Method and system for personalized recommendation of goods based on user behavior data analysis

Abstract

The application provides a commodity personalized recommendation method and system based on user behavior data analysis, first, a user behavior sequence and an interest stability model are constructed, then, the user behavior sequence is associated modeling, the behavior transfer correlation degree of adjacent interaction behavior units and the commodity attribute dynamic correlation degree which can be adjusted with the interest stability of the user are calculated in combination with the interest stability model, then, a dynamic commodity collaborative network is constructed based on the commodity attribute dynamic correlation degree, the node importance parameters are updated according to the real-time interaction behavior of the user, path optimization mining is carried out in the dynamic commodity collaborative network, a user potential behavior path set is generated, finally, the potential behavior path set is analyzed, a commodity personalized recommendation sequence is generated in combination with the path attribute correlation feature distribution and the current interaction behavior of the user, the accuracy and the personalized degree of the recommendation are greatly improved, and the shopping experience of the user is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a commodity personalized recommendation method and system based on user behavior data analysis. BACKGROUND

[0002] In the current booming e-commerce, commodity personalized recommendation system has become a key tool to improve user experience and platform sales performance. Traditional commodity recommendation methods are mainly divided into two categories. One is content-based recommendation, which generates recommendation results by analyzing the attribute characteristics of commodities and the attribute matching degree of user historical preference commodities. However, the above method ignores the time series information of user behavior and the dynamic change of interest, and is difficult to capture the transfer of user interest and the generation of new interest. The other is collaborative filtering-based recommendation, which uses the similarity between users or the similarity between commodities for recommendation. However, the collaborative filtering method usually calculates the similarity statically without considering the stability of user interest and the change of commodity attribute correlation with time, resulting in inaccurate and timely recommendation results.

[0003] In addition, few methods in the prior art can comprehensively consider user behavior sequence, interest stability and commodity attribute dynamic correlation to construct a recommendation model, which cannot meet the increasingly diversified and personalized needs of users. SUMMARY

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiments of the present application provide a commodity personalized recommendation method based on user behavior data analysis, which comprises:

[0005] A user behavior sequence and interest stability model is constructed, the user behavior sequence comprises user and commodity interaction behavior units arranged in time sequence, each interaction behavior unit is associated with interaction type and commodity attribute information, and the interest stability model is used to quantify the interest persistence characteristics of users to commodity attributes;

[0006] The user behavior sequence is subjected to association modeling processing, and the behavior transfer correlation and commodity attribute dynamic correlation of adjacent interaction behavior units are calculated in combination with the interest stability model, and the commodity attribute dynamic correlation is adjusted according to the change of user interest stability;

[0007] A dynamic commodity collaborative network is constructed based on the commodity attribute dynamic correlation, the nodes of the dynamic commodity collaborative network are commodities, the edges are commodity attribute dynamic correlations, and the node importance parameters are updated according to real-time user interaction behavior;

[0008] The path optimization mining processing is performed in the dynamic commodity collaborative network, a user potential behavior path set is generated by combining the behavior transition correlation degree and the node importance parameter, and the potential behavior path set includes multiple paths with different attribute correlation characteristics.

[0009] The potential behavior path set is analyzed, and a commodity personalized recommendation sequence is generated based on the attribute correlation characteristic distribution of the path and the current interaction behavior of the user.

[0010] In another aspect, the embodiment of the present application also provides a commodity personalized recommendation system based on user behavior data analysis, which comprises a processor and a machine readable storage medium, the machine readable storage medium is connected with the processor, the machine readable storage medium is used for storing programs, instructions or codes, and the processor is used for executing the programs, instructions or codes in the machine readable storage medium to realize the above-mentioned method.

[0011] Based on the above aspects, by constructing the user behavior sequence and the interest stability model, the time sequence characteristics of the user and the commodity interaction behavior and the continuous change of the user's interest in the commodity attributes can be comprehensively and accurately captured. When the association modeling processing is performed on the user behavior sequence, the behavior transition correlation degree and the commodity attribute dynamic correlation degree of adjacent interaction behavior units are calculated in combination with the interest stability model, so that the commodity attribute correlation degree can be adjusted in real time according to the stability of the user's interest, which is more in line with the actual change of the user's interest. The dynamic commodity collaborative network constructed based on the commodity attribute dynamic correlation degree can update the node importance parameter according to the real-time interaction behavior of the user, which ensures the dynamics and real-time performance of the network. The path optimization mining processing is performed in the dynamic commodity collaborative network, and the user potential behavior path set generated by combining the behavior transition correlation degree and the node importance parameter can more accurately predict the future behavior trend of the user. The commodity personalized recommendation sequence finally generated comprehensively considers the attribute correlation characteristic distribution of the path and the current interaction behavior of the user, greatly improves the accuracy and personalization degree of the recommendation, and effectively improves the shopping experience of the user. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 is the execution flow diagram of the commodity personalized recommendation method based on user behavior data analysis provided by the embodiment of the present application.

[0013] Figure 2 is the schematic diagram of the exemplary hardware and software components of the commodity personalized recommendation system based on user behavior data analysis provided by the embodiment of the present application. DETAILED DESCRIPTION

[0014] The present application will be specifically described below in combination with the drawings of the specification, Figure 1This is a flowchart illustrating a personalized product recommendation method based on user behavior data analysis according to an embodiment of the present invention. The following is a detailed description of this personalized product recommendation method based on user behavior data analysis.

[0015] Step S110: Construct a user behavior sequence and interest stability model. The user behavior sequence includes user and product interaction behavior units arranged in chronological order. Each interaction behavior unit is associated with interaction type and product attribute information. The interest stability model is used to quantify the sustained characteristics of user interest in product attributes.

[0016] In this embodiment, constructing a user behavior sequence can clearly present the interaction history between users and products, while the interest stability model can quantify the persistence of users' interest in different product attributes, helping to more accurately grasp users' interest preferences. The specific implementation process will be described in detail below.

[0017] Step S111: Collect the user's original interaction records during the entire interaction period. The original interaction records include the time of interaction, the products involved in the interaction, the type of interaction, and the duration of the interaction. The interaction types include browsing interaction, favorites interaction, add-to-cart interaction, and purchase interaction.

[0018] In practical applications, data acquisition systems can be used to collect raw interaction records from multiple data sources throughout the entire user interaction cycle. These data sources can include e-commerce platform log systems, mobile application event tracking data, etc. The interaction timestamps in the raw interaction records clearly indicate the chronological order of user interactions with products; the products involved in the interactions identify the specific products the user is interested in; the interaction types reflect different user actions on the products; and the interaction duration reflects the time and effort the user invests in that interaction. For example, in an online shopping platform, the system records information related to each time a user opens a product details page (browsing interaction), clicks the favorite button (favorite interaction), adds the product to the shopping cart (add to cart interaction), and completes the purchase (purchase interaction), including the specific time, the names of the products involved, the type of interaction, and the time the user spends on each action.

[0019] Step S112: Divide the original interaction records into behavioral units, merge the continuous interaction records of the same product within a preset time window into one interactive behavioral unit. The interaction type of the interactive behavioral unit is the set of all interaction types contained in the merged records, and the interaction duration is the total interaction duration of the merged records.

[0020] In order to analyze the user's interaction behavior more effectively, it is necessary to divide the original interaction record into behavior units. The setting of the preset time window is a key factor, which determines which continuous interaction records will be combined. In actual operation, a suitable time window can be determined according to business needs and data analysis experience. For example, for some high-frequency interaction goods, the time window can be set shorter; for low-frequency interaction goods, the time window can be appropriately extended. When there are continuous interaction records of the same goods within the preset time window, these records are combined into an interaction behavior unit. The interaction type of this unit is the set of all interaction types in the combined records, which can more comprehensively reflect the user's operation behavior on the goods. The interaction duration is the total interaction duration of the combined records, which reflects the user's total input time on the goods. For example, if a user first performs a browsing interaction and then a collection interaction on a certain mobile phone within a short period of time, these two records will be combined into an interaction behavior unit, and the interaction type will be the set of browsing interaction and collection interaction, and the interaction duration will be the sum of the two interaction times.

[0021] Step S113: Extract the product attribute information of the product involved in each interaction behavior unit, the product attribute information including product category attribute, product function attribute, product scene attribute and product style attribute, wherein the product scene attribute is used to indicate the use scene applicable to the product, and the product style attribute is used to indicate the design style characteristics of the product.

