A product optimization method based on user review data

Abstract

The present invention belongs to the field of mobile health technology, and specifically relates to a product optimization method based on user review data, comprising: obtaining review data of wearable smart products, preprocessing the review data to obtain product data information; analyzing the product data information to obtain factors affecting product quality; using a Bayesian causal inference model to extract causal relationship features of the factors affecting product quality to obtain a posterior probability tracking graph; generating a causal relationship graph based on the posterior probability tracking graph, and calculating the causal effect; improving and optimizing the wearable smart product based on the causal effect; the present invention uses a Bayesian causal inference method to analyze the causal relationship between various factors of the product, determines whether the causal relationship is stable from the obtained posterior probability tracking graph, determines factors affecting user experience, and improves the product based on the results obtained by causal inference.

Description

Technical Field

[0001] The present invention belongs to the field of mobile health technology, and in particular relates to a product optimization method based on user review data. Background Art

[0002] With the accelerated aging of the population, the health and safety of the elderly are attracting increasing attention. Falls are a common accidental event for the elderly and a leading cause of injury and death. Falls are a common difficulty in the lives of the elderly. Studies show that falls are the leading cause of death from injuries among people over 65 in my country. The older you are, the greater the probability of falling, and the higher the risk of injury and death from falls. Smart fall prevention devices are important assistive living devices and are crucial for the health and safety of the elderly. However, current smart fall prevention devices on the market still have some user experience issues, such as inconvenient operation, unclear prompts, and untimely data feedback. These issues reduce the device's practicality and user satisfaction, affect the user experience and comfort, and may also lead to a decline in the device's effectiveness and safety, thereby reducing user trust and acceptance of smart fall prevention devices. Summary of the Invention

[0003] To solve the above problems in the prior art, the present invention proposes a product optimization method based on user review data, comprising:

[0004] S1. Obtain review data of wearable smart products;

[0005] S2. Preprocess the review data to obtain product data information;

[0006] S3. Analyze product data information to obtain factors affecting product quality;

[0007] S4. Use the Bayesian causal inference model to extract causal relationship features of product quality influencing factors and obtain a posterior probability tracking graph;

[0008] S5. Determine whether the causal relationship in the posterior probability tracking graph is stable. If not, adjust the parameters in the Bayesian causal inference model and return to step S4. If stable, execute step S6.

[0009] S6. Generate a causal relationship graph based on the posterior probability tracking graph and calculate the causal effect;

[0010] S7. Improve and optimize wearable smart products based on cause-and-effect effects.

[0011] Beneficial effects of the present invention:

[0012] This invention uses Bayesian causal inference to analyze the causal relationships between various product factors. The resulting posterior probability tracking graph determines whether the causal relationships are stable, identifies factors influencing user experience, and improves the product based on the causal inference results. The goal of this invention's model is to identify the optimal direction for product improvement based on user experience as seen in product reviews. Compared to previous vague suggestions for improvement, this invention identifies the optimal direction for improvement by identifying user experience, which can improve user satisfaction and product market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of the user experience identification method based on online reviews of wearable smart products of the present invention; Figure 2 Product two-dimensional (horizontal-vertical) difference attribute graph extracted from online review text

[0014] Figure 3 Examples of tracking graphs and running average graphs of the number of DAG edges (graph size) of the present invention;

[0015] Figure 4 An example of a trace graph of the posterior probability of edge inclusion calculated in the MCMC iteration of the present invention;

[0016] Figure 5 An example of a posterior distribution graph of the estimated posterior probability and graph size for edge inclusion of the present invention;

[0017] Figure 6 This is an example of the posterior distribution graph of the DAG size of the present invention;

[0018] Figure 7 Examples of the DAG, MPM DAG and MAP DAG model diagrams of the present invention;

[0019] Figure 8 These are examples of the real DAG and modified DAG model diagrams of the present invention. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] A user experience identification method based on online reviews of wearable smart products involves obtaining review data for wearable smart products from e-commerce websites; analyzing and evaluating the reviews to extract information related to product quality, size, and other factors; and conducting data analysis based on the obtained review data to identify influential factors, providing a scientific basis for product improvement and optimization. Bayesian causal inference is used to analyze the causal relationships between various product factors. The resulting posterior probability trace graph determines whether the causal relationships are stable, identifies factors influencing user experience, and uses the causal inference results to improve product performance and enhance user satisfaction and market competitiveness.

