CN111488524B - Attention-oriented semantic-sensitive label recommendation method - Google Patents
Attention-oriented semantic-sensitive label recommendation method Download PDFInfo
- Publication number
- CN111488524B CN111488524B CN202010270909.4A CN202010270909A CN111488524B CN 111488524 B CN111488524 B CN 111488524B CN 202010270909 A CN202010270909 A CN 202010270909A CN 111488524 B CN111488524 B CN 111488524B
- Authority
- CN
- China
- Prior art keywords
- user
- label
- attention
- tag
- item
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention discloses an attention-oriented semantic sensitive label recommendation method, which comprises the following steps of 1), pre-training a label by using word2vec, and embedding and representing the label with semantics; 2) integrating the embedded representation of the label into the user characteristic and the project characteristic by using an attention mechanism, and modeling the dynamic label influence of the user characteristic and the commodity characteristic; 3) and combining the user characteristics and the item characteristics with the label information for prediction to obtain a prediction result based on the label information, and completing recommendation. A new model containing label semantic information is provided, the attention mechanism is utilized to model the dynamic label influence of the user and commodity characteristics, the label information is dynamically combined with the user and commodity characteristics to improve the recommendation performance, and the effectiveness of the method is verified through experiments.
Description
Technical Field
The invention relates to the technical field of computers, in particular to a semantic sensitive label recommendation method facing attention.
Background
With the increasing of network information, information overload becomes a main problem, and the recommendation system is generated. Recommendation systems play an important role in our lives, helping users to filter out useless information and select preferred items based on user preferences, interests and observed behavior about the items. The collaborative filtering method is the most widely applied recommendation method, but the method has the problems of sparsity and cold start, and the recommendation accuracy is influenced. Therefore, more and more research is being conducted to introduce auxiliary information into the recommendation system to improve the performance of the recommendation system.
The recommendation system mainly comprises explicit feedback and implicit feedback. In one aspect, user information (e.g., comments, ratings, and related feedback) that directly reflects the user's preferences is referred to as explicit feedback. On the other hand, information that cannot directly express the user's preference for purchase history, search mode, and click method is called implicit feedback. Because the recommendation method based on the implicit feedback is more adaptive due to the implicit feedback information, the recommendation of the implicit feedback is widely concerned in recent years. However, implicit feedback information does not directly reflect the user's preferences, which presents a significant challenge to recommendations.
Among the numerous auxiliary information, the tag information plays an important role in the recommendation because it contains rich information about the user and the item. Tags play an important role in websites, which are a non-hierarchical structure of keywords for describing information and embodying item semantics, e.g., the tag systems of Delcious, last. fm and MovieLens, etc., allow users to annotate web pages, songs and movies with keywords called tags. Tags can make it easier to manage and search for items on a website, and can be viewed as an implicit score that can identify not only the characteristics of the item, but also the user's preferences. Meanwhile, the label establishes a bridge between the user and the project, establishes a relation between the user preference and the commodity characteristics, and is beneficial to designing a more accurate and effective recommendation system. The existing label-based recommendation methods usually combine label information with collaborative filtering methods, which can improve recommendation accuracy by using auxiliary information provided by labels, but still have some problems: (1) conventional collaborative filtering methods have limited ability to capture user and project features. (2) The effect of tags on user and merchandise preferences is dynamic, the effect of different tags on user and merchandise preferences is different, and users are generally more concerned with certain tag information.
Disclosure of Invention
In view of the above-mentioned shortcomings or drawbacks, it is an object of the present invention to provide an attention-oriented semantic sensitive tag recommendation method.
In order to achieve the above purpose, the technical scheme of the invention is as follows:
an attention-oriented semantically sensitive tag recommendation method, comprising:
1) pre-training the label by using word2vec, and embedding and representing the label with semantics;
2) integrating the embedded representation of the label into the user characteristic and the project characteristic by using an attention mechanism, and modeling the dynamic label influence of the user characteristic and the commodity characteristic;
3) and combining the user characteristics and the item characteristics with the label information for prediction to obtain a prediction result based on the label information, and completing recommendation.
