CN113220994B - User personalized information recommendation method based on target object enhanced representation - Google Patents
User personalized information recommendation method based on target object enhanced representation Download PDFInfo
- Publication number
- CN113220994B CN113220994B CN202110498063.4A CN202110498063A CN113220994B CN 113220994 B CN113220994 B CN 113220994B CN 202110498063 A CN202110498063 A CN 202110498063A CN 113220994 B CN113220994 B CN 113220994B
- Authority
- CN
- China
- Prior art keywords
- feature vector
- target
- user
- articles
- recommended
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 124
- 239000013598 vector Substances 0.000 claims abstract description 139
- 238000005070 sampling Methods 0.000 claims abstract description 62
- 238000013507 mapping Methods 0.000 claims abstract description 23
- 230000006870 function Effects 0.000 claims description 23
- 238000012545 processing Methods 0.000 claims description 16
- 238000010606 normalization Methods 0.000 claims description 13
- 239000000126 substance Substances 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 5
- 238000010276 construction Methods 0.000 claims description 3
- 230000006399 behavior Effects 0.000 abstract description 9
- 238000013473 artificial intelligence Methods 0.000 abstract description 2
- 238000013135 deep learning Methods 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 10
- 230000008569 process Effects 0.000 description 9
- 238000004590 computer program Methods 0.000 description 8
- 230000008901 benefit Effects 0.000 description 6
- 230000006872 improvement Effects 0.000 description 6
- 238000013459 approach Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 230000003796 beauty Effects 0.000 description 4
- 238000004891 communication Methods 0.000 description 4
- 230000007246 mechanism Effects 0.000 description 4
- 230000003287 optical effect Effects 0.000 description 4
- 238000004422 calculation algorithm Methods 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000001914 filtration Methods 0.000 description 3
- 230000003993 interaction Effects 0.000 description 2
- 230000001737 promoting effect Effects 0.000 description 2
- 230000000644 propagated effect Effects 0.000 description 2
- 239000004065 semiconductor Substances 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000012733 comparative method Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000000835 fiber Substances 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 239000004973 liquid crystal related substance Substances 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention belongs to the technical field of artificial intelligence and deep learning, and particularly relates to a user personalized information recommendation method based on target object enhanced representation, aiming at solving the problems of low recommendation accuracy and poor robustness caused by the fact that the existing personalized information recommendation method is limited by user behavior records. The method comprises the steps of obtaining articles in a user history record and target articles to be recommended, and mapping the articles and the target articles to be recommended into a feature vector with a set dimension; acquiring an interest representation of a user about a target item; sampling a set number of articles from the intersection of the first data set and the target article to be recommended by a preset sampling method, and mapping the articles into a feature vector; obtaining an enhanced representation of a target item; predicting the interest degree of the target object to be recommended by the user, and recommending the target object to be recommended corresponding to the top N sorted interest degrees to the user. The invention improves the recommendation accuracy and robustness of the existing personalized information recommendation method.
Description
Technical Field
The invention belongs to the technical field of artificial intelligence and deep learning, and particularly relates to a user personalized information recommendation method, system and device based on target object enhanced representation.
Background
With the development of technology, the speed of information generation has far exceeded the degree that users can handle themselves. When searching for the information needed by the user, the user spends a great deal of time screening and filtering irrelevant information. The purpose of personalized information recommendation is to filter information and recommend information of potential interest to a user by analyzing relevant historical records and user profiles of the user instead of the user.
The personalized information recommendation model can be abstracted into a simple binary or regression model, wherein the input is the characteristics related to the user, including the characteristics of the related historical records, personal data and target information, and the output is whether the user is interested or the interest degree of the user. The model compares the characteristics of the user with the characteristics of the target information and makes a decision whether to recommend the information to the user. The binary classification model can be directly used as a decision according to the classification result, and the regression model can obtain the top N pieces of information most likely to be interested by sequencing the scores of all target information.
Other recommendation algorithms, such as collaborative filtering algorithms, exist that make recommendations for items by analyzing users for similar behavior records. However, this approach is limited to user behavior records, resulting in a model that is limited in generalization capability. Some recent methods use a deep neural network to encode user features and article features, and obtain deep nonlinear features through a multilayer perceptron, which achieves certain effects. Based on the method, the invention provides a user personalized information recommendation method based on the enhanced representation of the target object.
Disclosure of Invention
In order to solve the above problems in the prior art, that is, to solve the problem that the existing personalized information recommendation method is limited by user behavior records, and thus the recommendation accuracy is low and the robustness is poor, in a first aspect of the present invention, a user personalized information recommendation method based on target object enhanced representation is provided, the method comprising:
s10, acquiring the articles in the user history record and the target articles to be recommended, and mapping the articles and the target articles to be recommended into a feature vector with set dimensionality; taking the feature vector mapped by the object in the user history record as a first feature vector, and taking the feature vector mapped by the target object to be recommended as a second feature vector;
s20, calculating the similarity of the first feature vector and the second feature vector, and carrying out normalization processing on the similarity; carrying out weighted summation on each first feature vector through the normalized similarity, and taking the weighted summation as an interest expression of the user about the target object;
s30, acquiring an intersection of the first data set and the target object to be recommended; sampling a set number of articles from the intersection by a preset sampling method, and mapping the articles into a feature vector as a third feature vector; the first data set is a data set constructed based on items of a plurality of user history records;
s40, calculating the similarity of the second feature vector and the third feature vector, and normalizing the similarity; weighting and summing the third feature vectors through the normalized similarity to obtain an enhanced representation of the target object;
s50, splicing the second feature vector, the interest representation of the user about the target object and the enhanced representation about the target object, inputting the spliced second feature vector into a pre-constructed classifier, and predicting the interest degree of the user to the target object to be recommended; sorting according to the interest degrees, and recommending the target objects to be recommended corresponding to the first N interest degrees to the user after sorting; n is a positive integer.
