In music recommender systems, playlist continuation is the task of continuing a user's playlist with a tting next track, often also referred to as next-track or sequential recommendation. This work investigates the suitability and applicability of autoencoders for the task of playlist continuation. We utilize autoencoders and hence, representation learning to continue playlists. Our approach is inspired by the usage of autoencoders to denoise images and we consider the playlist without the missing next-track as a noisy input. Particularly, we design di erent autoencoders for this speci c task and investigate the e ects of di erent designs on the overall suitability of recommendations produced by the resulting recommender systems. To evaluate the suitability of recommendations produced by the proposed approach, we utilize the AotM-2011 and LFM-1b datasets. Based on those datasets, we show that n-grams are a well performing alternative baseline to kNN. Fruther, we show that it is possible to outperform a kNN as well as an n-gram baseline with our autoencoder approach.
Recommender systems are applicable to a broad spectrum of domains. The music domain is one such application area, where one speci c task is next-track music recommendation. Next-track music recommender systems are recommender systems that aim to nd a tting continuation (next-track) for a given playlist. In general, a playlist is an ordered list of music tracks, where the order is based on time, which means that the rst track in the list is expected to be listened rst, followed by the second track and so on. In other words, a playlist is a time series of tracks.
Multiple di erent approaches have been proposed for the
playlist continuation task. As mentioned in [
In the eld of music recommendation, deep learning
approaches are usually used to include additional features such
as textual information [
With this work, we investigate the general applicability of autoencoders for the playlist continuation task. We report the e ects of di erent parameter settings on the overall suitability of recommendations produced by the system and answer the following research questions:
RQ1: How can playlists be vectorized for an autoencoder? RQ2: Is there an alternative baseline to kNN that better utilizes the order of tracks in a playlist? RQ3: Which autoencoder design produces competitive results for the next-track music recommendation task?
In [
The remaining sections of this paper are organized as follows. First, related work will be discussed in Section 2. Afterwards, in Section 3, we describe the algorithm to convert a playlist into a corresponding vector and present our autoencoder approach. Following that, the experimental setup to evaluate our approach by comparing it with a kNN baseline is described in Section 4. Thereafter, we present the results in Section 5 and nally draw a conclusion in Section 6.
This section gives an overview of next-track music recommendation approaches.
We consider the playlist continuation task to be a special
case of the more general task of playlist generation.
According to Bonnin and Jannach [
Sedhain et al. [
Jannach et al. [
The evaluation approach presented by McFee and
Lanckriet [
In this section, the proposed recommendation approach based on an autoencoder is presented. First, we will explain how playlists are converted into a vector, which is necessary to use them as an input for the autoencoder. Afterwards, the structure and implementation details of the autoencoder are presented in Section 3.2. This includes the general training procedure used and a modi ed autoencoder layout to overcome over tting by simulating the continuation task during training.
Playlists are usually represented as an ordered lists of tracks. In the special case of the playlist continuation task, the playlist used as input for the algorithms is often referred to as \history". This history can be considered a list of past listening events.
To use playlists as an input for autoencoders it is necessary to convert the ordered list into a vector representation. We propose three di erent ways to determine the value of each dimension of the generated playlist vector, as presented in the following. An example of all three encodings is shown in Figure 1.
Binary Encoding is the simplest encoding and it is inspired
by the (one-hot) vector encoding used in [
Order Encoding is a modi ed version of the Binary Encoding and includes ordering information. We propose to use the track's index i in the playlist as the value of the dimension t assigned to the track. Therefore, the track vector encoding contains 0 for all dimensions except of the dimension with index t, to which the value i is assigned. To obtain a playlist vector p~, all track vectors are summed up.
Normalized-Order Encoding is an extension to Order
Encoding and takes the length of a playlist into account. The
playlist vector p~ is normalized by the number of tracks
contained in the corresponding playlist, which reduces the
effects of the playlist length on the encoding.