[0022] Each interaction behavior unit is associated with a specific product, and extracting the attribute information of these products helps to deeply understand the characteristics of the products and the interest preferences of the users. The product category attribute can clearly indicate the category to which the product belongs, such as electronic products, clothing, food, etc. The product function attribute describes various functions possessed by the product, such as the camera function of a mobile phone, the processing capability of a computer, etc., which are important basis for users to select products. The product scene attribute indicates the use scene applicable to the product, such as sports scene, office scene, leisure scene, etc., which reflects the practicality of the product in different environments. The product style attribute embodies the design style characteristics of the product, such as minimalist style, retro style, fashion style, etc., which meets the needs of users for the appearance and aesthetic aspects of the product. For example, for a sports watch, its product category attribute is electronic products-watches, its product function attribute may include heart rate monitoring, motion trajectory recording, etc., its product scene attribute is applicable to sports scene, and its product style attribute may be minimalist fashion style.

[0023] Step S114: Sort the divided interaction behavior units in chronological order according to the interaction occurrence time, generate a preliminary user behavior sequence, perform missing interaction verification on the preliminary user behavior sequence, supplement the interaction behavior units corresponding to the key product attributes caused by the interruption of the interaction records, and obtain a complete user behavior sequence.

[0024] The sorted interaction behavior units can form a clear timeline of user interaction with goods, i.e., a preliminary user behavior sequence. The preliminary user behavior sequence can intuitively show which goods the user interacted with at different time points, as well as the type and duration of the interaction, etc. However, due to various reasons such as system failure, network problems, etc., the interaction record may be interrupted, resulting in missing interaction behavior units corresponding to key goods attributes. In order to ensure the integrity and accuracy of the user behavior sequence, the preliminary user behavior sequence needs to be checked for missing interactions. In actual operation, by analyzing the user's historical interaction behavior patterns, the relevance between goods attributes, and the continuity of time sequence, etc., it can be determined whether there is a missing interaction behavior unit. If missing is found, according to the relevant rules and algorithms, the corresponding interaction behavior unit is supplemented, so as to obtain a complete user behavior sequence. For example, if in the user's interaction record, the interaction record of a certain type of popular goods is suddenly interrupted in a certain time period, and according to the user's historical preferences and the popularity of the goods, it is speculated that the user may have browsed the goods during this time period, at this time, the corresponding interaction behavior unit can be supplemented to the user behavior sequence.

[0025] Step S115: Based on the complete user behavior sequence, an interest stability model is constructed, the frequency of the same goods attribute information appearing in different time windows in the complete user behavior sequence is extracted, the fluctuation coefficient of the frequency change is calculated, and the reciprocal of the fluctuation coefficient is taken as the interest stability parameter of the user to the goods attribute. The higher the interest stability parameter is, the more stable the user's interest persistence feature for the goods attribute is.

[0026] The purpose of constructing the interest stability model is to quantify the user's interest persistence feature for the goods attribute, so as to better understand the user's interest preferences and behavior patterns. The specific implementation process is as follows:

[0027] Step S1151: The time span of the complete user behavior sequence is evenly divided into multiple continuous time windows, each time window has the same duration, and there is no overlap between the time windows.

[0028] In order to analyze the change of user interest in the commodity attribute in detail, the time span of the complete user behavior sequence needs to be reasonably divided. In actual operation, the length of the time window can be determined according to business requirements and data characteristics. For example, if the short-term change of user interest is analyzed, the time window can be set to be short; if the long-term user interest trend is analyzed, the time window can be set to be long. The time span is evenly divided into a plurality of continuous time windows, and the length of each time window is the same, and there is no overlap between the time windows, so that the data of each time window has independence and comparability. For example, the complete behavior sequence of a user in a month is divided into a plurality of time windows in units of days, and each window represents the interaction data of one day.

[0029] Step S1152: For each commodity attribute information, the number of interaction behavior units appearing in each time window is counted, and the ratio of the number to the total number of interaction behavior units in the time window is taken as the appearance frequency of the commodity attribute information in the time window.

[0030] In each divided time window, for each commodity attribute information, the number of interaction behavior units appearing is counted. The number of interaction behavior units reflects the degree of attention of the user to the commodity attribute in the time window. Then, the number is compared with the total number of interaction behavior units in the time window, and the ratio is calculated, which is the appearance frequency of the commodity attribute information in the time window. For example, in a time window of a day, there is a certain number of interaction behavior units, and the interaction behavior units related to a commodity attribute (such as the camera function of a mobile phone) are a certain number. By calculating the ratio of the two numbers, the appearance frequency of the commodity attribute in the day can be obtained.

[0031] Step S1153: The appearance frequencies of all time windows are arranged in time sequence to form a frequency sequence, and the mean value of the frequency sequence is calculated, which is the arithmetic mean of all appearance frequencies.

[0032] The appearance frequencies of each commodity attribute information in each time window are arranged in time sequence to form a frequency sequence. The frequency sequence can directly show the change of the appearance frequency of the commodity attribute information at different time points. The mean value of the frequency sequence, which is the arithmetic mean of all appearance frequencies, can reflect the average appearance frequency of the commodity attribute information in the entire time span. Through the calculation of the mean value, the overall attention degree of the user to the commodity attribute can be preliminarily understood. For example, for a commodity attribute, the appearance frequencies in a plurality of time windows are different values, and the sum of the values divided by the number of time windows can obtain the mean value of the appearance frequency of the commodity attribute.

[0033] Step S1154: square the difference between each frequency in the frequency sequence and the mean, sum all the squared values, and divide by the number of time windows to obtain the variance of the frequency change, and take the square root of the variance as the volatility coefficient of the frequency change.

[0034] To further analyze the stability of user interest in the product attribute, the volatility of the frequency sequence needs to be calculated. First, the square of the difference between each frequency in the frequency sequence and the mean is calculated. This squared value can amplify the difference between the frequency and the mean, highlighting the degree of volatility. Then, sum all the squared values and divide by the number of time windows to obtain the variance of the frequency change. The variance reflects the degree of dispersion of the frequency sequence. The larger the variance, the greater the volatility of the frequency, and the less stable the user's interest in the product attribute. Finally, take the square root of the variance as the volatility coefficient of the frequency change, which more intuitively reflects the volatility amplitude of the frequency. For example, for a frequency sequence of a product attribute, the square of the difference between each frequency and the mean is calculated, the sum of these squared values is divided by the number of time windows to obtain the variance, and the square root of the variance is taken to obtain the volatility coefficient.

[0035] Step S1155: take the reciprocal of the volatility coefficient as the interest stability parameter of the user for the product attribute information, and if the volatility coefficient is zero, set the interest stability parameter to a preset maximum value, and the value range of the interest stability parameter is consistent with the value range of the product attribute dynamic association degree.

[0036] Taking the reciprocal of the volatility coefficient as the interest stability parameter of the user for the product attribute information is because the smaller the volatility coefficient, the more stable the user's interest in the product attribute, and the larger the reciprocal. If the volatility coefficient is zero, it means that the frequency sequence has no volatility, and the user's interest in the product attribute is very stable. At this time, the interest stability parameter is set to a preset maximum value. At the same time, in order to ensure the consistency and reasonableness of subsequent calculations, the value range of the interest stability parameter needs to be consistent with the value range of the product attribute dynamic association degree. For example, if the value range of the product attribute dynamic association degree is from a minimum value to a maximum value, the interest stability parameter should also be valued within the same range.

[0037] Step S120: perform association modeling processing on the user behavior sequence, and calculate the behavior transition association degree of adjacent interaction behavior units and the product attribute dynamic association degree in combination with the interest stability model, wherein the product attribute dynamic association degree is adjusted according to the change of user interest stability.

[0038] After the user behavior sequence and the interest stability model are constructed, the user behavior sequence needs to be associated and modeled to calculate the behavior transition correlation degree and the commodity attribute dynamic correlation degree between adjacent interaction behavior units. The behavior transition correlation degree reflects the possibility and strength of the user transferring from one interaction behavior unit to the next interaction behavior unit, and the commodity attribute dynamic correlation degree reflects the correlation degree between the commodity attributes involved in adjacent interaction behavior units, and the correlation degree is adjusted with the change of the user interest stability. The specific implementation process is as follows:

[0039] Step S121: Identify the interaction type and the interaction duration of each interaction behavior unit in the user behavior sequence, assign a basic behavior weight to the interaction behavior unit according to the interaction type, and the basic behavior weight increases with the increase of the decision depth of the interaction type, wherein the basic behavior weight of the purchase interaction is greater than that of the add-to-cart interaction, the basic behavior weight of the add-to-cart interaction is greater than that of the collection interaction, and the basic behavior weight of the collection interaction is greater than that of the browsing interaction.

