Satisfaction evaluation method and system based on multi-dimensional data fusion

Abstract

The invention discloses a satisfaction assessment method and system based on multidimensional data fusion, and belongs to the technical field of satisfaction assessment, and the method specifically comprises the steps: obtaining multisource heterogeneous data, carrying out the preprocessing and feature extraction of the multisource heterogeneous data, and constructing a multimodal feature map with a timestamp, the method comprises the steps of scoring the current credibility of the multi-source heterogeneous data based on a credibility scoring strategy, generating a satisfaction expression vector based on a multi-modal characteristic spectrum and credibility scoring, generating quantized scoring values based on the satisfaction expression vector, and fusing all the quantized scoring values to obtain a satisfaction score. The generated quantized score value is obtained by the satisfaction expression vector through heterogeneous channel mapping, logic cluster scoring, symbol scoring sequence construction and decoding; the method effectively improves the evaluation precision of the complex user behavior state, and is suitable for various intelligent services and user feedback prediction scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of satisfaction evaluation, in particular to a satisfaction evaluation method and system based on multi-dimensional data fusion. BACKGROUND

[0002] As a core link in service quality management, human-computer interaction optimization and user experience research, satisfaction evaluation is widely used in customer service, e-commerce platforms, online education, intelligent terminals, smart cities and other scenarios. Existing satisfaction evaluation methods are mostly based on single modal data input, such as questionnaire survey, text evaluation, rating label or user click behavior, and use traditional machine learning models or rule-based expert systems for scoring and classification.

[0003] With the diversification of user interaction and the development of data collection technology, user feedback during the use of services or products has shown characteristics of multi-modal, weak structure, high noise and asynchronization, including text comments, voice tone, physiological perception signals (such as heart rate, skin electricity, facial expression), behavior path, historical click stream, etc. These data sources have problems such as non-uniform structure, asynchronous time stamp and unstable quality, which bring great challenges to satisfaction evaluation.

[0004] The existing technology has the following significant shortcomings: single modal fusion dimension, lack of credibility, limited expression ability of scoring model, etc. SUMMARY

[0005] In view of the shortcomings of the prior art, the application provides a satisfaction evaluation method and system based on multi-dimensional data fusion.

[0006] To achieve the above purpose, the application provides the following technical solutions:

[0007] The satisfaction evaluation method based on multi-dimensional data fusion comprises:

[0008] Obtaining multi-source heterogeneous data;

[0009] Preprocessing and feature extraction are performed on the multi-source heterogeneous data, a multi-modal feature map is constructed, and the multi-modal feature map has a time stamp;

[0010] Based on the credibility scoring strategy, the current credibility of the multi-source heterogeneous data is scored;

[0011] Based on the multi-modal feature map and the credibility score, a satisfaction representation vector is generated;

[0012] The satisfaction representation vector is used to generate a quantized score value, all quantized score values are fused to obtain a satisfaction score, and the satisfaction representation vector is mapped through a heterogeneous channel, a score logic cluster, a symbol score sequence is constructed, and decoding is performed.

[0013] Specifically, the multi-source heterogeneous data is preprocessed and feature extracted, and a multi-modal feature graph is constructed, the multi-modal feature graph has a timestamp and includes:

[0014] The collected multi-source heterogeneous data is preprocessed, including time synchronization processing, data normalization processing, and word segmentation;

[0015] The preprocessed multi-source heterogeneous data is input into a modal feature extraction network to extract multi-modal features, including structured features, semantic features, acoustic features, and physiological features;

[0016] Based on the multi-modal features and the corresponding timestamps, a multi-modal feature graph is constructed, each node in the multi-modal feature graph corresponds to a feature set of a type of data in a preset time slice, and the edge weight represents the time sequence or correlation information between the multi-modal features.

[0017] Specifically, the current credibility of the multi-source heterogeneous data is scored based on the credibility score strategy, including:

[0018] Based on historical data statistics, the completeness of the multi-source heterogeneous data in a specific time window is measured to obtain an integrity score;

[0019] Based on feature stability analysis, the coefficient of variation and deviation trend of the multi-modal features in a continuous time window are calculated to generate a feature volatility rate index;

[0020] The multi-source heterogeneous data at the current time is input into an error inversion model, and a credible residual error is calculated, the error inversion model generates a prediction confidence interval according to the historical input-output deviation;

[0021] According to the integrity score, the feature volatility rate index, and the credible residual error, a modal credibility score function is constructed;

[0022] The credibility of each multi-source heterogeneous data at the current time slice is quantitatively scored to obtain a credibility score.

[0023] Specifically, the multi-source heterogeneous data at the current time is input into an error inversion model, and a credible residual error is calculated, including:

[0024] A training sample set of the error inversion model is constructed, the training sample set is composed of the deviation between the multi-modal features in a historical time window and the corresponding satisfaction model output;

[0025] Based on the error inversion model training sample, a regression network is used to model the input error distribution, and the predicted error mean and confidence interval boundary corresponding to each multi-source heterogeneous data input are output.

[0026] In the current time slice, multi-modal features are obtained and input into the trained error inversion model to generate the current prediction error interval.

[0027] The actual output of the model at the current time is compared with the prediction error interval, and the deviation is calculated, which is defined as the credible residual.

