Keywords

1 Introduction

It is time-consuming for users to pick up a desirable travel itinerary on various travel websites. So, a recommendation system is helpful for them to generate personalized results and to save their time. Personalized recommenders in tourism can be mainly divided into three categories: (1) providing popular itineraries for a user without taking into consideration the user’s specific requirements (for instance, mafengwo.cnFootnote 1); (2) indicating all POIs (Points of Interests, or attractions) that match with user’s preference from which users have to choose and figure out a plan on their own (e.g. elong.comFootnote 2); and (3) accepting information on the origins, destinations, and travel date from users and in return, offering the driving routes and hotels (e.g. trippy.comFootnote 3). These systems are rarely seen to provide customized travel itineraries. Hence, we proposed a personalized itinerary recommender in [18] that leaves room for improvement. We enhanced that system by taking into account a few more factors that would potentially lead to a better itinerary. Initially, we allow for a flexible number of daily recommended POIs. Next, we improve our technique for selecting hotels and attractions so as to generate routes that involve shorter travel distance in total. Moreover, we consider the weather information and provide indoor POIs to cater for user needs for adverse weather. With all available routes, we can also provide users with information on the time that they would spend on commuting between sites.

A generic neural social collaborative ranking (NSCR) solution via the so-called “bridge” users (social media users who have accounts on two or three social networks) was proposed in [17], which seamlessly integrates user-item interactions of the travel domain and user-user social relations. Farseev et al. [5] proposed a recommendation framework C3R (cross-network collaborative recommendation framework) that utilizes both individual (user has visited in past) and group (Foursquare venue categories among user community members) knowledge to solve a task of venue category recommendation. Abel et al. [1] investigated the characteristic of tag-based profiles that resulted from tagging activities in social network websites such as Flickr, and cross-system user modeling strategies had significant impact on the performance of the recommendation quality within the scope of tag and resource recommendations in cold-start settings. The PERSTOUR algorithm was proposed in [9], which can reflect levels of user interest based on visit duration and demonstrate how POIs visit duration can be personalized using this time-based user interest.

The remainder of the paper is organized as follows. In Sect. 2, we briefly revisit our previous work and then examine the supporting techniques for our improved recommender. This is followed by Sect. 3 in which we introduce the method for working out improved travel routes that for users. Evaluation metrics and experiments on the proposed recommendation technique are described in Sect. 4. Section 5 concludes the paper and identifies our future work.

2 Related Work

2.1 Previous Work

We proposed a personalized recommender for travel itineraries [18], which uses the K-Means clustering and tag-based recommendation algorithms. Once the user’s travel plan is obtained, the recommender selects hotels according to user budget. Meanwhile, based on the favorite POIs specified by users, tags of all POIs are collected from travel websites and the recommender calculates the tag similarity (between all POIs harvested from websites and user’s favorite attractions) to suggest POIs that users may be keen on. When selecting attractions, the user ratings are taken into account. Then, the recommendation system clusters these selected POIs and calculates distances between POIs and hotels. Finally, daily itineraries are generated and evaluated. The SI (the Shortest Itinerary), MEI (the Most Effective Itinerary), and MAI (the Most Appealing Itinerary) are then displayed on map for user selection.

2.2 Word2Vec

Word2Vec is a series of models that are used to produce word embedding, including the Skip-gram Model [6], Neural Net Language Model (NNLM), Continuous Bag-of-Words Model (CBOW) and Recurrent Neural Net Language Model (RNNLM) [11]. It takes as its input a large corpus of text and produces a vector space, typically of several hundred dimensions, with each unique word in the corpus being assigned a corresponding vector in the space. The techniques are proposed for measuring the quality of the resulting vector representations, with the expectation that not only will similar word tend to be close to each other, but that words can have multiple degrees of similarity [12]. However, the amount of relevant in domain topics for automatically recognized is limited [11].

2.3 X-Means Clustering Algorithm

The X-Means clustering algorithm was proposed to solve the three major shortcomings of K-means [3, 16] and it can estimate the value of K quickly. The steps of X-Means are described as follows [14]:

  1. 1.

    Pick one centroid, and then produce a new centroid nearby via running K-means to completion;

  2. 2.

    Calculate the resulting model scores (BIC: Bayesian information criterion [8] or MDL: Minimum description length [15]) and see if the scores are better;

  3. 3.

    If it does, accept the new centroid; otherwise, return to the previous structure;

  4. 4.

    Repeat the former steps until the rounds reaching KMax (Maximum for numbers of clusters).

