A method for generating user profiles for e-commerce operation platforms based on big data.

Abstract

This invention relates to the field of user profile generation technology, specifically to a method for generating user profiles for e-commerce operation platforms based on big data. This method determines comprehensive correlation indicators based on the similarity and time difference of adjacent behavior vectors in a behavior vector sequence; groups adjacent behavior vectors with multiple consecutive comprehensive correlation indicators meeting preset session conditions into the same session, forming a session vector sequence; preliminarily groups sessions based on the directional differences between behavior vectors within a session and the total session duration, and assigns group labels to the sessions; determines core theme vectors based on the central tendency of behavior vectors within each session; groups sessions of the same user whose cosine similarity between core theme vectors is higher than a preset threshold into the same macro shopping path; generates an intent state sequence based on the mapping relationship between preset group labels and intent states; and generates user profile labels based on the intent state sequence, thereby improving the accuracy of user profiles.

Description

Technical Field

[0001] This invention relates to the field of user profile generation technology, specifically to a method for generating user profiles for e-commerce operation platforms based on big data. Background Technology

[0002] With the rapid development of the e-commerce industry and the continuous growth of platform users, extracting effective information from massive amounts of user data has become crucial for e-commerce platform operations. User profiling is a core technological tool for achieving this goal. It collects and analyzes user behavior data to abstract tagged user models, thereby supporting various downstream applications.

[0003] Currently, the mainstream user profiling methods for e-commerce operation platforms are mainly based on tag-based modeling and statistical analysis. They construct profiles by integrating multi-dimensional information such as basic user attributes and behavioral data, or rely on single-dimensional behavioral data (such as purchase records and browsing history). However, although these methods are simple to implement, they have significant limitations. They only perform static and isolated statistics on users' historical behavior, and only identify users' intentions in a one-sided way. They are prone to ignoring the continuous evolution and dynamic changes of users' intentions, resulting in profiles that are one-dimensional and too superficial, making it difficult to support the needs of subsequent refined applications. Summary of the Invention

[0004] To address the technical problem of decreased accuracy in user profiling caused by static, isolated statistics of user historical behavior or reliance on single-dimensional behavioral data, this invention provides a method for generating user profiles for e-commerce operation platforms based on big data. The specific technical solution adopted is as follows:

[0005] This invention proposes a method for generating user profiles for e-commerce operation platforms based on big data. The method includes:

[0006] Collect user interaction data on e-commerce platforms and convert it into behavior vectors, then generate a sequence of user behavior vectors in chronological order of the behavior occurrence.

[0007] Based on the similarity and time difference of adjacent behavior vectors in the behavior vector sequence, a comprehensive correlation index is determined; adjacent behavior vectors with multiple consecutive comprehensive correlation indices that meet the preset session conditions are classified into the same session and form a session vector sequence.

[0008] For each session vector sequence, based on the directional differences between behavioral vectors within the session, an intent similarity index is determined; and combined with the total session duration, all user sessions are initially grouped, and each session is assigned a group label representing the user's micro-intent.

[0009] Based on the central tendency of behavior vectors within each session, core topic vectors are determined; sessions with a cosine similarity between core topic vectors of the same user that is higher than a preset threshold are grouped into the same macro shopping path; based on the preset mapping relationship between group tags and intent states, the group tags of each session within the same macro shopping path are mapped to the corresponding intent states, and the user's intent state sequence is generated by arranging them in the order of session occurrence time.

[0010] State transition features are extracted from the intent state sequence, and labels representing the evolution of the user's core intent are generated based on the state transition features to form a user profile.

[0011] Furthermore, the process for determining the comprehensive correlation index includes:

[0012] For each sequence of action vectors, calculate the absolute difference between the timestamps of adjacent action vectors, and use this as the time difference.

[0013] Calculate the cosine similarity of adjacent action vectors as semantic similarity;

[0014] Based on preset weight coefficients, the normalized semantic similarity and normalized time difference are weighted and calculated to obtain a comprehensive association index, in which the weight of semantic similarity is greater than the weight of time difference.

