A Multimedia Recommendation Method and System for Products Combining RPA and AI

Abstract

This invention provides a product multimedia recommendation method and system combining RPA and AI. First, an RPA interaction capture module records the current interaction flow between the user and the product multimedia interface in real time, including the timing of operation triggers and attention trajectories. Next, based on a pre-set historical interaction pattern library, the intent evolution trajectory is extracted, covering the dynamic shift characteristics and intensity trends of user interests and preferences. A dynamic matching rule set is generated based on the intent evolution trajectory, including association constraints and priority ranking logic. This dynamic matching rule set is input into a pre-trained AI recommendation model to perform rule matching and filtering on a set of candidate product multimedia content, generating filtering results. Finally, an RPA display and arrangement module renders the recommendation sequence in real time based on the filtering results, adjusting the display size and text layout style, thereby improving the personalization of product multimedia recommendations and the user experience.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a product multimedia recommendation method and system that combines RPA and AI. Background Technology

[0002] In the e-commerce and multimedia content recommendation field, accurately grasping user interests and achieving personalized recommendations is key to improving user experience and platform efficiency. Existing product recommendation methods have significant shortcomings. On the one hand, most methods rely solely on users' historical purchase records or simple browsing behavior, lacking in-depth capture of real-time user interactions. When users browse product multimedia interfaces, their actions and attention spans contain rich information about their interests and preferences, but traditional methods fail to effectively utilize this dynamic information, resulting in recommendations that don't align well with users' current interests. On the other hand, existing recommendation rules are often statically set, unable to adjust in a timely manner according to dynamic changes in user interests, making it difficult to adapt to shifts in user interests during browsing. Furthermore, the presentation of recommendation results lacks flexibility and personalization, typically employing a uniform display method that cannot be optimized based on the characteristics of the recommended content and user preferences, thus affecting user acceptance of the recommended content. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a product multimedia recommendation method combining RPA and AI, the method comprising:

[0004] The RPA interaction capture module records the current interaction flow between the user and the product multimedia interface in real time. The current interaction flow includes the timing of the user's operation on the product image unit and the trajectory of the user's attention on the product description unit.

[0005] Based on a preset historical interaction pattern library, the intent evolution trajectory of the current interaction behavior flow is extracted. The intent evolution trajectory includes the dynamic shift features of user interest preferences and the trend of feature intensity changes.

[0006] A dynamic matching rule set is generated based on the intent evolution trajectory. The dynamic matching rule set includes the association constraints between product image features and user interest preferences, as well as the priority ranking logic of product description features.

[0007] The dynamic matching rule set is input into the pre-trained AI recommendation model to perform rule matching filtering on the candidate product multimedia content set, and generate product multimedia content filtering results containing matching degree scores.

[0008] The RPA display orchestration module performs real-time rendering of the recommended sequence based on the product multimedia content filtering results. The real-time rendering operation includes adjusting the display size of the product image unit and the text layout style of the product description unit.

[0009] In another aspect, embodiments of the present invention also provide a product multimedia recommendation system combining RPA and AI, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0010] Based on the above, this embodiment of the invention uses an RPA interaction capture module to record the user's interaction behavior flow with the product multimedia interface in real time, accurately capturing real-time dynamic information such as the timing of user operation triggers and attention trajectories. Based on a preset historical interaction pattern library, it extracts the intent evolution trajectory, clearly presenting the dynamic shift characteristics and intensity trends of user interests and preferences, making the system's grasp of user interests more comprehensive and dynamic. A dynamic matching rule set is generated based on the intent evolution trajectory, enabling real-time adjustment of recommendation rules according to user interests, improving the matching degree between recommended content and the user's current interests. The dynamic matching rule set is input into a pre-trained AI recommendation model for rule matching and filtering, generating filtering results including matching degree scores, enhancing the accuracy and scientific nature of the recommendations. Finally, the RPA display orchestration module renders the recommendation sequence in real time based on the filtering results, optimizing the display method of product images and descriptions, increasing user attention and acceptance of recommended content, thereby significantly improving the personalization and user experience of product multimedia recommendations. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the execution flow of the multimedia product recommendation method combining RPA and AI provided in an embodiment of the present invention.

[0012] Figure 2 This is a schematic diagram of exemplary hardware and software components of a multimedia product recommendation system combining RPA and AI provided in an embodiment of the present invention. Detailed Implementation

[0013] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a product multimedia recommendation method combining RPA and AI according to an embodiment of the present invention. The following is a detailed description of this product multimedia recommendation method combining RPA and AI.

[0014] Step S110: Record the current interaction behavior flow between the user and the product multimedia interface in real time through the RPA interaction capture module. The current interaction behavior flow includes the timing of the user's operation on the product image unit and the trajectory of the user's attention on the product description unit.

[0015] In e-commerce platform product recommendation scenarios, when users enter the product multimedia interface to browse products, the RPA interaction capture module records the interaction behavior between the user and the interface.

[0016] Before data collection, explicit and valid authorization from the user is required. When a user enters the product multimedia interface, relevant information about data collection will be displayed to the user in a prominent and easily understandable manner, including the data type collected (such as the trigger sequence of operations on product image units, the attention trajectory of product description units, etc.), the purpose of collection (for product multimedia recommendations to improve the user's shopping experience), the method of data use, and the scope of data sharing. Users can choose whether to agree to data collection. If the user agrees, authorization will be clearly indicated by electronic signature or click confirmation, ensuring the authenticity and validity of the authorization.

[0017] For user privacy-sensitive data, such as gaze trajectories and pupil diameter changes obtained from eye-tracking, privacy protection and leak prevention technologies are required. During the data acquisition phase, the data is encrypted using advanced encryption algorithms to convert it into ciphertext for transmission and storage, preventing the data from being stolen or tampered with during transmission. For example, symmetric encryption algorithms are used to encrypt eye-tracking data, and only authorized system components can decrypt it using the corresponding key.

[0018] Regarding data storage, data will be stored on secure and reliable servers with strict access control mechanisms. Only authorized personnel and system modules can access this data, and access activities will be logged in detail for auditing and traceability. Furthermore, stored data will be backed up regularly to prevent data loss due to hardware failures, natural disasters, or other reasons.

[0019] During data usage, we strictly adhere to the principle of minimum necessity, using only data relevant to product multimedia recommendations and refraining from using data for any other unauthorized purposes. Furthermore, we anonymize the data, removing information that can directly or indirectly identify users to reduce the risk of privacy breaches. For example, when using user action trigger sequences and attention trajectories, we replace or de-identify users' personal information.

[0020] When data is no longer needed, it will be securely destroyed according to the prescribed procedures to ensure that the data cannot be illegally recovered and used. Through these measures, it is ensured that the data collection process described in step S110 fully complies with legal requirements, protecting the legitimate rights and privacy of users.

[0021] Step S111: Start the RPA interaction capture module to load the interface element mapping table, which contains the unique identifier code of each product image unit in the product multimedia interface and the coordinate range of the text area of ​​the product description unit.

[0022] In this embodiment, after the RPA interaction capture module starts, it first loads the interface element mapping table from the data storage area. This interface element mapping table is obtained after the product multimedia interface is developed, through detailed element analysis and annotation of the interface. For product image units, developers assign a unique identifier code to each image. This identifier code is generated based on set rules, such as encoding according to product category, shelf order, etc., to ensure that each product image has a unique identifier. For product description units, coordinate positioning technology can be used to determine the coordinate range of its text area. This coordinate range is based on the pixel coordinate system of the interface. By determining the pixel coordinates of the upper left and lower right corners of the text area, the position of the product description on the interface can be accurately defined. During the loading process, the integrity of the data in the interface element mapping table can be checked to ensure that there is no missing or incorrect information, thus ensuring the accuracy of subsequent operations.

[0023] Step S112: Using the RPA interaction capture module, monitor the user's mouse operation events in the product multimedia interface through screen pixel scanning technology, identify the interface element type corresponding to the mouse operation event, and when the interface element type is a product image unit, record the operation trigger timestamp and operation type parameters to generate the operation trigger sequence of the product image unit.

[0024] In this embodiment, the RPA interaction capture module uses screen pixel scanning technology to monitor the multimedia interface of the product in real time. Screen pixel scanning technology relies on image sensors and image processing algorithms. The image sensor continuously scans the pixels of the interface. When the mouse moves on the interface or performs operations such as clicking or double-clicking, it can cause changes in pixel characteristics such as color and brightness. By detecting and analyzing these changes in real time, mouse operation events can be identified.

[0025] Once a mouse action event is identified, it is compared with a previously loaded interface element mapping table to determine the corresponding interface element type. This comparison is achieved by matching the pixel coordinates of the action location with the coordinate range of each element in the mapping table. If the interface element type is determined to be a product image unit, the timestamp and action type parameters are immediately recorded. The timestamp is obtained using the system's high-precision clock, accurate to a specific moment, accurately reflecting the sequence of actions. The action type parameters cover various mouse actions, such as single click, double click, right click, and drag, allowing for accurate determination of the action type based on its characteristics and patterns. By continuously recording this information, the action events can be arranged chronologically to generate the action trigger sequence for the product image unit. This sequence effectively displays the user's actions on the product image and the time intervals between them.

