An image recommendation method based on user psychological personality and progressive dissociation

Abstract

The present invention discloses an image recommendation method based on user psychology and personality and progressive dissociation, and belongs to the field of image recommendation. The method first constructs a user high-order strong and weak feature preference network to achieve deep preference characterization of users. Then a shopping psychology-personality extraction capsule based on a sliding window network is proposed to infer the psychological and personality differences of users in purchasing different commodities, and to deeply characterize the psychological and personality characteristics of user shopping behaviors in multiple dimensions. Finally, a progressive preference dissociation method is proposed, which uses click and add-to-cart behavior fusion respectively, and subsequently introduces auxiliary tasks of purchase, like and use shopping psychology and personality as basic tasks, and subsequently introduces auxiliary tasks to derive repurchase preferences. Multiple tasks simultaneously dissociate multiple preference factors to obtain significant user behavior characteristics between different behaviors. Starting from the user behavior sequence, the present invention comprehensively and deeply mines implicit information to improve the accuracy and diversification of image recommendations.

Description

Technical Field

[0001] The present invention relates to the field of image recommendation, and in particular to an image recommendation method based on user psychological characteristics and progressive dissociation. Background Art

[0002] In today's digital age, users generate a large amount of behavior and event data in various online platforms and applications. From browsing products to searching for information, to purchasing goods, users' behavior sequences record their interests, preferences, and needs. In order to better meet users' personalized needs and provide personalized experience, multi-behavior recommendation has become a key research direction in the field of personalized recommendation. Multi-Behavior Recommendation (MBR) comprehensively considers users' behaviors and preferences in multiple behavior sequences to provide users with more accurate recommendations. By deeply mining the characteristics and correlations of user behavior sequences, multi-behavior recommendation provides users with a richer and more diverse recommendation experience. In recent years, due to the rich user information implied in behavior sequences, multi-behavior recommendation has become a hot issue in the research of recommendation systems.

[0003] Psychological factors refer to the psychological characteristics and processes that influence people's use of e-commerce shopping platforms. The goal of the recommendation system is to provide customized recommendations based on users' interests and preferences, while users' psychological and personality characteristics can provide the system with richer information to gain a deeper understanding of users' needs and preferences. On the one hand, by analyzing users' psychological and personality characteristics, the recommendation system can better understand users' preferences and values. For example, some users may prefer to explore new areas, while others may prefer to be conservative and stable. On the other hand, the recommendation system can recommend products or services that are more in line with users' personality characteristics based on their personality characteristics, such as introversion or extroversion, emotional stability or susceptibility. For example, for introverted users, the system can recommend products suitable for independent thinking and relaxation; for emotionally stable users, the system can recommend products that can provide stable emotional support. In addition, understanding users' psychological and personality characteristics can also help improve user satisfaction. The recommendation system can predict users' deeper needs and expectations based on their psychological and personality characteristics, thereby providing recommendation results that are more in line with their preferences and enhancing users' trust and loyalty to the recommendation system.

[0004] Users have different psychology and personality when interacting with different items. The same user tends to be cautious when buying expensive items, but more decisive when buying ordinary items. When buying high-end lipsticks, users tend to constantly compare and collect, while when buying notebooks, the purchase behavior is more likely to occur.

[0005] Existing research has achieved certain results in solving the problem of image features and user data utilization by fusing information in multiple dimensions. However, current research ignores the characteristics between user behaviors. There are still some difficulties in how to utilize the micro-elements between behaviors, how to deeply extract the characteristics of users' deep shopping psychology, personality and derived repurchase preferences implied in user behavior sequences, and how to mine user sensitive characteristics. The main manifestations are:

[0006] Current methods need to be improved in terms of considering the temporal correlation between behaviors and the differences in the intensity of behaviors. The temporal correlation between behaviors includes information such as the order of user behaviors and user preference characteristics, which is very important for accurately understanding user behavior patterns and predicting future behaviors. The intensity differences between behaviors should also be taken into account, because some behaviors may have a greater impact on users' interests and decisions.

[0007] These micro-elements include the relative position of the behavior (for example, the position of the behavior in the sequence), the intensity of the behavior (for example, the number of purchases, etc.), and the correlation between behaviors (for example, other behaviors triggered after purchasing a product). In-depth exploration of these micro-elements will help us better understand the implicit characteristics and deep preferences in the user behavior sequence.

[0008] Traditional methods pay less attention to the normal deviation of user behavior patterns. However, user behavior patterns may deviate over time and in different situations, which may reflect changes in users' recent needs and behavioral responses. Therefore, how to obtain users' recent needs based on the deviation of their recent behavior patterns is an important research direction. By mining this deviation, more accurate and real-time recommendations can be provided.

[0009] To solve the problem of mining and utilizing user multi-behavior sequence data, the main current solutions include graph-based recommendation, deep learning-based recommendation and recommendation based on dissociated representation.

[0010] (1) Graph-based recommendation methods use user multi-behavior data to model and mine user interests and behavior patterns to provide personalized recommendations. Graph-based recommendations take into account the correlation and interactivity between users and items, and use the relationship graph structure between users and items to make recommendations. In graph-based recommendations, users and items can be viewed as nodes in the graph, and the relationship between users and items can be represented as edges in the graph. By constructing and analyzing this relationship graph, information such as user interests, item similarities and associations can be mined, thereby providing users with personalized recommendations. In response to the above problems, people have studied many solutions to analyze and process user and product information from different angles and mine user multi-behavior data.

[0011] A common method is the graph neural network method, which constructs the user behavior sequence into a graph structure and uses GNN to learn the user representation. GNN can capture the interaction relationship and behavior pattern between users through node aggregation and information transmission, thereby providing personalized recommendations. Another method is based on the random walk method, which uses random walks to simulate the user's behavior sequence in the recommendation of multi-behavior sequence data. By performing random walks on the graph and sampling neighbor nodes, more related user behavior information can be obtained and used to recommend candidate items. Some studies use path-based methods to convert users' multi-behavior sequences into path sequences, and make recommendations by analyzing and mining the features of different paths. For example, the user-behavior path can be constructed by counting the co-occurrence frequency in the user behavior sequence, and the correlation of the path can be used for recommendation. There is also a graph embedding-based method that maps nodes in the graph structure to a low-dimensional vector space. Users and items are mapped to vector representations through graph embedding methods, and recommendations are made using distance or similarity in the vector space. Commonly used graph embedding methods include DeepWalk, Node2Vec, and GraphSAGE. In addition, the graph attention network-based method uses the attention mechanism to dynamically learn the relationship between nodes in the graph structure, uses GAT to learn the association between users and behaviors, and uses attention weights to share user representations and capture the importance of behaviors. Another effective method is to make recommendations based on the social network graph structure, using the behaviors and preferences of friends or followers in the user's social network graph to recommend items that the user likes in his social network. For example, considering the mutual influence between users, user behaviors and preferences are extended to other users in their social network. Social network graphs have data sparsity and cold start problems, and considering social noise and false information, they will have a negative impact on recommendation methods based on social network graphs.

