Topic recommendation system using personalized fuzzy logic interest set

1 Introduction

As a special resource type, intelligent theme logic interest set resources have stronger specificity and particularity than other types of resources, and users of intelligent theme logic interest set resources are also quite different from other network service users. In recent years, various intelligent topic logical interest collection resource platforms have been built, which contain a large number of intelligent topic logical interest collection resources, but the means for users to obtain the required intelligent topic logical interest collection resources is limited to finding their own resources through search keywords Required resources. This method is not only inefficient, but also not conducive to discovering the potential interests of users. It makes users only use the intelligent theme logic interest collection resource platform when they need it, and a large amount of resources cannot be fully utilized.

Recommendation based on collaborative filtering assumes that users have similar preferences in the past, and they can be considered to have similar preferences in the future, so users can be considered to be implicitly collaborating with others [1–3]. Recommendations based on collaborative filtering can be divided into memory-based collaborative filtering and model-based collaborative filtering. Memory-based collaborative filtering is also called memory-based collaborative filtering. Memory-based collaborative filtering mainly includes collaborative filtering based on neighbors, which aims to calculate the similarity between users and interest sets and use the similarity to calculate the user’s score prediction value for the interest set for recommendation [4, 5]. The calculation methods of similarity mainly include Euclidean similarity, cosine similarity and Pearson similarity. Model-based collaborative filtering mainly includes collaborative filtering based on probabilistic methods and collaborative filtering based on matrix factorization [6]. Content-based recommendation can alleviate the problem of data sparsity to a certain extent, and has a strong explanatory nature [7, 8]. The processing of text content information can use topic model modeling methods to map text information to a low-dimensional topic space. The advantage of content-based recommendation over collaborative filtering-based recommendation is that it does not require large-scale users. However, the disadvantage of content-based recommendation is that it is difficult to obtain feature information, and it is difficult to implement novel recommendations [9, 10]. Knowledge-based recommendation uses knowledge information of users, interest sets and relationships between them to make recommendations [11, 12]. It is a method of reasoning based on specific domain knowledge rules or examples, and can be divided into constraint-based recommendation and instance-based recommendation. The difficulty of knowledge-based recommendation is that reasoning rules are difficult to obtain, and because the system requires additional causal knowledge and these causalities are usually provided by humans, it increases the burden on the system [13, 14]. In addition, it is usually necessary to obtain user needs through interactive guidance. With the continuous development of recommendation algorithms, new recommendation methods continue to be introduced. Recommendations based on social networks [15, 16], recommendations based on context information [17, 18]. Recommendations based on social networks mainly use information such as social relationships, trust relationships, and graph structures for personalized recommendations [19]; recommendations based on contextual information mainly use information such as time, geographic location, surrounding people, emotional factors, and activities for personalized recommendations. [20]. Literature [21, 22] is a recommendation method for education recommendation based on topic models. This method introduces machine learning methods to explore the potential topics of resources, which can effectively compress the number of topics for a type of resources and discover the true topics of resources, and effectively improve the computational efficiency of the algorithm and improve the push accuracy [23, 24]. Literature [25, 26] collected user data through a specially designed data sampling module to establish four databases of users, scientific papers, user relationship networks, and user ratings, and then calculated the degree of social relevance between users, Divide. After determining the neighbors of the target user, the system will calculate the community where the user is located, and then calculate the item recommendation list based on the user’s rating data [27, 28]. By using the concept of ontology to describe the resources of the logical interest set of intelligent topics, semantic recommendation is used to improve the problem of polysemy and multi-word single meaning [29, 30]. Use the Markov chain to establish the feature model of the intelligent topic logical interest set resource, obtain the feature points of the intelligent topic logical interest set resource through the TF-IDF algorithm, and achieve certain results in the optimization of the intelligent topic logical interest set resource model [31]. Use particle swarm optimization algorithm to design a recommendation method for intelligent topic logic interest set resources [32].