The step 1) pre-trains the label by using word2vec, and the label is embedded and expressed, wherein the loss function of the learning process is as follows:
wherein T is the number of words, w t Is the t-th word in the word sequence, w t-c :w t+c Represents the word w t Consecutive words of the context of (a).
The tags are added to an external corpus to train the tag embedded representation, and when a tag contains multiple words, the average of the embedded representations of all words in the tag is the final embedded representation of the tag.
The step 2) specifically comprises:
2.1, acquiring attention scores of a user-label and an item-label by using a multi-layer perceptron:
w(i,t)＝h 1 (ReLU(h u u i +h t t i +b u ))+b 1
w(j,t)＝h 2 (ReLU(h v v i +h t t j +b v ))+b 2
where w (i, t) represents the user-tag attention score and w (j, t) represents the item-tag attention score u i And v j Vector representation, t, representing user i and item j i And t j Tags representing users and items, respectively; h is u ,h v And h t Is a weight parameter in the attention network, b u And b v Denotes a bias parameter, h 1 And h 2 Weight parameter representing the outer layer, b 1 And b 2 Representing the weight parameter of the outer layer, wherein the nonlinear activation function uses a ReLU function;
2.2, normalizing the attention scores of the tags by using a softmax function:
wherein: alpha is attention size, i represents user, j represents item, t represents label, c is a certain label in label set;
2.3, according to the attention score, obtaining a user and item embedding vector containing label semantic embedding, and completing modeling on dynamic label influence of user and commodity characteristics:
wherein the user and item embedding vectors are represented as:
T i representing a user tag set, T j Representing a set of item tags, p i And q is j Feature vectors representing users and items.
The step 3) combines the user characteristics and the item characteristics with the label information for prediction, and the prediction result obtained based on the label information is as follows:
t ij ＝p i q j ＝(p i,1 q j,1 ,p i,2 q j,2 ,...,p i,d ,q j,d )
wherein p is i q j The user and item features are multiplied by the elements, d represents the vector dimension.
And finally, adopting a multilayer perceptron to predict a scoring result as follows:
wherein, w 1 ，w 2 And w is a weight parameter, b 1 ，b 2 And b is a bias parameter, and b is,
is the final predicted user item score.
Compared with the prior art, the invention has the beneficial effects that:
the invention provides an attention-oriented semantic sensitive label recommendation method, which learns user/item characteristics by using a deep learning method, aims to simulate dynamic label influence by using an attention mechanism, establishes an attention model based on labels, dynamically combines label information with user and commodity characteristics, provides a new model containing label semantic information, and utilizes the attention mechanism to model the dynamic label influence of the user and commodity characteristics so as to improve recommendation performance, and combines the user and item characteristics with the label information for prediction, so that the recommendation degree is improved.
Drawings
FIG. 1 is a flow chart of the attention-oriented semantic sensitive tag recommendation method of the present invention;
FIG. 2 is a model diagram of the attention-oriented semantic sensitive tag recommendation method of the present invention;
FIG. 3 is a graph of the effect of the model of the present invention on the performance variation with different feature dimensions;
FIG. 4 is a graph of the effect of the model of the present invention on the performance variation with different numbers of recommended items.
Detailed Description
The present invention will now be described in detail with reference to the drawings, wherein the described embodiments are only some, but not all embodiments of the invention. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, belong to the scope of the present invention.
As shown in fig. 1 and 2, the present invention provides an attention-oriented semantic sensitive tag recommendation method,
1) pre-training the label by using word2vec, and embedding and representing the label with semantics;
word2vec is used to extract the semantic information of the tags and integrate the semantic information into the user and item features. In order to model the influence of dynamic tags, an attention mechanism is introduced into a model to evaluate the influence of different tags on user and item preferences, and experimental results on two real-world data sets show that the model has effective recommendation performance.
