A personalized movie sequence recommendation method based on deep clustering algorithm

Abstract

The present invention discloses a personalized movie sequence recommendation method based on a deep clustering algorithm to achieve personalized movie recommendation. The process is as follows: obtain a movie data set and separate the required data; mine the user's high-level intention from the obtained data through deep clustering to form an interactive sequence that integrates the user's high-level intention; mine the user's static features from the obtained data and combine it with the user's movie viewing sequence to form a user feature dynamic sequence; combine the interactive sequence that integrates the user's high-level intention with the user's feature dynamic sequence and input it into a gated recurrent unit to obtain fused sequence information; use the fused sequence information to calculate the matching scores of all candidate items with this sequence, and select the top K items as the recommendation results. The present invention can effectively capture the higher-level intention features of users when watching movies, and alleviate the long-tail problem through clustered distribution, and finally generate personalized movie recommendations.

Description

Technical Field

[0001] The present invention relates to the technical field of movie sequence recommendation, and in particular to a method for recommending personalized movie sequences based on a deep clustering algorithm. Background Art

[0002] At present, the research on capturing user dynamic features in sequential recommendation mainly focuses on the discovery of user shallow features, lacking the mining and analysis of user high-level features. However, the user's behavior sequence often contains the user's higher-level intention features. For example, if a user watches and gives a good review of a certain movie, then the user may not only like the movie, but more likely like the subject matter, or even like this large category of movies. Therefore, the existing technical methods ignore the user's high-level intentions, which often leads to the recommendation results being less personalized and accurate.

[0003] In the field of sequential recommendation, there are often many items with little interaction with users in the candidate items for recommendation. These items are called long-tail items. Because there are too few interactions with them, it is very difficult to learn their distribution characteristics and recommend them in the recommendation system. This is the so-called long-tail problem in the field of sequential recommendation. The existence of the long-tail problem seriously affects the accuracy of recommendations. Summary of the invention

[0004] The purpose of the present invention is to solve the above defects in the prior art and provide a personalized movie sequence recommendation method based on a deep clustering algorithm. The sequence recommendation method extracts the user's high-level intention and alleviates the long-tail problem through deep clustering operations and user static feature mining, thereby achieving the goal of personalized sequence recommendation.

[0005] The purpose of the present invention can be achieved by adopting the following technical solutions:

[0006] A personalized movie sequence recommendation method based on a deep clustering algorithm, the personalized movie sequence recommendation method comprising the following steps:

[0007] S1. Obtain the sequence data of the user watching movies, the user's personal attribute data and the attribute data of the movie itself as a movie sequence data set, and extract the user static information, interaction sequence information and movie project attribute information from it, wherein the user static information includes the user's gender and age, the interaction sequence information includes the user's movie viewing sequence, and the movie project attribute information includes the movie category, director and the year of the movie release;

[0008] S2. All the movie projects in the sampled data are taken as movie project groups. The cluster center corresponding to each movie project is obtained from the movie project group through deep clustering. Each movie project in the interaction sequence information obtained in step S1 is concatenated with the cluster center corresponding to each movie project to form an interaction sequence embedding that integrates high-level intent.

[0009] S3, mining user static information, splicing user static information with interaction sequence information to form user feature dynamic sequence embedding;

[0010] S4, combining the interaction sequence embedding fused with the high-level intent and the dynamic sequence embedding of the user features into a gated recurrent unit to capture the fused sequence information;

[0011] S5. Use the fused sequence information to calculate the matching scores of all candidate movie projects and user interaction sequences, and select the top K projects as recommendation results to make personalized movie recommendations for users.

[0012] Furthermore, the process of step S1 is as follows:

[0013] S1.1. Extract user static information, interaction sequence information and movie project attribute information from the movie sequence dataset. Extracting user static information is helpful for personalized recommendation, extracting interaction sequence information is helpful for capturing user fusion sequence information, and extracting movie project attribute information is helpful for deep clustering.