In some preferred embodiments, in step S20, "calculating the similarity between the first feature vector and the second feature vector, and performing normalization processing on each similarity", the method includes:
calculating the similarity of the first feature vector and the second feature vector through a multilayer perceptron;
and (5) normalizing each similarity by using a softmax function.
In some preferred embodiments, in step S30, "a set number of articles are sampled from the intersection by a preset sampling method", which includes:
and sampling a set number of articles from the intersection by any one of a global sampling method, a local sampling method, a weighted sampling method and a uniform sampling method.
In some preferred embodiments, the weighted sampling method is:
wherein the content of the first and second substances,
representing the probability that item k is sampled under the global sampling method,
representing the probability that item k is sampled under the local sampling method, alpha represents the weight,
indicating the number of occurrences of item k in the first data set,
indicating the number of times item k co-occurred with the target item to be recommended, Oi representing a set of item i and item constructions in each user history of the first data set other than item i,
indicating the number of occurrences of item j in the first data set,
denotes an article j is Oi The number of occurrences in (c).
In some preferred embodiments, the uniform sampling method is:
in some preferred embodiments, the classifier is constructed based on a multi-tier perceptron; the multilayer perceptron is three layers, reLu activation functions are adopted among the layers, and the dimensions of the layers are 80, 40 and 1 in sequence.
In a second aspect of the present invention, a system for recommending personalized information of a user based on an enhanced representation of a target item is provided, the system comprising: the system comprises a characteristic mapping module, an interest representation acquisition module, a common article sampling module, an enhanced representation acquisition module and a target article recommendation module;
the characteristic mapping module is configured to acquire articles in the user history record and target articles to be recommended and map the articles and the target articles into a characteristic vector with set dimensionality; taking the feature vector mapped by the object in the user history record as a first feature vector, and taking the feature vector mapped by the target object to be recommended as a second feature vector;
the interest representation acquisition module is configured to calculate the similarity of the first feature vector and the second feature vector and carry out normalization processing on the similarity; carrying out weighted summation on each first feature vector through the normalized similarity, and taking the weighted summation as an interest expression of the user about the target object;
the common article sampling module is configured to obtain an intersection of the first data set and a target article to be recommended; sampling a set number of articles from the intersection by a preset sampling method, and mapping the articles into a feature vector as a third feature vector; the first data set is a data set constructed based on items of a plurality of user history records;
the enhancement representation acquisition module is configured to calculate the similarity of the second feature vector and the third feature vector and carry out normalization processing on the similarity; carrying out weighted summation on each third feature vector through the normalized similarity to serve as an enhanced representation about the target object;
the target article recommending module is configured to input the spliced second feature vector, the interest representation of the user about the target article and the enhanced representation about the target article into a pre-constructed classifier, and predict the interest degree of the user to the target article to be recommended; sorting according to the interest degrees, and recommending the target articles to be recommended corresponding to the first N interest degrees to the user after sorting; n is a positive integer.
In a third aspect of the invention, an apparatus is presented, at least one processor; and a memory communicatively coupled to at least one of the processors; wherein the memory stores instructions executable by the processor for implementing the method for recommending personalized information for a user based on an enhanced representation of a target item according to the claims.
In a fourth aspect of the present invention, a computer-readable storage medium is provided, which stores computer instructions for being executed by the computer to implement the method for recommending user-customized information based on enhanced representation of target item as claimed above.
The invention has the beneficial effects that:
the invention improves the recommendation accuracy and robustness of the existing personalized information recommendation method.
1) According to the method, the co-occurrence object of the target object is used, the enhanced representation of the target object is obtained after attention calculation, and the problems that the model representation is poor and the recommendation is limited by user behavior records due to low embedding representation quality of the low-frequency target object are solved;
2) According to the invention, the semantic information related to the target object is acquired from the co-occurrence object of the target object through the attention mechanism, and the semantic meaning of the target object is enriched through the information, so that a classifier in the model learns a more accurate mode. Compared with a baseline method, the method has the advantages that a better effect is achieved on the public data set, and the recommendation accuracy and robustness are improved.
Drawings
Other features, objects and advantages of the present application will become more apparent upon reading of the following detailed description of non-limiting embodiments thereof, made with reference to the accompanying drawings.