[0026] Step S113: Synchronously start the screen eye tracking component to monitor the movement trajectory of the user's gaze focus in the product multimedia interface. When the gaze focus falls within the coordinate range of the text area of ​​the product description unit, record the gaze entry time point, gaze exit time point, and pupil diameter change parameters during gaze dwell time. Calculate the difference between the gaze exit time point and the gaze entry time point as the attention dwell time and generate the attention dwell trajectory of the product description unit.

[0027] While the RPA interaction capture module monitors mouse operations, the screen eye-tracking component is simultaneously activated. This component employs advanced infrared reflection technology and image processing algorithms. It emits infrared light towards the user's eyes and tracks eye movements by detecting the reflected light, thereby determining the focal point of the gaze on the product's multimedia interface.

[0028] When a user's gaze enters the coordinate range of the product description unit's text area, the eye-tracking component immediately records the moment the gaze enters. This is achieved by comparing the detected gaze position with the coordinate range of the product description unit in the interface element mapping table in real time. While the user's gaze remains in this area, the component continuously monitors changes in pupil diameter. The component processes the captured eye images, analyzing the size and shape of the pupils in the images to obtain pupil diameter measurements. These measurements are recorded chronologically, forming a sequence of pupil diameter changes. When the user's gaze leaves the product description unit's text area, the component records the moment the gaze leaves.

[0029] By calculating the difference between the time the gaze leaves and the time the gaze enters, the duration of the user's attention on each product description unit can be obtained. Organizing and recording these attention durations, along with the corresponding product description unit identifiers and pupil diameter change sequences, allows the generation of an attention trajectory for each product description unit. This trajectory reflects the user's attention to different product descriptions and its temporal distribution.

[0030] Step S114: Align the operation triggering sequence of the product image unit with the attention lingering trajectory of the product description unit on the time axis so that operation events and gaze events under the same timestamp are associated and recorded.

[0031] To gain a comprehensive understanding of user interactions, it is necessary to correlate the timing of operations triggered by product image units with the attention trajectories of product description units over time. Timeline alignment is a crucial step in achieving this correlation.

[0032] First, the timestamps in the operation trigger sequence and attention lingering trajectory can be sorted. The sorting process is based on chronological order, ensuring that operation events and gaze events are arranged in chronological sequence. Then, operation events and gaze events at the same timestamp can be matched and associated based on the timestamp. Specifically, the operation trigger sequence and attention lingering trajectory can be traversed, and when operation events and gaze events have the same timestamp, they can be combined into an associated record. For example, if at a specific timestamp, a user clicks on a product image unit while their gaze is focused on the corresponding description unit, these two events can be associated to form an associated record containing information such as operation type, operation object, gaze position, and attention lingering duration. This method effectively demonstrates the user's comprehensive interaction with the product image and description at the same time.

[0033] Step S115: Perform noise filtering on the timeline-aligned associated records to remove invalid operation events caused by accidental touches and brief pause events caused by gaze drift, retaining valid associated records that reflect the user's true interaction intent, and finally forming the current interaction behavior flow that includes the user's operation trigger sequence for the product image unit and the attention pause trajectory for the product description unit.

[0034] The linked records after timeline alignment may contain some noisy data due to errors or unstable visual perception. To ensure data quality and validity, these linked records need to undergo noise filtering.

[0035] For invalid operation events caused by accidental touches, the determination can be made based on characteristics such as the duration and continuity of the operation. If the duration of an operation is too short, for example, the duration of a single click operation is much shorter than the average time of a normal operation, or if there is a lack of logical continuity between the operation and the preceding and following operations, such as a series of unrelated click operations performed in a short period of time, then the operation is likely caused by an accidental touch, and these operation events that meet the characteristics of accidental touches can be removed from the associated records.

[0036] For brief attention pauses caused by gaze drift, the severity can be assessed based on parameters such as attention duration and pupil diameter changes. If the attention duration is too short and there is no significant change in pupil diameter, it indicates that the user may not have been truly paying attention to the product description, and the pause is likely caused by gaze drift. A threshold can be set for both attention duration and pupil diameter change. When the attention duration and pupil diameter change of a particular pause are both less than the threshold, the pause can be filtered out.

[0037] After noise filtering, valid records that reflect the user's true interaction intent can be retained. These valid records constitute the current interaction flow, which accurately reflects the user's actual operations and attention distribution on the product multimedia interface.

[0038] Step S120: Extract the intent evolution trajectory of the current interaction behavior flow based on a preset historical interaction pattern library. The intent evolution trajectory includes the dynamic shift features of user interest preferences and the trend of feature intensity changes.

[0039] After obtaining the user's current interaction flow, it is necessary to further mine the user's interest and preference information contained within. The preset historical interaction pattern library stores a large amount of past user interaction data and corresponding interest intent tags. By comparing and analyzing the current interaction flow with historical patterns, dynamic changes in user interest and preferences can be extracted.

[0040] Step S121: Analyze the operation triggering sequence in the current interaction behavior flow, extract the time interval parameter and operation type sequence of continuous operation events, and convert the operation type sequence into the corresponding operation feature vector. The dimension of the operation feature vector is consistent with the preset operation dimension in the historical interaction mode library.

[0041] In this embodiment, the operation trigger sequence in the current interaction flow can be analyzed in detail. The operation trigger sequence contains a series of user operation information on the product image unit. First, the time interval parameters of consecutive operation events can be extracted. Specifically, the operation trigger sequence is traversed, and the time difference between two adjacent operation events is calculated sequentially. These time differences constitute a time interval parameter sequence. By analyzing this time interval parameter sequence, the rhythm and frequency of user operations can be determined. For example, if the time interval is short, it indicates that the user operates more frequently and may have a high interest in the product; conversely, if the time interval is long, it may indicate that the user operates more cautiously or has a lower level of interest.

[0042] Next, the operation type sequence can be extracted. The operation type sequence records the order in which the user performed operations during the process. The operation type of each operation event can be arranged chronologically to form an operation type sequence. Operation types include click, double-click, right-click, drag and drop, etc.

[0043] To facilitate subsequent analysis and comparison, the operation type sequence needs to be converted into corresponding operation feature vectors. During this conversion, the preset operation dimensions in the historical interaction pattern library can be referenced. These operation dimensions are determined based on extensive historical data and business requirements, covering various possible operation types and features. Each operation type in the sequence can be mapped to a preset operation dimension, converting it into a corresponding vector representation. For example, if the preset operation dimensions include click, double-click, right-click, and drag operations, then each operation type in the sequence can be converted into a vector element corresponding to these dimensions. For the "click" operation, the element corresponding to the "click" dimension in the vector will be set to a specific value to indicate the existence of the operation, while elements of other dimensions will be set according to specific mapping rules. Finally, the vector elements converted from all operation types are combined to generate an operation feature vector, whose dimensions are consistent with the preset operation dimensions in the historical interaction pattern library, thus ensuring comparability in subsequent comparisons and analyses.

[0044] Step S1211: Traverse the operation triggering sequence in the current interactive behavior flow, identify the triggering timestamps of two adjacent operation events, calculate the difference between the triggering timestamp of the later operation event and the triggering timestamp of the previous operation event, obtain the time interval parameter of the continuous operation events, and arrange all time interval parameters in chronological order to form a time interval sequence.

[0045] In this embodiment, the operation triggering sequence in the current interaction flow can be traversed one by one. During the traversal, the trigger timestamps of two adjacent operation events can be accurately identified. These trigger timestamps are recorded by the system's high-precision clock, ensuring high accuracy. Then, the difference between the trigger timestamp of the subsequent operation event and the trigger timestamp of the preceding operation event can be calculated; this difference is the time interval parameter for consecutive operation events. All calculated time interval parameters are arranged according to the chronological order of the operation events to generate a time interval sequence. This time interval sequence can intuitively display the temporal rhythm and frequency changes of user operations.

[0046] Step S1212: Extract the operation type parameter of each operation event in the operation trigger sequence. The operation type parameter includes click operation, double click operation, right click operation and drag operation. Map each operation type parameter to a preset type code value and arrange the type code values ​​in the time order of the operation events to form an operation type sequence.

[0047] In this embodiment, the operation type parameter for each operation event can be extracted from the operation trigger sequence. The operation type parameter clarifies the specific type of operation performed by the user, including click, double-click, right-click, and drag operations. For ease of processing and analysis, these operation type parameters can be mapped to preset type code values. These type code values ​​are predefined, with each operation type corresponding to a unique code. For example, a click operation might correspond to code value 1, a double-click operation to code value 2, a right-click operation to code value 3, and a drag operation to code value 4. The type code values ​​corresponding to each operation event can be arranged according to the chronological order of the operation events to form an operation type sequence. This operation type sequence can effectively demonstrate the changing order of user operation types.

[0048] Step S1213: Standardize the time interval sequence to generate a standardized time interval sequence.

[0049] To eliminate differences in units and scales within a time interval series, standardization can be performed. The purpose of standardization is to transform the data in the time interval series into data with the same scale and distribution for subsequent analysis and comparison. A pre-defined standardization method can be used, such as a standardization method based on the mean and standard deviation. First, the mean and standard deviation of the time interval series are calculated. Then, the mean is subtracted from each time interval parameter in the series, and the result is divided by the standard deviation to obtain the standardized time interval values. Arranging all the standardized time interval values ​​in their original order generates the standardized time interval series.