[0022] A specific implementation method of a product optimization method based on user review data, such as Figure 1 As shown, including:

[0023] S1. Obtain review data for a certain type of product (e.g., wearable smart anti-fall vest);

[0024] S2. Preprocess the review data to obtain product data information;

[0025] S3. Analyze product data information to obtain factors affecting product quality;

[0026] S4. Use the Bayesian causal inference model to extract causal relationship features of product quality influencing factors and obtain a posterior probability tracking graph;

[0027] S5. Determine whether the causal relationship in the posterior probability tracking graph is stable. If not, adjust the parameters in the Bayesian causal inference model and return to step S4. If stable, execute step S6.

[0028] S6. Generate a causal relationship graph based on the posterior probability tracking graph and calculate the causal effect;

[0029] S7. Improve and optimize wearable smart products based on cause-and-effect effects.

[0030] The review data of wearable smart products on e-commerce websites were obtained, and the multi-dimensional variables contained in the data were extracted, including review (RV), time (TM), reply (RP), number of words in the review (NV), presence or absence of pictures (IV), number of views (CV), number of useful clicks (UV), and number of words in the reply review (NP).

[0031] The review data analysis and evaluation process includes standardizing and preprocessing the review data and converting the review content into dummy variables, including product size (PS), senior-friendly (SA), product quality (PQ), product weight (PW), logistics (LS), product features (PF), and after-sales service (AS). If a review mentions content related to a dummy variable, the value is 1; otherwise, it is 0.

[0032] Get review data of wearable smart products on e-commerce websites, such as Figure 2 The following diagram shows the comments as a two-dimensional attribute diagram. The horizontal dimension displays variables contained in the crawled website data, including comments (RV), time (TM), replies (RP), comment word count (NV), presence of images (IV), number of views (CV), number of useful clicks (UV), and number of reply words (NP). The vertical dimension contains dummy variables related to the product user experience, including product size (PS), senior-friendly (SA), product quality (PQ), product weight (PW), logistics (LS), product function (PF), and after-sales service (AS).

[0033] The process of using Bayesian causal inference to analyze the causal relationships between various product factors involves importing standardized data into the program, selecting the variable (Useful Clicks UV) for analysis, and exploring the impact of the variables (Product Size PS, Elderly Applicability SA, Product Quality PQ, Product Function PF, Logistics LS, Product Weight PW, and After-Sales Service AS) on Useful Clicks UV. The program then generates a causal relationship diagram, identifies nodes in the diagram that have a causal relationship with the variable UV, calculates detailed causal effect values, and uses these values ​​to improve the product.

[0034] The Bayesian causal inference method includes: defining a Gaussian directed graph (DAG) model based on the likelihood and prior distribution of model parameters, and using the Monte Carlo (MCMC) algorithm to perform posterior inference on DAGs and DAG parameters. DAG D encodes the conditional independence between a set of nodes (variables), which can be read from the DAG using graph criteria. The result set of conditional independence embedded in D defines the DAG Markov property. First, define the Gaussian DAG model, let D = (V, E) be a DAG, (X1, ..., X q ) is a set of real-valued random variables associated with node V. Assume that (X1,……,X q ) belongs to the zero-mean Gaussian DAG model, that is:

[0035]