Research on the characteristics of personalized fuzzy logic interest set and user behavior characteristics of the existing personalized fuzzy logic interest set platform, establish a personalized fuzzy logic model for personalized fuzzy logic interest set, and adapt it The user interest model is modeled, and a personalized recommendation scheme based on the user’s personalized fuzzy logic interest model is proposed based on the resource and user model. The five-tuple formal model of the intelligent topic recommendation system model is proposed, and the conceptual framework of the intelligent topic recommendation system model based on domain ontology and data mining is given. The logical structure and hierarchical structure of the system are deeply analyzed, and imagine key technologies.

2 Personalized fuzzy logic interest set model

The intelligent recommendation system oriented to personalized fuzzy logic interest set is an important part of the personalized fuzzy logic interest set system. From the formal point of view, it is an information service software system; from the content point of view, it is based on information resource organization on the basis of processing, it contains information resource databases for specific applications or specific professional fields. The intelligent recommendation system oriented to personalized fuzzy logic interest sets needs to solve or make up for the “lack of personalized user service” and “lack of semantics in information retrieval matching” existing in traditional information retrieval technologies and information retrieval systems. From the perspective of the depth of information interaction, it should process information resources based on the level of knowledge content, that is, the expression and organization of knowledge units are converted from physical-level literature units to cognitive-level knowledge units, and the processing of knowledge units from simple vocabulary levels Transform to the conceptual semantic level; from the perspective of system functions and service modes, it emphasizes user-centered, based on the recommended retrieval interaction mode, to provide personalized and active information and knowledge resource services.

Figure 1 outlines the general process for knowledge workers to personalize fuzzy logic interest sets. From the perspective of the general process of personalized fuzzy logic interest set by knowledge workers, knowledge workers screen, sort, and refine various types of information resources, and finally reflect the value-added knowledge and services by providing users with interest set products Value-added. Among them, the basic work of screening and sorting is the description of external attributes and the indexing of content characteristics of information resources, such as classification indexing, subject indexing. while extraction is the learning and understanding of knowledge workers themselves. Therefore, the knowledge products that are finally reflected can be coded tangible products such as different carriers, different forms of information resource databases, information service software systems, reference books, research reports, etc., or they can be formed by knowledge workers through knowledge extraction and learning The intangible collection of intangible interests such as his wisdom, abilities and skills.

OPEN IN VIEWER

Fig. 1

General process of personalized fuzzy logic interest set.

2.1 User personalized logical interest model

After clarifying the personalized fuzzy logic of the fuzzy logic interest set and how to clearly encode the personalized fuzzy logic, the user’s interest model for personalized fuzzy logic can be established. Firstly, several key elements in the recommendation system are defined:

T ={ T₁, T₂ . . . T _n  }: user interest set for personalized fuzzy logic;

UT (t _i ) ={ w (u _i , t _i ) }: user is interested in personalized fuzzy logic set value. In the designed recommendation system, the user model will be established with personalized fuzzy logic interest set and user rating set. Through the user’s personalized fuzzy logic interest set, we can understand the user’s degree of interest in each personalized fuzzy logic at the current time, and it can also be said that the user is currently learning which personalized fuzzy logic. Through the user rating set, you can understand the user’s evaluation of specific resources. In the subsequent calculations, these two data sets are used as the core for recommendation calculations. Therefore, when a user no longer has any user behavior on a resource to which a personalized fuzzy logic belongs after a certain time interval, we judge that the user’s learning process for the resource has ended, and the interest value of the personalized fuzzy logic will follow. w ( u i , t i ) = 1 1 + e k * t - t 0 T(1)

T is the time interval for the user’s personalized fuzzy logic interest set to transfer, this interval is the time from the user’s interest in a personalized fuzzy logic to the completion of the personalized fuzzy logic learning; t is the current time; t. It is the last update time when the personalized fuzzy logic interest value is changed by the user’s behavior; the unit of T and t are both days, and the value of k is an integer. If the subsequent personalized fuzzy logic interest value of a personalized fuzzy logic is 0 when the transfer occurs, then after assigning the personalized fuzzy logic interest value, the subsequent personalized fuzzy logic is included in the personalized fuzzy logic interest concentration The last update time of is set to the current time, and the update time of its pre-personalized fuzzy logic remains unchanged.

