WO2022088417A1 - 一种基于异质特征深度残差网络的内容推荐方法 - Google Patents

一种基于异质特征深度残差网络的内容推荐方法 Download PDF

Info

Publication number
WO2022088417A1
WO2022088417A1 PCT/CN2020/136144 CN2020136144W WO2022088417A1 WO 2022088417 A1 WO2022088417 A1 WO 2022088417A1 CN 2020136144 W CN2020136144 W CN 2020136144W WO 2022088417 A1 WO2022088417 A1 WO 2022088417A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
content
residual network
deep residual
content recommendation
Prior art date
Application number
PCT/CN2020/136144
Other languages
English (en)
French (fr)
Inventor
蔡树彬
明仲
周槐枫
彭韬
Original Assignee
深圳大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 深圳大学 filed Critical 深圳大学
Publication of WO2022088417A1 publication Critical patent/WO2022088417A1/zh

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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Definitions

  • the invention relates to the field of communication technologies, and in particular, to a content recommendation method based on a heterogeneous feature deep residual network.
  • the technical problem to be solved by the present invention is to provide a content recommendation method based on heterogeneous feature deep residual network, aiming at solving the problem of cold start when encountering new users in the prior art. , and the prior art method directly performs operations on sparse data, resulting in high resource occupancy rate and low efficiency of the computing method.
  • an embodiment of the present invention provides a content recommendation method based on a deep residual network with heterogeneous features, wherein the method includes:
  • the predicted score data is obtained
  • performing data conversion processing on the source data to obtain weighted mixed embedded data includes:
  • a first triplet loss trainer is generated, and the embedding data is input into the first triplet loss trainer to obtain weighted mixed embedding data.
  • the extracting feature values of the source data of the content to be recommended to obtain feature value data includes:
  • obtaining the predicted score data includes:
  • the generation method of the deep residual network model is:
  • obtaining the recommendation result of the content according to the predicted scoring data includes:
  • the obtaining a recommendation result of the content according to the baseline algorithm result and the predicted score data includes:
  • the baseline score data, user deviation data and item deviation data are obtained;
  • the baseline score data, the user deviation data and the item deviation data adjust the predicted score data to obtain target predicted score data
  • the obtaining the recommendation result of the content according to the target prediction score data further includes:
  • the baseline score data is used as the target prediction score data to obtain a content recommendation result.
  • an embodiment of the present invention further provides a content recommendation device based on a heterogeneous feature deep residual network, the device comprising:
  • a source data acquisition unit of the content to be recommended configured to acquire the source data of the content to be recommended, and perform data conversion processing on the source data to obtain weighted mixed embedded data
  • a predictive scoring data acquisition unit configured to obtain predictive scoring data according to the weighted mixed embedded data and the deep residual network model
  • a content recommendation result obtaining unit configured to obtain a content recommendation result according to the predicted scoring data.
  • an embodiment of the present invention further provides an intelligent terminal including a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors
  • the one or more programs include methods for performing a heterogeneous feature deep residual network based content recommendation method as described in any of the above.
  • an embodiment of the present invention further provides a non-transitory computer-readable storage medium, when the instructions in the storage medium are executed by the processor of the electronic device, the electronic device can execute any of the above The described content recommendation method based on deep residual network with heterogeneous features.
  • the embodiment of the present invention first obtains the source data of the content to be recommended in a certain field, and performs data conversion processing on the source data to obtain weighted mixed embedded data; then according to the weighted mixed embedded data and the depth residual
  • the network model obtains the content prediction score data in the field; finally, according to the predicted score data, the final content recommendation result in the field can be given.
  • the source data of the domain content is processed, and the processed data is input into the residual network model to obtain the predicted score data, and then the domain content is accurately recommended according to the predicted score data, and the calculation method is highly efficient , the resource occupancy rate is low, and the diversified data can also avoid cold start when recommending new users.
  • FIG. 1 is a schematic flowchart of a content recommendation method based on a heterogeneous feature deep residual network according to an embodiment of the present invention
  • FIG. 2 is an overall network structure diagram of an HRN according to an embodiment of the present invention.
  • FIG. 3 is a structural diagram of heterogeneous feature extraction according to an embodiment of the present invention.
  • Fig. 4 is the variation trend diagram of the RMSE of the vector experiment result of the embodiment of the present invention
  • Fig. 4 is the variation trend diagram of LOSS of vector experiment results according to the embodiment of the present invention
  • FIG. 5 shows the variation trend diagram of MAE of vector experiment results according to the embodiment of the present invention
  • FIG. 6 Variation diagram of the deep residual network MAE according to the embodiment of the present invention
  • FIG. 7 Variation diagram of deep residual network RMSE according to an embodiment of the present invention
  • FIG. 8 is a change diagram of a deep residual network MAP according to an embodiment of the present invention.
  • FIG. 9 Variation diagram of the MRR of the deep residual network according to the embodiment of the present invention
  • Fig. 10 Variation diagram of the deep residual network NDCG according to the embodiment of the present invention
  • FIG. 11 is a schematic block diagram of a content recommendation device based on a heterogeneous feature deep residual network provided by an embodiment of the present invention.
  • FIG. 12 is a schematic block diagram of an internal structure of an intelligent terminal provided by an embodiment of the present invention.
  • the present invention discloses a content recommendation method based on heterogeneous feature deep residual network.
  • the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.
  • the content recommendation method may encounter a cold start problem when encountering a new user, and the prior art method directly performs operations on sparse data, resulting in high resource occupancy and low algorithm efficiency.
  • this embodiment provides a content recommendation method based on a heterogeneous feature deep residual network.
  • the source data of the content to be recommended is obtained, and then the user data, item data, and user item rating data are obtained.
  • data conversion processing is performed on the source data to obtain weighted mixed embedded data; then according to the weighted mixed embedded data, features are extracted, the features are vectorized, and converted into a recommendation algorithm.
  • the obtained embedding format, the weighted mixed embedded data and the extracted data are input into the deep residual network model to obtain the prediction score data; finally, the content recommendation result is obtained according to the prediction score data.
  • the core of the deep residual network in the embodiment of the present invention is HRN, which is the Chinese abbreviation of hybrid residual network.
  • the network accepts vectors and eigenvalues as input, and the vectors and eigenvalues enter a Dense layer (full connection layer) respectively.
  • Both Dense layers use tanh (hyperbolic tangent function) as the activation function to splicing the output neurons of the two Dense layers. , the input to the mixed residual network.
  • ResNet residual network
  • Dropout discretes neural network units from the network according to a certain probability
  • Batch_Normalization batch normalization
  • the embodiment of the present invention can be applied to various recommendation fields.
  • movies there may also be articles and even medical treatment.
  • movie recommendation it is first necessary to obtain items such as several movie rating data and movie-watching user data, and then Perform data conversion processing on the rating data of the above-mentioned several movies and the user data for watching movies, and after the data conversion processing, weighted mixed embedded data related to the movie content can be obtained, and feature values are extracted according to these weighted mixed embedded data, for example,