[0040] First, the interaction type and the interaction duration of each interaction behavior unit in the user behavior sequence need to be identified. The interaction type includes browsing interaction, collection interaction, add-to-cart interaction and purchase interaction, and different interaction types reflect different decision depths of the user on the commodity. Browsing interaction is usually the user's preliminary understanding of the commodity, and the decision depth is shallow; collection interaction indicates that the user has a certain interest in the commodity, and the decision depth increases; add-to-cart interaction indicates that the user is more inclined to purchase the commodity, and the decision depth is further deepened; and purchase interaction is the user's final decision behavior, and the decision depth is the deepest. According to the interaction type, a basic behavior weight is assigned to the interaction behavior unit, and the weight increases with the increase of the decision depth of the interaction type. For example, a relatively small basic behavior weight is assigned to browsing interaction, and a relatively large basic behavior weight is assigned to purchase interaction.

[0041] Step S122: Correct the basic behavior weight based on the interaction duration, calculate the ratio of the interaction duration to the average interaction duration of the interaction type, take the ratio as a duration correction factor, and multiply the basic behavior weight by the duration correction factor to obtain a corrected behavior weight value.

[0042] Interaction duration is also an important factor affecting the degree of user interest in the commodity. In order to more accurately reflect the user's interest, it is necessary to correct the basic behavior weight based on the interaction duration. The ratio of the interaction duration to the average interaction duration of the interaction type is calculated, and the ratio is the duration correction factor. If the interaction duration is greater than the average interaction duration of the interaction type, it means that the user spends more time on the interaction, and the interest in the commodity may be more intense, and the duration correction factor will be greater than 1; on the contrary, if the interaction duration is less than the average interaction duration, the duration correction factor will be less than 1. Multiply the basic behavior weight by the duration correction factor to get the corrected behavior weight value. For example, for a certain interaction behavior unit, its basic behavior weight has been allocated according to the interaction type, and the duration correction factor is obtained by calculating the ratio of its interaction duration to the average interaction duration of the interaction type, and then multiplying the two to get the corrected behavior weight value.

[0043] Step S123: Extract the interaction occurrence time of the adjacent two interaction behavior units in the user behavior sequence, calculate the time interval parameter, construct a time decay function based on the time interval parameter, the output value of the time decay function decreases with the increase of the time interval parameter, and the corrected behavior weight value is multiplied by the output value of the time decay function to obtain the time decay behavior weight.

[0044] There may be a certain time interval between the interaction behaviors of the user at different time points, and the time interval will affect the degree of association between adjacent interaction behavior units. Extract the interaction occurrence time of the adjacent two interaction behavior units in the user behavior sequence, and calculate the time interval parameter between them. Based on the time interval parameter, a time decay function is constructed, and the output value of the function will decrease with the increase of the time interval parameter. This is because the longer the time interval, the more likely the user's interest and decision will change, and the degree of association between adjacent interaction behavior units will also weaken accordingly. Multiply the corrected behavior weight value by the output value of the time decay function to get the time decay behavior weight. For example, for two adjacent interaction behavior units, calculate the time interval of their interaction occurrence, and get an output value through the time decay function according to the interval parameter, and then multiply the corrected behavior weight value by the output value to get the time decay behavior weight.

[0045] Step S124: Call the interest stability model to obtain the interest stability parameters corresponding to the respective commodity attribute information of the adjacent interaction behavior units, and take the product of the two interest stability parameters as the attribute stability factor.

[0046] The interest stability model has quantified the interest persistence characteristics of users on different commodity attributes. In calculating the dynamic correlation degree of commodity attributes, the interest stability model needs to be called to obtain the interest stability parameters of the commodity attribute information of each adjacent interaction behavior unit. These two parameters reflect the interest stability degree of the user on the commodity attributes involved in the adjacent interaction behavior units. Multiply these two interest stability parameters to obtain the attribute stability factor. The attribute stability factor reflects the interest stability correlation degree between the commodity attributes of adjacent interaction behavior units. For example, for two adjacent interaction behavior units, obtain the interest stability parameters corresponding to their commodity attribute information, multiply these two parameters to obtain the attribute stability factor.

[0047] Step S125: Extract the commodity attribute information of adjacent interaction behavior units, calculate the coincidence degree of commodity category attributes, the semantic matching degree of commodity function attributes, the scene matching degree of commodity scene attributes, and the style similarity of commodity style attributes, and weight sum the category coincidence degree, the function semantic matching degree, the scene matching degree, and the style similarity after assigning a preset weight to each to obtain the basic attribute correlation degree.

[0048] Step S1251: The commodity category attributes of adjacent interaction behavior units are respectively split into multiple category labels, the number of same category labels in the two commodity category attributes is counted, and the ratio of the number of same category labels to the total number of category labels of the two commodity category attributes is taken as the coincidence degree of the commodity category attributes.

[0049] In calculating the coincidence degree of commodity category attributes, the commodity category attributes of adjacent interaction behavior units are respectively split into multiple category labels. These labels can more specifically describe the category information of the commodity. Then, the number of same category labels in the two commodity category attributes is counted, which reflects the overlapping part of the two commodities in the category. Compare the number of same category labels with the total number of category labels of the two commodity category attributes to calculate the ratio, which is the coincidence degree of the commodity category attributes. For example, for two adjacent interaction behavior units, their commodity category attributes are respectively split into several category labels, the number of same labels is counted, and the ratio to the total number of labels is calculated to obtain the coincidence degree of the commodity category attributes.

[0050] Step S1252: The commodity function attributes of adjacent interaction behavior units are preprocessed, the core function description words are retained after removing stop words, the core function description words are input into a pre-trained semantic encoder to generate function attribute vectors, and the cosine similarity between the two function attribute vectors is calculated as the semantic matching degree of the commodity function attributes.

[0051] For the functional attributes of the commodities of the adjacent interaction behavior units, first, text preprocessing is performed. Stop words are some words that frequently appear in the text but do not help much in semantic understanding, such as “of”, “is”, “and” and the like. Removing these stop words can make the text more concise and highlight the core functional description words. The core functional description words that are retained are input into a pre-trained semantic encoder, and the encoder converts these words into a functional attribute vector. The functional attribute vector can more accurately represent the semantic information of the functional attributes of the commodities. The cosine similarity between the two functional attribute vectors is calculated, and the cosine similarity can measure the degree of similarity in direction of the two vectors. The closer the value is to 1, the more matched the semantics of the functional attributes of the two commodities are. The cosine similarity is taken as the semantic matching degree of the functional attributes of the commodities. For example, the functional attributes of the commodities of the adjacent two interaction behavior units are subjected to text preprocessing, and the core functional description words are obtained after the stop words are removed. The words are input into a semantic encoder to generate a functional attribute vector, the cosine similarity between the two vectors is calculated, and the semantic matching degree of the functional attributes of the commodities is obtained.

[0052] Step S1253: Extracting scene labels contained in the scene attributes of the commodities, each scene label corresponding to a preset scene weight, and calculating the sum of the scene weights of the same scene labels in the two scene attributes of the commodities. The ratio of the sum of the scene weights to the total sum of the scene weights of all scene labels of the two scene attributes of the commodities is taken as the scene matching degree of the scene attributes of the commodities.

[0053] Extracting scene labels contained in the scene attributes of the commodities, each scene label corresponding to a preset scene weight, and the scene weight reflecting the importance of the scene in the use of the commodity. The sum of the scene weights of the same scene labels in the two scene attributes of the commodities is calculated, which reflects the comprehensive importance of the two commodities in the same scene. Then, the sum of the scene weights is compared with the total sum of the scene weights of all scene labels of the two scene attributes of the commodities, and the ratio is calculated. The ratio is the scene matching degree of the scene attributes of the commodities. For example, for two adjacent interaction behavior units, the scene labels of the scene attributes of the commodities are extracted, the sum of the scene weights of the same scene labels is calculated, and the ratio of the sum to the total sum of the scene weights of all scene labels is calculated to obtain the scene matching degree of the scene attributes of the commodities.

[0054] Step S1254: Converting the style attributes of the commodities into style feature vectors, the dimensions of the style feature vectors corresponding to a preset set of style dimensions, and the value of each dimension being a feature value of the style dimension. The Euclidean distance between the two style feature vectors is calculated, and the reciprocal of the Euclidean distance is taken as the style similarity of the style attributes of the commodities. If the Euclidean distance is zero, the style similarity is set to a preset maximum value.