[0015] Furthermore, the preset session conditions include a high correlation threshold and a low correlation threshold, with the high correlation threshold being greater than the low correlation threshold; classifying adjacent behavior vectors that meet the preset session conditions for multiple consecutive comprehensive correlation indicators into the same session includes:

[0016] Using the first behavior vector in the behavior vector sequence as the starting vector of the current session, the comprehensive correlation index of each subsequent behavior vector with its preceding adjacent behavior vector is judged to be greater than the high correlation threshold or less than the low correlation threshold, and the judgment result is obtained.

[0017] If the judgment result indicates that the comprehensive correlation index is greater than the high correlation threshold, then the behavior vector of the current judgment will be included in the current session;

[0018] If the judgment result indicates that the comprehensive correlation index is less than the low correlation threshold, the current session is terminated, and the behavior vector of the current judgment is used as the starting vector of the new session.

[0019] Furthermore, the method also includes:

[0020] If the judgment result indicates that the comprehensive correlation index is between the low correlation threshold and the high correlation threshold, then the comprehensive correlation index between the currently judged behavior vector and the subsequent adjacent behavior vector is further judged; if the comprehensive correlation index between the currently judged behavior vector and the subsequent adjacent behavior vector is greater than the high correlation threshold, then the currently judged behavior vector is included in the current session; otherwise, the current session is terminated, and the currently judged behavior vector is used as the starting vector of the new session.

[0021] Furthermore, the process for determining the intent similarity index includes:

[0022] Calculate the cosine distance between any two behavior vectors within the session vector sequence, and use it as the directional difference value;

[0023] Calculate the arithmetic mean of all directional differences within the session vector sequence, and use it as the directional difference mean.

[0024] The difference between the positive integer 1 and the mean of directional differences is calculated as an indicator of intent similarity.

[0025] Furthermore, combining the total session duration, all user sessions are initially grouped, and each session is assigned a group label representing the user's micro-intent, including:

[0026] Using intent similarity metrics as the first dimension and total session duration as the second dimension, a two-dimensional feature point is constructed for each session;

[0027] Based on the two-dimensional feature points of all sessions, a clustering algorithm is used for grouping.

[0028] Based on the distribution characteristics of the two-dimensional feature points in each group in the first and second dimensions after grouping, a group label is assigned to each group.

[0029] Furthermore, the group label includes at least one of the following types: a first type corresponding to high intent similarity index and short total session consumption time; a second type corresponding to high intent similarity index and long total session consumption time; and a third type corresponding to low intent similarity index.

[0030] Furthermore, the process of determining the core topic vector includes:

[0031] Calculate the arithmetic mean of all behavior vectors within the session vector sequence, and use it as the core topic vector.

[0032] Furthermore, the step of extracting state transition features from the intent state sequence and generating labels representing the evolutionary pattern of the user's core intent based on the state transition features, as a user profile, includes:

[0033] Extract the transition patterns between all adjacent intent states from the user's intent state sequence;

[0034] The frequency of each transition pattern is counted, and the transition pattern with the highest frequency is identified as the core transition pattern; the Shannon entropy of all transition patterns is calculated as an indicator of the orderliness of intentional evolution.

[0035] The core transfer pattern is combined with the intent evolution orderliness index to serve as a user segmentation feature;

[0036] Based on the grouping characteristics of all users, a clustering algorithm is used to group users. According to the commonality of decision-making paths represented by the core transfer patterns of each user group after grouping, and the commonality of stability represented by the intention evolution orderliness index, each user group is assigned a corresponding user profile label as a user profile.

[0037] Furthermore, the rules for assigning user profile tags include: user groups with progressive decision paths and low intention evolution orderliness indicators are assigned a first type of tag; user groups with cyclical decision paths and high intention evolution orderliness indicators are assigned a second type of tag; and user groups with progressive decision paths but high intention evolution orderliness indicators are assigned a third type of tag.