  • the rating data of the movie the movie viewing rating data of N users is extracted, the data is converted into the weighted mixed embedded data, the eigenvalue data of the weighted mixed embedded data is extracted, and the eigenvalue data and the weighted mixed embedded data are extracted.
  • Input into the deep residual network, and the prediction score data can be obtained, that is, it can be obtained that model A user and B user have many similar movie viewing hobbies.
  • the model can be divided into The content of the items with high ratings for the viewing of user B that user A has not watched is recommended to user A.
  • This embodiment provides a content recommendation method based on a heterogeneous feature deep residual network, which can be applied to an IT intelligent terminal. Specifically, as shown in Figure 1, the method includes:
  • Step S100 Acquire source data of the content to be recommended, and perform data conversion processing on the source data to obtain weighted mixed embedded data.
  • data conversion processing is performed on the source data to obtain weighted mixed embedded data.
  • the rating data of the movie the viewing rating data of N users is extracted, and the data is converted into weighted mixed embedded data, and the obtained data is prepared for subsequent HRN processing.
  • the overall network structure of HRN is shown in Figure 2, and the algorithm flow is as follows: 1) Data feature extraction and vectorization; 2) Construct content-based embedding; 3) Use user-item synergy to enhance content-based embedding, and construct weighted hybrid embedding; 4)
  • performing data conversion processing on the source data to obtain weighted mixed embedded data includes the following steps:
  • S103 Perform content-based embedding construction on the feature value data and the collaborative relationship data between the user and the item to obtain embedded data
  • S104 Generate a first triplet loss trainer, and input the embedded data into the first triplet loss trainer to obtain weighted mixed embedded data.
  • the collaborative relationship data between users and items such as user id index, item id index, user vector, item vector, and the number of user recurrences, that is, each user scores each item, and the number of item recurrences, that is, each Items have received ratings from each user, content-based user vector, content-based item vector, collaborative filtering-based user vector, collaborative filtering-based item vector.
  • the feature values of the source data of the content to be recommended are extracted to obtain feature value data.
  • Commonly used statistical-based feature extraction methods include: extracting the maximum value, average value, standard deviation/variance, etc. of the data set.
  • the first Triplet Loss module is responsible for enhancing the generated content-based embedding with user-item system relationships, training it into a weighted hybrid embedding, and then generating a module in the displayed content-based embedding vector, which is responsible for converting the input user and item data into Content-based embedding vectors.
  • the extracting eigenvalues of the source data of the content to be recommended, and obtaining the eigenvalue data includes the following operations: performing linear dimension reduction processing on the source data of the content to be recommended according to the mapping matrix, Obtaining dimensionality reduction matrix data; extracting eigenvalues of the dimensionality reducing matrix data to obtain eigenvalue data.
  • the mapping matrix the source data of the content to be recommended is subjected to linear dimensionality reduction processing, the dimensionality reduction matrix adopts PCA (principal component analysis), and the PCA uses the orthogonal transformation matrix as the mapping matrix to transform the initial data into a
  • PCA principal component analysis
  • the linearly independent representation of each dimension of the group is used to extract the main feature components of the data to achieve the purpose of expressing the overall data with data with fewer features. It is a linear dimensionality reduction method.
  • the deep residual network uses the principal component analysis method PCA (Principle Component Analysis) to reduce the dimension of the high-dimensional matrix embedding after characterization, map it to a low-dimensional space, and convert the high-dimensional sparse embedding into a low-dimensional dense embedding to improve the space. Utilization, can solve the problem of high memory usage and long computing time.
  • PCA Principal component analysis method
  • This embodiment provides a content recommendation method based on a heterogeneous feature deep residual network, which can be applied to an IT intelligent terminal. Specifically, as shown in Figure 1, the method includes:
  • the data is input into the deep residual network model.
  • the core of the deep residual network model is HRN, and the design basis of HRN is: 1) In order to reduce the "cold start" "Influence on the recommendation results, HRN designed the Triplet Loss trainer to train a weighted hybrid embedding that takes into account both content-based and collaborative filtering-based, as the input of the prediction network. 2) Secondly, due to the huge number of users and items, and the number of items that interact with users only accounts for a very small part.
  • HRN uses PCA to reduce the dimension of the embedding, and uses Triplet Loss to enhance the content-based embedding with the user-item synergy, and converts the original user and item data into low-level embeddings. dimensional, while maintaining a good feature map for weighted mixture embedding vectors. Feeding the weighted mixed embedding data into a deep residual network model results in predictive scoring data.
  • obtaining the predicted score data according to the weighted mixed embedded data and the deep residual network model includes the following steps:
  • S201 Extract heterogeneous feature data of the weighted mixed embedded data according to the weighted mixed embedded data
  • S203 Input the deep residual network input data into a deep residual network model to obtain prediction score data.
  • the heterogeneous information network HIN contains a directed graph, with nodes representing entities and edges representing relationships, which are particularly similar to knowledge graphs.
  • the system fuses the learned feature vectors into heterogeneous feature vectors by linear splicing, and linearly splices the heterogeneous feature vectors and the learned homogeneous vectors (sequences), as the input to the deep residual network. Based on this, this paper constructs the heterogeneous feature extraction shown in Figure 3.
  • the weighted mixed embedded data and the heterogeneous feature data are linearly spliced, and the formed data is used as the input data of the deep residual network; since the deep residual network model is established, the deep residual network input data Input to the deep residual network model, you can get the prediction score data.
  • the deep residual network model is generated by: obtaining input sample data and output sample data; generating a second triplet loss trainer, and according to the second triplet loss trainer The loss trainer generates a network model; the input sample data is input into the network model to obtain the network model output data; when the average absolute error rate of the network model output data and the output sample data is less than a preset value, the iteration is stopped , the deep residual network model is obtained.
  • Triplet loss is a triplet loss function, and Triplet Loss is used for Learn more refined embedding of face images. Embedding is a way to convert discrete variables into continuous vector representations. Similar images are also similar in the embedding space. Triplet Loss has finer granularity than traditional trainers. Triplet Loss is more suitable for accurate identification and is mainly used in fine-grained identification problems such as face identification, identity identification, and vehicle identification. It can greatly reduce the output dimension of the network and obtain better feature embedding.
  • the LOSS, RMSE and MAE of the network gradually decrease.
  • the LOSS, RMSE and MAE of the network decrease from 3.6843, 0.8995, 0.7122 at the beginning to 3.6234, 0.8465, and 0.6694, respectively. It can be seen that after 30 iterations, the LOSS, RMSE and MAE of the deep residual network all drop to a relatively low level.
  • HRN, HSD, HSD++ are constructed based on the same weighted hybrid embedding that integrates content-based features and collaborative filtering features.
  • Hybrid_ResNet is compared with the latter two.
  • Qualitative features which are used to mine the hidden features of user items, and input heterogeneous features into the deep residual network.
  • the residual network improves the learning efficiency.
  • the deeper network structure also helps to learn more abstraction. The characteristics of , and at the same time will not affect the network performance, which is also the reason why the system chooses the residual network.
  • HSD and HSD++ use SVD and SVD++ algorithms to train the network on the basis of weighted hybrid embedding. From the performance comparison results, the indicators of HRN are better than the latter two, which proves that the use of deep residual network to mine the hidden features of user items is indeed helpful to improve the network prediction performance.