[0028] Specifically, the satisfaction representation vector is generated based on the multi-modal feature graph and the credibility score, including:

[0029] The feature nodes corresponding to each multi-source heterogeneous data in the current time slice are extracted from the multi-modal feature graph, and the credibility scores associated with the time slice are collected.

[0030] The modal feature vector and its corresponding credibility score are processed by weight association, and the weight association processing includes embedding the credibility score as an attention factor into the feature representation.

[0031] The weighted modal feature vector is input into the time-aware fusion network to obtain the fusion feature, and the time-aware fusion network fuses different weighted modal feature vectors in time sequence based on the time gating mechanism.

[0032] The context association coding of the fusion feature is performed through the cross-modal attention mechanism to capture the time sequence dependency relationship between different source heterogeneous data.

[0033] A fixed-dimensional satisfaction representation vector is generated.

[0034] Specifically, the weighted modal feature vector is input into the time-aware fusion network to obtain the fusion feature, including:

[0035] The weighted modal feature vector is arranged in time sequence, and the continuous time period is divided according to the time stamp.

[0036] A time sequence window is set for the weighted modal feature vector in each time period, and the input sequence between adjacent time slices is extracted.

[0037] A time gate is introduced in the time sequence window to filter the time state of the weighted modal feature vector, and the time gate generates a state gate value according to a preset time correlation function.

[0038] The weighted modal feature vector in each time window is recursively aggregated using a loop calculation to generate a time slice-level intermediate fusion representation.

[0039] Stacking the intermediate fusion representations at all time slice levels, a fused feature is obtained.

[0040] Specifically, the satisfaction representation vector is used to generate a quantized score value, and all quantized score values are fused to obtain a satisfaction score, including:

[0041] The generated satisfaction representation vector is copied to multiple score channels, including an emotion discrimination path, a behavior mapping path, a modal adversarial path, and a time sequence memory path.

[0042] In the emotion discrimination path, the satisfaction representation vector is projected and activated to extract signal dimensions related to the user's subjective attitude.

[0043] In the behavior mapping path, a mapping relationship between the satisfaction representation vector and the historical behavior label is established to generate user behavior bias features.

[0044] In the modal adversarial path, the proportion of each dimension data in the final representation is determined, and the weight vector is back propagated to each dimension data to refit the weight vector.

[0045] In the time sequence memory path, the sequence deviation between the historical satisfaction vector and the current satisfaction representation vector is captured to calculate a time stability encoding vector.

[0046] A satisfaction evaluation model is constructed based on the score channel to obtain a satisfaction score value.

[0047] Specifically, the satisfaction evaluation model is constructed based on the score channel to obtain a satisfaction score value, including:

[0048] The output results from the multiple score channels are mapped to a heterogeneous channel, including mapping the output results of different score paths to a discrimination space.

[0049] A score cluster is constructed, which is composed of parallel score sub-models, and each sub-model independently calculates based on different hypothetical score logic.

[0050] A symbolic score sequence is constructed for the score values output by each sub-model, which is combined in a non-numeric symbolic manner.

[0051] The symbolic score sequence is structurally compressed, indexed and path decoded to generate a quantized score value.

[0052] All quantized score values are fused to obtain a satisfaction score value.

[0053] The satisfaction evaluation system based on multi-dimensional data fusion is used for realizing the satisfaction evaluation method based on multi-dimensional data fusion, and comprises a data acquisition module, a graph construction module, a credibility scoring module, a vector generation module and a satisfaction evaluation module.

[0054] The data acquisition module is used for acquiring multi-source heterogeneous data.

[0055] The graph construction module is used for pre-processing and feature extraction of the multi-source heterogeneous data, and constructing a multi-modal feature graph.

[0056] The credibility scoring module is used for scoring the current credibility of the multi-source heterogeneous data based on a credibility scoring strategy.

[0057] The vector generation module is used for generating a satisfaction representation vector based on the multi-modal feature graph and the credibility score.

[0058] The satisfaction evaluation module is used for generating quantized score values based on the satisfaction representation vector, fusing all the quantized score values, and obtaining a satisfaction score.

[0059] Specifically, the satisfaction evaluation module comprises a sub-score unit, a heterogeneous channel mapping unit, a decoding unit and a satisfaction evaluation unit.

[0060] The sub-score unit is used for copying the generated satisfaction representation vector to multiple scoring channels and outputting sub-scores.

[0061] The heterogeneous channel mapping unit is used for performing heterogeneous channel mapping on the output results from the multiple scoring channels.

[0062] The decoding unit is used for constructing a score cluster, constructing a symbol score sequence for the score values output by each sub-model, and performing structure compression, index transformation and path decoding on the symbol score sequence to generate quantized score values.

[0063] The satisfaction evaluation unit is used for weighting and fusing all the quantized score values to obtain a satisfaction score value.

[0064] Compared with the prior art, the present application has the following advantages:

[0065] The application provides a satisfaction evaluation method and system based on multi-dimensional data fusion, which collects multi-source heterogeneous data of users, constructs a multi-modal feature map with a time stamp, quantifies the quality of each modal data by combining a dynamic credibility scoring mechanism, generates a unified satisfaction representation vector by using a time perception and cross-modal fusion network, and comprehensively evaluates the satisfaction by fusing a multi-scoring channel and a symbolic trajectory. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 A satisfaction evaluation method flowchart based on multi-dimensional data fusion is provided for the application.