The tag contains much semantic information, which is an important factor for improving the recommendation accuracy, and in order to obtain the semantic information of the tag, we use word2vec to pre-train the embedded representation of the tag. The skip-gram word2vec model is a simplified neural language model without any non-linear hidden layers.
Pre-training the label by using word2vec, embedding the label into a representation, wherein the loss function of the learning process is as follows:
wherein T is the number of words, w t Is the t-th word in the word sequence, w t-c :w t+c Represents the word w t Consecutive words of the context of (1).
Although tags contain much information, semantic embedding is not enough to be implemented with tags. Thus, tags are added to the external corpus to train the tag-embedded representation. In addition, for a tag containing a plurality of words, the average of the embedded representations of all the words in the tag is the final embedded representation of the tag.
2) Integrating the embedded representation of the label into the user characteristic and the project characteristic by using an attention mechanism, and modeling the dynamic label influence of the user characteristic and the commodity characteristic;
the method specifically comprises the following steps:
2.1, first obtain the attention scores of the user-tag and the item-tag using the attention mechanism:
w(i,t)＝h 1 (ReLU(h u u i +h t t i +b u ))+b 1
w(j,t)＝h 2 (ReLU(h v v i +h t t j +b v ))+b 2
where w (i, t) represents the user-tag attention score and w (j, t) represents the item-tag attention score u i And v j Vector representation, t, representing user i and item j i And t j Tags representing users and items, respectively; h is u ,h v And h t Is a weight parameter in the attention network, b u And b v Denotes a bias parameter, h 1 And h 2 Weight parameter representing the outer layer, b 1 And b 2 Representing the weight parameter of the outer layer, wherein the nonlinear activation function uses a ReLU function;
2.2, normalizing the attention scores of the tags by using a softmax function:
wherein: alpha is attention size, i represents user, j represents item, t represents label, c is a certain label in label set;
2.3, according to the attention score, obtaining a user and item embedding vector containing label semantic embedding, and completing modeling on dynamic label influence of user and commodity characteristics:
wherein the user and item embedding vectors are represented as:
T i representing a set of user tags, T j Representing a set of item tags, p i And q is j Feature vectors representing users and items.
3) And combining the user characteristics and the item characteristics with the label information for prediction to obtain a prediction result based on the label information, and completing label recommendation.
The step 3) combines the user characteristics and the item characteristics with the label information for prediction, and the prediction result obtained based on the label information is as follows:
t ij ＝p i q j ＝(p i,1 q j,1 ,p i,2 q j,2 ,...,p i,d ,q j,d )
wherein p is i q j The user and item features are multiplied by the elements, d represents the vector dimension.
And finally, adopting a multilayer perceptron to predict a scoring result as follows:
wherein, w 1 ，w 2 And w is a weight parameter, b 1 ，b 2 And b is a bias parameter, and b is,
is the final predicted user item score.
And (3) optimizing and solving the model: because the hidden feedback problem is processed, a loss function based on the sequencing idea is adopted, and the target function is as follows:
where O represents the training set, (i, j, k) is an instance of the training set, representing that user i has an interaction with item j, and item k has no interaction,
representing the difference in scores of the positive and negative sample predictions, r because we deal with the implicit feedback problem ijk ＝r ij -r ik 1-0-1, the final loss function can be derived in the rightmost form, with the goal of solving for the minimum of the loss function.
The experimental results are as follows:
our model was evaluated using two datasets with tagged information: movelens 2k and last. They are publicly accessible on websites and are widely used in the evaluation of recommendation systems. Table 1 gives the statistics for the four data sets.
TABLE 1 statistics of four data sets
Baseline:
ItemPop. The method is based on the popularity of the goods, and recommends highly popular goods to the user. This is a non-personalized approach.
BPR. This is a sampling-based algorithm that can optimize the pairwise ranking between observed instances and negative instances of sampling, aiming at learning from implicit feedback.
WMF. Weight matrix decomposition, which is a traditional method for implicit feedback recommendation.
MLP. This method applies a multi-layered perceptron on top of the user and item embedding to learn scores from the data for recommendations.