  • This embodiment provides a content recommendation method based on a heterogeneous feature deep residual network, which can be applied to an IT intelligent terminal. Specifically, as shown in Figure 1, the method includes:
  • HRN takes weighted mixed embeddings and heterogeneous feature vectors as the input of a deep residual network, which is trained to obtain prediction score data. HRN fuses various feature vectors of different properties into heterogeneous feature vectors, and then evaluates the similarity of new users, new items and other users and items, and makes content recommendations based on these. Data sources can alleviate the "cold start" problem.
  • the obtaining the recommendation result of the content according to the predicted score data includes the following steps:
  • the Baseline baseline algorithm is an algorithm independent of the deep residual network.
  • a Baseline baseline algorithm is established, and the final recommendation result will be the result obtained by the baseline algorithm and the result obtained by the HRN. combination of results.
  • the introduction of the baseline algorithm allows the network proposed by our system to have a baseline for comparison, which helps to improve the effect of the subsequent observation algorithm.
  • the predicted scoring data is revised according to the results of the baseline algorithm, and the final scoring data is obtained. According to the final scoring data, the recommendation result of the domain content can be obtained.
  • the obtaining the recommendation result of the content according to the baseline algorithm result and the prediction score data includes the following operations: obtaining baseline score data, user deviation data and item deviation data according to the baseline algorithm result; Scoring data, the user deviation data and the item deviation data, adjust the prediction score data to obtain target prediction score data; and obtain a content recommendation result according to the target prediction score data.
  • the basic idea of the Baseline algorithm is to establish a baseline, user bias and item bias are introduced.
  • the user deviation in the movie field is the deviation of each user's viewing habits
  • the item deviation in the movie field is the deviation of each movie's work itself.
  • the baseline score data, the user deviation data and the item deviation data adjust the predicted score data to obtain the target predicted score data; that is, according to each user's movie viewing habits, the score deviation and the work itself of each movie
  • the deviation of the predicted score can be adjusted to obtain the target predicted score data, and then according to the user's usual movie viewing habits, the content recommendation result is obtained, and it is decided whether to recommend the movie content to the user.
  • the baseline rating of the movie is 7, and "Titanic" is a good movie, and its item rating is increased by 0.5, but according to the predicted rating of user A output by the deep residual network Generally, it is 0.3 lower than others, that is, user A is a harsh user, so the user rating is subtracted by 0.3, and finally the rating of user A for the movie "Titanic” is 7.2, and then according to the actual rating of user A.
  • the movie "Titanic" can be recommended to user A.
  • obtaining the recommendation result of the content according to the target prediction score data also includes the following operations: when a new user or a new item is introduced, the baseline score data is used as The target predicts the rating data, and obtains the recommendation result of the content.
  • the baseline score data can be obtained, that is, the public has a benchmark score for watching a movie, so when new users and new projects are introduced, the baseline score data is used as the target prediction score data.
  • Target prediction rating data for content recommendation. For example, when a new user is introduced, there is no information about the user's past, and the similarity between the user and other users cannot be evaluated.
  • the baseline score data is used as the target prediction score data, and the target prediction score data is used. What content the user recommends.
  • an embodiment of the present invention provides a content recommendation device based on a heterogeneous feature deep residual network.
  • the device includes: a source data acquisition unit 401 of the content to be recommended, a prediction score data acquisition unit 402, a content recommendation unit result acquisition unit 403;
  • the source data acquisition unit 401 of the content to be recommended is configured to acquire the source data of the content to be recommended, and perform data conversion processing on the source data to obtain weighted mixed embedded data;
  • the predicted scoring data obtaining unit 402 obtains predicted scoring data according to the weighted mixed embedded data and the deep residual network model
  • the content recommendation result obtaining unit 403 obtains a content recommendation result according to the predicted score data.
  • the present invention also provides an intelligent terminal, the principle block diagram of which can be shown in FIG. 12 .
  • the intelligent terminal includes a processor, a memory, a network interface, a display screen, and a temperature sensor connected through a system bus.
  • the processor of the intelligent terminal is used to provide computing and control capabilities.
  • the memory of the intelligent terminal includes a non-volatile storage medium and an internal memory.
  • the nonvolatile storage medium stores an operating system and a computer program.
  • the internal memory provides an environment for the execution of the operating system and computer programs in the non-volatile storage medium.
  • the network interface of the intelligent terminal is used for communicating with external terminals through network connection.
  • the computer program when executed by a processor, implements a content recommendation method based on a heterogeneous feature deep residual network.
  • the display screen of the smart terminal may be a liquid crystal display screen or an electronic ink display screen, and the temperature sensor of the smart terminal is pre-set inside the smart terminal to detect the operating temperature of the internal equipment.
  • FIG. 12 is only a block diagram of a partial structure related to the solution of the present invention, and does not constitute a limitation on the intelligent terminal to which the solution of the present invention is applied.
  • the specific intelligent terminal may include There are more or fewer components than shown in the figures, or some components are combined, or have a different arrangement of components.
  • an intelligent terminal includes a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors
  • One or more programs include instructions for performing the following operations: obtaining source data of the content to be recommended, and performing data conversion processing on the source data to obtain weighted mixed embedded data; according to the weighted mixed embedded data and depth residuals
  • a network model is used to obtain prediction score data; according to the prediction score data, a content recommendation result is obtained.
  • Nonvolatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM is available in various forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Road (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
  • the present invention discloses a content recommendation method, an intelligent terminal, and a storage medium based on a heterogeneous feature deep residual network.
  • the method includes: acquiring source data of the content to be recommended, and performing a process on the source data.
  • Data conversion processing to obtain weighted mixed embedded data; according to the weighted mixed embedded data and the deep residual network model, to obtain prediction score data; according to the prediction score data, to obtain a content recommendation result.
  • the source data of the domain content is processed, and the processed data is input into the residual network model to obtain the predicted score data, and then the domain content is accurately recommended according to the predicted score data, the calculation method is efficient, and the resources The share is low, and the diversified data can avoid cold start when recommending new users.
  • the present invention discloses a content recommendation method based on a heterogeneous feature deep residual network. It should be understood that the application of the present invention is not limited to the above examples. The description is improved or transformed, and all such improvements and transformations should belong to the protection scope of the appended claims of the present invention.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种基于异质特征深度残差网络的内容推荐方法，所述方法包括：获取待推荐内容的源数据，并对所述源数据进行数据转换处理，得到加权混合嵌入数据(S100)；根据所述加权混合嵌入数据和深度残差网络模型，得到预测评分数据(S200)；根据所述预测评分数据，得到内容的推荐结果(S300)。通过对领域内容的源数据进行处理，并将处理后的数据输入到残差网络模型，得到预测评分数据，接着根据预测评分数据对领域内容进行精准推荐，计算方法效率高，资源占有率低，同时多元化的数据还能避免对新用户进行推荐时出现冷启动。