FIG. 1 is a flowchart illustrating a method for recommending personalized information for a user based on an enhanced representation of a target object according to an embodiment of the present invention;
FIG. 2 is a block diagram of a user personalized information recommendation system based on an enhanced representation of a target item according to an embodiment of the invention;
FIG. 3 is a diagram illustrating an exemplary structure of a personalized message recommendation model according to an embodiment of the invention;
FIG. 4 is a schematic diagram of the comparison of the effects of the Top-k recommendations of other models on different data sets according to an embodiment of the present invention;
FIG. 5 is a schematic illustration of a comparison of the recommendation effect on low frequency items on different data sets compared to other models according to an embodiment of the present invention;
fig. 6 is a schematic structural diagram of a computer system of an electronic device suitable for implementing the embodiments of the present application according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The present application will be described in further detail with reference to the following drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the relevant invention and not restrictive of the invention. It should be noted that, for convenience of description, only the portions related to the related invention are shown in the drawings.
It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict.
The invention relates to a user personalized information recommendation method based on enhanced representation of a target object, which comprises the following steps of:
s10, acquiring the articles in the user history record and the target articles to be recommended, and mapping the articles and the target articles to be recommended into a feature vector with set dimensionality; taking the feature vector mapped by the object in the user history record as a first feature vector, and taking the feature vector mapped by the target object to be recommended as a second feature vector;
s20, calculating the similarity of the first feature vector and the second feature vector, and carrying out normalization processing on the similarity; carrying out weighted summation on each first feature vector through the normalized similarity, and taking the weighted summation as an interest expression of the user about the target object;
s30, acquiring an intersection of the first data set and the target object to be recommended; sampling a set number of articles from the intersection by a preset sampling method, and mapping the articles into a feature vector as a third feature vector; the first data set is a data set constructed based on items of a plurality of user history records;
s40, calculating the similarity of the second feature vector and the third feature vector, and normalizing the similarity; carrying out weighted summation on each third feature vector through the normalized similarity to serve as an enhanced representation about the target object;
s50, splicing the second feature vector, the interest representation of the user about the target object and the enhanced representation about the target object, inputting the spliced second feature vector into a pre-constructed classifier, and predicting the interest degree of the user to the target object to be recommended; sorting according to the interest degrees, and recommending the target articles to be recommended corresponding to the first N interest degrees to the user after sorting; n is a positive integer.
In order to more clearly describe the method for recommending user personalized information based on enhanced representation of a target object, the following describes in detail various steps in an embodiment of the method of the present invention with reference to the drawings.
S10, acquiring the articles in the user history record and the target articles to be recommended, and mapping the articles and the target articles to be recommended into a feature vector with set dimensionality; taking the feature vector mapped by the object in the user history record as a first feature vector, and taking the feature vector mapped by the target object to be recommended as a second feature vector;
in the present embodiment, the user history S = { x =isacquired 1 ,x 2 ,...,x t Item x in (c) } i And target item x to be recommended e And embedding and representing the articles in the user history record and the target articles to be recommended through an embedding layer, wherein the articles in the user history record and the target articles to be recommended are represented by ID of the articles, namely mapping the ID of the articles in the user history record and the ID of the target articles to be recommended into the special article with dimension dA eigenvector used for representing the article.
And taking the feature vector mapped by the object in the user history record as a first feature vector, and taking the feature vector mapped by the target object to be recommended as a second feature vector.
S20, calculating the similarity of the first feature vector and the second feature vector, and carrying out normalization processing on the similarity; carrying out weighted summation on each first feature vector through the normalized similarity, and taking the weighted summation as an interest expression of the user about the target object;
in this embodiment, the similarity between the first feature vector and the second feature vector is calculated by a multi-layer perceptron MLP. The MLP is a three-layer multi-layer perceptron, the dimension size of each layer is 80, 40 and 1 respectively, and a ReLU activation function is adopted between layers.
After the relevance (namely the similarity) between the items in the user history record and the target items to be recommended is calculated, a softmax function is used for normalization processing on the whole user behavior history, namely, the softmax function is used for normalization processing on all the similarities, and the attention score is obtained. The attention score is used for adjusting the proportion of each article in the user history record in the final user interest expression, namely the normalized similarity is weighted and summed on the first feature vector to obtain the interest expression of the user about the target article. Specifically, the formula (1), (2) and (3) are shown as follows:
q i ＝MLP(concat(x i ,x e )) (3)
wherein q is i Denotes the degree of similarity, a i Representing the normalized similarity, i.e. attention score, h u Representing a representation of a user's interest in a target item, n representing a number of items in the user's history。
S30, acquiring an intersection of the first data set and a target object to be recommended; sampling a set number of articles from the intersection by a preset sampling method, and mapping the articles into a feature vector as a third feature vector; the first data set is a data set constructed based on items of a plurality of user history records;
in this embodiment, a fixed number of co-occurring items are first sampled from all items that co-occur with a target item in a first data set, the first data set being a data set constructed based on items of multiple user histories. The method comprises the steps of acquiring an intersection of a first data set and a target object to be recommended, sampling a set number of objects in the intersection, and mapping the objects to be recommended as feature vectors.
The sampling method comprises the following steps: any one of a global sampling method, a local sampling method, a weighted sampling method, and a uniform sampling method.
The probability of each item being sampled under the global sampling method is:
wherein the content of the first and second substances,
representing an article k Number of occurrences in the first data set, s k Representing the probability that the item k was sampled, Oi representing a set of item i and item constructions in each user history record of the first data set other than item i,
indicating the number of occurrences of item j in the first data set.
Using the global sampling approach, the model tends to use high frequency co-occurring artifacts for complementing the target commodity semantics. An advantage of high frequency co-occurring objects is that the model is more fully trained on its embedded representation due to the greater number of occurrences in the entire first data set. The enhanced representation semantics computed with these co-occurring article-embedded representations are also more unambiguous.
