Clothing trend crawler improvement system based on near-end strategy optimization model

Abstract

The invention relates to the technical field of crawler algorithms, particularly provides a clothing trend crawler improvement system based on a near-end strategy optimization model, and aims to solve the problem of poor data capture effect of a crawler algorithm in the prior art. The system comprises a keyword receiving unit for receiving clothes keyword information; the reinforcement learning unit is used for adjusting a crawler strategy through a near-end strategy optimization model; a reward mechanism is introduced to reward the clothing data captured by the crawler, and the crawler is driven to capture target clothing data which corresponds to the clothing keyword and has a fashion trend; and the data warehouse construction unit is used for classifying and storing the target clothing data. According to the method, the crawler strategy is optimized and adjusted through the near-end planning optimization model, so that the crawler can dynamically adapt to the change of the Internet environment, data in the fashion clothing industry can be flexibly and efficiently captured and processed, and reliable data support is provided for clothing trend analysis and prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of crawler algorithms, and in particular relates to an improved clothing trend crawler system based on a proximal strategy optimization model. Background Art

[0002] In the fashion apparel industry, timely acquisition and analysis of the latest trend data is crucial for design and market decision-making. Existing web crawler technologies usually rely only on predefined rules and keywords, but these rules and keywords may quickly become invalid when the network environment and information change rapidly. In addition, the execution efficiency and data processing speed of existing web crawler technologies are slow, making it difficult to acquire and process large amounts of data in real time or near real time. Summary of the invention

[0003] The present invention provides an improved clothing trend crawler system based on a proximal strategy optimization model, aiming to solve the problem of poor data capture effect of crawler algorithms in the prior art.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] The present invention provides a clothing trend crawler improvement system based on a proximal strategy optimization model, comprising:

[0006] A keyword receiving unit, receiving clothing keyword information;

[0007] A reinforcement learning unit adjusts the crawler strategy through a proximal strategy optimization model; and introduces a reward mechanism, rewards the clothing data captured by the crawler according to the reward mechanism, and drives the crawler to capture target clothing data with fashion trends corresponding to the clothing keywords;

[0008] The data warehouse construction unit classifies and stores the target clothing data captured by the crawler.

[0009] Further solution: The proximal strategy optimization model adjusts the crawler strategy including the following steps:

[0010] S100, define the crawler's state, action and reward mechanism respectively;

[0011] S200, the crawler captures clothing data according to the clothing keyword information and the strategy, and calculates a reward value for the captured clothing data;

[0012] S300, the proximal strategy optimization model adjusts the crawler strategy through the strategy network according to the reward value.

[0013] Based on the above scheme, the proximal strategy optimization model is used to enable the crawler to self-adjust and optimize in a constantly changing network environment. In addition, a reward mechanism is designed to reward the crawler according to the clothing it captures, thereby guiding the crawler to adjust its strategy when fashion trends change.

[0014] In a further solution, in the step S200, the reward mechanism uses a combined reward function to calculate the reward value of the clothing data captured by the crawler;

[0015] Combined reward function = data relevance reward function + data timeliness reward function + data novelty reward function.

[0016] Based on the above scheme, the combined reward function enables the crawler to not only prioritize the clothing data related to the clothing keyword information, but also evaluate the captured clothing according to the timeliness and novelty of the data, effectively guiding the crawler to capture more valuable clothing data.

[0017] Further solution: The formula of the data relevance reward function is:

[0018] R r =Similarity(P, T);

[0019] Wherein, Similarity() represents cosine similarity; P represents the content of the page currently crawled by the crawler; T represents clothing keyword information;

[0020] If R r >M, a positive reward is given to the clothing data captured by the crawler; otherwise, a zero or negative reward is given to the clothing data captured by the crawler; wherein M represents a set threshold;

[0021] The formula of the data timeliness reward function is:

[0022]

[0023] Wherein, max() represents the maximum value; T current Represents the current time; the T page Represents the publishing time of the page content; the T max Represents the maximum time difference set;

