A method for evaluating the quality of live streaming sales based on big data analysis

Abstract

The present invention discloses a method for evaluating the quality of live streaming with goods based on big data analysis, comprising: obtaining live streaming with goods data; obtaining a feature vector of live streaming attributes based on the live streaming with goods data; obtaining a preset number of live streaming with goods quality classifications based on the live streaming with goods data; constructing a live streaming with goods quality evaluation model based on the feature vector and the live streaming with goods quality classification; inputting the feature vector into the live streaming with goods quality evaluation model to obtain the live streaming with goods quality category. The present invention adopts a support vector machine for big data learning, establishes a quantitative evaluation model between the live streaming attribute feature vector and the goods carrying result, continuously optimizes the quantitative evaluation model through real-time analysis and feedback, and is used to help improve the quality of live streaming with goods.

Description

Technical Field

[0001] The present invention belongs to the field of live streaming sales, and in particular relates to a method for evaluating the quality of live streaming sales based on big data analysis. Background Art

[0002] With the rise of live streaming, more and more companies and individuals are starting to sell goods through live streaming platforms. However, different goods require different sales strategies. The live broadcast time, the host's expression, the host's voice characteristics and language style will also have an important impact on the sales of different types of goods. However, the current live streaming hot spot mining methods often only consider the category and sales data of the goods, ignoring the impact of the live broadcast time, the host's live broadcast emotions, and the live broadcast expressions on sales performance. For example, for sports products, it is recommended that the host use a high-pitched, dynamic tone and fast speaking speed, as well as positive and sunny emotions; for leisure products, it is recommended that the host use a relaxed, gentle tone and slow speaking speed, as well as relaxed and pleasant emotions. In live streaming, the impact of the host's voice and emotional expression on the live broadcast volume is mainly reflected in the following aspects: attracting user attention, enhancing the characteristics of the goods, and adjusting the live broadcast atmosphere.

[0003] Emotional communication refers to the expression, infection and sharing of emotions between individuals or groups. The starting point of the communication is the emotions of the communicator, with the evaluation and behavioral response to objective things as the content, and ultimately leads to the emotional response and communication behavior of both the communicator and the receiver. In live streaming sales, three emotions are more prominent:

[0004] Pleasant emotions: The more positive the consumer’s emotional state is, the more likely they are to ignore their own needs and generate irrational purchasing intentions.

[0005] Moral emotions: “Moral emotions” are emotions related to the interests or happiness of society or others. They are emotions that individuals experience when they comply with or violate moral standards and norms. They mainly include disgust, empathy, guilt, shame, empathy, embarrassment, pride, etc.

[0006] Emotional performance: refers to the host’s purposeful planning and control of emotional expression in order to increase traffic in the live broadcast room and product sales, and also refers to the performative emotional interaction of the audience in order to obtain prizes, coupons and other economic benefits.

[0007] The choice of live broadcast time also has a huge impact on the sales volume. You need to know when to start the broadcast to attract more viewers, how long the anchor can control the live broadcast room, how long the live broadcast should be so as not to make the audience feel tired, etc. Consider all factors comprehensively, and then set the daily or weekly live broadcast time and duration.

[0008] Facial expressions are the richest part of human language and reflect inner emotions. People express their inner feelings through expressions such as joy, anger, sadness, and happiness. In terms of live streaming, facial expressions play an important role. According to relevant research, the changes in the host's facial expressions during live streaming have a certain correlation with consumers' desire to buy. Summary of the invention

[0009] In order to solve the above technical problems, the present invention proposes a method for evaluating the quality of live streaming with goods based on big data analysis, which can maintain and improve the quality of live streaming with goods through real-time analysis and feedback.

[0010] To achieve the above object, the present invention provides a method for evaluating the quality of live streaming with goods based on big data analysis, comprising:

[0011] Obtain live streaming sales data;

[0012] Based on the live streaming sales data, a feature vector of live streaming attributes is obtained;

[0013] Based on the live streaming sales data, a preset number of live streaming sales quality classifications are obtained;

[0014] Based on the feature vector and the live streaming sales quality classification, a live streaming sales quality evaluation model is constructed;

[0015] The feature vector is input into the live streaming sales quality evaluation model to obtain the live streaming sales quality category.