The probability of each item being sampled under the local sampling method is as follows:
wherein the content of the first and second substances,
indicating the number of times item k co-occurred with the target item to be recommended,
denotes that the object j is at O i The number of occurrences in (c).
Using local sampling methods, models tend to use high co-occurrence frequency of items for completing the target item semantics. An advantage of high co-occurrence frequency articles is that the co-occurrence frequency can be roughly considered as a degree of correlation with the target article, the more relevant the article is to the higher its co-occurrence frequency will be. The enhanced representations computed with these co-occurring item-embedded representations are thus more relevant to the actual semantics of the target good.
The probability of each item being sampled under the weighted sampling method is as follows:
wherein the content of the first and second substances,
representing an article k The probability of being sampled under the global sampling method,
representing an article k The probability of being sampled under the local sampling method, α, represents a weight, and is a continuous value between 0 and 1, and is used to control the specific weight of the two different methods in the final sampling probability.
The weighted sampling method is used to adjust the clarity and accuracy of the final enhancement representation. When alpha is close to 1, the sampling method is more prone to generate enhanced representation with strong semantic definition, and when alpha is close to 0, the sampling method is more prone to generate enhanced representation with high semantic relevance.
The probability of each item being sampled under the uniform sampling method is as follows:
the uniform sampling method is an unbiased sampling method, and the sampled probability of each co-occurrence article is the same.
S40, calculating the similarity of the second feature vector and the third feature vector, and normalizing the similarity; carrying out weighted summation on each third feature vector through the normalized similarity to serve as an enhanced representation about the target object;
in this embodiment, the similarity of the second feature vector and the third feature vector is calculated by the multi-layer perceptron MLP. The MLP is a three-layer multi-layer perceptron, the dimension size of each layer is 80, 40 and 1 respectively, and a ReLU activation function is adopted between layers.
And (3) normalizing each similarity by using a softmax function, and performing weighted summation on each third feature vector according to the normalized similarity, wherein the weighted summation is used as an enhanced representation of the target object, and is specifically shown in formulas (10), (11) and (12):
q i ＝MLP(concat(x i ,x e )) (12)
wherein r is e Representing an enhanced representation of the target item, i.e. the target item to be recommended via weighted sum, O e A set of item constructs representing a set number of samples in the intersection.
S50, splicing the second feature vector, the interest representation of the user about the target object and the enhanced representation about the target object, inputting the spliced second feature vector into a pre-constructed classifier, and predicting the interest degree of the user to the target object to be recommended; sorting according to the interest degrees, and recommending the target articles to be recommended corresponding to the first N interest degrees to the user after sorting; n is a positive integer.
In this embodiment, the second feature vector, the interest representation of the user about the target object, and the enhanced representation about the target object are spliced, and the spliced vector is input into a pre-constructed classifier (the classifier is constructed based on a multi-layer perceptron, the structure is three layers, the dimensions of each layer are 80, 40, and 1, respectively, and a ReLU activation function is adopted between the layers), so as to obtain the interest degree of the user about the target object to be recommended. Specifically, as shown in formula (13):
y ue ＝MLP(concat(x e ,h u ,r e )) (13)
wherein, y ue The splicing function of the concat representation splices all the representations of the input into a single representation.
The classifier outputs the probability y that the target item to be recommended meets the user interest ue And for a large number of target items, after calculating the possibility of the target items, and then sorting the top N target items to be recommended according to the scores, recommending the top N target items to the user.
In the invention, the personalized information recommendation model adopts a back propagation algorithm in the training process, and cross entropy is used as a target function. As shown in equation (14):
wherein the content of the first and second substances,
the label and model representing the ith sample are based on the output of the ith sample, and N is the size of the training set. For the tag y ∈ {0,1},0 indicates that the user is not interested in the target item, and 1 indicates that the user is interested. The personalized information recommendation model comprises an embedding mapping module, a first attention module, a second attention module and a classification module, as shown in FIG. 3 (the co-occurrence item set in FIG. 3 is O in the above i )；
The embedded module is used for mapping the articles in the user history record and the target articles to be recommended into the feature vector; taking the feature vector mapped by the object in the user history record as a first feature vector, and taking the feature vector mapped by the target object to be recommended as a second feature vector;
the first attention module is used for calculating the similarity of the first feature vector and the second feature vector and normalizing the similarity; carrying out weighted summation on each first feature vector through the normalized similarity, and taking the weighted summation as an interest expression of the user about the target object;
the second attention module is used for calculating the similarity of the second feature vector and the third feature vector and normalizing the similarity; carrying out weighted summation on each third feature vector through the normalized similarity to serve as an enhanced representation about the target object;
the classification module is used for splicing the second feature vector, the interest representation of the user about the target object and the enhanced representation about the target object and then inputting the spliced second feature vector, the interest representation and the enhanced representation into a pre-constructed classifier, and predicting the interest degree of the user to the target object to be recommended; and sorting according to the interest degrees, and recommending the target articles to be recommended corresponding to the first N interest degrees to the user after sorting.
In addition, to verify the effectiveness of the present method, experiments were performed on 2 published amazon datasets using the present method. Amazon data set is a reference data set commonly used in the recommendation field, wherein two subsets of Beauty and boots are selected, and the characteristics of each commodity only comprise a commodity ID.
All the data sets above only retain user data with behavior record number greater than 5, and truncate the last 200 records retained for each user. The model is trained by predicting the (n-1) th item by using the first n-2 items, and the model is tested by predicting the nth item by using the first n-1 items, wherein n refers to the length of the user behavior record. We provide statistics for all datasets in table 1.
TABLE 1
The comparative method used in this experiment consisted of: (1) MF, a standard matrix decomposition model that utilizes user and item interaction information. (2) MLP, this approach uses a multi-layered perceptron instead of a dot product learning scoring function. Here we use the same MLP structure as in the present model. (3) Youtube Rec, youtube corporation used this method on their video recommendation system. This model takes an average of all the embedded representations of the user-interacted video and then feeds them to the multi-layered perceptron. (4) FISM is one of the best current methods based on item recommendation, which uses all items that the user has interacted with to represent the user and learns the scoring function as a dot product. (5) NAIS additionally introduced an attention mechanism to dynamically adjust the specific gravity of each historical item in the final interest representation. (6) DIN is a widely used method in the CTR field and is structurally similar to NAIS, but it uses MLP to learn the scoring function.
In this experiment, the dimension of the embedded representation for each article was 64. We chose SGD as the optimizer, with a learning rate of 1.0, setting the batch size to 32. For comparative fairness, we randomly select 100 negative examples for each user, and then sort the target items among the 100 negative examples. Our selected evaluation criteria include hit rate HR @10 and NDCG @10. We also used a relative improvement index RelaImpr to represent the degree of improvement in model effect over the baseline model. For all indices, a larger value indicates a better model performance.
Table 2 presents a comparison of experimental results for the baseline models of the invention and comparison, where RI represents RelaImpr versus FISM.
TABLE 2
From the experimental results, we can first see that our model CER performed best on all data sets and exceeded much the best methods NAIS and DIN in the baseline. On the Beauty dataset, our method showed about 10.42% improvement in HR and about 10.51% relative improvement in NDCG compared to FISM. On the Books dataset, our method has a relative boost on HR of about 17.87% and on NDCG of about 26.74% compared to FISM. For convenience of comparison with other methods, we can roughly classify all baseline models into three classes. The first category is user-based methods including MF and MLP. The second category uses items to represent users, including Youtube, FISM, NAIS, and DIN. The last category is the model we propose, using items to represent users and items. As shown in the table, the methods under each category may employ different modules, such as using MLP or attention mechanisms, which may result in some difference in the actual performance of the models. However, there is often a significant gap in performance between the different categories of methods.
For the user-based approach, whether inner product or MLP is used as the interaction function, it always achieves the worst results in all baseline models. This is because the user's interests are multidimensional, and representing the user's multiple interests using a single vector limits the model's fitting and generalization capabilities.
The second category of methods performs well on both datasets compared to the user-based methods. Their improvement comes mainly from the use of more information to encode the user's interests, which improves the quality of the user representation. The methods that use the attention mechanism perform a bit better than the methods that do not use attention. This demonstrates that dynamically adjusting the user's expression of interest for a target item plays an important role in promoting model performance.
Based on the second category of methods, our proposed CER method provides an enhanced representation of the target item based on co-occurring items. The enhancement expression is helpful for relieving the problem of semantic deletion of low-frequency secondary articles caused by long-tail distribution, and therefore, the enhancement expression plays an important role in promoting the article-based collaborative filtering method. As we see in table 2, our method is much better than the second class of methods on both datasets, demonstrating the effectiveness of our method.
FIG. 4 shows the performance of each model on Top-k recommendation tasks, where k has a value of 1 to 10. We can see from the figure that our proposed CER model is better than the second class model in the baseline in all positions. In these baseline models, NAIS and DIN performed relatively well on Beauty and boots datasets.
Fig. 5 illustrates the behavior of various models on low frequency secondary target items. We sort the target items in the test set by their global frequency. For each frequency, we calculated two evaluation indices for all test samples at that frequency. Notably, the 5-core property of a data set may be broken due to the truncation used in constructing the data set. Since we are only interested in low frequency secondary target items, here we have chosen a frequency of 0 to 50. First, we can see from the figure that in all article-based approaches, the model behaves better as the global frequency of the target article increases. This demonstrates that the frequency of the target item has a significant impact on the performance of the item-based method. Furthermore, we can see that the model performs much less frequently on the target item (frequency 0 to 30) than the model does on the data set as a whole. These target items can greatly pull down the overall appearance of the model due to the long tail distribution.
Second, on the Beauty dataset, we propose models that are superior to other baseline models, especially at frequencies in the 0 to 30 interval. On Books data sets, our method is much better than other baseline models in both evaluation indexes. This illustrates that the enhanced representation helps to alleviate the semantic missing problem of the target item. To explain this figure more clearly, we take Books dataset as an example. Considering the HR value at a frequency of 10, the best represented DIN in the baseline model is close to 0.41. However our model corresponds to values close to 0.57 with a relative improvement of close to 39% compared to DIN. Considering again the frequency at which the NDCG value is 0.5, DIN performs best in the baseline model, corresponding to a frequency of 44. For our model, this value is 21. That is, to achieve the same NDCG performance, our model only requires 21 occurrences of the target item in the data set, while the best baseline method requires at least 44 occurrences. Overall, experimental results show that our model has irreplaceable advantages over other item-based methods on low-frequency sub-target items.