[0012] Therefore, the current graph-based recommendation technology still has difficulty guaranteeing recommendation performance when mining and utilizing multi-behavior data. Multi-behavior data of users reveals the interactive relationship between users and items, and implies users' diverse interests and preferences. In the implementation of graph-based recommendation strategies, it is common to represent user behaviors as nodes and aggregate information without paying attention to the characteristics and associations between behaviors to design recommendation methods. These methods pay little attention to deep information such as micro-elements between behaviors and deviations from normalized patterns, and have very limited performance improvements on recommendation systems.

[0013] (2) Methods based on deep learning can discover the latent feature representations hidden in user behavior records, capture the interactive features of nonlinear relationships between users, users and items, and items and items, overcome some obstacles encountered in traditional recommendation technologies, and achieve more accurate recommendations. Recommendation strategies based on deep learning use the powerful representation learning ability of deep learning to learn effective feature representations from user and item data, so that deep learning-based recommendation systems can better understand user needs. Currently, many deep learning methods have been applied to recommendation systems, such as matrix decomposition, reinforcement learning, graph neural networks, and knowledge graphs. Recommendation methods based on deep learning can capture the order and time dependency of behaviors, and dynamically adjust behavior weights through attention mechanisms to focus on important behaviors. Useful feature representations are extracted from sequence data, multiple tasks are processed simultaneously through multi-task learning, and recommendation strategies are optimized using reinforcement learning.

[0014] However, deep learning-based recommendation systems often face some difficulties. First, deep learning models usually require a large amount of data to learn parameters during training. However, in recommendation systems, user behavior data is often very sparse, which may lead to poor generalization of the model and inaccurate recommendation results. Second, deep learning models usually require a large amount of training data to learn user representations. Therefore, in the cold start phase, when there is less historical behavior data for users and items, the model often cannot effectively generate personalized recommendations. Third, deep learning models are usually regarded as black box models, and it is difficult to provide interpretable recommendation results. This means that users and systems cannot accurately understand how the model makes recommendation decisions and lack explanation and understanding of the recommendation results. Fourth, deep learning models usually require a lot of computing resources for training and inference. Training a deep learning model may require a lot of GPU resources and time. For large-scale recommendation systems, this may bring significant computing costs and time overhead. Fifth, deep learning-based recommendation methods pay less attention to deep information such as micro-elements between behaviors, such as correlations between behaviors and deviations from normalized patterns. They cannot effectively identify and adapt to abnormal or novel patterns in user behaviors, and their ability to extract user implicit preferences is very limited. Therefore, how to mine multi-behavior sequence data to improve the accuracy of image recommendation remains a hot topic.

[0015] (3) Recommendation using dissociated representation: Recommendations are made by representing users and items as dissociated (i.e., binary) vectors. This representation method converts the attributes of users and items into binary feature vectors, where each dimension represents a specific attribute or tag. By calculating the similarity or matching degree between users and items, recommendations using dissociated representation can provide users with personalized recommendation results.

[0016] At present, the research on recommendation based on dissociated representation is still in its infancy, but some important research works and methods have emerged. For example, attribute-based dissociated representation method converts the attribute information of users and items into binary feature vectors, where each dimension indicates whether an attribute exists. Another method is hash-based dissociated representation method, which maps users and items into binary vectors through hash functions. These methods aim to convert the attribute information of users and items into binary representations for recommendation.

[0017] Recommendations based on dissociated representations rely on calculating the similarity or matching degree between users and items. In order to evaluate the similarity between binary vectors, researchers have proposed some similarity calculation methods, such as Hamming distance, Jaccard similarity, and cosine similarity. These methods can quantify the similarity between dissociated vectors and are used to discover similar users or similar items in recommendation systems.

[0018] For recommendations based on dissociated representations, researchers have proposed different recommendation algorithms. For example, the proximity-based dissociated recommendation method uses the dissociated vectors of users and items to perform proximity calculations to find similar users or items. Another method is hash-based dissociated recommendation, which maps users and items to binary vectors through hash functions and uses similarity calculation methods to make recommendations.

[0019] Recommendations based on dissociated representations face a trade-off between diversity and accuracy. Since the dissociated vector can only indicate whether an attribute exists, but cannot indicate the specific value of the attribute, it may lead to a lack of diversity in the recommendation results. To solve this problem, some research works have explored methods for diverse recommendations, such as introducing noise and randomness to increase the diversity of recommendation results.

[0020] Dissociated representation recommendation methods have been applied in some specific application scenarios. For example, anonymous dissociated data-based recommendations are used to protect user privacy, and binary representation-based recommendations are used to process large-scale data sets. These application scenarios put forward specific requirements and challenges for dissociated representation recommendations.

[0021] Although the recommendation of dissociated representation has potential in some aspects, it also faces some challenges. For example, the problem of feature loss and data sparsity still exists, and the processing of high-dimensional attributes and large-scale data sets still needs further research and improvement. Therefore, how to apply dissociated representation to recommendation systems still needs further research. Summary of the invention

[0022] The purpose of the present invention is to provide an image recommendation method based on user psychological characteristics and progressive dissociation, which can comprehensively and deeply mine implicit information from the user behavior sequence and improve the accuracy and diversity of image recommendation.

[0023] To achieve the above object, the present invention provides the following solutions:

[0024] An image recommendation method based on user psychological characteristics and progressive dissociation, comprising:

[0025] Collect a collection of projects containing various types of e-commerce images;

[0026] A multi-channel convolutional neural network is used to extract image features in the project set to obtain project representations of each project;

[0027] Convert the behavior data of the user to be recommended into a knowledge graph, and use a graph neural network to mine the high-order feature preferences of users and behaviors in the knowledge graph to obtain a high-order feature preference network of the user to be recommended;

[0028] Calculate the behavior intensity of each behavior of the recommended user based on the behavior frequency and behavior weight;

[0029] According to the strength of the behavior, all behaviors of the user to be recommended are divided into main behaviors and auxiliary behaviors, and the characteristics of the main behavior interaction items are defined as strong features, and the characteristics of the auxiliary behavior interaction items are defined as weak features;

[0030] The attention mechanism based on associated behaviors extracts the association relationship between behaviors from the behavior data of the recommended user and captures the derived repurchase preference;

[0031] According to the strong features, weak features, the correlation between behaviors and the derived repurchase preferences, the first user representation is obtained by using the high-order feature preference network;

[0032] Constructing shopping psychology-personality extraction capsules based on sliding window networks;

[0033] After extracting the global feature representation of the behavior data by using the sliding window network, the global feature representation is input into the shopping psychology-character extraction capsule to obtain the psychological and character fusion features of the user to be recommended;

[0034] According to the psychological and personality fusion characteristics of the user to be recommended, a progressive preference dissociation method is used to determine the second user representation; the progressive preference dissociation method includes a first branch and a second branch, the first branch is used to take the fusion of clicks and add to cart as the basic task, and gradually introduce purchase behavior and liking behavior; the second branch is used to take the fusion of user shopping psychology and personality as the basic task, and introduce the derived repurchase preference; the dissociation results of the first branch and the second branch are aggregated to obtain the second user representation;

[0035] Aggregate the first user representation and the second user representation to obtain a final user representation;

[0036] The inner product operation is performed on the final user representation and the item representation of each item to obtain the user's preference prediction score for each item.