[0055] The commodity style attribute is converted into a style feature vector, and the dimension of the style feature vector corresponds to a preset style dimension set. The value of each dimension is a characteristic value of the style dimension, and the characteristic values can quantify the style characteristics of the commodity. The Euclidean distance between two style feature vectors is calculated, and the Euclidean distance can measure the distance between two vectors in space. The closer the distance, the more similar the styles of the two commodities. The reciprocal of the Euclidean distance is taken as the style similarity of the commodity style attribute, because the smaller the Euclidean distance, the higher the similarity, and taking the reciprocal can more intuitively reflect the above relationship. If the Euclidean distance is zero, it means that the styles of the two commodities are completely the same, and at this time the style similarity is set to a preset maximum value. For example, the commodity style attributes of two adjacent interaction behavior units are converted into style feature vectors, the Euclidean distance between them is calculated, the reciprocal is taken to obtain the style similarity, and if the Euclidean distance is zero, it is set to a preset maximum value.

[0056] Step S126: Multiply the basic attribute association degree by the attribute stability factor to obtain a commodity attribute dynamic association degree, and multiply the time decay behavior weight by the commodity attribute dynamic association degree to obtain a behavior transition association degree of the adjacent interaction behavior unit.

[0057] The basic attribute association degree is multiplied by the attribute stability factor to obtain the commodity attribute dynamic association degree. The commodity attribute dynamic association degree not only considers the basic association degree between the commodity attributes of adjacent interaction behavior units, but also considers the interest stability of the user. The time decay behavior weight is multiplied by the commodity attribute dynamic association degree to obtain the behavior transition association degree of the adjacent interaction behavior unit. The behavior transition association degree comprehensively considers the behavior weight of the user, the time interval, and the association degree between the commodity attributes, and more comprehensively reflects the possibility and strength of the user transferring from one interaction behavior unit to the next interaction behavior unit. For example, the basic attribute association degree is multiplied by the attribute stability factor to obtain the commodity attribute dynamic association degree, and the time decay behavior weight is multiplied by the commodity attribute dynamic association degree to obtain the behavior transition association degree of the adjacent interaction behavior unit.

[0058] Step S130: Construct a dynamic commodity collaborative network based on the commodity attribute dynamic association degree, wherein the nodes of the dynamic commodity collaborative network are commodities, the edges are commodity attribute dynamic association degrees, and the node importance parameters are updated according to real-time interaction behaviors of the user.

[0059] The dynamic commodity collaborative network is a graph structure for representing the association relationship between commodities, which takes commodities as nodes and commodity attribute dynamic association degrees as edges, and can intuitively show the association degree between commodities. At the same time, the node importance parameters are updated according to the real-time interaction behaviors of the user to reflect the importance of the commodity in the current interest of the user. The specific construction process is as follows:

[0060] Step S131: Collect all the goods involved in the interaction behavior units in the user behavior sequence, remove the repeated goods to form a goods set, and take each good in the goods set as an initial node of the dynamic goods collaborative network.

[0061] All the goods involved in the interaction behavior units in the user behavior sequence are collected. Since the user may interact with the same good multiple times, it is necessary to remove the repeated goods to form a unique goods set. Each good in the goods set is taken as an initial node of the dynamic goods collaborative network, and these nodes constitute the basic elements of the network. For example, in the user's behavior sequence, multiple mobile phones, computers and other goods are involved, and after removing the repeated goods, these different goods are taken as the initial nodes of the dynamic goods collaborative network.

[0062] Step S132: For any two goods in the goods set, if there is at least one adjacent interaction behavior unit in the user behavior sequence that involves the two goods respectively, the dynamic correlation degree of the corresponding goods attribute of the adjacent interaction behavior unit is extracted as the initial weight of the edge connecting the two goods nodes, and if there are multiple connection edges between the two goods, the average value of all initial weights is taken as the final edge weight.

[0063] For any two goods in the goods set, check whether there is at least one adjacent interaction behavior unit in the user behavior sequence that involves the two goods respectively. If there is the above adjacent interaction behavior unit, the dynamic correlation degree of the corresponding goods attribute of the adjacent interaction behavior unit is extracted, which is taken as the initial weight of the edge connecting the two goods nodes. The initial weight reflects the correlation degree between the two goods. If there are multiple connection edges between the two goods (i.e. there are multiple adjacent interaction behavior units involving the two goods respectively), the average value of all initial weights is taken as the final edge weight to more accurately represent the comprehensive correlation degree between the two goods. For example, for goods A and goods B, if there are multiple adjacent interaction behavior units involving them in the user behavior sequence, the dynamic correlation degree of the corresponding goods attribute of each adjacent interaction behavior unit is extracted, and the average value of these correlation degrees is calculated as the final weight of the edge connecting goods A and goods B.

[0064] Step S133: Initialize the node importance parameter for each node in the dynamic goods collaborative network, and the initial node importance parameter is the ratio of the number of interaction behavior units of the good appearing in the user behavior sequence to the total number of interaction behavior units.

[0065] An initial node importance parameter of each node in the dynamic commodity collaborative network is initialized, which reflects the relative importance of the commodity in the user behavior sequence. The initial node importance parameter is calculated by the ratio of the number of interaction units of the commodity in the user behavior sequence to the total number of interaction units. For example, for a certain commodity, there are a certain number of interaction units of the commodity in the user behavior sequence, and there are a certain number of total interaction units. The ratio of the two numbers is calculated to obtain the initial node importance parameter of the commodity.

[0066] Step S134: setting a node importance update period, collecting real-time interaction behaviors of users in each update period, the real-time interaction behaviors being interaction units generated by the users in the current update period.

[0067] The node importance update period is set, which can be determined according to business requirements and the frequency of data updates. In each update period, real-time interaction behaviors of users are collected, which are interaction units generated by the users in the current update period. By continuously collecting real-time interaction behaviors, changes in user interests can be reflected in time, so as to update the importance parameters of nodes. For example, the update period is set to one day, and interaction units generated by the users in the day are collected every day.

[0068] Step S135: updating the node importance parameter of the corresponding node according to the commodity involved in the real-time interaction behavior, increasing the node importance parameter by a preset increment value if the interaction type of the real-time interaction behavior is purchase interaction, decreasing the node importance parameter by a preset decrement value if the interaction type is browsing interaction, and the increment value and the decrement value corresponding to the add-to-cart interaction and the collection interaction being between the purchase interaction and the browsing interaction.

[0069] The node importance parameter of the corresponding node is updated according to the commodity involved in the real-time interaction behavior. Different interaction types have different effects on the node importance parameter. If the interaction type of the real-time interaction behavior is purchase interaction, it indicates that the user has a high interest and purchase intention for the commodity, and the node importance parameter will increase by a preset increment value. If the interaction type is browsing interaction, it may only be the user's preliminary understanding, and the node importance parameter will decrease by a preset decrement value. The add-to-cart interaction and the collection interaction indicate that the user has a certain interest in the commodity, but it has not reached the purchase level, and the corresponding increment value and decrement value are between the purchase interaction and the browsing interaction. For example, when the user performs a purchase interaction, a certain commodity node is involved, and the importance parameter of the node will increase by a preset increment value. If it is a browsing interaction, the node importance parameter will decrease by a preset decrement value.

[0070] Step S1351: Set corresponding weight adjustment coefficients for different interaction types, wherein the weight adjustment coefficient of the purchase interaction is a first coefficient, the weight adjustment coefficient of the add-to-cart interaction is a second coefficient, the weight adjustment coefficient of the collection interaction is a third coefficient, and the weight adjustment coefficient of the browsing interaction is a fourth coefficient, and the first coefficient > the second coefficient > the third coefficient > the fourth coefficient, and the fourth coefficient is a negative value.

[0071] The weight adjustment coefficients are set for different interaction types, and these coefficients are used to adjust the node importance parameter. The weight adjustment coefficient of the purchase interaction is the largest because it represents the final decision behavior of the user and the importance of the product is the largest; the weight adjustment coefficient of the browsing interaction is a negative value because it may be just a casual browsing of the user and the importance of the product is reduced to a certain extent; the weight adjustment coefficients of the add-to-cart interaction and the collection interaction are between the two. For example, a larger weight adjustment coefficient is set for the purchase interaction, and a smaller negative weight adjustment coefficient is set for the browsing interaction.

[0072] Step S1352: Obtain the interaction duration of the real-time interaction behavior, and calculate the ratio of the interaction duration to the average interaction duration of the interaction type as a duration influence factor, and the value range of the duration influence factor is a preset interval.

[0073] The interaction duration of the real-time interaction behavior is obtained, which is compared with the average interaction duration of the interaction type, and the ratio is calculated, which is the duration influence factor. The duration influence factor reflects the situation that the time invested by the user in the interaction behavior is relative to the average time of the interaction type. The value range of the duration influence factor is a preset interval to ensure its rationality and effectiveness. For example, if the interaction duration of a purchase interaction of the user is longer, the duration influence factor will be larger accordingly; if the interaction duration is shorter, the duration influence factor will be smaller.