[0050] Step S1214: Perform one-hot encoding on the operation type sequence, convert each type encoding value into a binary vector of equal length, so that different operation types have orthogonality in the vector space, and generate a one-hot encoded operation sequence.

[0051] To convert the sequence of operation types into a vector form suitable for machine learning models, one-hot encoding can be performed. One-hot encoding is a commonly used encoding method that converts each type's code value into a binary vector of equal length. The length of the vector equals the number of preset operation types, and each vector contains only one element that is 1, with the rest being 0. The position of the 1 corresponds to the code for that operation type. For example, if there are four operation types (click, double-click, right-click, and drag), then the one-hot encoded vector for the type code value 1 corresponding to the click operation is [1, 0, 0, 0]; and the one-hot encoded vector for the type code value 2 corresponding to the double-click operation is [0, 1, 0, 0]. By performing one-hot encoding on each type's code value in the operation type sequence and then arranging these one-hot encoded vectors in chronological order of the operation events, a one-hot encoded operation sequence is generated. This encoding method ensures orthogonality between different operation types in the vector space, facilitating model differentiation and analysis.

[0052] Step S1215: Perform feature fusion processing on the standardized time interval sequence and the one-hot encoded operation sequence, and concatenate the standardized time interval value and the corresponding one-hot encoded vector into a high-dimensional vector according to the timestamp order to generate an operation feature vector that is consistent with the preset operation dimension in the historical interaction mode library.

[0053] In this embodiment, the standardized time interval sequence and the one-hot encoded operation sequence can be fused for feature processing. Specifically, according to the timestamp order of the operation events, each standardized time interval value in the standardized time interval sequence is concatenated with its corresponding one-hot encoded vector. For example, for a given operation event, its standardized time interval value is a numerical value, and its corresponding one-hot encoded vector is a binary vector. This numerical value and the binary vector are arranged sequentially to form a higher-dimensional vector. Arranging all the high-dimensional vectors corresponding to all operation events in chronological order generates the operation feature vector. The dimension of this operation feature vector is consistent with the preset operation dimension in the historical interaction pattern library, ensuring compatibility when comparing and analyzing with historical patterns later.

[0054] Step S122: Analyze the attention dwell trajectory in the current interactive behavior flow, extract the attention dwell time and pupil diameter change parameters of each product description unit, calculate the variance value of the pupil diameter change parameter as an indicator of attention concentration, and merge the attention dwell time and attention concentration indicator into an attention feature vector.

[0055] In this embodiment, the attention lingering trajectory in the current interaction flow can be analyzed in detail. The attention lingering trajectory records the user's attention to product description units. By analyzing this trajectory, it is possible to understand the user's level of interest and attention concentration on different product descriptions.

[0056] First, the attention dwell time for each product description unit can be extracted. Specifically, the attention dwell trajectory is traversed, and for each product description unit, the corresponding gaze entry time and gaze exit time are found, and the difference between the two is calculated to obtain the attention dwell time for that product description unit.

[0057] Next, the pupil diameter change parameters of each product description unit during attention lingering can be extracted. When a user focuses on a product description, the pupil diameter changes with the degree of attention concentration. The pupil diameter measurements of each product description unit during attention lingering can be collected; these measurements are recorded in real time by the eye-tracking component during the user's gaze lingering.

[0058] To measure the user's level of attention focus on product description units, the variance of the pupil diameter variation parameter can be calculated. The variance reflects the degree of fluctuation in pupil diameter during the period of attention retention; the smaller the fluctuation, the more focused the user's attention; conversely, the larger the fluctuation, the more scattered the user's attention may be. The pupil diameter measurements for each product description unit can be processed to calculate its variance, which can then be used as an indicator of attention focus.

[0059] Finally, the attention dwell time and corresponding attention concentration index of each product description unit can be combined into an attention feature vector. Specifically, the attention dwell time and attention concentration index of each product description unit can be arranged sequentially to form a vector. This vector integrates information on the user's attention time and attention concentration level towards the product description.

[0060] Step S1221: Traverse the attention dwell trajectory in the current interactive behavior flow, identify the gaze entry time and gaze exit time corresponding to each product description unit, calculate the difference between the gaze exit time and the gaze entry time, and obtain the attention dwell time of each product description unit.

[0061] In this embodiment, the attention lingering trajectory in the current interaction flow can be traversed one by one. During the traversal, the corresponding gaze entry time and gaze exit time can be accurately identified for each product description unit. These time points are recorded by the eye-tracking component when monitoring the user's gaze movement, and have high accuracy. Then, the difference between the gaze exit time and the gaze entry time can be calculated, and this difference is the attention lingering duration for that product description unit. By recording the attention lingering durations of all product description units in the order of the product description units, the distribution of user attention time for different product descriptions can be effectively obtained.

[0062] Step S1222: Extract pupil diameter sampling data for each product description unit in the attention dwell trajectory during the attention dwell period. The pupil diameter sampling data includes multiple pupil diameter measurements collected at fixed time intervals. Arrange the pupil diameter measurements of the same product description unit in the order of sampling time to form a pupil diameter sequence.

[0063] The attention dwell trajectory includes pupil diameter sampling data for each product description unit during the attention dwell period. This sampling data is obtained by the eye-tracking component measuring the user's pupil diameter at fixed time intervals. The pupil diameter sampling data for each product description unit can be extracted from the attention dwell trajectory, and then all pupil diameter measurements for the same product description unit are arranged in chronological order of sampling time to form a pupil diameter sequence. This pupil diameter sequence reflects the dynamic changes in pupil diameter when the user focuses on a particular product description.

[0064] Step S1223: Perform outlier detection processing on the pupil diameter sequence, identify and remove abnormal measurement values ​​that exceed the mean plus or minus K times the standard deviation, and retain the normal pupil diameter sequence.

[0065] To ensure the accuracy and reliability of the data, outlier detection processing can be performed on the pupil diameter sequence. Outliers may be caused by measurement errors, sudden blinking by the user, or other interfering factors, which can affect the subsequent calculation of attention concentration. An outlier detection method based on the mean and standard deviation can be used. First, the mean and standard deviation of the pupil diameter sequence are calculated, and then a threshold range is determined, which is the mean plus or minus K times the standard deviation (K is a pre-set coefficient). Each measurement value in the pupil diameter sequence is compared with this threshold range. If a measurement value exceeds the threshold range, it is considered an outlier and can be removed from the pupil diameter sequence. After removing outliers, the remaining pupil diameter sequence is the normal pupil diameter sequence, which more accurately reflects the true changes in pupil diameter when the user focuses on the product description.

[0066] Step S1224: Calculate the variance of all measurements in the normal pupil diameter sequence. The variance reflects the degree of fluctuation of pupil diameter during attention stagnation. Use this variance as an indicator of the degree of attention concentration of the user on the product description unit.

[0067] In this embodiment, the normal pupil diameter sequence can be further analyzed to calculate the variance of all measurements. Variance is a statistic that reflects the dispersion of data. In this scenario, the variance reflects the degree of fluctuation in pupil diameter during attention stagnation. A smaller variance indicates that the pupil diameter does not change significantly during attention stagnation, suggesting that the user's attention is relatively focused; a larger variance indicates greater fluctuation in pupil diameter, suggesting that the user's attention may be more scattered. The calculated variance can be used as an indicator of the user's attention concentration on the product description unit.

[0068] Step S1225: Associate and store the attention dwell time of each product description unit with the corresponding attention concentration index to form an attention feature record containing the product description unit identifier, attention dwell time, and attention concentration index.

[0069] In this embodiment, the attention dwell time of each product description unit can be associated and stored with the corresponding attention concentration index. Specifically, each product description unit is assigned a unique identifier, and then the attention dwell time, attention concentration index, and product description unit identifier are combined to form an attention feature record. All attention feature records of product description units are arranged in a certain order to generate an attention feature record set containing information on all product description units.

[0070] Step S123: Perform feature concatenation processing on the operation feature vector and the attention feature vector to generate the current interaction feature matrix. The row dimension of the current interaction feature matrix corresponds to the timestamp sequence, and the column dimension corresponds to the joint dimension of the operation feature and the attention feature.

[0071] After obtaining the operation feature vector and attention feature vector, feature concatenation can be performed to generate a more comprehensive feature representation. Specifically, the operation feature vector and attention feature vector can be aligned according to timestamps. Since both operation and attention feature vectors are generated based on the temporal order of operation events and gaze events, they can be mapped one-to-one according to timestamps.

[0072] Then, the operation feature vectors and attention feature vectors can be concatenated along the column direction. That is, for each timestamp, the elements of the operation feature vector and attention feature vector can be arranged sequentially to form a longer vector. Arranging all the concatenated vectors corresponding to all timestamps by row constitutes the current interaction feature matrix.

[0073] The row dimension of this current interaction feature matrix corresponds to a timestamp sequence, meaning each row represents a user interaction feature at a specific point in time. The column dimension corresponds to a joint dimension of operation and attention features, which includes various feature information such as operation type, time interval, attention duration, and attention concentration level. In this way, the current interaction feature matrix comprehensively reflects the user's operation and attention status at different points in time.

[0074] Step S124: Load a preset historical interaction pattern library. The historical interaction pattern library contains multiple historical interaction feature matrices and their corresponding intent tag sets. Calculate the similarity value between the current interaction feature matrix and each historical interaction feature matrix using a cosine similarity algorithm, and select the historical interaction feature matrix with the highest similarity value as the reference pattern matrix.