NCF. The method provides a neural network architecture to model potential features of users and projects, and designs a neural network-based universal depth framework for collaborative filtering.
Setting parameters:
the optimal parameter settings for each method are either determined by our experiments or are in accordance with the settings given in the previous methods. In the model of the present invention, the model parameters were randomly initialized using a gaussian distribution (mean 0, standard deviation 0.01) and the model was optimized using a random gradient SGD. The characteristic dimensions are tested to be [8, 16, 32, 64], the learning rates are [0.001, 0.01, 0.1, 1], and the regularization parameters are [0.0001, 0.001, 0.01, 0.1], respectively. The comparison was performed using the best parameters and the best results were selected in 20 experiments. For each test case, 100 negative items were randomly selected as negative examples.
Model comparison and analysis:
the performance of all methods were first compared and the top-N recommendation method was used, with the N value set to 10 by default. The experimental results on ML2K and last. fm data sets are summarized in table 2 and table 3.
TABLE 2 Experimental results on ML2K data set
Table 3 experimental results on last. fm dataset
The results show that the model of the invention outperforms the baseline on both datasets. It can also be seen from the results: (1) the performance of NCF and MLP is superior to other benchmarks, indicating that the deep learning approach has a good ability to capture user and project features. (2) ItemPop performs the worst of all benchmarks, demonstrating the importance of personalized recommendations. (3) The model of the invention is significantly improved compared with WMF, Itemp and BPR, and also has better performance compared with NCF and MLP, which shows that the recommendation accuracy can be improved by dynamically modeling the tag information (4) the model of the invention obtains similar performance on two data sets, and the performance of other benchmarks on last.
The effect of the feature dimension K on the model on both data sets was further investigated. Experiments were performed using different values of K in [4, 8, 16, 32, 64 ]. Fig. 3 shows the results of the model on both data sets. It can be seen from the figure that the performance of the model improves with increasing value of K, mainly because larger dimensions can encode more information in the feature vectors. However, when K is greater than 32, the growth trend is not significant, so K32 is taken as a default setting to reduce complexity.
Further studying the performance of models with different values of N on the two data sets, the invention selects the top 1, 5 and 10 high-grade items to recommend. Fig. 4 shows the results of the models with different values of N. It can be seen from the figure that with a larger value of N, the model has better performance. When the N value is more than 5, the improvement is not significant. When the value of N is 1, the model is poor in performance, but when N is larger than 1, the model still has good performance, so that the model has good capability for different values of N.
It will be appreciated by those skilled in the art that the above embodiments are merely preferred embodiments of the invention, and thus, modifications and variations may be made in the invention by those skilled in the art, which will embody the principles of the invention and achieve the objects and objectives of the invention while remaining within the scope of the invention.
Claims (5)
1. An attention-oriented semantically sensitive tag recommendation method, comprising:
1) pre-training the label by using word2vec, and embedding and representing the label with semantics;
2) integrating the embedded representation of the label into the user characteristic and the project characteristic by using an attention mechanism, and modeling the dynamic label influence of the user characteristic and the commodity characteristic;
the step 2) specifically comprises the following steps:
2.1, first obtain the attention scores of the user-tag and the item-tag using the attention mechanism:
w(i,t)＝h 1 (ReLU(h u u i +h t t i +b u ))+b 1
w(j,t)＝h 2 (ReLU(h v v i +h t t j +b v ))+b 2
where w (i, t) represents the user-tag attention score and w (j, t) represents the item-tag attention score u i And v j Vector representation, t, representing user i and item j i And t j Tags representing users and items, respectively; h is u ,h v And h t Is a weight parameter in the attention network, b u And b v Denotes a bias parameter, h 1 And h 2 Weight parameter representing the outer layer, b 1 And b 2 Representing the weight parameter of the outer layer, wherein the nonlinear activation function uses a ReLU function;
2.2, normalizing the attention scores of the tags by using a softmax function:
wherein: alpha is attention size, i represents user, j represents item, t represents label, c is a certain label in label set;
2.3, according to the attention score, obtaining a user and item embedding vector containing label semantic embedding, and completing modeling on dynamic label influence of user and commodity characteristics:
wherein the user and item embedding vectors are represented as:
T i representing a set of user tags, T j Representing a set of item tags, p i And q is j Feature vectors representing users and items;
3) and combining the user characteristics and the item characteristics with the label information for prediction to obtain a prediction result based on the label information, and completing recommendation.