[0074] Step S1353: Multiply the weight adjustment coefficient and the duration influence factor to obtain a node importance adjustment value, wherein the node importance adjustment value of the purchase interaction, the add-to-cart interaction, and the collection interaction is a positive value, and the node importance adjustment value of the browsing interaction is a negative value.

[0075] The weight adjustment coefficient and the duration influence factor are multiplied to obtain the node importance adjustment value. For the purchase interaction, the add-to-cart interaction, and the collection interaction, since the weight adjustment coefficient is a positive value and the duration influence factor is also a positive value, the node importance adjustment value is a positive value, which will increase the importance of the node; for the browsing interaction, the weight adjustment coefficient is a negative value, and the node importance adjustment value is a negative value, which will reduce the importance of the node. For example, for a purchase interaction, the weight adjustment coefficient and the duration influence factor are multiplied to obtain a positive node importance adjustment value, which is used to increase the importance of the corresponding node.

[0076] Step S1354: Extracting the real-time interaction behavior involves the current node importance parameter of the commodity in the dynamic commodity collaborative network, adding the current node importance parameter and the node importance adjustment value to obtain an updated node importance parameter.

[0077] Extracting the real-time interaction behavior involves the current node importance parameter of the commodity in the dynamic commodity collaborative network, adding the current node importance parameter and the node importance adjustment value to obtain an updated node importance parameter. This updating process can timely reflect the influence of the real-time interaction behavior of the user on the importance of the commodity. For example, for a certain commodity node, its current node importance parameter is known, and by calculating the node importance adjustment value, the two are added to obtain an updated node importance parameter.

[0078] Step S1355: Boundary constraint processing is performed on the updated node importance parameter. If the updated node importance parameter is greater than the preset upper limit value, it is set to the preset upper limit value, and if it is less than the preset lower limit value, it is set to the preset lower limit value, to ensure that the node importance parameter is within the preset effective range.

[0079] In order to ensure the rationality and effectiveness of the node importance parameter, boundary constraint processing needs to be performed on the updated node importance parameter. The preset upper limit value and the preset lower limit value define the effective range of the node importance parameter. If the updated node importance parameter is greater than the preset upper limit value, it is set to the preset upper limit value; if it is less than the preset lower limit value, it is set to the preset lower limit value. For example, if the updated node importance parameter exceeds the preset upper limit value, it is adjusted to the preset upper limit value to ensure that the parameter is within a reasonable range.

[0080] Step S136: Periodic normalization processing is performed on the weights of all edges in the dynamic commodity collaborative network, so that the sum of the out-edge weights of each node is a fixed value. The normalization processing period is consistent with the node importance updating period.

[0081] Periodic normalization processing is performed on the weights of all edges in the dynamic commodity collaborative network, with the purpose of making the sum of the out-edge weights of each node a fixed value. This can ensure that the weights of the edges in the network are within a reasonable range, avoiding the situation of excessively large or small weights. The normalization processing period is consistent with the node importance updating period to ensure that the weights of the edges and the importance parameters of the nodes can be updated synchronously. For example, after each node importance updating period ends, normalization processing is performed on the weights of all edges in the network to keep the sum of the out-edge weights of each node fixed.

[0082] Step S140: performing path optimization mining processing in the dynamic commodity collaborative network, combining behavior transition correlation degree and node importance parameter to generate a user potential behavior path set, the potential behavior path set containing multiple paths with different attribute correlation characteristics.

[0083] After the dynamic commodity collaborative network is constructed, path optimization mining processing needs to be performed therein to generate a user potential behavior path set. These potential behavior paths can reflect the possible behavior trajectory of the user under the current interest. The specific mining process is as follows:

[0084] Step S141: determining the commodity involved in the last interaction behavior unit in the user behavior sequence as the starting node of path mining, and extracting the node importance parameter of the starting node in the dynamic commodity collaborative network as the path starting weight.

[0085] The commodity involved in the last interaction behavior unit in the user behavior sequence is determined as the starting node of path mining, because this commodity is the latest commodity of interest of the user, and is likely to be the focus of the current interest of the user. The node importance parameter of the starting node in the dynamic commodity collaborative network is extracted as the path starting weight, which reflects the importance of the starting node in the current interest of the user. For example, in the user's behavior sequence, the last interaction involves a latest smart watch, and the smart watch is taken as the starting node of path mining, and the node importance parameter thereof in the dynamic commodity collaborative network is extracted as the path starting weight.

[0086] Step S142: taking the starting node as the starting point, traversing all adjacent nodes directly connected to the starting node in the dynamic commodity collaborative network, taking the adjacent nodes as candidate next hop nodes, and collecting the node importance parameter of each candidate next hop node and the commodity attribute dynamic correlation degree of the connection edge.

[0087] Taking the starting node as the starting point, all adjacent nodes directly connected to the starting node in the dynamic commodity collaborative network are found out. These adjacent nodes are taken as candidate next hop nodes, which are the commodities that the user may pay attention to next. The node importance parameter of each candidate next hop node and the commodity attribute dynamic correlation degree of the connection edge are collected, which will be used for subsequent path selection. For example, for the starting node (smart watch), other commodity nodes (such as mobile phones, earphones, etc.) directly connected thereto in the network are found as candidate next hop nodes, and their node importance parameters and commodity attribute dynamic correlation degrees of the connection edges are collected.

[0088] Step S143: For each candidate next-hop node, extract the behavior transition correlation degree of the interaction behavior unit corresponding to the starting node in the user behavior sequence, multiply the behavior transition correlation degree, the node importance parameter of the candidate next-hop node, and the commodity attribute dynamic correlation degree of the connection edge to obtain a path transition probability.

[0089] For each candidate next-hop node, extract the behavior transition correlation degree of the interaction behavior unit corresponding to the starting node in the user behavior sequence. This behavior transition correlation degree reflects the possibility of the user transitioning from the interaction behavior corresponding to the starting node to the interaction behavior corresponding to the candidate next-hop node. Multiply the behavior transition correlation degree, the node importance parameter of the candidate next-hop node, and the commodity attribute dynamic correlation degree of the connection edge to obtain a path transition probability. This path transition probability comprehensively considers the possibility of behavior transition, the importance of the node, and the correlation degree between commodities, and can more accurately evaluate the possibility of the user selecting this path. For example, for a certain candidate next-hop node (mobile phone), extract the behavior transition correlation degree of the interaction behavior unit corresponding to the starting node (smart watch), multiply it with the importance parameter of the mobile phone node and the commodity attribute dynamic correlation degree of the connection edge to obtain the path transition probability from the smart watch to the mobile phone.

[0090] Step S144: Select the candidate next-hop node with the highest path transition probability as the second node of the path, take this node as the new starting node, and repeat the steps of adjacent node traversal, path transition probability calculation, and node selection until the path length reaches a preset threshold or it is impossible to continue traversing adjacent nodes, to generate a preliminary potential behavior path.

[0091] Selecting the candidate next-hop node with the highest path transition probability as the second node of the path means that the user is most likely to choose this path. Taking this node as the new starting node, repeatedly performing the steps of adjacent node traversal, path transition probability calculation, and node selection, and constantly expanding the path. This process will continue until the path length reaches a preset threshold or it is impossible to continue traversing adjacent nodes, at which time a preliminary potential behavior path is generated. For example, among the candidate next-hop nodes, the path transition probability of the mobile phone is the highest, so the mobile phone is taken as the second node of the path, and then the mobile phone is taken as the new starting node to continue subsequent path mining until the termination condition is met, to generate a preliminary potential behavior path.

[0092] Step S145: Perform path pruning processing on the preliminary potential behavior path, remove nodes and corresponding subsequent path segments whose node importance parameters are lower than a preset node threshold, and retain paths whose all node importance parameters are higher than the preset node threshold as effective potential behavior paths.

[0093] The preliminary potential behavior path is pruned to remove nodes with low importance and corresponding path segments, so as to improve the quality and effectiveness of the path. The preset node threshold is a preset standard for judging whether the importance of a node is sufficient. If the node importance parameter of a node is lower than the preset node threshold, it means that the importance of the node in the current interest of the user is low, and the node and the corresponding subsequent path segment are removed. All paths with node importance parameters higher than the preset node threshold are retained as effective potential behavior paths. For example, in the preliminary potential behavior path, the node importance parameter of a node (such as a less popular accessory) is lower than the preset node threshold, and the node and the subsequent path segment are removed. The path that meets the conditions is retained as an effective potential behavior path.