[0075] In this embodiment, a pre-defined historical interaction pattern library can be loaded from a dedicated data storage area. This historical interaction pattern library is constructed through long-term data collection and organization, storing a large amount of past user interaction behavior data. It contains multiple historical interaction feature matrices, each corresponding to a set of users' historical interaction behavior features. Simultaneously, each historical interaction feature matrix is ​​also associated with a set of intent labels, which describe the user's interest intent during the historical interaction, such as a preference for a certain type of product or attention to specific product attributes.

[0076] To find the historical interaction pattern most similar to the current interaction feature matrix, the cosine similarity algorithm can be used. The cosine similarity algorithm is a commonly used method for calculating vector similarity; it measures the similarity between two vectors by calculating the cosine of the angle between them. In this scenario, the current interaction feature matrix and each historical interaction feature matrix in the historical interaction pattern library are treated as high-dimensional vectors, and the cosine similarity value between the current interaction feature matrix and each historical interaction feature matrix is ​​calculated.

[0077] The specific calculation process is as follows: First, the current interaction feature matrix and the historical interaction feature matrix are represented as vectors. Then, the dot product of the current interaction feature matrix and the historical interaction feature matrix is ​​calculated. Finally, this product is divided by the product of the magnitudes of the current interaction feature matrix and the historical interaction feature matrix to obtain the cosine similarity value. The closer the cosine similarity value is to 1, the more similar the current interaction feature matrix and the historical interaction feature matrix are; the closer it is to 0, the less similar the current interaction feature matrix and the historical interaction feature matrix are.

[0078] In this embodiment, all the calculated similarity values ​​can be compared, and the historical interaction feature matrix with the highest similarity value can be selected as the reference pattern matrix.

[0079] Step S125: Extract the set of intent tags corresponding to the reference pattern matrix and the rate of change of feature values ​​of each timestamp in the current interaction feature matrix. Calculate the feature offset of the current interaction feature matrix relative to the reference pattern matrix using a time series difference algorithm. Determine the dynamic steering feature of user interest preferences based on the feature offset. The dynamic steering feature is represented by a steering direction vector and a steering amplitude parameter.

[0080] In this embodiment, the corresponding intent tag set can be extracted from the reference pattern matrix. This intent tag set describes the user's interest intent under the reference pattern, such as a liking for a certain type of product or an interest in certain product attributes. Simultaneously, the rate of change of feature values ​​at each timestamp in the current interaction feature matrix can be analyzed. The rate of change of feature values ​​reflects the changes in the user's interactive behavior characteristics at different points in time; by calculating the rate of change of feature values, the dynamic trend of user interest can be determined.

[0081] In this embodiment, a time-series differencing algorithm can be used to calculate the feature offset of the current interaction feature matrix relative to the reference pattern matrix. The time-series differencing algorithm is a method for analyzing changes in time-series data; it reflects data changes by calculating the difference between data points at adjacent time points. In this scenario, the current interaction feature matrix and the reference pattern matrix can be compared along the same feature dimensions, and the difference between the current interaction feature matrix and the reference pattern matrix in each feature dimension can be calculated to obtain the feature offset.

[0082] Based on the calculated feature offsets, the dynamic shift characteristics of user interests and preferences can be determined. These dynamic shift characteristics are represented by a shift direction vector and a shift magnitude parameter. The shift direction vector indicates the direction of change in user interests and preferences; for example, a shift from focusing on one type of product to another, or an increase in attention to a particular attribute of a product. The shift magnitude parameter indicates the degree of change in interests and preferences; for example, whether it is a small or large change. By analyzing the shift direction vector and the shift magnitude parameter, the dynamic changes in user interests and preferences can be effectively understood.

[0083] Step S126: Construct a feature intensity decay function based on dynamic steering features. The feature intensity decay function uses timestamp as the independent variable and the cumulative sum of feature offsets as the dependent variable to calculate the feature intensity value at different timestamps, thereby generating a feature intensity change trend that reflects the change law of feature intensity over time.

[0084] In this embodiment, a feature intensity decay function can be constructed based on dynamic steering features. The purpose of this feature intensity decay function is to describe the change in the feature intensity of user interest preferences over time. The feature intensity decay function uses timestamps as independent variables and the cumulative sum of feature offsets as dependent variables.

[0085] The cumulative sum of feature offsets reflects the overall change in user interests and preferences from the beginning to the current time point. Over time, user interests may gradually shift, and the strength of early interest and preference features will gradually decay. The feature strength decay function adjusts the cumulative sum of feature offsets according to changes in timestamps, calculating the feature strength values ​​at different timestamps.

[0086] Specifically, the feature intensity decay function adjusts the cumulative sum of feature offsets according to a set decay rule. For example, it might use exponential decay, where the feature intensity value decreases exponentially over time. By arranging the feature intensity values ​​calculated at different timestamps in chronological order, a feature intensity trend reflecting the changing pattern of feature intensity over time can be generated. This trend can visually demonstrate how the feature intensity of user interests changes at different points in time, allowing for timely adjustments to recommendation strategies to adapt to changes in user interests.

[0087] Step S127: The dynamic steering feature and the feature intensity change trend are associated and encapsulated to form an intention evolution trajectory that includes the user's interest preferences, dynamic steering feature and feature intensity change trend.

[0088] In this embodiment, dynamic steering features and feature intensity change trends can be associated and encapsulated. Specifically, the steering direction vector and steering amplitude parameter in the dynamic steering features are combined with the feature intensity value in the feature intensity change trend according to a certain logical structure. For example, the steering direction vector, steering amplitude parameter, and feature intensity value can be stored as different attributes of a data object, and a timestamp can be added to this data object to indicate the time point corresponding to these information.

[0089] Through the aforementioned associative encapsulation process, an intent evolution trajectory is formed, encompassing dynamic shifting features of user interests and preferences, as well as the changing trends in feature intensity. This intent evolution trajectory comprehensively describes the dynamic changes in user interests and preferences, including the direction of interest shift, the magnitude of change, and the variation of feature intensity over time. Subsequently, recommendation rules and strategies that better align with the user's current interests can be generated based on this intent evolution trajectory.

[0090] Step S130: Generate a dynamic matching rule set based on the intent evolution trajectory. The dynamic matching rule set includes the association constraints between product image features and user interest preferences, as well as the priority ranking logic of product description features.

[0091] After obtaining the user's intention evolution trajectory based on the intention evolution trajectory, a dynamic matching rule set needs to be generated based on this intention evolution trajectory.

[0092] Step S131: Analyze the dynamic turning features in the intention evolution trajectory, extract the product feature dimension corresponding to the turning direction vector, the product feature dimension includes the product image feature dimension and the product description feature dimension, determine the preference polarity of each product feature dimension in the turning direction, the preference polarity includes positive preference and negative preference.

[0093] In this embodiment, the dynamic turning features in the intent evolution trajectory can be analyzed. The turning direction vector in the dynamic turning features indicates the direction of change in user interest preferences, and the corresponding product feature dimensions can be extracted from it. The product feature dimensions include product image feature dimensions and product description feature dimensions. The product image feature dimensions can include features such as color, texture, and shape; the product description feature dimensions can include features such as keywords, themes, and sentiments.

[0094] In this embodiment, the preference polarity of each product feature dimension in the turning direction can be determined based on the turning direction vector. Preference polarity is divided into positive and negative preferences. If the turning direction vector indicates an increase in user attention to a certain product feature dimension, then that feature dimension has a positive preference; if the turning direction vector indicates a decrease in user attention to a certain product feature dimension, then that feature dimension has a negative preference. For example, if the turning direction vector indicates an increase in user attention to a certain color in the color features of a product image, then the color feature dimension has a positive preference for that color; if it indicates a decrease in user attention to certain keywords in the product description, then the keyword feature dimension has a negative preference for those keywords.

[0095] Step S132: Based on the feature intensity value in the trend of preference polarity and feature intensity change, assign a preference weight coefficient to each product feature dimension. The higher the feature intensity value of the product feature dimension, the higher the preference weight coefficient. Generate a product feature dimension weight table.

[0096] In this embodiment, a preference weight coefficient can be assigned to each product feature dimension based on the feature intensity value in the trend of preference polarity and feature intensity change. The preference weight coefficient is used to represent the degree of importance that user attaches to different product feature dimensions. The higher the feature intensity value, the stronger the user's interest in that product feature dimension, and the higher the corresponding preference weight coefficient.

[0097] In this embodiment, all product feature dimensions can be traversed, and the corresponding preference weight coefficients can be determined based on their preference polarity and feature strength values. For product feature dimensions with positive preferences, a higher positive weight coefficient can be assigned based on their feature strength value; for product feature dimensions with negative preferences, a lower positive or negative weight coefficient can be assigned (the specific weight coefficient is determined based on business needs and algorithm design). Each product feature dimension and its corresponding preference weight coefficient are recorded to form a product feature dimension weight table. This product feature dimension weight table makes recommendations more aligned with user interests and preferences.

[0098] Step S133: For the product image feature dimension, select image feature sub-dimensions with weight coefficients higher than the preset threshold according to the product feature dimension weight table. The image feature sub-dimensions include color feature sub-dimensions, texture feature sub-dimensions and shape feature sub-dimensions. Set preference range constraints for each image feature sub-dimension to generate the association constraint conditions between product image features and user interest preferences.

[0099] Step S1331: Traverse the product image feature dimensions in the product feature dimension weight table, extract the weight coefficients corresponding to each image feature sub-dimension, compare the weight coefficients with the preset threshold, and select the image feature sub-dimensions with weight coefficients higher than the preset threshold as key image feature sub-dimensions.

[0100] In this embodiment, the product image feature dimensions in the product feature dimension weight table can be traversed. The product image feature dimensions include multiple image feature sub-dimensions, such as color, texture, and shape sub-dimensions. The preference weight coefficients corresponding to each image feature sub-dimension can be extracted, and then these weight coefficients are compared with a preset threshold. The preset threshold is a critical value set based on business needs and experience, used to filter out image feature sub-dimensions with high user attention. If the weight coefficient of a certain image feature sub-dimension is higher than the preset threshold, it indicates that the user has a strong interest in that sub-dimension, and it can be selected as a key image feature sub-dimension. For example, if the weight coefficient of the color feature sub-dimension is higher than the threshold, then the color feature sub-dimension will be determined as a key image feature sub-dimension.

[0101] Step S1332: For each key image feature sub-dimension, analyze the range of feature values ​​with the highest user interaction frequency under that key image feature sub-dimension in the historical interaction pattern library. The feature value range is obtained by statistical analysis of the feature value distribution in historical interaction data, and includes the lower limit and upper limit of the feature value.

[0102] For each selected key image feature sub-dimension, analysis can be performed in the historical interaction pattern library. This library stores a large amount of past user interaction data, allowing us to statistically analyze the range of feature values ​​with the highest user interaction frequency for that key image feature sub-dimension. Specifically, we perform distribution statistics on the feature values ​​of that key image feature sub-dimension in the historical interaction data. By analyzing the distribution of feature values, we determine a range that includes most of the high-frequency interaction feature values. This range includes both the lower and upper limits of the feature values. For example, for the color feature sub-dimension, we can statistically analyze the range of color values ​​with the highest historical user interaction frequency, determining the lower and upper limits of the color values. This range of color values ​​can represent the color intervals that users are most interested in.

[0103] Step S1333: Combining the dynamic steering features in the intention evolution trajectory, adjust the lower and upper limits of the feature value range. When the steering direction vector of the dynamic steering feature points to the direction of increasing feature value, increase the upper limit and keep the lower limit unchanged; when the steering direction vector points to the direction of decreasing feature value, decrease the lower limit and keep the upper limit unchanged.

[0104] In this embodiment, the feature value range of the key image feature sub-dimension can be adjusted by combining dynamic turning features in the intent evolution trajectory. The turning direction vector in the dynamic turning feature indicates the direction of change in user interest preferences. If the turning direction vector points in the direction of increasing feature value, it indicates that the user's interest in larger feature values ​​in that key image feature sub-dimension is increasing. Therefore, the upper limit of the feature value range can be increased while keeping the lower limit unchanged, thereby expanding the range of feature values ​​that the user might be interested in. For example, for the color feature sub-dimension, if the turning direction vector indicates that the user's interest in brighter colors is increasing, the upper limit of the color brightness feature value range can be increased.

[0105] If the steering direction vector points in the direction of decreasing feature values, it indicates that the user's interest in smaller feature values ​​within that key image feature sub-dimension is increasing. In this case, the lower bound of the feature value range can be lowered while keeping the upper bound unchanged. For example, for the shape feature sub-dimension, if the steering direction vector indicates increased user interest in smaller shapes, the lower bound of the shape size feature value range can be lowered. Through these adjustments, the feature value range can be made more aligned with the user's current interests and preferences.

[0106] Step S1334: Set constraint type identifiers for the adjusted feature value range. The constraint type identifiers include mandatory constraints and priority constraints. The key image feature sub-dimensions with the highest weight coefficients correspond to mandatory constraints, while the other key image feature sub-dimensions correspond to priority constraints.

[0107] In this embodiment, constraint type identifiers can be set for the adjusted feature value range. Constraint type identifiers are divided into mandatory constraints and priority constraints. They can be ranked according to the weight coefficients of the key image feature sub-dimensions, and the top-ranked key image feature sub-dimensions by a preset proportion of weight coefficients are selected, with mandatory constraints set for their corresponding feature value ranges. This means that when filtering products, the feature values ​​of these key image feature sub-dimensions must fall within the adjusted feature value range; otherwise, the product will be excluded.

[0108] For the remaining key image feature sub-dimensions, set priority constraints for their corresponding feature value ranges. This means that when filtering products, these feature values ​​will be given priority if they fall within the adjusted feature value range, but are not required to be met. For example, if a product meets the requirements for all key image feature sub-dimensions that must meet the constraints, but only partially meets the requirements for the key image feature sub-dimensions that prioritize the constraints, the product may still be selected. Setting constraint type identifiers allows for more flexible filtering of products that match user interests and preferences.

[0109] Step S1335: Combine the key image feature sub-dimensions, the adjusted feature value range, and the constraint type identifier into constraint condition statements. All constraint condition statements together constitute the association constraint conditions between product image features and user interest preferences.

[0110] In this embodiment, key image feature sub-dimensions, adjusted feature value ranges, and constraint type identifiers can be combined into constraint condition statements. For example, for the color feature sub-dimension, the constraint condition statement might be "The feature values ​​of the color feature sub-dimension must be within the range of [lower limit, upper limit] (must satisfy the constraint)". Summarizing the constraint condition statements corresponding to all key image feature sub-dimensions together constitutes the association constraint conditions between product image features and user interests and preferences. These association constraint conditions can be used to subsequently filter the image features of candidate products, ensuring that the filtered product image features match the user's interests and preferences.

[0111] Step S134: For the product description feature dimension, sort the description feature sub-dimensions according to the product feature dimension weight table. The description feature sub-dimensions include keyword feature sub-dimensions, topic feature sub-dimensions and sentiment feature sub-dimensions. Determine the importance order of each description feature sub-dimension according to the weight coefficient from high to low, and generate the priority sorting logic of product description features.

[0112] Step S1341: Extract all descriptive feature sub-dimensions under the product description feature dimension and their corresponding weight coefficients from the product feature dimension weight table, and construct a target comparison table.

[0113] In this embodiment, all descriptive feature sub-dimensions under the product description feature dimension and their corresponding preference weight coefficients can be extracted from the product feature dimension weight table. The descriptive feature sub-dimensions include keyword feature sub-dimensions, topic feature sub-dimensions, and sentiment feature sub-dimensions, etc. A target lookup table is constructed by mapping each descriptive feature sub-dimension to its corresponding weight coefficient. This target lookup table can effectively demonstrate the importance of each descriptive feature sub-dimension.

[0114] Step S1342: Sort the descriptive feature sub-dimensions in the target lookup table from high to low according to their weight coefficients to obtain the descriptive feature sub-dimension sorting sequence.

[0115] In this embodiment, the descriptive feature sub-dimensions in the target lookup table can be sorted from high to low according to their weight coefficients. The sorting process can employ common sorting algorithms, such as bubble sort or quicksort. By sorting, descriptive feature sub-dimensions with high weight coefficients are placed first, and those with low weight coefficients are placed last, resulting in a sorted sequence of descriptive feature sub-dimensions. This sorted sequence of descriptive feature sub-dimensions can intuitively reflect the order of importance of each descriptive feature sub-dimension.

[0116] Step S1343: Assign a priority level to each descriptive feature sub-dimension in the sorted sequence. The first descriptive feature sub-dimension in the sorted sequence corresponds to the highest priority level, and the priority levels of subsequent descriptive feature sub-dimensions decrease sequentially.

[0117] In this embodiment, a priority level can be assigned to each descriptive feature sub-dimension in the sorting sequence. The priority level indicates the importance of each descriptive feature sub-dimension in the matching process. The descriptive feature sub-dimension ranked first in the sorting sequence has the highest priority level, and subsequent descriptive feature sub-dimensions have progressively lower priority levels according to their sorting order. For example, if the keyword feature sub-dimension is ranked first in the sorting sequence, it will be assigned the highest priority level; the topic feature sub-dimension is ranked second, with the next highest priority level; and the sentiment feature sub-dimension is ranked third, with the lowest priority level. Assigning priority levels clearly defines the order and importance of each descriptive feature sub-dimension in the matching process.

[0118] Step S1344: For each priority level of the descriptive feature sub-dimension, set the priority rules for feature matching. The higher the priority level of the descriptive feature sub-dimension, the more it will be evaluated in the matching process, and its matching result will contribute more to the overall matching degree than the descriptive feature sub-dimension with lower priority.

[0119] In this embodiment, priority rules for feature matching can be set for each priority level of descriptive feature sub-dimensions. These rules specify the evaluation order of each descriptive feature sub-dimension and its contribution weight to the overall matching degree when matching product description features. Higher priority descriptive feature sub-dimensions are evaluated first during the matching process. For example, when matching product descriptions, the highest priority descriptive feature sub-dimensions can be matched first. If this sub-dimension matches successfully, it can be given a higher weight in the overall matching degree; if the match fails, the next lower priority descriptive feature sub-dimensions are then matched, and their matching results contribute a relatively lower weight to the overall matching degree. By setting the above priority rules, it can be ensured that descriptive features that better match user interests are given priority when matching product descriptions, thus improving the accuracy of recommendations.