Description

一种基于异质特征深度残差网络的内容推荐方法 技术领域
本发明涉及通信技术领域，尤其涉及的是一种基于异质特征深度残差网络的内容推荐方法。
背景技术
随着大数据时代的到来，人们对大数据时代中如电影，文章，医疗等领域内容推荐有极大的需求，为了对诸多领域内容进行很好的推荐，现有技术中涌现出许多方法，但是现有技术方法会存在分解举证步骤不确定性，分解后的低维矩阵会增加算法的时间复杂度，使得计算消耗高，效率低，在遇到新用户时会出现冷启动问题，并且，现有技术方法直接在稀疏数据上进行运算，资源占用率高，算法效率低。
因此，现有技术还有待改进和发展。
发明内容
本发明要解决的技术问题在于，针对现有技术的上述缺陷，提供一种基于异质特征深度残差网络的内容推荐方法，旨在解决现有技术中遇到新用户时会出现冷启动问题，并且，现有技术方法直接在稀疏数据上进行运算，资源占用率高，计算方法效率低的问题。
本发明解决问题所采用的技术方案如下：
第一方面，本发明实施例提供一种基于异质特征深度残差网络的内容推荐方法，其中，所述方法包括：
获取待推荐内容的源数据，并对所述源数据进行数据转换处理，得到加权混合嵌入数据；
根据所述加权混合嵌入数据和深度残差网络模型，得到预测评分数据；
根据所述预测评分数据，得到内容的推荐结果。
在一种实现方式中，其中，所述对所述源数据进行数据转换处理，得到加权混合嵌入数据包括：
获取用户与项目的协同关系数据；
提取所述待推荐内容的源数据的特征值，得到特征值数据；
将所述特征值数据和所述用户与项目的协同关系数据进行基于内容的嵌入构造，得到嵌入数据；
生成第一triplet loss训练器，并将所述嵌入数据输入到所述第一triplet loss训练器，得到加权混合嵌入数据。
在一种实现方式中，其中，所述提取所述待推荐内容的源数据的特征值，得到特征值数据包括：
根据映射矩阵，将所述待推荐内容的源数据进行线性降维处理，得到降维矩阵数据；
提取所述降维矩阵数据的特征值，得到特征值数据。
在一种实现方式中，其中，所述根据所述加权混合嵌入数据和深度残差网络模型，得到预测评分数据包括：
根据所述加权混合嵌入数据，提取所述加权混合嵌入数据的异质特征数据；
将所述加权混合嵌入数据和所述异质特征数据进行线性拼接，得到深度残差网络输入数据；
将所述深度残差网络输入数据输入到深度残差网络模型，得到预测评分数据。
在一种实现方式中，其中，所述深度残差网络模型生成方式为：
获取输入样本数据和输出样本数据；
生成第二triplet loss训练器，并根据所述第二triplet loss训练器，生成网络模型；
将所述输入样本数据输入到网络模型得到网络模型输出数据；
当所述网络模型输出数据和所述输出样本数据的平均绝对误差率小于预设值，则停止迭代，得到深度残差网络模型。
在一种实现方式中，其中，所述根据预测评分数据，得到内容的推荐结果包括：
根据Baseline基线算法，得到基线算法结果；
根据所述基线算法结果和所述预测评分数据，得到内容的推荐结果。
在一种实现方式中，其中，所述根据所述基线算法结果和所述预测评分数据，得到内容的推荐结果包括：
根据基线算法结果，得到基线评分数据，用户偏差数据和项目偏差数据；
根据所述基线评分数据，所述用户偏差数据和所述项目偏差数据，调整所述预测评分数据，得到目标预测评分数据；
根据所述目标预测评分数据，得到内容的推荐结果。
在一种实现方式中，其中，所述根据所述目标预测评分数据，得到内容的推荐结果还包括：
当有新用户、新项目引入时，将所述基线评分数据作为目标预测评分数据，得到内容的推荐结果。
第二方面，本发明实施例还提供一种基于异质特征深度残差网络的内容推荐装置，装置包括：
待推荐内容的源数据获取单元，用于获取待推荐内容的源数据，并对所述源数据进行数据转换处理，得到加权混合嵌入数据；
预测评分数据获取单元，用于根据所述加权混合嵌入数据和深度残差网络模型，得到预测评分数据；
内容推荐结果获取单元，用于根据所述预测评分数据，得到内容的推荐结果。
第三方面，本发明实施例还提供一种智能终端，包括有存储器，以及一个或者一个以上的程序，其中一个或者一个以上程序存储于存储器中，且经配置以由一个或者一个以上处理器执行所述一个或者一个以上程序包含用于执行如上述任意一项所述的基于异质特征深度残差网络的内容推荐方法。
第四方面，本发明实施例还提供一种非临时性计算机可读存储介质，当所述存储介质中的指令由电子设备的处理器执行时，使得电子设备能够执行如上述中任意一项所述的基于异质特征深度残差网络的内容推荐方法。
本发明的有益效果：本发明实施例首先获取某个领域的待推荐内容的源数据，并对所述源数据进行数据转换处理，得到加权混合嵌入数据；然后根据加权混合嵌入数据和深度残差网络模型得到该领域的内容预测评分数据；最后根据这个预测的评分数据，就可以给出最终该领域的内容推荐结果。可见，本发明实施例通过对领域内容的源数据进行处理，并将处理后的数据输入到残差网络模型，得到预测评分数据，接着根据预测评分数据对领域内容进行精准推荐，计算方法效率高，资源占有率低，同时多元化的数据还能避免对新用户进行推荐时出现冷启动。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明中记载的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1本发明实施例提供一种基于异质特征深度残差网络的内容推荐方法流程示意图
图2本发明实施例的HRN整体网络结构图
图3本发明实施例的异质特征提取结构图
图4左本发明实施例的向量实验结果RMSE变化趋势图
图4右本发明实施例的向量实验结果LOSS变化趋势图
图5本发明实施例的向量实验结果MAE变化趋势图
图6本发明实施例的深度残差网络MAE的变化图
图7本发明实施例的深度残差网络RMSE的变化图
图8本发明实施例的深度残差网络MAP的变化图
图9本发明实施例的深度残差网络MRR的变化图
图10本发明实施例的深度残差网络NDCG的变化图
图11本发明实施例提供的一种基于异质特征深度残差网络的内容推荐装置的原理框图。
图12本发明实施例提供的智能终端的内部结构原理框图。
具体实施方式
本发明公开了一种基于异质特征深度残差网络的内容推荐方法，为使本发明的目的、技术方案及效果更加清楚、明确，以下参照附图并举实施例对本发明进一步详细说明。应当理解，此处所描述的具体实施例仅用以解释本发明，并不用于限定本发明。
本技术领域技术人员可以理解，除非特意声明，这里使用的单数形式“一”、“一个”、“所述”和“该”也可包括复数形式。应该进一步理解的是，本发明的说明书中使用的措辞“包括”是指存在所述特征、整数、步骤、操作、元件和/或组件，但是并不排除存在或添加一个或多个其他特征、整数、步骤、操作、元件、组件和/或它们的组。应该理解，当我们称元件被“连接”或“耦接”到另一元件时，它可以直接连接或耦接到其他元件，或 者也可以存在中间元件。此外，这里使用的“连接”或“耦接”可以包括无线连接或无线耦接。这里使用的措辞“和/或”包括一个或更多个相关联的列出项的全部或任一单元和全部组合。
本技术领域技术人员可以理解，除非另外定义，这里使用的所有术语(包括技术术语和科学术语)，具有与本发明所属领域中的普通技术人员的一般理解相同的意义。还应该理解的是，诸如通用字典中定义的那些术语，应该被理解为具有与现有技术的上下文中的意义一致的意义，并且除非像这里一样被特定定义，否则不会用理想化或过于正式的含义来解释。
由于现有技术中，内容推荐方法在遇到新用户时会出现冷启动问题，并且，现有技术方法直接在稀疏数据上进行运算，资源占用率高，算法效率低。
为了解决现有技术的问题，本实施例提供了一种基于异质特征深度残差网络的内容推荐方法，先获取待推荐内容的源数据，在拿到用户数据，项目数据，用户项目评分数据后，为了方便后续对数据的处理，对所述源数据进行数据转换处理，得到加权混合嵌入数据；然后根据所述加权混合嵌入数据，抽取出特征，将特征向量化，转换成推荐算法需要用到的嵌入格式，将加权混合嵌入数据和提取的数据输入到深度残差网络模型，得到预测评分数据；最后根据所述预测评分数据，得到内容的推荐结果。本发明实施例中的深度残差网络的核心是HRN，HRN是混合残差网络的中文缩写。HRN基本网络具有四层残差网络结构，h＝4表示Principal Network(基本网络)网络深度为4，由四层ResNet(残差网络)堆叠而成。网络接受向量和特征值作为输入，向量和特征值分别进入一个Dense层(全连接层)，Dense层均采用tanh(双曲正切函数)作为激活函数，将两个Dense层的输出神经元进行拼接，输入到混合残差网络。
为了更好的提取到数据不同层次的特征，一般都会设计更深层次的网络，此外，更深的网络提取的特征越抽象，更能够捕捉数据隐藏特征。但是，如果简单的仅仅依靠增加神经网络的层数，不仅不能达到目的，而且会导致梯度爆炸。为解决这一问题，系统引入ResNet(残差网络)，在网络中加入了Dropout(按照一定的概率将神经网络单元从网路中丢弃)层来防止过拟合，引入Batch_Normalization(批规范化)层对输入数据和中间层的数据进行归一化操作，这样可以保证网络在反向传播中采用随机梯度下降，从而让网络达到收敛。
举例说明，本发明实施例可以应用于多种推荐领域，除了电影外还可以有文章甚至医疗等，比如以推荐电影为例，首先需要获取项目如若干电影评分数据，看电影的用户 数据，然后将上述若干电影的评分数据和看电影的用户数据进行数据转换处理，经过数据转换处理之后便可以得到与电影内容相关的加权混合嵌入数据，根据这些加权混合嵌入数据，提取出特征值，比如，根据电影的评分数据，提取出N个用户的观影评分数据，将这些数据进行数据转换处理得到加权混合嵌入数据，提取出加权混合嵌入数据的特征值数据，将特征值数据和加权混合嵌入数据输入到深度残差网络，可以得到预测评分数据，也即，可以得到模型A用户和B用户有着许多类似的观影爱好，最后根据预测评分数据，如B用户的观影项目很多，那么可以将A用户没有看过的B用户的观影评分高的项目内容推荐给A用户。
示例性方法
本实施例提供一种基于异质特征深度残差网络的内容推荐方法，该方法可以应用于IT智能终端。具体如图1所示，所述方法包括：
步骤S100、获取待推荐内容的源数据，并对所述源数据进行数据转换处理，得到加权混合嵌入数据。
具体地，获取待推荐内容的源数据，在拿到用户数据，项目数据，用户项目评分数据后，为了方便后续对数据的处理，对所述源数据进行数据转换处理，得到加权混合嵌入数据。例如，根据电影的评分数据，提取出N个用户的观影评分数据，将这些数据进行数据转换处理得到加权混合嵌入数据，得到的数据为后续进行HRN处理做准备。HRN整体网络结构图如图2所示，算法流程如下：1)数据特征提取与向量化；2)构造基于内容的嵌入；3)利用用户项目协同关系增强基于内容的嵌入，构造加权混合嵌入；4)
构建异质特征，拼接异质特征和加权混合嵌入；5)搭建深度残差预测网络；6)使用异质特征和加权混合嵌入对残差网络进行训练；7)得到最后的预测结果。
为了得到加权混合嵌入数据，所述对所述源数据进行数据转换处理，得到加权混合嵌入数据包括如下步骤：
S101：获取用户与项目的协同关系数据；
S102：提取所述待推荐内容的源数据的特征值，得到特征值数据；
S103：将所述特征值数据和所述用户与项目的协同关系数据进行基于内容的嵌入构造，得到嵌入数据；
S104：生成第一triplet loss训练器，并将所述嵌入数据输入到所述第一triplet loss训练器，得到加权混合嵌入数据。
具体地，获取用户与项目的协同关系数据，如用户id索引，项目id索引，用户 向量，项目向量，用户复现次数，即每个用户为每个项目打分，项目复现次数，即每个项目收到过来自每个用户的评分，基于内容的用户向量，基于内容的项目向量，基于协同过滤的用户向量，基于协同过滤的项目向量。然后提取待推荐内容的源数据的特征值，得到特征值数据。常用的基于统计的特征提取手段有：提取数据集的最值，平均值，标准差/方差等。接着将所述特征值数据和所述用户与项目的协同关系数据进行基于内容的嵌入构造，得到嵌入数据。也即将每个项目收到的来自每个用户的评分，基于内容的用户向量，基于内容的项目向量，基于协同过滤的用户向量，基于协同过滤的项目向量和数据集的最值，平均值，标准差/方差等进行嵌入式构造，得到嵌入数据。最后，生成第一triplet loss训练器并将所述嵌入数据输入到所述第一triplet loss训练器，triplet loss为三元组损失函数。第一Triplet Loss模块负责将生成的基于内容的嵌入利用用户项目系统关系进行增强，训练成加权混合嵌入，然后在展示的基于内容的嵌入向量中生成模块，负责将输入的用户和项目数据转换成基于内容的嵌入向量。
为了得到合适的特征值数据，所述提取所述待推荐内容的源数据的特征值，得到特征值数据包括如下操作：根据映射矩阵，将所述待推荐内容的源数据进行线性降维处理，得到降维矩阵数据；提取所述降维矩阵数据的特征值，得到特征值数据。
具体地，根据映射矩阵，将所述待推荐内容的源数据进行线性降维处理,降维矩阵采用PCA(主成分分析法),PCA利用正交变换矩阵作为映射矩阵，将初始数据变换为一组各维度线性无关的表示，用于提取数据的主要特征分量，达到以较少特征的数据表达整体数据的目的,是一种线性降维的方法。深度残差网络使用主成分分析方法PCA(Principle Component Analysis)对特征化后的高维矩阵嵌入进行降维，将其映射到低维空间，将高维稀疏嵌入转化为低维稠密嵌入，提高空间利用率，能够解决高内存占用和计算时间长的问题。
本实施例提供一种基于异质特征深度残差网络的内容推荐方法，该方法可以应用于IT智能终端。具体如图1所示，所述方法包括：
S200：根据所述加权混合嵌入数据和深度残差网络模型，得到预测评分数据。
具体地，本发明实施例获得所述加权混合嵌入数据后，会将数据输入到深度残差网络模型，深度残差网络模型的核心为HRN,HRN的设计依据为：1)为了减轻“冷启动”对推荐结果造成的影响，HRN设计了Triplet Loss训练器，训练出同时兼顾了基于内容和基于协同过滤的加权混合嵌入，作为预测网络的输入。2)其次，由于用户和项目的数目巨大，并且与用户有交互的项目的数量只占了极小部分。为了避免在极度稀疏的用户 到项目评价矩阵上直接进行运算，HRN利用PCA对嵌入进行降维，用Triplet Loss以用户项目协同关系去增强基于内容的嵌入，将原始的用户和项目数据转换成低维度的，同时保持良好的特征映射的加权混合嵌入向量。将加权混合嵌入数据输入到深度残差网络模型后会得到预测评分数据。
为了得到预测评分数据，所述根据所述加权混合嵌入数据和深度残差网络模型，得到预测评分数据包括如下步骤：
S201：根据所述加权混合嵌入数据，提取所述加权混合嵌入数据的异质特征数据；
S202：将所述加权混合嵌入数据和所述异质特征数据进行线性拼接，得到深度残差网络输入数据；
S203：将所述深度残差网络输入数据输入到深度残差网络模型，得到预测评分数据。
实际中，在获取到加权混合嵌入数据之后，会提取加权混合嵌入数据的异质特征数据，异质特征提取旨在挖掘不同类型特征间的隐含关系。异质信息网络HIN包含有向图，用节点表示实体，边来表示关系，与知识图谱特别相似。受到HIN的异质信息网络的启发，系统将学习到的各种特征向量通过线性拼接的方式融合成异质特征向量，将异质特征向量与学习到的同质向量(序列)进行线性拼接，作为深度残差网络的输入。基于此，本文构建了如图3所示的异质特征提取。然后将所述加权混合嵌入数据和所述异质特征数据进行线性拼接，形成的数据作为深度残差网络输入数据；由于深度残差网络模型是建立好的，将所述深度残差网络输入数据输入到深度残差网络模型，就可以得到预测评分数据。
为了使用深度残差网络模型，先要生成深度残差网络模型，深度残差网络模型生成方式为：获取输入样本数据和输出样本数据；生成第二triplet loss训练器，并根据所述第二triplet loss训练器，生成网络模型；将所述输入样本数据输入到网络模型得到网络模型输出数据；当所述网络模型输出数据和所述输出样本数据的平均绝对误差率小于预设值，则停止迭代，得到深度残差网络模型。
实际中，网络模型需要根据一组真实的样本数据来训练模型得到，因此，需要获取输入数据和输出数据，然后生成第二triplet loss训练器，triplet loss为三元组损失函数，Triplet Loss用于学习人脸图像更精细化的embedding，embedding为将离散变量转为连续向量表示的一个方式，相似的图像在embedding空间里也是相近的。Triplet Loss比起传统的训练器，拥有更细的粒度。Triplet Loss更适合精确的识别，主要应用于人脸识别、身份鉴定、车辆鉴定等细粒度识别问题中，可大大降低网络的输出维度，也能获 得更好的特征嵌入。然后将输入样本数据输入到网络模型得到网络模型输出数据，这个输出数据是网络模型根据输入数据得到的，为了得到最终的深度残差网络模型，还需要评估网络模型输出数据和所述输出样本数据的平均绝对误差率，当所述网络模型输出数据和所述输出样本数据的平均绝对误差率小于预设值，说明模型训练成功，则停止迭代，得到的深度残差网络模型就可以被使用。
本发明实施例中深度残差网络训练后，RMSE(均方根误差)随着epoch(一代训练)的增加的变化趋势如图4左所示，LOSS(损失)随着epoch(一代训练)的增加的变化趋势如图4右所示,MAE(平均绝对误差)随着epoch(一代训练)的增加的变化趋势如图5所示。
随着迭代次数的增加，网络的LOSS,RMSE以及MAE都逐渐下降，当epoch达到30时，网络的LOSS,RMSE以及MAE分别从开始的3.6843,0.8995,0.7122下降到3.6234,0.8465,0.6694。可以看到，经过30轮的迭代，深度残差网络的LOSS,RMSE以及MAE都下降到比较低的水平。
