TWI660315B - Sequence recommendation method - Google Patents

Abstract

The invention discloses a sequence recommendation method and a computer program product thereof. The method provided by the invention includes inputting usage history records of a plurality of users; using optimization methods to obtain user-to-item and article-to-item models respectively; and when needed When making a recommendation, the recommendation sequence can be obtained through the parameters of the model.

Description

   Sequence recommendation method   

The invention belongs to a sequence recommendation method. In order to obtain the user-to-item and article-to-item models by using an optimization method, and when a recommendation is needed, a recommended sequence can be obtained through the parameters of the model.

In recent years, due to the popularity of the Internet and the development of its infrastructure and related technologies, people have begun to complete many business activities through the Internet, such as online shopping and online auctions. Therefore, "recommendation systems" have become a very important and high-profile technology. With a large number of browsing history and transaction history of past users, we can think of ways to "predict" users' satisfaction and preference for a product or service, and then "recommend" products that users may be interested in , And this is usually the goal of the recommendation system.

Many online shopping platforms today use recommendation systems, and there have been many patented inventions and studies focusing on this. Most of these inventions have focused on recommending "single item as an entity". However, in some application scenarios, the object to be recommended should be "sequence", that is, "a series of ordered items". In this regard, the travel itinerary recommendation is an example. A travel itinerary should consist of a sequence of places; in addition, recommendations such as a music playlist are also examples of “sequence recommendations”.

Of course, for this kind of "sequence recommendation" problem, we can simply use the known technique of recommending a single item to make a recommendation, but if applied directly, it will directly face the problem of excessive complexity caused by the dimensional improvement. In addition, serial items have unique characteristics related to each item before and after. The present invention proposes an original probability model design, which is expected to solve this problem.

In fact, due to better performance evaluation, the recommendation problem in general is mainly the second type (predicted score) listed above. In this regard, the predecessors have proposed many methods that can be well solved, such as Matrix Factorization, Restricted Boltzmann Machine, and other Collaborative Filtering methods. However, it is rarely addressed specifically to the object of the present invention: the "generalized personalized sequence recommendation problem". For the "sequence recommendation problem", some studies have tried to use some rule-based or heuristic methods to solve this special recommendation problem.

For the problem of sequence recommendation system, Baccigalupo et al. [1] proposed a case-based method for music playlist recommendation. The method is as follows: first select some “observed” ones according to the case basis Sequences (sequences in the training data), and then recombining them into a sequence that is finally recommended to the user in a manner consistent with the sequence characteristics. Ge et al. [2] aimed at the itinerary recommendation problem, and pointed out that it is different from the traditional recommendation problem, but the focus is on the factors of money and time that are unique to the journey recommendation. Therefore, they designed and considered It takes a latent factor model to complete the sequence recommendation. Regarding the composition of the recommended results sequence, Ziegler et al. [3] proposed the topic diversification, and designed a new method to balance and make the recommended sequence changeable but not monotonous. In the above system, the common point is that the entire sequence is judged as a single entity, without considering each single element.

In view of the various shortcomings derived from the above-mentioned conventional methods, the inventor of this case has been eager to improve and innovate. After years of painstaking research, he finally successfully developed the method recommended in this sequence.

In order to achieve the above object, the present invention proposes to provide a sequence recommendation method, which uses two different matrices to solve the problem of personalized sequence recommendation. The personalization matrix records the number of purchases or preferences between users and items, and the item history matrix. It records the relationship between the last purchase of goods and this purchase. Using the combination of the two models, serial recommendation information can be obtained. Both models can be converted into general optimization problems and quickly calculated.

A sequence recommendation method includes: using system records collected by an online system to generate training data; using optimization algorithms to learn model parameters based on the training data; evaluating the performance of the current model parameters on the data set; whether the parameters are Optimization; if not, then use the optimization algorithm to learn the model parameters based on the training data; if so, establish an actual application system to complete the recommendation of the personalized sequence.

The collected system records are used to collect user sequence records. If there are no sequence records, the time information in the records is used to convert the records into sequences according to the chronological order of occurrence. The records are divided into training sets and verification sets. Is to train user preference model and item-to-item order model respectively, and store them in matrix.

To sum up, firstly, use the historical sequence records of each user recorded in the online system to generate sequence training data with personalized elements to collect user sequence records. If there is no sequence record, use the Time information, the records are converted into sequences according to the chronological order of occurrence, and the records are divided into training sets and verification sets. Second, the sequence training data with personalized elements is used to build a target machine learning model and collect the collected data. The training set of the user sequence record separately trains the user's preference model and item-to-item order model for the item. The second model is stored in a matrix, and then the machine learning model is used as the training set or verification set. For performance evaluation, it uses a machine learning model to make sequence recommendations, and uses a validation set to evaluate the results of this recommendation using the degree of divergence or inverse order method. If the convergence conditions are not met, return to the second step to continue training. If the convergence conditions are met, the model is stored. The convergence conditions can be relative or absolute conditions. Complete model to establish the actual sequence of personalized recommendation program, the convergence of user preference model of food items and food items to food items sequential model as radiation matrix and transfer matrix using dynamic programming algorithm can obtain the best recommendation sequence.

The sequence recommendation method provided by the present invention has the following advantages when compared with other conventional technologies:

1. Based on the training materials, automatically learn the way to evaluate the recommended effect.

2. According to the aforementioned evaluation method, establish a sequence recommendation model with a lower complexity method to generate the best sequence recommendation.

S110 ~ S150‧‧‧Recommended process

Please refer to the detailed description of the present invention and its accompanying drawings to further understand the technical content of the present invention and its purpose and effect. The related drawings are as follows: FIG. 1 is a flowchart of the method of the sequence recommendation of the present invention.

In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not intended to limit the present invention.

Hereinafter, the present invention will be further described with reference to the accompanying drawings. Please refer to FIG. 1, a sequence recommendation method includes: S110 uses system records collected by an online system to generate training data; S120 uses an optimization algorithm based on the training data Learning model parameters; S130 evaluates the performance of the current model parameters on the data set; whether the S140 parameters are optimized; if not, returns to S120 to learn the model parameters using the optimization algorithm based on the training data; if yes, then S150 is established Practical application system to complete the recommendation of personalized sequence.

The collected system records are used to collect user sequence records. If there are no sequence records, the time information in the records is used to convert the records into sequences according to the chronological order of occurrence. The records are divided into training sets and verification sets. , Is to train the user preference model and item-to-item order model respectively, and store them in a matrix.

In accordance with the above steps, reference may be made to the following embodiments at the same time.

Using the historical sequence records of each user recorded in the online system to generate sequence training data with personalized elements. The sample format is provided as follows: user-id element-0 element-1 element-2 ...

Where user-id is the user number, and element-i is the number of the i- th element in the sequence.

Use the generated sequence training data to build a target machine learning model. The present invention models the entire personalized sequence recommendation problem as an optimized mathematical problem:

Where L is the likelihood function value of the entire data set D when the model parameter is Theta (θ).

Under the first-order Markov assumption, the entire expression can be rewritten as:

Where N is the total number of sequences, and X _{i, k} i represents the sequence of the k-th element, U _i is the i th user represented by the sequence, _n-i-i is the total items in the user history.

For the optimization of this goal, the present invention designs the conditional probability to be constructed by a potential function r composed of (a) a personalized recommendation and (b) a sequence characteristic. The design of the overall conditional probability corresponding to the potential energy function is:

In the above equation, x represents an element extracted from the aggregate of x _{i, k} , u represents a user extracted from the aggregate of u _i , a represents the goods or services that the user last purchased in the record, and b represents the record in the record. The goods or services purchased by the user this time.

For this potential energy function r , a specific embodiment is constructed with two sets of latent matrices, which are used to describe the above two characteristics of (a) personalized recommendations and (b) sequence characteristics. . The example method is as follows:

The four matrices p, q, w, and h are the parameters of the model, Theta (θ); and alpha (α) is a constant determined in advance, and its value must be between 0 and 1 for control ( a), (b) The ratio of the two parts. Among them, p, q, and h are the ratings of goods and services for certain preference dimensions. Taking restaurants as an example, the preference dimensions may be whether the price is low, the environment is clean, whether parking is good, whether there is accessible space, whether the meal is delicious, etc. Wait. Because different types of goods and services are so different that it is impossible to do it in a way that lists all dimensions, it needs to be constructed in a machine learning manner. w is how much the user cares about the preference dimension. Taking the restaurant as an example, some users may be less concerned about whether there is accessibility space, but very concerned about the cleanliness of the environment. Another user may not care about the price and only care about the deliciousness. w This degree of preference must also be obtained through user learning through machine learning.

