CN114637909A - Film recommendation system and method based on improved deep structured semantic model - Google Patents

Abstract

A movie recommendation system based on an improved deep structured semantic model, comprising a user behavior collection and processing module, an offline training module and an online recall and sorting module, the user behavior collection and processing module collects the user's interactive behavior, the offline training module The module receives the combined data output by the user behavior collection and processing module, and the online recall and sorting module takes out the user feature vector that the user has trained to obtain according to the user's attribute characteristics, and uses the approximate nearest neighbor search technology in the movie vector library. Recall is a subset of movies recommended by users; a movie recommendation method based on an improved deep structured semantic model includes steps such as user behavior collection and processing, offline training, online recall and sorting, etc. Sexual characteristics and implicit interaction information can be used to effectively mine movies that meet the user's interests, provide users with personalized recommendation services, and obtain the user's recommendation results.

Description

A movie recommendation system and method based on an improved deep structured semantic model

technical field

The invention relates to the field of semantic models, in particular to a movie recommendation system and method based on an improved deep structured semantic model.

Background technique

With the continuous development of the Internet, the size of the online video market has grown year by year. While enjoying the video feast with rich and varied forms and contents, Internet users are constantly being impacted by a large amount of redundant and invalid information. These huge amounts of data and information are far beyond what users can bear, and seriously interfere with users' correct selection of the information they need, resulting in very low information utilization, and even causing confusion and disgust to users.

As an effective means to solve the problem of "information overload", the recommendation system has been developing rapidly in recent years, and its role in Internet services is also increasing. As a carrier of rich information, movies naturally become an important research object in personalized recommendation. With the continuous growth of the number of users and movies, how to deeply mine movie information, accurately match user interests, select suitable movies for users from the vast movie library, and provide accurate personalized services has become a research hotspot in the industry.

The movie recommendation algorithm is the core of the movie recommendation system. At present, collaborative filtering is widely used. It is based on the user's viewing history to find similar users who have watched the same video as the target user, and then find other videos that these similar users like to watch. Recommended to target users. The problem with this scheme is: when users have few ratings records for movies, the recommendation effect of using collaborative filtering is very poor, that is, the cold start problem; with the continuous increase of users and movies, the collaborative filtering matrix will be very large, using the matrix The cost of decomposing the feature vectors of users and movies is also increasing. Other solutions are to use deep neural networks or recurrent neural networks to learn the historical preferences of users. At present, there is still a lot of room for development in the research of movie recommendation algorithms.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a movie recommendation system and method based on an improved deep structured semantic model, so as to solve the problems in the prior art.

In order to achieve the above purpose, the present invention provides a movie recommendation system based on an improved deep structured semantic model, including a user behavior collection and processing module, an offline training module and an online recall and sorting module,

The user behavior collection and processing module collects the user's interactive behavior log, search behavior log and play record list by burying points in the front end, and stores it in the file system feature database as user feature data. Perform data cleaning to obtain the original data set of the basic sample;

According to the raw data obtained by cleaning, the system preprocesses the user's behavior log and merges the data of the user's behavior log;

The offline training module receives the merged data samples output by the user behavior collection and processing module, re-weights the data after encoding and dimensionality reduction of the data, and extracts and mines the deep hidden semantic features of the data. match between users and movies;

The online recall and sorting module retrieves the user feature vector that has been trained by the user according to the user's attribute characteristics, and uses the approximate nearest neighbor search technology to perform vector retrieval in the movie vector library, and recalls the movie subset recommended for the user.

The offline training module includes an input layer, a self-attention layer, a feature extraction layer and a matching layer,

The user feature data output by the user behavior collection and processing module is combined and sent to the input layer, the input layer includes an encoding module and a dimension reduction module, and the input layer inputs the user characteristic data to the encoding module and the dimension reduction module;

The self-attention layer adopts the compression excitation network SENET to re-weight the data of the input layer;

The feature extraction layer uses three fully connected networks to form a deep neural network, which is used to extract and mine deep latent semantic features for the input user and movie feature vectors;

The matching layer obtains the matching score between the user and the movie by calculating the cosine similarity between them according to the extracted latent semantic feature vector.

The user feature data includes user dense features and user sparse features, the user dense features are input to the encoding module, and the user sparse features are input to the dimension reduction module;

The user sparse feature includes a sparse feature with a definite value and a variable-length sparse feature, and after the sparse feature with a definite value is input into the encoding module, the output is a low-dimensional vector;

The variable-length sparse feature includes a viewing history and a search history, and a viewing vector is obtained after the movie embedding sequence corresponding to the viewing history is weighted and averaged by the dimensionality reduction module vector;

The search history is trained by a dimension reduction module to obtain an embedding vector, and the corresponding movie embedding sequence is weighted to obtain a search vector;

The search history and viewing history are interleaved for training, and the input layer splices the processed sparse features and dense features, and uses the spliced user and movie vectors as initial embedding vectors.

The self-attention layer includes a compression module and an excitation module, and the compression module performs data compression and information aggregation on the embedding vector of each feature received from the input layer to form an initial weight vector;

The excitation module is used to perform feature crossover and maintain the output size dimension on the initial weight vector output by the compression module;

The matching layer of the offline recommendation module calculates the cosine similarity between them according to the latent semantic feature vectors extracted by the feature extraction layer, and obtains the matching score between the user and the movie.

The online recall and sorting module retrieves the user feature vector that the user has trained to obtain according to the user's attribute characteristics, and uses the approximate nearest neighbor search technology to recall the movie subset recommended for the user in the movie vector library, and removes the user in the sorting stage. For the movies that have been watched, the similarity between the remaining movies and the user feature vector is calculated as the sorting basis, and a list of recommended results is returned.

A movie recommendation method based on an improved deep structured semantic model, including the following steps:

S1: User behavior collection and processing: Collect user interaction behavior logs, search behavior logs and play record lists by burying points in the front end, store them in the file system, and use data warehouse tools to clean the collected logs to obtain the original data. set;

According to the raw data obtained by cleaning, the system preprocesses the user's behavior log and merges the data of the user's behavior log;

S2: Offline training: receive the combined data output by the user behavior collection and processing module through S1, re-weight the data after encoding and dimensionality reduction, and extract and mine deep hidden semantic features of the data. Matching between features pair users and movies;

S3: Online recall and sorting: According to the user's attribute features, the user feature vector that has been trained by the user is extracted, and the approximate nearest neighbor search technology is used to recall the movie subset recommended for the user in the movie vector library.

In step S2, the following steps are also included:

A1: Feature input: after receiving the combined data output from the user behavior collection and processing, encode and reduce the dimension of the user feature data;

A2: Feature learning: use the compressed excitation network SENET to re-weight the data of the input layer;

A3: Feature extraction: use three fully connected networks to form a deep neural network to extract and mine deep latent semantic features from the input user and movie feature vectors;

A4: Feature matching: According to the extracted latent semantic feature vector, the matching score between the user and the movie is obtained by calculating the cosine similarity between them.

In step A1, the user sparse feature includes a sparse feature with a definite value and a variable-length sparse feature, and after the sparse feature with a definite value is encoded, the output is a low-dimensional vector;

The variable-length sparse features include viewing history and search history, and a viewing vector is obtained after the movie embedding sequence corresponding to the viewing history is weighted and averaged by a dimensionality reduction vector, and the formula is as follows:

Among them, t represents the current time, t ₀ represents the viewing time, and m _i is the embedding vector of the ith movie;

The search history obtains the embedding vector after dimensionality reduction training, and the corresponding movie embedding sequence is weighted to obtain the search vector, and the formula is as follows:

The search history and viewing history are interleaved for training, and the input layer splices the processed sparse features and dense features, and uses the spliced user and movie vectors as initial embedding vectors.

In step A2, it also includes a compression stage and an excitation stage. The compression stage performs data compression and information aggregation on the embedding vector of each feature received from step A1 to form an initial weight vector, and the formula is as follows:

In the excitation stage, feature crossover is performed on the initial weight vector output in the compression stage and the output size dimension is maintained. In the compression stage, a two-layer MLP network with a relatively narrow middle layer is introduced, which acts on the output vector Z in the excitation stage. The formula is as follows:

S=F _ex (Z,W)=δ(W ₂ δ(W ₁ Z));

Among them, δ is the activation function, the function of the first MLP is to do feature crossover, and the function of the second MLP is to maintain the size dimension of the output.

In step A3, three fully connected networks are used to form a deep neural network, and the input user and movie feature vectors are used to extract and mine deep latent semantic features. The latent semantic feature vector y is specifically expressed as:

l _i =f(W _i l _i-1 +b _i ),i=2,...,N-1;

y=f(W _N l _N-1 +b _N );

Among them, {l _i , i=1,2,...,N-1} represents the output of each fully connected layer, W _i , b _i represent the weight matrix and bias term of the i-th layer, respectively,

f represents the activation function tanh:

Step A4 extracts the latent semantic feature vector extracted according to the feature of step A3, calculates the cosine similarity between them, and obtains the matching score between the user and the movie;

Among them, y _U and y _M represent the final latent semantic feature vectors of users and movies, respectively, and ‖·‖ represents the modulo operation;

When the model is trained, the cosine similarity of the final feature vector of the user and the movie is converted into the posterior probability by the softmax function. The formula is as follows:

Among them, γ represents the smoothing factor of the softmax function, and the loss function is minimized by maximum likelihood estimation. By increasing the time weight, the goal is to maximize the user's viewing time. The formula is as follows:

L(Λ)=-log∏ _(U,M+) T _j ·P(M ⁺ |U);

Among them, Tj represents the duration of the _jth movie, M represents the candidate movie set, Λ represents the model parameters, and M ⁺ represents the positive samples in the candidate movie.

Step S3 extracts the user feature vector that the user has trained to obtain according to the user's attribute features, and uses the approximate nearest neighbor search technology to recall the movie subset recommended for the user in the movie vector library, and removes the movies that the user has watched in the sorting stage. , calculate the similarity between the remaining movies and the user feature vector, use this as the sorting basis, and return a list of recommended results.

Compared with the prior art, the beneficial effects of the present invention are:

1. The improved deep structured semantic recommendation model proposed in this patent can effectively mine movies that meet the user's interests according to the explicit features and implicit interaction information between users and movies, and provide users with personalized recommendation services;

2. This patent uses the Faiss framework to save the feature vectors of users and movies trained by the model according to their attribute characteristics. When the user accesses the system again, the vector can be quickly retrieved and used for ANN search to get the user's recommendation. result;

3. The movie recommendation system proposed in this patent will continuously collect user behavior logs, and regularly update these historical records to the user's portrait, so as to ensure that the system can learn the user's recent interests in time.

Description of drawings

1 is a schematic diagram of a movie recommendation system based on an improved deep structured semantic model of the present invention;

Fig. 2 is a kind of offline training model structure diagram of the movie recommendation system based on the improved deep structured semantic model of the present invention;

3 is a schematic diagram of solving the time travel problem of a movie recommendation system based on an improved deep structured semantic model of the present invention;

FIG. 4 is a flowchart of an embodiment of a movie recommendation method based on an improved deep structured semantic model of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described clearly and completely below. If the specific conditions are not indicated in the examples, it is carried out according to the conventional conditions or the conditions suggested by the manufacturer. The reagents or instruments used without the manufacturer's indication are conventional products that can be purchased from the market.

Example: As shown in Figure 1, a movie recommendation system based on an improved deep structured semantic model includes a user behavior collection and processing module, an offline training module and an online recall and sorting module,

The user behavior collection and processing module collects the user's interactive behavior log, search behavior log and play record list by burying points in the front end, and stores it in the file system feature database as user feature data. Warehouse tool, which cleans the collected logs to obtain the original data set of basic samples;

According to the raw data obtained by cleaning, the system preprocesses the user's behavior log and merges the data of the user's behavior log;

The offline training module receives the combined data samples output by the user behavior acquisition and processing module, re-weights the data after encoding and dimensionality reduction, and extracts and mines deep hidden semantic features of the data to complete the model training. Data features are matched between users and movies to complete model prediction;

The online recall and sorting module retrieves the user feature vector model prediction data that the user has trained based on the user's attribute characteristics, and uses the approximate nearest neighbor search technology to perform vector retrieval in the movie vector library. In this embodiment, the vector retrieval uses the nearest neighbor Search the ANN to recall the subset of movies recommended for the user.

As shown in Figure 2, the offline training module includes an input layer, a self-attention layer, a feature extraction layer and a matching layer.

The user feature data output by the user behavior collection and processing module is combined and sent to the input layer, the input layer includes an encoding module and a dimension reduction module, and the input layer inputs the user characteristic data to the encoding module and the dimension reduction module;

The self-attention layer adopts the compression excitation network SENET to re-weight the data of the input layer;

The feature extraction layer uses three fully connected networks to form a deep neural network, which is used to extract and mine deep latent semantic features from the input user and movie feature vectors;

The matching layer obtains the matching score between the user and the movie by calculating the cosine similarity between them according to the extracted latent semantic feature vectors.

The user feature data includes user dense features and user sparse features, the user dense features are input to the encoding module, and the user sparse features are input to the dimension reduction module;

User sparse features include sparse features with definite values and variable-length sparse features. After the sparse features with definite values are input into the encoding module, the output is a low-dimensional vector;

The variable-length sparse features include viewing history and search history. The viewing vector is obtained after the movie Embedding embedding sequence corresponding to the viewing history is weighted and averaged by the dimensionality reduction module vector;

The search history is trained by the dimensionality reduction module to obtain the Embedding embedding vector, and the corresponding movie Embedding embedding sequence is weighted and averaged to obtain the search vector;

The search history and viewing history are interleaved for training, and the input layer splices the processed sparse features and dense features, and uses the spliced user and movie vectors as the initial Embedding embedding vectors.

The self-attention layer includes a compression module and an excitation module. The compression module performs data compression and information aggregation on the Embedding embedding vector of each feature received from the input layer to form an initial weight vector;

The excitation module is used to perform feature crossover on the initial weight vector output by the compression module and maintain the output size dimension;

The matching layer of the offline recommendation module calculates the cosine similarity between the latent semantic feature vectors extracted by the feature extraction layer and obtains the matching score between the user and the movie.

The online recall and sorting module extracts the user feature vector that the user has trained according to the user's attribute characteristics, and uses the approximate nearest neighbor search technology ANN to recall the movie subset recommended for the user in the movie vector library, and removes the user's already in the sorting stage. Watched movies, calculate the similarity between the remaining movies and the user feature vector, use this as the sorting basis, and return a list of recommended results.

A movie recommendation method based on an improved deep structured semantic model, including the following steps:

S1: User behavior collection and processing: Collect user interaction behavior logs, search behavior logs and play record lists by burying points in the front end, store them in the file system, and use data warehouse tools to clean the collected logs to obtain the original data. set;

According to the raw data obtained by cleaning, the system preprocesses the user's behavior log and merges the data of the user's behavior log;

S2: Offline training: receive the combined data output by the user behavior collection and processing module through S1, re-weight the data after encoding and dimensionality reduction, and extract and mine deep hidden semantic features of the data. Matching between features pair users and movies;

S3: Online recall and sorting: According to the user's attribute features, the user feature vector that has been trained by the user is extracted, and the approximate nearest neighbor search technology is used to recall the movie subset recommended for the user in the movie vector library.

In step S2, the following steps are also included:

A1: Feature input: After receiving the combined data of user behavior collection and processing output, encode and reduce the dimension of user feature data;

A2: Feature learning: use the compressed excitation network SENET to re-weight the data of the input layer;

A3: Feature extraction: use three fully connected networks to form a deep neural network to extract and mine deep latent semantic features from the input user and movie feature vectors;

A4: Feature matching: According to the extracted latent semantic feature vector, the matching score between the user and the movie is obtained by calculating the cosine similarity between them.

In step A1, the user sparse features include sparse features with definite values and variable-length sparse features, and after the sparse features with definite values are encoded, the output is a low-dimensional vector;

The one-hot encoding operation is performed on the user dense features; the embedding dimension is reduced to a low-dimensional space operation on the user sparse features.

The processing of user sparse features can be divided into two categories:

One is to deal with sparse features with definite values: such features are mainly encoded features such as user ID, user age, and user occupation, and each user has only a unique definite value for such features. You can use the torch.nn.Embedding tool to create a word embedding model for coding with the help of machine learning libraries such as pytorch, and output sparse features with certain values into low-dimensional vectors.

The variable-length sparse features include viewing history and search history. The viewing vector is obtained after the movie Embedding embedding sequence corresponding to the viewing history is weighted and averaged by the dimension reduction vector. The formula is as follows:

Among them, t represents the current time, t ₀ represents the viewing time, and m _i is the Embedding embedding vector of the ith movie;

The processing of user search history is similar to the processing of viewing history. First, the keywords of historical search are segmented to obtain entry tokens, and the embedding vector of the token is obtained by training, and then the embedding vector corresponding to the token of the user's historical search is weighted. The search vector search vector is obtained on average, so that the overall state of the user's search history can be learned, and the shorter the time interval from the current time, the higher the weight of the Embedding embedding vector, the formula is as follows:

The search history is trained by dimensionality reduction to obtain the embedding vector, and the corresponding movie embedding sequence is weighted and averaged to obtain the search vector. The formula is as follows:

For the training of these two variable-length sequence features, the system has a mechanism to punish overactive users, and sets an upper limit of sequence features for each user to prevent the model from being represented by a small number of overactive users, and recommends each user equally. .

As shown in Figure 3, considering that users watching videos and searching are often sequential, and some viewing behaviors even have some causal relationships, the traditional recommendation model uses the context as input information, which is equivalent to revealing some future information for training , In order to solve such a time travel problem, the present invention completely separates future information from training samples, and uses T-2 search records and T-1 viewing history records for training to solve the time travel problem.

The search history and viewing history are interleaved for training, and the input layer splices the processed sparse features and dense features, and uses the spliced user and movie vectors as the initial embedding vectors.

In step A2, the importance of these features is dynamically learned by the compression excitation network SENET, the invalid low-frequency features are suppressed by reducing the weight, the important features are amplified by increasing the weight, and the initial embedding vectors of users and movies are re-weighted. Specifically, it includes the Squeeze compression stage and the Excitation excitation stage. The Squeeze compression stage performs data compression and information aggregation on the embedding vector of each feature received from step A1 to form an initial weight vector. The formula is as follows:

Assuming that the dimension of the embedding vector of a feature _ui is k, then we average the k numbers contained in the embedding vector to obtain a value _zi that can represent the summary information of this feature;

Through the Squeeze compression stage, each feature _ui is compressed into a single value _zi . Assuming that the spliced Embedding embedding vector has f features, the initial weight vector Z={z ₁ ,z ₂ ,…, z _f };

In the excitation stage, the initial weight vector output in the Squeeze compression stage is cross-featured and the output size is maintained. In the excitation stage, a two-layer MLP network with a relatively narrow middle layer is introduced, which acts on the output vector Z in the excitation stage. The formula is as follows :

S=F _ex (Z,W)=δ(W ₂ δ(W ₁ Z));

Among them, δ is the activation function, the function of the first MLP is to do feature crossover, and the function of the second MLP is to maintain the size dimension of the output.

In step A3, three fully connected network DNNs are used to form a deep neural network to extract and mine deep latent semantic features for the input user and movie feature vectors. The latent semantic feature vector y is specifically expressed as:

l _i =f(W _i l _i-1 +b _i ),i=2,...,N-1;

y=f(W _N l _N-1 +b _N );

Among them, {l _i , i=1,2,...,N-1} represents the output of each fully connected layer, W _i , b _i represent the weight matrix and bias term of the i-th layer, respectively,

f represents the activation function tanh:

Step A4 extracts the latent semantic feature vector extracted according to the feature of step A3, calculates the cosine similarity between them, and obtains the matching score between the user and the movie;

Among them, y _U and y _M represent the final latent semantic feature vectors of users and movies, respectively, and ‖·‖ represents the modulo operation;

The input of the model consists of a user and a set of multiple candidate movies. The movie set contains a certain proportion of positive and negative samples. The standard for the division of positive and negative proportions is based on the user's preference for movies. The preference is greater than 0.5. Positive samples, Less than 0.3 is a negative sample. The formula for calculating preference is as follows:

表示用户u_i对电影m_j的观看时长，如果观看时长小于电影时长的百分之三十则得1分，如果大于百分之七十则得2分，

和

分别表示用户u_i对电影m_j点赞和转发的得分，本专利设置点赞和转发的得分为3分和4分。in,

Represents the viewing time of the movie m _j by the user u _i . If the viewing time is less than 30% of the movie time, it will get 1 point, and if it is more than 70%, it will get 2 points.

and

respectively represent the scores of user _ui 's likes and reposts of the movie _mj , and the patent sets the likes and reposts as 3 points and 4 points.

When the model is trained, the cosine similarity of the final feature vector of the user and the movie is converted into the posterior probability by the softmax function. The formula is as follows:

Among them, γ represents the smoothing factor of the softmax function, and the loss function is minimized by maximum likelihood estimation. By increasing the time weight, the goal is to maximize the user's viewing time. The formula is as follows:

Among them, Tj represents the duration of the _jth movie, M represents the candidate movie set, Λ represents the model parameters, and M ⁺ represents the positive samples in the candidate movie.

Step S3 extracts the user feature vector that the user has trained to obtain according to the user's attribute features, and uses the approximate nearest neighbor search technology to recall the movie subset recommended for the user in the movie vector library, and removes the movies that the user has watched in the sorting stage. , calculate the similarity between the remaining movies and the user feature vector, use this as the sorting basis, and return a list of recommended results.

The specific implementation flowchart is shown in FIG. 4 . First, the user enters the system and logs in. The system judges whether the user is a new user. If the user is not a new user, it means that the system already has its historical data, the system has obtained the feature vector of the user in the latent semantic space, and establishes an index according to the user's ID and stores it in Faiss. in the frame. The system retrieves the feature vector according to the user ID, and uses this feature vector to perform an approximate nearest neighbor search (ANN) in the movie vector library to recall the movie candidate set that the user may be interested in. Since the movie candidate set may contain movies that the user has already watched, the system will filter the movie candidate set again in the sorting stage, remove the movies that have been watched, and calculate the similarity between the remaining movies and the user's feature vector, which is used as a ranking According to, and return the list of recommended results of TOP-N.

It will be apparent to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, but that the present invention may be embodied in other specific forms without departing from the spirit and essential characteristics of the present invention. Therefore, the embodiments are to be regarded in all respects as illustrative and not restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description, which are therefore intended to fall within the scope of the appended claims. All changes that come within the meaning and range of equivalents of , are intended to be embraced within the invention, and any reference signs in the claims shall not be construed as limiting the involved claim.

In addition, it should be understood that although this specification is described in terms of embodiments, not each embodiment only includes an independent technical solution, and this description in the specification is only for the sake of clarity, and those skilled in the art should take the specification as a whole , the technical solutions in each embodiment can also be appropriately combined to form other implementations that can be understood by those skilled in the art.

Claims (10)

1. A movie recommendation system based on an improved deep structured semantic model is characterized by comprising a user behavior acquisition and processing module, an offline training module and an online recall and sequencing module,

the user behavior acquisition and processing module collects an interactive behavior log, a search behavior log and a play record list of a user by embedding points at the front end, stores the interactive behavior log, the search behavior log and the play record list as user characteristic data into a file system characteristic library, and performs data cleaning on the collected log by means of a data warehouse tool to obtain a basic sample original data set;

according to the original data obtained by cleaning, the system preprocesses the behavior logs of the users and performs data combination on the behavior logs of the users;

the offline training module receives the combined data sample output by the user behavior acquisition and processing module, reweighs the data after coding and dimensionality reduction of the data, extracts and mines deep latent semantic features of the data, and matches the user with the movie according to the data features;

and the online recall and sorting module is used for taking out the user characteristic vector obtained by the training of the user according to the attribute characteristics of the user, performing vector retrieval in a movie vector library by adopting an approximate nearest neighbor search technology, and recalling the movie subset recommended for the user.

2. The movie recommendation system based on the improved deep-structured semantic model according to claim 1,

the off-line training module comprises an input layer, a self-attention layer, a feature extraction layer and a matching layer,

the user characteristic data output by the user behavior acquisition and processing module is merged and then sent to an input layer, the input layer comprises a coding module and a dimension reduction module, and the input layer inputs the user characteristic data to the coding module and the dimension reduction module;

the self-attention layer re-weights the data of the input layer by adopting a compressed excitation network SENET;

the feature extraction layer uses three fully-connected networks to form a deep neural network for extracting and mining deep latent semantic features of input user and movie feature vectors;

and the matching layer calculates the cosine similarity between the extracted latent semantic feature vectors according to the extracted latent semantic feature vectors to obtain the matching score between the user and the film.

3. The movie recommendation system based on the improved deep-structured semantic model according to claim 2,

the user feature data comprise user dense features and user sparse features, the user dense features are input to the coding module, and the user sparse features are input to the dimension reduction module;

the user sparse features comprise sparse features with determined values and variable-length sparse features, and the sparse features with the determined values are input into the coding module and then output as low-dimensional vectors;

the variable-length sparse features comprise viewing history and search history, and the movie embedding sequences corresponding to the viewing history are subjected to vector weighted average by a dimensionality reduction module to obtain viewing vectors;

the search history is trained by a dimensionality reduction module to obtain an embedded vector, and the corresponding film embedded sequence is weighted and averaged to obtain a search vector;

and the search history and the viewing history are trained in a staggered and corresponding mode, the input layer splices the processed sparse features and the processed dense features, and the spliced user and movie vectors are used as initial embedded vectors.

4. The movie recommendation system based on the improved deep-structured semantic model according to claim 2,

the self-attention layer comprises a compression module and an excitation module, and the compression module performs data compression and information summarization on the embedded vector of each feature received from the input layer to form an initial weight vector;

the excitation module is used for performing feature crossing on the initial weight vector output by the compression module and keeping the dimension of output size;

and the offline recommendation module matching layer calculates cosine similarity between the implicit semantic feature vectors extracted by the feature extraction layer according to the implicit semantic feature vectors extracted by the feature extraction layer to obtain a matching score between the user and the film.

5. The movie recommendation system based on the improved deep-structured semantic model according to claim 1, wherein the online recall and sorting module extracts a user feature vector obtained by training a user according to attribute features of the user, recalls a subset of movies recommended for the user in a movie vector library by using an approximate nearest neighbor search technique, removes movies already watched by the user in a sorting stage, calculates similarity between the remaining movies and the user feature vector, and returns a recommendation result list by using the similarity as a sorting basis.

6. A movie recommendation method based on an improved deep structured semantic model is characterized by comprising the following steps:

s1: user behavior acquisition and processing: the method comprises the steps that interactive behavior logs, search behavior logs and play record lists of users are collected through point burying at the front end and stored in a file system, and data cleaning is conducted on the collected logs by means of a data warehouse tool to obtain an original data set;

according to the original data obtained by cleaning, the system preprocesses the behavior logs of the users and performs data combination on the behavior logs of the users;

s2: off-line training: the combined data output by the user behavior acquisition and processing module is received by S1, the data is re-weighted after being encoded and dimensionality reduced, deep latent semantic features of the data are extracted and mined, and the user and the movie are matched according to the data features;

s3: online recall and ranking: and extracting the user feature vector obtained by the training of the user according to the attribute features of the user, and recalling the movie subset recommended for the user in the movie vector library by adopting an approximate nearest neighbor search technology.

7. The method for recommending movies based on the improved deep-structured semantic model as claimed in claim 6, wherein in step S2, the method further comprises the following steps:

a1: characteristic input: after the combined data output by the user behavior acquisition and processing is received, encoding and dimension reduction are carried out on the user characteristic data;

a2: and (3) feature learning: re-weighting the data of the input layer by adopting a compressed excitation network SENET;

a3: characteristic extraction: three full-connection networks are used for forming a deep neural network, and deep latent semantic features of input users and movie feature vectors are extracted and mined;

a4: characteristic matching: and according to the extracted latent semantic feature vectors, calculating the cosine similarity between the latent semantic feature vectors to obtain the matching score between the user and the movie.

8. The movie recommendation method based on the improved deep-structured semantic model according to claim 7, wherein in step a1, the sparse features of the user include sparse features with definite values and variable-length sparse features, and the sparse features with definite values are encoded and then output as low-dimensional vectors;

the variable-length sparse features comprise viewing history and search history, the movie embedding sequence corresponding to the viewing history is subjected to dimension reduction vector weighted average to obtain a viewing vector, and the formula is as follows:

where t denotes the current time, t₀Represents a viewing time, m_iIs the embedded vector for the ith movie;

the search history is subjected to dimensionality reduction training to obtain an embedded vector, the corresponding film embedded sequence is subjected to weighted average to obtain a search vector, and the formula is as follows:

and the search history and the viewing history are trained in a staggered and corresponding mode, the input layer splices the processed sparse features and the processed dense features, and the spliced user and movie vectors are used as initial embedded vectors.

9. The movie recommendation method based on the improved deep-structured semantic model according to claim 7, further comprising a compression stage and an excitation stage in step a2, wherein the compression stage performs data compression and information summarization on the embedded vector of each feature received in step a1 to form an initial weight vector, and the formula is as follows:

the excitation stage is used for performing feature crossing on the initial weight vector output in the compression stage and keeping the dimension of the output size, two layers of MLP networks with narrower middle layers are introduced in the compression stage and act on the output vector Z in the excitation stage, and the formula is as follows:

S＝F_ex(Z,W)＝δ(W₂δ(W₁Z))；

where δ is the activation function, the first MLP acts to do the feature crossing, and the second MLP acts to maintain the size dimension of the output.

In step a3, a deep neural network is formed by using three fully-connected networks, deep latent semantic features are extracted and mined from input user and movie feature vectors, and a latent semantic feature vector y is specifically represented as:

l_i＝f(W_il_i-1+b_i),i＝2,…,N-1；

y＝f(W_Nl_N-1+b_N)；

wherein, { l_iWhere i is 1,2, …, N-1, and W represents the output of each fully-connected layer_i，b_iRespectively representing the weight matrix and the bias term of the ith layer,

f represents the activation function tanh:

step A4, extracting the extracted implicit characteristic vectors according to the characteristics of the step A3, and calculating the cosine similarity between the implicit characteristic vectors and the implicit characteristic vectors to obtain the matching score between the user and the movie;

wherein, y_U、y_MLatent semantic feature vectors representing the resulting user and movie, respectively, | | represents a modulo operation;

during model training, the cosine similarity of the final feature vectors of the user and the film is converted into posterior probability through a softmax function, and the formula is as follows:

wherein γ represents a smoothing factor of the softmax function, and minimizes a loss function through maximum likelihood estimation, and by adding a time weight, the objective is to maximize the user viewing duration, and the formula is as follows:

wherein, T_jRepresenting the duration of the j-th movie, M representing the set of candidate movies, Lambda representing the model parameters, M⁺Representing positive samples in the candidate movie.

10. The method according to claim 6, wherein in step S3, the user feature vector obtained by training the user is extracted according to the attribute features of the user, the subset of movies recommended for the user is recalled in the movie vector library by using the approximate nearest neighbor search technique, the movies already watched by the user are removed in the sorting stage, the similarity between the remaining movies and the user feature vector is calculated, and the similarity is used as a sorting criterion, and the recommendation result list is returned.

Images