this. There are four possible situations regarding the historical aspect ratings and historical aspect weights of the target user's comments, namely, high weights and high scores, low weights and high scores, low weights and low scores, and high weights and low scores. The first and second situations will make the target user deepen his interest preference in this aspect after using the product. The third situation will not have a great impact on the user's interest preference, and the fourth situation will weaken the user's interest preference in this aspect. Therefore, aspect k is added at t j The interest impact factor at a moment is defined as:

[0071]

[0072] in, For aspect k in t j The interest influencing factor at the moment, aspect k is any aspect corresponding to the target product type, are the normalized historical aspect scores and historical aspect weights respectively.

[0073] It should be noted that the interest influence factor is to quantify the usage records of each type of product in the user's historical behavior, and whether it has a negative or positive impact on the user in various aspects of this type of product, as well as the degree of impact.

[0074] 104. Determine the aspect rating of the target user for each aspect corresponding to the target product type according to the interest influence factor of each behavior corresponding to the target aspect.

[0075] In this embodiment, after determining the interest influence factor of each behavior in each aspect, the product recommendation device for surprise can determine the target user's aspect rating for each aspect corresponding to the target product type based on the interest influence factor, that is, first determine the set of important behaviors corresponding to the target user based on the interest influence factor, and then determine the target user's aspect rating for each aspect corresponding to the target product type. Specifically, the feature sequence of each aspect corresponding to the target product type can be determined, and the set of important behaviors can be determined based on the feature sequence of each aspect corresponding to the target product type and the interest influence factor of each behavior in each aspect corresponding to the target product type, and then determine the target user's interest rating for each aspect of the target product type.

[0076] Specifically, we sort out the characteristics of all comments from target users. For target product type c m , sort the triples corresponding to these comments by time according to aspect type k, and get the target product type c m The feature sequence corresponding to each aspect k can be identified in the following form:

[0077] For each aspect feature sequence It can be obtained that for the target product category c m Aspect k, a set of important behaviors of the target user UI The historical behavior of the target user constantly affects the target user's current interest in the target candidate product. As time goes by, the interest will continue to evolve. By using the positive or negative interest influencing factors contained in the relevant historical behavior in the feature sequence of aspect k of the target user to measure the target user's preference, the target user's interest rating for aspect k is obtained. Specifically, the target user's interest rating for each aspect of the target product type can be measured by the following formula:

[0078]

[0079] in, is the interest rating of the target user for each aspect k in the target product type. δ and θ are two thresholds used to divide the interest rating. instantaneous_preference represents instantaneous preference, short-term_preference represents short-term preference, and long-term_preference represents long-term preference.

[0080] Assume that the following is a feature sequence of clothing materials in the target user's usage record of clothing products:

[0081]

[0082] The aspect score indicates the target user's evaluation or satisfaction with the clothing material, while the aspect weight indicates the target user's emphasis on the clothing material. Normalizing the two, ideally, if the score is higher than the weight, it means that the product has a positive impact on the target user in terms of clothing material, or it will promote the target user's preference for clothing material, and vice versa.

[0083] for example:

[0084] At t0: This means that the clothing products used (purchased) by the target user at time t0 do not meet his or her own standards, or do not match his or her level of importance, which will weaken the target user's preference for clothing materials, and vice versa.

[0085] At t1: the target user has used (purchased) clothing products twice, and each has a score and weight for the clothing material. At t1, the target user's preference for clothing material will be affected by the two uses at t0 and t1. It is not considered that the impact of each use on user preference will decay over time, and it is believed that the impact of each behavior will be equally distributed to all future behaviors, including the current behavior. Therefore, the impact of the behavior is directly divided by the total number of comments during this period, and then the impact of all behaviors on the current moment is added up to represent the user's preference for this aspect at this moment. Therefore, the interest influence factor at t1

[0086] Time t2: Similarly, the interest influence factor at time t2

[0087] By analogy, we can get t q Interest factors at the moment:

[0088] The set of important behaviors is defined as:

[0089] Since the interest influence factor can indicate the impact of the user's past behavior on the current aspect preference, it can also indicate the user's preference for the aspect at a certain moment. It represents t j At this moment, the user is interested in aspect k, and adds this behavior to B k collection, otherwise I am not interested.

[0090] Then we get:

[0091]

[0092] All added in B k The behavior in the set means that the user is interested in this aspect after the behavior occurs, so we can use B k The number of behaviors in The total number of behaviors relative to aspect k The size of the preference is used to distinguish the three preference types.