Figure 2 shows the transfer process of personalized fuzzy logic interest value, where personalized fuzzy logic I1, I2, I3 are three consecutive personalized fuzzy logics, and the interest value of personalized fuzzy logic I1 is assigned a value of 0.5. The user does not perform any operations on the resources under these personalized fuzzy logics, and time 0 is the first time interval after the user does not update I1 interests. We can observe: The first time interval: I2 obtains one third of the interest value of personalized fuzzy logic I2 in the first transfer of I1. The second time interval: I2’s interest value is strengthened, and I1 and I2 have similar interest values much lower than I3. The third time interval: I2 has a similarly lower level of interest value, while I1’s interest value has dropped to 1% before the transfer was not triggered.

OPEN IN VIEWER

Fig. 2

Personalized fuzzy logic interest set transfer.

As shown in Fig. 3, a relatively complete logical architecture of the intelligent topic recommendation system is given from the perspective of structural elements and processing flow. In this framework, the relationship between users, resources, and domain knowledge models can be better reflected, forming a dynamic, cyclical interactive processing process. Domain knowledge ontology, resource description and concept indexing, user retrieval and questioning, browsing and downloading, user access historical data mining and analysis, interest resource recommendation and collaborative recommendation constitute the main processing links of the intelligent topic recommendation system. In this cyclic interaction process, each processing link can call each other and benefit each other, which greatly improves the friendliness of the interaction between users and the system.

OPEN IN VIEWER

Fig. 3

Hierarchical structure of the intelligent theme recommendation system.

From the recommendation function and recommendation method, the recommendation in the intelligent topic recommendation system can be in multiple aspects. The document-based resource objects can be roughly divided into the following four categories: One is hot resource recommendation based on statistical analysis of historical access data (Resource objects with high user recent visits), classic resource recommendation (high-visit resource objects since the system has been running for a long time), etc.; the second is based on bibliometric data, such as the author, academic value, citation and Resource recommendations based on citations. This type of recommendation has been used in actual retrieval systems such as Wanfang, Weipu and CNKI; the third is content-based resource recommendation, which mainly includes the similarity when users browse a certain resource object Resource recommendation: After discovering the user’s interest, it recommends related resources for a certain interest of the user; the fourth is the collaborative recommendation based on group users, for example, after user similarity calculation or user clustering, the hot resources visited by similar user groups are compared with current users Recommendations that have not been visited to the current user can also be recommendations based on association rules based on the access relevance of the resource object.

Text similarity measurement is a relatively complex process. Although there are certain differences in the specific methods of similarity measurement, the entire processing flow from the generation of the dictionary to the final similarity measurement is basically the same, which can be roughly decomposed into the following five steps. As shown in Fig. 4:

OPEN IN VIEWER

Fig. 4

General process of text similarity measurement.

The first step: the generation of the dictionary, that is, a collection of words is obtained for processing the text. Dictionary is the foundation and key of text similarity measurement. It directly restricts the generation of feature items in text vectors and affects the accuracy of similarity measurement.

The second step: document preprocessing, that is, using automatic word segmentation technology or related methods to segment a complete text into words, separated by marks to form a sequence of word strings to generate the initial text vocabulary feature vector.

Step 3: Eliminate stop words and function words. The text feature vector obtained in the previous step is a high-dimensional sparse vector. There may be some feature items that have no effect on expressing the content of the text, such as particles, interjections, and function words that appear frequently. Therefore, the initial text vector Perform secondary processing to eliminate the above-mentioned useless words to reduce the dimensionality of the text vector and improve the accuracy of the text vector expression. The most common processing method is to generate a banned word list, and filter out the words appearing in the banned word list from the feature items of the text vector.

Step 4: Determine the weight of feature items. The contribution of each feature item to the content of the expressed text is different, and a certain weight should be assigned to it according to the contribution of the feature item to generate a weighted text feature vector. There are many methods for processing the weights of feature items, either by manual rules or by some statistical calculation methods. M ik = rf ik * itf k(3)

Step 5: Calculate text similarity. First, the text feature vector is normalized. Generally, the weight of each feature item in the text vector can be divided by the quotient of the maximum weight of the feature item in the vector as the normalized feature item weight; second, obtain two The feature item union of the vector, the two text feature vectors obtained are processed into a vector with the same feature item set, and the weight of the new feature item in the vector is recorded as zero.