[0038] The present invention has the following beneficial effects:

[0039] By comprehensively considering correlation indicators (integrating behavioral semantic similarity and time difference) and judgment rules to segment sessions, this approach effectively distinguishes between "continuous intent behavior" and "temporary distraction behavior," intelligently dividing user behavior data into semantically coherent sessions. This makes session boundaries more closely aligned with the user's actual decision-making process, laying a data foundation for subsequent intent analysis. Combining micro-level session characteristics (intent similarity indicators, total session duration) with macro-level evolutionary characteristics (core transfer patterns, intent evolution orderliness indicators), this approach captures the intent of a single session. Figure 1 It improves consistency and decision-making efficiency, and also uncovers the patterns of intent transfer across conversations, achieving a deep portrayal from "behavioral results" to "intent motivations." It deeply analyzes the complete lifecycle of users from generating interest, conducting in-depth research to making a final decision. Through multiple rounds of grouping, it ultimately generates dynamic user profile tags that reflect the user's decision-making process. This not only enriches the dimensions of the user profile but also improves the accuracy of the user profile, more truly reflecting the user's core intent. Attached Figure Description

[0040] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating a user profile generation method for an e-commerce operation platform based on big data, provided as an embodiment of the present invention;

[0042] Figure 2 An example diagram illustrating an intent similarity index determination process provided in one embodiment of the present invention;

[0043] Figure 3 This is an example diagram illustrating the user profile generation process provided in one embodiment of the present invention. Detailed Implementation

[0044] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a user profile generation method for an e-commerce operation platform based on big data proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0046] The following description, in conjunction with the accompanying drawings, details a specific scheme for a user profile generation method for an e-commerce operation platform based on big data, provided by this invention.

[0047] Please see Figure 1 The diagram illustrates a flowchart of a user profile generation method for an e-commerce operation platform based on big data, according to an embodiment of the present invention. The method includes:

[0048] S101: Collect user interaction data on e-commerce platforms and convert it into behavior vectors, then generate a sequence of user behavior vectors in chronological order of the behavior occurrence.

[0049] It should be noted that the interactive behavior data covers the entire user operation on the e-commerce platform, including core behaviors such as browsing product details, searching keywords, adding items to the shopping cart, and submitting orders, as well as auxiliary behaviors such as viewing reviews, saving products, and sharing links. In addition, when collecting user behavior data, metadata such as the timestamp of the behavior and the device type can also be associated.

[0050] It should be noted that a pre-trained embedding model can be used to transform unstructured interaction behavior data into structured behavior vectors. The pre-trained embedding model is trained based on the user's historical interaction behavior data. The specific training method of the pre-trained embedding model is a well-known technique to those skilled in the art and will not be described in detail in this embodiment. For example, Word2Vec or BERT models.

[0051] To fully preserve the temporal logic of user behavior and provide a coherent original data chain for subsequent session segmentation and intent analysis, and to avoid excessive data volume affecting user intent analysis, behavior vectors within a preset time window (such as the month preceding the current moment) can be selected to generate a behavior vector sequence.

[0052] S102: Determine the comprehensive correlation index based on the similarity and time difference of adjacent behavior vectors in the behavior vector sequence; classify adjacent behavior vectors with multiple consecutive comprehensive correlation indices that meet the preset session conditions into the same session, and form a session vector sequence.

[0053] It's important to understand that a user's complete decision typically begins with a triggering event (such as a search) and ends with a terminating event (such as a purchase or departure), with strong causal relationships between the actions in between. Generally speaking, the closer two actions are in time, the greater the probability that they belong to the same session (i.e., the same complete decision). However, due to the significant differences in behavior among different users and in different scenarios (e.g., actions taken in seconds during a flash sale versus minutes while browsing), it is also crucial to consider the semantic similarity of adjacent actions.

[0054] In this embodiment, for each behavior vector sequence, the absolute difference between the timestamps of adjacent behavior vectors is calculated as the time difference; the cosine similarity between adjacent behavior vectors is calculated as the semantic similarity; based on preset weight coefficients, the normalized semantic similarity and the normalized time difference are weighted and calculated to obtain a comprehensive association index.

[0055] Semantic similarity is used to characterize the degree of semantic association between two behaviors (e.g., "browsing phone A" and "adding phone A to shopping cart" have a strong semantic association, while "browsing phone" and "browsing snacks" have a weaker semantic association).

[0056] The time difference is used to characterize the degree of correlation between two actions in the time dimension (e.g., actions that occur consecutively within 1 minute are more correlated than actions that are 24 hours apart).

[0057] It should be noted that the algorithm for calculating cosine similarity is an existing technical method, and will not be described in detail in this embodiment.

[0058] It should be noted that, since the probability of behaviors with stronger semantic association belonging to the same session is greater than that of time difference, the weight of semantic similarity is greater than that of time difference.