A second embodiment of the present invention provides a system for recommending user personalized information based on enhanced representation of a target item, as shown in fig. 2, specifically including: the system comprises a feature mapping module 100, an interest representation acquisition module 200, a common article sampling module 300, an enhanced representation acquisition module 400 and a target article recommendation module 500;
the feature mapping module 100 is configured to obtain the articles in the user history record and the target articles to be recommended, and map the articles and the target articles into feature vectors with set dimensions; taking the feature vector mapped by the object in the user history record as a first feature vector, and taking the feature vector mapped by the target object to be recommended as a second feature vector;
the interest representation obtaining module 200 is configured to calculate similarities of the first feature vector and the second feature vector, and normalize the similarities; weighting and summing the first characteristic vectors through the normalized similarity to be used as the interest expression of the user about the target object;
the common item sampling module 300 is configured to obtain an intersection of the first data set and the target item to be recommended; sampling a set number of articles from the intersection by a preset sampling method, and mapping the articles into a feature vector as a third feature vector; the first data set is a data set constructed based on items of a plurality of user history records;
the enhanced representation obtaining module 400 is configured to calculate similarities of the second feature vector and the third feature vector, and perform normalization processing on the similarities; carrying out weighted summation on each third feature vector through the normalized similarity to serve as an enhanced representation about the target object;
the target item recommendation module 500 is configured to input a second feature vector, an interest representation of the user about the target item, and an enhanced representation about the target item after splicing into a pre-constructed classifier, and predict an interest degree of the user about the target item to be recommended; sorting according to the interest degrees, and recommending the target articles to be recommended corresponding to the first N interest degrees to the user after sorting; n is a positive integer.
It can be clearly understood by those skilled in the art that, for convenience and brevity of description, the specific working process and related description of the system described above may refer to the corresponding process in the foregoing method embodiment, and details are not described herein again.
It should be noted that, the system for recommending user personalized information based on enhanced representation of a target object provided in the foregoing embodiment is only illustrated by dividing the functional modules, and in practical applications, the functions may be allocated to different functional modules according to needs, that is, the modules or steps in the embodiment of the present invention are further decomposed or combined, for example, the modules in the foregoing embodiment may be combined into one module, or may be further split into multiple sub-modules, so as to complete all or part of the functions described above. Names of the modules and steps related in the embodiments of the present invention are only for distinguishing the modules or steps, and are not to be construed as unduly limiting the present invention.
An apparatus of a third embodiment of the invention, at least one processor; and a memory communicatively coupled to at least one of the processors; wherein the memory stores instructions executable by the processor for implementing the method for recommending personalized information for a user based on an enhanced representation of a target item according to the claims.
A computer-readable storage medium of a fourth embodiment of the present invention stores computer instructions for being executed by the computer to implement the method for recommending personalized information for a user based on an enhanced representation of a target item according to the claims.
It can be clearly understood by those skilled in the art that, for convenience and brevity, the specific working processes and related descriptions of the storage device and the processing device described above may refer to the corresponding processes in the foregoing method examples, and are not described herein again.
Referring now to FIG. 6, there is illustrated a block diagram of a computer system suitable for use as a server in implementing embodiments of the system, method and apparatus of the present application. The server shown in fig. 6 is only an example, and should not bring any limitation to the functions and the scope of use of the embodiments of the present application.
As shown in fig. 6, the computer system includes a Central Processing Unit (CPU) 601, which can perform various appropriate actions and processes according to a program stored in a Read Only Memory (ROM) 602 or a program loaded from a storage section 608 into a Random Access Memory (RAM) 603. In the RAM603, various programs and data necessary for system operation are also stored. The CPU601, ROM602, and RAM603 are connected to each other via a bus 604. An Input/Output (I/O) interface 605 is also connected to bus 604.
The following components are connected to the I/O interface 605: an input portion 606 including a keyboard, a mouse, and the like; an output portion 607 including a display such as a cathode ray tube, a liquid crystal display, and a speaker; a storage section 608 including a hard disk and the like; and a communication section 609 including a network interface card such as a lan card, a modem, or the like. The communication section 609 performs communication processing via a network such as the internet. The driver 610 is also connected to the I/O interface 605 as needed. A removable medium 611 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is mounted on the drive 610 as necessary, so that a computer program read out therefrom is mounted in the storage section 608 as necessary.
In particular, the processes described above with reference to the flow diagrams may be implemented as computer software programs, according to embodiments of the present disclosure. For example, embodiments of the present disclosure include a computer program product comprising a computer program embodied on a computer readable medium, the computer program comprising program code for performing the method illustrated in the flow chart. In such an embodiment, the computer program may be downloaded and installed from a network through the communication section 609, and/or installed from the removable medium 611. The computer program, when executed by the CPU601, performs the above-described functions defined in the method of the present application. It should be noted that the computer readable medium mentioned above in the present application may be a computer readable signal medium or a computer readable storage medium or any combination of the two. The computer-readable storage medium may be, for example but not limited to: an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of the computer readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a RAM, a ROM, an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present application, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. In this application, however, a computer readable signal medium may include a propagated data signal with computer readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, fiber optic cable, etc., or any suitable combination of the foregoing.