[0120] Step S1345: Combine the sorting sequence of the description feature sub-dimensions, the priority level, and the feature matching priority rules into a sorting logic statement to form the priority sorting logic of the product description features.

[0121] In this embodiment, the order sequence of descriptive feature sub-dimensions, priority levels, and feature matching priority rules can be combined into a sorting logic statement. This sorting logic statement can be expressed in natural language or a specific rule-based language, clearly defining the evaluation order, priority level, and contribution weight of each descriptive feature sub-dimension to the overall matching degree when matching product description features. For example, the sorting logic statement might be: "First, evaluate the keyword feature sub-dimension (highest priority level), whose matching result contributes X to the overall matching degree; if the keyword feature sub-dimension fails to match, then evaluate the topic feature sub-dimension (secondary priority level), whose matching result contributes Y to the overall matching degree; finally, evaluate the sentiment feature sub-dimension (lowest priority level), whose matching result contributes Z to the overall matching degree." Combining all relevant information into the above sorting logic statement generates the priority sorting logic for product description features.

[0122] Step S135: Perform rule formatting processing on the association constraints between the product image features and user interest preferences and the priority ranking logic of the product description features. Use a preset rule syntax to convert the constraints and ranking logic into machine-parseable rule statements, and merge all rule statements to form a dynamic matching rule set.

[0123] In this embodiment, the association constraints between product image features and user interest preferences, as well as the priority ranking logic of product description features, can be formatted using rules. The preset rule syntax is a predefined language specification used to convert rules described in natural language into rule statements that can be parsed and executed by machines.

[0124] For the associated constraints of product image features, each constraint statement can be transformed according to the rule syntax so that it can be understood and processed by machines. For example, "The feature values ​​of the color feature sub-dimension must be within the range of [lower limit, upper limit] (the constraint must be satisfied)" can be transformed into a statement that conforms to the rule syntax.

[0125] The same transformation is applied to the priority ranking logic of product description features, converting the ranking logic statements into machine-parseable rules. For example, the ranking logic statements regarding the evaluation order and contribution weight of the description feature sub-dimensions are converted into a rule syntax format.

[0126] Finally, all the transformed rule statements can be merged together to form a dynamic matching rule set. This dynamic matching rule set includes filtering rules for product image features and ranking rules for product description features. It can then be input into a recommendation model to filter and rank candidate products, thereby achieving more accurate product recommendations.

[0127] Step S140: Input the dynamic matching rule set into the pre-trained AI recommendation model, perform rule matching filtering on the candidate product multimedia content set, and generate product multimedia content filtering results containing matching degree scores.

[0128] After generating the dynamic matching rule set, it needs to be input into the pre-trained AI recommendation model to filter and score the candidate product multimedia content set in order to obtain the product multimedia content selection results that match the user's interests and preferences.

[0129] Step S141: Load the candidate product multimedia content set, which contains multiple product multimedia content units, each of which contains corresponding product image feature data and product description feature data.

[0130] In this embodiment, a candidate product multimedia content set can be loaded from a dedicated data storage area. This candidate product multimedia content set is pre-collected and organized, containing multiple product multimedia content units. Each product multimedia content unit contains corresponding product image feature data and product description feature data. Product image feature data may include feature information such as the image's color, texture, and shape; product description feature data may include descriptive information such as the product's keywords, theme, and sentiment.

[0131] Step S142: Input the dynamic matching rule set into the rule parsing layer of the AI ​​recommendation model, and use the rule parsing layer to perform syntax parsing on the association constraints and priority ranking logic in the dynamic matching rule set, and convert them into feature matching operators and weight allocation parameters that the model can execute.

[0132] In this embodiment, a dynamic matching rule set can be input into the rule parsing layer of the AI ​​recommendation model. The rule parsing layer is a crucial component of the AI ​​recommendation model, its main function being to perform syntactic parsing on the input rule set. The rule parsing layer performs a detailed analysis of the product image feature association constraints and product description feature priority ranking logic within the dynamic matching rule set, converting the natural language descriptions into feature matching operators and weight allocation parameters that the model can execute.

[0133] For the association constraints of product image features, the rule parsing layer parses the constraint statements into specific feature matching operators, such as operators that determine whether feature values ​​are within a certain range. For the priority ranking logic of product description features, the evaluation order of each description feature sub-dimension and its contribution weight to the overall matching degree can be parsed, and this information is converted into weight allocation parameters. Through the above parsing and transformation, the dynamic matching rule set can be understood and executed by the AI ​​recommendation model.

[0134] Step S143: Use the feature extraction layer of the AI ​​recommendation model to perform feature parsing processing on the product image feature data of each product multimedia content unit, extract the image feature values ​​corresponding to the key image feature sub-dimension in the association constraint conditions, and perform feature parsing processing on the product description feature data to extract the description feature values ​​corresponding to the description feature sub-dimension in the priority sorting logic.

[0135] In this embodiment, the feature extraction layer of the AI ​​recommendation model performs feature parsing processing on the product image feature data and product description feature data of each product multimedia content unit. For the product image feature data, the feature extraction layer extracts the corresponding image feature values ​​based on the key image feature sub-dimensions determined in the association constraints. For example, if the association constraints include color feature sub-dimensions and shape feature sub-dimensions, the feature extraction layer will extract color and shape feature values ​​from the product image.

[0136] For product description feature data, the feature extraction layer extracts the corresponding description feature values ​​based on the description feature sub-dimensions determined in the priority ranking logic. For example, if the priority ranking logic includes keyword feature sub-dimensions and topic feature sub-dimensions, the feature extraction layer will extract keyword and topic information from the product description. Through the above feature parsing process, the raw data of the product multimedia content unit is transformed into feature values ​​that can be used for rule matching.

[0137] Step S144: Perform matching and filtering processing using the rule matching layer of the AI ​​recommendation model. Compare the image feature values ​​with the feature value range in the associated constraints. Count the number of image feature sub-dimensions that meet the mandatory constraints and the number of image feature sub-dimensions that meet the priority constraints. Evaluate the matching degree of the descriptive feature values ​​according to the priority sorting logic, assign different matching weights according to the priority level, and calculate the comprehensive matching score.

[0138] The rule matching layer of the AI ​​recommendation model performs rule matching and filtering on the extracted image feature values ​​and descriptive feature values. For image feature values, they are compared with the feature value range in the associated constraints. The rule matching layer counts the number of image feature sub-dimensions that satisfy the mandatory constraints and the number that satisfy the preferred constraints. For example, if the color feature value of a product falls within the mandatory constraint range of the color feature sub-dimension, the color feature sub-dimension can be counted as satisfying the mandatory constraint; if the shape feature value falls within the preferred constraint range of the shape feature sub-dimension, the shape feature sub-dimension can be counted as satisfying the preferred constraint.

[0139] For descriptive feature values, the rule matching layer evaluates their matching degree sequentially according to a priority ranking logic. Based on priority levels, the matching of the highest priority descriptive feature sub-dimensions is evaluated first, followed by the next highest priority descriptive feature sub-dimensions. For each descriptive feature sub-dimension's matching result, different matching weights are assigned according to its priority level. For example, a successful match of the highest priority descriptive feature sub-dimension receives a higher weight, while a successful match of the lowest priority descriptive feature sub-dimension receives a lower weight.

[0140] Finally, the rule-based matching layer integrates the matching results of image features and descriptive features to calculate a comprehensive matching score for each product's multimedia content unit. This comprehensive matching score reflects the degree to which each product's multimedia content unit matches the user's interests and preferences.

[0141] Step S145: Sort the comprehensive matching scores of all product multimedia content units in descending order, select the first preset number of product multimedia content units as the filtering results, and retain the comprehensive matching score of each product multimedia content unit as the matching degree score, and finally generate the product multimedia content filtering results containing the matching degree score.

[0142] In this embodiment, the comprehensive matching scores of all product multimedia content units can be sorted in descending order. This descending order prioritizes products with high matching scores, facilitating subsequent filtering. Then, a preset number of product multimedia content units at the top of the list can be selected as the filtering results. This preset number is a threshold set based on business needs and recommendation strategies, used to control the number of recommended products.

[0143] In this embodiment, the overall matching score of each selected product multimedia content unit can be retained as a matching degree score. The matching degree score can intuitively show the degree of matching between each product and the user's interests and preferences. The selected product multimedia content units and their corresponding matching degree scores are combined to generate product multimedia content filtering results that include the matching degree scores.

[0144] Step S150: The RPA display arrangement module performs a real-time rendering operation of the recommended sequence based on the product multimedia content filtering results. The real-time rendering operation includes adjusting the display size of the product image unit and the text layout style of the product description unit.

[0145] After obtaining the multimedia content filtering results for the products, the RPA display orchestration module needs to render the filtered products in real time to display products that match the user's interests and preferences to the user in a suitable way.

[0146] For example, step S151: parse the product multimedia content units and their corresponding matching scores in the product multimedia content filtering results, and determine the recommended sequence order of the product multimedia content units from high to low according to the matching scores.

[0147] The RPA display and arrangement module parses the filtered product multimedia content results, extracting the product multimedia content units and their corresponding matching scores. Then, it sorts the product multimedia content units according to their matching scores from highest to lowest, determining their recommendation sequence order. Product multimedia content units with high matching scores are listed first, and those with low matching scores are listed later. This sorting ensures that users first see products that best match their interests and preferences.