The quality of the user interest model is directly related to the quality of the personalized recommendation service. Only when the user’s interests, preferences, and access patterns can be “understood” by the system, can the ideal personalized recommendation service be realized. User interest model is not only an accurate description of user interest. As a personalized intelligent recommendation service system based on a computer platform, computability is its basic requirement for user interest model. In other words, the user interest model is not a general simple description of individual users, but an algorithm-oriented, formal user interest description with a specific data structure. User interest content must be based on the user interest extraction method. The collection of user interest information is a process of obtaining data information related to user characteristics, preferences or activities. This process provides the necessary data source for the establishment of user interest model. The acquisition of user interest mainly includes two methods: explicit collection and implicit collection.

Bayesian classifier is a probabilistic method for classification. The system using this technology calculates the probability that the resource items browsed or visited by the user belong to a given category, and then classifies the resource items according to the probability to establish the user’s preference model for these resource items. Among them, for a given resource item I, if its attributes are probabilistically independent and the attribute values are known, the probability of belonging to class is calculated using the following Bayesian rule: P ( d j | T ) = p ( d j ) p ( T ) ∏ i = 1 k p ( b i | d j )(4)

In the recommendation system, the change of user interest is not only reflected in the content of interest, but also shows the decay of original interest and the generation of new interest over time, showing the phenomenon of interest drift. That is to say, in the process of the user interacting with the recommendation system, the search query questions given in different historical time periods and the content of the resource documents that have been browsed have different importance in the expression of the user’s current interests and preferences. Researchers have studied and explored the problems of user interest changes and interest drift. The typical way is to introduce the so-called window control method or forgetting function mechanism when calculating the user’s current interest.

The recommendation of interest resources based on the fine-grained interests of users should be more reliable and accurate, and the recommendation result list has better interpretability. It is generally believed that the coarse-grained user interest expression is a general and general list of interest points. In the user interest description, only two types of user interest feature expression methods are distinguished, as shown in Table 2. The fine-grained interest expression mainly distinguishes the user’s interest topic category in the user interest feature set, and further mines and divides the user’s coarse-grained interest, as shown in Table 3.

Interest set: Contains the trust relationship between users, so a directional trust network can be obtained. We randomly selected 1082 users, so we got a 1082×1082 binary asymmetric matrix. In the experiment, we use the leave-five-out strategy to construct the test set. Specifically, we selected all users who have at least 10 trust relationships; for each of these users, we took out 5 trusted users as positive examples, and randomly selected 45 untrusted users as negative examples. This constitutes the user’s test case.

Table 4 shows the AUC and standard deviation of all models on the test set. By virtue of heteroscedasticity selection modeling technology, BHCM has achieved significant performance improvement compared with other methods. This is because users always have different backgrounds and different interests, resulting in the heterogeneity of user preferences, and other methods are based on the assumption of independent and identical distribution, so such heterogeneity cannot be expressed. In contrast, because BHCM uses DP to adaptively group users and establish a specific distribution for each group to represent the potential characteristics of users, it can better capture the heterogeneity between users. In addition, we can find that BHCM has achieved better performance than BCM, which shows that different users have a certain degree of heterogeneity in the choices of different items, but the selection model based on the assumption of independent and identical distribution cannot well represent the choice. The heterogeneity between the two; BHCM is a heteroscedastic model based on the weak binary assumption, which is more reasonable than the traditional independent and identically distributed binary assumption model. In addition, we can find that the standard deviation of BHCM is also the smallest among all models, which shows that BHCM is a good representation of the heterogeneity between different users and different items, so these advantages make BHCM achieved in all comparison methods the best result.