In S120, it is a process of learning model parameters (such as p, q, w, h in a specific embodiment described above). Specific learning methods can use but are not limited to SGD (Stochastic Gradient Descent) optimization method. With the optimization algorithm of machine learning, we can get a set of local optimal model parameters.

Use the learned model parameters to evaluate the performance of each data set (training data, verification data). This case exemplifies two specific embodiments. The first specific embodiment is a value of Perplexity, the formula of which is as follows:

D data set is the data set, | D | is the total number of records in the data set, and s is one of the data.

This evaluation method can be used to know the performance of the entire system. Its mathematical meaning is average branching factor, which can be used as a criterion for this objective problem. The lower the value of the Perplexity, the better the effect of model learning. Another specific embodiment is Mean Reciprocal Rank. This performance measure can evaluate whether the model recommends the sequences that have been observed (that is, the sequences that appear in the test data set) to be preferred. order. In this question, the formula can be expanded as follows:

Where gamma ( i ) is the index of the element queried for the sequence i in the data set D.

To sum up, in this step P300, the performance of the current model parameter Theta (θ) on the actual data set is evaluated, and the evaluation result is used to determine whether to continue training and update the parameters (that is, return to step P200). If the effectiveness shows that it has reached the best performance in the local area, it will enter the complete personalized sequence recommendation system.

Use the trained model to learn and evaluate the completed model to establish an actual personalized sequence recommendation procedure. In this regard, an embodiment is: the system input: user ID (uid), the length of the sequence, the beginning element (element-0) of the requested sequence; the system output: the recommended sequence (element-1 to element-K) The way to actually generate a sequence can use but is not limited to one element in the sequence and one element is generated sequentially, which means that we can use heuristics based on the conditional probability constructed by the learned potential energy function. The method generates a sequence with a high probability (for the information entered by the user); practical application examples: for recommending music lists (elements = tracks), for recommending travel itineraries (elements = landscapes).

The above detailed description is a specific description of a feasible embodiment of the present invention, but this embodiment is not intended to limit the patent scope of the present invention. Any equivalent implementation or change that does not depart from the technical spirit of the present invention should be included in Within the scope of the patent in this case.

To sum up, this case is not only innovative in terms of technical ideas, but also has many of the above-mentioned effects that are not used by traditional methods. It has fully met the requirements of statutory invention patents that are novel and progressive. To approve this invention patent application, to encourage invention, to the utmost convenience.

Claims (6)

A sequence recommendation method includes the following steps: a computer generates training data based on system records collected by an online system; the computer learns model parameters using an optimization algorithm of machine learning based on the training data; and the computer evaluates The performance of the model parameters on the data set; the computer judges whether the model parameters are optimized; if not, the computer returns to the step of learning the model parameters based on the training data and using the optimization algorithm ; If so, the computer establishes an actual application system to complete the recommendation of the personalized sequence; wherein the computer obtains the sequence recommendation information by a combination of a personalization matrix and an item history matrix, and the personalization matrix records the user and the item Number of purchases or preferences, and the item history matrix records the relationship between the last purchased product and this purchased product. For example, the method of sequence recommendation described in item 1 of the scope of patent application, wherein the collected system records are to collect user sequence records. If there is no sequence record, the computer uses the time information in the record to record the record in accordance with The time sequence of occurrences is converted into a sequence, and the records are divided into a training set and a validation set. The method of sequence recommendation as described in item 2 of the scope of the patent application, wherein the training set is to separately train the user preference model and item-to-item order model for the user and store them in a matrix. A sequence recommendation method includes the following steps: generating training data based on system records collected by an online system; using the training data to learn model parameters using an optimization algorithm of machine learning; and evaluating the model parameters on the data set. Performance; judge whether the model parameters are optimized; if not, return to the step of learning the model parameters based on the training data and using the optimization algorithm; if so, establish an actual application system and complete the personalized sequence Recommendation; among which, the sequence recommendation information is obtained by a combination of a personalization matrix and an item history matrix, the personalization matrix records the number of purchases or preferences between the user and the item, and the item history matrix records the last purchased product and the current purchased product Relationship. The method of sequence recommendation as described in item 4 of the scope of patent application, wherein the collected system records are to collect user sequence records. If there is no sequence record, the time information in the record is used to record the record according to the time of occurrence. The sequence is converted into a sequence, and the records are divided into a training set and a validation set. The method of sequence recommendation as described in item 5 of the scope of the patent application, wherein the training set is to train a user-item preference model and an item-to-item order model separately and store them in a matrix manner.