[0016] Optionally, the live streaming sales data includes: host's voice emotion data, host's facial expression data, live streaming time, user comments, and product sales data.

[0017] Optionally, the feature vector includes: a voice emotion feature vector, a product feature vector, a live broadcast time feature vector and an expression feature vector.

[0018] Optionally, obtaining the voice emotion feature vector includes:

[0019] Divide the live broadcast process into a preset number of time periods, and obtain the sound information and live broadcast sales volume of each time period;

[0020] Perform Fourier transform on the sound information and the amount of goods sold through live streaming to obtain a sound feature vector;

[0021] Based on the sound feature vector, the sound emotion feature vector is obtained.

[0022] Optionally, obtaining the product feature vector includes:

[0023] Classify the characteristics of the goods and obtain the usage characteristics and price characteristics of the goods;

[0024] Allocate a preset weight ratio to the commodities with the aforementioned usage characteristics and price characteristics;

[0025] The allocated usage characteristics and price characteristics are feature encoded to obtain the commodity feature vector.

[0026] Optionally, obtaining the live broadcast time feature vector includes:

[0027] Divide the live broadcast time into a preset number of time periods;

[0028] Based on the time period, the live broadcast time feature vector is obtained.

[0029] Optionally, obtaining the facial expression feature vector includes:

[0030] Segment the image into a preset number of target pixel regions;

[0031] Obtaining the central pixel of each target pixel area;

[0032] Compare the central pixel with the remaining pixels in the target pixel area to obtain texture information of the central pixel;

[0033] Based on the texture information, an expression feature vector of the target pixel point area is obtained.

[0034] Based on the expression feature vector of the target pixel point area, the expression feature vector is obtained.

[0035] Optionally, obtaining the feature vector includes:

[0036] The voice emotion feature vector, the product feature vector, the live broadcast time feature vector and the expression feature vector are combined and spliced ​​to obtain the feature vector.

[0037] Optionally, obtaining a preset number of live streaming sales quality categories includes:

[0038] Based on the user comments and product sales data, obtaining comprehensive sales data;

[0039] Based on the comprehensive sales data, obtaining a preset weight ratio of the user comments and the product sales data;

[0040] Based on the preset proportion weight, obtaining the sales comprehensive data sample;

[0041] Selecting a preset number of the sales comprehensive data samples as the center point, and obtaining the distance between the remaining samples and the center point;

[0042] Based on the distance between the remaining samples and the center point, the samples are classified to obtain the live streaming sales quality classification.

[0043] Optionally, constructing the live streaming sales quality evaluation model includes:

[0044] The corresponding relationship between the live streaming sales quality classification and the feature vector is used as a training set, the support vector machine is trained using the training set to construct the live streaming sales quality evaluation model.

[0045] Compared with the prior art, the present invention has the following advantages and technical effects:

[0046] The present invention improves the effect of live streaming with goods by establishing a live streaming with goods evaluation model, comprehensively considers the influence of the host's live streaming emotions and expressions on the effect of live streaming with goods, and realizes accurate live streaming; the present invention uses an emotion recognition model to extract the host's emotions from the live streaming audio, and then uses an expression recognition model to extract the host's live streaming expressions and extract its expression features, and then vectorizes the basic attributes of the goods and the live streaming time period, and makes a feature fusion input into the defined classification model, uses an unsupervised clustering K-means algorithm to distinguish different qualities of goods, and uses expert scoring to evaluate its quality of goods, and finally obtains a relatively complete live streaming with goods quality evaluation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0048] Figure 1 The present invention is a flowchart of a method for evaluating the quality of live streaming sales based on big data analysis according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0050] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0051] Example

[0052] This embodiment proposes a method for evaluating the quality of live streaming sales based on big data analysis. Figure 1 As shown, including:

[0053] 1. Obtain live streaming sales data;

[0054] Collecting live streaming sales data: You can obtain live streaming sales data through the live streaming platform API or other channels, including the host's voice data, host's expression data, live streaming time, user comments, and product sales data.

[0055] This embodiment uses crawler technology to extract the live audio of the live broadcast platform, as well as the live broadcast time, user comments, product sales data and other related data, and segments the live audio of each anchor on that day to separate the audio sequence corresponding to each product. For the acquired live audio sequence, the expression recognition algorithm is used to obtain the live expression data of each audio. For sales data, focus on the sales volume and user evaluation of each product.