In our work, the X-Means clustering algorithm has been adjusted in step 1 and step 2. In step 1, K-Means++ (see below) can be used for calculation of the initial centers. Moreover, in step 2 we need to set a specified splitting criterion (tolerance) to control the process of splitting clusters. We measure the BIC as our scores. The steps of K-Means++ are presented as follows [2]:

  1. 1.

    Randomly select a data object as a center point from the data set;

  2. 2.

    Calculate the distance between all the data objects and the center points, respectively;

  3. 3.

    Use the distance from object to the closest center to calculate the probability for each point, then select a data object according to the probability as another center point;

  4. 4.

    Repeat step 2 and step 3 until the required number of centers is initialized.

BIC is a likelihood criterion: let be the dataset we are modeling; let be the candidates of desired parametric models calculated as follows [2, 8, 14]:

$$\begin{aligned} BIC(M_i) = \widehat{l_i(X)} - \frac{p_i}{2} * \log {N} \end{aligned}$$
(1)

where \(\widehat{l_i(X)}\) is the log-likelihood of the data according to the ith model and taken at the maximum-likelihood point, and \(p_i\) is the number of parameters in \(M_i\). N is the sample size.

The maximum likelihood estimate (MLE) for the variance, under the identical spherical Gaussian assumption, is:

$$\begin{aligned} \widehat{\sigma }^2 = \frac{1}{N-K}\sum _i{(x_i - \mu _i)^2} \end{aligned}$$
(2)

The log-likelihood of the data is as follows:

$$\begin{aligned} l(X) = \sum _i{(log(\frac{1}{\sqrt{2\pi }\sigma M}) - \frac{1}{2\sigma ^2}\Arrowvert {x_i - \mu _i}\Arrowvert ^2 + log{\frac{N_i}{N}})} \end{aligned}$$
(3)

2.4 Tag-Based Recommendation Algorithm

Being increasingly used in various networks, tags are seen as a potential source of user interest, preference, and user profile construction [8]. For example, a user may indicate her travel preferences for luxuries and arts. Meanwhile, the Marina Bay Sands happens to have been tagged with “luxury travel, family travel, nightlife”. A match can therefore be made [17]. To some extent this helps alleviate the cold start problem for recommender systems. Users tag an item when they develop an interest in it. These tags serve as resources for describing such items. We term this associated information as attributes, most of which are discrete categorical variables for the Web domain [4]. Others can also retrieve the item via its tags [7]. By calculating the similarity (use the cosine similarity equation [10]) between user tags, recommender system can predict the potential rating that a user would give to items, thus eventually being able to recommend the top-N items to her [13].

2.5 Summary

We combine the use of Word2Vec (w2v for short) with that of the X-Means clustering algorithm in our improvement. Firstly, we use the Chinese Wikipedia corpus to train our word embedding model and determine parameters, including the minimum count and the value of window, as 1 and 5, respectively. Before that, “mecab”Footnote 4 has been used to tackle the Chinese corpus. We need the w2v model to obtain similar tags of POIs. The X-Means algorithm is applied for clustering sites and hotels based on their geolocation, which is intended for minimizing distances between each location in real world. After getting all tags of attractions that travelers may be interested in, the tag-based recommendation algorithm is used for generate the final recommended candidate list. In fact, we also use an existing tool—“Synonyms”, which is utilized in obtaining near-synonym. We compared it with the w2v model we trained and eventually we chose one with the better effect.

3 An Itinerary Recommender

In this section, we introduce our improvement on the previously proposed recommender system. Users inform the recommender of their preferences, including their destination, planned travel date, hotel budget, favorite attraction sites, and etc. According to the requirement, itineraries with recommended hotels and POIs are then presented to the users.

At first, the recommender collects information on hotels, POIs, and tags from tripdavisor.com, and then uses X-Means to cluster the hotels and POIs according to their geolocations. Meanwhile, a trained w2v model is applied to calculate word similarity, which helps get similar tags of specific ones. In the phase of tag cleaning, the recommendation system divides all POIs into two categories, indoor POIs and outdoor POIs, based on their tags. According to user’s favorite sites, interested tags and similar tags are used in tag-based recommendation algorithm, combined with the number of POIs’ reviews, the rating of POIs, and the suggested visiting time, in order to produce a candidate sites list. Next, different clusters of all sites in the candidate list are counted. To recommend neighboring hotels with those sites, the recommender needs to calculate the distance between clusters of hotels and sites, and then selects the closest hotel cluster and the first three hotels from this cluster are picked up, with the highest rating and the price that suits user’s hotel budget. The weather information is then taken into account for working out a daily itinerary. With all available routes, the recommender also provides users with information on the time that they would spend on commuting between sites. The workflow is shown in Fig. 1.