[0094] Step S1451: Traverse each node in the preliminary potential behavior path, and extract the node importance parameter of the node in sequence according to the path order.

[0095] Each node in the preliminary potential behavior path is traversed, and the node importance parameter of the node is extracted in sequence according to the path order. This process can systematically check the importance of each node in the path. For example, starting from the starting node of the preliminary potential behavior path, the node importance parameter of each node is extracted in sequence.

[0096] Step S1452: Compare the node importance parameter of each node with the preset node threshold, and mark the first node whose node importance parameter is lower than the preset node threshold as a pruning node.

[0097] The node importance parameter of each node is compared with the preset node threshold, and the first node whose node importance parameter is lower than the preset node threshold is found and marked as a pruning node. The pruning node is the first node with insufficient importance in the path, and needs to be pruned. For example, during the traversal process, a node whose node importance parameter is lower than the preset node threshold is found and marked as a pruning node.

[0098] Step S1453: If there is no pruning node in the preliminary potential behavior path, the path is directly used as an effective potential behavior path.

[0099] If there is no pruning node in the preliminary potential behavior path, it means that all node importance parameters in the path are higher than the preset node threshold, and the path meets the requirements and can be directly used as an effective potential behavior path. For example, after comparison, all node importance parameters in the preliminary potential behavior path are higher than the preset node threshold, so the path is an effective potential behavior path.

[0100] Step S1454: If there is a pruning node, remove the pruning node and all path segments after the node, and keep the path segments before the pruning node as the candidate pruning path.

[0101] If there is a pruning node, remove the pruning node and all path segments after the node, and keep the path segments before the pruning node as the candidate pruning path. This operation can remove unimportant parts of the path and improve the quality of the path. For example, when a pruning node is found, the node and subsequent path segments are deleted, and the previous path segments are kept as the candidate pruning path.

[0102] Step S1455: Calculate the path length of the candidate pruning path, and if the path length is less than the preset minimum path length, discard the candidate pruning path, otherwise keep the candidate pruning path as an effective potential behavior path.

[0103] Calculate the path length of the candidate pruning path and compare it with the preset minimum path length. If the path length is less than the preset minimum path length, it means that the path is too short and may not provide enough information, so it is discarded; otherwise, the candidate pruning path is kept as an effective potential behavior path. For example, the length of the candidate pruning path is calculated, and if the length is less than the preset minimum path length, the path is discarded; if the length meets the requirements, it is kept as an effective potential behavior path.

[0104] Step S146: Repeat the above steps to generate multiple effective potential behavior paths, so that different potential behavior paths contain different combinations of associated features of product attributes, forming a set of user potential behavior paths.

[0105] Repeat the above path mining and pruning steps to generate multiple effective potential behavior paths. These paths contain different combinations of associated features of product attributes and can more comprehensively reflect the potential behavior trajectory of the user. Collect these effective potential behavior paths to form a set of user potential behavior paths. For example, through multiple iterations of the path mining and pruning process, multiple different effective potential behavior paths are generated, which cover different combinations of products and associated features of attributes, and together form a set of user potential behavior paths.

[0106] Step S150: Analyze the set of potential behavior paths, and generate a personalized product recommendation sequence based on the distribution of the path's attribute association features and the user's current interaction behavior.

[0107] Analyze the generated set of potential behavior paths, combine the distribution of the path's attribute association features and the user's current interaction behavior, and generate a product recommendation sequence that meets the user's personalized needs. The specific analysis and generation process is as follows:

[0108] For example, step S151: extract the product attribute association features of each path in the set of potential behavior paths, which include the sequence distribution of product category attributes, the combination features of product function attributes, the coverage range of product scene attributes, and the change trend of product style attributes.

[0109] The product attribute association features of each path in the set of potential behavior paths are extracted, which can reflect the characteristics and association of products in the path. The sequence distribution of product category attributes shows the distribution of product categories in the path in terms of order, reflecting the user's interest shift between different categories of products. The combination features of product function attributes embody the matching and combination of product functions in the path, meeting the user's diversified needs in terms of function. The coverage range of product scene attributes represents the breadth and depth of the applicable scenarios of products in the path, which can adapt to the user's use needs in different scenarios. The change trend of product style attributes reflects the evolution of product styles in the path, embodying the user's preference for different styles. For example, in a certain potential behavior path, the sequence distribution of product category attributes may be from mobile phones to earphones to smart watches, the combination features of product function attributes may be the combination of functions such as taking pictures, playing music, and monitoring sports, the coverage range of product scene attributes may include daily use, sports scenarios, and office scenarios, and the change trend of product style attributes may be the transition from minimalist style to fashionable style.

[0110] Step S152: Calculate the attribute association feature similarity between any two paths, which is determined by comparing the dynamic time warping distance of category sequence distribution, the Jaccard similarity coefficient of function combination features, the intersection-over-union of scene coverage range, and the Pearson correlation coefficient of style change trend.

[0111] The attribute association feature similarity between any two paths needs to consider multiple aspects of features. For the category sequence distribution, dynamic time warping distance is used for comparison, which can handle different sequence lengths and more accurately measure the similarity between category sequences. For the function combination features, the Jaccard similarity coefficient is used for calculation, which can reflect the similarity between two sets. For the scene coverage range, the intersection-over-union is used to measure, which can intuitively show the overlap between two scene coverage ranges. For the style change trend, the Pearson correlation coefficient is used to determine, which can measure the linear correlation between two variables. By integrating these indicators, the attribute association feature similarity between any two paths can be obtained. For example, for two potential behavior paths, the dynamic time warping distance of their category sequence distribution, the Jaccard similarity coefficient of their function combination features, the intersection-over-union of their scene coverage range, and the Pearson correlation coefficient of their style change trend are calculated, and the attribute association feature similarity between them is obtained by integrating these indicators.

[0112] Step S153: Perform clustering processing on the set of potential behavior paths based on the attribute association feature similarity, and divide the paths with a similarity higher than a preset clustering threshold into the same path cluster. Each path cluster contains paths with similar attribute association features.

[0113] Perform clustering processing on the set of potential behavior paths based on the attribute association feature similarity, and divide the paths with a similarity higher than a preset clustering threshold into the same path cluster. The preset clustering threshold is a pre-set standard for determining whether the similarity of two paths is high enough. The paths in the same path cluster have similar attribute association features, which represent the similar interests and behavior patterns of users in some aspects. For example, divide the paths with a similarity higher than a preset clustering threshold into a path cluster, and the paths in the path cluster may all have similar commodity category sequence distribution, function combination feature, scene coverage range, and style change trend.

[0114] Step S154: Select the path with the highest path comprehensive score from each path cluster as the representative path, and the path comprehensive score is obtained by multiplying the reciprocal of the path length by the cumulative value of the commodity attribute dynamic association degree in the path.

[0115] Select the path with the highest path comprehensive score from each path cluster as the representative path, and the calculation method of the path comprehensive score is to multiply the reciprocal of the path length by the cumulative value of the commodity attribute dynamic association degree in the path. The reciprocal of the path length reflects the simplicity of the path, and the shorter the path, the larger the reciprocal, indicating that the path is more efficient. The cumulative value of the commodity attribute dynamic association degree in the path reflects the degree of association between commodities in the path, and the larger the cumulative value, the closer the association between commodities. By multiplying the two, the simplicity of the path and the degree of association between commodities can be considered comprehensively to select the optimal representative path. For example, for each path cluster, calculate the comprehensive score of each path in the path cluster, and select the path with the highest score as the representative path.

[0116] Step S155: Collect all representative paths, extract the commodity nodes in the representative paths, and arrange them in the order of the paths to form an initial recommendation sequence.

[0117] Collect all representative paths of the path clusters, extract the commodity nodes from the representative paths, and arrange them in the order of the paths to form an initial recommendation sequence. The commodities in the initial recommendation sequence are selected based on the potential behavior paths and attribute association features of the user, and have certain relevance and rationality. For example, arrange the commodity nodes in all representative paths in the order of the paths to form an initial commodity recommendation sequence.

[0118] Step S156: Collect the current interaction behavior of the user, the current interaction behavior being the interaction behavior unit generated by the user within a preset time before generating the recommendation sequence, and extract the commodity attribute information of the commodities involved in the current interaction behavior.

[0119] The current interaction behavior of the user is collected, which is the interaction behavior unit generated by the user within a preset time before generating the recommendation sequence. The commodity attribute information of the commodities involved in the current interaction behavior is extracted, which can reflect the current interest and demand of the user. For example, within a period of time before generating the recommendation sequence, the user interacts with a certain mobile phone, and the commodity attribute information of the mobile phone is extracted, including category attribute, function attribute, scene attribute, and style attribute, etc.