[0037] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0038] The present invention discloses an image recommendation method based on user psychology and progressive dissociation. According to user multi-behavior sequence data, the method comprehensively and deeply mines implicit information, pays attention to the multi-dimensional characteristics of the project, constructs a user's high-order feature preference network, defines the user's main behavior and auxiliary behavior, and realizes the user's deep preference representation; proposes a shopping psychology-personality extraction capsule based on a sliding window network, and deeply represents the user's shopping behavior psychology and personality characteristics in multiple dimensions; uses a progressive preference dissociation method to obtain significant user behavior characteristics between different behaviors and dissociate multiple preference factors. Starting from the user behavior sequence, the present invention comprehensively and deeply mines implicit information to improve the accuracy and diversification of image recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0040] Figure 1 A flowchart of an image recommendation method based on user psychological characteristics and progressive dissociation provided by an embodiment of the present invention;

[0041] Figure 2 The overall framework diagram of the image recommendation method based on user psychological characteristics and progressive dissociation provided by an embodiment of the present invention;

[0042] Figure 3 A schematic diagram of the construction process of a high-order feature preference network provided by an embodiment of the present invention;

[0043] Figure 4 A user personality diagram provided by an embodiment of the present invention;

[0044] Figure 5 A schematic diagram of the parallel progressive preferential dissociation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] The present invention proposes a recommendation system based on user psychological-character extraction and progressive disentangled (PCPD), which aims to solve the problems that image recommendation pays more attention to micro-elements between behaviors, does not deeply explore the characteristics of users' deep shopping psychology, personality and derived repurchase preferences implied in user behavior sequences, pays less attention to the deviation of users' normalized behavior patterns, and ignores the mining of users' sensitive characteristics. Starting from the user behavior sequence, the implicit information is comprehensively and deeply mined to improve the accuracy and diversity of image recommendation.

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] like Figure 1 As shown, the embodiment of the present invention provides an image recommendation method based on user psychological characteristics and progressive dissociation, including:

[0049] Step 1: Collect a collection of items containing various types of e-commerce images.

[0050] Collection item set V = {v 1 ,v 2 ,…,v m}, where V represents the entire project set, including m items, which contains various types of e-commerce images.

[0051] The collection user set U = {u 1 ,u 2 ,…,u n}, where U represents the entire user set, including n individual users.

[0052] Step 2: Use a multi-channel convolutional neural network to extract image features in the item set and obtain the item representation of each item.

[0053] Use multi-channel CNN to extract image features of various e-commerce images in step 1. i ,y i 、z iCorresponding to the "type", "color" and "style" features of the image respectively, we get the feature D of image i i for:

[0054] D i =(x i ,y i ,z i )

[0055] The product attribute sequence is represented as [D 1 ,D 2 ,…,D M ], where M is the length of the attribute sequence. Through the embedding layer, it is converted into a product attribute embedding sequence, the continuous vector representation of the product is extracted, and the matching score between the end user and the project is calculated, which is expressed as Get the project representation i for:

[0056]

[0057] Among them, ω j By performing softmax calculation:

[0058]

[0059] In the formula, e i represents the item representation of the i-th item, d j represents the jth feature attribute in the product attribute embedding sequence, ω j Represents attribute d j The weight of M is the number of characteristic attributes in the product attribute embedding sequence, d m represents the mth characteristic attribute in the product attribute embedding sequence, and exp represents the exponential function.

[0060] Step 3: Convert the behavior data of the user to be recommended into a knowledge graph, and use a graph neural network to mine the high-order feature preferences of users and behaviors in the knowledge graph to obtain a high-order feature preference network of the user to be recommended.

[0061] Use multi-level attention to mine multi-behavior data and build a user's high-order feature preference network, such as Figure 3As shown in Figure 1. First, the preprocessed user behavior data is converted into the form of a knowledge graph. In the graph, each user and each behavior is a node, and the connection edge between the user and the behavior indicates that the user has performed a certain behavior. The weight of the connection can reflect the strength of the user's behavior, such as the number of purchases, clicks, etc. Such a graph can help better understand the associations and patterns between user behaviors. When constructing a user high-order feature preference network, the graph neural network learning method is used to mine the behavior patterns and preference relationships between users from the knowledge graph, and node representation learning is performed on the graph, and information is transmitted through the connection edges between nodes. In this network, both user nodes and behavior nodes have initialized feature representations, that is, each node is represented by a vector. Then, through the stacking and propagation of the graph neural network, the features of the user node will gradually evolve, including the user's high-order preference features for multiple behaviors. These features can be obtained by aggregating the information of neighboring nodes. For example, the neighbors of the user node are the behavior nodes that it has performed. By aggregating the features of these behavior nodes, the user's preference representation for different behaviors can be obtained. The user's deep preference representation is a high-dimensional vector that contains the user's strong and weak preference information for various behaviors. At the same time, it can capture the mutual influence between users, so these representations can also reflect the similarities and social relationships between users.

[0062] Convert user behavior data into knowledge graph form as the basis for mining user information and building a user high-order feature preference network:

[0063] G=(V,E,Φ)

[0064] Where G represents the knowledge graph. V represents a set of nodes, each node represents a user or a behavior. E represents a set of edges, which represents the connection between users and behaviors, and each edge represents that the user has performed a certain behavior. Φ represents the weight of the edge, which is used to reflect the strength of the user's behavior, such as the number of purchases, clicks, etc.

[0065] Mining user information from the knowledge graph, constructing a user high-order feature preference network. For each user, the project and project features are network nodes, and the user preference strength is the edge weight:

[0066] (1) Set the node feature initialization expression as:

[0067] X=x 1 ,x 2 ,…,x n

[0068] Where X represents the node feature matrix, and each row represents the feature vector of a node, such as user features or behavior features.

[0069] (2) Using graph neural networks for node representation learning:

[0070] X'=f(X,A,Ψ)

[0071] Where X' represents the learned node features, that is, the high-order feature preferences of users and behaviors. f is the graph attention mechanism, an aggregation function of the graph neural network, which is used to aggregate the information of neighboring nodes. A represents the adjacency matrix, which is used to describe the connection relationship between nodes and is used to transfer information in the aggregation function. Ψ represents the parameters of the network, which need to be learned through training.

[0072] Step 4: Calculate the behavior intensity of each behavior of the recommended user based on the behavior frequency and behavior weight.

[0073] The behavior intensity is calculated based on the behavior frequency, behavior interaction and behavior weight, and the user behavior is divided into main behavior and auxiliary behavior based on the behavior intensity. The corresponding multi-dimensional strong features and weak features are mined, and the image mutual information is maximized to obtain high-order information based on graph representation.

[0074] Behavior frequency: the number of times or frequency that a user performs a certain behavior, such as the number of purchases or browsing times. Behavior frequency can directly reflect the user's preference for a certain behavior.

[0075] Behavior weight: Different behavior types have different importance. For example, purchasing behavior is usually more valuable than browsing behavior. When calculating behavior strength, different behavior types are given different weights to reflect the higher importance users attach to certain behaviors.

[0076] When calculating the intensity of a behavior, we calculate the intensity of each user for each behavior type based on the behavior frequency and behavior weight indicators. For example, for purchase behavior, we can count the number of times each user purchases a product or the cumulative purchase amount. For browsing behavior, we can count the number of times each user browses. For interactive behavior, we count the number of times or frequency of interactions between users and other entities. Then, we normalize the intensity of the behavior based on the calculation indicators to get the intensity score of the user for various behaviors.

[0077] Definition 1: Primary behavior. The behavior in the user interaction sequence that explicitly or relatively better reflects the user's preference, such as purchase, comment, etc. If the user does not purchase an item, then collecting it is relatively the primary behavior.

[0078] Definition 2: Auxiliary behavior: The behavior in the user interaction sequence that relatively weakly reflects the user's preference, such as click, browse, etc.

[0079] First, for each behavior type, calculate the user's behavior intensity. For example, calculate the number of times or frequency of a user's execution of a certain behavior type as the behavior frequency intensity. Then take the calculated behavior intensity into consideration and distinguish the user's main behavior from the auxiliary behavior based on the following factors:

[0080] Behavior frequency: The main behavior is usually the behavior that users perform frequently, so the behavior with a higher frequency may be the main behavior.

[0081] Behavior weight: Purchase, rating and other behavior types have higher weights. Behaviors with higher behavior weights may be the main behaviors.

[0082] Combining these factors and comprehensively considering the indicators of behavior intensity, we use the weighted average method to make a judgment, that is, to weight the behavior frequency and behavior weight to obtain a comprehensive behavior intensity indicator to divide the main behavior and auxiliary behavior.

[0083] The user's behavior intensity for each behavior type is calculated based on the behavior frequency and behavior weight, and the behavior frequency α is calculated using the following formulas: freq and behavior weight α weight :

[0084]

[0085] α weight =α(k)

[0086] Among them, t freq Indicates the number of times a user performs a specific behavior type, T freq represents the total length of observation time. k is the user behavior type, and α represents the nonlinear activation function for user behavior type k.

[0087] Step 5: Based on the strength of the behavior, all behaviors of the user to be recommended are divided into main behaviors and auxiliary behaviors, and the features of the main behavior interaction items are defined as strong features, and the features of the auxiliary behavior interaction items are defined as weak features.

[0088] By analyzing the user's behavior data in the knowledge graph, the user's primary behavior and secondary behavior are defined according to the frequency and persistence of the behavior. The primary behavior is the user's frequent behavior, while the secondary behavior is the user's occasional or less frequent behavior, that is, the user's behavior reflects the strength of the user's preference for item features. For example, on an e-commerce platform, the user's purchase of goods is the primary behavior, while browsing goods and adding to the shopping cart are secondary behaviors.

[0089] The behavior intensity is used to distinguish user behavior preferences into strong features fs and weak features fw. Similarly, the type, color, and style features of the items are mined through the existing multi-channel CNN method as multi-dimensional strong and weak features corresponding to the main and auxiliary behaviors of the user. The user's high-order feature preference network is introduced to mine the user's item feature dimension preferences and obtain image information for dimensional migration and fusion recommendation. Figure 3 The node e in .

[0090] Taking into account the indicators of behavior strength, a weighted method is used to make a judgment, and the user's main behavior and auxiliary behavior are divided. α≥0.5 is the main behavior, and α<0.5 is the auxiliary behavior. The interaction item feature of the user's main behavior is the user's strong feature f s , the auxiliary behavior interaction item feature is the user's weak feature f w The user's preference strength for an item is the weight of the item's features.

[0091] α=∑α freq +α weight

[0092] Step 6: The attention mechanism based on associated behaviors extracts the association relationship between behaviors from the behavior data of the recommended user and captures the derived repurchase preference.

[0093] The attention mechanism based on associated behaviors targets multiple user behavior sequences, mines users' associated behaviors and candidate preferences, extracts the associations between behaviors, captures users' derived repurchase preferences, provides more personalized and accurate recommendation services, increases user loyalty, and promotes users' continued purchase and consumption behaviors. Derived repurchase preferences refer to users' continued preference for previously purchased products or services, that is, users maintain interest in previously purchased goods or related goods and are willing to buy or try similar products again.

[0094] Definition 3: Association relationship. Mining behaviors in user behavior sequences that can reflect user preference associations. For example, if a user first collects a yellow backpack and then buys a yellow water cup, there is an association relationship between the collection of the yellow backpack and the purchase of the yellow water cup, and yellow is their association feature.

[0095] After purchasing some items, they can often be used for a long time, and at this time the user does not need such goods and their complementary products. Blindly making recommendations based on the user's short-term preferences is often inaccurate and may even cause user disgust. Therefore, it is necessary to capture the user's temporarily disappearing preferences and temporarily reduce this part of the recommendations.

[0096] Mining behaviors that reflect user preference associations in user behavior sequences. For example, if a user first collects a yellow backpack and then buys a yellow water cup, there is an association between the yellow backpack collection behavior and the yellow water cup purchase behavior, and yellow is their association.1 and behavior b 2 The correlation relationship q and the derived repurchase preference d are obtained by the following formula:

[0097]

[0098]

[0099] where β 1 For behavior b 1 The weight, β 2 For behavior b 2 The weight of . and Behavior b 1 and behavior b 2 The embedding representation.

[0100] Step 7: Based on the strong features, weak features, associations between behaviors and derived repurchase preferences, the high-order feature preference network is used to obtain the first user representation.

[0101] Predict user preferences based on the strong and weak features of the items with which the user interacts obtained in step 5, and obtain the user embedding representation of the high-order feature preference network e u for:

[0102]

[0103] Among them, γ s , γ w They are strong features f s and weak feature f w The weight of .

[0104] Definition 4: Temporary disappearance preference. For the item v that the user has purchased n If the user does not click or browse in the next purchase behavior, the product v n is considered to be the user's temporarily disappearing preference. Then according to the formula Update the first user representation; where, The first user after the update indicates that For product v n Users said.

[0105] User interests change over time, so the dynamic strong and weak features and candidate sequences of users are constructed based on the user high-order feature preference network to better understand the user's behavior patterns and preferences, and provide a basis for applications such as personalized recommendations and user profiling. Using the trained user high-order feature preference network, each user's behavior sequence and other features are forward propagated to obtain the user's high-order feature preference vector. According to the user's high-order feature preference vector, combined with other user features, the user's dynamic strong and weak features are constructed. These features include the user's preference for different behavior types, the behavior trend in the recent period, etc. In order to provide personalized recommendations to users, a candidate sequence is constructed. The candidate sequence is a set of candidate products, which are potential recommendation objects of the recommendation system. Potential candidates are screened from the product pool based on the user's historical behavior, dynamic strong and weak features, and other contextual information. When the user interacts with the item, the user's strong and weak features change dynamically, and the candidate feature sequence is also updated in real time to obtain the user representation.

[0106] Step 8: Construct shopping psychology-personality extraction capsule based on sliding window network.

[0107] Construct a shopping psychology-personality extraction capsule based on a sliding window network to identify the psychological and personality representations based on the user behavior sequence level, infer the psychological and personality differences of users when purchasing different products, and characterize the user's shopping psychology and personality characteristics in multiple dimensions and depth. In general, the process of extracting user shopping psychology and personality by capsule network includes constructing a sliding window sequence, designing a sliding window network, constructing shopping psychology-personality feature extraction capsules, feature fusion and personality extraction, application, and recommendation generation. This process can extract the user's shopping psychology and personality characteristics, providing a basis for applications such as personalized recommendations and user profiling. The design of the capsule network enables it to capture the hierarchy and correlation between features, thereby performing well in feature extraction tasks. User shopping psychology is divided into the following categories:

[0108] Demand-based. Users’ purchasing behavior is usually driven by their actual needs for specific products. This demand psychology may cover basic life needs, work needs, entertainment needs, etc. Users’ interactions with daily items are often demand-based.

[0109] Desire type. The user's purchasing behavior may also be influenced by psychological desires and pursuits. For example, users may make purchases because they pursue fashion, personalization, or satisfy vanity. The user's interaction with some luxury goods is often desire-based.

[0110] Value-perceived. The user's perception of the balance between the price of a product and its perceived value will affect the purchase decision. The psychology of value perception may involve price sensitivity, cost-effectiveness, etc. By observing the categories, price ranges, and frequency of products purchased by users, we can infer the user's perceived value of the product. For example, frequent purchases of high-priced products may indicate that the user attaches importance to quality and value.

[0111] Emotional. User emotions play an important role in purchasing behavior. Positive emotions may prompt users to make purchasing decisions, while negative emotions may hinder purchases. By tracking users' purchasing behavior and comments over a period of time, you can analyze users' emotional trends, such as whether they gradually develop more positive or negative emotions towards a brand or product.

[0112] Social. Social factors also affect users’ shopping psychology. Users may choose to buy specific products based on social factors such as recommendations from others and reviews on social media. Behaviors that interact through recommendations from other users are often social.

[0113] like Figure 4 As shown, the user purchasing personality is divided into the following categories:

[0114] Decisive type. Decisive customers are decisive, straightforward, and have their own opinions. In the purchase process, they have clear goals, are proactive, and buy products according to their own intentions.

[0115] Impulsive type. They have strong emotional reactions and are easily influenced by various factors. The decision to buy or not is often dominated by emotions, and they will change their purchase decision in a short time if they are not satisfied.

[0116] Rational type. They like to think carefully and weigh the pros and cons of products rationally before buying. They will not buy easily before knowing all aspects of the product. They spend a relatively long time buying and choose products carefully.

[0117] Hesitant. Weak-willed, no imagined products, just collect them first. When buying, they often compare prices from different stores and spend a lot of time comparing.

[0118] Undetermined. Consumers’ purchasing intentions are uncertain and their purchasing behavior is quite random. Customers with this purchasing attitude usually lack purchasing experience and relevant product knowledge, so their purchasing psychology is not stable.

[0119] In order to process user multi-behavior sequence data, a sliding window network (SWN) is proposed to extract local user psychology and personality by using a sliding window on the sequence. Specifically, the sliding window network extracts local features within the window by sliding a fixed-size window on the input sequence, and uses these local features to capture the patterns and structures in the sequence. First, the user's behavior data is organized into a sliding window sequence in chronological order. The sliding window is a time window of fixed length that slides along the time axis, one time step at a time, to form a new sequence. The sliding window sequence can capture the user's behavior patterns and preferences in different time periods, and provide information of multiple time steps for subsequent feature extraction. The sliding window network allows the network to pay different degrees of attention and modeling to different parts of the sequence, thereby better understanding the sequence data. The sliding window network consists of two main components: a sliding window extractor and a psychological-personality feature fusion.

[0120] Definition 5: Sliding Window Extractor. The sliding window extractor is a recurrent neural network (RNN) module that is used to extract local features within a sliding window. The convolutional layer is used to process the data within the window to capture local features.

[0121] Definition 6: Psychological-Personality Feature Fusor. The psychological-personality feature fusor is used to fuse and integrate the local features extracted by the sliding window extractor to generate a global representation or sequence-level features. This paper uses weighted summarization of local features to generate a global representation or sequence-level features.

[0122] The design of sliding window network involves the following aspects:

[0123] Window size: The window size determines the scope of local contextual information that the model can capture. Smaller windows can capture local features in a more fine-grained manner, while larger windows can provide a broader context. The choice of window size should be weighed against the requirements of the task and the length of the sequence data.

[0124] Window step size: The window step size determines the distance the sliding window moves on the sequence. A smaller step size can provide more window overlap and increase the richness and diversity of features. A larger step size can reduce the computational complexity and computing resource consumption of the model.

[0125] The user shopping psychology and personality vectors output by the sliding window network are fed into the capsule network as the input of the capsule network. The capsule network will further encode and extract these feature representations to obtain more meaningful and discriminative features. Construct psychology and personality extraction capsules to extract psychology and personality representations at the user behavior sequence level. The capsule network consists of multiple capsule units, each of which is used to detect specific user features.

[0126] In the initial stage of the capsule network, low-level features are extracted through a series of convolutional layers and fully connected layers, and these features are organized into primary capsules. Primary capsules are a group of capsules, each of which represents a low-level feature, and each low-level feature is represented as a vector. The capsule network calculates the connection weights between capsules through negotiated dynamic routing to determine the importance and contribution of information transfer. During the routing process, the capsules measure the similarity between them by calculating the dot product between the vectors. Then the softmax function is used to convert these similarities into weights to allocate the weights of information transfer. This process is similar to a dynamic "negotiation", where different capsules "negotiate" weights by calculating correlations to determine the importance and contribution of information transfer. Dynamic routing updates the connection weights between capsules through multiple rounds of iterations. Each round of iteration recalculates the similarity between capsules and updates the weights based on the similarity. In each round of iteration, the weights between capsules are continuously adjusted based on their similarity so that more meaningful feature capsules receive more attention and weight. The number of iterations is usually pre-set and can be adjusted according to the complexity of the task and dataset. In multiple rounds of dynamic routing, the weights between capsules are continuously updated to ensure that more meaningful feature capsules get more weight. As the number of iterations increases, the weights between capsules gradually converge, and feature capsules are better encoded and extracted. Ultimately, the capsule network outputs a feature vector composed of high-level feature capsules that represent different high-level features.

[0127] Design a sliding window network to mine user behavior embedding, extract user shopping psychology and personality, and obtain user representation based on shopping behavior psychology and personality. t represents the input vector of the input layer at time t, o t represents the output of the input layer at time t, s t represents the hidden state of the neuron at time t. o represents the output, s represents the hidden layer, x represents the input, V, W, and U represent the weight matrices respectively. In the sliding window network, the weights of W, U, and V are shared and are the same at every moment. At time t in the sliding window, the following calculation is performed:

[0128] h t+1 =Ux t+1 +Ws t

[0129] s t+1 =f(h t+1 )

[0130] o t+1 =g(Vs t+1 )

[0131] The sliding window is set to 50, and the window step is 10. Define the user's shopping psychology and shopping personality. As time goes by, the sliding window extractor extracts the local psychology and personality features in the current window in turn, capturing the user's shopping psychology and personality performance in the short term, so as to better understand the user's attitude, emotion and behavior during the shopping process. Such local feature extraction can help identify the user's current shopping motivation, preference and personalized purchase intention.

[0132] The psychological-personality feature fuser aggregates local features using weights to produce a global representation as input to the capsule network. Where |P| is the input dimension:

[0133] g=∑o t+1

[0134] For the input g of the capsule network, each capsule unit first calculates the prediction vector, is the prediction vector of the capsule unit, and the softmax function is used to normalize the activation prediction value s, which is calculated by the following formula:

[0135] s=softmax(o)*g

[0136] A nonlinear “squashing” function is used to scale the activation prediction value s obtained in step 14 to a suitable range, and output a representation of the final user’s psychological and personality characteristics:

[0137] r=squash(s)*g

[0138] Step 9: After extracting the global feature representation of the behavior data using the sliding window network, input the shopping psychology-personality extraction capsule to obtain the psychological personality fusion features of the user to be recommended.