[0093] 105. Determine the product rating of the target candidate product at the current moment, the embedding vector corresponding to the target candidate product, and the interest cluster corresponding to the target candidate product.

[0094] In this embodiment, the product recommendation device for surprise can determine the product score of the target candidate product at the current moment, the embedding vector corresponding to the target candidate product, and the interest cluster of each. The following is a detailed description of how to determine the product score of the target candidate product, the embedding vector corresponding to the target candidate product, and the interest cluster corresponding to the target candidate product. The cluster corresponding to the target candidate product is the interest cluster based on the optimal aspect of all products that meet the target product type:

[0095] 1. Determine the product ratings of your target candidate products:

[0096] When determining the product score of the target candidate product, the product recommendation device for surprise degree can first determine the current aspect score and current aspect weight of each aspect corresponding to the target product type at the current moment; and determine the product score of the target candidate product according to the current aspect score and current aspect weight. The following is a specific description of the aspect feature sequence A obtained by the LARA algorithm using the product recommendation device for surprise degree. k As the input of GRU, it predicts the current aspect score of each aspect at the current moment and the current aspect weight And predict the basic score of the target candidate product based on the current aspect score and the current aspect weight. The specific score can be calculated by the following formula:

[0097]

[0098] Among them, r b Give the target candidate product a basic score, A m ={A 0 ,A 1 ,K,A n} is the target product type c m The corresponding aspect feature sequence set. The GRU model can predict the final state based on a sequence, so the score sequence of aspect k of the historical behavior and the weight sequence of aspect k can be used as the input of the GRU to predict the user's score for aspect k of the candidate product. and weight The final score of aspect k is The product rating is the sum of the final ratings of all aspects.

[0099] For different user groups, most users prefer products with high aspect ratings (i.e., high popularity), believing that such products can meet their needs for products in this aspect. However, as the maturity of user preferences increases, they can get rid of the influence of highly popular products and tend to make their own judgments. As the popularity of a product (i.e., aspect rating) increases, it is more attractive to low-maturity users, but its appeal to high-maturity users decreases. Therefore, user maturity and product popularity will affect the user's aspect rating and weight of the product. Before predicting the product rating, it is necessary to evaluate the impact of user maturity and product popularity on aspect k. Assuming that the target candidate product i * Belongs to target product type c m Specifically, the product recommendation device for surprise degree can recommend the target candidate product i by the following formula: * Make product rating predictions:

[0100]

[0101] Among them, r b Provide a basic score for the target candidate product, For target users in target product type c m The impact of the maturity of aspect k on user ratings, Indicates that the target user is in the target product type c m The maturity of aspect k ( The historical behavior records corresponding to the target user and c m The number of all comments related to aspect k), is the impact of the popularity of the target candidate product in aspect k on the user rating, represents the popularity of the target candidate product in aspect k ( and the number of comments related to the product aspect k in the history records of all users), and and Normalization is performed. On the basis of the basic rating prediction, the impact of each aspect of user maturity and product popularity on the user rating is further considered.

[0102] For each candidate product i * , we can get its embedding vector w in the latent space * , the original feature vector of the product is compressed and reduced by an autoencoder to reach the feature vector of the latent space, the target user's aspect-based interest in the same type of product is modeled as a cluster of historical behaviors in the latent space, and all candidate products of each type are embedded in the latent space based on their optimal aspects. The more similar the optimal aspects of the products are, the closer they will be in the latent space. Note that the target user usually has multiple preferences, so multiple different clusters should be used to represent the target user's multiple interests instead of using a single closure. Therefore, the mean shift algorithm is used to identify historical behavior clusters in the latent space to represent the user's product-based interest preferences, such as Figure 3 As shown, the target user u i Historical behavior records about target product category c m The product and all candidate products of this category are embedded into the latent space according to their optimal aspects, and the mean shift algorithm is used to obtain the interest clustering set C = {C 1 ,C 2 ,K,C n}, then each cluster C K Represents the target user based on the target product type c m The set of products that are of interest in aspect k.

[0103] It should be noted that latent space is a dimensionality reduction or compression representation of data. It is a concept in deep learning that can learn the characteristics of data and simplify the representation. An autoencoder can be used to obtain the characteristic representation of the product in the latent space. First, the characteristic vector representation of the product is learned, including the optimal aspect, as the input of the encoder. After passing through the latent space, the feature differences between similar samples are removed as redundant information, and only their core features are retained. Therefore, when the data points are mapped to the latent space, points with similar characteristics are closer, and products with the same optimal aspect will be closer in the latent space.