[0059] To unify dimensions and balance the influence of semantics and time, the semantic similarity and time difference are first normalized, and then a comprehensive association index is obtained by weighted summation. Therefore, the comprehensive association index can be expressed as: Comprehensive association index = α × normalized semantic similarity + (1-α) × (1- normalized time difference), where α represents the preset weight coefficient.

[0060] It should be noted that the specific value of the preset weight coefficient is determined according to the actual situation, and this embodiment does not impose a specific limitation. For example, it can be determined by historical session data, giving priority to the influence of semantic association, and the preset weight coefficient can be in the range of 0.6-0.8.

[0061] It is important to understand that if the normalized semantic similarity of two behavior vectors is greater, it indicates that the semantic relationship between the two behaviors is closer; if the normalized time difference between two behavior vectors is smaller, it indicates that the two behaviors are more continuous in time and the relationship may be stronger. Therefore, the larger the comprehensive correlation index, the stronger the overall correlation (semantics + time) between adjacent behaviors and the greater the probability that they belong to the same session.

[0062] It is important to understand that a session refers to a series of related behaviors that a user generates around the same core intent within a continuous period of time (such as "search → browse → compare prices → purchase" can be regarded as a "purchase decision session"). However, since different users have different focuses, their decision-making behaviors will also be different. Some users may also exhibit unrelated behaviors across different time periods when conducting the same decision session. Therefore, by comprehensively analyzing related indicators, we can further analyze behaviors belonging to the same session.

[0063] It should be noted that the preset session conditions include a high association threshold and a low association threshold, and the high association threshold is greater than the low association threshold.

[0064] In this embodiment, the first behavior vector in the behavior vector sequence is used as the starting vector of the current session. The comprehensive correlation index of each subsequent behavior vector and its preceding adjacent behavior vector is sequentially judged to determine whether it is greater than the high correlation threshold or less than the low correlation threshold, and the judgment result is obtained. If the judgment result indicates that the comprehensive correlation index is greater than the high correlation threshold, the currently judged behavior vector is included in the current session. If the judgment result indicates that the comprehensive correlation index is less than the low correlation threshold, the current session is terminated, and the currently judged behavior vector is used as the starting vector of the new session.

[0065] It should be noted that the specific values ​​of the high correlation threshold and the low correlation threshold are determined according to the actual situation, and this embodiment does not impose specific limitations. For example, the high correlation threshold can be 0.7 and the low correlation threshold can be 0.3.

[0066] To avoid including irrelevant behaviors across time periods in irrelevant sessions, as an example, if the judgment result indicates that the comprehensive correlation index is between the low correlation threshold and the high correlation threshold, then the comprehensive correlation index between the currently judged behavior vector and the subsequent adjacent behavior vector is further judged; if the comprehensive correlation index between the currently judged behavior vector and the subsequent adjacent behavior vector is greater than the high correlation threshold, then the currently judged behavior vector is included in the current session; otherwise, the current session is terminated, and the currently judged behavior vector is used as the starting vector of the new session.

[0067] It should be noted that all behavior vectors belonging to the same session are arranged in chronological order of their occurrence to form a corresponding session vector sequence.

[0068] For example, suppose a user's behavior vector sequence is [ , , , , High correlation threshold ( The correlation threshold is 0.7, which is low. The value is 0.3, so the behavior vector is... As the starting vector for the first session, the current session, i.e., session 1, is [ ],So, and The comprehensive correlation index is 0.8 (greater than 0.8). → Included in the current session, session 1 is [ , ]; and The comprehensive correlation index is 0.2 (less than 0.2). → Terminate Session 1 As the starting vector for the new session, session 2 is initially set to [ ]; and The overall correlation index is 0.9 (greater than 0.9). → Incorporated into Session 2, Session 2 updated to [ , ]; and The comprehensive correlation index is 0.8 (greater than 0.8). → Incorporated into Session 2, Session 2 ultimately becomes [ , , Therefore, we end up with two independent sessions: [ , ]and[ , , These correspond to two independent intent actions of the user.

[0069] S103: For each session vector sequence, determine the intent similarity index based on the directional difference characteristics between behavioral vectors within the session; and combine the total session duration to initially group all user sessions and assign group labels representing user micro-intents to the sessions.

