Market research system based on multi-source data integration and psychological analysis

Abstract

The application relates to the technical field of the cross of artificial intelligence and behavior science, in particular to a market research system based on multi-source data integration and psychological analysis, which comprises a processing module, an RAG module, a retrieval module, a summary module, an analysis module and a verification module; the processing module processes feedback data of market research; the RAG module retrieves to obtain multi-modal data; the retrieval module retrieves to obtain data samples; the summary module extracts data features; the analysis module analyzes the data samples and generates analysis conclusions; and the verification module verifies the results of the analysis module. The application generates multi-modal data of feedback data in the Internet through the retrieval module, and according to the obtained data samples, the analysis module analyzes to obtain analysis conclusions by combining virtual user models of psychology and / or sociology, so that the problem that the system has weak theoretical support and single data dimension is avoided.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and behavioral science, and in particular to a market research system based on multi-source data integration and psychological analysis. Background Technology

[0002] Before launching new products, services, advertisements, or strategies on a large scale into the market or online, companies should use scientific methods to test them in a small, real consumer group or collect consumer feedback on the internet, integrate, analyze, and evaluate consumer data, and then refine it into a concrete and labeled model data system.

[0003] On the internet, market research systems serve as the "brain" and "navigator" for modern enterprises to conduct refined operations and decision-making, playing a role in all consumer-related aspects such as marketing, product development, and service. Utilizing market research systems enables precise marketing and personalized recommendations, optimizes product development and selection, enhances user experience and service quality, drives data-driven business decisions, and facilitates the identification of potential customers and risk control.

[0004] In existing technologies, market research systems mainly rely on structured data such as transaction logs of advertising products, lacking the ability to integrate multimodal data such as social media and interview texts on the Internet, resulting in weak theoretical support and the problem of single data dimensions. Summary of the Invention

[0005] Therefore, it is necessary to provide a market research system based on multi-source data integration and psychological analysis to address the above-mentioned problems.

[0006] This invention is implemented as follows: a market research system based on multi-source data integration and psychological analysis, comprising:

[0007] A processing module, which is used to process the feedback data from market research;

[0008] The RAG module is used to generate multimodal data for feedback data;

[0009] A retrieval module is used to perform mixed retrieval of multimodal data in a database to obtain data samples;

[0010] The summary module is used to aggregate information and extract structured data features from multimodal data.

[0011] The analysis module analyzes multimodal data using a virtual user model to obtain analytical conclusions.

[0012] A verification module is used to verify the results of the multimodal data and / or analysis module.

[0013] The market research system based on multi-source data integration and psychological analysis provided in this invention processes market research feedback data through a processing module, retrieves data in the RAG module, calls a vector database to generate multimodal data, searches the internet for multimodal data using a retrieval module to obtain data samples related to the feedback data, extracts structured data from the multimodal data through a summary module, and then analyzes the multimodal data using a virtual user model generated by combining psychological and / or sociological theories in the analysis module to obtain analytical conclusions. A verification module verifies the multimodal data and / or the conclusions of the analysis module. Thus, by integrating multimodal data from market research feedback data on the internet to generate multimodal data, and retrieving data samples from the internet through the retrieval module, the summary module can extract structured data features from the data samples. Finally, the analysis module uses a virtual user model combining psychological and / or sociological feature models to analyze the multimodal data and obtain analytical conclusions. This avoids the problems of weak theoretical support and single data dimensions that often exist in market research for product advertising. Attached Figure Description

[0014] Figure 1 This is a structural block diagram of a market research system based on multi-source data integration and psychological analysis, provided in one embodiment. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0016] It is understood that the terms "first," "second," etc., used in this invention may be used to describe various elements herein, but unless specifically stated otherwise, these elements are not limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of this invention, a first script may be referred to as a second script, and similarly, a second script may be referred to as a first script.

[0017] Figure 1 Here is a structural block diagram of a market research system based on multi-source data integration and psychological analysis provided in one embodiment, such as... Figure 1 As shown, in one embodiment, a market research system based on multi-source data integration and psychological analysis is proposed. Specifically, it may include:

[0018] A processing module, which is used to process the feedback data from market research;

[0019] The RAG module is used to generate multimodal data for feedback data;

[0020] A retrieval module is used to perform mixed retrieval of multimodal data in a database to obtain data samples;

[0021] The summary module is used to aggregate information and extract structured data features from multimodal data.

[0022] The analysis module analyzes multimodal data using a virtual user model to obtain analytical conclusions.

[0023] A verification module is used to verify the results of the multimodal data and / or analysis module.

[0024] In this embodiment of the invention, preferably, the market research system based on multi-source data integration and psychological analysis can be a system that integrates feedback data from market research on products or creative advertising design, packaging, etc., and processes the feedback data based on multi-source data integration and psychological analysis. Based on data from real consumers surveyed in the market, it integrates multi-source data and uses psychological analysis to expand the data sample by creating digital twins of consumers, thereby gaining a deeper understanding of consumer needs and users, identifying unmet needs in various usage scenarios, and is suitable for application scenarios such as optimizing marketing strategies, verifying product design, mining consumer needs, and predicting behavior.

[0025] A market research system based on multi-source data integration and psychological analysis includes: a processing module, a RAG module, a retrieval module, a summary module, an analysis module, and a verification module. The processing module processes market research feedback data, which can be sourced from various sources, including but not limited to real consumer interview transcripts, focus groups, family notes, quantitative questionnaire results, social media text, and user facial expressions. Data from real consumers provides insights into emotional and irrational aspects, rapidly changing trends, content not typically posted on social media platforms, real-life scenarios, and reactions to specific content, enriching previously limited understanding of markets or product categories. Market research feedback data can also include social media content, facilitating real-time analysis. From this feedback data, real-time demand analysis, identification of usage scenarios, potential needs, and unmet needs can be performed.

[0026] After integrating and processing multimodal feedback data through the processing module, the RAG module is a technical framework for information retrieval and text generation. By retrieving the processed feedback data through the RAG module, it is possible to search in the database and construct multimodal data related to the feedback data. Multimodal data is generated by performing mixed retrieval and knowledge enhancement on the processed feedback data in the vector database and generating data forms related to the feedback data. Multimodal data can include, but is not limited to, text data, image data, and voice data.

[0027] The retrieval module initiates multimodal data retrieval requests across multiple databases, calling the most relevant data samples from the vector database. This expands the retrieval scope and integrates the multimodal data, avoiding data dimensionality limitations and improving the system's dynamic adaptability. The summarization module aggregates information from multiple dimensions of the obtained multimodal data and performs structured extraction to obtain data features. Structured extraction involves acquiring predefined types of information from the multimodal data and / or data samples and outputting it in a standardized format to obtain data features.

[0028] The analysis module combines multimodal data samples and their data characteristics with virtual user models based on psychological and / or sociological theories to predict consumer behavior and interpret and model psychological states. It analyzes market research feedback data to understand the psychological states or behavioral judgments. For example, it can analyze questionnaire data after consumers have viewed the packaging design of a product to analyze consumers' specific opinions on the product packaging design, such as liking the color of the product packaging but disliking the pattern design, and analyze the psychological state of consumers' liking or disliking of the product packaging, and predict whether to buy or not.