[0036] in is the precision (inverse covariance) matrix, is the space of symmetric positive definite precision matrices, and is the space of symmetric positive definite precision matrices D. Therefore, Satisfy the conditional independence (Markov property) encoded by D. Assume that DAG is fixed, so from the DAG parameters The dependence on D is omitted in . Another representation of model (1) is given by the related structural equation model (SEM). Specifically, let L be a (q, q) coefficient matrix such that for each (u, v) element L uv , u≠v, L uv ≠0, if and only if (u,v)∈E, and L uu = 1, for each u = 1, ..., q let D be a (q, q) diagonal matrix with (u, u) elements D uu Then the SEM expression of (1) is:

[0037]

[0038] This means reparameterizing Ω = LD -1 L T The latter is sometimes called the modified Cholesky decomposition of Ω and leads to the conclusion that Ω = Σ -1 In the node parameter {(D jj ,L <j ),j=1,…q}, which leads to:

[0039]

[0040]

[0041]

[0042]

[0043] In particular, the parameter D jj Corresponding to X j conditional variance of Based on (2), model (1) can be equivalently rewritten as:

[0044]

[0045] Consider n independent samples x i =(x i ,1,…x i,q ) T , i = 1, ..., n, from (4), and let X be an (n, q) data matrix with rows bound x i ,…,x n Then, the likelihood function is

[0046]

[0047] where X A is the (n,|A|) submatrix of X, corresponding to the set A of columns of X, I n Represents the (n,n) identity matrix. Assigns an appropriate prior distribution to the parameters (D, L) of the DAG.

[0048] Conditioned on D, a prior is specified for Ω via the DAG-Wishart prior (D, L) with rate hyperparameters U (a q×q symmetric positive definite matrix) and shape hyperparameters An important feature of the DAG-Wishart distribution is that the node parameter {(D jj ,L <j] ),j=1,…q} are a priori independent and distributed as:

[0049]

[0050] where IGa(α, β) denotes an inverse gamma distribution with shape α>0 and rate β>0, and whose expectation is β / (α-1), where α>1. And are specific parameters of the DAG model under consideration. The default parameters are: This setting ensures compatibility between the prior distributions of Markov-equivalent DAGs.

[0051] In particular, it can be shown that under this choice, any two Markov-equivalent DAGs are assigned the same marginal likelihood; then, the prior for the parameters (D, L) is given by

[0052]

[0053] The resulting prior is called a compatible DAG-Wishart distribution.

[0054] Given data X, p(D,L|D,X) has the following posterior distribution:

[0055]

[0056] in

[0057] Due to the prior independence of parameters in (7), the marginal likelihood in the DAG is

[0058]

[0059] The MCMC algorithm first performs a conditional independence test, inferring dependencies between variables by examining their conditional independence. A greedy search algorithm is then used to select the most appropriate DAG from the candidate relationships between the variables. Starting with an empty graph, edges are gradually added or removed to maximize the fit to the data and satisfy the assumption of conditional independence. A scoring function is then used to measure the quality of the DAG. Based on maximum likelihood estimation, it considers the trade-off between data fit and model complexity.

[0060] Accessed by the MCMC algorithm and DAG parameters The approximate posterior distribution of the set in space can be obtained as:

[0061]

[0062] For each (u, v), u ≠ v, the posterior probability of marginal inclusion.

[0063]

[0064] If and only if When it contains the edge u→v, The value of is 1.

[0065] The single DAG estimate output by MCMC can also be recovered. For example, consider the maximum a posteriori (MAP) DAG estimate, which corresponds to the DAG with the highest estimated posterior probability. Or construct a median probability DAG model (MPM), which is obtained by including all edges u→v with an estimated inclusion probability greater than 0.5.

[0066] Calculation of causal effect: using Ω=LD -1 L Τ The set of relationship recovery precision matrices For a given intervention goal From each causal effect coefficient h∈I is extracted from the posterior distribution, using Recover and derive causal effects.

[0067] A specific implementation of a user experience identification method based on online reviews of wearable smart products includes: constructing a model to recommend improvement solutions based on online reviews of wearable smart products, thereby achieving the best user experience and improving user satisfaction and product market competitiveness. The model acquires review data for wearable smart products. Review data for wearable smart products from e-commerce websites is obtained, and multi-dimensional variables are extracted from the data, including review (RV), time (TM), reply (RP), review word count (NV), image presence (IV), number of views (CV), number of useful clicks (UV), and number of reply words (NP). The review data is standardized and converted into dummy variables: product size (PS), age-appropriateness (SA), product quality (PQ), product weight (PW), logistics (LS), product function (PF), and after-sales service (AS). A program is used to generate a causal relationship diagram and determine whether the causal relationship diagram is the optimal solution. If it is the optimal solution, no parameters need to be changed. If it is not the optimal solution, the causal relationship diagram is regenerated after changing the parameters. The causal effect value is calculated, and an improvement solution is determined based on the results.

[0068] In this embodiment, based on the data obtained after preprocessing, the variable useful click number UV is selected for analysis to explore the influence of the variables product size PS, elderly suitability SA, product quality PQ, product function PF, logistics LS, product weight PW and after-sales service AS on the useful click number UV.

[0069] First, we approximate the joint posterior distribution of the DAG structure and the DAG parameters. Before using the MCMC output for posterior inference, it is common practice to perform some convergence checks. This function provides graphical diagnostics about the convergence of the MCMC output. The program generates a trace plot and a running average plot of the number of edges in the DAG (the size of the graph), such as Figure 3 As shown in , and a set of trace graphs of the posterior probability of edge inclusion calculated in the MCMC iteration, such as Figure 4 shown.

[0070] For each distinct pair of nodes (u, v), its posterior probability at time s (s=1,...,S) is estimated as the fraction of DAGs visited by MCMC before time s that contain directed edges u→v. The output is organized into q graphs (one for each node v=1,...,q), each summarizing the posterior probability of an edge u→v for u=1,...,q.

[0071] The tracking graph shows the evolution of the DAG size at each iteration. The horizontal axis represents the number of iterations, and the vertical axis represents the DAG size (number of edges). By observing the tracking graph, you can understand the convergence and trend of the DAG size during model training. If the tracking graph fluctuates slightly in a stable state, it indicates that the model has converged to a stable DAG.

[0072] The Running Average graph is a graphical representation of the average value calculated based on the Tracking Graph. It smooths the changes in the DAG size by taking the average of the values ​​within a certain range in the Tracking Graph. The Running Average graph helps you more clearly see the overall trend of the DAG size, reducing volatility in the Tracking Graph and making the trend more apparent.

[0073] The horizontal axis of the posterior probability trace represents the number of MCMC iteration steps, and the vertical axis represents the marginal posterior probability. By observing the trace, you can understand how the posterior probability of each marginal changes as the number of iteration steps increases during the MCMC iteration process.

[0074] The trace plot fluctuates less in the steady state, and the posterior probability fluctuates within a certain range, which indicates that the MCMC chain has converged and the final posterior probability estimate can be used for analysis.

[0075] The edge posterior probability represents the probability of each edge in the model existing under the given data. By tracking the graph, we can observe the changes in the edge posterior probability. Edges with higher edge posterior probabilities have higher confidence and importance in the model. Figure 4 As shown in the figure, nodes 5, 7, and 8 have high edge posterior probabilities and stable edges, which indicate high credibility and importance in the model. Other nodes have low posterior probabilities or unstable edges, which may indicate weak associations or uncertainty between variables.

[0076] Perform posterior inference on the DAG from the MCMC output. Use the program to generate the estimated posterior probability of edge inclusion and the posterior distribution of the graph size as follows Figure 5 shown.