[0014] S1.2. The interaction sequence information is divided. In the present invention, in order to ensure the uniformity of the interaction sequence information, the interaction sequence with a length of not less than 20 is intercepted with the last 20 interactions, and the interaction sequence with a length of less than 20 is padded with -1 at the head end to make the number of interactions reach 20, and the last interaction behavior in the interaction sequence is used as a test, the second last interaction behavior is used as a verification, and the other interaction behaviors are used as training;

[0015] S1.3, let S = {s 1 ,s 2 ,…,s |V|} represents the interaction items in all interaction sequences, where |V| is the sum of all the interaction items; let A = {s 1 ,s 2 ,…,s c} represents the interaction sequence of a user in a session, where s b represents the bth interaction behavior in this interaction sequence, s b ∈S, b=1,2,…,c, c represents the number of user interaction actions in a session;

[0016] S1.4. Create an embedding matrix E for all interaction sequences to facilitate the use of the feature information contained in the interaction sequence in subsequent steps. Set the constant zero vector 0 as the filler item for the blank part of the embedding matrix E. The creation process of the embedding matrix E is as follows: retrieve the first p items of each interaction sequence and stack them together to generate the embedding matrix E∈R t×d , where d is the potential dimension, here assumed to be 1, the embedding matrix E is defined as follows, where A pr Represents the first p items of the rth interaction sequence:

[0017] Furthermore, the process of step S2 is as follows:

[0018] S2.1. Create an embedding matrix W for the movie project group to facilitate the use of the feature information contained in the movie project group in the subsequent steps. Set the constant zero vector 0 as the filler item for the blank part of the embedding matrix W. The creation process of the embedding matrix W is as follows: retrieve the attribute information of each movie project and stack them together to generate the embedding matrix W∈R n×d , where d is the potential dimension, here assumed to be 1, and the embedding matrix W is defined as follows, where m n Represents the attribute information of the nth movie project in the movie project group:

[0019] S2.2, Assume that the entire movie project group is divided into k classes, each class has a cluster center, first initialize the k cluster centers, and then assume that the distance between each sample point and the cluster center to which the sample point belongs satisfies the following distribution:

[0020]

[0021] In the above formula, i represents the i-th sample, j represents the j-th cluster center in this round of clustering, and z i represents the spatial distribution of the original movie project attribute information after the deep clustering preliminary encoding process, μ j represents the center point of the jth cluster, α is the degree of freedom, q ij represents the probability that sample i belongs to cluster j. The above distribution is the initial distribution of the deep clustering process, which makes similar movie projects in the entire movie project group close to each other;

[0022] S2.3, determine the initial cluster center according to step S2.2, then use the distribution formula to measure whether the sample belongs to a certain cluster, and calculate a compact auxiliary distribution that conforms to the cluster center:

[0023]

[0024] In the above formula, p ijrepresents the probability that sample i belongs to cluster j, f j represents the soft allocation function, f j =∑ i q ij ,Compared with the initial distribution, each cluster is more compact and the clusters are farther apart;

[0025] S2.4, calculating the distance loss function according to the two different distributions obtained in step S2.2 and step S2.3, and continuously iterating to make the difference between the two distributions smaller and smaller, and the cluster center also changes with each round of iterative optimization, wherein the distance loss function is as follows:

[0026]

[0027] The gradient update principle formula for each iteration is as follows:

[0028]

[0029]

[0030] Concatenate each movie item in the interaction sequence embedding with the cluster center corresponding to this movie item to form an interaction sequence embedding that incorporates high-level intent;

[0031] At the same time, through deep clustering, the long-tail projects in the film project group also move closer to the cluster center to which they belong, which can solve the long-tail problem to a certain extent.

[0032] Furthermore, the process of step S3 is as follows:

[0033] S3.1. Create an embedding matrix U for the user static information to facilitate the use of the feature information contained in the user static information in the subsequent steps, and set the constant zero vector 0 as the filler item for the blank part of the embedding matrix U. The creation process of the embedding matrix is ​​as follows: retrieve the user static information of each user and stack them together to generate the embedding matrix U∈R v×d , where d is the potential dimension, which is assumed to be 1 here, and the matrix U is defined as follows, where u v Represents the user static information of the vth user:

[0034] S3.2. Concatenate the user static information embedding and the interaction sequence embedding to generate a user feature dynamic sequence embedding. This user feature dynamic sequence embedding contains both the user's static information and the user's dynamic information, which can effectively capture the user's viewing interest and generate accurate personalized recommendations for the user.

