E-commerce recommendation method based on false comment detection optimization

Abstract

The invention relates to the field of e-commerce recommendation systems, in particular to an e-commerce recommendation method based on false comment detection optimization, and the method comprises the steps: obtaining user behavior data, user comment data and user-commodity interaction data; constructing an emotion-behavior vectorization model, an emotion-behavior matching detection model and a sequence chart neural network recommendation model based on user credibility; and obtaining a recommendation list after weakening false comments according to the user behavior data, the user comment data, the user-commodity interaction data and the models. According to the method, the matching degree of the user emotion and the behavior is quantified and is used as an index of the user credibility, so that the influence of the user behavior with high credibility on the recommendation result is increased, the influence of false comment data is effectively weakened, and the accuracy of an e-commerce platform recommendation system is improved.

Description

Technical Field

[0001] The present invention relates to the field of e-commerce recommendation systems, and in particular to an e-commerce recommendation method based on false review detection optimization. Background Art

[0002] With the rapid development of e-commerce, the recommendation system of online shopping platforms has become one of the important bases for users' purchasing decisions.

[0003] At present, the recommendation system of e-commerce platforms mainly relies on user behavior data and comment data for prediction. Traditional recommendation systems rely on data such as users' historical behavior and advertising click-through rates to make recommendations. However, false comments often manipulate recommendation algorithms through unrealistic emotional expressions or behavioral patterns, resulting in the system being unable to effectively identify and filter these data, which in turn leads to a decrease in the accuracy of the system's recommendation results.

[0004] Therefore, how to reduce the impact of false reviews that do not match emotions and behaviors on the recommendation system has become one of the important challenges of e-commerce recommendation systems. Summary of the invention

[0005] In view of this, the present invention proposes an e-commerce recommendation method based on false review detection optimization to solve the above problem. The e-commerce recommendation method based on false review detection optimization of the present invention processes data including:

[0006] S1. Obtain user behavior data, user review data, and user-product interaction data from the e-commerce platform;

[0007] S2. Construct an emotion-behavior vectorization model; input user behavior data and user comment data into the emotion-behavior vectorization model to obtain a behavior vector and an emotion vector;

[0008] S21, obtaining a historical behavior sequence feature set B of the user in the user behavior data;

[0009] S22. Embed B to get the behavior vector B representing the user behavior emb ;

[0010] S23, build and optimize the emotion encoding module;

[0011] S231, converting the user comment data into a word sequence, and processing the word sequence into candidate words;

[0012] S232, clustering the candidate words by using the DBSCAN clustering method to obtain a sentiment seed word set;

[0013] S233, using the sentiment seed word set to identify sentiment words in the user comment data, and further identifying attribute words matching the sentiment words, wherein the attribute words and the sentiment words together form attribute-sentiment words, and masking the sentiment words and attribute-sentiment words;

[0014] S234, optimizing the sentiment encoding module according to the masked sentiment words and attribute-sentiment words;

[0015] S24, input the user comment data into the sentiment encoding module to obtain the sentiment vector S i ;

[0016] S3, constructing an emotion-behavior matching detection model; inputting the behavior vector and the feature vector into the trained emotion-behavior matching detection model to obtain the matching degree W between the emotion vector and the behavior vector;

[0017] S31, optimizing generator G;

[0018] S32, optimizing the discriminator D;

[0019] S33, using the optimized generator G and the optimized discriminator D to calculate the Wasserstein distance between the emotion vector and the behavior vector;

[0020] S4. Build a sequence graph neural network recommendation model based on user credibility; input user behavior data and user comment data into the sequence graph neural network recommendation model, calculate user activity UAI, combine W and UAI to obtain the user credibility quantitative index T; predict user preferences based on T, and obtain a recommendation list after weakening false comments;

[0021] S41, calculating the user activity UAI based on the user behavior data and the user comment data, and calculating the user credibility T based on the UAI and W;

[0022] S42. Construct an adjacency matrix A∈R based on T and user-product interaction data N×M , used to represent the interaction relationship between users and items;

[0023] S43. Perform graph convolution on A to obtain the updated embedding vector

[0024] S431, calculating the normalized weight of each neighbor node u relative to the node v;

[0025] S432, performing feature aggregation on features of neighbor node u according to normalized weights;

[0026] S433, performing feature transformation on the aggregated features to obtain weighted features;

[0027] S434, apply the activation function σ to the weighted features to obtain the feature representation of node v at the (i+1) layer And use it as the updated embedding vector;

[0028] S44. Calculate the similarity between the embedded vectors to obtain the preference score of user u for item v. u,v ;

[0029] S45, according to score u,v Generate a recommendation list after weakening false reviews;

[0030] S5. Recommend products to the user based on the product recommendation list.

[0031] The beneficial effects of the present invention include: by quantifying the degree of match between user emotions and behaviors and using it as an indicator of user credibility, the influence of highly credible user behaviors on recommendation results is increased, thereby effectively weakening the influence of false review data and improving the accuracy of the e-commerce platform recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a structural schematic diagram of the e-commerce recommendation method based on false review detection optimization in the present invention;

[0033] Figure 2 Schematic diagram of the structure of the emotion encoding module in the present invention;

[0034] Figure 3 It is a structural schematic diagram of the emotion-behavior matching detection model in the present invention;

[0035] Figure 4 It is a structural schematic diagram of the sequence graph neural network recommendation model in the present invention. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution, characteristics and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0037] This embodiment includes:

[0038] S1. Obtain user behavior data, user review data, and user-product interaction data from the e-commerce platform.

[0039] S2. Construct an emotion-behavior vectorization model; input user behavior data and user comment data into the emotion-behavior vectorization model to obtain a behavior vector and an emotion vector.

[0040] S21. Obtain a historical behavior sequence feature set B of the user in the user behavior data.

[0041] Specifically, the feature set B can be expressed as:

[0042] B={b1,b2,…b n ,ΔT}

[0043] ΔT={δt1,δt2,...,δt n |δt1=0,δt i =time(b i )-time(b i-1 )i>1}

[0044] Among them, ΔT represents the time range, which is a set of time intervals, B represents the set of user behavior sequence features within ΔT, and δt i represents the i-th behavior b i With the previous behavior b i-1 The time interval between i ) indicates user behavior b i Time of occurrence.

[0045] S22. Embed B to obtain a behavior vector representing the user behavior.

[0046] Specifically, the formula for embedding representation is:

[0047] B emb =W b b i

[0048] Among them, W b is the embedding matrix, B emb is the embedding vector of the behavior, that is, the user's behavior vector.

[0049] S23. Build and optimize the emotion encoding module.

[0050] Specifically, the TF-IDF (Term Frequency–Inverse Document Frequency) method and the DBSCAN clustering method are used to obtain the sentiment seed word set; the steps include:

[0051] S231, converting the user comment data into a word sequence, and processing the word sequence into candidate words.

[0052] Specifically, when the user comment data is converted into a word sequence, word segmentation and stop word removal operations are used; the TF-IDF method is used to mine keywords in the word sequence. The formula for mining keywords using the TF-IDF method is:

[0053] TF-IDF=TF×IDF

[0054] Among them, TF is term frequency, which indicates the frequency of a term appearing in a text, and IDF is inverse document frequency, which is used to measure the importance of a term in a document.

[0055] Furthermore, sentiment words in the keywords are identified as candidate words according to a preset sentiment dictionary.

[0056] S232. Cluster the candidate words using the DBSCAN clustering method to obtain a sentiment seed word set.

[0057] Specifically, the input of the DBSCAN clustering method includes: sample set D = {d1, d2, ... d m}, domain parameters (∈, MinPts), sample distance measurement method, the output result of DBSCAN clustering method is: sentiment seed word set C = {C1, C2, …C K}.

[0058] Furthermore, the DBSCAN clustering method processes data including the following steps:

[0059] Step 1: Initialize the core object collection Initialize the number of clusters k = 0, initialize the unvisited sample set Γ = D, and divide the clusters

[0060] Step 2: For j=1,2,…m, find all core objects:

[0061] By using the distance metric, find the sample d j Neighborhood∈-subsample set Nε(d j ); if the number of samples in the subsample set satisfies |Nε(d j )|≥Minpts, the sample x j Add core object sample set: Ω=Ω∪{d j}.

[0062] Step 3: If the core object collection The algorithm ends, otherwise it goes to step 4.

[0063] Step 4: Randomly select a core object O from the core object set Ω, initialize the current cluster core object queue Ωcur = {O}, initialize the category sequence k = k + 1, and initialize the current cluster sample set C k ={O}, update the unvisited sample set Γ=Γ-{O}.

[0064] Step 5: If the current cluster core object queue Ωcur = {O}, then the current cluster C k After generation, update the cluster partition C = {C1, C2, ... C K}, update the core object set Ω = Ω-Ck , go to step 3.

[0065] Step 6: Take out a core object O′ from the current cluster core object queue Ωcur, find all ∈-neighborhood subsample sets N∈(O′) through the neighborhood distance threshold ∈, let Δ=N∈(O′)∩Γ, and update the current cluster sample set C k =C k ∪Δ, update the unvisited sample set Γ=Γ-Δ, update Ωcur=Ωcur∪(Δ∩Ω)-O′, and go to step 5.

[0066] S233, using the sentiment seed word set to identify sentiment words in the user comment data, and further identifying attribute words matching the sentiment words, the attribute words and the sentiment words together form attribute-sentiment words, and the sentiment words and attribute-sentiment words are masked.

[0067] Specifically, dependency parsing is used when identifying attribute words that match sentiment words.

[0068] Furthermore, masking is performed based on sentiment words and attribute-sentiment words. Since sentiment words and attribute words are usually key information carriers in the text, masking these words can improve the model's generalization ability for unknown or unseen data. When performing masking, ensure that the number of masked tokens does not exceed 12% of the total number of tokens in the entire sentence; if it is less than 12%, continue to mask other sentiment words in the sentence until the ratio reaches 12%.

[0069] S234, optimizing the sentiment encoding module according to the masked sentiment words and attribute-sentiment words.

[0070] Specifically, the optimization function L is used to optimize the emotion encoding module. The formula of L is:

[0071] L=L sp +L sw +L E

[0072] Among them, L sp Optimize the function for sentiment polarity objective, L sw is the target optimization function for sentiment words, L E Optimize the function for the attribute-sentiment word objective.

[0073]

[0074]

[0075] Among them, p i Indicates whether it is a sentiment word, that is, if the i-th position is a word in the input sequence mask, then p i =1, otherwise 0, Vi is the input sequence, is the polarity of the hidden sentiment word, for The probability estimate, W sp and b sp is a trainable parameter in the sentiment polarity objective optimization function, is the original sentiment word x i Based on the one-hot encoding representation of the entire vocabulary, V i The probability distribution of W sw and b sw is a trainable parameter in the target optimization function of sentiment words, m represents the number of masked attribute words in the input sequence, is the sparse representation of attribute words, for The probability estimate, W E and b E is a trainable parameter in the attribute objective optimization function.

[0076] S24, input the user comment data into the sentiment encoding module to obtain the sentiment vector S i .

[0077] Specifically, the input sequence V of the emotion encoding module is i Character vector Z from user review data i With position vector P i Spliced ​​together. i It is obtained by embedding user comment data. The word sequence is obtained by the word embedding model Word2Vec. i , P i The formula is:

[0078] P i (pos,2i)=sin(pos / 10000 2i / d )

[0079] P i (pos,2i+1)=cos(pos / 10000 2i / d )

[0080] Z i With P i The splicing method is:

[0081] V i =P i +Z i

[0082] Among them, pos represents the position of a word with a word vector length of N, ranging from 0 to N, i represents the position of the smallest unit of a word representation vector, ranging from 0 to 253, and d represents the maximum length of the word vector, which is 254.

[0083] S3. Construct an emotion-behavior matching detection model; input the behavior vector and the feature vector into the trained emotion-behavior matching detection model to obtain the matching degree W between the emotion vector and the behavior vector.

[0084] Specifically, the emotion-behavior matching detection model based on Wasserstein distance is used. The emotion-behavior matching detection model consists of a generator and a discriminator. The generator G is a projection function, which is used to map the emotion vector and the behavior vector in the same feature space, and uses the emotion vector space as the target space, and maps the behavior vector to the target space, so as to achieve a unified space. The discriminator is used to calculate the Wasserstein distance between the emotion vector and the behavior vector, and use the distance as the degree of emotion-behavior matching. It includes the following steps:

[0085] S31. Optimize generator G.

[0086] Specifically, the goal of the generator G is to transform the behavior vector B emb Mapped into the emotional space, the mapped behavior vector B emb ′=G(B emb ) and the sentiment vector as much as possible in the same space.

[0087]

[0088]

[0089] Among them, l proj is the loss function of the generator G, is the sentiment vector S i and the mapped behavior vector G(B emb ) is the expected Wasserstein distance, W(S i ,G(B emb )) is the sentiment vector S i and the mapped behavior vector G(B emb ) is the Wasserstein distance between .

[0090] S32. Optimize discriminator D.

[0091] Specifically, the goal of the discriminator D is to quantify the degree of match between the sentiment vector and the mapped behavior vector, namely the Wasserstein distance. The Wasserstein distance of fake reviewers is larger than that of real reviewers. The formula for optimizing the discriminator D is:

[0092]

[0093]

[0094] Among them, l disc is the loss function of the discriminator D, D(S i ,G(B emb ) is the discriminator D for the emotion vector and the mapped behavior vector G(B emb ) output, which represents the matching score between the sentiment vector and the behavior vector.

[0095] S33. Use the optimized generator G and the optimized discriminator D to calculate the Wasserstein distance between the emotion vector and the behavior vector.

[0096] Specifically, the calculation formula is:

[0097]

[0098] Among them, γ∈(S i ,B emb ′) is the possible joint distribution, and d(p,q) is the distance measure between samples.

[0099] S4. Construct a sequence graph neural network recommendation model based on user credibility; input user behavior data and user comment data into the sequence graph neural network recommendation model, calculate the user activity UAI, combine W and UAI to obtain the user credibility quantitative index T; predict user preferences based on T, and obtain a recommendation list after weakening false comments.

[0100] S41. Calculate the user activity UAI based on the user behavior data and the user comment data, and calculate the user credibility T based on UAI and W.

[0101] Specifically, the calculation of user activity UAI includes the calculation of user interaction score UIS, purchase frequency PF, browsing to purchase conversion rate BBCR, purchase cart activity CA, and repeat purchase rate RPR:

[0102] UAI=a·UIS+b·BBCR+c·CA+d·RPR+e·PF

[0103] UIS=number of comments + number of likes + number of shares

[0104]

[0105]

[0106]

[0107]

[0108] Among them, a, b, c, d, and e represent the weights of user interaction score, conversion rate from browsing to purchase, purchase cart activity, repeat purchase rate, and purchase frequency, respectively, and the contribution of each indicator to activity is allocated according to its weight.

[0109] Furthermore, the formula for calculating user credibility T is:

[0110] T=μ·UAI+v·W(S i ,B emb ′)

[0111] Among them, μ and v represent weights, W(S i ,B emb ′) is the matching degree of the commentator’s emotion and behavior, that is, the Wasserstein distance W.

[0112] S42. Construct an adjacency matrix A∈R based on T and user-product interaction data N×M , used to represent the interaction relationship between users and items.

[0113] Specifically, the formula for constructing the adjacency matrix A is:

[0114] A′ uv =T = μ·UAI+v·W(S i ,B emb ′)

[0115] A∈R N×M

[0116] Where N is the number of users and M is the number of items. uv It is the weight of one of the elements in the adjacency matrix A, representing the interaction intensity between the user node u and the product node v.

[0117] S43. Perform graph convolution on A to obtain the updated embedding vector

[0118] Furthermore, the graph convolution operation includes:

[0119] S431. Calculate the normalized weight of each neighbor node u relative to the node v.

[0120] S432: Aggregate the features of the neighbor node u according to the normalized weights.

[0121] S433, performing feature transformation on the aggregated features to obtain weighted features.

[0122] S434, apply the activation function σ to the weighted features to obtain the feature representation of node v at the (ι+1)th layer And use it as the updated embedding vector.

[0123] Specifically, the calculation formula of the graph convolution operation is:

[0124]

[0125] in, is the feature representation of node v at the (ι+1)th layer, σ is the activation function, N(v) is the set of neighboring nodes u of node v, A′ uv′ is the weight of one of the elements in the matrix A, representing the interaction intensity between the user node u and the product node v′. is the feature representation of node u at layer ι, W ι is the weight matrix of the ιth layer, used to transform node features.

[0126] S44. Calculate the similarity between the embedded vectors to obtain the preference score of user u for item v. u,v .

[0127] Specifically, the similarity between embedding vectors is used to predict the user's preference for items, and the calculation formula is:

[0128]

[0129] in, is the feature representation of node u at layer ι, is the feature representation of node v at layer l

[0130] S45, according to score u,v Generate a recommendation list after weakening false reviews.

[0131] Specifically, the preference scores of all items v that user u may be interested in are calculated, and the K items with the highest scores are selected as recommendations:

[0132] Recommendations u =\topK({scre u,v |u∈χ})

[0133] Among them, χ represents the set of all items, and \topK means selecting the K items with the highest scores.

[0134] S5. Recommend products to the user based on the product recommendation list.

[0135] Finally, it should be noted that the above only describes some embodiments of the present invention. For those skilled in the art, it is conceivable that various changes, modifications, substitutions and deformations may be made to these embodiments without departing from the principles and spirit of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents, and the above-mentioned actions should be covered within the scope of protection of the present invention.

Claims

1. An e-commerce recommendation method based on false review detection, characterized in that: The model's processing of data includes: S1. Obtain user behavior data, user review data, and user-product interaction data from the e-commerce platform; S2. Construct an emotion-behavior vectorization model; input user behavior data and user comment data into the emotion-behavior vectorization model to obtain a behavior vector and an emotion vector; S3, constructing an emotion-behavior matching detection model; inputting the behavior vector and the feature vector into the trained emotion-behavior matching detection model to obtain the matching degree W between the emotion vector and the behavior vector; S4. Build a sequence graph neural network recommendation model based on user credibility; input user behavior data and user comment data into the sequence graph neural network recommendation model, calculate user activity UAI, combine W and UAI to obtain the user credibility quantitative index T; predict user preferences based on T, and obtain a product recommendation list after weakening false comments; S5. Recommend products to the user based on the product recommendation list.

2. The e-commerce recommendation method based on false review detection optimization according to claim 1 is characterized in that: The emotion-behavior vectorization model processes user behavior data and user comment data including: S21, obtaining a historical behavior sequence feature set B of the user in the user behavior data; S22. Embed B to get the behavior vector B representing the user behavior emb ; S23, build and optimize the emotion encoding module; S24, splicing sequence V according to user comment data i , V i Input the emotion encoding module for pre-training to obtain the emotion vector S i .

3. The e-commerce recommendation method based on false review detection optimization according to claim 2 is characterized in that: Constructing the emotion encoding module includes: S231, converting the user comment data into a word sequence, and processing the word sequence into candidate words; S232, clustering the candidate words by using the DBSCAN clustering method to obtain a sentiment seed word set; S233, using the sentiment seed word set to identify sentiment words in the user comment data, and identifying attribute words matching the sentiment words according to the sentiment words, combining the attribute words and the sentiment words into attribute-sentiment words, and masking the sentiment words and the attribute-sentiment words; S234, optimizing the sentiment encoding module according to the masked sentiment words and attribute-sentiment words.

4. The e-commerce recommendation method based on false review detection optimization according to claim 3 is characterized in that: Processing the word sequence includes: mining keywords in the word sequence using the TF-IDF method; identifying sentiment words in the keywords as candidate words according to a preset sentiment dictionary.

5. The e-commerce recommendation method based on false review detection optimization according to claim 3 is characterized in that: Dependency syntax is used to identify sentiment words and their matching attribute words.

6. The e-commerce recommendation method based on false review detection optimization according to claim 1 is characterized in that: The emotion-behavior matching detection model consists of a generator and a discriminator. The generator is used to map the emotion vector and the behavior vector in the same feature space, and the discriminator is used to calculate the Wasserstein distance between the emotion vector and the behavior vector.

7. The e-commerce recommendation method based on false review detection optimization according to claim 1 is characterized in that: The data processing of the sequential graph neural network recommendation model based on user credibility includes: S41, calculating the user activity UAI based on the user behavior data and the user comment data, and calculating the user credibility T based on the UAI and W; S42. Construct an adjacency matrix A∈R based on T and user-product interaction data N×M , used to represent the interaction relationship between users and items; S43. Perform graph convolution on A to obtain the updated embedding vector S44. Calculate the similarity between the embedded vectors to obtain the preference score of user u for item v. u,v ; S45, according to score u,v Generate a recommendation list after weakening false reviews.

8. The e-commerce recommendation method based on false review detection optimization according to claim 7 is characterized in that: Calculating user activity UAI includes calculating user interaction score UIS, purchase frequency PF, browse-to-purchase conversion rate BBCR, purchase cart activity CA, and repeat purchase rate RPR.

9. The e-commerce recommendation method based on false review detection optimization according to claim 7 is characterized in that: The graph convolution operation on A includes: S431, calculating the normalized weight of each neighbor node u relative to the node v; S432, performing feature aggregation on features of neighbor node u according to normalized weights; S433, performing feature transformation on the aggregated features to obtain weighted features; S434, apply the activation function σ to the weighted features to obtain the feature representation of node v at the (ι+1)th layer And use it as the updated embedding vector.

Images