[0077] Figure 5 The edge posterior probability represents the probability that the model believes each edge exists, given the data. The estimated posterior probability plot shows the model's confidence in the existence of each edge. A higher posterior probability indicates a higher model confidence in the edge's existence, while a lower posterior probability indicates a lower model confidence in the edge's existence. By observing the posterior probability plot, we can determine that the edges for nodes 4 and 6 have higher posterior probabilities, indicating that the model is more confident in the existence of these edges, given the data.

[0078] The edge posterior probabilities fluctuate little during the estimation process and tend to be stable, which means that the model's estimates of these edges are relatively stable. Stable posterior probabilities mean that the model consistently estimates the existence of edges across different MCMC iterations.

[0079] The posterior distribution reflects the uncertainty of the model's estimate of the DAG size. Figure 6 The posterior distribution graph shown has a high peak and is relatively concentrated near the DAG size of 6 and 7, so the model's estimation of the DAG size is relatively stable.

[0080] The program is used to calculate and return the set of posterior probabilities contained in the edge, arranged as a (q,q) matrix, where the (u,v)-element refers to the edge u→v, and generate a DAG model graph, such as Figure 7 shown

[0081] Figure 6 The presented model intuitively demonstrates the structure of a DAG, representing each variable as a node and using directed edges to represent dependencies between variables. The median probability DAG model (MPM) and the maximum a posteriori DAG model (MAP) are optimal DAG models derived based on posterior probability estimates. By observing the connections and directions between nodes, direct and indirect relationships between variables can be identified. This helps understand dependencies between variables and the paths of information transfer.

[0082] What is considered is the causal effect on the response variable of interest after a joint intervention on a given set of variables;

[0083] For a given DAG, the rules of do-calculus can be used to identify and estimate the causal effect of a hypothetical hard intervention on node j on node Y.

[0084] For a given response variable Y∈{X1,…,X q}, will intervene The total joint effect on Y is defined as

[0085]

[0086] Among them, for each h∈I,

[0087]

[0088] is the variable X in the joint intervention h The causal effect of Y on Y is related. In the Gaussian setting of equation (9), we get

[0089]

[0090] where Ω I =(LI )D -1 (L I ) T and

[0091]

[0092] Finally, in the j} j∈I In the combined intervention, X h The causal effect on Y(h∈I) is given by

[0093]

[0094] where ∑ I =(Ω I ) -1 ; Thus, the causal effect is a function of the precision matrix Ω (equivalent to L and D), which in turn depends on the underlying DAG D.

[0095] Consider the causal model represented by the previously generated DAG, where I = {2, 3, 4, 5, 6, 7, 8} (i.e., variables PS, SA, PQ, PF, LS, PW, AS) individually intervene in the total causal effect on node Y = 1 (i.e., variable UV). Then, calculate the total causal effect of the joint intervention on node Y = 1 (i.e., variable UV). I = {5, 7} (i.e., variables PF, PW) modify the DAG to add an edge from node PF to node PW. After modification, Figure 8 shown.

[0096] By intervening in the PW variable, we can eliminate the impact of PF on the PW variable and the UV variable. After eliminating the impact, we recalculate the causal effect. Analyzing the causal relationships between different factors reveals which factors have a significant impact on user experience, helping to prioritize improvements.

[0097] Based on the calculated causal effect values, the impact of different factors is ranked and improvements are prioritized to achieve a better user experience.

[0098] Table 1 User experience identification based on online reviews of wearable smart products (top three factors)

[0099]

[0100]

[0101] According to the ranking, the priority of product improvement is product weight first, followed by product function, and after-sales service is the least concerned factor. Based on the influencing factors, suggestions are provided for improving product design, function optimization, quality improvement, etc.

[0102] This invention identifies areas for improving the user experience of smart fall prevention devices. Based on user needs and usage habits, it proposes reasonable design solutions and improvement measures to enhance user satisfaction and performance, thereby increasing seniors' trust and acceptance of smart fall prevention devices. This not only helps improve the quality of life and safety of the elderly but also has positive implications for the development and promotion of the smart fall prevention device market. This invention provides a feasible solution for improving the user experience of smart fall prevention devices and promotes product iteration, upgrades, and updates.