Figure 5 shows the AUC of different groups. It can be found that BHCM achieves better and more stable results than other compared models. This is mainly caused by two reasons. On the one hand, BHCM is based on a weak binary assumption. It uses a heteroscedastic model to better represent the heterogeneity of each choice, while other models are established on the basis of the binary assumption of independent and identical distribution, it cannot model the uncertainty difference between positive and negative choices well, thus failing to accurately capture user preferences. On the other hand, the reason is that most users are tail users with less data. Under the assumption of independent and identical distribution, regularization makes the potential characteristics of all users shrink to a global value, which makes the heterogeneity of tail users Features cannot be represented well. In contrast, BHCM is a model based on the assumption of non-independent identical distribution. It uses DP and heteroscedasticity to better represent heterogeneity, so it achieves better and more stable performance than other models.

OPEN IN VIEWER

Fig. 5

AUC of all models is grouped by different numbers of trustees.

In order to improve the success rate of short-term prediction in the recommendation system, increasing the number of candidate recommendation items is an effective method. Assuming that a movie website currently has 5 movies in total, if only one movie is recommended to the user each time, the short-term prediction success rate will not be very high, but considering extreme cases, if 5 movies are recommended to the user each time, the short-term prediction success rate will become 1.

Figures 6 and 7 show that when the recommendation system uses the highest-scoring recommendation algorithm, as the number of recommended movies increases, the values of SPS and Recall can be seen to continuously increase, and the final short-term prediction success rate and recall rate are both 1. However, it can be easily seen from the figure that for the highest-scoring recommendation algorithm, if you want to obtain higher SPS and Recall, you need to recommend a large number of movies to users. When SPS and Recall are around 20%, you need to recommend movies. The number has exceeded 100, which is not practical for recommendation systems, and recommending a large number of movies to users is not personalized. What the research needs to do is to increase the SPS and Recall as much as possible under the minimum number of recommended movies.

OPEN IN VIEWER

Fig. 6

MovieLens 100K experimental diagram.

OPEN IN VIEWER

Fig. 7

MovieLens 1M experimental diagram.

For MovieLens 100K and MovieLens 1M, compare the SPS and Recall values of different models, as shown in Tables 5 and 6.

It can be seen from Fig. 8 that in the RMSE coefficient, the FUCF algorithm performs slightly worse when the recommended number is 3. The MUCF algorithm has always been low and relatively stable, so both FUCF and MUCF are better than UserCF in the error of predicting score algorithm, so the improved algorithm improves the accuracy of scoring.

OPEN IN VIEWER

Fig. 8

RMSE curve.

Experiments show that the fuzzy logic interest set personalized recommendation scheme proposed in this paper can reflect the user’s interest in other knowledge points according to the user’s behavior records. At the same time, as time goes by, the recommendation result will be slower through the transfer of the interest value of the knowledge points. Slowly tend to resources for follow-up knowledge points. In general, the fuzzy logic interest set personalized recommendation scheme proposed in this paper can recommend resources that users are really interested in, and can track the user’s learning process to a certain extent, and recommend resources backward along the learning chain of knowledge, the design of the algorithm basically achieves the expected purpose.

This paper proposes a fuzzy logic interest set personalized recommendation scheme based on the knowledge point interest model, which is different from the collaborative filtering basic algorithm that only predicts user ratings. The recommended results appear disorganized under the highly structured attributes of the fuzzy logic interest set, but based on knowledge personalized recommendation algorithm of the point interest model will recommend to users the resources they are really interested in. Through the transfer of user interest points, the system will recommend follow-up resources to the user along the user’s learning chain, without the user’s own tracking. The personalized recommendation technology has been researched and explored. The domain ontology is applied to fine-grained user interest modeling, and a fine-grained user interest discovery method based on user coarse interest vector and close neighbor concept aggregation is proposed. Experimental tests show that this method can better describe the composition and development of user interests, making the recommendation of interest resources for specific users more accurate and reliable. This paper studies the problem of collaborative resource recommendation in one step, and proposes a collaborative recommendation method based on users with similar interests and a collaborative recommendation method based on weighted association rules. The above collaborative recommendation method overcomes the high sparsity and computational scale of traditional recommendation methods to a certain extent. Problems such as large, poor recommendation effect, and have good applicability.