[0070] It's important to understand that since all of a user's searches and browsing within a single session revolve around the same theme, the behavioral vectors will cluster within a narrow area in the vector space. Therefore, by analyzing the degree of concentration of a user's intent within a single session, we can distinguish whether the user has a clear goal within that session. Having a clear goal means that the user's behavior revolves around the same core intent (e.g., searching → comparing prices → buying a certain mobile phone). If there is no clear goal, it indicates that the behavior is scattered and the goal is vague (e.g., randomly browsing clothing, snacks, and home appliances recommended on the homepage).

[0071] The process of determining the similarity index of intent is as follows: Figure 2 As shown, it includes:

[0072] S103-1: Calculate the cosine distance between any two behavior vectors within the session vector sequence, and use it as the directional difference value.

[0073] The directional difference value measures the degree of directional deviation between any two behavior vectors in a session in a high-dimensional space.

[0074] It should be noted that the calculation method of cosine distance is a well-known technique to those skilled in the art, and will not be elaborated upon in this embodiment. For example, the direction difference value (i.e., cosine distance) = 1 - ;in, This represents the vector of the i-th row; This represents the vector of the j-th row, where i is not equal to j; Represents a behavior vector The modulus length; Represents a behavior vector The length of the module.

[0075] It should be noted that, in practice, a behavior vector cannot be a zero vector, and therefore its magnitude cannot be zero.

[0076] S103-2: Calculate the arithmetic mean of all directional differences within the session vector sequence as the directional difference mean.

[0077] S103-3: Calculate the difference between the positive integer 1 and the mean of the directional difference as an indicator of intent similarity.

[0078] It is important to understand that if the mean directional difference within a certain session vector sequence is larger, it indicates that the overall direction of the behavioral vectors within the session is more dispersed, the user intent within the session is more dispersed, and the intent similarity index is smaller.

[0079] In this embodiment, a two-dimensional feature point is constructed for each session using the intent similarity index as the first dimension and the total session duration as the second dimension. Based on the two-dimensional feature points of all sessions, a clustering algorithm is used to group them. According to the distribution characteristics of the two-dimensional feature points in each group in the first and second dimensions after grouping, a group label is assigned to each group.

[0080] The total duration of a session represents the time a user invests in the session. The longer the total duration, the more cautious the user's decision-making process (e.g., "repeatedly comparing parameters and reading reviews"); the shorter the total duration, the more efficient the decision-making (e.g., "directly ordering frequently purchased items").

[0081] For example, each session corresponds to a two-dimensional coordinate point (x, y), where x represents the intent similarity index and y represents the total session duration. Suppose Session 1 (efficient decision-making): (0.9, 3 minutes), Session 2 (in-depth research): (0.8, 25 minutes), Session 3 (exploration and browsing): (0.3, 10 minutes).

[0082] It should be noted that clustering algorithms are well-known techniques in the field, and will not be elaborated upon in this embodiment. For example, density-based clustering (DBSCAN) is used to calculate the similarity (e.g., Euclidean distance) between two-dimensional feature points, automatically grouping sessions with similar features into the same group. Through iterative calculation, feature points that are close together are divided into the same cluster, while feature points that are far apart are divided into different clusters. For example, all points with high x-values ​​(intent similarity index > 0.7) and low y-values ​​(total duration < 5 minutes) will be clustered into one cluster; all points with high x-values ​​and high y-values ​​(duration > 15 minutes) will be clustered into another cluster; and all points with low x-values ​​(intent similarity index < 0.5) will be clustered into a third cluster.

[0083] It should be noted that the group label includes at least one of the following types: the first type, which corresponds to a high intent similarity index and a short total session duration; the second type, which corresponds to a high intent similarity index and a long total session duration; and the third type, which corresponds to a low intent similarity index.

[0084] It should be noted that the specific types of the first, second, and third categories are determined based on the actual situation, and this embodiment does not impose specific limitations. For example, based on the user behavior patterns in e-commerce scenarios, the first type is efficient decision-making (high intent similarity index and short total session duration): users have clear goals and quickly complete decisions; the second type is in-depth research type (high intent similarity index and long total session duration): users make cautious decisions and invest a lot of time in comparative analysis; the third type is exploratory browsing type (low intent similarity index): users have no clear goals and mainly discover products randomly.

