Kano evaluation method and system based on time window and deep learning

Abstract

This invention provides a Kano evaluation method and system based on time windows and deep learning, comprising the following steps: S1, adaptive data windowing and vectorization; S2, constructing a dual-channel gated fusion model; S3, calculating temporal sentiment features; S4, dynamic Kano type determination; S5, interpretability attribution prediction and product improvement closed loop. This invention solves the problems of poor timeliness and static lag in traditional Kano analysis by constructing a dynamic closed-loop mechanism and a multi-feature fusion model, significantly improving the accuracy of demand analysis and the interpretability of the model.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a Kano evaluation method and system based on time windows and deep learning. Background Technology

[0002] The Kano model is a classic tool for classifying and prioritizing user needs, but it suffers from the following problems in practice: First, time lag and static nature: Traditional Kano analysis relies heavily on periodic questionnaires to collect data, resulting in long data collection cycles and delayed results, failing to capture the dynamic evolution of user needs as the market environment changes. Second, insufficient semantic understanding and uninterpretable models: Traditional natural language processing methods are insufficient in understanding the semantics of complex contexts and metaphorical expressions when processing unstructured texts such as e-commerce reviews. Furthermore, statistical methods are highly susceptible to data sparsity and noise interference. Existing sentiment analysis models are mostly "black box" models, only outputting classification results without explaining the decision-making basis, making it difficult to establish a correlation between sentiment phenomena and specific Kano need types, thus limiting the guiding value of the analysis results for product improvement. Summary of the Invention

[0003] The purpose of this invention is to provide a Kano evaluation method and system based on time windows and deep learning, which can solve the technical problems of the traditional Kano evaluation method, such as time lag, insufficient extraction of sentiment features, and weak interpretability of sentiment analysis models.

[0004] The technical solution adopted by this invention to solve its technical problem is: In a first aspect, the present invention provides a Kano evaluation method based on time windows and deep learning, comprising the following steps: S1. Adaptive Data Windowing and Vectorization: Obtain time-series evaluation data of the target product, parse product attribute tags to match the preset time window length strategy, divide the evaluation data into multiple time window datasets arranged in time sequence, and perform three-level labeling of sentiment tags. S2. Construct a dual-channel gated fusion model: Build a deep neural network that includes a Bi-LSTM temporal feature extraction branch with parallel settings and a CNN local feature extraction branch, and establish the feature mapping relationship between the two branches through an attention fusion gating unit; S3. Calculation of temporal sentiment features: Input the datasets of each time window in step S1 into the model trained in step S2, use the attention fusion gating unit to dynamically calculate the fusion weight of temporal features and local features, and output the sentiment polarity probability distribution of each demand item under different time windows. S4. Dynamic Kano type determination: Based on the probability distribution of emotional polarity, the satisfaction coefficient and dissatisfaction coefficient of each time window are calculated. A time decay factor is introduced to weight and sum the coefficient values ​​of historical time windows to obtain the comprehensive weight of the demand item and the Kano type at the current moment. S5. Explainable Attribution Prediction and Product Improvement Closed Loop: Based on the weight distribution of the attention fusion gating unit, semantic attribution results are generated. The time series prediction model is used to identify the evolution trend of the emotional semantics of each demand item and generate product improvement feedback signals. New evaluation data in subsequent time windows are collected, the closed-loop health index is calculated, and when the index meets the preset threshold, the new data is used to incrementally train the deep neural network to achieve adaptive iteration of model parameters.

[0005] Furthermore, in step S1, the standard for the three-level labeling of the sentiment tag is as follows: Positive sentiment: Rated 4-5 stars and contains positive semantics; Neutral sentiment: 3 stars or no obvious sentiment tendency; Negative sentiment: Rated 1-2 stars and contains negative semantics; The annotation accuracy of the time window dataset is ≥95%.

[0006] Furthermore, the time window length of the time window dataset is adaptively set according to the product type, wherein it is 1-2 weeks for fast-moving consumer goods and 1-3 months for durable goods; when the business requirement is to evaluate promotional activities, specific time windows are set for 1 week before the activity, during the activity, and 2 weeks after the activity.

[0007] Furthermore, in step S2, the attention fusion gating unit implements feature fusion through the following instruction logic: configuring the processor to perform nonlinear mapping operations to obtain the fusion weight coefficients of temporal dependent features and local phrase features, and adjusting the signal strength contribution of temporal dependent features and local phrase features in the final classification decision through the fusion weight coefficients.

[0008] Furthermore, in step S2, the Bi-LSTM branch and the CNN branch are parallel input structures. The Bi-LSTM branch is used to capture the long-range dependent features of the contextual semantics of the text, and the CNN branch captures local sentiment phrase features through multi-scale convolutional kernels and achieves adaptive information complementarity fusion in the feature vector dimension through the attention fusion gating unit.