[0029] Finally, the verification module verifies the analysis conclusions of the analysis module. Combining psychological and sociological theories can avoid the impact of weak theoretical support on the research results when conducting data analysis. In this way, the market research feedback data is output as verified analysis conclusions through a market research system based on multi-source data integration and psychological analysis.

[0030] After receiving feedback data, the processing module can process the data and ensure its quality and consistency. The RAG module extracts and organizes semantic information from the processed data to generate multimodal data. The retrieval module uses its search function, the summary module performs structured extraction, and the analysis module combines psychological and sociological theories to analyze the psychological variables and motivational mechanisms within the data. The results are then validated by the verification module, finally outputting analytical conclusions and a virtual user model for decision-making. For example, a typical task, "analyzing the dietary preferences of Generation Z women in specific scenarios," involves processing a large amount of social media comments and in-depth interview texts through the processing module, retrieving relevant multimodal content and performing semantic aggregation through the RAG module, and then using the retrieval module... The database searches for data samples related to multimodal data to avoid missing multimodal data and integrating data with a single dimension. It can also aggregate multi-dimensional information through the summary module and extract structured data features of multimodal data to obtain specific characteristics of the dietary preferences of Generation Z women in the current specific scenario, such as the workplace environment. These characteristics may include a preference for healthy meals, a need for weight loss, and the ability to supplement and maintain a balanced diet. By using the virtual user model in the analysis module to model with multimodal data, it is possible to analyze the variables of the psychological state and the structure of the behavioral motivations behind Generation Z women. The verification module then verifies the entire analysis process and the data acquisition process, and finally generates conclusions and behavioral judgment results that can be used for business decision-making.

[0031] The market research system based on multi-source data integration and psychological analysis provided in this invention processes market research feedback data through a processing module, retrieves data in the RAG module, calls a vector database to generate multimodal data, searches the internet for multimodal data using a retrieval module to obtain data samples related to the feedback data, extracts structured data from the multimodal data through a summary module, and then analyzes the multimodal data using a virtual user model generated by combining psychological and / or sociological theories in the analysis module to obtain analytical conclusions. A verification module verifies the multimodal data and / or the conclusions of the analysis module. Thus, by integrating multimodal data from market research feedback data on the internet to generate multimodal data, and retrieving data samples from the internet through the retrieval module, the summary module can extract structured data features from the data samples. Finally, the analysis module uses a virtual user model combining psychological and / or sociological feature models to analyze the multimodal data and obtain analytical conclusions. This avoids the problems of weak theoretical support and single data dimensions that often exist in market research for product advertising.

[0032] In one embodiment of the present invention, the processing of market research feedback data includes:

[0033] The feedback data from the market research was divided into text data, image data, and voice data.

[0034] Text data is processed using a regular expression library combined with a bidirectional LSTM time anomaly detection model.

[0035] Image processing is performed using a mid-range filtering algorithm on image data.

[0036] Wavelet transform is used to filter noise in speech data.

[0037] Data alignment is performed on text data after text processing, image data after image processing, and speech data after noise filtering using a unified timestamp system, a spatiotemporal alignment engine, and a spatial coordinate transformation matrix.

[0038] The quality of the aligned feedback data is verified using data credibility assessment metrics.

[0039] In this embodiment of the invention, preferably, the processing module is mainly used to uniformly clean, structure, and standardize various data types in different formats. Market research feedback data can include, but is not limited to, text interview transcripts, focus groups, family notes, quantitative questionnaire results, social media text, user facial expressions, and other data sources based on real consumer data. This data may contain a large amount of noise, redundant content, missing information, and ambiguous spatial positioning. Given the pursuit of data breadth rather than individuality, unified processing of various types of data is necessary. The sampled feedback data is categorized into text data, voice data, and image data. After cleaning each type of data, data alignment is performed, and finally, quality verification is conducted. The data is first input into the AI ​​decoding model through an adaptive parser. An adaptive parser refers to the process by which the system modifies its parameters or structure in real time according to changes in operating status and environment to optimize performance or adapt to new requirements. Data can be categorized according to type and format. The dynamic adjustment of the adaptive parser is based on a feedback control mechanism, which monitors key indicators (such as response time and resource utilization) and applies algorithms (such as PID control and reinforcement learning) for adaptive optimization. The core of dynamic adjustment lies in balancing resource consumption and efficiency to ensure that the system can maintain stability in a changing environment.

[0040] For example, text data (interview transcripts, social media comments, etc.) is processed using regular expression cleaning, part-of-speech tagging, and named entity recognition based on AI decoding models before being input into a Natural Language Processing (NLP) model to extract semantic intent, sentiment polarity, and verb subject structure. Regular expression libraries can be considered miniature languages ​​for matching patterns in strings. A regular expression is a method for describing character sequences and a powerful computational tool for processing character sequences. Regular expression libraries typically contain a series of classes and functions that encapsulate regular expressions and the results of matching them within a target sequence of characters. These libraries provide powerful functionalities such as searching, replacing, and splitting strings and are widely used in various programming languages ​​and tools. Bidirectional LSTM (Long Short-Term Memory) networks are deep learning models suitable for processing sequential data, particularly well-suited for capturing long-term dependencies in time series. LSTM time series anomaly detection is an important data analysis technique that helps discover sudden anomalies or anomalous patterns in time series data. A function can be defined to generate training samples and labels, where the training samples are a subset of data the size of a time window, and the labels are the values ​​at the next time step. Suppose we have a dataset containing time and numerical values, where the numerical columns are for anomaly detection. However, a well-designed LSTM (Layered Stochastic Mechanism) recurrent neural network can be used instead of a multilayer perceptron structure. This requires transforming the two-dimensional data (rows representing different times and columns representing different variables) into a three-dimensional recurrent neural network structure. This structure is then fed into the neural network to train an autoencoder, and finally decoded to predict whether the reconstruction error is an outlier. This process involves transforming the data shape.

[0041] The voice data primarily consists of recorded audio from user interviews. Wavelet transform denoising is an effective signal processing technique that utilizes the time-frequency localization properties of wavelet transform to separate useful components from noise in the signal. The basic principle of wavelet transform denoising is to decompose the signal into different scales using wavelet transform. At each scale, the noise coefficient amplitude is typically smaller than the signal coefficient amplitude. Therefore, by setting a threshold, the noise coefficient can be removed or reduced while preserving the useful information of the signal.

[0042] Image data (such as screenshots) combined with eye-tracking data can generate interest heatmaps, which can also be combined with emotion recognition models to further analyze visual emotion tendencies. The main role of median filtering algorithms in image processing is to remove noise from images. Median filtering is a non-linear digital filtering technique that is commonly used to remove noise from images or signals. This noise reduction is a typical preprocessing step to improve the results of subsequent processing (e.g., edge detection of images).

[0043] Cross-modal semantic fusion is performed using a trained AI decoding model. Based on a unified Transformer model architecture, data alignment is performed on data of different data types. Data alignment is achieved by establishing a unified timestamp system, developing a spatiotemporal alignment engine, and using a spatial coordinate transformation matrix. Data alignment includes two operations: time alignment and spatial alignment. The purpose is to unify the reference dimensions of data from different modalities and ensure synchronization and comparability in subsequent modeling.