2. The attention-oriented semantic sensitive tag recommendation method according to claim 1, wherein the step 1) pre-trains the tag features by using word2vec, and embeds the tags into the representation, and the loss function of the pre-training process is:
wherein T is the number of words, w t Is the t-th word in the word sequence, w t-c :w t+c Represents the word w t Consecutive words of the context of (1).
3. The attention-oriented semantic sensitive tag recommendation method according to claim 2, wherein tags are added to an external corpus to train tag embedded representations, and when a tag contains multiple words, the average of the embedded representations of all words in the tag is the final embedded representation of the tag.
4. The attention-oriented semantic sensitive tag recommendation method according to claim 1, wherein the step 3) combines user features and item features with tag information for prediction, and obtains prediction results with tag-based information as follows:
t ij ＝p i q j ＝(p i,1 q j,1 ,p i,2 q j,2 ,...,p i,d ,q j,d )
wherein p is i q j The user and item features are multiplied by the elements, d represents the vector dimension.
5. The attention-oriented semantic sensitive tag recommendation method according to claim 1, wherein the final multi-layer perceptron prediction scoring result is:
wherein w 1 ，w 2 And w is a weight parameter, b 1 ，b 2 And b is a bias parameter, and b is,
is the final predicted user item score.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010270909.4A CN111488524B (en) | 2020-04-08 | 2020-04-08 | Attention-oriented semantic-sensitive label recommendation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010270909.4A CN111488524B (en) | 2020-04-08 | 2020-04-08 | Attention-oriented semantic-sensitive label recommendation method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111488524A CN111488524A (en) | 2020-08-04 |
| CN111488524B true CN111488524B (en) | 2022-08-16 |
Family
ID=71810994
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010270909.4A Active CN111488524B (en) | 2020-04-08 | 2020-04-08 | Attention-oriented semantic-sensitive label recommendation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111488524B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2022165714A1 (en) * | 2021-02-04 | 2022-08-11 | 深圳大学 | Information recommendation method, electronic device, and storage medium |
| CN113434761B (en) * | 2021-06-25 | 2024-02-02 | 平安科技（深圳）有限公司 | Recommendation model training method, device, computer equipment and storage medium |
| CN113420221B (en) * | 2021-07-01 | 2022-09-09 | 宁波大学 | Interpretable recommendation method integrating implicit article preference and explicit feature preference of user |
| CN114493535B (en) * | 2022-01-28 | 2022-09-06 | 北京乐开科技有限责任公司 | Asset model-based data center system for data driving |
| TWI818501B (en) * | 2022-04-07 | 2023-10-11 | 創鑫智慧股份有限公司 | Embedding table generation method and embedding table condensation method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109903138A (en) * | 2019-02-28 | 2019-06-18 | 华中科技大学 | A kind of individual commodity recommendation method |
| CN110008377A (en) * | 2019-03-27 | 2019-07-12 | 华南理工大学 | A method of film recommendation is carried out using user property |
| CN110517121A (en) * | 2019-09-23 | 2019-11-29 | 重庆邮电大学 | Method of Commodity Recommendation and the device for recommending the commodity based on comment text sentiment analysis |
| CN110717037A (en) * | 2019-08-28 | 2020-01-21 | 阿里巴巴集团控股有限公司 | Method and device for classifying users |
| CN110879864A (en) * | 2019-10-29 | 2020-03-13 | 南京大学 | Context recommendation method based on graph neural network and attention mechanism |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7630986B1 (en) * | 1999-10-27 | 2009-12-08 | Pinpoint, Incorporated | Secure data interchange |
| CN109670121A (en) * | 2018-12-18 | 2019-04-23 | 辽宁工程技术大学 | Project level and feature level depth Collaborative Filtering Recommendation Algorithm based on attention mechanism |
| CN110188272B (en) * | 2019-05-27 | 2023-04-21 | 南京大学 | Community question-answering website label recommendation method based on user background |
| CN110569353B (en) * | 2019-07-03 | 2023-04-07 | 重庆大学 | Attention mechanism-based Bi-LSTM label recommendation method |
-
2020
- 2020-04-08 CN CN202010270909.4A patent/CN111488524B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109903138A (en) * | 2019-02-28 | 2019-06-18 | 华中科技大学 | A kind of individual commodity recommendation method |
| CN110008377A (en) * | 2019-03-27 | 2019-07-12 | 华南理工大学 | A method of film recommendation is carried out using user property |
| CN110717037A (en) * | 2019-08-28 | 2020-01-21 | 阿里巴巴集团控股有限公司 | Method and device for classifying users |
| CN110517121A (en) * | 2019-09-23 | 2019-11-29 | 重庆邮电大学 | Method of Commodity Recommendation and the device for recommending the commodity based on comment text sentiment analysis |
| CN110879864A (en) * | 2019-10-29 | 2020-03-13 | 南京大学 | Context recommendation method based on graph neural network and attention mechanism |
Non-Patent Citations (3)
| Title |
|---|
| 一种融合社交信息的注意力推荐模型;柴超群;《中国优秀硕士学位论文全文数据库（信息科技辑）》;20191215(第12期);I138-705 * |
| 融合注意力机制的深度协同过滤推荐算法;王永贵 等;《计算机工程与应用》;20190418;第55卷(第13期);8-14 * |
| 采用词向量注意力机制的双路卷积神经网络句子分类模型;郭宝震 等;《浙江大学学报（工学版）》;20180816;第52卷(第09期);1729-1737 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111488524A (en) | 2020-08-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111488524B (en) | Attention-oriented semantic-sensitive label recommendation method | |
| WO2020211566A1 (en) | Method and device for making recommendation to user, computing apparatus, and storage medium | |
| CN111931062B (en) | Training method and related device of information recommendation model | |
| Qiu et al. | Predicting customer purchase behavior in the e-commerce context | |
| Gao et al. | An attention-based unsupervised adversarial model for movie review spam detection | |
| US20190392082A1 (en) | Comprehensive search engine scoring and modeling of user relevance | |
| CN110879864B (en) | Context recommendation method based on graph neural network and attention mechanism | |
| CN106462626A (en) | Modeling interestingness with deep neural networks | |
| CN111553754A (en) | Updating method and device of behavior prediction system | |
| US9317584B2 (en) | Keyword index pruning | |
| CN112016002A (en) | Mixed recommendation method integrating comment text level attention and time factors | |
| CN109598586A (en) | A kind of recommended method based on attention model | |
| Liu et al. | Using collaborative filtering algorithms combined with Doc2Vec for movie recommendation | |
| CN110110218B (en) | Identity association method and terminal | |
| CN116521906B (en) | Meta description generation method, device, equipment and medium thereof | |
| Elahi et al. | Hybrid recommendation by incorporating the sentiment of product reviews | |
| CN116680363A (en) | Emotion analysis method based on multi-mode comment data | |
| CN111897999B (en) | Deep learning model construction method for video recommendation and based on LDA | |
| CN113239159A (en) | Cross-modal retrieval method of videos and texts based on relational inference network | |
| CN111985520A (en) | Multi-mode classification method based on graph convolution neural network | |
| Zahrawi et al. | Implementing recommender systems using machine learning and knowledge discovery tools | |
| CN112434130B (en) | Multi-task label embedded emotion analysis neural network model construction method | |
| CN112463966B (en) | False comment detection model training method, false comment detection model training method and false comment detection model training device | |
| Ye et al. | A collaborative neural model for rating prediction by leveraging user reviews and product images | |
| CN112214511A (en) | API recommendation method based on WTP-WCD algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |