Intelligent product recommendation method based on large model

Abstract

The invention relates to the technical field of intelligent product recommendation, in particular to an intelligent product recommendation method based on a large model, which comprises the following steps of: firstly, receiving a user dominant demand text and real-time behavior data, analyzing the behavior data through a dynamic window self-adaptive mechanism, identifying a time sequence correction logic chain for adding, comparing clicking and canceling adding, and finally recommending the user dominant demand text to the real-time behavior data. Combining attribute-level focus capture and visual thermodynamic diagram feedback positioning core attributes, quantifying a behavior weight offset effect, generating a composite label fusing explicit and implicit demands, adopting time-space density clustering statistics attributes to compare effective click frequencies, combining an attribute dependency relationship network to rearrange attribute priorities, and obtaining a real-time feature; according to the method, dynamic attribute weight distribution is formed by differentially distributing weights through a weight rebalance tree model, finally, three types of information are integrated through a large model, the weights are calibrated through antagonism verification, a product recommendation list conforming to real preferences of users is generated, and recommendation accuracy and individuation degree are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent product recommendation technology, and more specifically, to an intelligent product recommendation method based on a large model. Background Technology

[0002] Intelligent product recommendation is an important technology, specifically applied to the accurate capture and matching of users' real needs in product recommendation scenarios. Its core is to improve the fit of large-scale model recommendations by analyzing the temporal correction logic of user behavior and dynamically adapting demand tags and attribute priorities. This aligns with the core requirements of intelligent recommendation for accuracy and personalization. Users generate initial and subsequent corrective actions on the product recommendation interface. These actions form a corrective logic chain reflecting changes in real needs over time. Because this temporal correction intent is difficult to accurately identify and quantify, implicit user demand tags cannot be effectively integrated with behavioral correction information, and product attribute priorities cannot be dynamically adjusted according to user focus. Consequently, large-scale models struggle to align with users' real preferences when integrating needs, affecting the accuracy of recommendation results. To address this technical problem, we provide an intelligent product recommendation method based on large-scale models. Summary of the Invention

[0003] The purpose of this invention is to provide a smart product recommendation method based on a large model to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, one of the objectives of this invention is to provide an intelligent product recommendation method based on a large model, comprising the following steps:

[0005] S1. Receive explicit demand text and real-time user behavior data input by the user. The explicit demand text includes product attribute requirements directly proposed by the user. The real-time user behavior data includes the user's operation behavior and corresponding timestamp on the product recommendation interface. S2. Analyze real-time user behavior data based on a preset time window, identify and quantify the time-series correction features of user behavior, whereby the time-series correction features refer to the correction logic chain formed by user operations over time, including: Starting with the user's initial addition behavior, a dynamic time window is set to monitor subsequent behaviors. When it is detected that the user performs multiple comparison clicks on the specific attribute parameter module of the second product within the first time window after adding the first product, and the dwell time of each comparison click exceeds a preset threshold, and the user triggers the behavior of canceling the addition of the first product within the second time window, the behavior sequence is extracted as a correction logic chain, and the behavior weight offset effect in the correction logic chain is quantified. This is achieved through a behavior correction weight offset threshold mechanism. That is, when the cancellation of the addition behavior occurs, the weight of the initial addition behavior in the demand tag is automatically reduced according to the preset offset ratio threshold, and the corrected user implicit demand tag is generated. S3. Based on the user behavior time sequence correction features, rearrange the priority of product attribute dimensions, count the comparison click frequency for each product attribute in the behavior sequence, and calculate its proportion of the total comparison click frequency of all attributes. When the comparison click frequency proportion of any attribute exceeds the preset priority rearrangement threshold, trigger the attribute priority dynamic adjustment mechanism, raise the priority of the attribute from the original weight position to the highest weight position, and at the same time reduce the weight of other attributes proportionally to form a rearranged dynamic attribute weight distribution. S4. Using a large model, integrate the explicit demand text, the corrected implicit user demand tags, and the rearranged dynamic attribute weight distribution to generate the final product recommendation list.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes a dynamic window adaptive mechanism to adjust the time window span in real time and encode the intent of add and compare click behaviors in a multimodal manner. It identifies the user behavior temporal correction logic chain, solving the problem of difficulty in capturing temporal correction intent. Combined with attribute-level focus capture technology and visual heatmap feedback, it locates high-confidence attention attributes, providing accurate basis for demand analysis. It quantifies the behavior weight offsetting effect through a behavior chain integrity model, and generates composite tags that integrate explicit and implicit needs with a tag conflict resolution strategy to avoid demand bias. It uses spatiotemporal density clustering to count effective clicks, and combines attribute dependency relationship networks and weight rebalancing trees to dynamically rearrange attribute priorities to ensure that weights adapt to user attention. Finally, the weights are calibrated through adversarial verification of a large model. The generated recommendation list can deeply match the user's true preferences, significantly improving the accuracy and personalization of intelligent recommendations. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating the overall workflow of the present invention. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] Please see Figure 1 As shown, this embodiment provides an intelligent product recommendation method based on a large model, including the following steps: S1. Receive explicit demand text and real-time user behavior data input by the user. The explicit demand text includes the product attribute requirements directly put forward by the user. The real-time user behavior data includes the operation behavior performed by the user on the product recommendation interface and the corresponding timestamp. S2. Analyze real-time user behavior data based on a preset time window, identify and quantify the time-series correction features of user behavior, whereby the time-series correction features refer to the correction logic chain formed by user operations over time, including: Starting with the user's initial addition behavior, a dynamic time window is set to monitor subsequent behaviors. When it is detected that the user performs multiple comparison clicks on the specific attribute parameter module of the second product within the first time window after adding the first product, and the dwell time of each comparison click exceeds a preset threshold, and the user triggers the behavior of canceling the addition of the first product within the second time window, the behavior sequence is extracted as a correction logic chain, and the behavior weight offsetting effect in the correction logic chain is quantified. This is achieved through a behavior correction weight offsetting threshold mechanism. That is, when the cancellation behavior occurs, the weight of the initial addition behavior in the demand tag is automatically reduced according to the preset offsetting ratio threshold, and the corrected user implicit demand tag is generated. S3. Based on the time-series correction features of user behavior, the priority of product attribute dimensions is rearranged. The frequency of comparison clicks for each product attribute in the behavior sequence is counted, and its proportion in the total frequency of comparison clicks of all attributes is calculated. When the proportion of comparison clicks of any attribute exceeds the preset priority rearrangement threshold, the attribute priority dynamic adjustment mechanism is triggered, which raises the priority of the attribute from the original weight position to the highest weight position, while the weights of other attributes are reduced proportionally, forming a dynamic attribute weight distribution after rearrangement. S4. Use the large model to integrate explicit demand text, corrected implicit user demand tags, and rearranged dynamic attribute weight distribution to generate the final product recommendation list.

[0010] The analysis of real-time user behavior data based on a preset time window further includes the introduction of a dynamic window adaptive mechanism. This mechanism automatically adjusts the span between the first and second time windows by monitoring the density and interval distribution of user operation behaviors in real time. Specifically, when the time interval of continuous user operation behaviors is detected to be compressing, the time window span is dynamically shortened to capture high-frequency correction intentions; conversely, the window span is extended to associate long-cycle behavior chains. At the same time, the behavior pattern recognition module performs multimodal intent encoding on add, cancel add, and comparison click behaviors, transforming the timestamp sequence of operation behaviors into a weighted behavior vector, which is then input into a temporal convolutional network to identify the start and stop boundaries of the correction logic chain.