[0067] Figure 2 A satisfaction representation vector generation flowchart is provided for the application.

[0068] Figure 3 A satisfaction evaluation flowchart is provided for the application.

[0069] Figure 4 A satisfaction evaluation system architecture diagram based on multi-dimensional data fusion is provided for the application. DETAILED DESCRIPTION

[0070] The application will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the application. These are within the scope of protection of the application.

[0071] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with reference to the drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0072] It should be noted that the features in the embodiments of the application can be combined with each other without conflict, and are within the scope of protection of the application. In addition, although the functional modules are divided in the device schematic diagram, and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order from the module division in the device or the order in the flowchart. In addition, the "first", "second", "third" and the like used in the application do not limit the data and execution order, but only distinguish the same items or similar items with basically the same function and effect.

[0073] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application. All publications, patent applications, patents, figures, and other references mentioned in this specification are herein incorporated by reference in their entirety for the teachings relevant to the sentence in which the reference is presented.

[0074] Embodiment 1

[0075] Referring to Figures 1-3 An embodiment provided by the present application: a satisfaction evaluation method based on multi-dimensional data fusion, comprising the following specific steps:

[0076] Step S1: acquiring multi-source heterogeneous data.

[0077] In a typical user interaction scenario, multi-source heterogeneous data generated by users is collected through the joint collection of the server and the edge, including but not limited to the following data types: structured behavior data such as clicks, dwell time, bounce rate; text feedback data such as comments, questions and answers, and dialogue records; audio signal data such as voice content, tone characteristics, and speech rhythm; physiological perception data such as heart rate fluctuations, skin electrical changes, and eye movement trajectories; and environmental context data such as geographic location, use time period, and device type.

[0078] Step S2: preprocessing and feature extraction of the multi-source heterogeneous data, and constructing a multi-modal feature graph with a timestamp.

[0079] The specific steps of step S2 are:

[0080] Step S201: preprocessing the collected multi-source heterogeneous data, including time synchronization processing, data normalization processing, and word segmentation.

[0081] It should be noted that the preprocessing is processed by existing means, which will not be described in detail here.

[0082] Step S202: inputting the preprocessed multi-source heterogeneous data into a modal feature extraction network to extract multi-modal features, including structured features, semantic features, acoustic features, and physiological features.

[0083] Step S203: constructing a multi-modal feature graph based on the multi-modal features and their corresponding timestamps, wherein each node in the multi-modal feature graph corresponds to a feature set of a type of data in a preset time slice, and the edge weight represents the time sequence or correlation information between the multi-modal features.

[0084] In the embodiment, the feature vectors of different modalities are organized as a graph structure to construct a multi-modal feature graph; specifically, each time slice is defined as a group of slice windows of the graph, and in each time slice, the structured behavior features, semantic text features, speech emotion features, and physiological signal features are respectively regarded as nodes of different modalities, each node records the feature set of the modality at the current time slice, and the time synchronization and type separation management of heterogeneous modalities are realized through the modality identifier and timestamp of the node.

[0085] In the graph construction, the edge weight needs to be calculated. On the one hand, for different modality nodes in the same time slice, the weight of the edge is determined by a similarity function between modalities, a collaborative change rate, or an empirical mutual information, to represent the degree of horizontal association between modalities; on the other hand, for nodes of the same modality in adjacent time slices, a longitudinal edge connection is established according to the time sequence continuity, feature change trend, or residual stability, to represent the time evolution path.

[0086] Step S3: Based on the credibility scoring strategy, the current credibility of the multi-source heterogeneous data is scored.

[0087] The specific steps of step S3 are:

[0088] Step S301: Based on historical data statistics, the integrity of the multi-source heterogeneous data in a specific time window is measured to obtain an integrity score.

[0089] In the embodiment, a sliding time window is set, and the data collection of each modality is counted in the time period, and the missing proportion, sampling frequency consistency, and collection delay are comprehensively calculated; for structured behavior data, the availability is evaluated by the field filling rate in consecutive time slices; for audio modalities, the frame loss ratio and effective speech ratio of speech paragraphs can be detected; for physiological signal data, whether the signal sampling meets the minimum time domain sampling threshold and whether there are typical abnormalities such as long-time flat value or high-frequency peak are counted.

[0090] Step S302: Based on feature stability analysis, the coefficient of variation and deviation trend of multi-modal features in a continuous time window are calculated to generate a feature volatility index.

[0091] In the embodiment, a fixed-length time window is drawn, and the modality feature vector in each time slice in the window is extracted; for numerical features, the coefficient of variation of the feature sequence in the window is calculated to measure the relative fluctuation degree in a short time range, and the coefficient of variation is defined as the ratio of the standard deviation to the mean, reflecting the numerical distribution dispersion of the modality in the time period.

[0092] Further, to capture the directional shift of feature evolution, the difference sequence between adjacent time slices is calculated, and its trend is fitted, including sliding linear regression slope estimation, trend score normalization processing, etc.; if the features of a certain mode show unstationary characteristics such as one-sided rapid rise, sudden drop or frequent reversal in continuous time, it is marked as a high-volatility mode, and on this basis, the coefficient of variation and the trend shift measure are fused to generate a feature volatility rate index.