[0035] Furthermore, in step S4, the gated recurrent unit is referred to as a GRU unit, and the σ() used in the GRU unit is a sigmoid activation function, and the formula of the sigmoid activation function is as follows: Where x is the input;

[0036] The GRU unit can effectively memorize the long-term information in the interaction sequence and combine it with the short-term information through the reset gate and the update gate. In this step, the GRU unit is used to capture the fused sequence information. The working process of the GRU unit is as follows:

[0037] First, the reset gate R is calculated in each GRU unit t , controls whether the candidate state depends on the implicit state of the GRU unit at the previous moment, X t represents the sequence information of the GRU unit at time t, h t-1 represents the hidden state of the GRU unit at time t-1, W rx With W rh represents the first and second weight matrices to be learned, b r Indicates the first bias that needs to be learned:

[0038] R t =σ(W rx X t +W rh h t-1 +b r );

[0039] Calculate the update gate Z in each GRU unit t , controls how much information the state of the GRU unit at this moment needs to retain from the state of the previous moment, where W zx With W zh represents the third and fourth weight matrices that need to be learned, b z Represents the second bias that needs to be learned:

[0040] Z t =σ(W zx X t +W zh h t-1 +b z );

[0041] Next, the candidate state H of the input information is calculated in each GRU unit t ', where W hx With W hh represents the fifth and sixth weight matrices that need to be learned, b h Indicates the third bias that needs to be learned:

[0042] H t '=tanh(Whx X t +W hh (R t ·h t-1 +b h ));

[0043] Finally, the memory cell state H at time t is calculated in each GRU unit t , controls the information that is ultimately memorized in the GRU unit:

[0044] H t =Z t ·h t-1 +(1-Z t )·H t ';

[0045] The working process of the above GRU unit is abbreviated into the function GRU(). The interaction sequence embedding that integrates high-level intent and the dynamic sequence embedding of user features are combined and input into the GRU unit. The formula is as follows:

[0046] h t =GRU(h t-1 ,[a t ,g t ,r t ]);

[0047] Among them, a t , g t 、r t They represent the interaction sequence embedding at time t, the user feature dynamic sequence embedding, and the interaction sequence embedding fused with high-level intent. The three together form the input X at time t in the above GRU unit formula t , so that the GRU unit can jointly capture the interaction sequence information, user static information and the interaction information generated by deep clustering that integrates high-level intentions, h t This is the final fusion sequence information.

[0048] Further, the fused sequence information is used as a sequence embedding representation, and then the sequence embedding representation is projected into the embedding matrix of the movie project group in step S2.1. This step can combine the candidate movie projects in the movie project group with the fused sequence information, so as to facilitate the calculation of the matching score;

[0049] For the projected sequence embedding, y ul represents the matching score between the fused sequence information of user u and the movie project l, where W represents the movie project matrix and σ() is the sigmoid activation function: ul =σ(Wh t ), select the top K movie projects with the highest scores and recommend them to users.

[0050] Furthermore, the overall objective function in step S5 is defined as: L = L rec +L kl ,in,

[0051] In the formula, W represents the embedding matrix of the movie project group, z represents the positive sample, f represents the negative sample, and y uz represents the positive sample matching score of user u, y uf represents the negative sample matching score of user u, σ() is the sigmoid activation function, and the above total objective function is used for training to finally generate a more accurate personalized movie sequence recommendation result.

[0052] Compared with the prior art, the present invention has the following advantages and effects:

[0053] 1) The present invention uses deep clustering algorithms and user static feature extraction to fully mine the user's high-level intentions implicit in the interaction sequence and combines the user's high-level intention information with the user's static information and sequence information to jointly guide the recommendation process, making the recommendation results more accurate and personalized.

[0054] 2) The present invention uses a deep clustering algorithm to make the relatively long-tail items in the movie project group move closer to their corresponding cluster centers, so that the distribution characteristics of this part of the long-tail content can be effectively learned and thus incorporated into the recommendation process, alleviating the long-tail problem in the recommendation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a specific flow chart of the personalized movie sequence recommendation method based on the deep clustering algorithm disclosed in the present invention;

[0056] Figure 2 It is a structural diagram of the personalized movie sequence recommendation method based on the deep clustering algorithm disclosed in the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] Example 1

[0059] This embodiment discloses a personalized movie sequence recommendation method based on a deep clustering algorithm. Figure 1As shown, the personalized sequence recommendation method includes the following steps:

[0060] Step S1, obtaining the sequence data of the user watching movies, the user's personal attribute data and the attribute data of the movie itself as a movie sequence data set, and extracting the user static information, interaction sequence information and movie project attribute information from it, wherein the user static information includes the user's gender and age, the interaction sequence information includes the user's movie viewing sequence, and the movie project attribute information includes the movie category, director and the year of the movie release;

[0061] In specific applications, movie datasets may have different sequence lengths. Figure 2 As shown, any length-classified movie dataset can be used as input to extract user static information, interaction sequence information, and movie project attribute information from the movie sequence dataset;

[0062] Then the interaction sequence information is divided, the penultimate interaction behavior in the interaction sequence is used as a test, the penultimate interaction behavior is used as a verification, and the other interaction behaviors are used as training;

[0063] Let S = {s 1 ,s 2 ,…,s |V|} represents the interaction items in all interaction sequences, where |V| is the sum of all the interaction items; let A = {s 1 ,s 2 ,…,s c} represents the interaction sequence of a user in a session, where s b represents the bth interaction behavior in this interaction sequence, s b ∈S, b=1,2,…,c, c represents the number of user interaction actions in a session;

[0064] like Figure 2 As shown in the left side, the interaction sequence information is embedded. An embedding matrix E is created for all interaction sequences, and a constant zero vector 0 is set as the filler item for the blank part of the embedding matrix E. The creation process of the embedding matrix E is as follows: the first p items of each interaction sequence are retrieved and stacked together to generate the embedding matrix E∈R t×d , where d is the potential dimension, here assumed to be 1, the embedding matrix E is defined as follows, where A pr Represents the first p items of the rth interaction sequence:

[0065] Step S2: All movie projects in the sampled data are taken as a movie project group, such as Figure 2As shown in the deep clustering, the cluster center corresponding to each movie project is obtained from the movie project group through deep clustering, and each movie project in the interaction sequence information obtained in step S1 is concatenated with the cluster center corresponding to each movie project to form an interaction sequence embedding that integrates high-level intentions.

[0066] First, we create an embedding matrix W for the movie project group, and set the constant zero vector 0 as the filler for the blank part of the embedding matrix W. The creation process of the embedding matrix W is as follows: retrieve the attribute information of each movie project and stack them together to generate the embedding matrix W∈R n×d , where d is the potential dimension, here assumed to be 1, and the embedding matrix W is defined as follows, where m n Represents the attribute information of the nth movie project in the movie project group:

[0067] Next, assume that the entire movie project group is divided into k classes, and each class has a cluster center, such as Figure 2 As shown in the deep clustering, first initialize k cluster centers, and then assume that the distance between each sample point and the cluster center to which the sample point belongs satisfies the following distribution:

[0068]

[0069] In the above formula, i represents the i-th sample, j represents the j-th cluster center in this round of clustering, and z i represents the spatial distribution of the original movie project attribute information after the deep clustering preliminary encoding process, μ j represents the center point of the jth cluster, α is the degree of freedom, q ij represents the probability that sample i belongs to cluster j;

[0070] According to the above steps, the initial cluster center is determined, and then the distribution formula is used to measure whether the sample belongs to a certain cluster, and a compact auxiliary distribution that conforms to the cluster center is calculated:

[0071]

[0072] In the above formula, p ij represents the probability that sample i belongs to cluster j, f j represents the soft allocation function, f j =∑ i q ij ;

[0073] The distance loss function is calculated based on the two different distributions obtained in the above steps, and it is continuously iterated until convergence. The cluster center also changes with each round of iterative optimization. The distance loss function is as follows:

[0074]

[0075] The gradient update principle formula for each iteration is as follows:

[0076]

[0077]

[0078] Each movie item in the interaction sequence embedding is concatenated with the cluster center corresponding to this movie item to form an interaction sequence embedding that integrates high-level intent.

[0079] Step S3: mining user static information, and splicing user static information with interaction sequence information to form user feature dynamic sequence embedding;

[0080] like Figure 2 As shown in the left side, the user static information embedding creates an embedding matrix U for the user static information, and sets the constant zero vector 0 as the filler item for the blank part of the embedding matrix U. The creation process of the embedding matrix is ​​as follows: retrieve the user static information of each user and stack them together to generate the embedding matrix U∈R v×d , where d is the potential dimension, which is assumed to be 1 here, and the matrix U is defined as follows, where u v Represents the user static information of the vth user:

[0081] Then, the user static information embedding is concatenated with the interaction sequence embedding to generate the user feature dynamic sequence embedding.

[0082] Step S4, combining the interaction sequence embedding fused with the high-level intent and the user feature dynamic sequence embedding into a gated recurrent unit to capture the fused sequence information;

[0083] The gated recurrent unit is referred to as the GRU unit. The σ() used in the GRU unit is the sigmoid activation function. The formula of the sigmoid activation function is as follows: Where x is the input;

[0084] The working process of the GRU unit is as follows:

[0085] First, the reset gate R is calculated in each GRU unit t , X t represents the sequence information of the GRU unit at time t, h t-1 represents the hidden state of the GRU unit at time t-1, W rx With W rh represents the first and second weight matrices to be learned, b r Indicates the first bias that needs to be learned:

[0086] R t =σ(W rx X t +W rh h t-1 +b r );

[0087] Calculate the update gate Z in each GRU unit t , where W zx With W zh represents the third and fourth weight matrices that need to be learned, b z Represents the second bias that needs to be learned:

[0088] Z t =σ(W zx X t +W zh h t-1 +b z );

[0089] Next, the candidate state H of the input information is calculated in each GRU unit t ', where W hx With W hh represents the fifth and sixth weight matrices that need to be learned, b h Indicates the third bias that needs to be learned:

[0090] H t '=tanh(W hx X t +W hh (R t ·h t-1 +b h ));

[0091] Finally, the memory cell state H at time t is calculated in each GRU unit t :

[0092] H t =Z t ·h t-1 +(1-Z t )·H t ';

[0093] The working process of the above GRU unit is abbreviated into the function GRU(). The interaction sequence embedding that integrates high-level intent and the dynamic sequence embedding of user features are combined and input into the GRU unit. The formula is as follows:

[0094] h t =GRU(h t-1 ,[a t ,g t ,r t ]);

[0095] Among them, a t , g t 、r t They represent the interaction sequence embedding at time t, the user feature dynamic sequence embedding, and the interaction sequence embedding fused with high-level intents, respectively. Figure 2 As shown, the three are combined to form the input X at time t in the above GRU unit formula t ,h t This is the final fusion sequence information.

[0096] Step S5: using the fused sequence information to calculate the matching scores of all candidate movie projects and the user interaction sequence, and selecting the top K projects as recommendation results to make personalized movie recommendations for the user;

[0097] First, the fused sequence information is used as the sequence embedding representation, and then the sequence embedding representation is projected into the embedding matrix of the movie project group. For the projected sequence embedding, y ul represents the matching score between the fused sequence information of user u and the movie project l, where W represents the movie project matrix and σ() is the sigmoid activation function: ul =σ(Wh t ), select the top K movie projects with the highest scores and recommend them to users.

[0098] The overall objective function is defined as: L = L rec +L kl , using the above total objective function for training, and finally generating accurate personalized movie sequence recommendation results, where L rec =∑ (z,f)∈W -lnσ(y uz -y uf ),

[0099] In the formula, W represents the embedding matrix of the movie project group, z represents the positive sample, f represents the negative sample, and y uz represents the positive sample matching score of user u, y uf represents the negative sample matching score of user u, and σ() is the sigmoid activation function.

[0100] Example 2

[0101] The experimental results of the personalized movie sequence recommendation method based on the deep clustering algorithm disclosed in Example 1 on the commonly used movie sequence dataset MovieLens are shown in Tables 1, 2, and 3, where the MovieLens dataset is divided into ML-100k, ML-1m, and ML-25m according to the amount of data contained. In order to illustrate the superiority of this method in personalized movie sequence recommendation, this embodiment implemented experiments on the above three datasets.

[0102] Among them, the selected comparison methods are all more advanced recommendation methods in recent years, such as the FPMC method based on Markov chain, the GRU4Rec method based on GRU unit, the Caser method based on convolutional neural network, and the DIB-PEB method that adds recommendation enhancement based on GRU unit.

[0103] The selected evaluation indicators are recall rate Recall and normalized discounted cumulative gain NDCG. The larger the value of these two indicators, the better the recommendation effect. In all experiments, the top 10 movie projects are selected for recommendation to users, that is, K = 10.

[0104] Table 1. Experimental results of the present invention and other methods on the ML-100k movie sequence dataset

[0105] FPMC GRU4Rec Caser DIB-PEB The present invention Recall 0.2500 0.2725 0.2618 0.3294 0.3755 NDCG 0.1247 0.1404 0.1345 0.1717 0.1939

[0106] Table 2. Experimental results of the present invention and other methods on the ML-1m movie sequence dataset

[0107] FPMC GRU4Rec Caser DIB-PEB The present invention Recall 0.2585 0.4435 0.4471 0.4585 0.5041 NDCG 0.1284 0.2466 0.2577 0.2558 0.2815

[0108] Table 3. Experimental results of the present invention and other methods on the ML-25m movie sequence dataset

[0109] FPMC GRU4Rec Caser DIB-PEB The present invention Recall 0.0102 0.6124 0.6013 0.6135 0.6628 NDCG 0.0045 0.3749 0.3803 0.3656 0.3984

[0110] From the above experimental results, we can see that in movie sequence data sets with different data sizes, the method proposed in the present invention has a certain improvement compared with the sequence recommendation methods based on different algorithms. This can show that the personalized movie sequence recommendation method based on the deep clustering algorithm proposed in the present invention has certain advancement, and the deep clustering and user static feature mining proposed in the present method can effectively mine the user's high-level intentions and alleviate the long tail problem in sequence recommendation to a certain extent.

[0111] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A personalized movie sequence recommendation method based on deep clustering algorithm, It is characterized in that The personalized movie sequence recommendation method comprises the following steps: S1. Obtain the sequence data of the user watching movies, the user's personal attribute data and the attribute data of the movie itself as a movie sequence data set, and extract the user static information, interaction sequence information and movie project attribute information from it, wherein the user static information includes the user's gender and age, the interaction sequence information includes the user's movie viewing sequence, and the movie project attribute information includes the movie category, director and the year of the movie release; S2. All movie projects in the sampled data are taken as movie project groups. The cluster center corresponding to each movie project is obtained from the movie project group through deep clustering. Each movie project in the interaction sequence information obtained in step S1 is spliced ​​with the cluster center corresponding to each movie project to form an interaction sequence embedding that integrates high-level intentions. The process of step S2 is as follows: S2.

1. Create an embedding matrix W for the movie project group, and set the constant zero vector 0 as the filler for the blank part of the embedding matrix W. The creation process of the embedding matrix W is as follows: retrieve the attribute information of each movie project and stack them together to generate the embedding matrix W∈R n×d , where d is the potential dimension, here assumed to be 1, and the embedding matrix W is defined as follows, where m n Represents the attribute information of the nth movie project in the movie project group: ; S2.2, Assume that the entire movie project group is divided into k classes, each class has a cluster center, first initialize the k cluster centers, and then assume that the distance between each sample point and the cluster center to which the sample point belongs satisfies the following distribution: In the above formula, i represents the i-th sample, j represents the j-th cluster center in this round of clustering, and z i represents the spatial distribution of the original movie project attribute information after the deep clustering preliminary encoding process, μ j represents the center point of the jth cluster, α is the degree of freedom, q ij represents the probability that sample i belongs to cluster j; S2.3, determine the initial cluster center according to step S2.2, and then use the distribution formula to measure whether the sample belongs to a certain cluster, and calculate a compact auxiliary distribution that conforms to the cluster center: In the above formula, p ij represents the probability that sample i belongs to cluster j, represents the soft allocation function, ; S2.4, calculate the distance loss function according to the two different distributions obtained in step S2.2 and step S2.3, and continue to iterate until convergence, and the cluster center also changes with each round of iterative optimization, wherein the distance loss function is as follows: ； The gradient update principle formula for each iteration is as follows: Concatenate each movie item in the interaction sequence embedding with the cluster center corresponding to this movie item to form an interaction sequence embedding that incorporates high-level intent; S3, mining user static information, splicing user static information with interaction sequence information to form user feature dynamic sequence embedding; S4, combining the interaction sequence embedding fused with the high-level intent and the dynamic sequence embedding of the user features into a gated recurrent unit to capture the fused sequence information; S5. Use the fused sequence information to calculate the matching scores of all candidate movie projects and user interaction sequences, and select the top K projects as recommendation results to make personalized movie recommendations for users.

2. According to claim 1, a personalized movie sequence recommendation method based on a deep clustering algorithm, It is characterized in that The process of step S1 is as follows: S1.1, extracting user static information, interaction sequence information and movie project attribute information from the movie sequence dataset; S1.2, dividing the interaction sequence information, using the penultimate interaction behavior in the interaction sequence as a test, the penultimate interaction behavior as a verification, and the other interaction behaviors as training; S1.3, let S = {s 1 ,s 2 ,…,s |V| } represents the interaction items in all interaction sequences, where |V| is the sum of all the interaction items; let A={s 1 ,s 2 ,…,s c } represents the interaction sequence of a user in a session, where s b represents the bth interaction behavior in this interaction sequence, s b ∈S, b=1,2, …,c, c represents the number of user interaction actions in a session; S1.

4. Create an embedding matrix E for all interaction sequences and set the constant zero vector 0 as the filler for the blank part of the embedding matrix E. The creation process of the embedding matrix E is as follows: retrieve the first p items of each interaction sequence and stack them together to produce the embedding matrix E∈R t×d , where d is the potential dimension, here assumed to be 1, the embedding matrix E is defined as follows, where A pr Represents the first p items of the rth interaction sequence: .

3. The method for personalized movie sequence recommendation based on deep clustering algorithm according to claim 1, It is characterized in that The process of step S3 is as follows: S3.

1. Create an embedding matrix U for the user static information, and set the constant zero vector 0 as the filler for the blank part of the embedding matrix U. The creation process of the embedding matrix is ​​as follows: retrieve the user static information of each user and stack them together to generate the embedding matrix U∈R v×d , where d is the potential dimension, which is assumed to be 1 here, and the matrix U is defined as follows, where u v Represents the user static information of the vth user: ; S3.

2. Concatenate the user static information embedding and the interaction sequence embedding to generate a user feature dynamic sequence embedding.

4. The method for personalized movie sequence recommendation based on deep clustering algorithm according to claim 1, It is characterized in that The gated recurrent unit in step S4 is referred to as a GRU unit. The σ() used in the GRU unit is a sigmoid activation function. The formula of the sigmoid activation function is as follows: , where x is the input; The working process of the GRU unit is as follows: First, the reset gate R is calculated in each GRU unit t , X t represents the sequence information of the GRU unit at time t, h t-1 represents the hidden state of the GRU unit at time t-1, W rx With W rh represents the first and second weight matrices to be learned, b r Indicates the first bias that needs to be learned: ； Calculate the update gate Z in each GRU unit t , where W zx With W zh represents the third and fourth weight matrices that need to be learned, b z Represents the second bias that needs to be learned: ； Next, the candidate states of the input information are calculated in each GRU unit , where W hx With W hh represents the fifth and sixth weight matrices that need to be learned, b h Indicates the third bias that needs to be learned: ； Finally, the memory cell state H at time t is calculated in each GRU unit t : ； The working process of the above GRU unit is abbreviated into the function GRU(). The interaction sequence embedding that integrates high-level intent and the dynamic sequence embedding of user features are combined and input into the GRU unit. The formula is as follows: ； Among them, a t , g t 、r t They represent the interaction sequence embedding at time t, the user feature dynamic sequence embedding, and the interaction sequence embedding fused with high-level intent. The three together form the input X at time t in the above GRU unit formula t ,h t This is the final fusion sequence information.

5. The personalized movie sequence recommendation method based on deep clustering algorithm according to claim 1, It is characterized in that In step S5, the fused sequence information is used as the sequence embedding representation, and then the sequence embedding representation is projected into the embedding matrix of the movie project group in step S2.

1. For the projected sequence embedding, y ul represents the matching score between the fused sequence information of user u and the movie project l, where W represents the movie project matrix and σ() is the sigmoid activation function: , select the top K movie projects with the highest scores to recommend to users.

6. The personalized movie sequence recommendation method based on deep clustering algorithm according to claim 1, It is characterized in that The overall objective function in step S5 is defined as: , using the above total objective function for training, and finally generating accurate personalized movie sequence recommendation results, where , ; In the above formula, W represents the embedding matrix of the movie project group, z represents the positive sample, f represents the negative sample, and y uz represents the positive sample matching score of user u, y uf represents the negative sample matching score of user u, and σ() is the sigmoid activation function.

Images