[0009] Furthermore, in step S4, the processor is configured to perform time-series weighted compensation calculations on the satisfaction coefficients and dissatisfaction coefficients corresponding to different time windows to obtain the comprehensive weight of each demand item.

[0010] Furthermore, in step S4, the processor is configured to perform quantification mapping based on the number of evaluations of different emotional tendencies within different time windows, and calculate the satisfaction coefficient and dissatisfaction coefficient.

[0011] Furthermore, in step S5, identifying the evolutionary trend of sentiment semantics using a time series prediction model includes: The sentiment semantic feature index was decomposed into trend, seasonal and residual terms using the time series decomposition method (STL). The trend term is modeled and predicted using the autoregressive integral moving average (ARIMA) model to identify the direction of change in sentiment semantics. A three-dimensional feedback report is generated based on the prediction results, including demand priorities, improvement directions, and expected goals.

[0012] Furthermore, in step S5, the processor is configured to collect feedback data in real time, thereby promoting the monitoring of the closed-loop health index and the termination of incremental training.

[0013] Secondly, the present invention provides a system for the Kano evaluation method based on time windows and deep learning as described in any one of the above-mentioned methods, characterized in that it comprises: Data Windowing and Preprocessing Module: Configured to acquire time-series evaluation data of target products, parse product attribute tags to match preset time window length strategies, divide the evaluation data into multiple time window datasets arranged in time sequence, and perform three-level sentiment labeling; Dual-channel gated fusion sentiment analysis module: It has a built-in deep neural network, which includes a parallel Bi-LSTM temporal feature extraction branch and a CNN local feature extraction branch, as well as an attention fusion gating unit. The module is configured to input the datasets of each time window into the deep neural network, use the attention fusion gating unit to dynamically calculate the feature fusion weights, and output the sentiment polarity probability distribution of each required item. Dynamic Kano Analysis and Weight Calculation Module: Configured to calculate the satisfaction and dissatisfaction coefficients for each time window based on the probability distribution of emotional polarity, and introduce a time decay factor to weight and sum the coefficient values ​​of historical time windows to determine the comprehensive weight of the demand item and the Kano type at the current moment; The interpretable attribution prediction and product improvement closed-loop iteration module is configured to generate semantic attribution results based on the weight distribution of the attention fusion gating unit, and to identify the evolution trend of sentiment semantics using the internally integrated evolution trend modeling unit to generate product improvement feedback signals; it is also configured to collect new evaluation data in subsequent time windows, calculate the closed-loop health index, and when the index meets a preset threshold, use the new data to incrementally train the deep neural network to achieve adaptive iteration of model parameters.

[0014] In summary, the beneficial effects of the present invention are as follows: The Kano evaluation method and system based on time windows and deep learning of this invention significantly improves the timeliness, accuracy, interpretability, and decision-making value of customer needs analysis through a series of synergistic technical means, specifically reflected in: 1. Real-time, highly sensitive capture of customer demand evolution trajectory: By configuring the processor to execute an adaptive time window segmentation mechanism based on product attributes, and using a time decay factor to exponentially compensate for the decay of historical data, this not only solves the data lag problem of the traditional Kano method, but also enables the analysis results to reflect the subtle shifts in user demand within the current time window (such as the prediction of conversion from "attractive" to "essential") through a mathematical weighting mechanism, shortening the demand identification lag for FMCG products to the weekly level.

[0015] 2. Significantly enhanced robustness and accuracy of sentiment feature extraction in complex contexts: By employing a parallel architecture of Bi-LSTM and CNN, along with an attention fusion gating unit, the model overcomes the limitation of single neural networks in simultaneously addressing local features and long-range semantics when processing text. The gating unit adaptively calculates feature contribution weights, improving the negative sentiment recognition score to over 85% when handling complex evaluations containing transitions, negations, and mixed emotions. Furthermore, by incorporating domain-specific fine-tuning of the pre-trained model, the generalization error across different product categories is significantly reduced.

[0016] 3. This invention enables visualized tracing of the emotion judgment logic, providing precise evidence for product improvement: By extracting the weight distribution of attention fusion gating units, a correlation mechanism between text features and classification decisions is constructed. This allows the system to output word-level attention heatmaps, intuitively displaying the core keywords affecting satisfaction. Compared to the "black box" model of existing technologies, this invention can transform abstract emotion tags into specific, traceable functional feedback points (e.g., accurately locating the specific feature of "voice wake-up delay"), significantly reducing the difficulty of transforming data analysis into implementation decisions.

[0017] 4. An adaptive closed-loop mechanism based on real-time steady-state monitoring was constructed to improve the reliability of model iteration: By introducing a "closed-loop health index" and "feedback interruption logic," automated quality monitoring of the model evolution process was achieved. The system effectively identifies and shields abnormal data from interfering with model parameters through multi-dimensional quantitative monitoring of improvement verification efficiency, model performance stability, and early warning accuracy. When the monitored health score falls below a preset safety threshold, the system automatically triggers an interruption signal to suspend the iteration process, preventing performance degradation of the model in an unattended environment and ensuring the robustness of parameter updates. This closed-loop architecture significantly shortens the cycle from requirement identification to effect verification, optimizing product improvement efficiency in practical applications and supporting a 15%-20% quantitative improvement in customer satisfaction. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Figure 1 This is a flowchart of the Kano evaluation method based on time windows and deep learning of the present invention; Figure 2 This is a schematic diagram of the structure of the dual-channel gated fusion sentiment analysis model of the present invention; Figure 3 This is a flowchart of the dynamic Kano evaluation module of the present invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, S1-S5: the main steps of the evaluation method; S301-S305: the specific execution steps of the dynamic Kano evaluation module; H: the closed-loop health index, used to measure the performance status of the model in the current time window; T: the preset safety threshold matrix, used as the criterion for triggering feedback interruption. Time decay factor, used to adjust the influence of historical data on the current weight; : Long-range feature vectors of contextual semantics extracted by the Bi-LSTM branch; Local phrase feature vectors extracted by CNN branches; : Dynamic feature fusion weights generated by attention fusion gating.

[0022] Please see Figure 1-3 In a first aspect, the present invention provides a Kano evaluation method based on time windows and deep learning, comprising the following steps: S1. Adaptive Data Windowing and Vectorization: Obtain time-series evaluation data of the target product, parse product attribute tags to match the preset time window length strategy, divide the evaluation data into multiple time window datasets arranged in time sequence, and perform three-level labeling of sentiment tags. S2. Construct a dual-channel gated fusion model: Build a deep neural network that includes a Bi-LSTM temporal feature extraction branch with parallel settings and a CNN local feature extraction branch, and establish the feature mapping relationship between the two branches through an attention fusion gating unit; S3. Calculation of temporal sentiment features: Input the datasets of each time window in step S1 into the model trained in step S2, use the attention fusion gating unit to dynamically calculate the fusion weight of temporal features and local features, and output the sentiment polarity probability distribution of each demand item under different time windows. S4. Dynamic Kano type determination: Based on the probability distribution of emotional polarity, the satisfaction coefficient and dissatisfaction coefficient of each time window are calculated. A time decay factor is introduced to weight and sum the coefficient values ​​of historical time windows to obtain the comprehensive weight of the demand item and the Kano type at the current moment. S5. Explainable Attribution Prediction and Product Improvement Closed Loop: Based on the weight distribution of the attention fusion gating unit, semantic attribution results are generated. The time series prediction model is used to identify the evolution trend of the emotional semantics of each demand item and generate product improvement feedback signals. New evaluation data in subsequent time windows are collected, the closed-loop health index is calculated, and when the index meets the preset threshold, the new data is used to incrementally train the deep neural network to achieve adaptive iteration of model parameters.

[0023] Furthermore, in step S1, the standard for the three-level labeling of the sentiment tag is as follows: Positive sentiment: Rated 4-5 stars and contains positive semantics; Neutral sentiment: 3 stars or no obvious sentiment tendency; Negative sentiment: Rated 1-2 stars and contains negative semantics; The annotation accuracy of the time window dataset is ≥95%.

[0024] Furthermore, the time window length of the time window dataset is adaptively set according to the product type, wherein it is 1-2 weeks for fast-moving consumer goods and 1-3 months for durable goods; when the business requirement is to evaluate promotional activities, specific time windows are set for 1 week before the activity, during the activity, and 2 weeks after the activity.

[0025] Furthermore, in step S2, the attention fusion gating unit implements feature fusion through the following instruction logic: configuring the processor to perform nonlinear mapping operations. To obtain the fusion weight coefficients of temporal dependency features and local phrase features. ; in, Let W be the sigmoid activation function, W1 and W2 be learnable weight matrices, and b be the bias term. The time-dependent feature vector output by the Bi-LSTM branch. The local phrase feature vector output by the CNN branch; based on the The feature vectors are weighted and fused to obtain a comprehensive feature vector. Through the The contribution of temporal-dependent features and local phrase features to signal strength in the final classification decision is dynamically adjusted.

[0026] Furthermore, in step S2, the Bi-LSTM branch and the CNN branch are parallel input structures. The Bi-LSTM branch is used to capture long-range semantic dependencies of the text context, while the CNN branch captures local sentiment phrase features through multi-scale convolutional kernels and achieves adaptive information complementarity fusion in the feature vector dimension through the attention fusion gating unit. Specifically, the Bi-LSTM branch is used to process the contextual information of the text sequence, capturing long-range semantic dependencies (such as contrast relationships and negation prefixes) through two LSTM layers (forward and backward); the CNN branch extracts local key phrase features (such as emotionally charged adjective phrases) in the text through convolutional kernels of different sizes (such as 3, 4, and 5).

[0027] To achieve effective complementarity between these two types of features, a new attention fusion gating unit is added. This unit does not directly concatenate features, but instead configures the processor to perform non-linear mapping operations. This formula utilizes the Sigmoid function ( The fusion weights are normalized to the (0, 1) interval. The model then adjusts the weights accordingly. Dynamically adjust information flow: When processing long and complex sentences, automatically increase the value. The weighting; when processing phrase fragments, increase The weights. The final generated comprehensive feature vector. It incorporates both global and local semantic information, thereby significantly improving the robustness of classification.

[0028] Furthermore, in step S4, the processor is configured to perform time-series weighted compensation calculations on the satisfaction and dissatisfaction coefficients corresponding to different time windows to obtain the comprehensive weight of each demand item. : Where t is the time window number, n is the latest time window number, and α is the preset time decay constant, where 0 < α < 1. and These are the satisfaction coefficient and dissatisfaction coefficient of the j-th demand item within the t-th time window, respectively.

[0029] Furthermore, in step S4, the configuration processor performs quantification mapping based on the number of evaluations with different emotional tendencies within different time windows, and calculates the satisfaction coefficient S and the dissatisfaction coefficient D: .

[0030] Specifically, this invention introduces a time decay weighting mechanism to address the problem in traditional analysis methods where recent and distant data have equal weights, failing to highlight the value of recent data. The system calculates the weights based on the formula... The overall weight of each demand item is calculated. A time decay factor α (0 < α < 1) is used to adjust the dynamic influence of historical data: as the distance between the historical time window t and the current time window n increases, the contribution of this sample to the overall weight decreases exponentially. This mechanism can sensitively capture the non-linear drift characteristics of user demands over time, thereby more accurately identifying the dynamic evolution of demand attributes, such as the degradation process from attractive demands to essential demands.

[0031] Furthermore, in step S5, identifying the evolutionary trend of sentiment semantics using a time series prediction model includes: The sentiment semantic feature index was decomposed into trend, seasonal and residual terms using the time series decomposition method (STL). The trend term is modeled and predicted using the autoregressive integral moving average (ARIMA) model to identify the direction of change in sentiment semantics. A three-dimensional feedback report is generated based on the prediction results, including demand priorities, improvement directions, and expected goals.

[0032] Furthermore, in step S5, the monitoring and incremental training termination logic for the closed-loop health index H specifically involves configuring the processor to collect feedback data in real time and calculating the closed-loop health index H using the following formula: Among them, H1 is the score for improved verification efficiency, H2 is the score for model performance stability, and H3 is the score for demand early warning accuracy. The preset weighting coefficients satisfy the following conditions: The system presets a security threshold matrix T=[T1,T2,T3], where T1,T2,T3 correspond to the minimum security scores of H1, H2, and H3, respectively. When any sub-indicator is detected... When the processor detects a logic verification anomaly, it immediately generates a feedback interrupt signal and suspends the gradient update process of the deep neural network to stop automatic incremental training.

[0033] Specifically, the Kano evaluation method based on time windows and deep learning in this invention, on the one hand, utilizes the STL (Seasonal-Trend Decomposition using Loess) algorithm to decompose the time series of sentiment semantics into trend, seasonal, and residual terms, and uses the ARIMA model to perform linear extrapolation prediction on the trend term, thereby identifying potential demand evolution directions in advance. On the other hand, to prevent erroneous samples from causing model performance degradation, the system introduces a closed-loop health index. This index is composed of a weighted average of an improved validation efficiency score (measuring feedback response speed), a model performance stability score (measuring model performance fluctuations), and a demand early warning accuracy score (measuring prediction accuracy). The system monitors changes in this index in real time. When any sub-index falls below a preset safety threshold, the system generates a feedback interruption signal and, based on this signal, suspends the backpropagation gradient update process of the deep neural network through control logic instructions to stop the incremental training stream. This mechanism achieves state monitoring and anomaly shielding during automated operation through logical verification, ensuring the reliability of model iteration.

[0034] Secondly, the present invention provides a Kano evaluation system based on time windows and deep learning, comprising: Data Windowing and Preprocessing Module: Configured to acquire time-series evaluation data of target products, parse product attribute tags to match preset time window length strategies, divide the evaluation data into multiple time window datasets arranged in time sequence, and perform three-level sentiment labeling; Dual-channel gated fusion sentiment analysis module: It has a built-in deep neural network, which includes a parallel Bi-LSTM temporal feature extraction branch and a CNN local feature extraction branch, as well as an attention fusion gating unit. The module is configured to input the datasets of each time window into the deep neural network, use the attention fusion gating unit to dynamically calculate the feature fusion weights, and output the sentiment polarity probability distribution of each required item. Dynamic Kano Analysis and Weight Calculation Module: Configured to calculate the satisfaction and dissatisfaction coefficients for each time window based on the probability distribution of emotional polarity, and introduce a time decay factor to weight and sum the coefficient values ​​of historical time windows to determine the comprehensive weight of the demand item and the Kano type at the current moment; The interpretable attribution prediction and product improvement closed-loop iteration module is configured to generate semantic attribution results based on the weight distribution of the attention fusion gating unit, and to identify the evolution trend of sentiment semantics using the internally integrated evolution trend modeling unit to generate product improvement feedback signals; it is also configured to collect new evaluation data in subsequent time windows, calculate the closed-loop health index, and when the index meets a preset threshold, use the new data to incrementally train the deep neural network to achieve adaptive iteration of model parameters.

[0035] In one embodiment, the present invention provides a Kano evaluation method based on time windows and deep learning, comprising the following steps: S1, Adaptive Data Windowing and Vectorization The system first acquires full evaluation data for the target product. Based on a pre-built metadata rule base, the system automatically identifies product attribute tags: for fast-moving consumer goods (such as daily chemicals and food), the system configures the processor to divide the data into sliding time windows with a step size of 7-14 days; for durable goods (such as home appliances and digital products), the division is performed with a step size of 1-3 months. During the data annotation phase, a three-level sentiment labeling system (positive, neutral, and negative) is used. To ensure data quality, the system constructs a multi-level verification mechanism: each data point is independently annotated by three annotation subjects, and the system automatically calculates the Fleiss' Kappa consistency coefficient. When the coefficient is below 0.8, the system triggers conflict handling logic, introducing a verification node to perform the final decision.

[0036] To adapt to the business attributes of different products, this embodiment uses a processor to match a preset time window strategy. As shown in the table, the processor automatically determines the corresponding recommended time window length and segmentation logic based on the target product attributes (such as fast-moving consumer goods or durable goods) to ensure accurate capture of emotional polarity characteristics.

[0037] ; S2. Construct a dual-channel gating fusion model The constructed model adopts an architecture of "domain pre-training + dual-channel parallel feature extraction + dynamic weight fusion": Embedding layer: A fine-tuned BERT-base-chinese model is used to transform the evaluation text into a 768-dimensional high-dimensional feature vector. For example... Figure 2 As shown, the parallel feature extraction layer contains two parallel branches: Bi-LSTM and CNN. The Bi-LSTM branch has two hidden layers, each containing 128 units, and is responsible for capturing long-range semantic dependencies of the text (such as transitional sentences) through a bidirectional recurrent structure. The CNN branch uses convolutional kernels of sizes 3, 4, and 5, each with 128 filters, and is responsible for extracting local key sentiment phrases (such as "amazing sound quality" and "poor battery life").

[0038] Attention Fusion Gating Unit: This unit receives the output vectors from two branches. and Configure the processor to perform nonlinear mapping operations By using the sigmoid function to compress the weights to between 0 and 1, a dynamic weighted superposition of global semantics and local phrase features is achieved, generating a fused feature vector that maximizes information gain. .

[0039] Classification output layer: The fused feature vector The input is a fully connected layer, which is then classified using the Softmax activation function. The output is the probability distribution of the current evaluation text corresponding to the three sentiment polarities of positive, neutral, and negative, providing a quantitative basis for feature calculation in the subsequent S3 step.

[0040] S3, Calculation of Temporal Sentiment Features The multi-time-window dataset after S1 is divided is input in batches into the trained deep sentiment analysis model, and the following calculation logic is executed: Demand item identification and extraction: Based on the high-frequency words output by the model, the configuration processor automatically extracts target demand items (such as key functional attributes such as "voice wake-up", "sound quality", "battery life" and other key functional attributes) from the evaluation text by combining the TF-IDF algorithm with the pre-set domain dictionary.

[0041] Sentiment polarity mapping: For each evaluation record involving a specific need item, the probability distribution corresponding to the positive, neutral, and negative sentiment categories is calculated using the Softmax function of the classification layer.

[0042] Structured data output: The unstructured text corpus is transformed into a structured dataset containing demand item ID, time window ID, sentiment label, and confidence, providing basic data support for the quantification of Kano coefficient in the subsequent S4 step.

[0043] S4. Dynamic Kano Classification and Weight Iteration The processor is configured to perform quantization mapping on each time window dataset based on the sentiment polarity probability distribution of the demand items output by S3: Satisfaction quantification: The processor counts the number of positive reviews, negative reviews, and indifferent reviews for the j-th requirement within the t-th time window. According to the formula... Calculate the satisfaction coefficient (Better coefficient) according to the formula. Calculate the Worse coefficient.

[0044] Kano type dynamic positioning: Based on preset coordinate determination rules, the (S, D) coordinate point is mapped to different demand attribute ranges to realize dynamic tracking of the evolution trajectory of demand items from attractive to essential.

[0045] Comprehensive weight calculation: To reflect the novelty of the data, the processor introduces a time decay factor α. n-1 (Where α takes values ​​in the range 0 < α < 1, such as 0.95). According to the formula... The overall weight of the demand items is calculated. This mechanism ensures that the analysis results can sensitively reflect the core needs of users at the current stage by performing exponential decay compensation on the historical time window coefficients.

[0046] After obtaining the satisfaction coefficient S and dissatisfaction coefficient D, the system performs the mapping of demand attributes, such as... Figure 3 As shown in Table 2, the processor determines the Kano demand type of the demand item in the current time window by performing logical jumps based on the interval in which the calculated coefficients fall and in conjunction with the decision matrix.

[0047] ; S5. Explainable Attribution Prediction and Product Improvement Closed Loop This embodiment forms a complete technical closed loop by constructing interpretable output and feedback control logic: Interpretable semantic attribution: The processor extracts the weight distribution of the attention fusion gating units in the S2 deep neural network and maps it to the original text corpus to generate word-level attention heatmaps. By identifying high-weight feature words (such as "slow wake-up" and "crisp voice"), the transparent tracing of the basis for sentiment decisions is achieved.

[0048] Evolutionary Trend Prediction: The configuration processor uses the Time Series Decomposition (STL) method to decompose sentiment characteristic indicators into trend, seasonal, and residual terms. The trend term is modeled using an Autoregressive Integrated Moving Average (ARIMA) model to identify potential directions of sentiment semantic evolution and generate a three-dimensional feedback signal that includes demand priorities, improvement directions, and expected goals.

[0049] Closed-loop health monitoring and self-iteration: The system improves and verifies its performance by collecting new data in subsequent time windows in real time. The processor is configured according to the formula. The closed-loop health index is calculated, where H1 is the improved validation efficiency score, H2 is the model performance stability score, and H3 is the demand early warning accuracy score. When any sub-index falls below a preset safety threshold (e.g., model performance decline ≥ 5%), the system automatically generates a feedback interruption signal. This feedback interruption signal, through control logic circuits or software scheduling instructions, suspends the backpropagation gradient update process of the deep neural network trainer, such as... Figure 1As shown, this physically halts the incremental training flow to prevent parameter contamination caused by erroneous samples until an external verification instruction is received. If the health indicators are normal, the verified new data is added to the training set to perform incremental training on the model, achieving self-learning evolution of the domain-adaptive model.

[0050] This invention uses e-commerce review analysis of a certain brand of smart speaker as an example to explain in detail the entire application process of the above method: 1. Experimental preparation and windowing We collected full evaluation data for the speaker over 12 consecutive months after its market launch. Since the speaker is a digital durable good, the system set the time window step to 1 month, for a total of 12 time windows. A training set of 8400 records, a validation set of 2400 records, and a test set of 1200 records were constructed using three levels of annotation, achieving an annotation accuracy of 96%.

[0051] 2. Model Performance Validation The model employs Python 3.8 and the PyTorch framework, loading BERT-base-chinese pre-trained weights. In the dual-channel parallel feature extraction layer, the Bi-LSTM hidden layer is set to 128 units, and the CNN convolutional kernel size is set to 3, 4, and 5. After 50 rounds of training, the model achieves a sentiment classification accuracy of 92.5% on the test set and an F1 score of 88.7% for negative sentiment recognition.

[0052] 3. Dynamic Kano Evolution Analysis The analysis results for the "voice wake-up" function under different time windows (t) are shown in Table 3 below: Phase 1 (t=1 to 3): This function falls within the "attractive demand" range, with a satisfaction coefficient s≥ 0.8, indicating that users show extremely high positive sentiment towards the intelligence of the wake-up function.

[0053] Phase Two (t=4 to 9): With the launch of competing products and the increase in user expectations, the proportion of negative reviews rose from 5% to 35%, and Kano's attributes shifted from "attractive" to "essential".

[0054] Phase 3 (Current Moment): Introduce a time decay factor (α=0.95) to calculate the comprehensive weight, and the system identifies it as a "P1 level core improvement requirement".

[0055] 4. Closed-loop iteration example The word-level attention heatmap generated by the system showed that "delayed arousal" and "slow response" were the core attributions for negative emotions. Based on this, the product team optimized the algorithm, and the system detected a health index H=0.88 (above the safe threshold of 0.8) in the 12th time window. The model automatically collected new feedback data for incremental training. Final comparative data showed that the proportion of negative reviews decreased from 35% to 12% after the improvement, and overall customer satisfaction increased by approximately 18%.

[0056] To further verify the technical effects of the present invention, this embodiment compares the present invention's solution with traditional statistical methods and single-channel neural network models in the prior art. The experimental results are shown in Table 3, and the present invention's solution demonstrates significant advantages in all key indicators.

[0057] ; As shown in Table 3, the present invention, through dual-channel feature fusion and closed-loop monitoring mechanism, not only improves the accuracy of emotion recognition, but also significantly enhances the system's decision support capability in actual business scenarios.

[0058] The present invention relates to a Kano evaluation method and system based on time windows and deep learning, comprising the following steps: S1, adaptively dividing time windows according to product attributes and performing three-level sentiment labeling; S2, constructing a deep model including parallel branches of Bi-LSTM and CNN and attention fusion gating units; S3, calculating the sentiment polarity probability distribution of demand items within each time window; S4, calculating satisfaction and dissatisfaction coefficients, and introducing a time decay factor to dynamically determine Kano type and comprehensive weight; S5, identifying sentiment evolution trends and generating feedback signals, and achieving incremental model training based on closed-loop health indicators. This invention solves the problems of poor timeliness and static lag in traditional Kano analysis by constructing a dynamic closed-loop mechanism and a multi-feature fusion model, significantly improving the accuracy of demand analysis and the interpretability of the model.

[0059] The Kano evaluation method and system based on time windows and deep learning of this invention significantly improves the timeliness, accuracy, interpretability, and decision-making value of customer needs analysis through a series of synergistic technical means, specifically reflected in: 1. Real-time, highly sensitive capture of customer demand evolution trajectory: By configuring the processor to execute an adaptive time window segmentation mechanism based on product attributes, and using a time decay factor to exponentially compensate for the decay of historical data, this not only solves the data lag problem of the traditional Kano method, but also enables the analysis results to reflect the subtle shifts in user demand within the current time window (such as the prediction of conversion from "attractive" to "essential") through a mathematical weighting mechanism, shortening the demand identification lag for FMCG products to the weekly level.

[0060] 2. Significantly enhanced robustness and accuracy of sentiment feature extraction in complex contexts: By employing a parallel architecture of Bi-LSTM and CNN, along with an attention fusion gating unit, the model overcomes the limitation of single neural networks in simultaneously addressing local features and long-range semantics when processing text. The gating unit adaptively calculates feature contribution weights, improving the negative sentiment recognition score to over 85% when handling complex evaluations containing transitions, negations, and mixed emotions. Furthermore, by incorporating domain-specific fine-tuning of the pre-trained model, the generalization error across different product categories is significantly reduced.

[0061] 3. This invention enables visualized tracing of the emotion judgment logic, providing precise evidence for product improvement: By extracting the weight distribution of attention fusion gating units, a correlation mechanism between text features and classification decisions is constructed. This allows the system to output word-level attention heatmaps, intuitively displaying the core keywords affecting satisfaction. Compared to the "black box" model of existing technologies, this invention can transform abstract emotion tags into specific, traceable functional feedback points (e.g., accurately locating the specific feature of "voice wake-up delay"), significantly reducing the difficulty of transforming data analysis into implementation decisions.

[0062] 4. An adaptive closed-loop mechanism based on real-time steady-state monitoring was constructed to improve the reliability of model iteration: By introducing a "closed-loop health index" and "feedback interruption logic," automated quality monitoring of the model evolution process was achieved. The system effectively identifies and shields abnormal data from interfering with model parameters through multi-dimensional quantitative monitoring of improvement verification efficiency, model performance stability, and early warning accuracy. When the monitored health score falls below a preset safety threshold, the system automatically triggers an interruption signal to suspend the iteration process, preventing performance degradation of the model in an unattended environment and ensuring the robustness of parameter updates. This closed-loop architecture significantly shortens the cycle from requirement identification to effect verification, optimizing product improvement efficiency in practical applications and supporting a 15%-20% quantitative improvement in customer satisfaction.

[0063] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may utilize the disclosed technical content to make changes or equivalent variations to other fields. However, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the present invention, shall still fall within the protection scope of the present invention. In the description of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation" and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in the present invention can be understood through specific circumstances.

Claims

1. A Kano evaluation method based on time windows and deep learning, characterized in that, Includes the following steps: S1. Adaptive Data Windowing and Vectorization: Obtain time-series evaluation data of the target product, parse product attribute tags to match the preset time window length strategy, divide the evaluation data into multiple time window datasets arranged in time sequence, and perform three-level labeling of sentiment tags. S2. Construct a dual-channel gated fusion model: Build a deep neural network that includes a Bi-LSTM temporal feature extraction branch with parallel settings and a CNN local feature extraction branch, and establish the feature mapping relationship between the two branches through an attention fusion gating unit; S3. Calculation of temporal sentiment features: Input the datasets of each time window in step S1 into the model trained in step S2, use the attention fusion gating unit to dynamically calculate the fusion weight of temporal features and local features, and output the sentiment polarity probability distribution of each demand item under different time windows. S4. Dynamic Kano type determination: Based on the probability distribution of emotional polarity, the satisfaction coefficient and dissatisfaction coefficient of each time window are calculated. A time decay factor is introduced to weight and sum the coefficient values ​​of historical time windows to obtain the comprehensive weight of the demand item and the Kano type at the current moment. S5. Explainable Attribution Prediction and Product Improvement Closed Loop: Based on the weight distribution of the attention fusion gating unit, semantic attribution results are generated. The time series prediction model is used to identify the evolution trend of the emotional semantics of each demand item and generate product improvement feedback signals. New evaluation data in subsequent time windows are collected, the closed-loop health index is calculated, and when the index meets the preset threshold, the new data is used to incrementally train the deep neural network to achieve adaptive iteration of model parameters.

2. The Kano evaluation method based on time windows and deep learning as described in claim 1, characterized in that, In step S1, the standard for the three-level labeling of the sentiment tag is as follows: Positive sentiment: Rated 4-5 stars and contains positive semantics; Neutral sentiment: 3 stars or no obvious sentiment tendency; Negative sentiment: Rated 1-2 stars and contains negative semantics; The annotation accuracy of the time window dataset is ≥95%.

3. The Kano evaluation method based on time windows and deep learning as described in claim 1, characterized in that, The time window length of the time window dataset is adaptively set according to the product type, with 1-2 weeks for fast-moving consumer goods and 1-3 months for durable goods; when the business requirement is to evaluate promotional activities, specific time windows are set for 1 week before the activity, during the activity, and 2 weeks after the activity.

4. The Kano evaluation method based on time windows and deep learning as described in claim 1, characterized in that: In step S2, the attention fusion gating unit implements feature fusion through the following instruction logic: configure the processor to perform nonlinear mapping operation to obtain the fusion weight coefficients of temporal dependent features and local phrase features, and adjust the signal strength contribution of temporal dependent features and local phrase features in the final classification decision through the fusion weight coefficients.

5. The Kano evaluation method based on time windows and deep learning as described in claim 1, characterized in that: In step S2, the Bi-LSTM branch and the CNN branch are parallel input structures. The Bi-LSTM branch is used to capture the long-range semantic features of the text context, and the CNN branch captures local sentiment phrase features through multi-scale convolution kernels and achieves adaptive information complementarity fusion in the feature vector dimension through the attention fusion gating unit.

6. The Kano evaluation method based on time windows and deep learning as described in claim 1, characterized in that: In step S4, the processor is configured to perform time-series weighted compensation calculations on the satisfaction coefficients and dissatisfaction coefficients corresponding to different time windows to obtain the comprehensive weight of each demand item.

7. The Kano evaluation method based on time windows and deep learning as described in claim 1, characterized in that: In step S4, the processor is configured to perform quantitative mapping based on the number of evaluations of different emotional tendencies within different time windows, and calculate the satisfaction coefficient and dissatisfaction coefficient.

8. The Kano evaluation method based on time windows and deep learning as described in claim 1, characterized in that: In step S5, identifying the evolution trend of sentiment semantics using a time series prediction model includes: The sentiment semantic feature index was decomposed into trend, seasonal and residual terms using the time series decomposition method (STL). The trend term is modeled and predicted using the autoregressive integral moving average (ARIMA) model to identify the direction of change in sentiment semantics. A three-dimensional feedback report is generated based on the prediction results, including demand priorities, improvement directions, and expected goals.

9. The Kano evaluation method based on time windows and deep learning as described in claim 1, characterized in that: In step S5, the processor is configured to collect feedback data in real time, thereby promoting the monitoring of the closed-loop health index and the termination of incremental training.

10. A system employing the Kano evaluation method based on time windows and deep learning as described in any one of claims 1-9, characterized in that: include: Data Windowing and Preprocessing Module: Configured to acquire time-series evaluation data of target products, parse product attribute tags to match preset time window length strategies, divide the evaluation data into multiple time window datasets arranged in time sequence, and perform three-level sentiment labeling; Dual-channel gated fusion sentiment analysis module: It has a built-in deep neural network, which includes a parallel Bi-LSTM temporal feature extraction branch and a CNN local feature extraction branch, as well as an attention fusion gating unit. The module is configured to input the datasets of each time window into the deep neural network, use the attention fusion gating unit to dynamically calculate the feature fusion weights, and output the sentiment polarity probability distribution of each required item. Dynamic Kano Analysis and Weight Calculation Module: Configured to calculate the satisfaction and dissatisfaction coefficients for each time window based on the probability distribution of emotional polarity, and introduce a time decay factor to weight and sum the coefficient values ​​of historical time windows to determine the comprehensive weight of the demand item and the Kano type at the current moment; Interpretable Attribution Prediction and Product Improvement Closed-Loop Iteration Module: Configured to generate semantic attribution results based on the weight distribution of the attention fusion gating unit, and to identify the evolution trend of sentiment semantics using the internally integrated evolution trend modeling unit to generate product improvement feedback signals; It is also configured to collect new evaluation data in subsequent time windows, calculate the closed-loop health index, and when the index meets a preset threshold, use the new data to perform incremental training on the deep neural network to achieve adaptive iteration of model parameters.

Images