[0148] Step S152: Load a preset interface layout template using the RPA display orchestration module. The interface layout template includes grid division parameters for the recommended area and upper and lower limits for the size of each grid cell.

[0149] The RPA display and arrangement module loads a preset interface layout template. This template is pre-designed and defines the layout and grid division of the product recommendation area. The template includes grid division parameters for the recommendation area, such as the number of rows and columns, as well as the upper and lower limits of the size of each grid cell. The grid division parameters determine how products are arranged on the interface, while the upper and lower limits restrict the display size range of each product's multimedia content unit.

[0150] Step S153: Based on the recommended sequence order and matching score, assign a corresponding grid unit to each product multimedia content unit, calculate the actual display size of the grid unit, adjust the pixel resolution of the product image unit to fit the actual display size, and assign grid units with larger upper limits to the product multimedia content units with higher matching scores.

[0151] The RPA display layout module assigns a corresponding grid cell to each product's multimedia content unit based on the recommended sequence order and matching score. Product multimedia content units with higher matching scores are assigned to grid cells with larger upper size limits to highlight these products that are more aligned with user interests and preferences.

[0152] For each assigned grid cell, its actual display size is calculated. This actual display size is adjusted based on the specific product within the upper and lower limits of the grid cell's size. Then, the pixel resolution of the product image cell can be adjusted to fit the actual display size of the grid cell. For example, if a product's multimedia content cell is assigned a larger grid cell size, the pixel resolution of the product image cell can be increased to ensure a clear and sharp image when displayed on a large screen; if the grid cell size is smaller, the pixel resolution can be reduced accordingly to prevent the image from becoming too large and exceeding the grid's boundaries. Through this method, it is ensured that product images have a good display effect in grid cells of different sizes.

[0153] Step S154: Extract the text data of the product description unit in the product multimedia content unit, determine the text layout style parameters according to the matching score, and adjust the text line spacing and character spacing to optimize the reading experience. The product description unit with a higher matching score corresponds to a larger font size and a more eye-catching font color.

[0154] The RPA display layout module extracts the text data of product description units from the product multimedia content units. Then, it determines the text layout style parameters based on the matching score of the product multimedia content unit. Product description units with higher matching scores can have larger font sizes and more prominent font colors to attract user attention. Simultaneously, line spacing and character spacing can be adjusted to optimize the reading experience. Appropriate line spacing and character spacing make the product description text clearer and easier to read, avoiding overly crowded or sparse text. For example, for product description units with high matching scores, larger fonts and more prominent colors can be used, along with increased line spacing and character spacing to make the text more prominent and readable; for product description units with low matching scores, the font size and color are relatively less prominent, and the line spacing and character spacing are appropriately reduced.

[0155] Step S155: Generate a rendering instruction set containing grid cell coordinates, product image cell display size, and product description cell layout style parameters. Send the rendering instruction set to the interface rendering engine to drive the interface rendering engine to render the product image cells and product description cells in the corresponding grid cells according to the recommended sequence, thus completing the real-time rendering operation of the recommended sequence.

[0156] The RPA display orchestration module combines grid cell coordinates, product image cell display dimensions, and product description cell layout style parameters to generate a rendering instruction set. This rendering instruction set contains all the necessary information for displaying the product on the interface. The rendering instruction set can then be sent to the interface rendering engine.

[0157] The interface rendering engine renders product image units and product description units in the corresponding grid cells according to the recommended sequence of the product multimedia content units, based on the received rendering instruction set. During the rendering process, images can be rendered according to the display size and pixel resolution of the product image units, and text can be rendered according to the layout style parameters of the product description units. In this way, the real-time rendering operation of the recommended sequence is completed, presenting products that match the user's interests and preferences to the user with the best display effect.

[0158] Figure 2 The illustration shows exemplary hardware and software components of a multimedia product recommendation system 100 combining RPA and AI, which can implement the ideas of this application, according to some embodiments of this application. For example, processor 120 can be used in the multimedia product recommendation system 100 combining RPA and AI and to perform the functions in this application.

[0159] The product multimedia recommendation system 100 combining RPA and AI can be a general-purpose server or a special-purpose server; both can be used to implement the product multimedia recommendation method combining RPA and AI of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0160] For example, a multimedia product recommendation system 100 combining RPA and AI may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the multimedia product recommendation system 100 combining RPA and AI may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The multimedia product recommendation system 100 combining RPA and AI also includes an I / O interface 150 between the computer and other input / output devices.

[0161] For ease of explanation, only one processor is described in the product multimedia recommendation system 100 combining RPA and AI. However, it should be noted that the product multimedia recommendation system 100 combining RPA and AI in this application may also include multiple processors. Therefore, the steps performed by one processor as described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the product multimedia recommendation system 100 combining RPA and AI performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0162] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned product multimedia recommendation method combining RPA and AI is implemented.

[0163] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A multimedia product recommendation method combining RPA and AI, characterized in that, The method includes: The RPA interaction capture module records the current interaction flow between the user and the product multimedia interface in real time. The current interaction flow includes the timing of the user's operation on the product image unit and the trajectory of the user's attention on the product description unit. Based on a preset historical interaction pattern library, the intent evolution trajectory of the current interaction behavior flow is extracted. The intent evolution trajectory includes the dynamic shift features of user interest preferences and the trend of feature intensity change. A dynamic matching rule set is generated based on the intent evolution trajectory. The dynamic matching rule set includes the association constraints between product image features and user interest preferences, as well as the priority ranking logic of product description features. The dynamic matching rule set is input into the pre-trained AI recommendation model to perform rule matching filtering on the candidate product multimedia content set, and generate product multimedia content filtering results containing matching degree scores. The RPA display orchestration module performs real-time rendering of the recommended sequence based on the product multimedia content filtering results. The real-time rendering operation includes adjusting the display size of the product image unit and the text layout style of the product description unit. The process of extracting the intent evolution trajectory of the current interaction behavior flow based on a preset historical interaction pattern library includes: The operation triggering sequence in the current interaction behavior flow is analyzed, the time interval parameter and operation type sequence of continuous operation events are extracted, and the operation type sequence is converted into the corresponding operation feature vector. The dimension of the operation feature vector is consistent with the preset operation dimension in the historical interaction pattern library. The attention dwell trajectory in the current interactive behavior flow is analyzed, and the attention dwell time and pupil diameter change parameters of each product description unit are extracted. The variance of the pupil diameter change parameter is calculated as an indicator of attention concentration. The attention dwell time and the attention concentration indicator are merged into an attention feature vector. The operation feature vector and the attention feature vector are concatenated to generate the current interaction feature matrix. The row dimension of the current interaction feature matrix corresponds to the timestamp sequence, and the column dimension corresponds to the joint dimension of the operation feature and the attention feature. Load a preset historical interaction pattern library, which contains multiple historical interaction feature matrices and their corresponding intent tag sets. Calculate the similarity value between the current interaction feature matrix and each historical interaction feature matrix using a cosine similarity algorithm, and select the historical interaction feature matrix with the highest similarity value as the reference pattern matrix. Extract the set of intent tags corresponding to the reference pattern matrix and the rate of change of feature values ​​of each timestamp in the current interaction feature matrix. Calculate the feature offset of the current interaction feature matrix relative to the reference pattern matrix using a time series difference algorithm. Determine the dynamic steering feature of the user's interest preference based on the feature offset. The dynamic steering feature is represented by a steering direction vector and a steering amplitude parameter. Based on dynamic steering features, a feature intensity decay function is constructed. The feature intensity decay function uses timestamp as the independent variable and the cumulative sum of feature offsets as the dependent variable to calculate the feature intensity value at different timestamps, thereby generating a feature intensity change trend that reflects the change law of feature intensity over time. The dynamic steering features and the trend of feature intensity change are associated and encapsulated to form an intention evolution trajectory that includes user interest preferences and the trend of feature intensity change.

2. The product multimedia recommendation method combining RPA and AI according to claim 1, characterized in that, The method of recording the current interaction flow between the user and the product multimedia interface in real time through the RPA interaction capture module includes: The RPA interaction capture module is started to load the interface element mapping table, which contains the unique identifier code of each product image unit in the product multimedia interface and the coordinate range of the text area of ​​the product description unit. The RPA interaction capture module uses screen pixel scanning technology to monitor user mouse operation events in the product multimedia interface, identifies the interface element type corresponding to the mouse operation event, and when the interface element type is a product image unit, records the operation trigger timestamp and operation type parameters to generate the operation trigger timing sequence of the product image unit. The screen eye-tracking component is activated synchronously to monitor the movement trajectory of the user's gaze focus in the product multimedia interface. When the gaze focus falls within the coordinate range of the text area of ​​the product description unit, the time point when the gaze enters, the time point when the gaze leaves, and the pupil diameter change parameters during the gaze dwell time are recorded. The difference between the time point when the gaze leaves and the time point when the gaze enters is calculated as the attention dwell time, and the attention dwell trajectory of the product description unit is generated. The operation triggering sequence of the product image unit is aligned with the attention lingering trajectory of the product description unit on the time axis so that operation events and gaze events under the same timestamp are associated and recorded. Noise filtering is performed on the timeline-aligned associated records to remove invalid operation events caused by accidental touches and brief pause events caused by gaze drift, while retaining valid associated records that reflect the user's true interaction intent. This results in a current interaction behavior flow that includes the user's operation trigger sequence for the product image unit and the attention lingering trajectory for the product description unit.

3. The product multimedia recommendation method combining RPA and AI according to claim 1, characterized in that, The process of parsing the operation trigger timing in the current interaction behavior flow, extracting the time interval parameters and operation type sequences of continuous operation events, and converting the operation type sequences into corresponding operation feature vectors includes: Traverse the operation triggering sequence in the current interactive behavior flow, identify the triggering timestamps of two adjacent operation events, calculate the difference between the triggering timestamp of the later operation event and the triggering timestamp of the previous operation event, obtain the time interval parameter of the continuous operation events, and arrange all time interval parameters in chronological order to form a time interval sequence. Extract the operation type parameter of each operation event in the operation trigger sequence. The operation type parameter includes click operation, double-click operation, right-click operation and drag operation. Map each operation type parameter to a preset type code value. Arrange the type code values ​​according to the time order of the operation events to form an operation type sequence. The time interval sequence is standardized to generate a standardized time interval sequence; One-hot encoding is performed on the operation type sequence to convert each type encoding value into a binary vector of equal length, so that different operation types have orthogonality in the vector space, thereby generating a one-hot encoded operation sequence. The standardized time interval sequence and the one-hot encoded operation sequence are subjected to feature fusion processing. The standardized time interval values ​​and the corresponding one-hot encoded vectors are concatenated into a high-dimensional vector in the order of timestamps to generate an operation feature vector that is consistent with the preset operation dimension in the historical interaction mode library.

4. The product multimedia recommendation method combining RPA and AI according to claim 1, characterized in that, The process of analyzing the attention lingering trajectory in the current interaction flow, extracting the attention lingering duration and pupil diameter change parameters for each product description unit, and calculating the variance of the pupil diameter change parameters as an indicator of attention concentration includes: Traverse the attention dwell trajectory in the current interaction behavior flow, identify the gaze entry time and gaze exit time corresponding to each product description unit, calculate the difference between the gaze exit time and the gaze entry time, and obtain the attention dwell time of each product description unit; Extract pupil diameter sampling data for each product description unit in the attention dwell trajectory during the attention dwell period. The pupil diameter sampling data includes multiple pupil diameter measurements collected at fixed time intervals. Arrange the pupil diameter measurements of the same product description unit in the order of sampling time to form a pupil diameter sequence. The pupil diameter sequence is subjected to outlier detection processing to identify and remove abnormal measurement values ​​that exceed the mean plus or minus K times the standard deviation, while retaining the normal pupil diameter sequence; Calculate the variance of all measurements in the normal pupil diameter sequence. The variance reflects the degree of fluctuation of pupil diameter during attention dwell time. Use this variance as an indicator of the degree of attention concentration of users on the product description unit. The attention dwell time of each product description unit is associated with the corresponding attention concentration level index and stored to form an attention feature record that includes the product description unit identifier, attention dwell time, and attention concentration level index.

5. The product multimedia recommendation method combining RPA and AI according to claim 1, characterized in that, The step of generating a dynamic matching rule set based on the intent evolution trajectory includes: The dynamic turning features in the intention evolution trajectory are analyzed, and the product feature dimensions corresponding to the turning direction vector are extracted. The product feature dimensions include product image feature dimensions and product description feature dimensions. The preference polarity of each product feature dimension in the turning direction is determined, and the preference polarity includes positive preference and negative preference. Based on the feature intensity value in the trend of preference polarity and feature intensity change, a preference weight coefficient is assigned to each product feature dimension. The product feature dimension with a higher feature intensity value corresponds to a higher preference weight coefficient, and a product feature dimension weight table is generated. For product image feature dimensions, image feature sub-dimensions with weight coefficients higher than a preset threshold are selected according to the product feature dimension weight table. The image feature sub-dimensions include color feature sub-dimensions, texture feature sub-dimensions, and shape feature sub-dimensions. Preference range constraints are set for each image feature sub-dimension to generate association constraints between product image features and user interest preferences. For product description feature dimensions, the description feature sub-dimensions are sorted according to the product feature dimension weight table. The description feature sub-dimensions include keyword feature sub-dimensions, topic feature sub-dimensions, and sentiment feature sub-dimensions. The importance order of each description feature sub-dimension is determined according to the weight coefficient from high to low, and the priority sorting logic of product description features is generated. The association constraints between the product image features and user interest preferences, as well as the priority ranking logic of the product description features, are processed by rule formatting. The constraints and ranking logic are converted into machine-parseable rule statements using a preset rule syntax, and all rule statements are merged to form a dynamic matching rule set.

6. The product multimedia recommendation method combining RPA and AI according to claim 5, characterized in that, The process involves selecting image feature sub-dimensions with weight coefficients higher than a preset threshold based on a product feature dimension weight table, setting preference range constraints for each image feature sub-dimension, and generating association constraints between product image features and user interest preferences, including: Traverse the product image feature dimensions in the product feature dimension weight table, extract the weight coefficients corresponding to each image feature sub-dimension, compare the weight coefficients with a preset threshold, and select the image feature sub-dimensions with weight coefficients higher than the preset threshold as key image feature sub-dimensions. For each key image feature sub-dimension, the range of feature values ​​with the highest user interaction frequency under that key image feature sub-dimension is analyzed in the historical interaction pattern library. The range of feature values ​​is obtained by statistical analysis of the feature value distribution in historical interaction data, and includes the lower limit and upper limit of the feature values. By combining the dynamic steering features in the intention evolution trajectory, the lower and upper limits of the feature value range are adjusted. When the steering direction vector of the dynamic steering feature points to the direction of increasing feature value, the upper limit is increased while the lower limit remains unchanged; when the steering direction vector points to the direction of decreasing feature value, the lower limit is decreased while the upper limit remains unchanged. Set a constraint type identifier for the adjusted feature value range. The constraint type identifier includes mandatory constraints and priority constraints. The key image feature sub-dimensions with the top preset proportion of weight coefficients correspond to mandatory constraints, while the other key image feature sub-dimensions correspond to priority constraints. The key image feature sub-dimensions, adjusted feature value ranges, and constraint type identifiers are combined into constraint condition statements. All constraint condition statements together constitute the association constraints between product image features and user interest preferences.

7. The product multimedia recommendation method combining RPA and AI according to claim 5, characterized in that, The step of prioritizing product description features involves sorting the sub-dimensions according to a product feature dimension weight table, determining the importance order of each sub-dimension based on its weight coefficient from high to low, and generating a priority ranking logic for product description features, including: Extract all descriptive feature sub-dimensions under the product description feature dimension and their corresponding weight coefficients from the product feature dimension weight table, and construct a target comparison table; The descriptive feature sub-dimensions in the target lookup table are sorted from high to low according to their weight coefficients to obtain the descriptive feature sub-dimension sorting sequence; Assign a priority level to each descriptive feature sub-dimension in the sorted sequence. The first descriptive feature sub-dimension in the sorted sequence corresponds to the highest priority level, and the priority levels of subsequent descriptive feature sub-dimensions decrease in order. For each priority level of the descriptive feature sub-dimension, a priority rule for feature matching is set. The higher the priority level of the descriptive feature sub-dimension, the more it is evaluated in the matching process, and the contribution weight of its matching result to the overall matching degree is higher than that of the lower priority level of the descriptive feature sub-dimension. The sorting sequence of description feature sub-dimensions, priority levels, and feature matching priority rules are combined into sorting logic statements to form the priority sorting logic for product description features.

8. The product multimedia recommendation method combining RPA and AI according to claim 1, characterized in that, The step of inputting the dynamic matching rule set into a pre-trained AI recommendation model, performing rule matching filtering on the candidate product multimedia content set, and generating product multimedia content filtering results including matching score includes: Load a candidate product multimedia content set, which contains multiple product multimedia content units, each of which contains corresponding product image feature data and product description feature data; The dynamic matching rule set is input into the rule parsing layer of the AI ​​recommendation model. The rule parsing layer performs syntax parsing on the association constraints and priority ranking logic in the dynamic matching rule set, and converts them into feature matching operators and weight allocation parameters that the model can execute. The AI ​​recommendation model's feature extraction layer performs feature parsing processing on the product image feature data of each product multimedia content unit, extracting image feature values ​​corresponding to the key image feature sub-dimensions in the association constraints. It also performs feature parsing processing on the product description feature data, extracting description feature values ​​corresponding to the description feature sub-dimensions in the priority sorting logic. The matching and filtering process is performed using the rule matching layer of the AI ​​recommendation model. The image feature values ​​are compared with the feature value range in the associated constraints. The number of image feature sub-dimensions that meet the mandatory constraints and the number of image feature sub-dimensions that meet the priority constraints are counted. The matching degree of the descriptive feature values ​​is evaluated in turn according to the priority sorting logic. Different matching weights are assigned according to the priority level, and the comprehensive matching score is calculated. The comprehensive matching scores of all product multimedia content units are sorted in descending order. A preset number of product multimedia content units at the top of the list are selected as the filtering results, and the comprehensive matching score of each product multimedia content unit is retained as the matching degree score. Finally, a product multimedia content filtering result containing the matching degree score is generated.

9. A multimedia product recommendation system combining RPA and AI, characterized in that, The device includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the product multimedia recommendation method combining RPA and AI as described in any one of claims 1-8.

Images