[0139] The object of the sliding window network processing is the user's multi-behavior sequence. The process of obtaining the user's multi-behavior sequence involves steps such as data collection, preprocessing, behavior sequence construction and feature extraction. First, collect user behavior data from the e-commerce platform, including user clicks, browsing, purchases, comments, favorites, sharing and other behaviors. Data collection can be carried out through API interfaces, crawler technology or data authorization. The collected data contains information such as user ID, behavior type, behavior timestamp, etc. After collecting the data, the original data is preprocessed to ensure the quality and consistency of the data. The preprocessing process includes data cleaning, deduplication, processing missing values, and removing abnormal data to ensure the accuracy of subsequent analysis. The process of constructing user multi-behavior sequences is to organize the multiple behaviors of each user into sequences in chronological order. The specific method is to sort the behavior data according to user ID and timestamp, and then organize the behaviors of the same user into sequences in chronological order. Since the length of the user's behavior sequence varies depending on the user's activity level and the data collection time, in order to unify the length of the sequence, the sequence is truncated or padded. Truncation refers to removing redundant behaviors in the sequence so that the length of all sequences is the same. Padding refers to adding specific placeholders or markers at the end of the sequence so that all sequences have the same length.

[0140] The user psychological personality extracted by the shopping psychology-personality extraction capsule based on the sliding window network is combined with the user fine-grained sequence knowledge graph to construct a behavior sequence-shopping psychological personality binary knowledge graph, which consists of two parts: the user high-order feature preference network and the user purchase behavior knowledge graph. The two knowledge graphs are connected through the same user u, such as Figure 5 As shown. n1 ,…,d ni} is the mined user u n The derived repurchase preference node, {v n1 ,…,v nl} is the mined user u n The purchase character node, {p n1 ,…,p nj} is the mined user u n For example, by comparing capsule networks, we can predict the user u in the multi-behavior fine-grained sequence knowledge graph. 1 The derivative purchase preference is white French dress d 11 , black casual style short sleeves 12 And white ins style crossbody bag d 13 . User u 1 When buying, they like to buy luxury goods. 1 The shopping psychology of 11 ; During the interaction process, the goal is clear, and the purchase behavior is more likely to occur. 1The purchasing personality is decisive type 11 , then user u 1 Build 11 d 12 d 13 、p 11 and v 11 node. Figure 5 In the figure, Binary knowledge graph represents binary knowledge graph.

[0141] Step 10: According to the psychological and personality fusion characteristics of the user to be recommended, a progressive preference dissociation method is used to determine the second user representation; the progressive preference dissociation method includes a first branch and a second branch, the first branch is used to take the fusion of clicks and add to cart as the basic task, and gradually introduce purchase behavior and liking behavior; the second branch is used to take the fusion of user shopping psychology and personality as the basic task, and introduce the derived repurchase preference; the dissociation results of the first branch and the second branch are aggregated to obtain the second user representation.

[0142] Progressive preference dissociation is divided into two branches. In the first branch, click and add to cart are integrated as the basic task. Based on user representation and product representation, the results of click and add to cart are predicted, which can be expressed as:

[0143] y basic1 =f basic1 (E c ,E a ,u,i)

[0144] The embedding sequences of click and add to cart are represented as N c 、N a are the number of corresponding behaviors, f basic1 ( ) represents the nonlinear activation function used to predict click and add-to-cart results.

[0145] Introduce the purchase task into the result y basic1 , predict the result of introducing purchase behavior, expressed as:

[0146] y b =f b (y basic1 ,E b ,u,i)

[0147] The purchased embedding sequence is expressed as N b is the number of purchases, f b ( ) represents the nonlinear activation function used to predict the purchase behavior outcome.

[0148] Introducing the liking task into the result y b , predict the result of introducing liking behavior, and get the dissociation result y of the first branch task 1 , expressed as:

[0149] y 1 =f f (y b ,E f ,u,i)

[0150] The preferred embedding sequence is expressed as N f is the number of liking behaviors, f f ( ) represents the nonlinear activation function used to predict the outcome of liking behavior.

[0151] Similarly, in the second branch, the user shopping psychology and personality fusion representation r is used as the basic task. Based on the user representation and product representation, the results of predicting shopping psychology and personality can be expressed as:

[0152] y basic2 =f basic2 (r,u,i)

[0153] Introduce the derived repurchase preference d task into the result y basic2 , predict the result of introducing the derived repurchase preference, and obtain the dissociated result y of the second branch task 2 , expressed as:

[0154] y 2 =f d (y basic2 ,d,u,i)

[0155] Among them, f basic2 ( ) represents the nonlinear activation function used to predict purchasing psychology and personality results, f d ( ) represents the nonlinear activation function used to predict the derived repurchase preference results.

[0156] By connecting the branches of each task to form a multi-task learning structure, the input features are shared and the corresponding task result predictions are output:

[0157]

[0158] Parallel progressive preference dissociation gradually introduces new tasks according to their importance in personalized recommendations, so that the model can gradually learn more information. The basic tasks are integrated using click and add-to-cart behaviors, and then auxiliary tasks such as purchase, like, and use shopping psychology and personality are introduced as basic tasks. Auxiliary tasks are then introduced to derive repurchase preferences. At the same time, a branch is added to the model to share input features and output corresponding task result predictions. After introducing auxiliary tasks for all preference factors, the model becomes a multi-task learning structure, dissociating multiple preference factors at the same time. By training all tasks together, the model gradually learns the correlation and feature representation between different tasks.

[0159] Step 11: Aggregate the first user representation and the second user representation to obtain a final user representation.

[0160] The user representation obtained through the user high-order feature preference network and user representations obtained through progressive preference dissociation Aggregate and get the final user representation e u :

[0161]

[0162] Step 12: Perform inner product operations on the final user representation and the item representation of each item to obtain the user's preference prediction score for each item.

[0163] The user is represented by u and the item representation extracted by multi-channel CNN i Perform the inner product operation to get the final prediction score:

[0164]

[0165] In the formula, represents the preference prediction score of user u for the i-th item.

[0166] The overall framework of the PCPD method of the present invention is as follows Figure 2 As shown in the figure, it is divided into four steps: (1) User high-order feature preference network, which mines multi-behavioral data through multi-level attention and distinguishes users' strong and weak feature preferences; (2) Shopping psychology-personality extraction capsule based on sliding window network, which mines user behavior embedding through designing sliding window network, extracts user shopping psychology and personality, and obtains user representation based on shopping behavior psychology and personality; (3) Progressive preference dissociation, constructs behavior sequence-shopping psychology personality binary knowledge graph, dissociates it progressively, and adaptively migrates it to other projects; (4) Prediction layer, aggregates user and project representations, and outputs prediction scores.

[0167] Figure 2In the above, item embedding means item embedding, user embedding means user embedding, User highorder feature preference network means user high-order feature preference network, Multi-level attention means multi-level attention, Progressive Disentangled means progressive preference disentanglement, character, psychology, derivative preference means character, psychology, derivative purchase preference, character-psychology extraction capsule means character psychology, character extraction capsule, embedding layer means embedding layer. MCNN means multi-channel convolutional neural network, i 1 、i n Indicates the first and nth pictures, c 1 、c n 、a 1 、a n 、b 1 、b n 、f 1 、f n Represents user behavior data, and t represents time.