[0085] S104: Determine the core topic vector based on the central tendency of behavior vectors within each session; group sessions with a cosine similarity higher than a preset threshold among the core topic vectors of the same user into the same macro shopping path; based on the preset mapping relationship between group labels and intent states, map the group labels of each session within the same macro shopping path to the corresponding intent states, and generate a sequence of user intent states by arranging them in the order of session occurrence time.

[0086] The core topic vector is a quantitative representation of the central trend of all behavioral vectors within the same session, used to extract the core intent of the session (such as "buy sneakers" or "learn about laptops").

[0087] In this embodiment, the arithmetic mean of all behavior vectors within the session vector sequence is calculated as the core topic vector.

[0088] It should be noted that the method for calculating the average value of vectors is an existing technical means, and will not be described in detail in this embodiment.

[0089] It should be noted that even for the same user, the intent reflected in different sessions can vary.

[0090] It's important to understand that, because the direction of the core topic vector represents the core semantics of the conversation in a high-dimensional space, conversation 1 includes "searching for basketball shoes → browsing product A → viewing product A reviews," with the core topic vector leaning towards "basketball shoe selection." Conversation 2 includes "browsing the mobile phone section → comparing models B and C → consulting customer service for configuration," with the core topic vector leaning towards "mobile phone selection research." To objectively analyze the dynamic evolution of user intent, it's necessary to first identify the common macro-level shopping path from multiple conversation vector sequences.

[0091] A macro shopping path refers to a series of related conversations that a user generates over a period of time around the same macro goal (for example, "selecting athletic shoes → purchasing athletic shoe accessories" can be classified as a "sports equipment purchase" path).

[0092] It should be noted that the specific value of the preset threshold is determined according to the actual situation, and this embodiment does not impose a specific limitation. For example, by combining the historical session vector sequence, it can be found that when the preset threshold is 0.6, if the cosine similarity between two sessions is greater than 0.6, these two sessions clearly point to the same macro goal.

[0093] It is important to understand that the greater the cosine similarity between core topic vectors, the stronger the connection between the core intentions of the two conversations (such as "choosing a mobile phone" and "buying a phone case"); the closer the cosine similarity between core topic vectors is to 0, the weaker the connection between intentions (such as "choosing a mobile phone" and "buying snacks").

[0094] It should be noted that the specific content of the mapping relationship between the preset group labels and the intent state can be determined by combining business rules or historical data training results, and this embodiment does not impose specific limitations.

[0095] It should be noted that the group labels are mapped one-to-one with the intent states.

[0096] For example, the mapping relationship between group labels and intent states is shown in Table 1. Assuming that the group label of a certain session is "efficient decision-making", then based on Table 1, the corresponding intent state is "clear goal - quick decision".

[0097] Table 1

[0098]

[0099] S105: Extract state transition features from the intent state sequence, and generate labels representing the evolution pattern of the user's core intent based on the state transition features, as a user profile.

[0100] The intent state sequence is the user's intent states arranged chronologically in the macro shopping path (e.g., no clear goal → careful price comparison → quick purchase). The state transition feature is the quantification and refinement of the transformation relationship between adjacent states in the intent state sequence.

[0101] The user profile generation process is as follows Figure 3 As shown, it includes:

[0102] S105-1: Extract the transition patterns between all adjacent intent states from the user's intent state sequence.

[0103] The transition pattern of adjacent intention states refers to the transformation relationship between two consecutive states in a sequence, and is the smallest unit that describes the evolution of user intentions.

[0104] For example, the intent state sequence is [ , , , ],in, For "exploration", For "research", For "decision making", the extracted transfer pattern is: → (Exploration → Research) (Research → Decision) → (Decision → Exploration). These patterns directly reflect the user's intention to switch from one state to another.

[0105] S105-2: Count the frequency of occurrence of each transition mode, and determine the transition mode with the highest frequency as the core transition mode; calculate the Shannon entropy of all transition modes as an indicator of the intended evolutionary orderliness.

[0106] Core transfer patterns refer to the most frequent transfer patterns that occur during the evolution of user intent. For example, the transfer patterns that a particular user exhibits most frequently (such as...) → The core transition pattern was identified as the core transition pattern, reflecting the dominant trend in the evolution of user intent (for example, the transition from exploration to in-depth research is the most typical behavioral transition for this user).