[0120] Step S157: Filtering the initial recommendation sequence according to the commodity attribute information of the current interaction behavior, removing the commodity nodes whose attribute correlation with the commodity attribute information of the current interaction behavior is lower than a preset filtering threshold, and then sorting the filtered commodity nodes according to the order in the representative path and the path comprehensive score to generate the final commodity personalized recommendation sequence.

[0121] The initial recommendation sequence is filtered according to the commodity attribute information of the current interaction behavior. The preset filtering threshold is a pre-set standard for judging whether the attribute correlation between commodities is high enough. If the attribute correlation between a certain commodity node in the initial recommendation sequence and the commodity attribute information of the current interaction behavior is lower than the preset filtering threshold, it means that the commodity is not closely related to the current interest of the user, and it is removed. Then, the filtered commodity nodes are sorted according to the order in the representative path and the path comprehensive score to generate the final commodity personalized recommendation sequence. The commodity personalized recommendation sequence can better meet the personalized needs of the user and provide more suitable commodity recommendations for the current interest of the user. For example, for the commodity nodes in the initial recommendation sequence, the attribute correlation between them and the commodity attribute information of the current interaction behavior is calculated, the nodes with correlation lower than the preset filtering threshold are removed, and the remaining nodes are sorted to generate the final commodity personalized recommendation sequence.

[0122] Figure 2 A schematic diagram of exemplary hardware and software components of the commodity personalized recommendation system 100 based on user behavior data analysis that can implement the idea of the present application provided by some embodiments of the present application is shown. For example, the processor 120 can be used in the commodity personalized recommendation system 100 based on user behavior data analysis and used to perform the functions in the present application.

[0123] The user behavior data analysis based commodity personalized recommendation system 100 can be a general server or a special purpose server, both of which can be used to implement the user behavior data analysis based commodity personalized recommendation method of the present application. The present application only shows one server, but for the sake of convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0124] For example, the user behavior data analysis based commodity personalized recommendation system 100 can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, a ROM, or a RAM, or any combination thereof. The user behavior data analysis based commodity personalized recommendation system 100 can also include program instructions stored in a ROM, a RAM, or other types of non-transitory storage media, or any combination thereof, for example. The method of the present application can be implemented according to these program instructions. The user behavior data analysis based commodity personalized recommendation system 100 also includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0125] For the sake of illustration, only one processor is described in the user behavior data analysis based commodity personalized recommendation system 100. However, it should be noted that the user behavior data analysis based commodity personalized recommendation system 100 in the present application can also include multiple processors, so the steps performed by one processor described in the present application can also be jointly performed or separately performed by multiple processors. For example, if the processor of the user behavior data analysis based commodity personalized recommendation system 100 performs steps A and B, it should be understood that steps A and B can also be jointly performed by two different processors or separately performed in one processor. For example, a first processor performs step A and a second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0126] In addition, the present application also provides a readable storage medium, in which computer executable instructions are pre-set, and when the processor executes the computer executable instructions, the user behavior data analysis based commodity personalized recommendation method described above is implemented.

[0127] It should be noted that, in order to simplify the description of the present application and to help understand one or more embodiments of the present application, in the foregoing description of the embodiments of the present application, various features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A method for personalized product recommendation based on user behavior data analysis, characterized in that, The method includes: Construct a user behavior sequence and interest stability model. The user behavior sequence includes user and product interaction behavior units arranged in chronological order. Each interaction behavior unit is associated with interaction type and product attribute information. The interest stability model is used to quantify the sustained characteristics of user interest in product attributes. The user behavior sequence is subjected to association modeling processing. The behavior transfer correlation degree of adjacent interactive behavior units and the dynamic correlation degree of product attributes are calculated by combining the interest stability model. The dynamic correlation degree of product attributes is adjusted with the change of user interest stability. A dynamic product collaboration network is constructed based on the dynamic correlation of product attributes. The nodes of the dynamic product collaboration network are products, the edges are the dynamic correlation of product attributes, and the node importance parameters are updated according to the real-time interaction behavior of users. In a dynamic product collaboration network, path optimization mining is performed to generate a set of potential user behavior paths by combining behavior transfer correlation degree and node importance parameters. The set of potential behavior paths contains multiple paths with different attribute correlation characteristics. Analyze the set of potential behavioral paths, and generate personalized product recommendation sequences based on the distribution of attribute association features of the paths and the user's current interaction behavior; The process of performing path optimization and mining in the dynamic product collaboration network, combining behavior transfer correlation and node importance parameters to generate a set of potential user behavior paths, includes: The product involved in the last interactive behavior unit in the user behavior sequence is determined as the starting node of path mining, and the node importance parameter of the starting node in the dynamic product collaboration network is extracted as the starting weight of the path. Starting from the initial node, traverse all adjacent nodes directly connected to the initial node in the dynamic product collaboration network, and use the adjacent nodes as candidate next-hop nodes. Collect the node importance parameters of each candidate next-hop node and the dynamic correlation degree of product attributes of the connecting edges. For each candidate next-hop node, extract the behavior transition correlation degree of the interactive behavior unit corresponding to the starting node in the user behavior sequence, and multiply the behavior transition correlation degree, the node importance parameter of the candidate next-hop node and the dynamic correlation degree of the product attribute of the connecting edge to obtain the path transition probability. Select the candidate next-hop node with the highest path transition probability as the second node of the path, and use this node as the new starting node. Repeat the steps of traversing adjacent nodes, calculating path transition probability, and selecting nodes until the path length reaches a preset threshold or it is no longer possible to traverse adjacent nodes, thus generating a preliminary potential behavioral path. The initial potential behavior paths are pruned to remove nodes whose importance parameters are lower than the preset node threshold and their corresponding subsequent path segments, and retain paths whose importance parameters are all higher than the preset node threshold as valid potential behavior paths. Repeat the above steps to generate multiple valid potential behavior paths, so that different potential behavior paths contain different combinations of product attribute associated features, forming a set of user potential behavior paths.

2. The personalized product recommendation method based on user behavior data analysis according to claim 1, characterized in that, The construction of the user behavior sequence and interest stability model includes: Collect users' original interaction records throughout the entire interaction cycle. The original interaction records include the time of interaction, the products involved in the interaction, the type of interaction, and the duration of the interaction. The interaction types include browsing interaction, favorites interaction, add to cart interaction, and purchase interaction. The original interaction records are divided into behavioral units. Continuous interaction records of the same product within a preset time window are merged into one interactive behavioral unit. The interaction type of the interactive behavioral unit is the set of all interaction types contained in the merged records, and the interaction duration is the total interaction duration of the merged records. Extract the product attribute information of each interactive behavior unit involving the product. The product attribute information includes product category attribute, product function attribute, product scenario attribute and product style attribute. The product scenario attribute is used to indicate the applicable usage scenario of the product, and the product style attribute is used to indicate the design style characteristics of the product. The segmented interactive behavior units are sorted according to the order of their occurrence time to generate a preliminary user behavior sequence. The preliminary user behavior sequence is then checked for missing interactions. Interactive behavior units corresponding to key product attributes that are missing due to interruption of interaction records are added to obtain a complete user behavior sequence. An interest stability model is constructed based on the complete user behavior sequence. The frequency of occurrence of the same product attribute information in different time windows in the complete user behavior sequence is extracted, and the fluctuation coefficient of the frequency change is calculated. The reciprocal of the fluctuation coefficient is used as the user's interest stability parameter for the product attribute. The higher the interest stability parameter, the more stable the user's interest in the product attribute.

3. The personalized product recommendation method based on user behavior data analysis according to claim 2, characterized in that, The interest stability model constructed based on the complete user behavior sequence extracts the frequency of occurrence of the same product attribute information in different time windows within the complete user behavior sequence, calculates the fluctuation coefficient of the frequency change, and uses the reciprocal of the fluctuation coefficient as the user's interest stability parameter for that product attribute, including: The time span of the complete user behavior sequence is evenly divided into multiple consecutive time windows, each with the same duration and no overlap between time windows; For each product attribute information, the number of interactive behavior units that appear within each time window is counted, and the ratio of this number to the total number of interactive behavior units within the time window is taken as the frequency of occurrence of the product attribute information within that time window. Arrange the occurrence frequencies of all time windows in chronological order to form a frequency sequence, and calculate the mean of the frequency sequence, which is the arithmetic mean of all occurrence frequencies. Calculate the square of the difference between each occurrence frequency and the mean in the frequency sequence, sum all the squares and divide by the number of time windows to obtain the variance of the frequency change, and use the square root of the variance as the fluctuation coefficient of the frequency change. The reciprocal of the volatility coefficient is used as the user's interest stability parameter for the product attribute information. If the volatility coefficient is zero, the interest stability parameter is set to the preset maximum value. The value range of the interest stability parameter is consistent with the value range of the dynamic correlation degree of the product attribute.