Computer program code for carrying out operations for aspects of the present application may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, smalltalk, C + +, and conventional procedural programming languages, such as the C language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network or a wide area network, or the connection may be made to an external computer (for example, through the Internet using an Internet service provider).
The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
The terms "first," "second," and the like are used for distinguishing between similar elements and not necessarily for describing or implying a particular order or sequence.
The terms "comprises," "comprising," or any other similar term are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
So far, the technical solutions of the present invention have been described in connection with the preferred embodiments shown in the drawings, but it is easily understood by those skilled in the art that the scope of the present invention is obviously not limited to these specific embodiments. Equivalent changes or substitutions of related technical features can be made by those skilled in the art without departing from the principle of the invention, and the technical scheme after the changes or substitutions can fall into the protection scope of the invention.
Claims (9)
1. A user personalized information recommendation method based on target object enhanced representation is characterized by comprising the following steps:
s10, acquiring the articles in the user history record and the target articles to be recommended, and mapping the articles and the target articles to be recommended into a feature vector with set dimensionality; taking the feature vector mapped by the object in the user history record as a first feature vector, and taking the feature vector mapped by the target object to be recommended as a second feature vector;
s20, calculating the similarity of the first feature vector and the second feature vector, and carrying out normalization processing on the similarity; carrying out weighted summation on each first feature vector through the normalized similarity, and taking the weighted summation as an interest expression of the user about the target object;
s30, acquiring an intersection of the first data set and the target object to be recommended; sampling a set number of articles from the intersection by a preset sampling method, and mapping the articles into a feature vector as a third feature vector; the first data set is a data set constructed based on items of a plurality of user history records;
s40, calculating the similarity of the second feature vector and the third feature vector, and normalizing the similarity; carrying out weighted summation on each third feature vector through the normalized similarity to serve as an enhanced representation about the target object;
s50, splicing the second feature vector, the interest representation of the user about the target object and the enhanced representation about the target object, inputting the spliced second feature vector into a pre-constructed classifier, and predicting the interest degree of the user to the target object to be recommended; sorting according to the interest degrees, and recommending the target articles to be recommended corresponding to the first N interest degrees to the user after sorting; n is a positive integer.
2. The method for recommending user personalized information based on target item enhanced representation according to claim 1, wherein in step S20, "calculating the similarity of the first feature vector and the second feature vector, and performing normalization processing on each similarity", the method comprises:
calculating the similarity of the first feature vector and the second feature vector through a multilayer perceptron;
and (5) normalizing each similarity by using a softmax function.
3. The method for recommending user personalized information based on enhanced representation of target item according to claim 1, wherein "a set number of items are sampled from the intersection by a preset sampling method" in step S30, the method is as follows:
and sampling a set number of articles from the intersection by any one of a global sampling method, a local sampling method, a weighted sampling method and a uniform sampling method.
4. The method for recommending user personalized information based on enhanced representation of target item according to claim 3, wherein said weighted sampling method is:
wherein the content of the first and second substances,
representing the probability that item k is sampled under the global sampling method,
representing the probability that item k is sampled under the local sampling method, alpha represents the weight,
indicating the number of occurrences of item k in the first data set,
represents an item k andnumber of co-occurrences of target items to be recommended, o i Representing a set of item i and item constructions in each user history record of the first data set other than item i,
indicating the number of occurrences of item j in the first data set,
indicating that item j is at o i The number of occurrences in (c).
5. The method for recommending user personalized information based on enhanced representation of target item according to claim 4, wherein said uniform sampling method is:
6. the method for recommending user personalized information based on enhanced representation of target item according to claim 1, wherein said classifier is constructed based on multi-layer perceptron; the multilayer perceptron is three layers, reLu activation functions are adopted among the layers, and the dimensions of the layers are 80, 40 and 1 in sequence.
7. A system for recommending user personalized information based on enhanced representation of target object, the system comprising: the system comprises a characteristic mapping module, an interest representation acquisition module, a common article sampling module, an enhanced representation acquisition module and a target article recommendation module;
the characteristic mapping module is configured to acquire articles in the user history record and target articles to be recommended and map the articles and the target articles into a characteristic vector with set dimensionality; taking the feature vector mapped by the object in the user history record as a first feature vector, and taking the feature vector mapped by the target object to be recommended as a second feature vector;
the interest representation acquisition module is configured to calculate the similarity of the first feature vector and the second feature vector and carry out normalization processing on the similarity; carrying out weighted summation on each first feature vector through the normalized similarity, and taking the weighted summation as an interest expression of the user about the target object;
the common article sampling module is configured to acquire an intersection of the first data set and the target article to be recommended; sampling a set number of articles from the intersection by a preset sampling method, and mapping the articles into a feature vector as a third feature vector; the first data set is a data set constructed based on items of a plurality of user history records;
the enhancement representation acquisition module is configured to calculate the similarity of the second feature vector and the third feature vector and carry out normalization processing on the similarity; carrying out weighted summation on each third feature vector through the normalized similarity to serve as an enhanced representation about the target object;
the target article recommending module is configured to input the spliced second feature vector, the interest representation of the user about the target article and the enhanced representation about the target article into a pre-constructed classifier, and predict the interest degree of the user to the target article to be recommended; sorting according to the interest degrees, and recommending the target articles to be recommended corresponding to the first N interest degrees to the user after sorting; n is a positive integer.
