Recommendation method and device based on BLOOM model

Abstract

The invention discloses a recommendation method and device based on a BLOOM model, and belongs to the technical field of recommendation, and the method comprises the steps: collecting historical behavior data and project information data of a user, carrying out the cleaning and preprocessing, coding the preprocessed data text, extracting feature vectors, and fusing the feature vectors into a comprehensive feature vector, the similarity between user historical behaviors and project features is calculated through algorithms such as cosine similarity, the generated initial recommendation list is closer to actual preferences and demands of users, the deep learning model performs secondary screening according to the initial recommendation list to generate the final recommendation list, and the recommendation efficiency is improved through a comprehensive feature vector and cosine similarity algorithm. The similarity between the historical behaviors of the user and the project features is calculated, the generated initial recommendation list better fits the actual preference and demand of the user, secondary screening is performed on the initial recommendation list through the deep learning model, screening is performed through the threshold value, and the recommendation accuracy and efficiency are improved.

Description

Technical Field

[0001] The present invention belongs to the field of recommendation technology, and specifically refers to a recommendation method and device based on a BLOOM model. Background Art

[0002] With the rapid development of big data and artificial intelligence technologies, recommendation systems have become an indispensable part of Internet applications;

[0003] However, the existing recommendation methods still have certain defects. When constructing feature vectors, the existing recommendation methods only rely on relatively superficial features and fail to deeply mine the deep semantic information in text data, resulting in the feature vectors not being comprehensive and detailed enough when expressing user preferences and project features, which in turn affects the accuracy of recommendations. The model cannot fully capture the complexity and diversity of user preferences. The lack of a secondary screening mechanism leads to inaccurate recommendation results, which cannot fully meet the personalized needs of users and cannot adjust the recommendation list in time according to the real-time feedback of users. For this reason, a recommendation method and device based on the BLOOM model are proposed. Summary of the invention

[0004] The purpose of the present invention is to provide a recommendation method and device based on the BLOOM model to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a recommendation method and device based on the BLOOM model, comprising the following steps:

[0006] S1. Collect the user's historical behavior data and project information data, and perform cleaning and preprocessing;

[0007] S2, encode the preprocessed data text, extract feature vectors and fuse them into comprehensive feature vectors;

[0008] S3, making preliminary recommendations based on the extracted comprehensive feature vectors and generating an initial recommendation list;

[0009] S4. Build a deep learning model and train it. The training results are used to perform a secondary screening of the initial recommendation list.

[0010] S5. Sort the secondary screening results output by deep learning to generate a final recommendation list;

[0011] S6. Dynamically adjust the recommendation list based on user interaction feedback.

[0012] Among them, S1 obtains the user's historical behavior data and project data information in real time through the API interface, and performs data preprocessing operations such as data cleaning, format conversion, data integration, data segmentation and data conversion on the acquired historical behavior data and project data.

[0013] Among them, S2 performs text encoding according to the preprocessed data, decomposes the text into words and excludes meaningless words, converts the words into word embedding vectors, captures semantic information, and performs feature extraction after text encoding. The implementation formula is:

[0014]

[0015] In the formula, v d represents the comprehensive feature vector of text d, w i Represents the word embedding vector of the i-th word, tf(w i ,d) represents the frequency of the i-th word in text d, idf(w i ) represents the inverse text frequency of the i-th word, s(w i ) represents the sentiment score of the i-th word, N represents the total vocabulary in the text, d represents the total number of texts in the text collection, and α and β represent weight coefficients.

[0016] By converting text data into word embedding vectors that can capture deep semantic information, and through the feature extraction formula, comprehensively considering word frequency, inverse text frequency and sentiment score, it significantly improves the comprehensive capture of user preferences and project characteristics.

[0017] Among them, S3 makes preliminary recommendations based on the comprehensive feature vector extracted by S2, assuming that the historical behavior data vector is v d(u) , the historical behavior data vector is encoded, and the formula is v d(u) =BLOOM(h 1 ,h 2 ,...,h i ), h i represents the i-th item of historical behavior;

[0018] According to the project feature vector v d(i) and historical behavior data vector v d(u) The similarity is calculated by cosine similarity, and the implementation formula is:

[0019]

[0020] In S3, based on the calculated comprehensive feature vector similarity, k items are used as a recommendation list, and the initial recommendation list is implemented by the formula:

[0021]

[0022] In the formula, R(u) represents u recommendation, d k Represents a set of k items, sorted according to the generated recommendation list, and the implementation formula is:

[0023] R(u)=Top-k i∈I (sim(V d(u) ,V d(i) )).

[0024] Among them, S4 performs secondary screening based on the initial recommendation list generated and sorted by S3, and constructs a deep learning model M by integrating the feature vector v d Train and output the predicted score p. The implementation formula is:

[0025] p=M(V d );

[0026] Let a loss function L be used to maximize the positive sample score and minimize the negative sample score. Let y be the binary label. The implementation formula of the loss function L is:

[0027]

[0028] In the formula, L represents the loss function, which indicates the gap between the prediction result of the deep learning model and the true label, y represents the binary label, 1 represents interest, 0 represents disinterest, λ∥M∥ 2 represents regularization, ∥M∥ 2 represents the sum of squares of all weight parameters;

[0029] Based on the preliminary recommendation list, the deep learning model performs u Perform secondary screening and screen by threshold τ. The implementation formula is:

[0030] R′(u)={i∈R(u)|M([V d(u) ,V d(i) ,V d(c) ])>τ},

[0031] In the formula, R′(u) represents the final recommendation list, R(u) represents the initial recommendation list, τ represents the threshold, V d(c) Represents the context feature vector.

[0032] In S5, the secondary screening results output by deep learning are sorted to generate a final recommendation list, and the implementation formula is:

[0033] R′(u)=Top N ({i∈R(u)|M([V d (u),V d (i),V d (c)])>τ}),

[0034] In the formula, R′(u) represents the final recommendation list, Top N Means to select the top N items with the highest scores.

[0035] By combining algorithms such as feature vectors and cosine similarity, we calculate the similarity between user historical behavior and project features, and generate an initial recommendation list that is more in line with the user's actual preferences and needs. The initial recommendation list is then screened for a second time using a deep learning model and filtered using thresholds, which improves the accuracy and efficiency of recommendations.

[0036] Among them, S6 dynamically adjusts the recommendation list according to user interaction feedback, collects the interaction data between the user and the recommendation list in real time and provides real-time feedback, and the deep learning model is fine-tuned and updated according to the real-time feedback information and generates a new recommendation list.

[0037] Among them, the recommended device based on the BLOOM model is characterized by: comprising a data collection device, a processor, an integrated server and a data output device;

[0038] The data collection device is used to collect the user's historical behavior data and project information data in real time and pre-process the collected data;

[0039] The processor is used to perform text encoding, preliminary recommendation, model training and final recommendation on the preprocessed data text;

[0040] The integrated server is used to integrate various functional modules to form a recommendation system and is responsible for communication and data exchange between modules;

[0041] The data output device displays the final recommendation list to the user, and collects the interaction data between the user and the recommendation list in real time and feeds it back to the processor for real-time updating.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. In order to solve the problems of accuracy and real-time performance of existing recommendations and improve personalized, accurate, timely and dynamic recommendations, the present invention makes preliminary recommendations through comprehensive feature vectors and calculates the similarity between user historical behaviors and project features through algorithms such as cosine similarity. The generated initial recommendation list is closer to the user's actual preferences and needs. The deep learning model performs secondary screening based on the initial recommendation list to generate the final recommendation list.

[0044] 2. The present invention converts text data into word embedding vectors that can capture deep semantic information. Through the feature extraction formula, word frequency, inverse text frequency and sentiment score are comprehensively considered, which significantly improves the comprehensive capture of user preferences and project features;

[0045] 3. The present invention calculates the similarity between the user's historical behavior and the project features by using algorithms such as comprehensive feature vectors and cosine similarity, and generates an initial recommendation list that is more in line with the user's actual preferences and needs. The initial recommendation list is screened again by a deep learning model and screened by a threshold, thereby improving the accuracy and efficiency of the recommendation;

[0046] 4. The present invention quickly adjusts the recommendation list according to the real-time feedback of the user, and the dynamic update mechanism enables the recommendation system to respond to the changes in user needs in a timely manner and provide more personalized and timely recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The operation flow of the recommended method based on the BLOOM model of the present invention Figure 1 ;

[0048] Figure 2 The operation flow of the recommended method based on the BLOOM model of the present invention Figure 2 ;

[0049] Figure 3 It is a structural schematic diagram of the recommended device based on the BLOOM model of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0051] Example

[0052] See also Figure 1-Figure 3 As shown, the present invention provides a technical solution: comprising the following steps:

[0053] S1. Collect the user's historical behavior data and project information data, and perform cleaning and preprocessing;

[0054] S2, encode the preprocessed data text, extract feature vectors and fuse them into comprehensive feature vectors;

[0055] S3, making preliminary recommendations based on the extracted comprehensive feature vectors and generating an initial recommendation list;

[0056] S4. Build a deep learning model and train it. The training results are used to perform a secondary screening of the initial recommendation list.

[0057] S5. Sort the secondary screening results output by deep learning to generate a final recommendation list;

[0058] S6. Dynamically adjust the recommendation list based on user interaction feedback.

[0059] Among them, S1 obtains the user's historical behavior data and project data information in real time through the API interface, and performs data preprocessing operations such as data cleaning, format conversion, data integration, data segmentation and data conversion on the acquired historical behavior data and project data.

[0060] Among them, S2 performs text encoding according to the preprocessed data, decomposes the text into words and excludes meaningless words, converts the words into word embedding vectors, captures semantic information, and performs feature extraction after text encoding. The implementation formula is:

[0061]

[0062] In the formula, v d Represents the comprehensive feature vector of text d, w i Represents the word embedding vector of the i-th word, tf(w i ,d) represents the frequency of the i-th word in text d, idf(w i ) represents the inverse text frequency of the i-th word, s(w i ) represents the sentiment score of the i-th word, N represents the total vocabulary in the text, d represents the total number of texts in the text collection, and α and β represent weight coefficients.

[0063] Among them, S3 makes preliminary recommendations based on the comprehensive feature vector extracted by S2, assuming that the historical behavior data vector is v d(u) , the historical behavior data vector is encoded, and the formula is v d(u) =BLOOM(h 1 ,h 2 ,...,h i ), h i represents the i-th item of historical behavior;

[0064] According to the project feature vector v d(i) and historical behavior data vector v d(u) The similarity is calculated by cosine similarity, and the implementation formula is:

[0065]

[0066] In S3, based on the calculated comprehensive feature vector similarity, k items are used as a recommendation list, and the initial recommendation list is implemented by the formula:

[0067]

[0068] In the formula, R(u) represents u recommendation, dk Represents a set of k items, sorted according to the generated recommendation list, and the implementation formula is:

[0069] R(u)=Top-k i∈I (sim(V d(u) ,V d(i) )).

[0070] Among them, S4 performs secondary screening based on the initial recommendation list generated and sorted by S3, and constructs a deep learning model M by integrating the feature vector v d Train and output the predicted score p. The implementation formula is:

[0071] p=M(V d );

[0072] Let a loss function L be used to maximize the positive sample score and minimize the negative sample score. Let y be the binary label. The implementation formula of the loss function L is:

[0073]

[0074] In the formula, L represents the loss function, which indicates the gap between the prediction result of the deep learning model and the true label, y represents the binary label, 1 represents interest, 0 represents disinterest, λ∥M∥ 2 represents regularization, ∥M∥ 2 represents the sum of squares of all weight parameters;

[0075] Based on the preliminary recommendation list, the deep learning model performs u Perform secondary screening and screen by threshold τ. The implementation formula is:

[0076] R′(u)={i∈R(u)|M([V d(u) ,V d(i) ,V d(c) ])>τ},

[0077] In the formula, R′(u) represents the final recommendation list, R(u) represents the initial recommendation list, τ represents the threshold, V d(c) Represents the context feature vector.

[0078] In S5, the secondary screening results output by deep learning are sorted to generate a final recommendation list, and the implementation formula is:

[0079] R′(u)=Top N ({i∈R(u)|M([V d (u),V d (i),V d (c)])>τ}),

[0080] In the formula, R′(u) represents the final recommendation list, Top N Means to select the top N items with the highest scores.

[0081] Among them, S6 dynamically adjusts the recommendation list according to user interaction feedback, collects the interaction data between the user and the recommendation list in real time and provides real-time feedback, and the deep learning model is fine-tuned and updated according to the real-time feedback information and generates a new recommendation list.

[0082] Among them, the recommended device based on the BLOOM model is characterized by: comprising a data collection device, a processor, an integrated server and a data output device;

[0083] The data collection device is used to collect the user's historical behavior data and project information data in real time and pre-process the collected data;

[0084] The processor is used to perform text encoding, preliminary recommendation, model training and final recommendation on the preprocessed data text;

[0085] The integrated server is used to integrate various functional modules to form a recommendation system and is responsible for communication and data exchange between modules;

[0086] The data output device displays the final recommendation list to the user, and collects the interaction data between the user and the recommendation list in real time and feeds it back to the processor for real-time updating.

[0087] In this example, specifically: S2 performs text encoding based on the preprocessed data, decomposes the text into words and excludes meaningless words, converts the words into word embedding vectors, captures semantic information, and performs feature extraction after text encoding. The implementation formula is:

[0088]

[0089] In the formula, v d represents the comprehensive feature vector of text d, w i Represents the word embedding vector of the i-th word, tf(w i ,d) represents the frequency of the i-th word in text d, idf(w i ) represents the inverse text frequency of the i-th word, s(w i ) represents the sentiment score of the i-th word, N represents the total vocabulary in the text, d represents the total number of texts in the text collection, and α and β represent weight coefficients.

[0090] In this example, specifically: S3, based on the calculated comprehensive feature vector similarity, selects k items as the recommendation list, and the initial recommendation list is implemented as follows:

[0091]

[0092] In the formula, R(u) represents u recommendation, d k Represents a set of k items, sorted according to the generated recommendation list, and the implementation formula is:

[0093] R(u)=Top-k i∈I (sim(V d(u) ,V d(i) )).

[0094] According to the initial recommendation list generated and sorted by S3, a secondary screening is performed to build a deep learning model M through the comprehensive feature vector v d Train and output the predicted score p. The implementation formula is:

[0095] p=M(V d );

[0096] Let a loss function L be used to maximize the positive sample score and minimize the negative sample score. Let y be the binary label. The implementation formula of the loss function L is:

[0097]

[0098] In the formula, L represents the loss function, which indicates the gap between the prediction result of the deep learning model and the true label, y represents the binary label, 1 represents interest, 0 represents disinterest, λ∥M∥ 2 represents regularization, ∥M∥ 2 represents the sum of squares of all weight parameters;

[0099] Based on the preliminary recommendation list, the deep learning model performs u Perform secondary screening and screen by threshold τ. The implementation formula is:

[0100] R′(u)={i∈R(u)|M([V d(u) ,V d(i) ,V d(c) ])>τ},

[0101] In the formula, R′(u) represents the final recommendation list, R(u) represents the initial recommendation list, τ represents the threshold, V d(c) Represents the context feature vector.

[0102] In this example, specifically: sorting is performed based on the secondary screening results output by deep learning to generate the final recommendation list. The implementation formula is:

[0103] R′(u)=Top N ({i∈R(u)|M([V d (u),Vd (i),V d (c)])>τ}),

[0104] In the formula, R′(u) represents the final recommendation list, Top N Means to select the top N items with the highest scores.

[0105] Working principle: Obtain the user's historical behavior data and project information data in real time through the API interface, and perform cleaning, format conversion, data integration, data segmentation and data conversion preprocessing operations on the collected data;

[0106] The preprocessed text data is decomposed into words, meaningless words are excluded, and the words are converted into word embedding vectors to capture semantic information. The word embedding vectors, word frequency, inverse text frequency and sentiment scores are combined to form a comprehensive feature vector. According to the item feature vector and the user historical behavior data vector, the similarity between them is calculated by methods such as cosine similarity. According to the similarity calculation results, the k items with the highest similarity are selected as the initial recommendation list and sorted. A deep learning model is constructed, and the comprehensive feature vector is used for training. The prediction score is output. The prediction accuracy of the model is improved by optimizing the loss function. The preliminary recommendation list is screened for the second time by the trained deep learning model. By setting a threshold, only items with scores higher than the threshold are retained. The items that have been screened for the second time are sorted, and the top N items with the highest scores are selected as the final recommendation list. The interaction data between users and the recommendation list is collected in real time. According to user feedback, the deep learning model is fine-tuned to optimize the recommendation results. A new recommendation list is generated and displayed to the user through the updated model.

[0107] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0108] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. The recommendation method based on the BLOOM model is characterized by: The following steps are involved: S1. Collect the user's historical behavior data and project information data, and perform cleaning and preprocessing; S2, encode the preprocessed data text, extract feature vectors and fuse them into comprehensive feature vectors; S3, making preliminary recommendations based on the extracted comprehensive feature vectors and generating an initial recommendation list; S4. Build a deep learning model and train it. The training results are used to perform a secondary screening of the initial recommendation list. S5. Sort the secondary screening results output by deep learning to generate a final recommendation list; S6. Dynamically adjust the recommendation list based on user interaction feedback.

2. The recommendation method based on the BLOOM model according to claim 1, characterized in that: The S1 obtains the user's historical behavior data and project data information in real time through the API interface, and performs data preprocessing operations such as data cleaning, format conversion, data integration, data segmentation and data conversion on the obtained historical behavior data and project data.

3. The recommendation method based on the BLOOM model according to claim 1, characterized in that: S2, according to the preprocessed data, performs text encoding, decomposes the text into words and excludes meaningless words, converts the words into word embedding vectors, captures semantic information, and performs feature extraction after text encoding. The implementation formula is: In the formula, v d Represents the comprehensive feature vector of text d, w i Represents the word embedding vector of the i-th word, tf(w i ,d) represents the frequency of the i-th word in text d, idf(w i ) represents the inverse text frequency of the i-th word, s(w i ) represents the sentiment score of the i-th word, N represents the total vocabulary in the text, d represents the total number of texts in the text collection, and α and β represent weight coefficients.

4. The recommendation method based on the BLOOM model according to claim 1, characterized in that: S3 makes a preliminary recommendation based on the comprehensive feature vector extracted by S2. Let the historical behavior data vector be v d(u) , the historical behavior data vector is encoded, and the formula is v d(u) =BLOOM(h1,h2,...,h i ), h i represents the i-th item of historical behavior; According to the project feature vector v d(i) and historical behavior data vector v d(u) The similarity is calculated by cosine similarity, and the implementation formula is:

5. The recommendation method based on the BLOOM model according to claim 1, characterized in that: In S3, based on the calculated comprehensive feature vector similarity, k items are used as a recommendation list, and the initial recommendation list is implemented as follows: In the formula, R(u) represents u recommendation, d k Represents a set of k items, sorted according to the generated recommendation list, and the implementation formula is: R(u)=Top-k i∈I (sim(V d(u) ,V d(i) ))。 6. The recommendation method based on the BLOOM model according to claim 1, characterized in that: S4 performs secondary screening based on the sorted initial recommendation list generated by S3, and constructs a deep learning model M by integrating the feature vector v d Train and output the predicted score p. The implementation formula is: p=M(V d ); Let a loss function L be used to maximize the positive sample score and minimize the negative sample score. Let y be the binary label. The implementation formula of the loss function L is: In the formula, L represents the loss function, which indicates the gap between the prediction result of the deep learning model and the true label, y represents the binary label, 1 represents interest, 0 represents disinterest, λ∥M∥ 2 represents regularization, ∥M∥ 2 represents the sum of squares of all weight parameters; Based on the preliminary recommendation list, the deep learning model performs u Perform secondary screening and screen by threshold τ. The implementation formula is: R′(u)={i∈R(u)|M([V d(u) ,V d(i) ,V d(c) ])>τ}, In the formula, R′(u) represents the final recommendation list, R(u) represents the initial recommendation list, τ represents the threshold, V d(c) Represents the context feature vector.

7. The recommendation method based on the BLOOM model according to claim 1, characterized in that: S5, sorting the secondary screening results output by deep learning to generate a final recommendation list, the implementation formula is: R′(u)=Top N ({i∈R(u)|M([V d (u),V d (i),V d (c)])>τ}), In the formula, R′(u) represents the final recommendation list, Top N Means to select the top N items with the highest scores.

8. The recommendation method based on the BLOOM model according to claim 1, characterized in that: The S6 dynamically adjusts the recommendation list according to user interaction feedback, collects the interaction data between the user and the recommendation list in real time and provides real-time feedback, and the deep learning model is fine-tuned and updated according to the real-time feedback information and generates a new recommendation list.

9. Recommended equipment based on the BLOOM model, characterized by: It includes a data collection device, a processor, an integrated server and a data output device; The data collection device is used to collect the user's historical behavior data and project information data in real time and pre-process the collected data; The processor is used to perform text encoding, preliminary recommendation, model training and final recommendation on the preprocessed data text; The integrated server is used to integrate various functional modules to form a recommendation system and is responsible for communication and data exchange between modules; The data output device displays the final recommendation list to the user, and collects the interaction data between the user and the recommendation list in real time and feeds it back to the processor for real-time updating.

Images