Exploiting Sequential Music Preferences through Optimisation-Based Sequencing

October 26, 2023 Published by Dmitrii Moor, Yi Yuan, Rishabh Mehrotra, Zhenwen Dai, Mounia Lalmas

We research the issue of establishing customized playlists for customers of music streaming providers, in our case Spotify. To ship finest at school expertise for such customers not solely understanding the customers’ preferences over particular person music tracks but in addition understanding how customers eat the tracks as they navigate by playlists is necessary.  In distinction to many different advice domains (equivalent to films or books) consumption of music is inherently sequential: customers not often take heed to solely a single monitor from a playlist; as a substitute, they sometimes eat the complete classes, or sequences of tracks, one after one other. In our work we study the sequential music preferences of customers and take a look at how we are able to leverage such sequential preferences in an optimisation-based approach.

We present that even easy ML fashions are able to capturing probably the most important sequential points of customers’ preferences. Through in depth offline off-policy analysis, we reveal how our optimisation-based sequencing method can leverage such ML fashions to extend consumption. Finally, counting on a large-scale A/B take a look at we present that we are able to considerably improve the variety of accomplished tracks in customers’ classes. 

Sequential Music Preferences

Generally, there exist quite a few totally different points of customers’ sequential preferences. Consumption of any monitor by the person could also be affected by the person’s previous short-term and long-term consumption in some ways; it could additional rely on the person’s anticipated future consumption, and so forth. In our paper, we argue that whereas capturing extra of such sequential points is feasible by constructing extra complicated ML-based desire fashions, the issue of establishing optimum playlists by using such fashions turns into computationally intractable. 

Therefore, as a substitute of attempting to seize all of the refined points of the customers’ sequential habits, we solely determine two primary sequential hypotheses that a lot of the customers’ preferences would have. We then attempt to leverage these particular hypotheses within the sequencer in an optimisation-based approach.

To this finish, we first posit that there exists a relationship between customers’ consumption decisions and the positions of the tracks within the playlists. Intuitively, the primary monitor within the playlist could also be indicative for the person to resolve whether or not they wish to begin listening to the playlists. As an instance, if the monitor on the primary place is acquainted to the person, the person begins consuming tracks from the playlist perhaps extra possible. As a second instance, if the person has reached the additional positioned tracks, the person persevering with consuming tracks from the playlist is extra possible. We confer with such preferences as position-aware preferences.

Second, we hypothesize that whether or not the person consumes a monitor on a sure place relies upon not solely on that particular monitor but in addition on the monitor on the earlier place. Intuitively, if the adjoining tracks are considerably totally different of their acoustic options, the ensuing music movement might turn out to be much less coherent and consequently, much less gratifying for the person. Figure 1 illustrates the acoustic power of the tracks on positions 2 and three of the playlist. We can see that if the acoustic power of the respective tracks is extra comparable (blue line), it makes it extra possible that the person completes each tracks. However, because the distinction within the acoustic power will get bigger (purple line), the person would extra possible skip the latter monitor. In what follows, we confer with such preferences as local-sequential preferences.

Figure 1. Mean acoustic power of two adjoining tracks. Track on place 2 is accomplished by the person. If the acoustic power of the tracks is comparable, it’s extra possible that the monitor on place 3 is accomplished (blue); in any other case, it’s possible that monitor 3 is skipped (purple). 

Optimisation-Based Sequencing 

We body the playlist sequencing drawback as a stochastic optimization drawback. We mannequin the rewards r of the person interacting with the music tracks as random variables. These rewards might rely on the person u, the monitor t in addition to on the complete sequence of tracks π (see Figure 2). In this case, we intention at discovering such a sequence of tracks π that maximizes the anticipated complete discounted reward. Equation 1 illustrates our method.

Equation 1. Optimal Sequencing Problem

Such a generic formulation permits us to explicitly introduce our position-aware and local-sequential assumptions into the ML-based desire mannequin f(.) that predicts the possibilities of the totally different rewards r. In our paper, we illustrate how this may be achieved with pretty easy ML fashions. The ensuing possibilities estimated by such fashions permit us to resolve the sequencing drawback in an optimisation-based approach. 

Experiments

To higher perceive whether or not introducing the position-aware and local-sequential assumptions into the desire mannequin permits to extend consumption we depend on offline off-policy analysis. To this finish, we first ran a randomized knowledge assortment A/B take a look at after we uncovered the customers to the playlists ranked with a uniform random logging coverage. We then relied on the collected randomized knowledge and on inverse propensity scoring (IPS@ok) to estimate the anticipated variety of accomplished tracks throughout the primary ok positions of the goal insurance policies outlined by our optimum sequencing drawback (see Equation 1). We relied on two modifications of IPS, particularly Independent IPS (IIPS@ok) and Reward-Interaction IPS (RIPS@ok) that permits us to explicitly mannequin the sequential rewards.  

Table 1 illustrates the estimated anticipated numbers of accomplished tracks throughout the primary ok positions. Introducing the position-aware and the local-sequential assumptions into the desire mannequin helped growing the anticipated variety of accomplished tracks in comparison with the straightforward relevance sequencer (i.e, a easy mannequin that ranks tracks utilizing the cosine similarity between the person/monitor latent representations) in addition to to the state-of-art neural non-sequential Myopic sequencer. 

When trying into the acoustic coherence of the playlists constructed with our local-sequential mannequin we additionally noticed that our mannequin allocates acoustically comparable tracks nearer to one another in comparison with the non-sequential Myopic sequencer. Figure 2 illustrates the distinction in acoustic options between the 2 adjoining tracks for the primary 9 positions of the playlist. We see that the acoustic distinction for the sequential mannequin (blue) is usually smaller than for the non-sequential one (purple) for all twelve acoustic options. This means that our sequential mannequin can assist to assemble playlists that ship a extra acoustically coherent music movement to the customers in comparison with the non-sequential fashions.

Figure 2. Mean distance between the acoustic options of consecutive tracks throughout totally different positions of the playlists.

Finally, we relied on our best-performing local-sequential mannequin to hold out an internet A/B experiment. We noticed that our mannequin permits us to extend the variety of accomplished tracks per session by +2.52% relative to the non-sequential baseline. Skip charges decreased by -2.73%.  

These outcomes are much more pronounced for  Free customers than for Premium ones. This permits us to considerably enhance the expertise of free customers who can solely skip a restricted variety of tracks. Table 2 illustrates our outcomes. 

In conclusion, we see that counting on sequential fashions and optimisation-based sequencing permits us to considerably improve the variety of accomplished tracks in customers’ classes and reduce skips. We additionally see that the ensuing playlists are extra acoustically coherent. 

For extra data, please confer with our paper: 
Exploiting Sequential Music Preferences through Optimisation-Based Sequencing
Dmitrii Moor, Yi Yuan, Rishabh Mehrotra, Zhenwen Dai, Mounia Lalmas 
CIKM 2023