The Short-term User Modeling for Predictive Applications

Abstract

One of the important purposes of data mining on the web is to reveal hidden characteristics of users including their behavior. These characteristics are often used to analyze previous user actions, his/her preferences, and also to predict the future behavior. An average user session consists of only few actions, which brings several complications for the user modeling and also for subsequent prediction tasks. Such tasks are usually researched from the long-term point of view (e.g., contract renewal or course quit). On the contrary, the short-term user modeling plays an important role in the context of web applications, where it helps to improve user experience. Its shortcoming is that it often requires rich data, which availability is rather rare. For this reason, we propose a novel user model focused on the capturing changes in the user’s behavior on the level of specific actions. The model idea is based on the enrichment of user actions by a comparison of actual user session data with previous sessions. As the model basis on generally available data sources, the approach is applicable to wide scale of existing systems. We evaluate our model by the task of session end intent prediction in the e-learning and news domain. Thanks to reflecting differences in user behavior we are able to predict the intent to end the session for particular user in the scale of his/her next couple of actions. Obtained results clearly show that the proposed model brings higher precision, accuracy and session hit ratio than baseline models.

Appendix

Proposed model consists of two attributes types (Fig. 1):

• 1. Comparative attributes $A_C$

• 2. Descriptive attributes $A_D$

There are in total nine comparative attributes $ac_1,\ldots ac_9$, which are described in Table 1. Each comparative attribute is further computed for each of the eight time layers $tl \in TL$ and each of the two model part $mp \in MP$. Resulting in $9 \times 8 \times 2$ values of comparative attributes. The formal definition of nine comparative attributes (Table 1) is following.

Attribute. Let $S_{u,n}$, $n\in {\mathbb {N}}$ be the actual session of a user u and $S_{u,n} \in S_u$. Then $Act_{u,n}$ is the set of actions user performed in a session n$S_{u,n}$.

$$\begin{aligned} Attr1= & {} \frac{\left| Act_{u,n} \right| }{\overline{\left| Act_{u,1..n-1} \right| }} \end{aligned}$$

(5)

$$\begin{aligned} Attr2= & {} \left| Act_{u,n} \right| - \overline{\left| Act_{u,1..n-1} \right| } \end{aligned}$$

(6)

$$\begin{aligned} Attr3= & {} if \,\, \left| Act_{u,n} \right| \ge \overline{\left| Act_{u,1..n-1} \right| } \, \, \, then \, \, 1\, \, else\, \, 0 \end{aligned}$$

(7)

Let $ts_{u,n}$ be the time spent by user u in the session $S_{u,n}$.

$$\begin{aligned} Attr4= & {} \frac{\left| ts_{u,n} \right| }{\overline{\left| ts_{u,1..n-1} \right| }} \end{aligned}$$

(8)

$$\begin{aligned} Attr5= & {} \left| ts_{u,n} \right| - \overline{\left| ts_{u,1..n-1} \right| } \end{aligned}$$

(9)

$$\begin{aligned} Attr6= & {} if \,\, \left| ts_{u,n} \right| \ge \overline{\left| ts_{u,1..n-1} \right| } \, \, \, then \, \, 1\, \, else\, \, 0 \end{aligned}$$

(10)

Let $S_{cat,u}$ be the set of all sessions terminating after visiting a page with category cat and $S_{u}$ set of all sessions of user u.

$$\begin{aligned} Attr7=\frac{\left| S_{cat,u} \right| }{\left| S_u \right| } \times \left| Act_{u,n} \right| \end{aligned}$$

(11)

Let $Act_{cat,u,n}$ be the subset of actions $Act_{u,n}$ with category cat.

$$\begin{aligned} Attr8= & {} \frac{\left| S_{cat,u} \right| }{\left| S_u \right| } \times \left| Act_{cat,u,n} \right| \end{aligned}$$

(12)

$$\begin{aligned} Attr9= & {} \frac{\left| S_{cat,u} \right| }{\left| S_u \right| } \times ts_{u,n} \end{aligned}$$

(13)

Let $pageseq_m$ be the sequence of length m of consequent visited pages, i.e., user actions $Act_{u,n}$. Let $pageseq_{m,end}$ be the sequence of length m of consequent visited pages, i.e., user actions $Act_{u,n}$ which result in session end.

$$\begin{aligned} Attr10= & {} \frac{\left| pageseq_{m,end} \right| }{\left| pageseq_m \right| }; \, \, \, m=4 \end{aligned}$$

(14)

$$\begin{aligned} Attr11= & {} \frac{\left| pageseq_{m,end} \right| }{\left| pageseq_m \right| }; \, \, \, m=3 \end{aligned}$$

(15)

$$\begin{aligned} Attr12= & {} \frac{\left| pageseq_{m,end} \right| }{\left| pageseq_m \right| }; \, \, \, m=2 \end{aligned}$$

(16)

$$\begin{aligned} Attr13= & {} \frac{\left| pageseq_{m,end} \right| }{\left| pageseq_m \right| }; \, \, \, m=1 \end{aligned}$$

(17)

Let $catseq_{m}$ be the sequence of length m of consequent visited categories, i.e., user actions $Act_{u,n}$ with category cat. Let $catseq_{m,end}$ be the sequence of length m of consequent visited categories, i.e., user actions $Act_{u,n}$ with category cat, which result in session end.

$$\begin{aligned} Attr14= & {} \frac{\left| catseq_{m,end} \right| }{\left| catseq_m \right| }; \, \, \, m=4 \end{aligned}$$

(18)

$$\begin{aligned} Attr15= & {} \frac{\left| catseq_{m,end} \right| }{\left| catseq_m \right| }; \, \, \, m=3 \end{aligned}$$

(19)

$$\begin{aligned} Attr16= & {} \frac{\left| catseq_{m,end} \right| }{\left| catseq_m \right| }; \, \, \, m=2 \end{aligned}$$

(20)

$$\begin{aligned} Attr17= & {} \frac{\left| catseq_{m,end} \right| }{\left| catseq_m \right| }; \, \, \, m=1 \end{aligned}$$

(21)

$$\begin{aligned} Attr18= & {} \left| Act_{u,n} \right| \end{aligned}$$

(22)

$$\begin{aligned} Attr19= & {} ts_{u,n} \end{aligned}$$

(23)

Let $Cat_distinct$ be the set of distinct categories visited in actions $Act_{u,n}$ within session $S_{u,n}$

$$\begin{aligned} Attr20=\left| Cat_distinct \right| \end{aligned}$$

(24)

Let $cat_last$ be the category of the page actually visited by the user.

$$\begin{aligned} Attr21= & {} \left| Act_{cat_last,u,n} \right| \end{aligned}$$

(25)

$$\begin{aligned} Attr22= & {} \left| Act_{cat_last,u,n} \right| \times \left| Cat_distinct \right| \end{aligned}$$

(26)

As described in Sect. 3, descriptive attributes are calculated for two model parts (personal and global). We defined these attributes (Attr1–Attr9) in Eqs. 5–13 from the personal part perspective. The global part attributes 1–9 are calculated in the same way, but all users not only user u is considered. In other words, all metrics for user u are altered by average of all users. For example, in personal part $ts_{u,n}$ is defined as time spent by the user u within the session n (last session). On the contrary, in global part $ts_{u,n}$ results in average time spent by all users within last sessions.

Following the idea of proposed model, each of descriptive attributes (Attr1–Attr9) are also calculated for several time layers. In other words, according to the specific time layer (e.g., month) all attributes are calculated only based on last month data.