[0168] The present invention first constructs a user multi-behavior fine-grained sequence knowledge graph, focuses on user micro-behaviors, defines user main behaviors and auxiliary behaviors based on the intensity of user behaviors, constructs a user high-order strong and weak feature preference network, and realizes the deep preference characterization of users. Then, a shopping psychology-personality extraction capsule based on a sliding window network is proposed, and a sliding window network is designed to process sequence data, extract user psychology and personality, and construct shopping psychology and shopping personality extraction capsules to infer the psychological and personality differences of users when purchasing different commodities, and deeply characterize the psychological and personality characteristics of user shopping behaviors in multiple dimensions. Finally, a progressive preference dissociation method is proposed, which uses the fusion of click and add-to-cart behaviors respectively, and subsequently introduces auxiliary tasks of purchase, liking and using shopping psychology and personality as basic tasks, and subsequently introduces auxiliary tasks to derive repurchase preferences. Multiple tasks simultaneously dissociate multiple preference factors to obtain significant user behavior characteristics between different behaviors.

[0169] Referring to the ideas of RNN and sliding window algorithm graph convolutional network, a sliding window network is proposed. The representation vector of nodes is learned through information transmission and aggregation operations, so as to capture the relationship and dependency between nodes, and the system establishes a user high-order feature preference network. Through multi-level attention mining of multi-behavior data, user behaviors are divided into main behaviors and auxiliary behaviors according to the interactive items and behavior intensity of different user behaviors, and the multi-dimensional strong and weak features of the items are mined. The high-order information based on graph representation is obtained by maximizing the image mutual information. A sliding window network is proposed to process user multi-behavior sequences, and combined with a capsule network, a psychological and personality extraction capsule is constructed to extract the psychological and personality representation of the user behavior sequence level. The user psychology and personality extracted by the shopping psychology-personality extraction capsule based on the sliding window network are combined with the user fine-grained sequence knowledge graph to construct a behavior sequence-shopping psychology personality binary knowledge graph, and it is dissociated to dissociate the user's interest in different items and obtain more accurate preferences.

[0170] The key point of this method is to consider the impact of information between multiple behavioral data on user preferences, use the shopping psychology-personality extraction capsule based on the sliding window network to extract the user's psychological and personality characteristics, and propose a progressive dissociation method to dissociate multiple preference factors to improve the accuracy and adaptability of the recommendation system. The main key points are:

[0171] (1) A user high-order feature preference network was constructed. A user multi-behavior fine-grained sequence knowledge graph was constructed to capture the user's micro-behavior and more comprehensively understand the user's behavior patterns and preferences. The user's behavior intensity was introduced into the knowledge graph to distinguish between primary and secondary behaviors, and a user's high-order strong and weak feature preference network was constructed to provide a deeper user preference representation for the subsequent recommendation process.

[0172] (2) A shopping psychology-personality extraction capsule based on a sliding window network is proposed. By using a sliding window on a sequence to extract local user psychology and personality, a shopping psychology and purchasing personality extraction capsule is constructed to infer the psychological and personality differences of users when shopping for different products, so that the model can deeply characterize the psychological and personality characteristics of user shopping behavior in multiple dimensions.

[0173] (3) A progressive dissociation method is proposed. Click and add-to-cart behaviors are fused separately, and auxiliary tasks such as purchase, liking, and using shopping psychology and personality are subsequently introduced as basic tasks. Auxiliary tasks are then introduced to derive repurchase preferences. This method can simultaneously dissociate multiple preference factors and obtain significant user behavior characteristics between different behaviors. Through progressive dissociation, continuous behavior sequences are dissociated into binary feature vectors for calculation and matching in the subsequent recommendation process.

[0174] The above three points have a certain effect on improving the performance of the image recommendation system. Focusing on the multi-dimensional characteristics of the project, constructing a user high-order feature preference network, defining the user's main behavior and auxiliary behavior, and realizing the deep preference representation of the user; proposing to build a shopping psychology-personality extraction capsule based on a sliding window network, and deeply characterizing the user's shopping behavior psychology and personality characteristics in multiple dimensions; using the progressive preference dissociation method, obtain the significant user behavior characteristics between different behaviors. The three modules proposed in this method jointly construct the entire recommendation process, ensuring the accuracy and effectiveness of the recommendation results, and greatly improving the reliability of the recommendation results.

[0175] Based on the user's multi-behavior sequence data, the present invention comprehensively and deeply mines implicit information, builds a user's high-order strong and weak feature preference network, extracts the user's deep-level representation, pays attention to the user's shopping psychology and purchasing personality, infers the user's personalized differences, and dissociates multiple preference factors, so as to achieve the personalization and adaptability of the system, improve the accuracy of the recommendation system, and propose a new solution for better mining and utilizing user multi-behavior data. The advantages of this solution are mainly:

[0176] (1) Deeply explore the user's multi-behavior sequence and reveal the interactive relationship between users and items. Compared with traditional methods, it pays more attention to the temporal correlation between behaviors, the difference in behavior intensity, and micro-elements, such as the relative position, intensity, and correlation of behaviors, and more accurately captures the user's diverse interests and preference intensity, thus improving the accuracy of recommendations.

[0177] (2) Drawing on the ideas of RNN and sliding window algorithm, a sliding window network is proposed, which consists of a sliding window extractor and a psychological-personality feature fuser. By sliding a fixed-size window on the input sequence, local features within the window are extracted, allowing the network to pay different degrees of attention to and model different parts of the sequence, thereby better understanding the sequence data.

[0178] (3) By constructing a user's high-order strong and weak feature preference network and a shopping psychology-personality extraction capsule based on a sliding window network, we can deeply explore the user's deep shopping psychology, personality, and derived repurchase preferences implied in the user behavior sequence, characterize the user's shopping behavior psychology and personality characteristics in multiple dimensions, and provide a more comprehensive and in-depth user representation.

[0179] (4) By gradually introducing different preference factors and dissociating multiple preference features, we can obtain significant user behavior characteristics between different behaviors. By integrating click and add-to-cart behaviors, and introducing auxiliary tasks such as purchase, liking, shopping psychology, and personality as basic tasks, we can further introduce auxiliary tasks to derive repurchase preferences, and more accurately reveal user behavior characteristics and preferences.

[0180] By utilizing the above points, we have achieved remarkable technical results in multi-behavior sequence mining and deep user feature representation of image recommendation systems, improved the accuracy of recommendation systems, and made the systems effective and personalized, thus promoting the further development of e-commerce image recommendations.