[0056] 2. Based on the live streaming sales data, obtain the feature vector of live streaming attributes;

[0057] Live attribute feature vector generation: collect live data from a specified time period, including voice emotion features, product features, live time, and host expressions, and normalize them to form feature vectors. Different types of feature extraction use the following methods:

[0058] 2.1 Extraction of Voice Emotional Features

[0059] The emotion feature extraction module is mainly used to extract the emotion features of the input speech signal, including digital preprocessing, endpoint detection and emotion features. The emotion pattern recognition module is mainly composed of two parts: emotion recognition model training and emotion recognition. The emotion recognition model training part mainly uses the SVM method to establish the model of the required emotion recognition. The emotion recognition part classifies the test speech signal according to the established emotion recognition model to realize the emotion recognition of the speaker.

[0060] Since emotions belong to the large category of feelings, the classification of emotions can be equivalent to the classification of speech emotions. This patent mainly focuses on the classification of three emotions: pleasure, moral emotions, and emotional performance. The extracted emotional features are energy and resonance peak features.

[0061] Input: Live audio sequence

[0062] The live streaming sales process is divided into k time periods, and the voice information and live streaming sales volume of each time period are obtained.

[0063] Fourier transform is performed on each sampling sequence, and the variation coefficient is extracted to form the audio feature. Specifically, for each speech frame, it is windowed and Fourier transformed to obtain its power spectrum. The power spectrum is then converted to a Mel frequency spectrum and discrete cosine transform (DCT) is performed to obtain the MFCC coefficients. All MFCC coefficients are connected in series to obtain the feature vector T of the speech signal.

[0064] T=[t1,t2,t3,t4,t5...t k ], t k Represented as the MFCC coefficients of the kth frame

[0065] The energy characteristics of speech signals are strongly correlated with the expression of emotions. Speech signal energy usually has two types: short-time energy and short-time average amplitude energy. Since short-time energy requires a lot of calculation and is sensitive to high levels, the short-time average amplitude function is used here. Assume that the short-time average amplitude function of the n-th frame speech signal xn(m) is M n , then M n The estimated expression of is:

[0066] N is the total number of frames

[0067] The short-time average amplitude energy is calculated for each frame of each sequence to obtain the short-time average amplitude energy sequence M[m1,m2,m3....m k ]

[0068] Weight the eigenvector by coefficients T1[t1*m1,t2*m2,t3*m3,...t k *m k ]

[0069] Output: Emotional feature vector T1 of each sampling sequence

[0070] 2.2 Product Features

[0071] Product classification is the process of categorizing products according to certain classification standards. Usually, product classification can be classified according to different characteristics. The following are two characteristics that consumers prioritize when purchasing products:

[0072] Usage characteristics: Products can be classified according to their usage, such as sporting goods, household goods, food, etc.

[0073] Price characteristics: Goods can be classified according to price range, such as high-end luxury goods, mid-range goods, low-end goods, etc.

[0074] When users buy goods, they have certain priorities for different product features. In other words, users will assign different weights to different product features when considering buying a product. After big data analysis, the present invention assigns weights of 0.6 and 0.4 to the two product features respectively.

[0075] For usage features, one-hot encoding is used for feature encoding. For price features, Min-max normalization is used for feature normalization. Let minA and maxA be the minimum and maximum values ​​of the price of commodity A, respectively. An original value x of commodity A is mapped to a value x' in the interval [0,1] through min-max normalization. The formula is:

[0076] x' = (original data - minimum value) / (maximum value - minimum value)

[0077] The encoded feature vectors are concatenated to form a two-dimensional vector T2[t1, t2,], and the feature vectors are weighted T2[t1*0.6, t2*0.4]. Since different features usually have different numerical ranges and scales, T2 is finally normalized.

[0078] 2.3 Live broadcast time characteristics

[0079] 7:00-10:00 is the time to stay in bed. This group of people have free time and relatively stable income. At this time, there are few anchors broadcasting on the platform and little competition, which is a good time to attract fans.

[0080] Most of the people who take lunch breaks from 12:00 to 14:00 are office workers with stable and considerable income. At this time, the number of anchors broadcasting on the platform gradually increases, and the competition gradually becomes fierce. It is a time to maintain fans. Office workers take lunch breaks at this time, and those who can't sleep go to the live broadcast room.

[0081] 19:00~24:00 Peak period: The platform welcomes a peak in traffic. Both anchors and viewers flock to the platform during this period. There are many wealthy people and consumption reaches a peak. The war for anchors begins, and it is a time to stimulate consumption.

[0082] We generate a one-dimensional feature vector T3[t1], where t1∈{1, 2, 3}, 1, 2, and 3 correspond to three time periods respectively.

[0083] 2.4 Expression Feature Extraction

[0084] Because the local binary pattern (LBP) feature will not change significantly due to rotation or illumination changes, and the algorithm of the LBP feature is relatively easy to calculate, the processing of video facial expression recognition can be more effective and more real-time. Therefore, the present invention uses the LBP operator to extract the facial expression features of the characters. Local refers to the texture characteristics of a certain pixel point on the image, and in many cases refers to the relationship between this point and the surrounding pixels; the binary pattern refers to using the binarization of the central pixel as the threshold.

[0085] The simplest LBP mode divides an image into several 3×3 areas in units of pixels. The values ​​of the surrounding 8 pixels are compared with the central pixel. If they are not less than the value of the central pixel, they are set to 1, otherwise they are set to 0. In this way, the 8 pixels in the 3×3 range can generate an 8-bit binary number 00010011 (converting the binary to a decimal number is the LBP code, which has 256 types in total) after comparison with the central pixel, and the LBP value of the central pixel of the window is obtained. This value can reflect the texture information of the area. The anchor expression feature T4[t1, t2, t3....t k ], the feature dimension is related to the size of the video frame.

[0086] 2.5 Feature Vector Concatenation

[0087] VectorAssembler is a transformer that combines a given list of columns into a single vector column. In order to train ML models such as logistic regression and decision trees, it is useful to combine the original features and the features generated by different feature transformers into a feature vector. VectorAssembler accepts the following input column types: all numeric types, boolean types, and vector types. In each row, the values ​​of the input columns will be concatenated into a vector in the specified order.

[0088] 3. Based on the live streaming sales data, obtain a preset number of live streaming sales quality classifications;

[0089] Live streaming sales quality scoring: Combined with the sales volume and user comments of each time period, a weighted unsupervised clustering model is used to divide the videos into 10 categories. Then, professionals are invited to score the 10 categories, with 1 representing poor quality and 10 representing excellent quality.

[0090] 3.1 Data Preprocessing

[0091] The data we currently have includes product sales and user comments. For user comments, we need to further explore the available information. We perform keyword searches on the user comments obtained, and then conduct positive feedback statistics on the obtained keywords. Then we assign a weight of 0.9 to the live broadcast sales volume and a weight of 0.1 to the number of positive feedback obtained.

[0092] 3.2 Unsupervised Clustering

[0093] This embodiment uses the K-Means algorithm for cluster analysis. First, we randomly select 10 samples from the sample set as the center points of 10 classes, then check the distances from the remaining samples to the center points, and then divide the other samples into these 10 classes according to their distances from the center points. Here, we use the Euclidean distance as the distance standard, calculate the distance from each sample to the 10 center points, and then divide them according to the distance. For example, the volume of commodity A is 300, and the number of positive feedbacks obtained from the evaluation is 20. Then, after weighting, the sample points obtained are (300*0.9, 20*0.1). If the center point of class 1 is the closest to its Euclidean distance at this time, then the sample belongs to class 1. Calculating the similarity between two samples is equivalent to calculating the Euclidean distance between them.

[0094] 3.3 Expert Rating

[0095] For the classified samples, we invited relevant live broadcast practitioners to evaluate the 10 types with scores ranging from [1-10].

[0096] 4. Based on the feature vector and live streaming quality classification, a live streaming quality evaluation model is constructed;

[0097] Big data evaluation model generation: Based on the correspondence between the feature vector of each time period and the quality of 10 categories of live streaming sales as training data, support vector machine is used to train the big data model to build a live streaming sales evaluation model.

[0098] 4.1 Dataset Generation

[0099] For the 10 categories of live streaming sales quality, 100 labeled samples are given as training sets for each category, and 2,000 unlabeled samples (200 for each category) are tested.

[0100] 4.2 Classification Model Training

[0101] Support vector machine (SVM) is a new generation of learning algorithm developed on the basis of statistical learning theory. This algorithm has been widely used in text classification, handwriting recognition, image classification, bioinformatics and other fields. For a binary classification problem, SVM constructs an optimal hyperplane C0, which separates positive and negative samples in the form of a maximum margin. In the case of linear inseparability, SVM uses the kernel function K(P i ,P j ) maps the feature vector to a high-dimensional space, in which the linearly indistinguishable problem is transformed into a linearly separable problem.

[0102] The basic two-class SVM was expanded and 10 two-class SVMs were designed to solve the classification problem of 10 types of live streaming quality. i,x) represents the inner product function. There are three common inner product functions: polynomial function, radial basis function and sigmoid kernel function. The radial basis function is used in the experiment.

[0103] By establishing a matrix R composed of one-to-one recognition template results, a / b represents the recognition of live streaming quality category a and live streaming quality category b. If there is a feature vector x = (x1, x2, x3), the 10 output results generated after passing through 10 binary SVM classifiers are represented as vectors h (h1, h2, ..., h 10 ), using the nearest neighbor principle, the h vector is compared with each row in the matrix R to obtain the live streaming quality category to which the feature vector x belongs.

[0104] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for evaluating the quality of live streaming sales based on big data analysis, characterized in that: include: Obtain live streaming sales data; The live streaming sales data includes: the host’s voice emotion data, the host’s expression data, the live streaming time, user comments, and product sales data; Based on the live streaming sales data, a feature vector of live streaming attributes is obtained; Obtaining the sound emotion feature vector includes: Divide the live broadcast process into a preset number of time periods, and obtain the sound information and live broadcast sales volume of each time period; Perform Fourier transform on the sound information and the amount of goods sold through live streaming to obtain a sound feature vector; Based on the sound feature vector, obtaining the sound emotion feature vector; Obtaining product feature vectors includes: Classify the characteristics of the goods and obtain the usage characteristics and price characteristics of the goods; Allocate a preset weight ratio to the usage characteristics and price characteristics of the commodity; Performing feature coding on the allocated usage features and price features to obtain the commodity feature vector; Obtaining the expression feature vector includes: Segment the image into a preset number of target pixel regions; Obtaining the central pixel of each target pixel area; Compare the central pixel with the remaining pixels in the target pixel area to obtain texture information of the central pixel; Based on the texture information, an expression feature vector of the target pixel area is obtained. Based on the expression feature vector of the target pixel point area, obtaining the expression feature vector; Based on the live streaming sales data, a preset number of live streaming sales quality classifications are obtained; Get a preset number of live streaming sales quality categories including: Based on the user comments and the product sales data, obtaining comprehensive sales data; Based on the comprehensive sales data, obtaining a preset weight ratio of the user comments to the product sales data; Based on the preset proportion weight, obtaining the sales comprehensive data sample; Selecting a preset number of the sales comprehensive data samples as the center point, and obtaining the distance between the remaining samples and the center point; Based on the distance between the remaining samples and the center point, the samples are classified to obtain the live streaming sales quality classification; Based on the feature vector and the live streaming sales quality classification, a live streaming sales quality evaluation model is constructed; Input the feature vector into the live streaming sales quality evaluation model to obtain the live streaming sales quality category; The feature vectors include: a voice emotion feature vector, a product feature vector, a live broadcast time feature vector and an expression feature vector; Obtaining the live broadcast time feature vector includes: Divide the live broadcast time into a preset number of time periods; Based on the time period, obtaining the live broadcast time feature vector; Obtaining the feature vector includes: The voice emotion feature vector, the product feature vector, the live broadcast time feature vector and the expression feature vector are combined and spliced ​​to obtain the feature vector; Constructing the live streaming sales quality evaluation model includes: The corresponding relationship between the live streaming sales quality classification and the feature vector is used as a training set, the training set is used to train the support vector machine, and the live streaming sales quality evaluation model is constructed.

Images