4. The personalized product recommendation method based on user behavior data analysis according to claim 1, characterized in that, The process of performing association modeling on user behavior sequences, and calculating the behavioral transition association degree and product attribute dynamic association degree of adjacent interactive behavior units in conjunction with the interest stability model, includes: Identify the interaction type and duration of each interactive behavior unit in the user behavior sequence, and assign basic behavior weights to the interactive behavior units according to the interaction type. The basic behavior weights increase with the decision depth of the interaction type. Among them, the basic behavior weight of purchase interaction is greater than that of add-to-cart interaction, the basic behavior weight of add-to-cart interaction is greater than that of favorite interaction, and the basic behavior weight of favorite interaction is greater than that of browse interaction. The basic behavior weight is corrected based on the interaction duration. The ratio of the interaction duration to the average interaction duration of this interaction type is calculated. This ratio is used as the duration correction factor. The basic behavior weight is multiplied by the duration correction factor to obtain the corrected behavior weight value. Extract the interaction time of two adjacent interactive behavior units in the user behavior sequence, calculate the time interval parameter, construct a time decay function based on the time interval parameter, the output value of the time decay function decreases as the time interval parameter increases, and multiply the corrected behavior weight value with the output value of the time decay function to obtain the time decay behavior weight. The interest stability model is called to obtain the interest stability parameters corresponding to the product attribute information of each adjacent interactive behavior unit, and the product of the two interest stability parameters is used as the attribute stability factor. Extract product attribute information from adjacent interactive behavior units, calculate the overlap of product category attributes, the semantic matching degree of product function attributes, the scene fit of product scenario attributes, and the style similarity of product style attributes. After assigning preset weights to the overlap of categories, the semantic matching degree of functions, the scene fit, and the style similarity, the basic attribute correlation degree is obtained by weighted summation. Multiplying the basic attribute correlation degree by the attribute stability factor yields the dynamic correlation degree of the product attribute, and multiplying the time decay behavior weight by the dynamic correlation degree of the product attribute yields the behavior transfer correlation degree of adjacent interactive behavior units.

5. The personalized product recommendation method based on user behavior data analysis according to claim 4, characterized in that, The step of extracting product attribute information from adjacent interactive behavior units and calculating the overlap of product category attributes, the semantic matching degree of product functional attributes, the scene fit of product scene attributes, and the style similarity of product style attributes includes: The product category attributes of adjacent interactive behavior units are split into multiple category labels. The number of identical category labels in the two product category attributes is counted. The ratio of the number of identical category labels to the total number of category labels in the two product category attributes is used as the overlap of product category attributes. Text preprocessing is performed on the product functional attributes of adjacent interactive behavior units. Stop words are removed and core functional description words are retained. The core functional description words are input into a pre-trained semantic encoder to generate functional attribute vectors. The cosine similarity between two functional attribute vectors is calculated as the semantic matching degree of the product functional attributes. Extract scene tags contained in the product scene attributes. Each scene tag corresponds to a preset scene weight. Calculate the sum of scene weights of the same scene tags in the scene attributes of two products. The ratio of the sum of scene weights to the sum of scene weights of all scene tags in the scene attributes of the two products is taken as the scene fit of the product scene attributes. The product style attributes are converted into style feature vectors. The dimensions of the style feature vectors correspond to a preset set of style dimensions. The value of each dimension is the feature value of that style dimension. The Euclidean distance between two style feature vectors is calculated. The reciprocal of the Euclidean distance is used as the style similarity of the product style attributes. If the Euclidean distance is zero, the style similarity is set to the preset maximum value.

6. The personalized product recommendation method based on user behavior data analysis according to claim 1, characterized in that, The construction of a dynamic product collaboration network based on the dynamic correlation of product attributes includes: Collect all products involved in the interactive behavior units in the user behavior sequence, remove duplicate products to form a product set, and use each product in the product set as the initial node of the dynamic product collaboration network. For any two products in the product set, if there are at least one adjacent interactive behavior unit in the user behavior sequence that involves the two products respectively, then the dynamic correlation degree of the product attribute corresponding to the adjacent interactive behavior unit is extracted as the initial weight of the edge connecting the two product nodes. If there are multiple connecting edges between the two products, then the average of all the initial weights is taken as the final edge weight. Initialize the node importance parameter for each node in the dynamic product collaboration network. The initial node importance parameter is the ratio of the number of interactive behavior units that the product appears in the user behavior sequence to the total number of interactive behavior units. Set the node importance update cycle, and collect the user's real-time interaction behavior in each update cycle. The real-time interaction behavior is the interaction behavior unit generated by the user in the current update cycle. Based on the node importance parameter of the corresponding node for product updates involved in real-time interactive behavior, if the interaction type of real-time interactive behavior is purchase interaction, the node importance parameter is increased by a preset incremental value; if it is browsing interaction, the node importance parameter is decreased by a preset decrement value. The incremental and decrement values ​​corresponding to add-to-cart and favorite interactions are between purchase interaction and browsing interaction. The weights of all edges in the dynamic commodity collaborative network are periodically normalized so that the sum of the weights of the outgoing edges of each node is a fixed value. The normalization period is consistent with the node importance update period.

7. The personalized product recommendation method based on user behavior data analysis according to claim 6, characterized in that, The node importance parameter is updated based on the product update involved in the real-time interaction. If the interaction type is a purchase interaction, the node importance parameter is increased by a preset increment value; if it is a browsing interaction, the node importance parameter is decreased by a preset decrement value. The increment and decrement values ​​corresponding to add-to-cart and favorite interactions are between those of purchase and browsing interactions, including: Set corresponding weight adjustment coefficients for different interaction types. The weight adjustment coefficient for purchase interaction is the first coefficient, the weight adjustment coefficient for add-to-cart interaction is the second coefficient, the weight adjustment coefficient for favorite interaction is the third coefficient, and the weight adjustment coefficient for browse interaction is the fourth coefficient. The first coefficient > the second coefficient > the third coefficient > the fourth coefficient, and the fourth coefficient is a negative value. The interaction duration of real-time interactive behavior is obtained, and the ratio of the interaction duration to the average interaction duration of this interaction type is calculated as the duration influence factor. The value range of the duration influence factor is a preset interval. The node importance adjustment value is obtained by multiplying the weight adjustment coefficient by the duration influence factor. The node importance adjustment values ​​for purchase interaction, add-to-cart interaction, and favorite interaction are positive, while the node importance adjustment value for browsing interaction is negative. Extract the current node importance parameters of the products involved in real-time interactive behavior in the dynamic product collaboration network, and add the current node importance parameters to the node importance adjustment value to obtain the updated node importance parameters; Boundary constraints are applied to the updated node importance parameters. If the updated node importance parameters are greater than the preset upper limit, they are set to the preset upper limit; if they are less than the preset lower limit, they are set to the preset lower limit, so that the node importance parameters are within the preset valid range.

8. The personalized product recommendation method based on user behavior data analysis according to claim 1, characterized in that, The step of pruning the initial potential behavior paths involves removing nodes whose importance parameters are below a preset threshold and their corresponding subsequent path segments, while retaining paths where all node importance parameters are above the preset threshold as valid potential behavior paths. This includes: Traverse each node in the initial potential behavior path and extract the node importance parameters in the order of the path. The node importance parameter of each node is compared with a preset node threshold, and the first node whose node importance parameter is lower than the preset node threshold is marked as a pruning node; If there are no pruned nodes in the initial potential behavior path, then the path is directly regarded as a valid potential behavior path. If a pruning node exists, remove the pruning node and all path segments after it, and keep the path segments before the pruning node as candidate pruning paths. Calculate the path length of the candidate pruning path. If the path length is less than the preset minimum path length, discard the candidate pruning path; otherwise, treat the candidate pruning path as a valid potential behavior path. Repeat the above pruning process for all preliminary potential behavioral paths, collect all valid potential behavioral paths to form an intermediate path set, so that the node importance parameters of the paths in the intermediate path set all meet the preset node threshold requirements.

9. A personalized product recommendation system based on user behavior data analysis, characterized in that, The device includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the personalized product recommendation method based on user behavior data analysis as described in any one of claims 1-8.

Images