Research On The Recommendation Algorithms Based On Spatial And Temporal Data Characteristics

Many recommendation application scenarios will generate time-series data,or data containing spatial location coordinates,and These data are closely related to their service types.One example is in the scenario of travel websites.Its products are travel destinations(usually with services nowadays).Travel destinations have inherent spatial attributes,that refer to,latitude and longitude here.Before traveling,users will give priority to the distances and use them as search criteria to find the best targets on the website,for example:outside the province or in the suburbs?Domestic or go abroad?And the travel histories of tourists will also form trajectory logs with distance attributes,which became a unique data source.Another example is e-commerce scenario,a large number of time-ordered browsing and consumption logs will be generated on e-commerce websites.When consumers shop online,the recommendation system starts to work.It is necessary to predict the products that users will be interested in within a short period of time,and users' preferences are also reflected in the current browsing sequence.In the historical data of these two fields,the time series and distance are not the inherited attributes of the item itself,but are formed dynamically along with user activities,while the traditional recommendation field focuses on the profile mining of users and items by considering the additional information such as labels and texts.These two features are quite less integrated into the algorithm and usually not used effectively.However,in such real economy fields,spatio or temporal information is closely related to product services' quality.Therefore,in the problem of using the history of user-item interaction to make recommendations,it is important to design a recommendation algorithm that better suits the product based on the characteristics of spatio or temporal data.In the course of our research,we happened to encounter two unique scenarios:Ctrip Travel Website and Hisense Smart TV.In view of the above problems and the specific characteristics,we proposed new recommendation algorithms separately.For travel recommendation,although tourists can currently use web search engines to select tourist attractions,it is often difficult to obtain a travel plan that better meets travelers' needs.The backend recommender system is an effective way to solve the above problems.A good travel recommendation model should be personalized and take into account the user's filtering conditions,such as time and distance constraints and cost restrictions.Our research shows that when users choose a tourist attraction,the distance between the destination and the user's habitual residence is often a first issue to be considered,and the travel distance often indirectly reflects the impact of time and cost.For the website's aspect,the distance or city area is usually used as the initial condition in the retrieval tool,and then the recommendation system starts to work.It can be seen that the distance factor is so important in the field of travel recommendation.Therefore,in the part of travel recommendation,we first assume that each user has a most preferred travel distance according to the analysis of the real Ctrip travel logs.then based on the Bayesian model and the Probability Matrix Decomposition model,we propose a travel recommendation model that combines the characteristic of travel distance.It is abbreviated as G-PMF(Probabilistic Matrix Factorization with Geographical Distance Feature).The innovation of the model lies in:(1)We map each user's distance preference as a weight,and combine it to the Probability Matrix Factorization.(2)By using data discretization technology,the distance between the scenic spot and the user's usual residence is divided into sections,which are introduced into the learning process as additional information.(3)Experiments on the travel data set of Ctrip website show that compared with the benchmark method,the root mean square error(RMSE)of G-PMF has been significantly reduced.For Smart TV scenario,which enables viewers to watch both live TVs and on-demand network contents on a single platform,is becoming increasingly popular.The study of sequence prediction problems is described as the recommendation engine provides a list of videos that the user will probably see in the next when the user finishes the current one.In the process of our scientific research road,we have obtained Hisense users' viewing logs and analyzed them to solve the problem of sequence prediction.In other related issues such as e-commerce session-based recommendation,recurrent neural network(RNN)has been used to capture sequential patterns,and the deep sequential recommendation algorithm has achieved better recommendation results than traditional models.However,migrating this method to a smart TV solution directly is feasible but not effective because the smart TV dataset has different characteristics from other Web datasets,such as account sharing,sparseness caused by time span,and the discontinuities of user viewing behavior and more.And as far as the algorithm itself is concerned,RNNs cannot capture the uniqueness of user interest.Therefore,we have developed a hybrid method here to capture sequence patterns and combine the traditional collaborative filtering to mine unique personalities of sequences,which we call the Hybrid Sequence Prediction Model(HSPM).It contains two modules.The first one uses the Parallel Gated Recurrent Unit(GRU)model to find general sequence representations using clicks and post images information.The second module is called the TCSKNN(Time Context Sequence K Nearest Neighbor),which takes into account that the user's choice in the current time range is usually affected by the TV promotion page,so the traditional KNN is extended to a time-sensitive method to simulate online real-time recommendation.Finally,we combine the two modules' prediction scores and mix them in a weighted manner.Through experiments on a Hisense Smart TV on-demand data set and Youtube dateset,we prove that the model can achieve a certain improvement on the HR(hit rate)and MRR(reciprocal rank of the mean)indicators compared with the latest baseline model.In summary,we have performed data mining,problem modeling,and experimental analysis on travel destination recommendation scenarios with spatial distance features and smart TV VOD sequence prediction scenarios with time-series features.For the model,two new recommendation algorithms for each characteristic are proposed separately,and they combined traditional collaborative filtering or deep learning independently.The validity of the models are verified on large-scale real-word data sets.The works done are of great help to the fields of smart TV and travel recommendation.