Fig. 1.
figure 1

The workflow for a recommender

Full size image

3.1 Selection of Attractions

Having obtained user’s favorite attraction sites, the recommender calculates the tag similarityFootnote 5 between the user’s favorite sites and all the other attractions in the same city. Attractions are ranked in reverse order of their tag similarity and a threshold of the number of reviews is set. Because the number of travel days (d) is provided by users, the recommender can select the first d attractions. With the same similarity, a site with a higher rating is chosen. If chosen attraction sites do not belong to different clusters, we continue to add the first d (the number of clusters) attractions to the candidate sites list, until all sites on the list can be divided into d clusters or the length of the list is equal to 4*d (The default number of recommended daily attractions is 4).

The number of clusters in the candidate sites list is counted. If the number is equal to d, Scheme1 is used. Otherwise, we use Scheme 2.

Scheme 1:

Count the number of attractions in each cluster:

  1. 1.

    If there is one attraction in a cluster, the recommender determine whether the suggested visiting time of this site is greater than three hours. If so, in one day just one attraction is recommended to users; otherwise, the first three attractions, with the highest tag similarity and the suggested visiting time that is less than three hours, are selected;

  2. 2.

    If there are more than or equal to four attractions in a cluster, user’s favorite sites are taken into priority. If the number of user’s favorite sites is greater than 4, in one day all those sites are recommended to users. Otherwise, the first three attractions, with the highest tag similarity and the suggested visiting time that is less than three hours, are added into the candidate list, until the number of attractions in this cluster is equal to 4;

  3. 3.

    If there are less than four attractions in a cluster, all those attractions are recommended to users for a day.

Scheme 2:

Determine whether all user’s favorite sites belong to the same cluster:

  1. 1.

    If so, several attractions, with the highest tag similarity in this cluster and the suggested visiting time that is less than three hours, are added into the candidate list, until the length of the list is equal to 4*d;

  2. 2.

    Otherwise, one site, with the highest tag similarity in its cluster and the suggested visiting time that is less than three hours, is picked up in these clusters in turn, until 4*d attractions are selected.

3.2 Planning of Daily Itineraries

After a candidate sites list is generated, weather information is taken into consideration for producing daily itineraries. According to user’s travel date, weather information is harvested using the Xinzhi APIFootnote 6. However, the validity of weather information is time-dependent.

If adverse weather is expected during the travel period, for example a storm is on the way, the weather will be given top priority by our recommender when it works on generating daily itineraries. In the Sect. 3.1, two categories of attractions are labeled. If the number of days with adverse weather is equal to the number of clusters of indoor attractions, the recommender plans daily itineraries for indoor sites and outdoor sites, respectively. Finally, the system combines them in a whole daily itinerary. If the number of days with adverse weather is less than the number of clusters of indoor attractions, the same method is applied. In the final step of combining routes, the system needs to arrange indoor attractions on some days with fine weather.

After taking weather information into consideration, the recommender is expected to yield daily itineraries for users. If there is only one cluster in the candidate sites list, the recommendation system recommends four sites to users every day (the order of recommended sites is based on their geo distance), until all sites in the list are planned. If the number of clusters is equal to the number of travel days, the system plans itineraries for each cluster respectively, until all clusters are planned. If the number of clusters is greater than the number of travel days, excludes nFootnote 7 clusters that not include user’s favorite sites. However, if all clusters include user’s favorite sites, excludes n clusters randomly. The method of planning itineraries is same to the above condition. If the number of clusters is less than the number of travel days, the system selects the first n clusters with the most number of attractions and plans itineraries. Finally, a whole daily itinerary is generated.

3.3 Selection of Hotels

In Sect. 3.2, the recommender generates a complete daily itinerary for users. Next, the system calculates distance between different clusters of attractions in the itinerary and all clusters of hotels, respectively. Clusters of hotels are ranked in order of their average distance between each hotel cluster and all clusters of POIs in the route, and the cluster with the smallest average value is chosen. In this cluster, the recommender selects hotels that their price is within user’s hotel budget and the number of reviews is greater than a threshold, and then they are ranked in reverse order of ratings. The first three hotels are recommended. If there are less than three hotels, the system recommends all hotels that meet the requirements to users.