[0104] 106. Recommend candidate products in the target candidate product set to target users based on interest clustering, interest influencing factors, aspect ratings of each aspect in the target product type, and product ratings.

[0105] In this embodiment, the product recommendation device oriented to surprise degree can determine the attenuation function according to the interest cluster and the aspect rating of each aspect in the target product type, which is specifically described below:

[0106] According to the Ebbinghaus forgetting curve, people will have a deeper memory for things that are mentioned frequently, and thus will be less likely to forget. The target user's interest in the target candidate product can also be reflected in its historical usage records. Just like the Ebbinghaus forgetting curve, for instantaneous preference, the target user will only have a short-term interest and will decay rapidly; for short-term preference, the target user will be interested for a longer time, but it will also decay over time; and long-term preference can keep the target user interested for a long time. Compared with instantaneous preference and short-term preference, it is not easy to be abandoned. However, although most people's memory conforms to the Ebbinghaus forgetting curve, different people forget at different speeds. Usually, people are more mature in a certain aspect, so the relevant memory about that aspect will decay more slowly. At the same time, when the popularity of things is higher, people will find it more difficult to forget the things. Based on this, different decay functions are specified for instantaneous preference, short-term preference and long-term preference respectively. Based on the above decay ideas, different decay functions can be defined for different preference types. Here, the decay functions of instantaneous preference, short-term preference and long-term preference are defined as follows:

[0107]

[0108] in, λ is the forgetting speed, and Inversely proportional, the higher the maturity and popularity, the slower it is forgotten. is the target user’s maturity with respect to aspect k, is the popularity of the target candidate product in aspect k, and Δt is normalized to the range [0, 1]. After obtaining the interest clusters based on each aspect and defining different decay functions, the decay function corresponding to each aspect-based cluster can be determined. For example, if the instantaneous preference of an interest cluster based on a certain aspect in the target product type is defined, the decay function of the aspect is defined as decay instant (Δt) = (1-Δt) λ ∈[0,1].

[0109] It should be noted that the above formula is only for illustration, and other attenuation functions may also be defined without specific limitation.

[0110] After selecting the attenuation function for each aspect, the product recommendation device for surprise can calculate the interest clustering and interest influencing factor according to the attenuation function to obtain the surprise degree of the target user for the candidate products in the target candidate product set at the current moment, as follows:

[0111] Since users have different interest levels for different aspects of the same type of product, including instantaneous preference, short-term preference and long-term preference, different types of preferences will have different decay modes, which will affect the degree of surprise that the candidate product brings to the user at the current moment. Therefore, for different interest clusters, corresponding decay should be made. At the same time, the user's historical behavior will affect the user's interest at the current moment, so t c The surprise metric at the moment (current moment) is the embedding vector of the target candidate product and each interest cluster C based on the optimal aspect K The weighted distance between them is calculated, and the corresponding attenuation and processing are performed according to the aspect preference type to obtain the surprise degree of the candidate products in the target candidate product set:

[0112]

[0113] Among them, decay ype (Δt) is the decay function selected according to the decay type of aspect k, t c is the current recommendation time, t0 is the time when the target user's first comment on this type of product appears, N is the number of aspects of the target candidate product, and w * is the embedding vector of the target candidate product, is the interest impact factor, C K The first part is to calculate the weighted distance between the embedding of the candidate product and other clusters in the latent space according to the clustering algorithm. An appropriate distance means that it is similar to the cluster but has certain differences, which can bring surprises to users. This is also a commonly used method in current work. The attenuation function is selected according to the preference type represented by the cluster, which can make the original surprise degree based on distance calculation have appropriate attenuation, which is also in line with the forgetting law of different preferences of users. Finally, all the previous historical behaviors of the user in the same aspect will affect the current user's surprise degree, either enhancing or reducing it.

[0114] Finally, based on the surprise degree and product rating, the candidate products in the target candidate product set are recommended to the target user. Specifically, the following utility function can be used to recommend the target user: in, is the final recommendation score, β and 1-β represent the product score and surprise weight of the target recommended product, respectively, r u,i Product ratings for target recommended products, Seredipity i,u When the final recommendation score of the target recommended product is greater than a certain threshold, the target recommended product will be recommended to the target user.