[0044] Data alignment for the feedback data includes assigning a unified timestamp (millisecond level) to text, images, and audio, and converting geographic information to a unified coordinate system (such as WGS84). The output is that each data point has standardized time and spatial markers, facilitating the insertion of a unified behavioral timeline or spatial distribution analysis. The Spatiotemporal Alignment Engine (STA-Engine) first extracts the original time and spatial fields (such as EXIF, logs, and location) from various data types and converts them uniformly to UTC millisecond timestamps and standardized latitude and longitude. It calls different extraction modules to parse metadata for text, audio, and images respectively. The processing result is that each sample data point is accompanied by a timestamp (ms) and (lat, lon) format, outputting structured data records as input for subsequent timeline reconstruction and spatial modeling.

[0045] All input data requires millisecond-level timestamps to support behavioral chain analysis and cross-modal synchronization. This precision ensures the model can distinguish between events occurring consecutively within a short period (such as speech and action frames) and input them into the RAG module retrieval or behavioral simulation process in strict chronological order. The timescale is used to restore the sequential nature and temporal causal relationships of data occurrences. Without time information, the model cannot determine the order of events, making it difficult to build behavioral evolution models and support time-driven insight tasks such as commuting scenarios and emotional fluctuations. A unified timescale is a prerequisite for scene recognition and decision path tracking. After noise filtering, the retained data enters the timestamp module. This module converts the original acquisition time (such as file generation time, system log time, or voice stream time encoding) to millisecond-level UTC format. A high-precision clock reference and rules are used to identify abnormal times (such as future times, time reversals, etc.), and after correction, unified modeling is performed. In the spatial coordinate transformation matrix, the system supports conversion between multiple coordinate systems (such as WGS84, GCJ02, BD09). When the data contains GPS or address fields, the system identifies its coordinate system and converts it to WGS84 latitude and longitude format using a preset transformation matrix (or an external API). Text addresses are parsed into coordinates by the geocoding module, and the converted results are used for geographic clustering and spatial behavior analysis.

[0046] The unified timestamp system refers to a method of representing and processing time using Unix timestamps in computer science and network technology; the spatiotemporal alignment engine is a technique for processing and analyzing spatiotemporal data; the spatial coordinate transformation matrix is ​​used to describe the transformation process of points, vectors, or any other data items from one coordinate system to another in three-dimensional space. The unified Transformer model architecture customizes tasks and adjusts the structure within a general pre-trained large model. The system employs Transformer's multi-head attention mechanism, residual connections, and positional encoding structure, but its input interface and output tasks have been optimized at the module level to adapt to the actual needs of market preference prediction. The model architecture does not rely on behavioral comparison data between multiple products for the same user, but rather builds a preference trend model based on the emotional responses, language expressions, and behavioral signals of a large number of distributed users to various product designs. Through the statistical distribution differences of a large number of user behaviors, it constructs a relative advantage judgment of a product category involving the overall preference of the target population, rather than a comparative judgment of a particular individual. When various types of feedback data are input into the unified Transformer model architecture through independent embedding networks, cross-modal data alignment can be performed.

[0047] After data alignment, the quality of the aligned feedback data is verified using data credibility assessment metrics. These metrics primarily include the following dimensions: 1. Temporal integrity: Does the data have a standard timestamp, and are there any time jumps or missing times? 2. Spatial consistency: Does the geographical location match the task's defined area, and are there any offsets or outliers? 3. Data density: Is the data sufficiently covered within the analysis period, and are there any long-term gaps in data? 4. Noise residual rate: Based on the signal quality ratio before and after filtering (e.g., speech signal-to-noise ratio, image blur). 5. Modal consistency: Are multimodal data synchronized and complete (e.g., image-speech-text triplet integrity)? These metrics are derived from intermediate parameters calculated by the data cleaning and alignment modules, and a credibility score (e.g., 0~1 floating-point number) is generated after comprehensive analysis. This score is used for subsequent filtering and weighted processing. By classifying, processing, aligning, and verifying the feedback data, the preprocessing of the feedback data is completed.

[0048] In one embodiment of the present invention, the step of vectorizing the processed feedback data includes:

[0049] The processed text data is linked to the Multi-BERT-XL model for vectorization, resulting in vectorized text data.

[0050] The processed image data is linked to the Vision Transformer model for vectorization, resulting in vectorized image data.

[0051] The noise-filtered speech data is linked to the Wav2Vec model for vectorization processing to obtain vectorized speech data.

[0052] The vector space alignment module maps vectorized text data, vectorized image data, and vectorized speech data to obtain structured vector data.

[0053] In this embodiment of the invention, preferably, the processing module further includes a vectorization processing module. First, the feedback data after data processing, data alignment, and quality verification in the preprocessing module is vectorized. For example, for text data after text processing, semantic vectors are input into the MultiBERT-XL model. The MultiBERT-XL model is a multilingual pre-trained model supporting 128 languages ​​and combining the features of BERT (Bidirectional Encoder Representations from Transformers) and XLNet to vectorize the text data after text processing. For image data, the Vision Transformer-384, an improved version of CLIP, is used in the image editor. The CLIP visual encoder part adopts the Vision Transformer architecture, and 384x384 resolution images are used during pre-training to improve the model's performance when processing high-resolution images. It may also achieve better results in tasks such as zero-shot classification and image retrieval. For speech data, the Wav2Vec 3.0 temporal encoder is used in speech feature extraction. The temporal encoder is one of the core components of the model. It is responsible for converting the raw audio signal into a series of feature vectors and capturing the temporal characteristics of speech by processing these feature vectors. In Wav2Vec 3.0, the temporal encoder typically uses a combination of convolutional neural network (CNN) and transformer encoder layers to vectorize speech data.

[0054] The vectorized data from these different modalities are not used directly after the vectors are generated. Instead, they are uniformly sent to a dimension vector space alignment module. This module aligns different types of data in the semantic space using a specific loss function and organizes them into structured vector data. It can also be input into a vector database to ensure that the feature dimensions of all modal outputs are consistent, usually fixed at 1024 dimensions, so as to ensure that they can be easily queried in the database, stored side by side, and compared across modalities.

[0055] In one embodiment of the present invention, the multimodal data for generating feedback data includes:

[0056] The feedback data processed by the processing module is vectorized.

[0057] The vectorized feedback data is input into the multimodal query module to construct multimodal vectors;

[0058] The HNSW+PQ hybrid index is used to retrieve vector data from the vector database and feedback data from vectorization processing in parallel.

[0059] The retrieved vector data related to the feedback data will be embedded in a preset prompt word template by an adaptive load balancer.

[0060] The Delta Update algorithm is applied to generate multimodal data related to the feedback data.

[0061] In this embodiment of the invention, preferably, the RAG module mainly includes a storage layer, a service layer, and an update layer. The vectorized structured vector data is output to the multimodal query module, which facilitates the construction of multimodal vectors from the structured vector data obtained after the feedback data is processed by the processing module. In the storage layer of the RAG module, the HNSW+PQ hybrid indexing method is used to retrieve relevant vector data from the vector knowledge base, which improves the retrieval speed while ensuring high recall. The data is then input into the language model to enhance the model's ability to handle knowledge-intensive tasks. Furthermore, the vector data in the vector database and the vectorized feedback data are retrieved in parallel. The HNSW+PQ hybrid indexing method is an indexing method that combines the advantages of both HNSW and PQ indexing algorithms. HNSW (Hierarchical Navigable Small World) is a graph-based indexing algorithm that generates a graph structure from the index data, enabling searches to quickly converge to the nearest neighbor. PQ (Product Quantization) is a product quantization indexing algorithm that compresses and stores vectors by decomposing high-dimensional vectors into multiple sub-vectors and quantizing and encoding each sub-vector.

[0062] The feedback data after parallel retrieval, along with the vector data in the related vector database, are embedded into a preset prompt word template in the service layer of the RAG module through the application adaptive load balancer. The service layer supports high-concurrency access scenarios through query routing and load balancing mechanisms, ensuring that the call is not affected by latency. The adaptive equalizer used is a device or software that can dynamically adjust resource allocation according to real-time demand to optimize system performance and resource utilization. The update layer in the RAG module introduces a Delta-Update incremental update mechanism, which can quickly insert newly generated vectors into the database at a relatively low cost, ensuring the timeliness and accuracy of search results. It can integrate feedback data and vector data linked to a language model based on preset prompt templates into the update generation process, and use the Delta Update algorithm to generate the required output content to update and generate multimodal data related to the feedback data. Multimodal data is a collection of various forms of feedback data samples related to the products in market research. For example, feedback data on product packaging design can generate data information about the product itself or related data such as the design concept of the packaging design, which can be integrated. The Delta Update algorithm is a technique used to generate and apply incremental update packages, which can improve the efficiency and stability of the incremental update package process. These database structures directly correspond to multiple retrieval stages within the RAG module.

[0063] In the Multimodal Query Understanding (MMQU) phase, the system first constructs multimodal vectors from the vectorized input feedback data. This is followed by the Hybrid Retrieval phase in the storage layer, where parallel retrieval is performed using multiple index structures in the database. Background knowledge is then supplemented by Knowledge Augment in the generation layer. Finally, at least one of the retrieval, summarization, analysis, and summary modules performs generation and feedback optimization, either sequentially or in parallel. Therefore, the multiple data sources from market research are transformed into structured, uniformly characterized input items through the processing module. This data, through embedding models and spatial mapping, constitutes the core content of the vector database. The database's structural design is crucial for ensuring search efficiency and system response performance, providing a fundamental guarantee for each sub-process of the RAG module, enabling it to truly complete the "retrieval + generation" closed loop. This transforms massive amounts of heterogeneous data into an efficient "information network," providing readily available knowledge support for the retrieval, summarization, analysis, and summary modules.

[0064] As one embodiment of the present invention, such as Figure 1 As shown, the step of retrieving multimodal data from the database to obtain data samples includes:

[0065] The Hybrid-Retriever retrieval method is applied to read and identify multimodal data;

[0066] The vector data in the vector database and the virtual user model are used to perform a similarity search on the reading and recognition results in the vector space to obtain data samples that are consistent with the multimodal data modality.

[0067] In this embodiment of the invention, preferably, after generating multimodal data, a retrieval module can be invoked to retrieve the multimodal data and the Hybrid-Retriever retrieval method can be used to read and identify the multimodal data of the feedback data. The Hybrid-Retriever retrieval method performs cross-database hybrid retrieval across multiple databases. These databases can be databases of feedback data on a product during market research, historical databases of product packaging design conducted at other times or in other scenarios, or professional databases of the same category as the product. The semantic information of the multimodal data of the feedback data is captured to facilitate similarity searches of the reading and identification results. Searches are performed within the vector space of a vector database, and similarity searches can also be performed by invoking virtual user models. The retrieval can yield several virtual user models consistent with the multimodal data modalities, such as historical data of the same market research product as the multimodal data, such as mobile phone casing design from last month's market research. Research data can be obtained from surveys among students or working professionals, resulting in data samples consistent with the multimodal data modality. Similarity searches can be based on independent feedback from at least one predefined virtual user model, whose content aligns with the multimodal data modality and produces a structured response. This predefined virtual user model can be consistent with the multimodal data modality, such as calling virtual user models belonging to either the student or working professional groups. Based on the relevant data samples obtained from the virtual user models, other data samples consistent with the multimodal data modality and producing a structured response to the product can also be retrieved from the database. This expands the data base for summarizing or analyzing, avoiding the problem of a single data dimension.

[0068] The Hybrid-Retriever retrieval method, derived from Hybrid-Retriever v2.1, is a hybrid retrieval tool combining dense and sparse vector search techniques. Hybrid-Retriever v2.1 operates based on Milvus's hybrid search functionality. Milvus is an open-source vector search engine that allows users to search using both dense and sparse vectors simultaneously. Dense vector search typically uses embeddings generated by deep learning models, which capture the semantic information of the text. Sparse vector search, on the other hand, typically uses traditional information retrieval techniques such as TF-IDF or BM25, which can precisely match keywords in the query.

[0069] As one embodiment of the present invention, such as Figure 1 As shown, the information aggregation and extraction of structured data features from multimodal data includes:

[0070] Data analysis and feature extraction of multimodal data were performed using the SummaGPT module and a virtual user model.

[0071] Perform dimensional aggregation on the results of data feature extraction;

[0072] The results of dimensional aggregation are subjected to structured extraction to obtain structured data features.

[0073] In this embodiment of the invention, preferably, the summarization module mainly associates the data samples retrieved by the retrieval module with multimodal data, and then uses the SummaGPT module and a virtual user model to perform data analysis and extract data features from the association results. The SummaGPT module can quickly obtain the core content of the associated data, and is used for data analysis and extraction of the required data features. Multiple dimensions are set to aggregate and analyze the extracted data features. These multiple dimensions can include, but are not limited to, time, source, sentiment, and entity dimensions. After aggregating information from multiple dimensions, structured extraction is performed, which can include, but is not limited to, entity extraction, relationship extraction, and event extraction. Features extraction and keyword extraction can be used to extract structured features of product A from multimodal data, such as the color features, pattern design features, and emotional tendencies of consumers towards product A. After extraction, several structured data features can be obtained. Alternatively, several virtual user models with the same psychological behavior as real consumers can be used to analyze multimodal data, which can also yield data features consistent with the modalities and structure of the multimodal data. Furthermore, by combining the data features obtained from the virtual user models with the data features obtained from the multimodal data through multi-dimensional information aggregation, a more comprehensive set of data features can be summarized, thereby expanding the data scope of market research and resulting in more uniform and comprehensive data features.

[0074] The SummaGPT module is a tool based on GPT technology designed to help users quickly extract the core content of articles and videos, saving time. Through intelligent algorithms, SummaGPT provides concise and clear summaries, making complex information easy to understand. Using SummaGPT mainly involves: entering the content you want to summarize; clicking the summarize button to generate a brief summary; and reading and utilizing these summaries to quickly obtain key information.

[0075] In one embodiment of the present invention, the steps for establishing the virtual user model include:

[0076] The parameters of psychological and sociological theories are vectorized and then input into the mapping matrix.

[0077] The multimodal data is input into a mapping matrix, and the InsightEngine method is used to combine the multimodal data with vectorized psychological and sociological theoretical parameters, respectively.

[0078] The SNA algorithm is used to perform multimodal data psychological state modeling and behavior prediction modeling on the combined results, and at least one virtual user model is obtained.