8. An apparatus, comprising:
at least one processor; and
a memory communicatively coupled to at least one of the processors; wherein the content of the first and second substances,
the memory stores instructions executable by the processor to implement the method of recommending user-customized information based on enhanced representations of target items according to any one of claims 1-6.
9. A computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions for execution by the computer to implement the method for recommending user-customized information based on enhanced representation of target item according to any one of claims 1-6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110498063.4A CN113220994B (en) | 2021-05-08 | 2021-05-08 | User personalized information recommendation method based on target object enhanced representation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110498063.4A CN113220994B (en) | 2021-05-08 | 2021-05-08 | User personalized information recommendation method based on target object enhanced representation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113220994A CN113220994A (en) | 2021-08-06 |
| CN113220994B true CN113220994B (en) | 2022-10-28 |
Family
ID=77091638
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110498063.4A Active CN113220994B (en) | 2021-05-08 | 2021-05-08 | User personalized information recommendation method based on target object enhanced representation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113220994B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106776881A (en) * | 2016-11-28 | 2017-05-31 | 中国科学院软件研究所 | A kind of realm information commending system and method based on microblog |
| CN107967320A (en) * | 2017-11-23 | 2018-04-27 | 南京邮电大学 | A kind of matrix decomposition project recommendation algorithm of user's social status enhancing |
| CN110110094A (en) * | 2019-04-22 | 2019-08-09 | 华侨大学 | Across a network personage's correlating method based on social networks knowledge mapping |
| CN111310056A (en) * | 2020-03-11 | 2020-06-19 | 腾讯科技（深圳）有限公司 | Information recommendation method, device, equipment and storage medium based on artificial intelligence |
| CN112069408A (en) * | 2020-06-15 | 2020-12-11 | 北京理工大学 | Recommendation system and method for fusion relation extraction |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9336546B2 (en) * | 2014-03-27 | 2016-05-10 | Microsoft Technology Licensing, Llc | Recommendation system with multi-dimensional discovery experience |
-
2021
- 2021-05-08 CN CN202110498063.4A patent/CN113220994B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106776881A (en) * | 2016-11-28 | 2017-05-31 | 中国科学院软件研究所 | A kind of realm information commending system and method based on microblog |
| CN107967320A (en) * | 2017-11-23 | 2018-04-27 | 南京邮电大学 | A kind of matrix decomposition project recommendation algorithm of user's social status enhancing |
| CN110110094A (en) * | 2019-04-22 | 2019-08-09 | 华侨大学 | Across a network personage's correlating method based on social networks knowledge mapping |
| CN111310056A (en) * | 2020-03-11 | 2020-06-19 | 腾讯科技（深圳）有限公司 | Information recommendation method, device, equipment and storage medium based on artificial intelligence |
| CN112069408A (en) * | 2020-06-15 | 2020-12-11 | 北京理工大学 | Recommendation system and method for fusion relation extraction |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113220994A (en) | 2021-08-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Gabriel De Souza et al. | Contextual hybrid session-based news recommendation with recurrent neural networks | |
| Lau et al. | Parallel aspect‐oriented sentiment analysis for sales forecasting with big data | |
| US11436487B2 (en) | Joint embedding of corpus pairs for domain mapping | |
| WO2022041979A1 (en) | Information recommendation model training method and related device | |
| US10657189B2 (en) | Joint embedding of corpus pairs for domain mapping | |
| CN112231569B (en) | News recommendation method, device, computer equipment and storage medium | |
| US11276099B2 (en) | Multi-perceptual similarity detection and resolution | |
| CN112395487B (en) | Information recommendation method and device, computer readable storage medium and electronic equipment | |
| Maupomé et al. | Early Detection of Signs of Pathological Gambling, Self-Harm and Depression through Topic Extraction and Neural Networks. | |
| US10642919B2 (en) | Joint embedding of corpus pairs for domain mapping | |
| Kastrati et al. | Performance analysis of machine learning classifiers on improved concept vector space models | |
| CN111563158A (en) | Text sorting method, sorting device, server and computer-readable storage medium | |
| Yang et al. | Enhancing recommendation on extremely sparse data with blocks-coupled non-negative matrix factorization | |
| CN111429161B (en) | Feature extraction method, feature extraction device, storage medium and electronic equipment | |
| Ertekin et al. | Approximating the crowd | |
| US11361031B2 (en) | Dynamic linguistic assessment and measurement | |
| US20220366295A1 (en) | Pre-search content recommendations | |
| US20220277031A1 (en) | Guided exploration for conversational business intelligence | |
| Bhatnagar et al. | A novel aspect based framework for tourism sector with improvised aspect and opinion mining algorithm | |
| CN112784157A (en) | Training method of behavior prediction model, behavior prediction method, device and equipment | |
| Vielma et al. | Sentiment Analysis with Novel GRU based Deep Learning Networks | |
| CN113220994B (en) | User personalized information recommendation method based on target object enhanced representation | |
| US20230237093A1 (en) | Video recommender system by knowledge based multi-modal graph neural networks | |
| Ghosh et al. | Understanding Machine Learning | |
| US20200226159A1 (en) | System and method of generating reading lists |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |