serenade 0.3.1

Serenade: Session-based recommender system
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196

extern crate hashbrown;

use std::collections::BinaryHeap;

use hashbrown::{HashMap, HashSet};
//use std::time::Instant;
use hashbrown::hash_map::DefaultHashBuilder as RandomState;
use itertools::Itertools;

use crate::io::{ItemId, Time, TrainingSessionId};
use crate::vmisknn::{SessionScore, SessionTime};
use crate::vmisknn::similarity_hashed::{idf, SimilarityComputationHash};

pub struct VSkNNIndex {
    session_index: HashMap<TrainingSessionId, HashSet<ItemId>>,
    session_max_order: HashMap<TrainingSessionId, Time>,
    item_index: HashMap<ItemId, HashSet<TrainingSessionId>>,
    item_idfs: HashMap<ItemId, f64>,
}

impl SimilarityComputationHash for VSkNNIndex {
    fn items_for_session(&self, session: &TrainingSessionId) -> &HashSet<ItemId, RandomState> {
        &self.session_index[session]
    }

    fn idf(&self, item: &ItemId) -> f64 {
        self.item_idfs[item]
    }

    fn find_neighbors(
        &self,
        evolving_session: &[ItemId],
        k: usize,
        m: usize,
    ) -> BinaryHeap<SessionScore> {
        let num_items_in_evolving_session = evolving_session.len();

        let mut most_recent_neighbors: BinaryHeap<SessionScore> = BinaryHeap::with_capacity(m);

        for session_item in evolving_session.iter() {
            if let Some(sessions) = self.sessions_for_item(session_item) {
                for session in sessions {
                    let max_order = self.max_order_for_session(session);

                    let session_with_age = SessionScore::new(*session, max_order as f64);

                    if most_recent_neighbors.len() < m {
                        most_recent_neighbors.push(session_with_age);
                    } else {
                        let mut top = most_recent_neighbors.peek_mut().unwrap();
                        if session_with_age.score > top.score {
                            *top = session_with_age;
                        }
                    }
                }
            }
        }

        let mut closest_neighbors: BinaryHeap<SessionScore> = BinaryHeap::with_capacity(k);

        for neighbor_session in most_recent_neighbors.into_iter() {
            let mut similarity = 0_f64;

            let other_session_items = self.items_for_session(&neighbor_session.id);

            //            let start_time = Instant::now();
            // Decayed dot product
            for (pos, item_id) in evolving_session.iter().enumerate() {
                if other_session_items.contains(item_id) {
                    let decay_factor = (pos + 1) as f64 / num_items_in_evolving_session as f64;
                    similarity += decay_factor;
                }
            }
            // let duration = start_time.elapsed();
            // let duration_as_micros:f64 = duration.as_micros() as f64;
            // if duration_as_micros > 500_f64 {
            //     println!("slow session matching:{} micros, evolving_session.len():{}, other_session_items.len:{}", duration_as_micros, evolving_session.len(), other_session_items.len());
            // }

            if similarity > 0.0 {
                // Update heap holding top-n scored items for this item
                let scored_session = SessionScore::new(neighbor_session.id, similarity);

                if closest_neighbors.len() < k {
                    closest_neighbors.push(scored_session);
                } else {
                    let mut bottom = closest_neighbors.peek_mut().unwrap();
                    if scored_session.score > bottom.score {
                        *bottom = scored_session;
                    }
                }
            }
        }

        closest_neighbors
    }
}

impl VSkNNIndex {
    pub fn new(
        interactions: Vec<(TrainingSessionId, ItemId, Time)>,
        sample_size_m: usize,
        max_qty_session_items: usize,
    ) -> Self {
        // start only need to retain sample_size_m sessions per item
        let valid_session_ids: HashSet<u32> = interactions
            .iter()
            .cloned()
            .map(|(session_id, item_id, time)| (item_id, SessionTime::new(session_id, time as u32)))
            .into_group_map()
            .into_iter()
            .flat_map(|(_item_id, mut session_id_with_time)| {
                session_id_with_time.sort();
                session_id_with_time.dedup();
                session_id_with_time.sort_unstable_by(|left, right| {
                    // We keep the sessions with the largest time values
                    left.cmp(right).reverse()
                });
                if session_id_with_time.len() > sample_size_m {
                    // we remove the sessions per item with the lowest time values
                    session_id_with_time.truncate(sample_size_m);
                }
                if session_id_with_time.len() > max_qty_session_items {
                    // this training session has too many items and does not contribute to improving predictions
                    session_id_with_time.clear();
                }

                let session_ids: HashSet<u32> = session_id_with_time
                    .iter()
                    .map(|session_id_time| session_id_time.session_id)
                    .collect();
                session_ids
            })
            .collect();
        // end only need to retain sample_size_m sessions per item

        let mut historical_session_index: HashMap<TrainingSessionId, HashSet<ItemId>> =
            HashMap::new();
        let mut historical_session_max_order: HashMap<TrainingSessionId, Time> = HashMap::new();
        let mut historical_item_index: HashMap<ItemId, HashSet<TrainingSessionId>> = HashMap::new();

        //let mut ignored_training_rows = 0;
        for (session_id, item_id, order) in interactions.into_iter() {
            if !valid_session_ids.contains(&session_id) {
                //ignored_training_rows += 1;
                continue;
            }
            let session_items = historical_session_index
                .entry(session_id)
                .or_insert(HashSet::new());
            session_items.insert(item_id);

            let current_max_order = historical_session_max_order
                .entry(session_id)
                .or_insert(order);
            if order > *current_max_order {
                *current_max_order = order;
            }

            let item_sessions = historical_item_index
                .entry(item_id)
                .or_insert(HashSet::new());
            item_sessions.insert(session_id);
        }

        //println!("Ignored training events: {} ", ignored_training_rows);
        //println!("Sessions used: {}", valid_session_ids.len());
        //println!("Items used: {}", historical_item_index.len());

        let num_historical_sessions = historical_session_index.len();

        let item_idfs: HashMap<u64, f64> = historical_item_index
            .iter()
            .map(|(item, session_ids)| {
                let item_idf = idf(num_historical_sessions, session_ids.len());

                (*item, item_idf)
            })
            .collect();

        VSkNNIndex {
            session_index: historical_session_index,
            session_max_order: historical_session_max_order,
            item_index: historical_item_index,
            item_idfs,
        }
    }

    fn sessions_for_item(&self, item: &u64) -> Option<&HashSet<u32, RandomState>> {
        self.item_index.get(item) // move object and ownership to function that call us
    }

    fn max_order_for_session(&self, session: &TrainingSessionId) -> Time {
        self.session_max_order[session]
    }
}