[0079] In this embodiment of the invention, preferably, a virtual user model is constructed by combining psychological and / or sociological theoretical features to analyze multimodal data and / or data features, and to predict their psychological state, behavior, or business insights. The virtual user model in the analysis module can be a model that exhibits psychological and behavioral consistency with real consumers. The virtual user model primarily encodes psychological and sociological theoretical parameters into feature vectors, which are then used as input mapping matrices to a neural network. Simultaneously, the data features of the multimodal data are input into the mapping matrix, and the data within the matrix is ​​combined using the InsightEngine method to generate analytical conclusions. The mapping matrix can be 15-dimensional. N, where 15 represents the number of psychological variables and N is the total number of corresponding data feature dimensions, constitutes the psychological characteristics of the data sample. Psychological theoretical parameters include, but are not limited to, the Big Five personality theory and Maslow's hierarchy of needs; sociological theoretical parameters include, but are not limited to, knowledge of group social relationship networks and cultural value orientation theory. The psychological characteristics of the constructed multimodal data are used to generate analytical conclusions about market research feedback data. The psychological characteristics are then processed through the SNA algorithm, making node behavior driven by psychological characteristics, and the psychological characteristics are mapped through a neural network to generate behavioral judgment results, thereby simulating a virtual user model with psychological consistency with real consumption. The resulting virtual user model can be called in modules including, but not limited to, the retrieval module, the summary module, the analysis module, and the verification module. The virtual user model is like a virtual user twinned from a real consumer. This virtual user is constructed based on psychological and sociological theoretical parameters, possessing the same psychological state, personality traits, social relationships, and behavioral characteristics as the real consumer. The virtual user model can be called in the retrieval module, the summary module, the analysis module, and the verification module to retrieve data, summarize data characteristics, analyze the psychological state and behavioral prediction of real consumers, or verify the obtained analytical conclusions from the feedback data obtained from real consumers in market research.

[0080] Furthermore, multiple virtual user models can be generated. These models can be multiplied or increased in number based on different psychological characteristics. For example, if the multimodal data on real consumers' perceptions of product design includes ten dimensions of information, such as whether real consumers prefer square patterns, green backgrounds, or white patterns, ten corresponding virtual user models can be constructed using psychological and sociological theories. Moreover, based on the personality traits or behavioral predictions of these virtual user models, data features not found in the multimodal data can be predicted. For instance, if the multimodal data does not mention whether real consumers like cartoon or anime patterns, corresponding data can be obtained through virtual user models that align with real consumer psychology. Virtual user models not only expand the sample size of multimodal data—from 100 real consumers to 1000—by generating virtual user models consistent with their psychological behavior, the overall survey sample size can be increased, resulting in more comprehensive and balanced market research results, leading to more complete and reliable analytical conclusions. The SNA algorithm can be the SNA-Pro algorithm. SNA, or Social Network Algorithm, can be used to build multi-intelligence systems. It helps analysts understand complex interpersonal networks, thus improving intelligence gathering. By analyzing the connections between individuals, the SNA algorithm can reveal patterns and structures hidden behind the data, allowing for the simulation of virtual user profiles that are psychologically and behaviorally consistent with the market research subjects, driven by psychological characteristics of the nodes.

[0081] In one embodiment of the present invention, the step of analyzing multimodal data through a virtual user model to obtain analytical conclusions includes:

[0082] Multimodal data is matched and combined with virtual user models using the InsightEngine method;

[0083] The SNA algorithm is used to perform multimodal data psychological state and behavior prediction analysis on the combined results and draw analytical conclusions.

[0084] In this embodiment of the invention, preferably, the multimodal data is matched and combined with the virtual user model using the InsightEngine 5.0 system with the InsightEngine method. The process of analysis and comparison after combination can be an analysis of consumers' psychological state and behavioral prediction regarding the product. Furthermore, the combination with the virtual user model can expand the number of data analyzed by the analysis module and extend the dimensions of consumer analysis. For example, if the multimodal data does not mention the degree of liking for cartoon characters in the product's graphic design, by matching and combining it with the virtual user model, the percentage of people who like cartoon characters can be analyzed. Alternatively, it can be used to mine business insights from market research objects, predict the changing trends of market acceptance of various objects, and help companies avoid more than 80% of potential design defects. For example, if the target is an advertisement or packaging design, by analyzing multimodal consumer data, predictions can be made regarding the psychological state or preference for the product, as well as consumer purchasing or clicking behavior. For instance, it can reveal whether consumers like or are interested in the product, and provide corresponding data percentages, such as 50% liking the product itself, 60% liking the design, and 50% liking the color. Furthermore, it can predict whether consumers will choose to purchase or click to learn more, with 60% choosing to purchase and 80% choosing to click to learn more. The system can also align with the feedback data of the advertisement and conduct business and marketing analysis to determine the commercial value of the advertisement in the market. This analysis can also improve the quality and efficiency of content marketing for a specific product. InsightEngine 5.0 is an AI-powered creative generation and management platform. According to search results, InsightEngine is applied to multiple marketing scenarios, including intelligent generation of print ads, social media user-driven marketing, intelligent e-commerce product detail pages, intelligent creative diversification and extension, and intelligent short video creation. The platform fully leverages the company's expertise in brand marketing, maintaining high-quality output while integrating AI, big data, and other intelligent technologies to provide clients with efficient, cost-effective, and high-quality intelligent creative solutions. INSIGH Tengine is used for developing AI-generated creative content generation and management, combining brand marketing content materials, brand user data, and brand marketing methodologies, and based on AI algorithms such as convolutional neural networks to produce high-quality AI-generated marketing content (AIGC).

[0085] In another embodiment of the present invention, the analysis conclusions are obtained by analyzing and comparing multimodal data through a virtual user model, including:

[0086] Data aggregation was performed on multimodal data using the InsightEngine method.

[0087] The data aggregation results of multimodal data are analyzed by performing at least one of the following: weighted preference difference index, relative sentiment intensity index, and risk assessment index, and the analysis results are generated.

[0088] Multiple versions of multimodal data across the time dimension were analyzed and compared to obtain the comparison results;

[0089] Multimodal data is analyzed and compared in the database based on the comparison results using a virtual user model.

[0090] In this embodiment of the invention, preferably, the analysis module analyzes and compares multimodal data through a virtual user model. The analysis process can be based on the output of the summary results from the summary module, or on the search results from the retrieval module, or it can directly analyze and compare the multimodal response samples of the feedback data. Alternatively, it can use the InsightEngine 5.0 system with the InsightEngine method to aggregate the feedback data, that is, to perform horizontal aggregation of multimodal response samples after aligning the feedback data with data that is consistent with its modality and cross-modal semantics. In the system, the system calculates various indicators such as sentiment scores, focus, and triggering factors, weighted preference differences, relative emotional intensity, and risk assessment.

[0091] Comparing multimodal data primarily involves optimizing the sensitivity to design differences across multiple product versions through adversarial training. For instance, for the same phone case design, where the multimodal data is obtained from market research, comparing it with other market research data on the same phone case design from historical databases can identify differences and further optimize the phone's design. Comparing multimodal data using virtual user models involves analyzing and comparing data obtained from virtual user models accessed in the retrieval module. These virtual user models can share the same user criteria as those used in the research that yielded the feedback data. For example, by defining the criteria for virtual user models, several models targeting the 20-30 age group can be selected, all based on data obtained from research on the same product (the same phone case packaging design).

[0092] Analyzing and comparing multiple versions of multimodal data along the time dimension involves considering the different order of appearance of phone case design version A during market research, historical research data on the same product and version from a historical database, or data from research conducted using a virtual user model. To avoid discrepancies in multimodal data due to different orderings, the number of times phone case design version A appears first among other design versions (such as phone case design versions B / C) is consistent with the number of times it appears last. The resulting multimodal data for multiple versions obtained through this time dimension represents the multimodal count of phone case design version A when it is in the first order. The analysis compares the multimodal data of design version A of the phone case when it was second in the rankings. This helps avoid bias from users who may have preconceived notions based on the order in which different versions of the product were released. The multimodal data of design version A can be analyzed and compared between two versions of the product from the same person within the same testing period, or between different people analyzing the same product at different time points. This process allows for the discovery of business insights into the product, predicting market acceptance trends for various products through before-and-after comparison techniques, and helping companies avoid over 80% of potential design flaws.

[0093] As one embodiment of the present invention, the psychological state modeling includes a digital mapping model of the Big Five personality model and a quantitative evaluation matrix model of Maslow's hierarchy of needs theory.

[0094] The behavioral prediction modeling includes a group relationship network model based on the GraphSAGE algorithm and a cultural value orientation analysis model.

[0095] In this embodiment of the invention, preferably, the digital mapping model of the Big Five personality model is a data-driven computational framework. It analyzes data samples (such as text, consumption records, and social interactions) to predict indicators in the five dimensions of openness, conscientiousness, extraversion, agreeableness, and neuroticism, and transforms them into signals that can be captured from numbers. By identifying stable behavioral patterns related to personality traits in the data samples through the digital mapping model of the Big Five personality model, relatively stable psychological traits can be inferred, and then the psychological states that may occur in specific situations can be predicted. For example, it can predict the psychological state of users regarding product design, and the predicted psychological states may be curiosity, excitement, anxiety, resistance, etc. The quantitative assessment matrix model of Maslow's hierarchy of needs is a dynamic model for assessing user needs. It maps user behavior data at different points in time to the seven levels of Maslow's hierarchy of needs and quantifies the "satisfaction" or "activation intensity" of each level's needs. This model allows for the identification of psychological motivations within data samples, precise market positioning and product design, and analysis of social trends. By defining behavioral indicators for each level of needs in the data sample, constructing a quantitative matrix, and finally predicting psychological motivation states, it can also combine the digital mapping model of the Big Five personality model with the quantitative assessment matrix model of Maslow's hierarchy of needs to construct a psychological panorama of users in market research from data samples, generating a more comprehensive virtual user model for predicting psychological states.

[0096] In the GraphSAGE-based group relationship network model, GraphSAGE is an inductive learning algorithm for graph neural networks. It generates embedding representations of nodes in the graph by sampling data samples and aggregating features of neighboring nodes. In the group relationship network model, individuals in the data samples are represented as nodes in the graph, and social relationships between individuals (such as friends, colleagues, and family) are represented as edges. The GraphSAGE model learns the embedding of each node, which captures the node's position and local neighborhood structure within the social network. The model's role is to capture the social influences on individuals, as individual behavior is influenced not only by their own characteristics but also by their social circles. Through GraphSAGE, we can obtain the embedding representation of each individual in the social network. This representation can be used to predict individual behavior, including the social influences they experience. This can be achieved by concatenating the data samples with the social embedding vectors generated by GraphSAGE, inputting this fused vector into a classifier (such as a fully connected neural network), and then outputting a behavioral prediction of the target audience. The cultural value orientation analysis model is a psychometric model that infers users' intrinsic values ​​and cultural tendencies by analyzing their behavior, consumption, and content preferences. Based on psychological theories such as Hofstede's cultural dimensions theory and the Schwarz value scale, the model identifies user data characteristics in a data sample and predicts how users will choose in different situations based on their stable value orientations, thus maintaining consistency between their behavior and values. First, it infers value orientations from the data characteristics of the data sample, mapping corresponding behavioral characteristics to various cultural value dimensions and outputting value orientation vectors, before making behavioral predictions. Furthermore, it can co-model with a group relationship network model based on the GraphSAGE algorithm. The two models provide strong behavioral prediction signals from the dimensions of "external social influence" and "internal value drive," respectively, making the generated virtual user model closer to real consumers. This expands the scope and data dimensions of market research and avoids weaknesses in theoretical support.

[0097] In one embodiment of the present invention, the verification of the results of the multimodal data and / or analysis module includes:

[0098] Input multimodal data and / or analysis results into the TrustGuard model;

[0099] The TrustGuard model and the virtual user model are used to verify the generation process and / or analysis conclusions of multimodal data and output the verification results.

[0100] In this embodiment of the invention, preferably, the verification module can be used to verify multimodal data and the analysis conclusions of the multimodal data. This can be achieved by calling a suitable virtual user model to verify the analysis conclusions. For example, if the analysis conclusion is that students prefer the version of the product design with a red cartoon pattern, then a virtual user model of the student group with the same psychological and behavioral consistency as real consumers can be called to verify the analysis conclusions. Alternatively, the module can conduct a credibility review of the analysis process or the generation process. Furthermore, the verification module includes, but is not limited to, verifying the analysis conclusions of the analysis module; it can also verify the summary results of the summary module, and can also verify... The system verifies the retrieval results of the authentication module, the retrieval process of the RAG module, and the vectorization process of the vectorization module on the preprocessed data. The TrustGuard model belongs to the TrustGuard system, a network access control system (NAC) and a dynamic credit evaluation model based on GNN (Graph Neural Network). Its main function is to assess and quantify trust in the field of cybersecurity to reduce potential risks. It combines traditional device authentication with risk-based dynamic access control. The TrustGuard system quantifies trust by incorporating factors affecting trust into the digital trust assessment and expression. It possesses characteristics such as subjectivity, dynamism, context awareness, asymmetry, and conditional transitivity, and is influenced by various factors, including context and the subjective and objective attributes of the trustor and the trusted party.

[0101] The validation module uses the TrustGuard model to verify the interpretability of models such as Multi-BERT-XL, Vision Transformer, Wav2Vec, and the vector space alignment module during vectorization processing. It assesses the clarity of the vectorization process, verifies the accuracy of these models using historical data, and verifies their stability on different datasets through cross-validation and adversarial testing. In the RAG module, validation is performed based on its internal storage, service, and update layers. Performance evaluation is also conducted on the structured vector data's retrieval process within the RAG module and the technical implementation methods used in the multimodal data generation workflow. When the retrieval module performs similarity searches between vector data from the vector database and multimodal data, vector data quality is assessed, verifying the reliability and breadth of the vector data sources, evaluating the source, diversity, and quality of the vector data used. There are issues such as sample bias, completeness of vector data, and timeliness. In the summary module, when using the SummaGPT module to analyze and extract features from associated multimodal data and data samples, the model is validated to verify the association and extraction processes. This validation process verifies the model's feasibility, the clarity and accuracy of the data analysis and feature extraction processes, and the stability of the module on different datasets through cross-validation and adversarial testing. This verifies the summary process and results of the summary module. In the analysis module, during the vectorization of psychological and sociological theoretical parameters, the data sources and reliability of these parameters are verified, along with the completeness and timeliness of the data. The process of combining and analyzing data using the InsightEngine method to obtain conclusions verifies the feasibility, accuracy, and robustness of the method. The conclusions obtained from the analysis are validated for logical contradictions and internal consistency. The analysis conclusions are compared with external information such as industry reports, authoritative research, and market trends to see if they are consistent. Furthermore, the basic common sense and business logic of the analysis conclusions can be verified.When using the SNA algorithm to model and generate virtual user models, the algorithm itself can be validated, along with the generated virtual user models. This validation process and model verification can include assessing transparency (interpretability, process clarity, etc.), model stability and data reliability for psychological state and behavioral prediction modeling, and verifying internal consistency (logical inconsistencies between attributes within the generated virtual user model). Comparing profiles of the same consumer at different time points can check for reasonable continuity and abrupt changes. Comparative testing can also be conducted. Personalized marketing based on the virtual user model can be implemented to observe whether the actual conversion rate matches expectations. Furthermore, a monitoring mechanism can be established to regularly evaluate the analytical conclusions obtained during market research and the predictive ability of the virtual user model. Profiles can be updated, and multi-dimensional or multi-process evaluations can be performed to verify the analytical conclusions and the credibility of the virtual user model. This ensures the reliability and stability of the final conclusions and the virtual user model, guaranteeing the stability and traceability of the results.

[0102] This invention also provides a market research method based on multi-source data integration and psychological analysis, the market research method based on multi-source data integration and psychological analysis including:

[0103] Obtain feedback data A for object one and feedback data B for object two;

[0104] Feedback data A and feedback data B are preprocessed, and the preprocessed feedback data is retrieved from the RAG knowledge base to generate multimodal response samples.

[0105] The analysis module is invoked to analyze and compare the response samples of feedback data A and feedback data B.

[0106] Call the verification module to verify the response samples and / or the results of the analysis and comparison;

[0107] The evaluation is based on the analysis and comparison results according to the evaluation system;

[0108] Based on the analysis and comparison results, output recommendations and / or optimization suggestions for Object 1 and Object 2.

[0109] In this embodiment of the invention, preferably, the feedback data A for object one and the feedback data B for object two are obtained. The market research method based on multi-source data integration and psychological analysis can be an application method of a market research system based on multi-source data integration and psychological analysis. This mainly involves inputting the sampled feedback data collected after surveying different objects or product advertisements into the application of the market research system based on multi-source data integration and psychological analysis. Specifically, the obtained feedback data can be sent to an AI large model, which drives multimodal data analysis of the feedback data A for object one and the feedback data B for object two. Then, the analysis module is called to analyze and compare the feedback data of different products (object one and object two) within the same testing period and / or the feedback data of object one and object two for the same test product. The system analyzes and compares feedback data, and generates analysis results based on the results of the analysis and comparison according to a multi-dimensional dynamic evaluation system. Based on the analysis results, it generates recommendations and / or optimization suggestions for Object 1 and Object 2. This can be a comparison report of multiple design versions of the product within the same testing period, or a comparison result of a single design version of the product at different time dimensions. The output can be the analysis results of recommendations and / or optimization suggestions presented in the form of graphics and text. The feedback data obtained for Object 1 and Object 2 includes, but is not limited to, feedback data from Object 1 and Object 2, and can also be feedback data from multiple objects. After testing, recommendations and optimization requirements are obtained. Through market research methods based on multi-source data integration and psychological analysis, a large amount of feedback data from materials or products that need to be tested can be analyzed, evaluated, and predicted.

[0110] Feedback data A for Object 1 and feedback data B for Object 2 can both be multi-source heterogeneous data, integrating interview transcripts, quantitative survey data, social media comments, eye-tracking data, and other data from internet users regarding product advertising versions. For example, feedback data A for user feedback on mobile phone case advertising design version A and feedback data B for design version B. The target audience can be advertising creatives for product categories such as advertising, key visual (KV), concept, and packaging. After the acquired multi-source heterogeneous data is input into the system, it is used to extract semantic features and sentiment tendencies through AI large-scale model tagging, training, natural language processing (NLP), and eye-tracking prediction technology. The sentiment tendency extraction of all feedback data is completed through cross-modal semantic fusion using a trained AI large-scale model. Multiple types of sampled feedback data can be preprocessed using the AI ​​large-scale model. The preprocessing process can be based on... In a market research system integrating multi-source data and psychological analysis, the processing module processes feedback data A from object one and feedback data B from object two, extracting semantic features and sentiment tendencies from the feedback data. The RAG knowledge base is a technical framework combining information retrieval and text generation. The pre-processed feedback data is vectorized and input into the RAG knowledge base in parallel. By calling vector data from the vector database, retrieval is performed to obtain multimodal response samples related to the product and the feedback data. The vector database is a historically semantically similar sample group established based on the feedback data, facilitating the formation of behavioral benchmarks and obtaining multimodal response samples related to the product. The multimodal response samples can be in various data formats, such as text, image, eye-tracking, or voice data.

[0111] Then, the evaluation and prediction of multimodal response samples related to product feedback data can be performed through any of the retrieval, summary, analysis, and verification modules in a market research system based on multi-source data integration and psychological analysis. This can be achieved by sequentially or in parallel using multiple modules. For example, the retrieval module can retrieve data from the corresponding product database or perform cross-database hybrid retrieval using the RAG knowledge base to predict preferences for multimodal response samples. The summary module can aggregate information from multiple databases to summarize response samples and predict preferences, or it can summarize the retrieval results from the retrieval module and predict preferences accordingly. The analysis module can analyze the semantic features and sentiment tendencies extracted from the response samples, comparing and contrasting response samples related to feedback data from different objects to obtain comparative analysis and preference predictions. Alternatively, it can directly compare and analyze the results of multi-dimensional information aggregation from the summary module using comparative techniques. For instance, it can analyze and conclude that users prefer the color of design version A of a phone case to the pattern design of design version B. The comparison process is related to the order in which users see the phone case design versions; the comparison process can adjust the time order in which users see different designs to ensure that the order in which the design versions of the same product appear remains consistent across different times. The verification module can review the credibility of information, features, and sentiment tendencies involved in the analysis and comparison process to ensure the stability and traceability of the results. It can also review the credibility of the analysis module's comparison process. Based on a preset evaluation system, the system performs multi-dimensional dynamic evaluation of the analysis and comparison results. For each dimension, the system can use a 0-1 rating or a percentage system to standardize the output of the analysis results. The scoring calculation for each dimension can be linked to structured data and model inference results provided by the AI ​​model. The system then generates full-dimensional performance analysis results or subdivided dimension weight analysis results through evaluation. It can also dynamically adjust the weight bias and generate recommendations and / or optimization suggestions for Object 1 and Object 2 presented in graphic and textual form. For example, users who value color matching can choose design version A, while optimization suggestions can combine the colors of phone case design version A with the patterns of phone case design version B to obtain a new design scheme.

[0112] In one embodiment of the present invention, the evaluation system includes at least one of user experience, functional performance and market performance, and each evaluation dimension includes several quantifiable indicators. The evaluation dimensions are used to score the response samples.

[0113] The evaluation dimension linking AI decoding model generates full-dimensional and / or sub-dimensional analyses of Object 1 and Object 2 by integrating at least one quantifiable indicator.

[0114] In this embodiment of the invention, preferably, the evaluation system can be a quantitative evaluation index covering user experience dimensions (ease of use, emotional resonance), functional dimensions (response speed, failure rate), and market performance dimensions (purchase intention, word-of-mouth dissemination). The multi-dimensional dynamic evaluation system evaluates and scores the results of comparative analysis of response samples. The AI ​​decoding model linked to the evaluation dimensions can provide structured data and model inference results for the calculation of each indicator value, thereby generating a full-dimensional analysis result including all evaluation dimensions or an analysis result for a specific evaluation dimension based on the comparative analysis results. The multi-dimensional dynamic evaluation system is an evaluation method that reveals the development trend and potential risks of a specific object by tracking and comparing data changes at different points in time. The evaluation system not only considers the current state of the object but also focuses on its historical performance and future trends, thus achieving comprehensive and dynamic analysis. It has real-time and dynamic adjustment capabilities, enabling rapid response to environmental or market changes and providing real-time decision-making references for managers or investors. The multi-dimensional dynamic evaluation system can be applied to multiple fields, such as medical equipment management and stock investment, helping to improve management efficiency and grasp market dynamics. The evaluation system consists of three dimensions: user experience, functional performance, and market performance. Each dimension contains several quantifiable indicators, which are then standardized and output by the system using a 0-1 rating or a percentage system.

[0115] User experience dimensions include usability scores (calculated from operation success rate and error rate), emotional resonance index (calculated from the fusion of emotional word frequency and eye-tracking metrics), and user satisfaction scores (calculated from semantic clustering of user feedback and word polarity scores). Functional performance dimensions include system metrics such as average response time, loading error rate, and bounce rate, which are collected through client-side logs. Market performance dimensions include intention scores (generated from questionnaires or semantic scores), purchase prediction scores (predicted through Transformer model regression), and social media buzz (such as the number of comments and repost rate). The calculation of each metric value relies on structured data and model inference results provided by the AI ​​decoding model. The system ultimately generates a comprehensive performance report by integrating multiple metrics through an evaluation engine, and can perform adjustable analysis of subdivided dimensional weights. It can also dynamically adjust metric weights based on reinforcement learning to adapt to different testing scenarios (such as new product testing vs. mature product optimization), and what is dynamically adjusted is the relative weight of each evaluation metric under different testing task scenarios. For example, when Object 1 and / or Object 2 are in the new product concept verification stage, the evaluation system will increase the weight of indicators such as "user first impression" and "emotional resonance"; while when Object 1 and / or Object 2 are in the optimization stage, the weight of "usability" and "performance stability" will be increased. This weight adjustment is driven by a reinforcement learning algorithm. The evaluation system uses the feedback results of historical test tasks as a reward function to guide the policy network to learn the optimal combination of indicators in different contexts. The weight adjustment follows these rules: first, it calls the initial empirical value according to the current task type, and then it performs gradient fine-tuning based on real-time test feedback (such as anomaly rate, user preference deviation). Ultimately, it achieves adaptive evaluation strategy optimization under different test objectives. Finally, it generates market research test results for Object 1 and / or Object 2 through an interpretable AI large model and outputs recommendations and / or optimization suggestions (e.g., when the object is product packaging, the optimization suggestion could be "changing the main color of the packaging to blue can increase purchase intention by 10%").

[0116] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0117] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0118] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0119] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A market research system based on multi-source data integration and psychological analysis, characterized in that, The market research system based on multi-source data integration and psychological analysis comprises: a processing module for processing feedback data of market research; an RAG module for generating multi-modal data of feedback data; a retrieval module for mixed retrieval of multi-modal data in a database to obtain data samples; a summary module for information aggregation and extraction of structured data features of multi-modal data; an analysis module for analyzing multi-modal data through a virtual user model to obtain analysis conclusions; a verification module for verifying multi-modal data and / or results of the analysis module; the multi-modal data of the feedback data comprises: vectorization processing of the feedback data processed by the processing module; input of the vectorization-processed feedback data into a multi-modal query module for multi-modal vector construction; application of a HNSW+PQ hybrid index to call vector data in a vector database and perform parallel retrieval with the vectorization-processed feedback data; embedding of the retrieved vector data related to the feedback data on a preset prompt word template through an adaptive load balancer; application of a Delta Update algorithm to generate multi-modal data related to the feedback data; the analysis conclusions obtained by analyzing the multi-modal data through the virtual user model comprise: matching and combining of the multi-modal data with the virtual user model through an InsightEngine method; psychological state and behavior prediction analysis of the combined results through an SNA algorithm to obtain analysis conclusions; the establishment steps of the virtual user model comprise: vectorization processing of psychological theory parameters and sociology theory parameters and input into a mapping matrix; input of the multi-modal data into the mapping matrix to combine the multi-modal data with the vectorization-processed psychological theory parameters and sociology theory parameters through an InsightEngine method; psychological state modeling and behavior prediction modeling of the combined results through an SNA algorithm to obtain at least one virtual user model. 2.The market research system based on multi-source data integration and psychological analysis of claim 1, wherein, the processing of the feedback data of market research comprises: division of the feedback data of market research into text data, image data and voice data; text processing of the text data through a regular expression library combined with a bidirectional LSTM time anomaly detection model; image processing of the image data through a median filter algorithm; noise filtering of the voice data through wavelet transform denoising; data alignment of the text data processed through text processing, the image data processed through image processing and the voice data filtered through noise filtering through a unified timestamp system, a space-time alignment engine and a spatial coordinate conversion matrix; quality verification of the feedback data after data alignment through a data credibility evaluation index. 3.The market research system based on multi-source data integration and psychological analysis of claim 1, wherein, the steps of vectorization processing of the processed feedback data comprise: linking of the text data processed through text processing to a Multi-BERT-XL model for vectorization processing to obtain vectorization text data; Link the image data processed to the Vision Transformer model for vectorization processing to obtain vectorized image data; Link the noise filtered voice data to the Wav2Vec model for vectorization processing to obtain vectorized voice data; Map the vectorized text data, vectorized image data and vectorized voice data through the vector space alignment module to obtain structured vector data. 4.The market research system based on multi-source data integration and psychological analysis of claim 1, wherein, The mixed retrieval of multi-modal data in the database obtains data samples, including: Applying the Hybrid-Retriever retrieval method to read and identify multi-modal data; Calling vector data and virtual user models in the vector database to search for similarity in the vector space based on the results of reading and identifying to obtain data samples consistent with the modal of multi-modal data. 5.The market research system based on multi-source data integration and psychological analysis of claim 1, wherein, The information aggregates and extracts structured data features of multi-modal data, including: Through the SummaGPT module and calling the virtual user model, data analysis and data feature extraction are performed on multi-modal data; The results of data feature extraction are dimensionally aggregated; The results of dimensional aggregation are structured to obtain structured data features. 6.The market research system based on multi-source data integration and psychological analysis of claim 1, wherein, The psychological state modeling includes a digital mapping model of the Big Five Personality Model and a quantitative evaluation matrix model of Maslow's Hierarchy of Needs Theory; The behavior prediction modeling includes a group relationship network model based on the GraphSAGE algorithm and a cultural value orientation analysis model. 7.The market research system based on multi-source data integration and psychological analysis of claim 1, wherein, The verification of multi-modal data and / or the results of the analysis module includes: Input multi-modal data and / or analysis conclusions into the TrustGuard model; Through the TrustGuard model and the virtual user model, the generation process of multi-modal data and / or the analysis conclusion process is verified and the verification result is output.

Images