[0024] If R t ≤T max , a positive reward is given to the clothing data captured by the crawler; otherwise, a negative reward is given to the clothing data captured by the crawler;

[0025] The formula of the data novelty reward function is:

[0026]

[0027] if P∈D,R n = -λ;

[0028] Wherein, P represents the content of the current crawler crawling page; D represents the existing clothing data set; Novelty Score (P, D) calculates the difference between P and D; λ represents the penalty factor for duplicate clothing data;

[0029] If the P has not been captured and is significantly different from the D, a positive reward is given to the clothing data captured by the crawler; if the P already exists in the D, a negative reward is given to the clothing data captured by the crawler;

[0030] The formula of the combined reward function is:

[0031] R = αR r +βR t +γR n ;

[0032] Wherein, α represents the data relevance reward R r The weight parameter of β represents the data timeliness reward R t The weight parameter is γ, and the γ represents the data novelty reward R. n The weight parameter of .

[0033] Further solution: In step S300, the step of adjusting the crawler strategy through the strategy network includes:

[0034] S301, calculate the probability distribution πθ of each action selected by the crawler according to the old strategy old (a t |s t );

[0035] S302: Train the current strategy based on the reward value of the clothing data captured by the crawler according to the old strategy, and calculate the probability distribution πθ of each action selected by the crawler according to the current strategy new (a t |s t );

[0036] S303: Calculate the probability ratio between the action probability distribution of the current strategy and the action probability distribution of the old strategy Adjusting the crawler's strategy according to the probability ratio;

[0037] Among them, the a t Represents an action; t Represents the state; the πθ old represents the old strategy; the πθnew Represents the current policy.

[0038] Based on the above scheme, by comparing the mouth movement probability distribution of the old strategy and the action probability distribution of the current strategy, the degree of improvement of the current strategy compared with the old strategy can be more effectively evaluated, thereby accelerating the crawler learning and data grabbing process and avoiding invalid exploration.

[0039] In a further solution, the proximal policy optimization model uses a shear objective function to limit the amplitude of policy adjustment, and the shear objective function formula is:

[0040]

[0041] Wherein, θ represents the parameters of the policy network; At represents the event function; clip() represents the clipping objective function used to limit the range of the probability ratio; and ε represents a small constant between 0.1 and 0.2.

[0042] Based on the above scheme, the shearing objective function is used to avoid large-scale update and adjustment of the crawler strategy, thereby ensuring the stability of the crawler strategy during the update process.

[0043] A further solution: in step S100, a graph neural network is added when defining the state and action of the crawler;

[0044] The graph neural network aggregates the features of nodes and neighbor nodes through a message passing mechanism, updates the status of nodes, obtains the source of clothing fashion trends, and drives the crawler to capture the clothing data from the source of clothing fashion trends.

[0045] Based on the above scheme, the graph neural network can build a relationship network on the Internet, helping the crawler strategy to more intelligently discover information sources with fashion trends, optimize data acquisition paths, and improve the efficiency and accuracy of clothing data acquisition.

[0046] A further solution: the reinforcement learning unit also includes a multimodal model for extracting and analyzing features of the target clothing data captured by the crawler.

[0047] A further solution: the multimodal model includes a text encoder, an image encoder and a contrastive learning mechanism;

[0048] The text encoder encodes the text description in the target clothing data captured by the crawler to obtain a text feature vector;

[0049] The image encoder encodes the image content in the target clothing data captured by the crawler to obtain an image feature vector;

[0050] The text feature vector is aligned with the image feature vector through a contrast learning mechanism, and the correlation between the text feature vector and the image feature vector is analyzed.

[0051] Based on the above scheme, the multimodal model can improve the comprehensiveness and accuracy of the crawler in acquiring fashion trend information and enhance the ability to understand and process unstructured clothing data by extracting and analyzing the correlation between text and images in the target clothing data captured by the crawler.

[0052] A further solution: In the data warehouse construction unit, a star schema or a snowflake schema is used to classify and store the target clothing data captured by the crawler.

[0053] Based on the above solution, the target clothing data captured by building a data warehouse can improve the subsequent user query speed.

[0054] The beneficial effects of the present invention are:

[0055] 1. The present invention optimizes and adjusts the crawler strategy through a proximal planning optimization model, so that the crawler can dynamically adapt to changes in the Internet environment, can flexibly and efficiently capture and process data on the fashion clothing industry, and provide reliable data support for clothing trend analysis and prediction.

[0056] 2. The present invention introduces a reward mechanism, which can evaluate the clothing data captured by the crawler and guide the crawler to optimize the crawling behavior in a complex and changeable network environment. The reward mechanism ensures the quality and value of the captured clothing data, so that the crawler can capture clothing data more intelligently and accurately.

[0057] 3. The data warehouse in the present invention can realize real-time data storage, updating and analysis, and provide real-time insights into fashion trends. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying creative work.

[0059] Figure 1 It is a structural schematic diagram of a clothing trend crawler improvement system based on a proximal strategy optimization model of the present invention;

[0060] Figure 2 It is a flowchart of the steps of adjusting the crawler strategy by the proximal strategy optimization model in the present invention;

[0061] Figure 3 It is a flow chart of the steps of adjusting the crawler strategy of the strategy network in the present invention. DETAILED DESCRIPTION

[0062] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work belong to the protection scope of the present invention.

[0063] like Figure 1 As shown, this embodiment provides an improved clothing trend crawler system based on a proximal strategy optimization model, including:

[0064] A keyword receiving unit, receiving clothing keyword information;

[0065] A reinforcement learning unit adjusts the crawler strategy through a proximal policy optimization model (PPO); and introduces a reward mechanism, rewards the clothing data captured by the crawler according to the reward mechanism, and drives the crawler to capture target clothing data with fashion trends corresponding to the clothing keywords;

[0066] The data warehouse construction unit classifies and stores the target clothing data captured by the crawler.

[0067] A more specific example of the keyword receiving unit is that the keyword information received by the keyword receiving unit is natural language, for example, ancient style dress, red coat, etc.

[0068] A more specific example of the reinforcement learning unit is: the reinforcement learning unit includes a proximal optimization model and a multimodal model. The proximal strategy optimization model adjusts the crawler strategy; and introduces a reward mechanism, according to which the clothing data captured by the crawler is rewarded, driving the crawler to capture the target clothing data with fashion trends corresponding to the clothing keywords. The multimodal model is used to extract and analyze the features of the target clothing data captured by the crawler.

[0069] like Figure 2 As shown, the following are steps for adjusting the crawler strategy by the proximal strategy optimization model, including:

[0070] S100, respectively defining the state of the crawler, the web page, network environment or captured data set state where the crawler is currently located; for example, a specific web page content or a data set that the crawler has acquired;

[0071] Define the crawler's actions, which are the next steps the crawler can take, such as clicking a web link, parsing web content, or jumping to another website; and,

[0072] Define the crawler's reward mechanism. Each time the crawler performs an action, it will give a reward or negative reward based on the quality of the captured clothing data (such as timeliness, relevance, or novelty). The crawler accumulates these rewards and continuously optimizes future strategies;

[0073] S200, the crawler captures clothing data according to the clothing keyword information and the strategy, and calculates a reward value for the captured clothing data;

[0074] S300, the proximal strategy optimization model adjusts the crawler strategy through the strategy network according to the reward value.

[0075] Specifically, in step S100, a graph neural network is added when defining the state and action of the crawler;

[0076] The graph neural network aggregates the features of nodes and neighbor nodes through a message passing mechanism, updates the status of nodes, obtains the source of clothing fashion trends, and drives the crawler to capture the clothing data from the source of clothing fashion trends.

[0077] It should be noted that the graph neural network (GNN) is a deep learning model specifically used to process graph structure data. Through the message passing mechanism, each node obtains information from neighboring nodes and updates its own features, so that it can identify complex structural relationships.

[0078] The crawler can view the entire Internet as a large graph, where nodes are information sources such as web pages, social media users, bloggers, and edges are the connections between these nodes (such as social media follow-up relationships, hyperlinks between articles, etc.). The graph neural network identifies potential sources of fashion trend information by aggregating the features of nodes and their neighbor nodes layer by layer. For example, it can analyze the influence of fashion bloggers, or identify new fashion trends through users' social relationships. Fashion trend information sources refer to data sources that can provide key fashion trend signals, such as fashion bloggers' articles, popular topics on social media, and new product releases from brands. For example, when the graph neural network recognizes that a photo of a new clothing item posted by a fashion blogger has caused a large number of user interactions, then the blogger's information may become an important source of fashion trends.

[0079] The process of the message transmission mechanism includes:

[0080] Step 1: Initialize node features; each node initially has its own features (such as content features of a web page, attributes of a social media user, etc.), and this feature vector is used to represent the attributes of the node itself.

[0081] Step 2: Message passing: Each node will pass its own features to its neighboring nodes. The information passed can be the current state of the node or some weighted feature information.

[0082] Step 3: Neighbor node aggregation: After each node receives feature information from its neighbor nodes, it will integrate the feature information transmitted by the neighbor nodes according to the aggregation function (such as sum, average or maximum value, etc.), and update the state in combination with its own features, so as to gradually build its global feature representation.

[0083] In addition, in the step S200, the reward mechanism uses a combined reward function to calculate the reward value of the clothing data captured by the crawler;

[0084] Combined reward function = data relevance reward function + data timeliness reward function + data novelty reward function.

[0085] The data relevance reward is used to evaluate the matching degree between the captured clothing data and the current clothing keyword information. The formula of the data relevance reward function is:

[0086] R r =Similarity(P, T);

[0087] Wherein, Similarity() represents cosine similarity; P represents the content of the page currently crawled by the crawler; T represents clothing keyword information;

[0088] If R r >M, a positive reward is given to the clothing data captured by the crawler; otherwise, a zero or negative reward is given to the clothing data captured by the crawler; wherein M represents a set threshold;

[0089] The data timeliness reward is used to evaluate the timeliness of the captured data. The newer the captured clothing data, the higher the reward. The formula of the data timeliness reward function is:

[0090]

[0091] Wherein, max() represents the maximum value; T current Represents the current time; the T page Represents the publishing time of the page content; the T max Represents the maximum time difference set;

[0092] If R t ≤T max, a positive reward is given to the clothing data captured by the crawler; otherwise, a negative reward is given to the clothing data captured by the crawler;

[0093] The novelty reward is used to evaluate the incremental value of the captured clothing data, that is, whether the captured clothing data provides new data that has not been captured before. The formula of the data novelty reward function is:

[0094]

[0095] if P∈D,R n = -λ;

[0096] Wherein, P represents the content of the current crawler crawling page; D represents the existing clothing data set; Novelty Score (P, D) calculates the difference between P and D; λ represents the penalty factor for duplicate clothing data;

[0097] If the P has not been captured and is significantly different from the D, a positive reward is given to the clothing data captured by the crawler; if the P already exists in the D, a negative reward is given to the clothing data captured by the crawler;

[0098] The formula of the combined reward function is:

[0099] R = αR r +βR t +γR n ;

[0100] Wherein, α represents the data relevance reward R r The weight parameter of β represents the data timeliness reward R t The weight parameter is γ, and the γ represents the data novelty reward R. n The weight parameter of .

[0101] In addition, if Figure 3 As shown, in step S300, the step of adjusting the crawler strategy through the strategy network includes:

[0102] S301, calculate the probability distribution πθ of each action selected by the crawler according to the old strategy old (a t |s t );

[0103] S302: Train the current strategy based on the reward value of the clothing data captured by the crawler according to the old strategy, and calculate the probability distribution πθ of each action selected by the crawler according to the current strategy new (a t |s t );

[0104] S303: Calculate the probability ratio between the action probability distribution of the current strategy and the action probability distribution of the old strategy Adjusting the crawler's strategy according to the probability ratio;

[0105] Among them, the a t Represents an action; t Represents the state; the πθ o1d represents the old strategy; the πθ new Represents the current policy.