从图4左、图4右、图5可以看出，对于各项指标，参与对比的算法大体上分成了三个阵营。其中，Normal(正轨算法)、Baseline(基线算法)、CoClustering(聚类算法)归为一个阵营，SVD(奇异值分解算法)、SVD++(奇异值分解改进算法)算法归为一个阵营，还有就是基于异质特征的深度残差网络，使用加权混合嵌入训练出来的Hybrid_ResNet(HRN)算法，以及同样使用了加权混合嵌入，不过分别使用了SVD和SVD++训练出来的Hybrid_SVD(混合奇异值分解算法)以及Hybrid_SVD++(混合奇异值分解改进算法)算法所构成的阵营。可以看到，随着用户评分数的增大，各个算法的各项指标都变好了，这也说明，用户与项目的交互数据越多，算法确实越能够捕捉到用户的偏好，能够做出更好的预测。通过横向比较，发现用系统提出的方法训练出来的三个算法HRN,HSD以及HSD++表现明显优于其他算法如SVD、SVD++算法。其次，Baseline、CoClustering、Normal算法表现较差。SVD++作为SVD的改进版本，在除了系统提出的第一阵营外，表现最好，各项指标相较于SVD都有所提升。
对于系统提出的三个网络，HRN,HSD,HSD++，是基于同样的融合了基于内容特征和协同过滤特征的加权混合嵌入构建出来的，不同点在于，Hybrid_ResNet相比较后两者，还加入了异质特征，用于挖掘用户项目的隐含特征，将异质特征输入到深度残差网络，对比普通深度网络，残差网络提高了学习效率的同时，更深的网络结构也有助于学习到更抽象的特征，并且同时不会影响网络性能，这也是系统选择残差网络的原因。而 HSD和HSD++，则是在加权混合嵌入基础上，分别使用SVD和SVD++算法训练网络。从性能比较结果来看，HRN的指标比后两者都要好，证明利用深度残差网络挖掘用户项目隐含特征确实有助于提升网络预测性能。
用户与项目的交互数据越多，推荐算法就越能学习到用户的偏好，就拿ML(规范数据集)(1M)数据集来说，用户参与评分的电影越多，HRN给出的预测评分误差就越小。随着user-rating-0count用户评分数的增加，图6给出了MAE(平均绝对误差)变化情况，图7给出了RMSE(均方根误差)变化情况,图8给出了MAP(平均精度均值)的变化情况，图9给出了MRR(平均倒数排名)的变化情况，图10给出了NDCG(归一化折损累计增益)的变化情况，随着用户的评分数据越丰富，算法确实更能够捕捉到用户的行为特征偏好，做出更准确的预测。
本实施例提供一种基于异质特征深度残差网络的内容推荐方法，该方法可以应用于IT智能终端。具体如图1所示，所述方法包括：
S300：根据所述预测评分数据，得到内容的推荐结果。
具体地，HRN将加权混合嵌入和异质特征向量作为深度残差网络的输入，经过训练得到预测评分数据。HRN将各种不同性质的特征向量融合成异质特征向量，再对新用户、新项目与其他用户、项目进行相似度评估并据此做出内容的推荐，本发明实施例中的多元化的数据来源能够缓解“冷启动”问题。
为了得到详细的推荐结果，所述根据预测评分数据，得到内容的推荐结果包括如下步骤：
S301：根据Baseline基线算法，得到基线算法结果；
S302：根据所述基线算法结果和所述预测评分数据，得到内容的推荐结果。
在本实施例中，Baseline基线算法是一个独立于深度残差网络的算法，在构建深度残差网络之前，先建立一个Baseline基线算法，最终的推荐结果将是基线算法得到的结果与HRN得到的结果的结合。基线算法的引入，让我们系统提出的网络有了对比基线，有助于后续观察算法效果的提升。最后，根据基线算法结果来修正预测评分数据，得到最终的评分数据，根据最终的评分数据，就可以得到领域内容的推荐结果。
具体实施时，所述根据所述基线算法结果和所述预测评分数据，得到内容的推荐结果包括如下操作：根据基线算法结果，得到基线评分数据,用户偏差数据和项目偏差数据；根据所述基线评分数据，所述用户偏差数据和所述项目偏差数据，调整所述预测评分数据，得到目标预测评分数据；根据所述目标预测评分数据，得到内容的推荐结果。
在一种实施方式中，由于Baseline算法的基本思想是：设立基线，引入用户偏差和项目偏差。比如，电影领域的用户偏差为每个用户观影习惯评分偏差，电影领域的项目偏差为每个电影的作品本身的偏差。根据所述基线评分数据，所述用户偏差数据和所述项目偏差数据，调整所述预测评分数据，得到目标预测评分数据；也即根据每个用户观影习惯评分偏差和每个电影的作品本身的偏差，就可以对预测评分进行调整，得到目标预测评分数据，再根据此用户平时的观影习惯，得到内容的推荐结果，决定是否向该用户推荐电影内容。例如，对A用户进行观影评分，电影的观影基线评分是7，而《泰坦尼克号》是部不错的电影，其项目评分增加0.5，但是根据深度残差网络输出的A用户的预测评分普遍比别人低0.3，也即A用户是个苛刻的用户，因此用户评分减去0.3，最后得到A用户对电影《泰坦尼克号》的评分是7.2，再根据实际中A用户一般观看评分多少以上的电影，比如A用户只观看评分6.5以上的电影，此时根据评分7.2的结果，就可以给用户A推荐此电影《泰坦尼克号》。
为了解决当新用户引入时产生冷启动的问题，所述根据所述目标预测评分数据，得到内容的推荐结果还包括如下操作：当有新用户、新项目引入时，将所述基线评分数据作为目标预测评分数据，得到内容的推荐结果。
具体地，根据基线算法，是可以得到基线评分数据的，就是大众对一个电影的观影有一个基准的评分，这样当新用户、新项目引入时，将基线评分数据作为目标预测评分数据，根据目标预测评分数据来进行内容的推荐。如，当新用户引入时，没有关于用户过往的任何信息，无法将用户与其他用户进行相似度评估，这个时候根据基线评分数据来作为目标预测评分数据，根据目标预测评分数据来决定给该新用户推荐何种内容。
示例性设备
如图11中所示，本发明实施例提供一种基于异质特征深度残差网络的内容推荐装置，该装置包括：待推荐内容的源数据获取单元401,预测评分数据获取单元402,内容推荐结果获取单元403；
待推荐内容的源数据获取单元401，用于获取待推荐内容的源数据，并对所述源数据进行数据转换处理，得到加权混合嵌入数据；
预测评分数据获取单元402，根据所述加权混合嵌入数据和深度残差网络模型，得到预测评分数据；
内容推荐结果获取单元403，根据所述预测评分数据，得到内容的推荐结果。
基于上述实施例，本发明还提供了一种智能终端，其原理框图可以如图12所示。该 智能终端包括通过系统总线连接的处理器、存储器、网络接口、显示屏、温度传感器。其中，该智能终端的处理器用于提供计算和控制能力。该智能终端的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统和计算机程序。该内存储器为非易失性存储介质中的操作系统和计算机程序的运行提供环境。该智能终端的网络接口用于与外部的终端通过网络连接通信。该计算机程序被处理器执行时以实现一种基于异质特征深度残差网络的内容推荐方法。该智能终端的显示屏可以是液晶显示屏或者电子墨水显示屏，该智能终端的温度传感器是预先在智能终端内部设置，用于检测内部设备的运行温度。
本领域技术人员可以理解，图12中的原理图，仅仅是与本发明方案相关的部分结构的框图，并不构成对本发明方案所应用于其上的智能终端的限定，具体的智能终端可以包括比图中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。
在一个实施例中，提供了一种智能终端，包括有存储器，以及一个或者一个以上的程序，其中一个或者一个以上程序存储于存储器中，且经配置以由一个或者一个以上处理器执行所述一个或者一个以上程序包含用于进行以下操作的指令：获取待推荐内容的源数据，并对所述源数据进行数据转换处理，得到加权混合嵌入数据；根据所述加权混合嵌入数据和深度残差网络模型，得到预测评分数据；根据所述预测评分数据，得到内容的推荐结果。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，所述的计算机程序可存储于一非易失性计算机可读取存储介质中，该计算机程序在执行时，可包括如上述各方法的实施例的流程。其中，本发明所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限，RAM以多种形式可得，诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
综上所述，本发明公开了一种基于异质特征深度残差网络的内容推荐方法、智能终端、存储介质，所述方法包括：获取待推荐内容的源数据，并对所述源数据进行数据转 换处理，得到加权混合嵌入数据；根据所述加权混合嵌入数据和深度残差网络模型，得到预测评分数据；根据所述预测评分数据，得到内容的推荐结果。本发明实施例通过对领域内容的源数据进行处理，并将处理后的数据输入到残差网络模型，得到预测评分数据，接着根据预测评分数据对领域内容进行精准推荐，计算方法效率高，资源占有率低，同时多元化的数据还能避免对新用户进行推荐时出现冷启动。
应当理解的是，本发明公开了一种基于异质特征深度残差网络的内容推荐方法，应当理解的是，本发明的应用不限于上述的举例，对本领域普通技术人员来说，可以根据上述说明加以改进或变换，所有这些改进和变换都应属于本发明所附权利要求的保护范围。