[0093] Step S303: inputting the multi-source heterogeneous data at the current time into an error inversion model, and calculating a credible residual error, wherein the error inversion model generates a prediction confidence interval according to historical input-output deviation.

[0094] The specific steps of step S303 are as follows:

[0095] Step S3031: constructing an error inversion model training sample set, wherein the training sample set is composed of the deviation between the multi-modal features in the historical time window and the corresponding satisfaction model output.

[0096] In the embodiment, the multi-modal feature data in the historical time window is collected as the input part, the output result after being processed by the satisfaction evaluation model in the period is extracted, and is compared with the corresponding true label (such as user actual score, behavior feedback, etc.), and the prediction deviation is calculated; the deviation is defined as the residual error between the model output and the reference label.

[0097] Further, by constructing a feature-residual pair, a mapping relationship between the modal input and the prediction error is established, and the sample set covers multiple time periods, multiple user instances and multiple modal combinations, and has error representation capability across modalities and time periods.

[0098] Step S3032: based on the error inversion model training sample, using a regression network to model the input error distribution, and outputting the prediction error mean and confidence interval boundary corresponding to each multi-source heterogeneous data input.

[0099] In the embodiment, the input is the multi-modal feature vector in the historical time window, and the output is the corresponding prediction residual value, which is used to depict the distribution characteristics of the deviation, not only to fit a single residual value, but also to output the expected value and distribution boundary of the prediction error through multi-objective regression, such as upper and lower confidence limits or quantile range.

[0100] In the model training stage, the prediction output is supervised and optimized using the label composed of the historical residual value; after the final training is completed, when a new input feature is given, the model returns the mean estimation and confidence interval boundary of the potential prediction deviation, which is used to evaluate the error risk corresponding to the current input.

[0101] Step S3033: Obtain the multi-modal feature at the current time slice, and input it into the trained error inversion model to generate the current prediction error interval.

[0102] In this embodiment,

[0103] Step S3034: Compare the actual output of the model at the current time with the prediction error interval, and calculate the deviation, which is defined as the credible residual.

[0104] In this embodiment, under the current time slice, the feature vectors of each modality after standardization and feature extraction are extracted, and are spliced synchronously according to the time stamp to form a multi-modal joint input sample, which is input into the trained error inversion model; the model has learned the probability mapping relationship between different modal combinations and their corresponding prediction residuals through historical samples, and can infer the possible prediction error behavior based on the input features.

[0105] In the reasoning process, the model performs forward propagation on the joint feature vector, and outputs the corresponding error estimation value and its confidence boundary; the error estimation value can be understood as the expected value of the deviation of the satisfaction prediction result from the true reference label under the current feature state; the confidence boundary reflects the upper and lower limits of the deviation interval, that is, the maximum acceptable error range that the model may produce under the current input state.

[0106] Step S304: Construct a modality credibility scoring function based on the integrity score, feature fluctuation rate indicator and credible residual.

[0107] In this embodiment, based on the above three types of input variables, a combined scoring function is constructed, which adopts a normalized weighted structure or an interval mapping model based on fuzzy logic; each input dimension is mapped to the common credibility score interval after normalization conversion; further, different weight parameters are configured according to the modality type, reflecting the experienced importance of the contribution degree of each modality to the satisfaction evaluation model; in the combination stage, a nonlinear function is introduced to enhance the response ability of the scoring function to extreme fluctuations and confidence shifts.

[0108] Step S305: Quantitatively score the credibility of each multi-source heterogeneous data at the current time slice to obtain the credibility score.

[0109] In this embodiment, for each type of multi-source heterogeneous data at the current time slice, the scoring function is called to calculate the input state item by item, and the quality state of each modality feature is mapped to a continuous value in the [0, 1] interval, which is used to represent the current credibility.

[0110] In the quantification process, each index is given a corresponding weight through the function structure, and weighted synthesis is performed; this quantification process outputs a single credibility score for each modality at each time slice.

[0111] As shown in Figure 2 Step S4: generating a satisfaction representation vector based on the multi-modal feature graph and the credibility score.

[0112] The specific steps of step S4 are as follows:

[0113] Step S401: extracting the feature nodes corresponding to the current time slice of each multi-source heterogeneous data from the multi-modal feature graph, and collecting the credibility score associated with the time slice.

[0114] Step S402: performing weight association processing on the modal feature vector and its corresponding credibility score, the weight association processing including embedding the credibility score as an attention factor into the feature representation.

[0115] In this embodiment, the original feature vector of each modality at the current time slice is subjected to a consistent dimension transformation, so as to meet the multiplicative requirement between the attention factor; then the credibility score is expanded into a weight vector consistent with the feature dimension, and the original feature vector is subjected to element-by-element weighting through the broadcast mechanism; this operation does not change the representation structure of the original feature, but adjusts the amplitude of its components dynamically.

[0116] Step S403: inputting the weighted modal feature vector into a time-aware fusion network to obtain a fusion feature, the time-aware fusion network being based on a time gating mechanism to fuse different weighted modal feature vectors in time sequence.

[0117] The specific steps of step S403 are as follows:

[0118] Step S4031: arranging the weighted modal feature vectors in time sequence, and dividing continuous time periods according to the time stamp.

[0119] In this embodiment, the full time period sequence is divided into multiple continuous and non-overlapping time periods, each time period serving as a local processing unit of the fusion network; the division principle can be set according to absolute time interval, sliding window strategy or key event trigger point, to ensure that the data within each time period has strong correlation, while the data between different time periods is relatively independent.

[0120] Step S4032: setting a time sequence window for the weighted modal feature vectors in each time period, and extracting the input sequence between adjacent time slices.

[0121] In this embodiment, after completing the time period division, a local timing context is constructed within each time period to capture the dynamic change trend of the modal feature under the fine-grained time slice; a sliding timing window is set within each time period to extract the modal feature sequence between consecutive time slices, the window slides step by step within a set length and assembles the weighted feature vectors of adjacent time points to generate an input sub-sequence with short-time dependence structure.

[0122] Step S4033: Introducing a time gate within the timing window to perform time state filtering on the weighted modal feature vector, and the time gate generates a state gate value according to a preset time correlation function.

[0123] In this embodiment, a time gate mechanism is introduced to perform state filtering on the input sequence within each timing window to model the importance difference of the modal feature to time and control the information flow path; the basic function of the time gate is to give adjustable passing weights to the modal features of different time slices, which can dynamically select which information of time points to retain and which fluctuations or abnormal values of time points to suppress.

[0124] The generation of the time gate value depends on a class of preset time correlation functions, which are usually constructed according to the position of the time slice, the relative distance from the center of the window, the feature change gradient, the modal lag, etc.; the function is of linear decay type, Gaussian concentration type or periodic activation type, and its output value changes in the interval [0, 1], corresponding to the activation degree of the current time slice feature.

[0125] Step S4034: Using recurrent calculation to recursively aggregate the weighted modal feature vectors in each time window to generate the intermediate fusion representation at the time slice level.

[0126] In this embodiment, the recurrent calculation is sequentially unfolded in the time dimension, and the input vector of the current time slice and the state vector of the last time slice are jointly transformed through the introduction of a recurrent unit to generate the hidden state at the current time; the hidden state not only encodes the information of the current modal feature, but also inherits the historical memory of the last step, so that the sequence context is gradually compressed into the state vector at the time slice level.

[0127] Step S4035: Stacking all the intermediate fusion representations at the time slice level to obtain the fusion feature.

[0128] In the embodiment, after completing the recursive aggregation in each time window, a series of intermediate fusion representations at the time slice level are obtained, each of which carries the local dynamic characteristics of the multi-modal features in the time period, and then a fusion input sequence with global time awareness is constructed, the intermediate representations of all time slices are stacked in chronological order to form a unified fusion feature matrix. The stacking operation is essentially feature splicing along the time axis, which preserves the sequence structure while maintaining the consistency of the modal.

[0129] Step S404: context-related encoding of the fusion features through a cross-modal attention mechanism to capture the temporal dependency between different source heterogeneous data.

[0130] In the embodiment, for any modal or time slice feature, the influence proportion of each modal feature on the final representation is dynamically adjusted by calculating the association weight between it and other modal features to identify the most informationally valuable interaction path in the context environment.

[0131] During operation, an attention query vector is constructed, the representation of the current modal at a specific time slice is selected as the query, and the feature vectors of other modalities are selected as the key and value inputs; the generation of attention weight is driven by a similarity function, and the attention score is calculated by vector dot product or high-dimensional projection; the higher the weight, the more critical the information provided by the modal in the current context; the information from other modalities is integrated by weighted summation to form a context-enhanced representation.

[0132] Step S405: generating a satisfaction representation vector with fixed dimensions.

[0133] In the embodiment, the sequence pooling mechanism is used to compress the time dimension, and linear mapping or multi-layer perception transformation is introduced to align the inter-modal dimension. For modal features with asynchronous or redundant differences, the weight distribution in the pooling stage is adaptively adjusted by the gating mechanism to strengthen the proportion of key modal in the final representation, and the output dimension is constrained to be fixed length.

[0134] Step S5: generating a quantized score value based on the satisfaction representation vector, fusing all quantized score values to obtain a satisfaction score, wherein the generation of the quantized score value is obtained by heterogeneous channel mapping, score logic cluster, constructing a symbolic score sequence, and decoding from the satisfaction representation vector.

[0135] As shown in Figure 3 , the specific steps of step S5 are:

[0136] Step S501: copying the generated satisfaction representation vector to a plurality of score channels, the score channels including an emotion discrimination path, a behavior mapping path, a modal adversarial path, and a time sequence memory path.

[0137] Step S502: In the emotion discrimination path, the satisfaction representation vector is projected and activated to extract the signal dimension related to the user's subjective attitude.

[0138] In this embodiment, in the emotion discrimination path, the original representation vector is reconstructed in dimension by using the emotion label driven attention weight matrix, so as to extract the feature dimension containing subjective emotional information;

[0139] The calculation method of the activation projection is that, in the pre-training stage, a mapping matrix from the representation space to the emotion label space is learned by constructing an emotion perception sub-model; in the inference stage, the matrix performs linear projection on the input satisfaction representation vector, and simultaneously cooperates with the activation function to perform nonlinear adjustment on the result, so as to form a sub-vector with enhanced emotion correlation. This path can explicitly extract the emotional component in the user's satisfaction.