[0106] Based on the above solution, the proximal strategy optimization model uses a shear objective function to limit the amplitude of strategy adjustment. The shear objective function formula is:

[0107]

[0108] Wherein, θ represents the parameters of the policy network; At represents the event function; clip() represents the clipping objective function used to limit the range of the probability ratio; and ε represents a small constant between 0.1 and 0.2.

[0109] The multimodal model includes a text encoder, an image encoder, and a contrastive learning mechanism. The text encoder processes text descriptions in fashion blogs and social media posts; the image encoder processes image content, such as photos of clothing or models wearing them; and the contrastive learning mechanism allows the crawler to understand how text descriptions and images with the same semantics are related. For example, the description of "red dress" will match pictures containing red dresses.

[0110] The following are steps for the multimodal model to extract and analyze features of the target clothing data captured by the crawler, including:

[0111] The text encoder encodes the text description in the target clothing data captured by the crawler to obtain a text feature vector;

[0112] The image encoder encodes the image content in the target clothing data captured by the crawler to obtain an image feature vector;

[0113] The text feature vector is aligned with the image feature vector through a contrast learning mechanism, and the correlation between the text feature vector and the image feature vector is analyzed.

[0114] A more specific example of the data warehouse construction unit is: in the data warehouse construction unit, the target clothing data captured by the crawler is classified (can be classified by time, brand or type) and stored using a star schema or a snowflake schema.

[0115] Case 1: The core of the star schema is a fact table, which is organized around multiple dimension tables, each of which is directly connected to the fact table, shaped like a star. For example, clothing sales data can be used as a fact table, while brand, time, type, etc. can be used as dimension tables.

[0116] Case 2: The snowflake schema is an extension of the star schema. The dimension table in the snowflake schema is further normalized into multiple sub-tables, which resemble snowflakes. Compared with the star schema, the snowflake schema reduces data redundancy by decomposing the dimension table, but the query complexity will increase.

[0117] It should be noted that: before the target clothing data captured by the crawler is classified and stored, the target data is first processed; the data processing includes cleaning, deduplication and format conversion, etc., to ensure the quality and consistency of the target clothing data.

[0118] Based on the above scheme, an improved clothing trend crawler system based on the proximal strategy optimization model also includes a fashion trend prediction unit; a time series prediction model is trained using a deep learning model to analyze and predict the evolution of fashion trends and consumer behavior. The causal reasoning method is used to analyze the causes and influencing factors of fashion trend changes, and provide decision support driven by the captured target clothing data.

[0119] The present invention is not limited to the above-mentioned optional implementation modes. Anyone can derive other various forms of products under the inspiration of the present invention. However, no matter what changes are made in the shape or structure, all technical solutions that fall within the scope defined by the claims of the present invention fall within the protection scope of the present invention.

Claims

1. An improved clothing trend crawler system based on a proximal strategy optimization model, characterized in that: include: A keyword receiving unit, receiving clothing keyword information; Reinforcement learning unit, which adjusts crawler strategy through proximal strategy optimization model; A reward mechanism is introduced, and the clothing data captured by the crawler is rewarded according to the reward mechanism, so as to drive the crawler to capture target clothing data with fashion trends corresponding to the clothing keywords; The data warehouse construction unit classifies and stores the target clothing data captured by the crawler.

2. According to claim 1, a clothing trend crawler improvement system based on a proximal strategy optimization model is characterized in that: The proximal strategy optimization model adjusts the crawler strategy including the following steps: S100, define the crawler's state, action and reward mechanism respectively; S200, the crawler captures clothing data according to the clothing keyword information and the strategy, and calculates a reward value for the captured clothing data; S300, the proximal strategy optimization model adjusts the crawler strategy through the strategy network according to the reward value.

3. The clothing trend crawler improvement system based on the proximal strategy optimization model according to claim 2 is characterized in that: In the step S200, the reward mechanism uses a combined reward function to calculate the reward value of the clothing data captured by the crawler; Combined reward function = data relevance reward function + data timeliness reward function + data novelty reward function.

4. The clothing trend crawler improvement system based on the proximal strategy optimization model according to claim 3 is characterized in that: The formula of the data relevance reward function is: R r =Similarity(P,T); Wherein, Similarity() represents cosine similarity; P represents the content of the page currently crawled by the crawler; T represents clothing keyword information; If R r >M, a positive reward is given to the clothing data captured by the crawler; otherwise, a zero or negative reward is given to the clothing data captured by the crawler; wherein M represents a set threshold; The formula of the data timeliness reward function is: Wherein, max() represents the maximum value; T current Represents the current time; the T page Represents the publishing time of the page content; the T max Represents the maximum time difference set; If R t ≤T max , a positive reward is given to the clothing data captured by the crawler; otherwise, a negative reward is given to the clothing data captured by the crawler; The formula of the data novelty reward function is: if P ∈ D,R n =-λ; Wherein, P represents the content of the current crawler crawling page; D represents the existing clothing data set; Novelty Score (P, D) calculates the difference between P and D; λ represents the penalty factor for duplicate clothing data; If the P has not been captured and is significantly different from the D, a positive reward is given to the clothing data captured by the crawler; if the P already exists in the D, a negative reward is given to the clothing data captured by the crawler; The formula of the combined reward function is: R=α R r +β R t +γ R n ; Among them, α represents the weight parameter of the data relevance reward Rr; β represents the data timeliness reward R t The weight parameter is γ, and the γ represents the data novelty reward R. n The weight parameter of .

5. The clothing trend crawler improvement system based on the proximal strategy optimization model according to claim 2 is characterized in that: In step S300, the step of adjusting the crawler strategy through the strategy network includes: S301, calculate the probability distribution πθ of each action selected by the crawler according to the old strategy old (a t |s t ); S302: Train the current strategy based on the reward value of the clothing data captured by the crawler according to the old strategy, and calculate the probability distribution πθ of each action selected by the crawler according to the current strategy new (a t |s t ); S303: Calculate the probability ratio between the action probability distribution of the current strategy and the action probability distribution of the old strategy Adjusting the crawler's strategy according to the probability ratio; Among them, the a t Represents an action; t Represents the state; the πθ old represents the old strategy; the πθ new Represents the current policy.

6. The clothing trend crawler improvement system based on the proximal strategy optimization model according to claim 5 is characterized in that: The proximal strategy optimization model uses a shear objective function to limit the magnitude of strategy adjustment. The shear objective function formula is: Wherein, θ represents the parameters of the policy network; At represents the event function; clip() represents the clipping objective function used to limit the range of the probability ratio; and ε represents a small constant between 0.1 and 0.

2.

7. The clothing trend crawler improvement system based on the proximal strategy optimization model according to claim 2 is characterized in that: In the step S100, a graph neural network is added when defining the state and action of the crawler; The graph neural network aggregates the features of nodes and neighbor nodes through a message passing mechanism, updates the status of nodes, obtains the source of clothing fashion trends, and drives the crawler to capture the clothing data from the source of clothing fashion trends.

8. The clothing trend crawler improvement system based on the proximal strategy optimization model according to claim 1 is characterized in that: The reinforcement learning unit also includes a multimodal model for extracting and analyzing features of the target clothing data captured by the crawler.

9. The clothing trend crawler improvement system based on the proximal strategy optimization model according to claim 8 is characterized in that: The multimodal model includes a text encoder, an image encoder, and a contrastive learning mechanism; The text encoder encodes the text description in the target clothing data captured by the crawler to obtain a text feature vector; The image encoder encodes the image content in the target clothing data captured by the crawler to obtain an image feature vector; The text feature vector is aligned with the image feature vector through a contrast learning mechanism, and the correlation between the text feature vector and the image feature vector is analyzed.

10. The clothing trend crawler improvement system based on the proximal strategy optimization model according to claim 1 is characterized in that: In the data warehouse construction unit, a star schema or a snowflake schema is used to classify and store the target clothing data captured by the crawler.

Images