[0181] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0182] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. An image recommendation method based on user psychological characteristics and progressive dissociation, characterized in that: include: Collect a collection of projects containing various types of e-commerce images; A multi-channel convolutional neural network is used to extract image features in the project set to obtain project representations of each project; Convert the behavior data of the user to be recommended into a knowledge graph, and use a graph neural network to mine the high-order feature preferences of users and behaviors in the knowledge graph to obtain a high-order feature preference network of the user to be recommended; Calculate the behavior intensity of each behavior of the recommended user based on the behavior frequency and behavior weight; According to the strength of the behavior, all behaviors of the user to be recommended are divided into main behaviors and auxiliary behaviors, and the characteristics of the main behavior interaction items are defined as strong features, and the characteristics of the auxiliary behavior interaction items are defined as weak features; The attention mechanism based on associated behaviors extracts the association relationship between behaviors from the behavior data of the recommended user and captures the derived repurchase preference; According to the strong features, weak features, the correlation between behaviors and the derived repurchase preferences, the first user representation is obtained by using the high-order feature preference network; Constructing shopping psychology-personality extraction capsules based on sliding window networks; The sliding window network in the shopping psychology-character extraction capsule based on the sliding window network includes: a sliding window extractor and a psychology-character feature fuser; the sliding window extractor is a recurrent neural network module for extracting local features within the sliding window; the psychology-character feature fuser is used to weightedly fuse the local features extracted by the sliding window extractor to generate a global representation; the shopping psychology-character extraction capsule in the shopping psychology-character extraction capsule based on the sliding window network is used to calculate the activation prediction value s according to the global representation using the formula s=softmax(o)*g; then a nonlinear squashing function is used to scale the activation prediction value according to r=squash(s)*g, and output the psychology-character fusion feature r of the user to be recommended; wherein o represents the prediction vector of the capsule unit, softmax() represents the softmax function, g is the global representation, and squash() represents the squashing function; After extracting the global feature representation of the behavior data by using the sliding window network, the global feature representation is input into the shopping psychology-character extraction capsule to obtain the psychological and character fusion features of the user to be recommended; According to the psychological and personality fusion characteristics of the user to be recommended, a progressive preference dissociation method is used to determine the second user representation; the progressive preference dissociation method includes a first branch and a second branch, the first branch is used to take the fusion of clicks and add to cart as the basic task, and gradually introduce purchase behavior and liking behavior; the second branch is used to take the fusion of user shopping psychology and personality as the basic task, and introduce the derived repurchase preference; the dissociation results of the first branch and the second branch are aggregated to obtain the second user representation; Aggregate the first user representation and the second user representation to obtain a final user representation; The inner product operation is performed on the final user representation and the item representation of each item to obtain the user's preference prediction score for each item.

2. The image recommendation method based on user psychological characteristics and progressive dissociation according to claim 1, characterized in that: A multi-channel convolutional neural network is used to extract image features in the project set to obtain project representation of each project, specifically including: A multi-channel convolutional neural network is used to extract image features of each e-commerce image in the project set; the image features include type, color and style; The image features of each item that the user has interacted with are converted into a product attribute embedding sequence through the embedding layer, and the formula is used and Calculate the item representation for each item; where e i represents the item representation of the i-th item, d j represents the jth feature attribute in the product attribute embedding sequence, ω j Represents attribute d j The weight of M is the number of characteristic attributes in the product attribute embedding sequence, d m represents the mth characteristic attribute in the product attribute embedding sequence, and exp represents the exponential function.

3. The image recommendation method based on user psychological characteristics and progressive dissociation according to claim 1, characterized in that: The calculation formula of the behavior intensity is: α＝∑α freq +α weight ； α weight ＝α(k)； In the formula, α represents the behavior intensity, α freq represents the behavior frequency, α weight Indicates the weight of behavior; t freq Indicates the number of times a user performs a specific behavior type, T freq represents the total length of observation time; k represents the user behavior type, and α(k) represents the nonlinear activation function for user behavior type k.

4. The image recommendation method based on user psychological characteristics and progressive dissociation according to claim 1, characterized in that: Based on the intensity of the behavior, all behaviors of the recommended user are divided into primary behaviors and secondary behaviors, including: The behavior with a behavior intensity greater than or equal to the intensity threshold is determined as the main behavior, and the behavior with a behavior intensity less than the intensity threshold is determined as the auxiliary behavior.

5. The image recommendation method based on user psychological characteristics and progressive dissociation according to claim 2, characterized in that: The calculation formula for the correlation between behaviors is: In the formula, q represents the correlation between behavior b1 and behavior b2. Represents the exclusive OR operation; The calculation formula for the derived repurchase preference is: In the formula, d represents the derived repurchase preference, β1 is the weight of behavior b1, and β2 is the weight of behavior b2. and They are the embedding representations of behavior b1 and behavior b2 respectively.

6. The image recommendation method based on user psychological characteristics and progressive dissociation according to claim 5, characterized in that: The formula for determining the first user representation is: In the formula, e u is the first user representation, γ s , γ w They are strong features f s and weak feature f w The weight of Represents the inner product operation.

7. The image recommendation method based on user psychological characteristics and progressive dissociation according to claim 6, characterized in that: According to the strong features, weak features, the correlation between behaviors and the derived repurchase preferences, the first user representation is obtained by using the high-order feature preference network, and then the following is further included: For products that users have purchased n If the user does not click or browse in the next purchase behavior, the product v n is considered as the user's temporary disappearance preference, then according to the formula Update the first user representation; where, The first user after the update indicates that For product v n Users said.

8. The image recommendation method based on user psychological characteristics and progressive dissociation according to claim 7, characterized in that: According to the psychological and personality fusion characteristics of the user to be recommended, the progressive preference dissociation method is used to determine the second user representation, which specifically includes: For the first branch: The integration of click and add to cart is used as the basic task, according to the formula y basic1 =f basic1 (E c ,E a ,u,i), predict the results of click and add to cart; where y basic1 represents the prediction results of click and add-to-cart, f basic1 ( ) represents the nonlinear activation function used to predict click and add-to-cart results, E c 、E a Represent the embedding sequences of clicks and add-to-carts, u represents user u, and i represents the i-th item; Introduce the purchase task into the prediction results of clicks and add-to-cart, according to the formula y b =f b (y basic1 ,E b ,u,i), predict the result of introducing purchase behavior; where y b represents the prediction result of introducing purchase behavior, f b () represents the nonlinear activation function used to predict the purchase behavior results, E b Embedding sequence representation for the purchased embedding; Introduce the like task into the prediction results of purchase behavior, according to the formula y1 = f f (y b ,E f ,u,i), predict the result of introducing liking behavior as the dissociation result y1 of the first branch; where f f ( ) represents the nonlinear activation function used to predict the outcome of liking behavior, E f Embed the sequence representation for the favorite embedding; For the second branch: Take the psychological and personality fusion features r of the recommended user as the basic task, according to the formula y basic2 =f basic2 (r,u,i), predicting the results of purchasing psychology and personality; where y basic2 represents the prediction results of purchasing psychology and personality, f basic2 ( ) represents the nonlinear activation function used to predict purchase psychology and personality outcomes; The prediction results of introducing the derived repurchase preference task into the purchase psychology and personality basic2 , according to the formula y2 = f d (y basic2 ,d,u,i), predict the result of introducing the derived repurchase preference as the dissociation result y2 of the second branch; where f d ( ) represents the nonlinear activation function used to predict the derived repurchase preference results; Using the formula Connect the dissociation result y1 of the first branch and the dissociation result y2 of the second branch to output the second user representation 9. The image recommendation method based on user psychological characteristics and progressive dissociation according to claim 8, characterized in that: The formula for determining the end user representation is: In the formula, e u for end users; The formula for determining the preference prediction score is: In the formula, represents the preference prediction score of user u for the i-th item.

Images