4 Experiments

We anticipated to recommend different kinds of itineraries that best meet the user’s travel needs. A series of experiments were therefore carried out to demonstrate the efficiency of the proposed recommender for itinerary planning.

4.1 Evaluation Metrics

We used the same evaluation metrics with our previous work [18], including the tag similarity (\(tag\_sim_{rou}\)), the travel distance (\(dis\_sim_{rou}\)) , and the efficiency (\(comp_{rou}\)) that takes into account the hotel price, tag similarity, and the travel distance. The most appealing itinerary is formed when \(tag\_sim_{rou}\) reaches its maximum. The shortest itinerary and the most effective one are formed when \(dis\_sim_{rou}\) and \(comp_{rou}\) reach their minimum, respectively.

4.2 Methodologies, and Experimental Settings

We used a desktop Dell with a 64-bit operating system. The CPU is intel(R) Core(TM) i7-7700, and the software supporting our experiments includes Python 3.5 and MongoDB. We also utilized the package googlemap to generate distance and commuting time between all sites and hotels. The Xinzhi Weather API was used to acquire weather information around the world.

Table 1. Travel Questionnaire
Full size table
Table 2. Parameters in the clustering process for hotels and POIs
Full size table

We simulated the requirements of users Alva, Bob, Carlo, Peter, and Tom, and different travel questionnaires are filled out for them, which are shown in Table 1 (BJ: Beijing, HZ: Hangzhou, WH: Wuhan, SZ: Suzhou, QZ: Quanzhou). Information on 2,017 sites and 8,885 hotels was harvested from tripadvisor.cn and then saved in MongoDB. “Synonyms” was a Python package that was used to get similar tags. In addition, we applied the X-Means clustering to the grouping of different hotels and sites from different cities. After clustering has been performed for several times, the best parameters of each city are picked up. This is shown in Table 2. Figures 2, 3, 4, 5 and Fig. 6 illustrate the clustering results.

Fig. 2.
figure 2

BJ: the clustering result of hotels (left) and of attractions (right)

Full size image
Fig. 3.
figure 3

HZ: the clustering result of hotels (left) and of attractions (right)

Full size image
Fig. 4.
figure 4

SZ: the clustering result of hotels (left) and of attractions (right)

Full size image
Fig. 5.
figure 5

WH: the clustering result of hotels (left) and of attractions (right)

Full size image
Fig. 6.
figure 6

QZ: the clustering result of hotels (left) and of attractions (right)

Full size image

4.3 Results and Analysis

Taking Hangzhou as an example, the itineraries for Bob have been displayed in Table 3. We also evaluated them in terms of the metrics introduced in Sect. 4.1. The result is shown in Table 4.

Table 3. Daily itineraries for Bob (Hangzhou)
Full size table
Table 4. The result of evaluation for Bob (Hangzhou)
Full size table

As shown in Table 4, we acquired the shortest itinerary for Bob that sets off from the Hofang Youth Hostel, and an itinerary with the highest performance/price ratio that also sets off from the Hofang Youth Hostel.

Finally, we compared these itineraries for Bob given by our proposed recommender with the routes presented in [18], in terms of the three evaluation metrics. Only the first two daily itineraries of Bob’s recommendation results are considered, because in the previous experiment only two daily itineraries were provided.

Table 5. A comparison between itineraries provided by our previous recommender and the current one
Full size table

The comparison result is shown in Table 5, from which we can tell that the itinerary provided by this recommender is shorter in distance and higher efficiency.

5 Conclusions

We improved our recommender of personalized itineraries in the sense that more factors are taken into account, such as the suggested visiting time for each POI and future weather data. The X-Means clustering algorithm was used to cluster all sites and hotels to help minimize the entire distance involved in the travel among them. Then a Word2Vec model was applied to obtain several similar tags of specific ones. In addition, a tag-based algorithm was utilized to create a list of candidate attractions that best match with the user’s favorite sites. Eventually the itineraries that would meet different user needs were created and the results were displayed in a format with complete distances and commuting time for user selection.

We performed experiments with a few cities in China, including Beijing, Hangzhou, Suzhou, Wuhan, and Quanzhou, and found out some parameters for clustering to help produce the shortest itinerary. On our server (118.89.196.180:8888), we will deploy our proposed travel recommender shortly.

In the future, we plan to enhance accuracy in finding similar tags by training corpus in travel domain. Some conditions related to weather should be detailed in order to obtain more flexible itineraries.