[0107] The intent evolution orderliness index is used to measure the degree of disorder or stability of user intent state transitions.

[0108] It should be noted that the Shannon entropy calculation method is an existing technical means, which will not be elaborated in this embodiment. For example, assuming a = 1, 2...n; the intentional evolutionary orderliness index can be expressed by the following formula:

[0109]

[0110] in, Indicators representing the orderliness of intentional evolution; Let a represent the a-th transition pattern; n represents the total number of all transition patterns.

[0111] It's important to understand that the lower the Shannon entropy value, i.e., the smaller the intentional evolutionary orderliness index, the more concentrated the transition patterns (e.g., 80% of the transition patterns are...). → The higher the Shannon entropy value, the more orderly and stable the user intent evolution is; the higher the Shannon entropy value, the greater the orderliness index of intent evolution, indicating that the transfer pattern is more dispersed (such as multiple patterns appearing randomly), and the more chaotic and volatile the user intent evolution is.

[0112] S105-3: Combine the core transfer pattern with the intent evolution orderliness index to serve as a user segmentation feature.

[0113] It is important to understand that the core shift pattern reflects the most common path of user intent change (e.g., "explore → research → decision" or "research → hesitate → give up"); the intent evolution orderliness index reflects the stability of the user intent change path (e.g., whether the decision is always made along a fixed path or changes frequently); the combination of the two fully describes the typicality and stability of user intent evolution, providing a comprehensive basis for subsequent grouping.

[0114] S105-4: Based on the clustering characteristics of all users, a clustering algorithm is used to group users; according to the commonality of decision-making paths represented by the core transfer patterns of each user group after clustering, and the commonality of stability represented by the orderliness index of intent evolution, corresponding user profile labels are assigned to each user group as user profiles.

[0115] It should be noted that the specific method for using clustering algorithms (such as density clustering DBSCAN) to group users can refer to the specific method for initially grouping all user sessions in step 103, which will not be repeated in this embodiment. For example, a two-dimensional feature point is constructed for each user with the core transfer mode as the first dimension and the intent evolution orderliness index as the second dimension; based on the two-dimensional feature points of all users, a clustering algorithm is used for grouping; according to the distribution characteristics of the two-dimensional feature points in each user group in the first and second dimensions after grouping, a user profile label is assigned to each user group.

[0116] For example, if user A's core transition pattern is "exploration → research → decision-making" with a Shannon entropy of 0.3 (low entropy, stable), and user B has the same core transition pattern with a Shannon entropy of 0.4 (low entropy), then users A and B may be grouped into the same user group. If user C's core transition pattern is "research → hesitation → exploration" with a Shannon entropy of 1.8 (high entropy, unstable), then user C will be grouped into another group. The grouping results ensure that the intention evolution patterns of users within the same user group are highly similar, while the differences between different user groups are significant.

[0117] It should be noted that the rules for assigning user profile tags include: user groups with progressive decision-making paths and low intention evolution orderliness indicators are assigned the first type of tag; user groups with cyclical decision-making paths and high intention evolution orderliness indicators are assigned the second type of tag; and user groups with progressive decision-making paths but high intention evolution orderliness indicators are assigned the third type of tag.

[0118] The decision-making path is progressive, meaning that the intentional state advances step by step in a logical order, showing a trend of "moving closer to the target state from the initial state" (such as "exploration → research → decision" or "no clear goal → clear goal → place an order"). This reflects that the user's decision-making process has a clear direction and gradually moves towards the final result.

[0119] The decision-making path is cyclical, meaning that the intention state repeatedly switches back and forth, showing a trend of "switching repeatedly between multiple states" (such as "research → hesitation → exploration → research" or "decision → abandonment → re-exploration"). This reflects that the user's decision-making process lacks a stable direction and is prone to repetition or interruption.

[0120] The first type of label is used to reflect that users are "goal-oriented and have stable processes". They usually have clear decision-making steps when shopping, are not easily disturbed, and their behavior is highly predictable.

[0121] The second type of label is used to reflect that the user belongs to the "hesitant and changeable type", whose shopping goals are vague, easily influenced by external factors, and whose decision-making cycle is long and unstable.

[0122] The third type of tag is used to reflect that users belong to the "goal-oriented but process-fluctuating type". They have a clear final goal when shopping, but the process of achieving the goal lacks a fixed pattern and may change the steps due to factors such as scenario and time.