Autoencoders are an unsupervised learning method used
for representation learning [
In contrast to a representation learning task, where the decoder is removed after the training phase to get the code as an output of the network, we use the whole network to compute recommendations. Recommendations are computed based on the following procedure: Given a playlist that should be continued with a tting next-track as an input, it is necessary to convert this playlist to a vector. Afterwards, this vector encoded playlist is used as an input for the trained network that produces an output vector represented by the output layer. The output vector holds rating values for all tracks (number of dimensions of the vector) contained in the dataset, which was used to train the network. This output vector is then converted to a prioritized list of tracks by creating a list of indexes (track id's) ordered by their corresponding value in the vector. The index with the highest value is the rst in the list while the index with the lowest value is last. Further, all tracks contained in the input playlist are removed from this prioritized list to get next-track recommendations di ering from the tracks contained in the input playlist. Mostly, it is necessary to compute a given number of possible continuations. This is achieved by chopping o the list after the given number of tracks.
Using Keras2, we implemented an autoencoder in Python. Our implementation allows to set the number of epochs, the hidden code layer activation function, the output layer activation function and the used loss function. The input layer size equals the output layer size and is determined automatically base on the dataset. Further, we decided to automatically adapt the code layer size based on the input size divided by 40, which is a result from preliminary experiments. Keep in mind that dense layers are used to build the network, which means that all nodes of one layer are connected to each node of the neighboring layers. Our network consists of one input layer followed by a ltering layer that removes the last track of the input during training. This ltering layer is followed by an optional dropout layer with 0.5 dropout rate that can be disabled. This layer is then followed by a hidden code layer and an output layer.
To train the neural network, a training set is used, which consists of an input and the expected output, which are both the same for autoencoders. The autoencoder learning method is depicted in Figure 2. To train an autoencoderbased on the previously introduced playlist vectors, it rst is necessary to create an autoencoder with an input/output layer size tting the dimensionality of the playlist vectors in the dataset. The training process is further con gured to use 614 of the total number of playlists the training set as
its batch size and uses Adam as an optimizer because preliminary experiments showed that these settings work well and that they show similar performance compared to other parameter choices.
This setup allows to compare di erent parameter con gurations of an autoencoder where the basic structure of the used neural network, which can be seen in the center part of Figure 2, remains the same. 4.
In this section, we present the setup used for the conducted experiments to evaluate the suitability of recommendations produced by the autoencoder-based approach in comparison to a kNN baseline. In Section 4.1, we introduce the used datasets. Thereafter, in Section 4.2 the kNN baseline recommender is introduced. Followed by an explanation of the n-gram baseline in Section 4.3, continued by the overall experimental setup in Section 4.4. 4.1
To evaluate the recommender systems, we aim for datasets
that are based on user interaction such as listening logs
and playlists because next-track music recommender
systems must satisfy the needs of users. Along the lines of
previous work [
Based on the LFM-1b dataset [
This results in approximately 62 million sessions which we ltered further. In a rst step, a session chunk containing the rst 3 million sessions with a minimum length of three was created. This chunk shows a high number of di erent tracks. Reducing the number of tracks contained in a dataset is an important step to keep the size of the resulting neural network low enough to train it in reasonable time. This was achieved by dropping all playlists that contain rarely occurring tracks. We dropped all playlists containing tracks with 840 or fewer occurrences.
Further, we use the AotM-2011 dataset [
A k-Nearest Neighbors (kNN) approach [
has been used for the playlist continuation task in multiple
other papers such as [
The basic idea behind kNN is to nd k di erent items that
are the nearest neighbors of a given item. Neighborhood
of items in general is de ned based on a distance-function,
which allows to take di erent properties into account.
Further, it is necessary to get to a conclusion from those k
neighbors which can, for example, be done with a majority
vote [
In [
N-grams are a common statistical model in natural
language processing. This technique is for example used in [
Therefore, the simplest model is a unigram model (n = 1) which only counts track occurrences in playlists.