[0140] Step S503: In the behavior mapping path, the mapping relationship between the satisfaction representation vector and the historical behavior label is established to generate the user behavior bias feature.

[0141] In this embodiment, in the behavior mapping path, the current representation vector is taken as a high-dimensional representation of the user's current interaction state, and a multi-layer nonlinear relationship between the current representation vector and the historical behavior label is fitted by a supervised learning method, so as to infer the user's behavior tendency in a similar scenario.

[0142] Specifically, a set of historical behavior labels is constructed, and the label types can include click frequency, interaction depth, dwell time, complaint record or repurchase probability, etc.; these labels are taken as target variables, and a regression or classification sub-network input by the current satisfaction representation vector is designed to predict the behavior index or behavior category; in the training stage, the network learns the mapping function between the input representation and the behavior label; in the inference stage, the function outputs a set of continuous or discrete values as the behavior bias feature, which is used to reflect the projection result of the current satisfaction state in the user behavior space.

[0143] Step S504: In the modal confrontation path, the proportion of each dimension data in the final representation is judged, and the weight vector is re-fitted by reverse propagation to each dimension data.

[0144] In this embodiment, in the modal confrontation path, the structural bias of multi-modal features in the satisfaction representation vector is identified, and the contribution proportion of each modality is dynamically adjusted, and a modal discrimination mechanism and an adversarial back propagation are introduced; specifically, a modal discriminator structure is designed, the discriminator receives the final satisfaction representation vector as input, and outputs a modal probability distribution to characterize the dominant degree of each modality in the representation composition.

[0145] When the modal discriminator completes the forward discrimination, that is, the proportion of each modal signal in the current representation has been estimated, in the backward propagation stage, the error gradient of the discriminator is not optimized, but is transferred to the feature weight distribution module that generates the representation; finally, through the modal confrontation mechanism, the fusion proportion of different source data in the representation vector is dynamically controlled, which is especially suitable for real scenarios with fluctuating multi-source data quality or modal sparsity.

[0146] Step S505: In the time memory path, the sequence deviation between the historical satisfaction vector and the current satisfaction representation vector is captured, and the time stability encoding vector is calculated.

[0147] In the embodiment, a sequence of satisfaction representation vectors within a time window is extracted from the user behavior history as a baseline reference path; then the satisfaction representation vector generated at the current time is analyzed for sequence level difference with the historical sequence, and a recursive structure or time comparison encoder is used to calculate its deviation trajectory in the representation space. The deviation includes two parts: one is the amplitude difference, which is used to reflect the degree of change in the current satisfaction intensity; the other is the directional deviation, which is used to capture the reversal or enhancement of semantic trend.

[0148] Step S506: Construct a satisfaction evaluation model based on the scoring channel to obtain a satisfaction score value.

[0149] The specific steps of step S506 are:

[0150] Step S5061: Perform heterogeneous channel mapping on the output results from multiple scoring channels, which includes mapping the output results of different scoring paths to a discrimination space respectively.

[0151] In the embodiment, the channel level mapping function projects the output results of each scoring path to a unified constructed discrimination space respectively; the construction principle of the discrimination space is to ensure that different paths can express their biased scoring basis in a distinguishable but coexisting manner in the space; specifically, an independent mapping module such as a linear transformer or a nonlinear decoupling encoder is introduced for each path output to capture the internal structure of the channel and compress its dimension, so that the output falls into a multi-channel representation set under the same scoring semantics.

[0152] Through the mapping process, the scoring sub-results with different sources and structures are normalized under the same discrimination framework, enhancing the logical separability and multi-path consistency of the final scoring result.

[0153] Step S5062: Construct a scoring cluster, which is composed of parallel scoring sub-models, each of which is independently calculated based on different hypothetical scoring logic.

[0154] In this embodiment, the score assumption can cover the following types: the emotion dominant score assumption emphasizes the importance of the emotion dimension in the current representation; the behavior driven score assumption focuses on the mapping relationship between past behavior and current score; the modal robust score assumption prefers inter-modal consistency as the evaluation standard; the time stable score assumption emphasizes the continuity of time trends on the impact of score deviation; each sub-model selects different depths or attention mechanisms in structure, allowing it to focus on capturing the score structure in a specific dimension or specific mode.

[0155] Through the score cluster structure, the same input can obtain multiple score outputs with reasoning differences, which are not redundant with each other, but a candidate score set constructed based on independent logical paths.

[0156] Step S5063: Construct a symbolic score sequence for the score values output by each sub-model, and the symbolic score sequence is combined in a non-numeric symbolic manner.

[0157] In this embodiment, the numerical output of each score sub-model is normalized to a unified score reference scale, and then mapped to a limited symbol set such as ↑, ↓, ≈, →, etc. according to the set interval division rule, which respectively represent score rising, falling, tending to be consistent, and existing drift, etc. At the same time, the sequence is constructed according to the relative position relationship between the score values, for example, arranged in the order of the score path to form a score trajectory, or constructed in the form of deviation mode according to the score fluctuation change. The symbolic score sequence not only compresses the numerical complexity of the score space, but also has a structured difference expression capability.

[0158] Step S5064: Perform structure compression, index transformation, and path decoding on the symbolic score sequence to generate quantized score values.

[0159] In this embodiment, the structure compression first performs clustering and redundancy elimination operations on the symbolic sequence, merging adjacent repeated or trend consistent symbols into high-order segment identifiers, thereby reducing the sequence complexity. Then, the index transformation is performed, which identifies the score path type to which the current symbolic sequence belongs by constructing a pre-defined symbol-behavior mapping table or using a training obtained symbolic trajectory mode library, such as convergence type, divergence type, and oscillation type. Finally, the path decoding stage is entered, which maps the identified sequence structure to a set of quantized score rules, such as average interval regression, trend weight accumulation, or trajectory probability projection methods, to calculate the final numerical score.

[0160] Step S5065: Weighted fusion of all quantized score values to obtain the satisfaction score value.

[0161] Embodiment 2

[0162] Please refer to Figure 2In another embodiment, the application provides a satisfaction evaluation system based on multi-dimensional data fusion, comprising a data acquisition module, a graph construction module, a credibility scoring module, a vector generation module, and a satisfaction evaluation module.

[0163] The data acquisition module is configured to acquire multi-source heterogeneous data.

[0164] The graph construction module is configured to preprocess and extract features of the multi-source heterogeneous data, and construct a multi-modal feature graph.

[0165] The credibility scoring module is configured to score the current credibility of the multi-source heterogeneous data based on a credibility scoring strategy.

[0166] The vector generation module is configured to generate a satisfaction representation vector based on the multi-modal feature graph and the credibility score.

[0167] The satisfaction evaluation module is configured to generate quantized score values based on the satisfaction representation vector, fuse all the quantized score values, and obtain a satisfaction score.

[0168] The satisfaction evaluation module comprises a sub-score unit, a heterogeneous channel mapping unit, a decoding unit, and a satisfaction evaluation unit.

[0169] The sub-score unit is configured to copy the generated satisfaction representation vector to multiple scoring channels and output sub-scores.

[0170] The heterogeneous channel mapping unit is configured to perform heterogeneous channel mapping on the output results from the multiple scoring channels.

[0171] The decoding unit is configured to construct a scoring cluster, construct a symbolic score sequence for the score values output by each sub-model, and perform structural compression, index transformation, and path decoding on the symbolic score sequence to generate quantized score values.

[0172] The satisfaction evaluation unit is configured to weight and fuse all the quantized score values to obtain a satisfaction score value.

[0173] In addition, the parts of the above technical solutions in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive repetition.

[0174] The specific embodiments described above further illustrate the purposes, technical solutions, and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A satisfaction evaluation method based on multi-dimensional data fusion, characterized in that, The method comprises the following steps: acquiring multi-source heterogeneous data; preprocessing and feature extraction of the multi-source heterogeneous data, and constructing a multi-modal feature graph with a timestamp; scoring the current credibility of the multi-source heterogeneous data based on a credibility scoring strategy; generating a satisfaction representation vector based on the multi-modal feature graph and the credibility score; generating a quantized score value based on the satisfaction representation vector, and fusing all the quantized score values to obtain a satisfaction score, wherein the generation of the quantized score value is obtained by mapping the satisfaction representation vector through a heterogeneous channel, a scoring logic cluster, constructing a symbolic score sequence, and decoding.

2. The satisfaction evaluation method based on multi-dimensional data fusion according to claim 1, characterized in that, The preprocessing and feature extraction of the multi-source heterogeneous data, and the construction of the multi-modal feature graph with a timestamp, comprise the following steps: preprocessing the collected multi-source heterogeneous data, including time synchronization processing, data normalization processing, and word segmentation; inputting the preprocessed multi-source heterogeneous data into a modal feature extraction network to extract multi-modal features, including structured features, semantic features, acoustic features, and physiological features; constructing a multi-modal feature graph based on the multi-modal features and their corresponding timestamps, wherein each node in the multi-modal feature graph corresponds to a feature set of a type of data in a preset time slice, and the edge weight represents the time sequence or correlation information between the multi-modal features.

3. The satisfaction evaluation method based on multi-dimensional data fusion according to claim 2, wherein, The scoring of the current credibility of the multi-source heterogeneous data based on the credibility scoring strategy comprises the following steps: based on historical data statistics, measuring the collection integrity of the multi-source heterogeneous data within a specific time window to obtain an integrity score; based on feature stability analysis, calculating the coefficient of variation and deviation trend of the multi-modal features within a continuous time window to generate a feature volatility rate index; inputting the multi-source heterogeneous data at the current time into an error inversion model and calculating a credible residual error, wherein the error inversion model generates a prediction confidence interval based on the historical input-output deviation; constructing a modal credibility scoring function based on the integrity score, the feature volatility rate index, and the credible residual error; quantitatively scoring the credibility of each multi-source heterogeneous data at the current time slice to obtain a credibility score.

4. The satisfaction evaluation method based on multi-dimensional data fusion according to claim 3, wherein, The inputting of the multi-source heterogeneous data at the current time into the error inversion model and the calculation of the credible residual error comprise the following steps: constructing an error inversion model training sample set, wherein the training sample set is composed of the deviation between the multi-modal features within a historical time window and the corresponding satisfaction model output; based on the error inversion model training sample, using a regression network to model the input error distribution, and outputting the predicted error mean and confidence interval boundary corresponding to each multi-source heterogeneous data input; acquiring the multi-modal features at the current time slice and inputting them into the trained error inversion model to generate the current prediction error interval; comparing the actual output of the model at the current time with the prediction error interval to calculate the deviation degree, which is defined as the credible residual error.

5. The satisfaction evaluation method based on multi-dimensional data fusion according to claim 4, characterized in that, The generation of a satisfaction representation vector based on the multi-modal feature graph and the credibility score comprises the following steps: extracting the feature nodes corresponding to each multi-source heterogeneous data at the current time slice from the multi-modal feature graph, and collecting the credibility scores associated with the time slice; The modal feature vector is associated with the corresponding credibility score by weight processing, and the weight processing includes embedding the credibility score as an attention factor into the feature representation; The weighted modal feature vector is input into a time-aware fusion network to obtain a fusion feature, and the time-aware fusion network fuses different weighted modal feature vectors in time sequence based on a time gating mechanism; The fusion feature is contextually associated and encoded through a cross-modal attention mechanism to capture the time sequence dependency between different source heterogeneous data; A fixed-dimension satisfaction representation vector is generated.

6. The satisfaction evaluation method based on multi-dimensional data fusion according to claim 5, wherein, The inputting of the weighted modal feature vector into the time-aware fusion network to obtain the fusion feature includes: The weighted modal feature vector is arranged in time sequence and divided into consecutive time periods according to timestamps; A time window is set for the weighted modal feature vector in each time period, and an input sequence between adjacent time slices is extracted; A time gate is introduced in the time window to filter the time state of the weighted modal feature vector, and the time gate generates a state gate value according to a preset time correlation function; The weighted modal feature vector in each time window is recursively aggregated using a loop calculation to generate an intermediate fusion representation at a time slice level; All intermediate fusion representations at the time slice level are stacked to obtain the fusion feature.

7. The satisfaction evaluation method based on multi-dimensional data fusion according to claim 6, wherein, The generation of the quantized score value based on the satisfaction representation vector includes: The generated satisfaction representation vector is copied to multiple score channels, and the score channels include an emotion discrimination path, a behavior mapping path, a modal adversarial path and a time sequence memory path; In the emotion discrimination path, the satisfaction representation vector is projected and activated to extract a signal dimension related to the user's subjective attitude; In the behavior mapping path, a mapping relationship between the satisfaction representation vector and a historical behavior label is established to generate a user behavior bias feature; In the modal adversarial path, the proportion of each dimension data in the final representation is judged, and the weight vector is inversely propagated to each dimension data to refit; In the time sequence memory path, the sequence deviation between the historical satisfaction vector and the current satisfaction representation vector is captured to calculate a time stability encoding vector; A satisfaction evaluation model is constructed based on the score channels to obtain a satisfaction score value.

8. The satisfaction evaluation method based on multi-dimensional data fusion according to claim 7, wherein, The construction of the satisfaction evaluation model based on the score channels to obtain the satisfaction score value includes: The output results from the multiple score channels are mapped in a heterogeneous channel, and the mapping includes mapping the output results of different score paths to a discrimination space respectively; A score cluster is constructed, and the score cluster is composed of parallel score sub-models, and each sub-model independently calculates based on different hypothetical score logic; A symbolic score sequence is constructed for the score values output by each sub-model, and the symbolic score sequence is combined in a non-numeric symbolic manner; The symbolic score sequence is structure-compressed, index-transformed and path-decoded to generate a quantized score value; All quantized score values are weighted and fused to obtain a satisfaction score value.

9. A satisfaction evaluation system based on multi-dimensional data fusion for implementing the satisfaction evaluation method based on multi-dimensional data fusion according to any one of claims 1 to 8, characterized in that, It includes: a data acquisition module, a graph construction module, a credibility scoring module, a vector generation module and a satisfaction evaluation module; The data acquisition module is configured to acquire multi-source heterogeneous data. The graph construction module is configured to preprocess and extract features of the multi-source heterogeneous data, and construct a multi-modal feature graph. The credibility scoring module is configured to score a current credibility of the multi-source heterogeneous data based on a credibility scoring strategy. The vector generation module is configured to generate a satisfaction representation vector based on the multi-modal feature graph and the credibility score. The satisfaction evaluation module is configured to generate quantized score values based on the satisfaction representation vector, fuse all the quantized score values, and obtain a satisfaction score.

10. The satisfaction assessment system based on multi-dimensional data fusion as claimed in claim 9, wherein, The satisfaction evaluation module includes a sub-score unit, a heterogeneous channel mapping unit, a decoding unit, and a satisfaction evaluation unit. The sub-score unit is configured to copy the generated satisfaction representation vector to multiple scoring channels and output sub-scores. The heterogeneous channel mapping unit is configured to perform heterogeneous channel mapping on output results from the multiple scoring channels. The decoding unit is configured to construct a scoring cluster, construct a symbolic score sequence for score values output by each sub-model, and perform structural compression, index transformation, and path decoding on the symbolic score sequence to generate quantized score values. The satisfaction evaluation unit is configured to fuse all the quantized score values by weighting to obtain a satisfaction score value.

Images