Claims (10)

  1. 一种基于异质特征深度残差网络的内容推荐方法，其特征在于，所述方法包括：
    获取待推荐内容的源数据，并对所述源数据进行数据转换处理，得到加权混合嵌入数据；
    根据所述加权混合嵌入数据和深度残差网络模型，得到预测评分数据；
    根据所述预测评分数据，得到内容的推荐结果。
  2. 根据权利要求1所述的基于异质特征深度残差网络的内容推荐方法，其特征在于，所述对所述源数据进行数据转换处理，得到加权混合嵌入数据包括：
    获取用户与项目的协同关系数据；
    提取所述待推荐内容的源数据的特征值，得到特征值数据；
    将所述特征值数据和所述用户与项目的协同关系数据进行基于内容的嵌入构造，得到嵌入数据；
    生成第一triplet loss训练器，并将所述嵌入数据输入到所述第一triplet loss训练器，得到加权混合嵌入数据。
  3. 根据权利要求2所述的基于异质特征深度残差网络的内容推荐方法，其特征在于，所述提取所述待推荐内容的源数据的特征值，得到特征值数据包括：
    根据映射矩阵，将所述待推荐内容的源数据进行线性降维处理，得到降维矩阵数据；
    提取所述降维矩阵数据的特征值，得到特征值数据。
  4. 根据权利要求3所述的基于异质特征深度残差网络的内容推荐方法，其特征在于，所述根据所述加权混合嵌入数据和深度残差网络模型，得到预测评分数据包括：
    根据所述加权混合嵌入数据，提取所述加权混合嵌入数据的异质特征数据；
    将所述加权混合嵌入数据和所述异质特征数据进行线性拼接，得到深度残差网络输入数据；
    将所述深度残差网络输入数据输入到深度残差网络模型，得到预测评分数据。
  5. 根据权利要求4所述的基于异质特征深度残差网络的内容推荐方法，其特征在于，所述深度残差网络模型生成方式为：
    获取输入样本数据和输出样本数据；
    生成第二triplet loss训练器，并根据所述第二triplet loss训练器，生成网络模型；
    将所述输入样本数据输入到网络模型得到网络模型输出数据；
    当所述网络模型输出数据和所述输出样本数据的平均绝对误差率小于预设值，则停止迭代，得到深度残差网络模型。
  6. 根据权利要求5所述的基于异质特征深度残差网络的内容推荐方法，其特征在于，所述根据预测评分数据，得到内容的推荐结果包括：
    根据Baseline基线算法，得到基线算法结果；
    根据所述基线算法结果和所述预测评分数据，得到内容的推荐结果。
  7. 根据权利要求6所述的基于异质特征深度残差网络的内容推荐方法，其特征在于，所述根据所述基线算法结果和所述预测评分数据，得到内容的推荐结果包括：
    根据基线算法结果，得到基线评分数据,用户偏差数据和项目偏差数据；
    根据所述基线评分数据，所述用户偏差数据和所述项目偏差数据，调整所述预测评分数据，得到目标预测评分数据；
    根据所述目标预测评分数据，得到内容的推荐结果。
  8. 根据权利要求7所述的基于异质特征深度残差网络的内容推荐方法，其特征在于，所述根据所述目标预测评分数据，得到内容的推荐结果还包括：
    当有新用户、新项目引入时，将所述基线评分数据作为目标预测评分数据，得到内容的推荐结果。
  9. 一种智能终端，其特征在于，包括有存储器，以及一个或者一个以上的程序，其中一个或者一个以上程序存储于存储器中，且经配置以由一个或者一个以上处理器执行所述一个或者一个以上程序包含用于执行如权利要求1-8中任意一项所述的方法。
  10. 一种非临时性计算机可读存储介质，其特征在于，当所述存储介质中的指令由电子设备的处理器执行时，使得电子设备能够执行如权利要求1-8中任意一项所述的方法。
PCT/CN2020/136144 2020-10-29 2020-12-14 一种基于异质特征深度残差网络的内容推荐方法 WO2022088417A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011180629.0 2020-10-29
CN202011180629.0A CN112287222B (zh) 2020-10-29 2020-10-29 一种基于异质特征深度残差网络的内容推荐方法

Publications (1)

Publication Number Publication Date
WO2022088417A1 true WO2022088417A1 (zh) 2022-05-05

Family

ID=74373921

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/136144 WO2022088417A1 (zh) 2020-10-29 2020-12-14 一种基于异质特征深度残差网络的内容推荐方法

Country Status (2)

Country Link
CN (1) CN112287222B (zh)
WO (1) WO2022088417A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117975764B (zh) * 2024-03-29 2024-06-07 烟台富士智能科技有限公司 一种智能停车场的车位预约方法、设备及介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109726806A (zh) * 2017-10-30 2019-05-07 上海寒武纪信息科技有限公司 信息处理方法及终端设备
CN109933678A (zh) * 2019-03-07 2019-06-25 合肥工业大学 艺术品推荐方法、装置、可读介质及电子设备
US10360900B1 (en) * 2011-07-03 2019-07-23 Reality Analytics, Inc. System and method for taxonomically distinguishing sample data captured from sources
CN110674265A (zh) * 2019-08-06 2020-01-10 上海孚典智能科技有限公司 面向非结构化信息的特征判别与信息推荐系统
CN111651613A (zh) * 2020-07-08 2020-09-11 海南大学 一种基于知识图谱嵌入的动态推荐方法及系统

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109785062B (zh) * 2019-01-10 2021-09-24 电子科技大学 一种基于协同过滤模型的混合神经网络推荐系统
CN111506811A (zh) * 2020-03-19 2020-08-07 上海理工大学 一种基于深度残差网络的点击率预测方法
CN111488529B (zh) * 2020-06-28 2020-09-29 腾讯科技（深圳）有限公司 信息处理方法、装置、服务器及存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10360900B1 (en) * 2011-07-03 2019-07-23 Reality Analytics, Inc. System and method for taxonomically distinguishing sample data captured from sources
CN109726806A (zh) * 2017-10-30 2019-05-07 上海寒武纪信息科技有限公司 信息处理方法及终端设备
CN109933678A (zh) * 2019-03-07 2019-06-25 合肥工业大学 艺术品推荐方法、装置、可读介质及电子设备
CN110674265A (zh) * 2019-08-06 2020-01-10 上海孚典智能科技有限公司 面向非结构化信息的特征判别与信息推荐系统
CN111651613A (zh) * 2020-07-08 2020-09-11 海南大学 一种基于知识图谱嵌入的动态推荐方法及系统

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117975764B (zh) * 2024-03-29 2024-06-07 烟台富士智能科技有限公司 一种智能停车场的车位预约方法、设备及介质

Also Published As

Publication number Publication date
CN112287222A (zh) 2021-01-29
CN112287222B (zh) 2023-12-15

Similar Documents

Publication Publication Date Title
US10062039B1 (en) Methods and apparatus for asynchronous and interactive machine learning using word embedding within text-based documents and multimodal documents
WO2019140863A1 (en) Method of calculating relevancy, apparatus for calculating relevancy, data query apparatus, and non-transitory computer-readable storage medium
WO2021184902A1 (zh) 图像分类方法、装置、及其训练方法、装置、设备、介质
WO2022105117A1 (zh) 一种图像质量评价的方法、装置、计算机设备及存储介质
Xu et al. Transfer representation learning with TSK fuzzy system
Chen et al. Deep attention user-based collaborative filtering for recommendation
Birlutiu et al. Efficiently learning the preferences of people
Wei et al. Sequential transformer via an outside-in attention for image captioning
WO2023231753A1 (zh) 一种神经网络的训练方法、数据的处理方法以及设备
Tong et al. A deep discriminative and robust nonnegative matrix factorization network method with soft label constraint
Xia et al. Attention neural collaboration filtering based on GRU for recommender systems
Jing et al. Knowledge-enhanced attentive learning for answer selection in community question answering systems
CN113221003B (zh) 一种基于对偶理论的混合过滤推荐方法及系统
Chen et al. Distributed support vector machine in master–slave mode
Cheng et al. Multi-view multi-label learning with view feature attention allocation
He et al. Multi-label feature selection based on correlation label enhancement
Wang et al. Task recommendation method combining multimodal cognition and collaboration in mobile crowdsensing systems
Mai et al. Multimodal graph for unaligned multimodal sequence analysis via graph convolution and graph pooling
Chen et al. Edge data based trailer inception probabilistic matrix factorization for context-aware movie recommendation
WO2022088417A1 (zh) 一种基于异质特征深度残差网络的内容推荐方法
CN112749557A (zh) 文本处理模型的构建方法和文本处理方法
Liu et al. Semi‐supervised uncorrelated dictionary learning for colour face recognition
Sun et al. Rumour detection technology based on the BiGRU_capsule network
CN116108363A (zh) 基于标签引导的不完备多视图多标签分类方法和系统
CN114238715A (zh) 基于社会救助的问答系统、构建方法、计算机设备及介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20959576

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 110823)

122 Ep: pct application non-entry in european phase

Ref document number: 20959576

Country of ref document: EP

Kind code of ref document: A1