[0103] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation plans of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A product optimization method based on user review data, characterized in that: include: S1. Obtain review data of wearable smart products; S2. Preprocess the review data to obtain product data information; S3. Analyze product data information to obtain factors affecting product quality; S4. Use the Bayesian causal inference model to extract causal relationship features of product quality influencing factors and obtain a posterior probability tracking diagram; specifically, the following steps are included: Step 1: Define a Gaussian directed graph model and construct an initial Gaussian directed graph D = (V, E) based on product quality influencing factors, where V is a node of the Gaussian directed graph and E is an edge of the Gaussian directed graph; Step 2: Get the set of real-valued random variables associated with node V (X1, ..., X q ), where (X1,……,X q ) satisfies the zero-mean Gaussian DAG model; Step 3: Input the random variable set into the structural equation model to obtain the structural equation expression: in, is a zero-mean distribution; Step 4: Convert the structural equation into a likelihood function; Step 5: Determine the prior parameters of the Gaussian directed graph, and use the likelihood function to verify the edges of the Gaussian directed graph based on the prior parameters to obtain the marginal likelihood value; Step 6: Use MCMC algorithm to sample the parameters of Gaussian directed graph model; Step 7: Calculate the posterior distribution based on the sampled parameter values, and calculate the expected value and maximum a posteriori estimate of the posterior distribution; Step 8: Verify the quality of sampling and the stability of parameter estimation, and output the posterior probability map; S5. Determine whether the causal relationship in the posterior probability tracking graph is stable. If not, adjust the parameters in the Bayesian causal inference model and return to step S4. If stable, execute step S6. S6. Generate a causal relationship graph based on the posterior probability tracking graph and calculate the causal effect; S7. Improve and optimize wearable smart products based on cause-and-effect effects.

2. A product optimization method based on user review data according to claim 1, characterized in that: Wearable smart product review data includes: review, time, reply, comment word count, presence or absence of pictures, number of views, number of useful clicks, and number of words in reply comments.

3. The product optimization method based on user review data according to claim 1, characterized in that: The preprocessing of the review data includes: screening the review data to obtain review data related to product information; and removing redundancy and completing the screened review data to obtain product data information.

4. The product optimization method based on user review data according to claim 1, characterized in that: The analysis of product data information includes: standardizing the product data information and converting the standardized data into virtual variables, where the virtual variables are factors affecting product quality; the virtual variables include product size, applicable age, product quality, product weight, logistics, product function and after-sales service.

5. The product optimization method based on user review data according to claim 1, characterized in that: The likelihood function is: Among them, X A is the (n,|A|) submatrix of X, where A is the set of columns of X; I n represents the (n,n) identity matrix.

6. The product optimization method based on user review data according to claim 1, characterized in that: The prior distribution is: where IGa(α, β) represents an inverse gamma distribution with shape α>0 and rate β>0, is a specific parameter of the DAG model.

7. The product optimization method based on user review data according to claim 1, characterized in that: The boundary likelihood function is: Where X represents a real-valued random variable, D is a Gaussian directed graph, and L is a coefficient matrix.

8. The product optimization method based on user review data according to claim 1, characterized in that: The posterior probability value is: Where X is a real-valued random variable associated with node V, u is the node u in the Gaussian directed graph, v is the node in the Gaussian directed graph, (u, v) is the edge from node u to node v, and S is the total number of edges.

9. The product optimization method based on user review data according to claim 1, characterized in that: Calculating the causal effect includes: recovering the precision matrix set according to the relationship of the symmetric positive definite precision matrix Ω For a given intervention goal Draw each causal effect coefficient from the posterior distribution use Restore the coefficient and get the causal effect.

Images