[0123] It should be noted that the specific meaning of user profile tags can be determined based on business industry knowledge. This embodiment does not impose specific limitations. For example, user group 1 has a core transfer pattern of "exploration → research" and a Shannon entropy value of 0.2, which corresponds to the user profile tag "stable and progressive type," representing users who are "patient and make decisions according to procedures." User group 2 has a core transfer pattern of "research → hesitation → exploration → research" and a Shannon entropy value of 1.5, which corresponds to the user profile tag "easily giving up type," representing users whose "decision-making is easily disturbed and who often give up halfway." User group 3 has a core transfer pattern of "exploration → hesitation → decision" and a Shannon entropy value of 1.5, which corresponds to the user profile tag "flexible shopper," representing users whose "ultimate direction is progressive, but the specific path is not fixed."

[0124] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0125] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for generating a user portrait of an e-commerce operation platform based on big data, characterized in that, The method comprises: Collecting interaction behavior data of users on an e-commerce platform and converting the data into behavior vectors to generate a behavior vector sequence of the user in chronological order; According to the similarity and time difference between adjacent behavior vectors in the behavior vector sequence, a comprehensive correlation index is determined; adjacent behavior vectors whose continuous multiple comprehensive correlation indexes meet a preset session condition are classified into the same session to form a session vector sequence; For each session vector sequence, an intent similarity index is determined according to the direction difference between the behavior vectors in the session; and all sessions of the user are preliminarily grouped by combining the total duration of the session, and a group label representing the micro-intent of the user is assigned to the session; According to the concentration trend of the behavior vectors in each session, a core theme vector is determined; sessions whose cosine similarity between the core theme vectors in the same user is higher than a preset threshold are merged into the same macro shopping path; according to a mapping relationship between the preset group label and the intent state, the group label of each session in the same macro shopping path is mapped to the corresponding intent state, and an intent state sequence of the user is generated in chronological order of the session occurrence time; State transition features are extracted from the intent state sequence, and a label representing the core intent evolution rule of the user is generated according to the state transition features as the user portrait; The state transition features are extracted from the intent state sequence, and the label representing the core intent evolution rule of the user is generated according to the state transition features as the user portrait, comprising: From the intent state sequence of the user, the transition mode between all adjacent intent states is extracted; The occurrence frequency of each transition mode is counted, and the transition mode with the highest occurrence frequency is determined as the core transition mode; the Shannon entropy of all transition modes is calculated as an intent evolution orderliness index; The core transition mode and the intent evolution orderliness index are combined to serve as the grouping feature of the user; Based on the grouping features of all users, a clustering algorithm is used to group the users; according to the decision path commonality represented by the core transition mode of each user group after grouping and the stability commonality represented by the intent evolution orderliness index, a corresponding user portrait label is assigned to each user group as the user portrait of the user; The assignment rule of the user portrait label comprises: a user group with a progressive decision path and a low intent evolution orderliness index is assigned a first type of label; a user group with a cyclic decision path and a high intent evolution orderliness index is assigned a second type of label; a user group with a progressive decision path but a high intent evolution orderliness index is assigned a third type of label. 2.The method of claim 1, wherein, The comprehensive correlation index determination process comprises: For each behavior vector sequence, the absolute difference between the time stamps of adjacent behavior vectors is calculated as the time difference value; The cosine similarity of adjacent behavior vectors is calculated as the semantic similarity; Based on a preset weight coefficient, the normalized semantic similarity and the normalized time difference value are weighted to obtain the comprehensive correlation index, wherein the weight of the semantic similarity is greater than the weight of the time difference value. 3.The method of claim 1, wherein, The preset session condition comprises a high correlation threshold and a low correlation threshold, and the high correlation threshold is greater than the low correlation threshold. The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method further comprises the following steps:

4. The e-commerce operation platform user portrait generation method based on big data according to claim 3, characterized in that, The method further comprises the following steps: The method further comprises the following steps: 5.The method of claim 1, wherein, The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: 6.The method of claim 5, wherein, The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps:

7. The user portrait generation method of the e-commerce operation platform based on big data according to claim 6, characterized in that, The method further comprises the following steps: The method further comprises the following steps: 8.The method of claim 1, wherein, The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further

Images