[0115] In summary, it can be seen that in the embodiments provided by the present application, when recommending products to users, the products that the users are interested in are preliminarily selected as candidate products in combination with the elasticity factor, and then the product ratings are predicted based on aspect features, user maturity and commodity popularity. In addition, the evolution of different types of aspect preferences is taken into account, and the user's surprise for the candidate products is determined through interest influence factors, interest clustering, user aspect ratings of product types, and attenuation functions. This can achieve more surprising and smaller-granular surprise recommendations for users, thereby improving user experience.

[0116] The present application is described above from the perspective of a product recommendation method oriented toward surprise, and the present application is described below from the perspective of a product recommendation device oriented toward surprise.

[0117] See also Figure 4 , Figure 4 A virtual structural diagram of a product recommendation device oriented toward surprise provided in an embodiment of the present application, wherein the product recommendation device oriented toward surprise 400 includes:

[0118] A determination unit 401 is used to determine, according to the historical behavior record corresponding to the target user, the elasticity factor between the target user and each initial candidate product in the initial candidate product set and the similarity between the historical products in the historical behavior record and each initial candidate product;

[0119] A screening unit 402, configured to screen the initial candidate product set according to the elasticity factor and the similarity to obtain a target candidate product set;

[0120] An influence factor determination unit 403 is used to determine the interest influence factor of each behavior corresponding to the target aspect according to the historical behavior record, wherein the target aspect is any aspect corresponding to the target product type, the target product type is the product type corresponding to the target candidate product, and the target candidate product is any candidate product in the target candidate product set;

[0121] An aspect rating determination unit 404 is used to determine the aspect rating of the target user for each aspect corresponding to the target product type according to the interest influence factor of each behavior corresponding to the target aspect;

[0122] A comprehensive determination unit 405 is used to determine the product score of the target candidate product at the current moment, the embedding vector corresponding to the target candidate product, and the interest cluster corresponding to the target candidate product;

[0123] The recommendation unit 406 is configured to recommend the candidate products in the target candidate product set to the target user according to the interest cluster, the interest influencing factor, the aspect rating of each aspect in the target product type, and the product rating.

[0124] In a possible design, the comprehensive determination unit 405 determines the product score of the target candidate product at the current moment by:

[0125] Determine a current aspect score and a current aspect weight for each aspect corresponding to the target product type at the current moment;

[0126] A product score of the target candidate product is determined according to the current aspect score and the current aspect weight.

[0127] In a possible design, the comprehensive determination unit 405 determines the product score of the target candidate product according to the current aspect score and the current aspect weight, including:

[0128] Determine a basic score of the target candidate product according to the current aspect score and the current aspect weight;

[0129] Determine the maturity of the target user in each aspect corresponding to the target product type and the popularity of the target candidate product in each aspect corresponding to the target product type;

[0130] A product score of the target candidate product is determined according to the basic score, the maturity and the popularity.

[0131] In a possible design, the impact factor determination unit 403 is specifically used for:

[0132] Determine the historical aspect score and historical aspect weight of each aspect corresponding to the target product type according to the historical behavior record;

[0133] The interest influence factor of each behavior corresponding to the target aspect is determined according to the historical aspect score and the historical aspect weight.

[0134] In a possible design, the recommendation unit 406 is specifically used for:

[0135] Determine a decay function for each aspect of the target product type according to the interest cluster and the aspect rating of each aspect of the target product type;

[0136] Determining a target distance between the interest cluster and the embedding corresponding to the target candidate product;

[0137] Calculating the target distance and the interest influence factor based on the attenuation function to obtain the surprise degree of the target user for the candidate products in the target candidate product set at the current moment;

[0138] The candidate products in the target candidate product set are recommended to the target user based on the surprise degree and the product score.

[0139] In a possible design, the screening unit 402 is specifically used for:

[0140] Determining an update factor according to the similarity and the elasticity factor;

[0141] Determining the interest status of the target user for each of the initial candidate products based on the update factor;

[0142] The target candidate product set is determined according to the interest status of the target user in each of the initial candidate products.

[0143] In a possible design, the determining unit 401 is specifically configured to:

[0144] Determine, according to the historical behavior records, the diversity of product categories corresponding to the target user and the diversity of aspects corresponding to the target user;

[0145] Determining user elasticity corresponding to the target user according to the product category diversity and the aspect diversity;

[0146] Determining the user group diversity corresponding to each initial candidate product in the initial candidate product set and the aspect popularity corresponding to each initial candidate product according to the historical behavior record;

[0147] Determining the product elasticity corresponding to the candidate product according to the diversity of the user group and the popularity of the aspect;

[0148] The elasticity factor is determined according to the user elasticity of the target user and the product elasticity corresponding to each of the initial candidate products.

[0149] In a possible design, the aspect rating determination unit 404 is specifically used to:

[0150] Determine the feature sequence of each aspect corresponding to the target product type;

[0151] The interest rating of the target user for each aspect of the target product type is determined according to the feature sequence of each aspect corresponding to the target product type and the interest influencing factor of each behavior corresponding to the target aspect.

[0152] Figure 5This is a schematic diagram of the structure of the application server, such as Figure 5 As shown, the server 500 of this embodiment includes at least one processor 501, at least one network interface 504 or other user interface 503, a memory 505, and at least one communication bus 502. The server 500 optionally includes a user interface 503, including a display, a keyboard or a clicking device. The memory 505 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 505 stores execution instructions. When the server 500 is running, the processor 501 communicates with the memory 505, and the processor 501 calls the instructions stored in the memory 505 to execute the above-mentioned product recommendation method for surprise. The operating system 506 includes various programs for implementing various basic businesses and processing tasks based on hardware.

[0153] The server provided in the embodiment of the present application can execute the technical solution of the embodiment of the above-mentioned product recommendation method oriented to surprise degree, and its implementation principle and technical effect are similar, which will not be repeated here.

[0154] The present application also provides a computer-readable storage medium on which a computer program is stored. When the computer executes the computer program, the method flow related to the product recommendation device for surprise degree in any of the above method embodiments is implemented. Correspondingly, the computer can be the above product recommendation device for surprise degree.

[0155] The present application also provides a computer program or a computer program product including the computer program, which, when executed on a computer, enables the computer to implement the method flow related to the product recommendation device for the degree of surprise in any of the above method embodiments. Correspondingly, the computer may be the above-mentioned product recommendation device for the degree of surprise.

[0156] In the above Figure 1 In the corresponding embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0157] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A product recommendation method oriented to surprise, characterized in that: include: Determine, according to the historical behavior record corresponding to the target user, the elasticity factor between the target user and each initial candidate product in the initial candidate product set and the similarity between the historical products in the historical behavior record and each initial candidate product; Screening the initial candidate product set according to the elasticity factor and the similarity to obtain a target candidate product set; Determine, according to the historical behavior record, an interest influencing factor of each behavior corresponding to a target aspect, wherein the target aspect is any aspect corresponding to a target product type, wherein the target product type is a product type corresponding to the target candidate product, and wherein the target candidate product is any candidate product in the target candidate product set; Determine, according to the interest influencing factor of each behavior corresponding to the target aspect, the aspect rating of the target user for each aspect corresponding to the target product type; Determine a product score of the target candidate product at a current moment, an embedding vector corresponding to the target candidate product, and an interest cluster corresponding to the target candidate product; recommending candidate products in the target candidate product set to the target user according to the interest cluster, the interest influencing factor, the aspect rating of each aspect in the target product type, and the product rating, The step of determining the elasticity factor between the target user and each of the initial candidate products in the initial candidate product set according to the historical behavior record corresponding to the target user includes: Determine, according to the historical behavior records, the diversity of product categories corresponding to the target user and the diversity of aspects corresponding to the target user; determining a user elasticity corresponding to the target user according to the product category diversity and the aspect diversity, wherein the user elasticity represents the user's ability to accept different products; Determining the user group diversity corresponding to each initial candidate product in the initial candidate product set and the aspect popularity corresponding to each initial candidate product according to the historical behavior record; Determining the product elasticity corresponding to the candidate product according to the diversity of the user group and the popularity of the aspect, wherein the product elasticity represents the probability of the product being accepted by different users; The elasticity factor is determined according to the user elasticity of the target user and the product elasticity corresponding to each of the initial candidate products, wherein the elasticity factor is obtained by weighting the user elasticity and the product elasticity. The determining of the interest influencing factor of each behavior corresponding to the target aspect according to the historical behavior record includes: Determine the historical aspect score and historical aspect weight of each aspect corresponding to the target product type according to the historical behavior record; The interest influencing factor of each behavior corresponding to the target aspect is determined according to the historical aspect score and the historical aspect weight.

2. The method according to claim 1, characterized in that Determining the product score of the target candidate product at the current moment includes: Determine a current aspect score and a current aspect weight for each aspect corresponding to the target product type at the current moment; A product score of the target candidate product is determined according to the current aspect score and the current aspect weight.

3. The method according to claim 2, characterized in that Determining the product score of the target candidate product according to the current aspect score and the current aspect weight includes: Determine a basic score of the target candidate product according to the current aspect score and the current aspect weight; Determine the maturity of the target user in each aspect corresponding to the target product type and the popularity of the target candidate product in each aspect corresponding to the target product type; A product score of the target candidate product is determined according to the basic score, the maturity and the popularity.

4. The method according to any one of claims 1 to 3, characterized in that The recommending the candidate products in the target candidate product set to the target user according to the interest cluster, the interest influencing factor, the aspect rating of each aspect in the target product type, and the product rating comprises: Determine a decay function for each aspect of the target product type according to the interest cluster and the aspect rating of each aspect of the target product type; Determining a target distance between the interest cluster and the embedding corresponding to the target candidate product; Calculating the target distance and the interest influence factor based on the attenuation function to obtain the surprise degree of the target user for the candidate products in the target candidate product set at the current moment; The candidate products in the target candidate product set are recommended to the target user based on the surprise degree and the product score.

5. The method according to any one of claims 1 to 3, characterized in that The screening of the initial candidate product set according to the elasticity factor and the similarity to obtain a target candidate product set includes: Determining an update factor according to the similarity and the elasticity factor; Determining the interest status of the target user for each of the initial candidate products based on the update factor; The target candidate product set is determined according to the interest status of the target user in each of the initial candidate products.

6. The method according to any one of claims 1 to 3, characterized in that Determining the aspect rating of the target user for each aspect corresponding to the target product type according to the interest influence factor of each behavior corresponding to the target aspect includes: Determine the feature sequence of each aspect corresponding to the target product type; The interest rating of the target user for each aspect of the target product type is determined according to the feature sequence of each aspect corresponding to the target product type and the interest influencing factor of each behavior corresponding to the target aspect.

7. A product recommendation device oriented to surprise, characterized in that: include: A determination unit, configured to determine, according to the historical behavior record corresponding to the target user, an elasticity factor between the target user and each initial candidate product in the initial candidate product set and a similarity between the historical product in the historical behavior record and each initial candidate product; A screening unit, configured to screen the initial candidate product set according to the elasticity factor and the similarity to obtain a target candidate product set; an influence factor determination unit, configured to determine, based on the historical behavior record, an interest influence factor of each behavior corresponding to a target aspect, wherein the target aspect is any aspect corresponding to a target product type, wherein the target product type is a product type corresponding to the target candidate product, and wherein the target candidate product is any candidate product in the target candidate product set; an aspect rating determination unit, configured to determine an aspect rating of the target user for each aspect corresponding to the target product type according to an interest influence factor of each behavior corresponding to the target aspect; A comprehensive determination unit, used to determine the product score of the target candidate product at the current moment, the embedding vector corresponding to the target candidate product, and the interest cluster corresponding to the target candidate product; a recommendation unit, configured to recommend a candidate product in the target candidate product set to the target user according to the interest cluster, the interest influencing factor, the aspect rating of each aspect in the target product type, and the product rating, The step of determining the elasticity factor between the target user and each of the initial candidate products in the initial candidate product set according to the historical behavior record corresponding to the target user includes: Determine, according to the historical behavior records, the diversity of product categories corresponding to the target user and the diversity of aspects corresponding to the target user; determining a user elasticity corresponding to the target user according to the product category diversity and the aspect diversity, wherein the user elasticity represents the user's ability to accept different products; Determining the user group diversity corresponding to each initial candidate product in the initial candidate product set and the aspect popularity corresponding to each initial candidate product according to the historical behavior record; Determining the product elasticity corresponding to the candidate product according to the diversity of the user group and the popularity of the aspect, wherein the product elasticity represents the probability of the product being accepted by different users; The elasticity factor is determined according to the user elasticity of the target user and the product elasticity corresponding to each of the initial candidate products, wherein the elasticity factor is obtained by weighting the user elasticity and the product elasticity. The determining of the interest influencing factor of each behavior corresponding to the target aspect according to the historical behavior record includes: Determine the historical aspect score and historical aspect weight of each aspect corresponding to the target product type according to the historical behavior record; The interest influencing factor of each behavior corresponding to the target aspect is determined according to the historical aspect score and the historical aspect weight.

8. A computer storage medium, characterized in that: include: The instructions, when executed on a computer, cause the computer to execute the steps of the product recommendation method oriented to surprise level according to any one of claims 1 to 6.

Images