Increasing the number n of the n-gram model makes it possible to consider the previous n 1 tracks for the prediction by calculating probabilities as given in the following equation:
Pn-gram tijti (n 1) =
F ti (n 1); : : : ; ti F ti (n 1); : : : ; ti 1 (1) where ti is the ith track in a sequence of tracks t1; t2; : : : ; tn. F (seq) is the frequency of occurrences of a given sequence of tracks seq in the training set. Ranking the tracks by their probabilities (highest rst) allows to make predictions based on the probabilities learned by an n-gram model. 4.4
We used scikit-learn7 to run a grid search on the parameters of the approaches. To ensure that the reported results are not bound to a speci c train-test split of the datasets, we run a ve-fold cross-validation. The k-fold splitting procedure of scikit-learn was used with a xed random state (seed) to ensure the reproducibility and comparability of the results. The metrics used for the evaluation are recall (r) and mean reciprocal rank (mrr).
For the evaluation of all recommender systems, a two-step
process is used for each of the two presented datasets. In
the rst step, each recommender system is trained using the
training data. For this purpose, a new instance of the
recommender system is created for each run and then trained
6https://pypi.org/project/implicit/0.3.8/
7https://scikit-learn.org/
using the training procedure. The second step of the
evaluation process computes recommendations with the previously
trained recommender. Therefore, each playlist in the
training set is decomposed into a history (all tracks except the
last one) and the last track [
The autoencoder implementation presented in Section 3.2 has a high degree of freedom in terms of modi able parameters. Therefore, we decided to x the reduction factor (code layer size), the batch size and the optimizer. The used values were obtained from preliminary experiments. Based on the knowledge gained from those preliminary experiments we determined value ranges for the other parameters used in the grid search. 5.
In the following, we rst report the recommendation suitability of the di erent encoding types in Section 5.1. After that, the results achieved by the n-gram baseline are shown in Section 5.2. Finally, we compare the suitability of recommendations produced by our approach on di erent datasets in Section 5.3. 5.1
In a rst evaluation step, the recommendation suitability of the di erent encoding types introduced in Section 3.1 was compared using the kNN baseline, based on the AotM and LFM-1b datasets. Table 2 shows the results for kNN using the item-item distance metric. We observe that the normalized-order encoding outperforms both other encodings on both datasets and for di erent values of k. Interestingly, order encoding without normalization has a negative e ect on the performance of the kNN implementation. This can be explained by the fact that the length of a playlist is encoded as well. In addition, order encoding has a larger vector space than binary and normalized-order encoding as the values in each dimension have a bigger range. Due to space reasons we do not include results for the cosine and tfidf distance metric and other values of k, as these show that normalized-order encoding works best. Further experiments showed, that the Autoencoders show similar behavior when the encoding type is changed.
Based on these ndings, we argue that the normalizedorder encoding should be used among the three encodings introduced in Section 3.1, which also answers RQ1. Therefore, we use normalized-order encoding for all further evaluations. 5.2
To evaluate the suitability of an n-gram model we decided to compare a bigram (n = 2) and a trigram (n = 3) model with the best performing kNN (found using a grid search) con guration on each dataset. The kNN baseline using the item-item as a distance metric with 20 neighbors (kNNi20) works best on the AotM dataset while the cosine distance with 50 neighbors (kNNc50) works best on the LFM-1b dataset. In addition, we give the results of a unigram (n = 1) model that always recommends the most popular tracks.
Table 3, shows that a bi- and trigram is able to outperform a unigram model. Further, it can be seen that those n-gram models work much better on our LFM-1b dataset variation than on the AotM dataset. This can be lead back to the fact that each track on average occurs in 0:22 playlists in the AotM dataset compared to 7:79 occurrences in the LFM-1b dataset, as stated in Table 1. Therefore, common sequences of tracks among playlists are more likely in the LFM-1b dataset than in the AotM dataset. We argue that this is also the reason why the absolute values of all metrics di er that much when comparing the results on both datasets. Compared to both kNN baselines it can be seen that the ngram models work better on the LFM-1b dataset especially for short recommendation lengths. In contrast, they are less e ective on the AotM dataset than kNN models.
It can be seen that n-gram models form a strong baseline for the LFM-1b dataset. Note that the kNN models operate on the normalized-order encoding while the n-gram models utilize the track sequences directly without any encoding, which answers RQ2. 5.3
In Table 3 we depict the results of our autoencoder approach in comparison to the kNN and n-gram baseline. The results show, that it is possible to outperform a kNN baseline on both datasets using the proposed autoencoder approach. This is especially true for longer recommendation lengths. To give a better overview of the capabilities of our autoencoder approach we give the results of four autoencoder con gurations. To distinguish the di erent parameter con gurations of our autoencoder approach we decided to name each con guration (AE1{AE4), where we report results. For each dataset we include results for one autoencoder including the dropout layer (see Section 3.2) and one without dropout. AE1 without the dropout layer and AE2 with the dropout layer are respectively the best performing autoencoder con gurations for the AotM dataset, while AE3 (without dropout) and AE4 (with dropout) perform best on LFM-1b according to the mrr@1. AE1 uses tanh as the code layer and output layer activation function with cosine proximity as the loss function and is trained over 5 epochs. AE2 utilizes relu as the code layer activation and softmax as the output layer activation with categorical crossentropy loss and was trained over 40 epochs. AE3 and AE4 are both trained over 40 epochs use tanh as the code layer activation and cosine proximity as loss. While AE3 is con gured with a softmax output activation, AE4 is con gured to use sigmoid for output activation.
It can be seen in the results that autoencoders outperform both the kNN and n-gram baselines on the AotM dataset while they are not able to outperform n-gram models on the LFM-1b dataset. Surprisingly, AE1 works best on AotM when trained for 5 epochs. While AE2, AE3 and AE4 reveal similar results per dataset AE1 only produces comparable results on the AotM dataset which answers RQ3. One possible explanation is that it over ts on the particular training set which is tried to be prevented using a dropout layer.
In the above section, the recommendation suitability impact of multiple con gurations of an autoencoder were presented. Additionally, results of the best performing con gurations on di erent datasets are given. The results show, that autoencoders can be used for the playlist continuation tasks when con gured correctly. 6.
In this work, we proposed a novel autoencoder approach for the playlist continuation task. To use playlists as an input for autoencoders, we introduced a procedure to encode playlists as vectors. The evaluation shows that the proposed autoencoder approach outperforms a basic kNN approach. Particularly, the results show that this is the case regardless of the playlists/track ratio of the used dataset.
This work solely focuses on determining if an
autoencoder approach can be used for the playlist continuation
task. We showed that outperforming basic kNN is
possible for datasets, that we consider small in comparison to
the amount of data given in a real-world scenario. One
possible source of improvement is the training procedure.
Autoencoders are usually trained to reconstruct the input
which we modi ed slightly. We introduced a ltering layer
in the training phase that removes the last track of the
input. This trains the autoencoder to \reconstruct" playlists
including the last track (next-track) ltered in the input.
Additionally, it would be possible to speci cally designing
a loss function for the continuation task. Strub et al. [
Additionally, a more advanced training procedure could
lead to a well performing deep autoencoder. One way of
creating a deep autoencoder would be to rst train an
autoencoder with one input and one output layer (as the ones
proposed in this work) and then use the learned code as an
input to train another autoencoder. After that, it is possible
to split the rst autoencoder into the encoder and decoder
parts and insert the second autoencoder in between. The
resulting autoencoder then can be ne-tuned and extended
in the same way. This process is like the one proposed by
Vincent et al. [
In addition, a user-study to evaluate the approaches should be conducted in future work. This is important to get an impression of the user-perceived quality of the approaches.