[0011] When it is detected that a user performs multiple comparative clicks on a specific attribute parameter module of the second product within the first time window after adding the first product, attribute-level focus capture technology is further deployed. By recording the depth of the user's interaction trajectory when clicking on the parameter module of the second product, the specific dimensions of the specific attribute parameter module can be identified. When a user repeatedly triggers the expand, collapse, or parameter comparison controls of the same attribute module on the same product page, and the coverage density of the cursor hover trajectory in the module area exceeds the preset clustering threshold, it is determined that the core comparison dimension in the user's correction logic chain is established, and a difference mapping relationship is established between the core comparison dimension and the corresponding attribute of the first product, which serves as the conflict attribute identifier of the correction logic chain.

[0012] The attribute-level focus capture technology further integrates with the visual heatmap feedback mechanism. By capturing the user's eye-tracking data and page scrolling speed when browsing the second product parameter module, it dynamically corrects the recognition results of specific attribute parameter modules. When it is detected that the duration of the user's eye focus on a certain attribute description area is positively correlated with the comparison click behavior, and the page scrolling shows an alternating pattern between adjacent attributes, the attribute is automatically marked as a high-confidence correction dimension, and the conflict attribute association engine is activated. The parameter difference between the first product and the second product in this dimension is quantified into the implicit demand intensity coefficient of the correction logic chain.

[0013] The behavioral weight offsetting effect in the quantitative correction logic chain is further realized through the behavioral chain integrity assessment model. The behavioral chain integrity assessment model calculates the chain confidence based on the spatiotemporal density distribution of behavioral events within the correction logic chain. When the addition behavior, multiple comparison click behavior, and cancellation of addition behavior form a continuous and uninterrupted sequence within the time window, and the time decay coefficient between behaviors is lower than the preset threshold, the behavioral chain integrity is determined to be high. At this time, the weight offsetting ratio adaptive algorithm is triggered, which dynamically increases the baseline value of the offsetting ratio threshold according to the frequency and duration of the comparison click behavior, so that the weight reduction of the initial addition behavior is positively correlated with the clarity of the user's correction intention.

[0014] The generated modified user implicit demand tags are further integrated with tag conflict resolution strategies. When the behavior correction weight offset threshold mechanism reduces the initial addition behavior weight, the target attribute parameter value of the click behavior is compared with the corresponding attribute parameter value of the canceled addition product to generate reverse constraint conditions. Specifically, the attribute parameter value that is frequently compared in the second product is set as the upper or lower limit threshold of the implicit demand tag, and a logical AND operation is performed with the original tag in the explicit demand text to output a composite tag that integrates explicit demand and behavior correction constraints.

[0015] The statistical analysis further employs a spatiotemporal density clustering algorithm to compare click frequencies for each product attribute within the behavioral sequence. This algorithm eliminates noise clicks generated by non-intent user actions. By analyzing and comparing the clustering of click behaviors on the time axis and the spatial proximity of page positions, discrete click events are merged into intent enhancement signals of the same attribute dimension. Only the effective frequencies after clustering are included in the percentage calculation. Simultaneously, the depth interaction index of each click is recorded, and the frequency value is accumulated by weighting the interaction depth.

[0016] The triggering attribute priority dynamic adjustment mechanism further links the attribute dependency network. When the comparison click frequency ratio of a certain attribute exceeds the priority reordering threshold, the pre-built product attribute knowledge graph is queried to identify the derivative dimensions that are strongly related to the attribute. Cross-dimensional priority adjustment is initiated, and the main attribute is promoted to the highest weight position. At the same time, according to the correlation strength ratio marked in the knowledge graph, the weight level of its derivative dimensions is simultaneously increased. Moreover, the weights that are reduced are preferentially extracted from non-related attributes to ensure the weight aggregation effect of the core attribute cluster.

[0017] The proportional reduction of other attribute weights is specifically achieved through a weight rebalancing tree model. The weight rebalancing tree model takes the dynamically increased highest weight attribute as the root node. Based on the semantic similarity of the attributes of each leaf node in the original weight distribution tree and the user's historical behavior preferences, the weight reduction amount is differentially distributed to low-association attributes. Among them, attributes that semantically conflict with the increased attribute receive the largest reduction ratio, while attributes that are compatible with the increased attribute retain the baseline weight, ultimately generating a dynamic weight distribution with attribute coupling relationships.

[0018] The final product recommendation list generated by the large model is further enhanced with an adversarial verification mechanism. Before outputting the recommendation results, the explicit demand text, the corrected implicit demand tags of users, and the dynamic attribute weight distribution are input into the adversarial verification layer of the large model. This layer constructs a virtual product set and injects weight conflict conditions. If the large model still recommends products that violate the implicit tags under conflict conditions, the weight distribution feedback calibration is triggered. The dynamic attribute weight distribution is adjusted in reverse according to the verification results until the recommendation list fully meets the composite tag constraints.

[0019] Further explanation is needed regarding the large model used in this invention: a lightweight large language model (LLM) with a 12-layer encoder-decoder architecture based on the Transformer architecture. It has 768 hidden layer dimensions, 12 attention heads, and fewer than 300 million parameters. Pre-training is based on a dataset of 5 million e-commerce product parameters, and the mapping ability between attributes and requirements is enhanced through an attribute semantic alignment task. Its multi-source data fusion architecture features a dedicated multimodal input layer. Explicit requirement text is converted into text feature sequences via jieba word segmentation and Word2Vec word embedding and input into the text channel. Implicit requirement labels are represented by "attribute name". The threshold constraint key-value pair one-hot vector is transformed and input into the label feature channel. The dynamic attribute weight distribution is mapped to a 768-dimensional feature input into the weight feature channel using the "attribute name-weight value-coupling relationship" triple. The three types of features are fused through a cross-attention mechanism. The decoder generates a candidate list by beam search. The output controllability is defined by embedding constraint instructions in the prompt to clarify the inference direction. An attribute matching verification module is added to the decoder output layer to remove products that do not meet the constraints. The matching degree (attribute fit × total weight ratio) is calculated by combining the weight distribution after adversarial verification. Finally, the recommendation list is output in descending order of matching degree.

[0020] After receiving explicit user request text and real-time user behavior data, in order to accurately capture the temporal correction intent behind user operations and avoid omissions or misjudgments in the correction logic chain due to the inability of a fixed time window to adapt to different user operation rhythms, a dynamic window adaptive mechanism is further introduced when parsing real-time user behavior data based on a preset time window. The specific implementation method is as follows: The core of the dynamic window adaptive mechanism is to automatically adjust the span between the first time window (the monitoring window from when an action is added to the comparison click) and the second time window (the monitoring window from when an action is compared to the cancellation of the action) by monitoring the density and interval distribution of user actions in real time. The density of actions refers to the number of times a user performs a valid action (add, cancel, or compare) on the recommendation interface per unit of time. This is counted every 30 seconds by the system's built-in behavior statistics tool, covering the action data for one minute prior to the current time. The interval distribution is the set of time differences between two consecutive valid actions. The standard deviation of this set is calculated to determine the trend of interval changes. The specific adjustment logic is as follows: When the standard deviation of the operation interval shows a downward trend for three consecutive statistical analyses, indicating a compression of the time interval between consecutive user actions, it suggests that the user's intention to correct behavior is clear and frequent. In this case, the time window span is dynamically shortened: the first time window is reduced from the default 3 minutes to 1 minute, and the second time window from the default 5 minutes to 2 minutes, ensuring rapid capture of high-frequency behavior correction chains. Conversely, when the standard deviation of the operation interval shows an upward trend, it indicates a longer user decision-making cycle, requiring an extended window span: the first time window is extended to 5 minutes, and the second time window to 8 minutes, to avoid breakage of long-cycle behavior chains due to excessively short windows. While adjusting the window span, the behavior pattern recognition module is used to track the addition, removal, and comparison of points. Multimodal intent encoding is performed on click behavior. The behavior pattern recognition module is a processing unit that integrates behavior classification and feature extraction functions. Multimodal intent encoding is the process of combining different types of operation behaviors with corresponding timestamps and interaction attributes, and converting them into weighted behavior vectors in a unified format. First, a fixed basic weight is assigned to each operation behavior. Based on a large amount of user behavior data, a behavior weight of 0.3 is added to represent the initial demand tendency. The click behavior weight of 0.2 represents demand verification and correction. The deselection behavior weight of 0.5 represents demand rejection and explicit correction. Then, the key attributes of each operation are extracted. Among them, adding behavior extracts the added product ID and added timestamp, and comparing click behavior extracts the click product ID, etc. The attributes include module name and dwell time. The "cancel add" behavior extracts the "cancel add" product ID and "cancel" timestamp. Finally, the operation type weight, product association information, and timestamp are combined according to fixed dimensions to form a weighted behavior vector, achieving a unified quantitative representation of different operation behaviors. The generated weighted behavior vectors are arranged in timestamp order and then input into a temporal convolutional network to identify the start and stop boundaries of the correction logic chain. The temporal convolutional network is a deep learning network specifically designed for processing time-series data. It captures the temporal dependencies between behavior vectors through three convolutional layers (kernel size 3, ReLU activation function). The network has been pre-trained on 100,000 sets of labeled user behavior sequences (including complete correction logic chains and invalid behavior sequences). To accurately distinguish between effective correction links and random operations, during identification, the network first locates the initial addition behavior vector (weight 0.3 and containing the addition identifier) ​​as the starting boundary of the correction logic chain. Then, it traverses subsequent behavior vectors along the time axis. When multiple consecutive comparison click behavior vectors are detected, containing the same attribute module name and having a dwell time exceeding a preset threshold, and a subsequent cancellation addition behavior vector (weight 0.5 and containing the cancellation addition identifier) ​​appears, the cancellation addition behavior vector is located as the termination boundary of the correction logic chain. If no cancellation addition behavior is detected even after traversing to the end of the window span, it is determined that a complete correction logic chain has not been formed, and only the current behavior sequence is recorded without triggering subsequent weight adjustments, thus avoiding the inclusion of random clicks or misoperations in the correction analysis.

[0021] After determining the first time window through a dynamic window adaptation mechanism and capturing the user's comparison click behavior for the second product, relying solely on click frequency is insufficient to accurately pinpoint the attribute dimensions that the user is truly interested in. Some users may accidentally click on non-target attribute modules or superficially browse multiple attributes, making it impossible to directly associate them with the intention to make corrections. Therefore, it is necessary to further deploy attribute-level focus capture technology. By deeply analyzing the interaction details between the user and the attribute module, the specific dimensions of specific attribute parameter modules can be identified. The specific implementation method is as follows: When the system detects that a user performs multiple comparison clicks on a specific attribute parameter module of a second product within the first time window after adding the first product, it immediately activates attribute-level focus capture technology. This technology is specifically designed for precise identification of user-focused attribute dimensions in interaction analysis. Its core is achieved by recording the depth of the user's interaction trajectory when clicking on the second product's parameter module. This interaction trajectory depth is a quantitative indicator reflecting the degree of user interaction with the attribute module, including not only the number of clicks but also the interaction method (expand, collapse, parameter comparison) and interaction duration. The system uses page tracking technology to embed listening code in the HTML tags of each attribute parameter module, recording every user interaction action in real time. If a user clicks the "Expand Details" button on the memory capacity module, it's recorded as an "Expand" interaction; clicking the "Collapse" button, it's recorded as a "Collapse" interaction; and clicking the comparison control below the module with added products, it's recorded as a "Parameter Comparison" interaction. Simultaneously, a timer records the duration of each interaction. This data collectively constitutes the foundational information for the depth of the interaction trajectory, preventing misjudgments of user focus solely based on click count. Based on the recorded interaction trajectory depth, the system further identifies specific dimensions of particular attribute parameter modules. The judgment criteria revolve around repeated user interactions with the same attribute module and cursor trajectory. First, it monitors the user's actions on the same page of the second product. If the user repeatedly triggers the expand, collapse, or parameter comparison control of the same attribute module, it indicates that the user... This module is continuously monitored. Simultaneously, the system uses the browser's mouse event listening interface to collect the cursor's coordinates on the page at a frequency of 100 milliseconds per event, generating a cursor hover trajectory and calculating the coverage density of this trajectory within the target attribute module area. The cursor hover trajectory coverage density is the ratio of the number of cursor sampling points falling within the attribute module area to the total number of sampling points corresponding to the total pixel area of ​​that module. For example, if the battery capacity module has a pixel area of ​​200×100 pixels on the page, corresponding to a total of 20,000 sampling points, and the cursor falls on 14,000 sampling points in that area, the coverage density is 70%. The system pre-stores a preset clustering threshold, which is calibrated using 100,000 sets of user attribute interaction data and set to... A value below 70% indicates shallow browsing, while a value above 70% indicates deep engagement. When the cursor hover trajectory coverage density exceeds 70%, combined with previous repeated interactions, this attribute module can be clearly identified as the core comparison dimension in the user's correction logic chain, ensuring that the recognition results match the user's true focus. After determining the core comparison dimension, the system establishes a difference mapping relationship between it and the corresponding attribute of the first product, serving as a conflict attribute identifier in the correction logic chain. The difference mapping relationship refers to extracting the specific parameter values ​​of the second and first products under the core comparison dimension and recording their differences. For example, if the core comparison dimension is battery capacity, the second product's parameter is 5000mAh, and the first product's parameter is 4500mAh.The system generates a mapping record showing the difference of 500mAh between the battery capacity of the second product (5000mAh) and the battery capacity of the first product (4500mAh). If the core comparison dimension is screen resolution, and the second product is 2.5K while the first product is 1080P, then the resolution level difference is recorded. This mapping relationship is marked as a conflict attribute identifier in the correction logic chain. The conflict attribute identifier is used to clarify the core reason why the user canceled adding the first product, i.e., the first product cannot meet the user's needs in this attribute dimension. When generating implicit demand tags subsequently, this identifier will serve as a key basis to ensure that the demand tags accurately reflect the user's correction intentions and avoid generalized attribute dimension judgments.

[0022] After initially identifying specific attribute parameter modules based on interaction trajectory depth using attribute-level focus capture technology, to further improve the confidence of the identification results and avoid misjudgment of core dimensions due to user cursor mis-touches or unconscious repetitive operations, attribute-level focus capture technology needs to be further combined with a visual heatmap feedback mechanism. By capturing more intuitive user visual attention and page browsing behavior, the identification results of specific attribute parameter modules can be dynamically corrected. The specific implementation method is as follows: The core of the visual heatmap feedback mechanism is to construct a heatmap distribution of user visual attention by capturing eye-tracking data and page scrolling speed when the user browses the second product parameter module. This provides supplementary information for attribute dimension identification. The visual heatmap feedback mechanism is a technology that transforms the user's eye focus position, dwell time, and page scrolling status into a visual heatmap (red represents high attention areas, blue represents low attention areas) and uses it to infer interaction intent. Its data collection relies on two types of hardware and algorithms: first, the collection of eye-tracking data, which involves capturing the coordinate position of the user's eye focus on the page in real time through the user's device and recording the dwell time of the focus in each coordinate area, i.e., the time the eye stays in a certain attribute description area, in seconds; second, page scrolling data. The scrolling speed is calculated by collecting the pixel offset of the page scrolling at a frequency of 50 milliseconds per instance through the browser's scroll event listener interface. This offset is then divided by the sampling interval to obtain the real-time scrolling speed. This speed reflects the rhythm of the user's browsing of the attribute module. A slow scrolling speed indicates that the user is carefully examining the corresponding attribute, while a fast speed may indicate that the user is quickly skipping non-focused areas. After inputting the collected eye-tracking data and page scrolling speed into the visual heatmap analysis module, the system first dynamically corrects the identification results of the previously initially identified specific attribute parameter module. The correction logic revolves around the consistency between visual attention and interactive behavior. The system first maps the eye-tracking data to the parameter page layout of the second product, marking the attribute descriptions where the eye lingers for more than 1 second. The system defines the area as having a default effective attention threshold of 1 second; anything below this threshold is considered saccaded. It then compares this area with the attribute modules previously identified through interaction trajectory depth recognition. If the initial identification is the memory capacity module, but the longest eye lingers on the processor model module, it indicates a possible cursor misclick. In this case, the system prioritizes the visual attention area and corrects the specific attribute parameter module to the processor model. If both match, the system strengthens the recognition confidence of that module, avoiding bias from a single data dimension. During dynamic correction, the system focuses on detecting two key behavioral features to determine high-confidence correction dimensions: first, the duration of the user's eye focus on a specific attribute description area is positively correlated with the compared click behavior; a positive correlation means that the longer the dwell time, the more likely the user is to engage with that attribute. The more comparison clicks a feature module receives, the more positive the correlation is considered. The system analyzes the relationship between the most recent 10 comparison clicks and the corresponding dwell time. If five consecutive clicks follow the pattern of increasing dwell time leading to increasing click count, it's considered a positive correlation. Secondly, if page scrolling exhibits an alternating pattern between adjacent attributes, meaning the user's scrolling direction switches back and forth between the page positions corresponding to two adjacent attribute modules, the system can determine this by detecting three consecutive scrolling direction changes (down-up-down) and dwell positions (all within the two adjacent attribute areas). This pattern indicates the user is actively comparing two adjacent attributes, further confirming their focus on the target attribute. When both of these key behavioral characteristics are detected simultaneously...The system automatically marks this attribute as a high-confidence correction dimension. A high-confidence correction dimension is an attribute dimension that simultaneously satisfies the triple verification of interaction trajectory depth, visual attention, and browsing rhythm. Its recognition accuracy is more than 60% higher than that of single interaction trajectory recognition. For example, if the page simultaneously satisfies eye-dwelling time and positive correlation with clicks, and scrolls alternately between 'battery capacity' and 'charging power', then battery capacity is marked as a high-confidence correction dimension, completely eliminating the possibility of misjudgment. After marking, the system immediately activates the conflict attribute association engine. The conflict attribute association engine is a processing unit specifically used to quantify the parameter differences between the first and second products on the high-confidence correction dimension. The core function is to convert parameter differences into coefficients reflecting the intensity of users' implicit needs. First, the engine retrieves the specific parameter values ​​of the first product (the product initially added by the user) and the second product from the product database on the high-confidence correction dimension, calculates the parameter difference between the two, and then combines this with the intensity of the user's interaction behavior on this dimension, comparing 5 clicks and a total dwell time of 15 seconds. This is quantified into an implicit need intensity coefficient according to preset rules. The larger the parameter difference and the higher the intensity of the interaction behavior, the larger the coefficient. This coefficient directly reflects the strength of the user's implicit need for higher battery capacity, providing a core quantitative basis for subsequently generating corrected implicit need tags and reordering attribute priorities.

[0023] After determining the high-confidence correction dimension through the visual heatmap feedback mechanism and quantifying the implicit demand intensity coefficient, in order to accurately measure the degree to which the user's cancellation behavior weakens the initial addition behavior demand weight and avoid erroneous weight adjustment due to incomplete behavior sequence, the behavioral weight offsetting effect in the quantification correction logic chain needs to be further implemented through the behavior chain integrity assessment model. The specific implementation method is as follows: The Behavioral Chain Completeness Assessment Model is an algorithm specifically designed to determine the validity of add, comparison click, and cancel add behavior sequences. Its core principle is to calculate chain confidence by correcting the spatiotemporal density distribution of behavioral events within the logical chain. Spatiotemporal density distribution refers to the interval distribution of behavioral events in the time dimension and the correlation distribution in the spatial dimension. In the time dimension, it statistically analyzes the time intervals between add behaviors, multiple comparison click behaviors, and cancel add behaviors. In the spatial dimension, it determines whether all comparison click behaviors target the same high-confidence correction dimension. If comparison clicks target other attributes, the spatial correlation is considered low. The chain confidence is a dimensionless 0-1 value generated by comprehensively analyzing the spatiotemporal density distribution, used to characterize the completeness and validity of the behavioral chain. The higher the spatiotemporal density, the closer the confidence is to 1, and vice versa. The key threshold condition for the model to determine the completeness of the behavioral chain is: Adding actions, multiple comparison clicks, and canceling add actions form a continuous and uninterrupted sequence within the time window. Continuity and uninterruptedness mean the sequence contains only these three core actions and no other irrelevant operations. The system uses a behavior filtering module to remove invalid actions from the sequence. If the remaining actions are strictly arranged in the order of adding, comparison clicks (at least twice), and canceling add, they are considered continuous. The time decay coefficient between actions is below a preset threshold. The time decay coefficient is a parameter reflecting the impact of the action interval on the relevance of requirements; the longer the interval, the larger the coefficient. The preset threshold is set to 0.5. When the interval exceeds 3 minutes, the coefficient... If the decay coefficients for adding to the comparison click and the time decay coefficients from comparison click to cancellation are both below 0.5, it indicates a strong correlation between the needs of the behaviors. When both of the above conditions are met simultaneously, the model determines that the behavior chain has high completeness (chain confidence greater than or equal to 0.8), which is the prerequisite for triggering weight adjustment. If either condition is not met, the behavior chain is determined to be incomplete, and weight offset is not executed temporarily to avoid including random behaviors in the requirement correction. After the determination result of high behavior chain completeness is generated, the system immediately triggers the weight offset ratio adaptive algorithm. The weight offset ratio adaptive algorithm is based on the clarity of the user's correction intention. The algorithm for dynamically adjusting the offsetting ratio threshold works by positively correlateding the initial weight reduction of the added behavior with the clarity of the user's correction intent. The clearer the correction intent, the higher the offsetting ratio threshold and the more significant the weight reduction. The algorithm first extracts key features of the comparison click behavior, comparing click frequency (e.g., 5 clicks for battery capacity comparison) and total duration (the sum of the dwell times of all comparison clicks, 15 seconds). These two features directly reflect the clarity of the correction intent. Then, it dynamically increases the baseline value of the offsetting ratio threshold according to preset rules. The default baseline value is set to 0.3 (i.e., an initial weight reduction of 30%). For every 2 additional clicks or 5 additional seconds of total duration compared to the previous click, the baseline value increases by 0.1, up to a maximum of 0.8 (to avoid excessive weight reduction that could invalidate the demand tag). For example, compared to 5 clicks and a total duration of 15 seconds, the baseline value increases from 0.3 to 0.3 + (5-2) ÷ 2 × 0.1 + (15-5) ÷ 5 × 0.1 = 0.3 + 0.15 + 0.2 = 0.65. At this point, the initial weight of the added behavior in the demand tag (e.g., the original weight of 0.8) is reduced proportionally, resulting in a weight of 0.8 × (1-0.65) = 0.28, accurately matching the user's clear intention to make corrections.

[0024] After quantifying the behavioral weight offsetting effect and weakening the demand weight of the initially added behavior, it is necessary to generate implicit demand tags that better reflect the actual needs based on the user's correction logic chain. However, these tags may conflict with the original tags in the user's explicit demand text. Therefore, the generation of corrected implicit user demand tags needs to further integrate tag conflict resolution strategies. The specific implementation method is as follows: The core of the tag conflict resolution strategy is to generate reverse constraints by analyzing the attribute parameter preferences in user behavior, and then merge them with explicit demand tags to output a conflict-free composite tag. Its execution begins after the behavior correction weight offset threshold mechanism reduces the initial addition behavior weight. At this point, the demand tag weight corresponding to the initial addition behavior has been significantly reduced, freeing up weight space for new tag generation. The system first analyzes and compares the target attribute parameter value of the click behavior with the corresponding attribute parameter value of canceling the addition of a product. The target attribute parameter value is the specific parameter of the second product under the high-confidence correction dimension. The specific value and attribute unit of this parameter are retrieved from the product database. The corresponding attribute parameter value of canceling the addition of a product is the parameter of the first product under the same high-confidence correction dimension. Similarly, complete parameter information is retrieved. Based on the analyzed two parameter values, the system generates reverse constraints. These reverse constraints refer to inferring the user's parameter preference constraints for the target attribute by comparing the second product and canceling the first product. The specific rules are as follows: If the parameter value of the second product is greater than that of the first product and the user frequently compares this attribute of the second product, then the constraint condition is set to the target attribute parameter value being greater than or equal to the parameter value of the second product. If the parameter value of the second product is less than that of the first product and the user frequently compares this attribute, then the constraint condition is set to the target attribute parameter value being less than or equal to the parameter value of the second product. This constraint condition directly excludes the parameter range that the user has already denied through cancellation behavior, clearly defining the parameter boundaries of implicit needs. After generating the reverse constraint condition, the system performs a logical AND operation with the original label in the explicit need text. The logical AND operation means taking the intersection of the constraint ranges of the two labels to ensure that the merged label satisfies both the explicit needs explicitly stated by the user and the implicit preferences reflected by the behavior modification. For example, the original label in the explicit need text is battery capacity greater than or equal to 4800mAh, and the reverse constraint condition is battery capacity greater than or equal to 5000mAh. After the logical AND operation, the intersection of the two is taken, i.e., battery capacity greater than or equal to 5000mAh. If the explicit... The requirement is a price of 3000 yuan or less, and the reverse constraint is memory of 16G or more. Since there is no conflict between the two, they are directly merged into a price of 3000 yuan or less and memory of 16G or more. This calculation process is executed by the tag fusion module. The module has a built-in attribute parameter logic judgment unit that can automatically identify parameter types to ensure that the calculation rules are adapted to different attributes. The final output of the calculation is a composite tag that integrates explicit requirements and behavioral correction constraints. This tag is the corrected implicit user requirement tag, which retains the user's explicitly expressed demand tendency, incorporates implicit preferences mined through behavioral correction, and eliminates conflicts between tags. For example, if the user's explicit requirement is a thin and light laptop with 8G of memory, the correction logic chain reflects that they compare products with 16G of memory and cancel products with 8G of memory. The composite tag is a thin and light laptop with memory of 16G or more. This tag will serve as the core requirement basis for subsequent reordering of product attribute priorities and large model recommendations to ensure that the recommendation results match the user's real needs.

[0025] After generating the corrected user implicit demand tags, in order to accurately quantify the degree of user attention to each product attribute and avoid including unintentional actions such as accidental touches and quick swipes in valid clicks, which could lead to bias in attribute priority judgment, the spatiotemporal density clustering algorithm needs to be further used to optimize the statistical process when comparing the click frequency for each product attribute in the statistical behavior sequence. The specific implementation method is as follows: Spatiotemporal density clustering is an intelligent statistical algorithm that integrates the clustering of the temporal dimension and the proximity of the spatial dimension to filter valid clicks and eliminate noise. Its core logic is to group click events with similar spatiotemporal characteristics into the same intent signal, while classifying isolated events as noise. First, the algorithm performs data preprocessing on all comparative click behaviors in the behavior sequence: extracting key information for each click, including the timestamp and page coordinates, and arranging them in timestamp order to form a click dataset with both time and coordinate dimensions, laying the foundation for subsequent clustering analysis. The core step of this algorithm is to eliminate noise clicks generated by non-intent-based user operations, specifically including: The analysis compares the clustering of click behavior along the timeline. Temporal clustering refers to whether the time interval between consecutive click events is within a preset threshold. This threshold is calibrated to 5 seconds through user behavior experiments; that is, the interval between two clicks is less than or equal to 5 seconds, which is considered temporal clustering. The system calculates the time difference between adjacent clicks. If the time difference between a click and the clicks before and after it is greater than 10 seconds, it is initially judged as an isolated click. The analysis also analyzes the spatial proximity of page positions. Spatial proximity refers to whether the click coordinates fall within the pixel range of the same product attribute module. This is done by calculating the Euclidean distance between the click coordinates and the center coordinates of each attribute module. If the distance is greater than 150 pixels... Clicks exceeding 50 pixels beyond the module boundary are considered spatially isolated. Only clicks satisfying both temporal and spatial isolation are identified as noise clicks and removed from the dataset. Examples include accidental clicks on blank areas of the page or random clicks on different attributes at 15-second intervals. For retained non-noise clicks, the algorithm merges discrete click events into intent enhancement signals of the same attribute dimension using spatiotemporal density correlation. If the time interval of multiple discrete clicks is less than or equal to 5 seconds and their coordinates all fall within the pixel range of the processor model module, the algorithm considers these three clicks as a group. The continued intent expression is merged into a single intent reinforcement signal for the processor model attribute. This avoids inflated frequency due to repeated clicks on the same module. If an attribute has two such intent reinforcement signals (e.g., the battery capacity module has 5 clicks), they are merged into two groups, and the initial effective frequency for that attribute is recorded as 2. This ensures that frequency statistics reflect the user's true attention intent rather than the number of operations. During the statistical process, only the clustered effective frequencies are included in the percentage calculation. The system first classifies and counts by attribute dimension, then calculates the percentage of each attribute's effective frequency to the total effective frequency of all attributes. This percentage accurately reflects the impact of different attributes on the user's corrective logic. The attention priority in the chain provides a quantitative basis for subsequent attribute priority adjustments. At the same time, the system records the depth interaction index of each click and accumulates the frequency value by weighting the interaction depth. The depth interaction index is a quantitative value assigned based on the degree of interaction between the user and the attribute module. During weighted accumulation, the depth interaction index of all clicks within the same intent reinforcement signal is added together as the weighted frequency of that signal. Finally, the weighted frequency replaces the initial effective frequency in the calculation of the proportion, ensuring that the frequency statistics not only reflect the number of clicks, but also the interaction depth, avoiding shallow clicks and deep interactions being counted equally, and further improving the accuracy of statistics.

[0026] After obtaining the weighted frequency proportion of each attribute through spatiotemporal density clustering algorithm, when the proportion of any attribute exceeds the preset priority reordering threshold, a dynamic attribute priority adjustment mechanism needs to be triggered. However, simply increasing the priority of the main attribute may ignore the influence of its associated derived dimensions, otherwise it will lead to the recommendation focusing only on the main attribute and missing key association requirements. Therefore, this mechanism needs to further associate the attribute dependency network. The specific implementation is as follows: The core of the attribute priority dynamic adjustment mechanism is to achieve coordinated priority adjustment between primary attributes and derived dimensions through a network of attribute dependency relationships. This network is an attribute association system based on a pre-built product attribute knowledge graph. The product attribute knowledge graph is a structured relational database built by the system by crawling industry product manuals, e-commerce platform product parameter pages, and user review data. It stores the primary and secondary relationships and association strengths of each product attribute. This knowledge graph is pre-stored in the system's attribute association library, allowing for quick lookup of the corresponding derived dimensions and association strength using the attribute's unique ID. When the weighted frequency of a certain attribute exceeds the priority reordering threshold, the system automatically triggers an attribute association query. The API interface of the attribute association library, when inputting the ID of the main attribute, immediately returns all derived dimensions with an association strength greater than or equal to 0.6 with the main attribute, along with the association strength value of each derived dimension. This ensures that the selected derived dimensions are highly relevant to the main attribute, avoiding association with irrelevant attributes. After the query is completed, the system initiates cross-dimensional priority adjustment. First, the priority of the main attribute (processor model) is directly elevated from its original weight position to the highest weight position, making it the primary consideration for product recommendations. Then, according to the association strength ratio marked in the knowledge graph, the weight levels of the derived dimensions are synchronously increased. The weight increase of the main attribute is set as the baseline value, and the increase of the derived dimension = increase of the main attribute × relevance. To ensure the correlation strength between the weight increase of derived dimensions and the correlation strength of the main attribute, a core attribute cluster is formed with the main attribute as the core and derived dimensions as support. While increasing the weight of the core attribute cluster, the weights of other attributes need to be reduced proportionally to maintain the total weight sum of 1, avoiding weight overflow that would affect the recommendation calculation. The reduction rule is clearly defined as prioritizing extraction from non-related attributes. Non-related attributes refer to attributes in the knowledge graph with a correlation strength of less than 0.3 with the main attribute. The system first calculates the total weight increase of the core attribute cluster, and then allocates the reduction amount according to the original weight proportion of the non-related attributes. For example, if the original weight of screen size is 0.15, body color is 0.05, and interface type is 0.1, the total non-related weight is 0.3, and the reduction amount of screen size is 0. 0.96 × (0.15 ÷ 0.3) = 0.48, down from 0.15 to 0.07; Body color down by 0.96 × (0.05 ÷ 0.3) = 0.16, down from 0.05 to 0; Interface type down by 0.96 × (0.1 ÷ 0.3) = 0.32, down from 0.1 to 0. If the total weight of non-related attributes is insufficient, the remaining down by 0.96 × (0.1 ÷ 0.3) = 0.32, down from 0.1 to 0. If the total weight of non-related attributes is insufficient, the remaining down by 0.96 × (0.15 ÷ 0.3) = 0.48, down from 0.15 to 0.32, down from 0.1 ...

[0027] After determining the weight increase range of core attribute clusters and the reduction range of non-related attributes through the attribute dependency network, a simple average reduction method may lead to excessive reduction of compatible attributes or insufficient reduction of conflicting attributes, which may interfere with the accuracy of subsequent recommendations. Therefore, proportionally reducing the weights of other attributes needs to be implemented through a weight rebalancing tree model. This model can achieve differentiated allocation based on attribute semantic association and user preferences. The specific implementation method is as follows: The weight rebalancing tree model is an algorithmic model based on a hierarchical tree structure that dynamically allocates weight adjustments based on semantic association and behavioral preferences. Its construction logic uses the highest dynamically increased weight attribute as the root node. The weight value of the root node is determined as the highest weight after adjustment. The model radiates from this root node to connect all other attributes to be downgraded, forming a two-level tree structure of root and leaf nodes. The leaf nodes represent the attributes whose weights need to be downgraded, and each leaf node pre-stores semantic similarity and user historical behavioral preference scores with the root node. Semantic similarity is extracted from the product attribute knowledge graph and calculated using a cosine similarity algorithm, ranging from 0 to 1. User historical behavioral preference scores are calculated based on the frequency of user attention to the attribute in the past 3 months, also ranging from 0 to 1. These two data points are the core basis for subsequent differentiated allocation. After the model starts, it first reads the leaf node attribute data from the original weight distribution tree, which was constructed by classifying each product attribute according to its semantic association before adjustment. The hierarchical structure consists of a root node representing the core product category and leaf nodes representing specific attributes. Each leaf node is labeled with an initial weight value. The model associates and binds these initial weights with the semantic similarity of the root node and the user's historical preference score, forming a four-dimensional dataset of attributes, initial weights, semantic similarity, and preference scores. This ensures the integrity of the basic information for each attribute to be downgraded. Subsequently, the model classifies all leaf node attributes into three categories based on the four-dimensional dataset. The first category consists of attributes that semantically conflict with the boosting attribute (root node), with the criteria being a semantic similarity of less than 0.3 and a user preference score of less than 0.4. The second category consists of attributes with low relevance, with the criteria being a semantic similarity of 0.3-0.6 and a user preference score of 0.4-0.6. The third category consists of attributes that are compatible with the boosting attribute, with the criteria being a semantic similarity of greater than 0.6 or a user preference score of greater than 0.7. The purpose of this classification is to assign differentiated downgrade ratios to attributes with different degrees of relevance, avoiding a one-size-fits-all adjustment approach.

[0028] After classification, the model calculates the total weight reduction, which is the sum of the weight increases of all attributes in the core attribute cluster. For example, if the main attribute increases by 0.4, the derived attributes (processor clock speed increases by 0.32, and number of cores increases by 0.28) result in a total increase of 0.96. Since the total weight needs to be maintained at 1, the total reduction is 0.96. This total reduction is then distributed differentially to each attribute based on the classification results. For semantically conflicting attributes, due to their contradiction with core attribute requirements and low user attention, the largest reduction ratio is assigned. For example, the original weight of body thickness is 0.15, and the allocated reduction is 0.96 × 50% × ( 0.15 ÷ Total weight of all conflicting attributes). If only the fuselage thickness is a conflicting attribute, its reduction amount = 0.96 × 50% = 0.48, and the weight after reduction = 0.15 - 0.48 (since the weight cannot be negative, it is taken as 0). For low-relevance attributes, a medium reduction ratio is assigned (accounting for 30% of the total reduction amount). For example, the original weight of the interface type is 0.1, and considering its higher user preference score (0.7), it is adjusted according to the reduction ratio = base ratio × (1 - preference score × 0.5), that is, the base ratio is 30% × (1 - 0.7 × 0.5) = 30% × 0.65 = 19.5%, with an allocated reduction of 0.96 × 19.5% = 0.187. After the reduction, the weight = 0.1 - 0.187 (taken as 0.013). To avoid overlooking requirements due to user concerns, for compatibility attributes, only the minimum reduction percentage (accounting for 20% of the total reduction) is assigned, and the reduction amount does not exceed 20% of its original weight (retaining the baseline weight). For example, the original weight of memory capacity is 0.2, the base percentage is 20% × (1 - 0.8 × 0.5) = 20% × 0.6 = 12%, and the allocated reduction is 0.96 × 12% = 0.115. Since 0.11... Since 5 is greater than 0.2 × 20% = 0.04, it is ultimately adjusted down by 0.04. After the adjustment, the weight = 0.2 - 0.04 = 0.16, ensuring that its synergy with the core attribute is not disrupted. After all the adjustment amounts for all attributes are distributed, the model performs normalization verification on the adjusted weights to ensure that the sum of all attribute weights is 1. The coupling relationship between each attribute and the root node is marked in the final output dynamic weight distribution. This dynamic weight distribution with attribute coupling relationship can provide a basis for association when the large model integrates demand information in the future, further improving the rationality and accuracy of the recommendation results.

[0029] After generating a dynamic attribute weight distribution with attribute coupling relationships using a weight rebalancing tree model, although the attribute weights have been differentiated based on semantic association and user preferences, the large model may still recommend products that do not meet implicit needs when integrating explicit demand text, corrected implicit user demand tags, and dynamic weights due to parameter association bias or improper weight boundary value settings. Therefore, an adversarial verification mechanism needs to be added to generate the final product recommendation list using the large model. This mechanism tests the demand-meeting capability of the large model by constructing extreme conflict scenarios, corrects weight bias in advance, and ensures that the recommendation results fully match the user's real needs. The specific implementation method is as follows: The key validation step in generating the final product recommendation list using a large model is to input the explicit demand text, the corrected implicit user demand tags, and the dynamic attribute weight distribution into the adversarial validation layer of the large model before outputting the recommendation results. The adversarial validation layer is a demand compliance detection module built into the large model, specifically designed to simulate extreme scenarios where weight configuration and demand constraints conflict. It shares the same demand parsing interface with the recommendation generation layer of the large model, ensuring that the input data format is completely consistent. The explicit demand text is input in a structured format according to attributes and parameters, the corrected implicit user demand tags are labeled with attributes and threshold constraints, and the dynamic attribute weight distribution is input in the form of a list of attributes, weight values, and coupling relationships. This avoids validation deviations caused by differences in data format and ensures that the demand parsing logic of the validation scenario is consistent with that of the real recommendation scenario.

[0030] After receiving the input data, the adversarial validation layer first constructs a set of virtual products and injects weight conflict conditions. The construction of the virtual product set aims to cover various scenarios where attributes violate implicit needs. The adversarial validation layer extracts all attribute dimensions that users care about from the product attribute knowledge graph and generates virtual products according to the rule of taking extreme values ​​for each attribute. The injection of weight conflict conditions involves deliberately amplifying the weight of attributes that conflict with implicit needs and suppressing the weight of attributes that meet the needs in the weight configuration of the virtual products. After completing the configuration of the virtual product set and conflict conditions, the adversarial validation layer inputs it into the large model, requiring the large model to output a recommendation ranking according to the current dynamic attribute weight distribution. If the large model still fails to meet the conflict conditions... If a product is recommended that violates the implicit demand label, it indicates a deviation in the demand analysis of the large model, immediately triggering a weight distribution feedback calibration. The core basis for calibration is the type of attribute that violates the recommendation. If a product is recommended due to excessive body thickness, it means that although the weight of body thickness is low, it is not enough to not affect the recommendation decision and needs to be further reduced. If a product is recommended due to an incompatible processor model, it means that although the weight of the processor model is high, it is not enough to suppress the interference of conflicting attributes and needs to be appropriately increased to ensure that the calibration direction accurately matches the root cause of the deviation. The weight distribution feedback calibration is performed by the system's built-in weight calibration module. This module first locates the attribute dimension associated with the violation based on the verification results, and then queries the attribute. The semantic conflict level with the core attribute (processor model) is adjusted according to the rule that the higher the conflict level, the greater the reduction. For example, the current weight of the body thickness is 0.013, but due to the high semantic conflict level and the resulting recommendation violation, it is further reduced to 0.005. For core attributes that meet implicit requirements but whose weights are suppressed, causing violations, the weights are finely adjusted according to the rule that the higher the compatibility level, the greater the increase, increasing from 0.6 to 0.65. At the same time, by reducing the weights of other low-relevance attributes, the total weight of all attributes is ensured to remain at 1 to avoid weight overflow. The adjusted dynamic attribute weight distribution is then re-input into the adversarial verification layer to repeatedly construct virtual products, inject conflict conditions, and perform recommendation verification. The process involves iterative calibration if another violation occurs, continuing until, in two consecutive verifications, the top 5 virtual products recommended by the large model fully meet the implicit demand tags and still prioritize products that meet the constraints even under weight conflict conditions. At this point, it is determined that the dynamic attribute weight distribution has fully adapted to the demand constraints. Finally, the large model calls this calibrated weight distribution, combines the explicit demand text with the corrected implicit demand tags, selects products from the real product library that meet all constraints, calculates the product matching degree according to the dynamic weights, and generates a sorted final product recommendation list. This ensures that the recommendation results not only match the user's explicit expression but also fit the implicit preferences after behavior correction, while avoiding attribute coupling conflicts.

[0031] This method first receives explicit user needs text and real-time behavioral data. It then parses the behavioral data using a dynamic window adaptive mechanism, identifying the temporal correction logic chain of adding, comparing clicks, and canceling additions. Combining attribute-level focus capture and visual heatmap feedback, it locates core attributes, quantifies the behavioral weight offsetting effect, and generates composite tags that integrate explicit and implicit needs. It uses spatiotemporal density clustering to statistically compare the effective click frequency of attributes, and combines attribute dependency networks to rearrange attribute priorities. A weight rebalancing tree model is used to differentiate weights, forming a dynamic attribute weight distribution. Finally, the large model integrates the three types of information, calibrates the weights through adversarial validation, and generates a product recommendation list that matches the user's true preferences, improving the accuracy and personalization of recommendations.

[0032] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart product recommendation method based on a large model, characterized by: Includes the following steps: S1. Receive explicit demand text and real-time user behavior data input by the user. The explicit demand text includes product attribute requirements directly proposed by the user. The real-time user behavior data includes the user's operation behavior and corresponding timestamp on the product recommendation interface. S2. Analyze real-time user behavior data based on a preset time window, identify and quantify the time-series correction features of user behavior, whereby the time-series correction features refer to the correction logic chain formed by user operation behavior over time, including: Starting with the user's initial product addition behavior, a dynamic time window is set to monitor subsequent behaviors. When it is detected that the user performs multiple comparison clicks on a specific attribute parameter module of the second product within the first time window after adding the first product, and the dwell time of each comparison click exceeds a preset threshold, and the user triggers the behavior of canceling the addition of the first product within the second time window, the behavior sequence is extracted as a correction logic chain, and the behavior weight offsetting effect in the correction logic chain is quantified. This is achieved through a behavior correction weight offsetting threshold mechanism. That is, when the cancellation of the addition behavior occurs, the weight of the initial addition behavior in the demand tag is automatically reduced according to the preset offsetting ratio threshold, and a corrected user implicit demand tag is generated. S3. Based on the user behavior time sequence correction features, rearrange the priority of product attribute dimensions, count the comparison click frequency for each product attribute in the behavior sequence, and calculate its proportion of the total comparison click frequency of all attributes. When the comparison click frequency proportion of any attribute exceeds the preset priority rearrangement threshold, trigger the attribute priority dynamic adjustment mechanism, raise the priority of the attribute from the original weight position to the highest weight position, and at the same time reduce the weight of other attributes proportionally to form a rearranged dynamic attribute weight distribution. S4. Using a large model, integrate the explicit demand text, the corrected implicit user demand tags, and the rearranged dynamic attribute weight distribution to generate the final product recommendation list.

2. The intelligent product recommendation method based on a large model according to claim 1, characterized in that: The method of parsing real-time user behavior data based on a preset time window further includes introducing a dynamic window adaptive mechanism. This mechanism automatically adjusts the span between the first and second time windows by monitoring the density and interval distribution of user operation behaviors in real time. Specifically, when the time interval of continuous user operation behaviors is detected to be compressing, the time window span is dynamically shortened to capture high-frequency correction intentions; conversely, the window span is extended to associate long-cycle behavior chains. At the same time, the behavior pattern recognition module performs multimodal intent encoding on add, cancel add, and comparison click behaviors, transforming the timestamp sequence of operation behaviors into a weighted behavior vector, which is then input into the temporal convolutional network to identify the start and stop boundaries of the correction logic chain.

3. The intelligent product recommendation method based on a large model according to claim 1, characterized in that: When it is detected that a user performs multiple comparative clicks on a specific attribute parameter module of the second product within the first time window after adding the first product, attribute-level focus capture technology is further deployed to identify the specific dimensions of the specific attribute parameter module by recording the depth of the interaction trajectory when the user clicks on the parameter module of the second product. When a user repeatedly triggers the expand, collapse, or parameter comparison controls of the same attribute module on the same product page, and the coverage density of the cursor hover trajectory in the module area exceeds the preset clustering threshold, it is determined that the core comparison dimension in the user's correction logic chain is established, and a difference mapping relationship is established between the core comparison dimension and the corresponding attribute of the first product, which serves as the conflict attribute identifier of the correction logic chain.

4. The intelligent product recommendation method based on a large model according to claim 3, characterized in that: The attribute-level focus capture technology further combines a visual heatmap feedback mechanism. By capturing eye-tracking data and page scrolling speed when the user browses the second product parameter module, it dynamically corrects the recognition results of the specific attribute parameter module. When it is detected that the duration of the user's eye focus in a certain attribute description area is positively correlated with the comparison click behavior, and the page scrolling shows an alternating pattern between adjacent attributes, the attribute is automatically marked as a high-confidence correction dimension, and the conflict attribute association engine is activated. The parameter difference between the first product and the second product in this dimension is quantified as the implicit demand intensity coefficient of the correction logic chain.

5. The intelligent product recommendation method based on a large model according to claim 1, characterized in that: The behavioral weight offsetting effect in the quantitative correction logic chain is further realized through the behavioral chain integrity assessment model. The behavioral chain integrity assessment model calculates the chain confidence based on the spatiotemporal density distribution of behavioral events within the correction logic chain. When the addition behavior, multiple comparison click behavior, and cancellation of addition behavior form a continuous and uninterrupted sequence within the time window, and the time decay coefficient between behaviors is lower than the preset threshold, the behavioral chain integrity is determined to be high. At this time, the weight offsetting ratio adaptive algorithm is triggered, which dynamically increases the baseline value of the offsetting ratio threshold according to the frequency and duration of the comparison click behavior, so that the weight reduction of the initial addition behavior is positively correlated with the clarity of the user's correction intention.

6. The intelligent product recommendation method based on a large model according to claim 5, characterized in that: The generated and corrected user implicit demand tags further integrate tag conflict resolution strategies. When the behavior correction weight offset threshold mechanism reduces the initial addition behavior weight, reverse constraint conditions are generated by parsing and comparing the target attribute parameter value of the click behavior with the corresponding attribute parameter value of the canceled product. Specifically, the attribute parameter value that is frequently compared in the second product is set as the upper or lower limit threshold of the implicit demand tag, and a logical AND operation is performed with the original tag in the explicit demand text to output a composite tag that integrates explicit demand and behavior correction constraints.

7. The intelligent product recommendation method based on a large model according to claim 1, characterized in that: The statistics further employ a spatiotemporal density clustering algorithm to compare the click frequency for each product attribute in the behavioral sequence. The spatiotemporal density clustering algorithm eliminates noise clicks generated by non-intent user operations. By analyzing and comparing the clustering of click behaviors on the time axis and the spatial proximity of page positions, discrete click events are merged into intent enhancement signals of the same attribute dimension. Only the effective frequencies after clustering are included in the proportion calculation. At the same time, the depth interaction index of each click is recorded, and the frequency value is accumulated by weighting the interaction depth.

8. The intelligent product recommendation method based on a large model according to claim 7, characterized in that: The trigger attribute priority dynamic adjustment mechanism further links the attribute dependency network. When the comparison click frequency ratio of a certain attribute exceeds the priority reordering threshold, the pre-built product attribute knowledge graph is queried to identify the derivative dimensions that are strongly related to the attribute. Cross-dimensional priority adjustment is initiated, and the main attribute is promoted to the highest weight position. At the same time, according to the correlation strength ratio marked in the knowledge graph, the weight level of its derivative dimensions is simultaneously increased. Moreover, the weights that are reduced are preferentially extracted from non-related attributes to ensure the weight aggregation effect of the core attribute cluster.

9. The intelligent product recommendation method based on a large model according to claim 8, characterized in that: The proportional reduction of other attribute weights is specifically achieved through a weight rebalancing tree model. The weight rebalancing tree model takes the dynamically increased highest weight attribute as the root node. Based on the semantic similarity of the attributes of each leaf node in the original weight distribution tree and the user's historical behavior preferences, the weight reduction amount is differentially allocated to low-association attributes. Among them, attributes that semantically conflict with the increased attribute receive the largest reduction ratio, while attributes that are compatible with the increased attribute retain the baseline weight, ultimately generating a dynamic weight distribution with attribute coupling relationships.

10. The intelligent product recommendation method based on a large model according to claim 8, characterized in that: The process of generating the final product recommendation list using a large model further incorporates an adversarial verification mechanism. Before outputting the recommendation results, the explicit demand text, the corrected implicit user demand tags, and the dynamic attribute weight distribution are input into the adversarial verification layer of the large model. This layer constructs a virtual product set and injects weight conflict conditions. If the large model still recommends products that violate the implicit tags under conflict conditions, it triggers weight distribution feedback calibration. Based on the verification results, the dynamic attribute weight distribution is adjusted in reverse until the recommendation list fully satisfies the composite tag constraints.

Images